Actual source code: aij.c
1: /*
2: Defines the basic matrix operations for the AIJ (compressed row)
3: matrix storage format.
4: */
6: #include <../src/mat/impls/aij/seq/aij.h>
7: #include <petscblaslapack.h>
8: #include <petscbt.h>
9: #include <petsc/private/kernels/blocktranspose.h>
11: /* defines MatSetValues_Seq_Hash(), MatAssemblyEnd_Seq_Hash(), MatSetUp_Seq_Hash() */
12: #define TYPE AIJ
13: #define TYPE_BS
14: #include "../src/mat/impls/aij/seq/seqhashmatsetvalues.h"
15: #include "../src/mat/impls/aij/seq/seqhashmat.h"
16: #undef TYPE
17: #undef TYPE_BS
19: MatGetDiagonalMarkers(SeqAIJ, 1)
21: static PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A)
22: {
23: PetscBool flg;
24: char type[256];
26: PetscFunctionBegin;
27: PetscObjectOptionsBegin((PetscObject)A);
28: PetscCall(PetscOptionsFList("-mat_seqaij_type", "Matrix SeqAIJ type", "MatSeqAIJSetType", MatSeqAIJList, "seqaij", type, sizeof(type), &flg));
29: if (flg) PetscCall(MatSeqAIJSetType(A, type));
30: PetscOptionsEnd();
31: PetscFunctionReturn(PETSC_SUCCESS);
32: }
34: static PetscErrorCode MatGetColumnReductions_SeqAIJ(Mat A, PetscInt type, PetscReal *reductions)
35: {
36: PetscInt i, m, n;
37: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
39: PetscFunctionBegin;
40: PetscCall(MatGetSize(A, &m, &n));
41: PetscCall(PetscArrayzero(reductions, n));
42: if (type == NORM_2) {
43: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i] * aij->a[i]);
44: } else if (type == NORM_1) {
45: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i]);
46: } else if (type == NORM_INFINITY) {
47: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]), reductions[aij->j[i]]);
48: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
49: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscRealPart(aij->a[i]);
50: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
51: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscImaginaryPart(aij->a[i]);
52: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
54: if (type == NORM_2) {
55: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
56: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
57: for (i = 0; i < n; i++) reductions[i] /= m;
58: }
59: PetscFunctionReturn(PETSC_SUCCESS);
60: }
62: static PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A, IS *is)
63: {
64: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
65: PetscInt i, m = A->rmap->n, cnt = 0, bs = A->rmap->bs;
66: const PetscInt *jj = a->j, *ii = a->i;
67: PetscInt *rows;
69: PetscFunctionBegin;
70: for (i = 0; i < m; i++) {
71: if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) cnt++;
72: }
73: PetscCall(PetscMalloc1(cnt, &rows));
74: cnt = 0;
75: for (i = 0; i < m; i++) {
76: if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) {
77: rows[cnt] = i;
78: cnt++;
79: }
80: }
81: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, is));
82: PetscFunctionReturn(PETSC_SUCCESS);
83: }
85: PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A, PetscInt *nrows, PetscInt **zrows)
86: {
87: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
88: const MatScalar *aa;
89: PetscInt i, m = A->rmap->n, cnt = 0;
90: const PetscInt *ii = a->i, *jj = a->j, *diag;
91: PetscInt *rows;
93: PetscFunctionBegin;
94: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
95: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
96: for (i = 0; i < m; i++) {
97: if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) cnt++;
98: }
99: PetscCall(PetscMalloc1(cnt, &rows));
100: cnt = 0;
101: for (i = 0; i < m; i++) {
102: if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) rows[cnt++] = i;
103: }
104: *nrows = cnt;
105: *zrows = rows;
106: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
107: PetscFunctionReturn(PETSC_SUCCESS);
108: }
110: static PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A, IS *zrows)
111: {
112: PetscInt nrows, *rows;
114: PetscFunctionBegin;
115: *zrows = NULL;
116: PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(A, &nrows, &rows));
117: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), nrows, rows, PETSC_OWN_POINTER, zrows));
118: PetscFunctionReturn(PETSC_SUCCESS);
119: }
121: static PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A, IS *keptrows)
122: {
123: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
124: const MatScalar *aa;
125: PetscInt m = A->rmap->n, cnt = 0;
126: const PetscInt *ii;
127: PetscInt n, i, j, *rows;
129: PetscFunctionBegin;
130: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
131: *keptrows = NULL;
132: ii = a->i;
133: for (i = 0; i < m; i++) {
134: n = ii[i + 1] - ii[i];
135: if (!n) {
136: cnt++;
137: goto ok1;
138: }
139: for (j = ii[i]; j < ii[i + 1]; j++) {
140: if (aa[j] != 0.0) goto ok1;
141: }
142: cnt++;
143: ok1:;
144: }
145: if (!cnt) {
146: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
147: PetscFunctionReturn(PETSC_SUCCESS);
148: }
149: PetscCall(PetscMalloc1(A->rmap->n - cnt, &rows));
150: cnt = 0;
151: for (i = 0; i < m; i++) {
152: n = ii[i + 1] - ii[i];
153: if (!n) continue;
154: for (j = ii[i]; j < ii[i + 1]; j++) {
155: if (aa[j] != 0.0) {
156: rows[cnt++] = i;
157: break;
158: }
159: }
160: }
161: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
162: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, keptrows));
163: PetscFunctionReturn(PETSC_SUCCESS);
164: }
166: PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y, Vec D, InsertMode is)
167: {
168: PetscInt i, m = Y->rmap->n;
169: const PetscInt *diag;
170: MatScalar *aa;
171: const PetscScalar *v;
172: PetscBool diagDense;
174: PetscFunctionBegin;
175: if (Y->assembled) {
176: PetscCall(MatGetDiagonalMarkers_SeqAIJ(Y, &diag, &diagDense));
177: if (diagDense) {
178: PetscCall(VecGetArrayRead(D, &v));
179: PetscCall(MatSeqAIJGetArray(Y, &aa));
180: if (is == INSERT_VALUES) {
181: for (i = 0; i < m; i++) aa[diag[i]] = v[i];
182: } else {
183: for (i = 0; i < m; i++) aa[diag[i]] += v[i];
184: }
185: PetscCall(MatSeqAIJRestoreArray(Y, &aa));
186: PetscCall(VecRestoreArrayRead(D, &v));
187: PetscFunctionReturn(PETSC_SUCCESS);
188: }
189: }
190: PetscCall(MatDiagonalSet_Default(Y, D, is));
191: PetscFunctionReturn(PETSC_SUCCESS);
192: }
194: PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
195: {
196: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
197: PetscInt i, ishift;
199: PetscFunctionBegin;
200: if (m) *m = A->rmap->n;
201: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
202: ishift = 0;
203: if (symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) {
204: PetscCall(MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, ishift, oshift, (PetscInt **)ia, (PetscInt **)ja));
205: } else if (oshift == 1) {
206: PetscInt *tia;
207: PetscInt nz = a->i[A->rmap->n];
209: /* malloc space and add 1 to i and j indices */
210: PetscCall(PetscMalloc1(A->rmap->n + 1, &tia));
211: for (i = 0; i < A->rmap->n + 1; i++) tia[i] = a->i[i] + 1;
212: *ia = tia;
213: if (ja) {
214: PetscInt *tja;
216: PetscCall(PetscMalloc1(nz + 1, &tja));
217: for (i = 0; i < nz; i++) tja[i] = a->j[i] + 1;
218: *ja = tja;
219: }
220: } else {
221: *ia = a->i;
222: if (ja) *ja = a->j;
223: }
224: PetscFunctionReturn(PETSC_SUCCESS);
225: }
227: PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
228: {
229: PetscFunctionBegin;
230: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
231: if ((symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) || oshift == 1) {
232: PetscCall(PetscFree(*ia));
233: if (ja) PetscCall(PetscFree(*ja));
234: }
235: PetscFunctionReturn(PETSC_SUCCESS);
236: }
238: PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
239: {
240: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
241: PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n;
242: PetscInt nz = a->i[m], row, *jj, mr, col;
244: PetscFunctionBegin;
245: *nn = n;
246: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
247: if (symmetric) {
248: PetscCall(MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, 0, oshift, (PetscInt **)ia, (PetscInt **)ja));
249: } else {
250: PetscCall(PetscCalloc1(n, &collengths));
251: PetscCall(PetscMalloc1(n + 1, &cia));
252: PetscCall(PetscMalloc1(nz, &cja));
253: jj = a->j;
254: for (i = 0; i < nz; i++) collengths[jj[i]]++;
255: cia[0] = oshift;
256: for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i];
257: PetscCall(PetscArrayzero(collengths, n));
258: jj = a->j;
259: for (row = 0; row < m; row++) {
260: mr = a->i[row + 1] - a->i[row];
261: for (i = 0; i < mr; i++) {
262: col = *jj++;
264: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
265: }
266: }
267: PetscCall(PetscFree(collengths));
268: *ia = cia;
269: *ja = cja;
270: }
271: PetscFunctionReturn(PETSC_SUCCESS);
272: }
274: PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
275: {
276: PetscFunctionBegin;
277: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
279: PetscCall(PetscFree(*ia));
280: PetscCall(PetscFree(*ja));
281: PetscFunctionReturn(PETSC_SUCCESS);
282: }
284: /*
285: MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from
286: MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output
287: spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ()
288: */
289: PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done)
290: {
291: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
292: PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n;
293: PetscInt nz = a->i[m], row, mr, col, tmp;
294: PetscInt *cspidx;
295: const PetscInt *jj;
297: PetscFunctionBegin;
298: *nn = n;
299: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
301: PetscCall(PetscCalloc1(n, &collengths));
302: PetscCall(PetscMalloc1(n + 1, &cia));
303: PetscCall(PetscMalloc1(nz, &cja));
304: PetscCall(PetscMalloc1(nz, &cspidx));
305: jj = a->j;
306: for (i = 0; i < nz; i++) collengths[jj[i]]++;
307: cia[0] = oshift;
308: for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i];
309: PetscCall(PetscArrayzero(collengths, n));
310: jj = a->j;
311: for (row = 0; row < m; row++) {
312: mr = a->i[row + 1] - a->i[row];
313: for (i = 0; i < mr; i++) {
314: col = *jj++;
315: tmp = cia[col] + collengths[col]++ - oshift;
316: cspidx[tmp] = a->i[row] + i; /* index of a->j */
317: cja[tmp] = row + oshift;
318: }
319: }
320: PetscCall(PetscFree(collengths));
321: *ia = cia;
322: *ja = cja;
323: *spidx = cspidx;
324: PetscFunctionReturn(PETSC_SUCCESS);
325: }
327: PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done)
328: {
329: PetscFunctionBegin;
330: PetscCall(MatRestoreColumnIJ_SeqAIJ(A, oshift, symmetric, inodecompressed, n, ia, ja, done));
331: PetscCall(PetscFree(*spidx));
332: PetscFunctionReturn(PETSC_SUCCESS);
333: }
335: static PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A, PetscInt row, const PetscScalar v[])
336: {
337: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
338: PetscInt *ai = a->i;
339: PetscScalar *aa;
341: PetscFunctionBegin;
342: PetscCall(MatSeqAIJGetArray(A, &aa));
343: PetscCall(PetscArraycpy(aa + ai[row], v, ai[row + 1] - ai[row]));
344: PetscCall(MatSeqAIJRestoreArray(A, &aa));
345: PetscFunctionReturn(PETSC_SUCCESS);
346: }
348: #include <petsc/private/isimpl.h>
350: /*@
351: MatSeqAIJSetValuesLocalFast - An optimized version of `MatSetValuesLocal()` for `MATSEQAIJ` matrices, valid under
352: several restrictive assumptions.
354: Not Collective
356: Input Parameters:
357: + A - the `MATSEQAIJ` matrix
358: . m - the number of rows being set (must be 1)
359: . im - array of length `m` giving the local row index
360: . n - the number of columns being set
361: . in - array of length `n` giving the local column indices
362: . v - array of length `n` of values to add
363: - is - the insert mode (must be `ADD_VALUES`)
365: Level: developer
367: Notes:
368: This routine requires that a single row of values is set with each call, that no row or column
369: index is negative or larger than the number of rows or columns, that values are always added
370: (not inserted), and that no new nonzero locations are introduced.
372: The global column indices are not assumed to be sorted.
374: .seealso: `Mat`, `MATSEQAIJ`, `MatSetValuesLocal()`, `MatSetValues()`
375: @*/
376: PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
377: {
378: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
379: PetscInt low, high, t, row, nrow, i, col, l;
380: const PetscInt *rp, *ai = a->i, *ailen = a->ilen, *aj = a->j;
381: PetscInt lastcol = -1;
382: MatScalar *ap, value, *aa;
383: const PetscInt *ridx = A->rmap->mapping->indices, *cidx = A->cmap->mapping->indices;
385: PetscFunctionBegin;
386: PetscCall(MatSeqAIJGetArray(A, &aa));
387: row = ridx[im[0]];
388: rp = aj + ai[row];
389: ap = aa + ai[row];
390: nrow = ailen[row];
391: low = 0;
392: high = nrow;
393: for (l = 0; l < n; l++) { /* loop over added columns */
394: col = cidx[in[l]];
395: value = v[l];
397: if (col <= lastcol) low = 0;
398: else high = nrow;
399: lastcol = col;
400: while (high - low > 5) {
401: t = (low + high) / 2;
402: if (rp[t] > col) high = t;
403: else low = t;
404: }
405: for (i = low; i < high; i++) {
406: if (rp[i] == col) {
407: ap[i] += value;
408: low = i + 1;
409: break;
410: }
411: }
412: }
413: PetscCall(MatSeqAIJRestoreArray(A, &aa));
414: return PETSC_SUCCESS;
415: }
417: PetscErrorCode MatSetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
418: {
419: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
420: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N;
421: PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen;
422: PetscInt *aj = a->j, nonew = a->nonew, lastcol = -1;
423: MatScalar *ap = NULL, value = 0.0, *aa;
424: PetscBool ignorezeroentries = a->ignorezeroentries;
425: PetscBool roworiented = a->roworiented;
427: PetscFunctionBegin;
428: PetscCall(MatSeqAIJGetArray(A, &aa));
429: for (k = 0; k < m; k++) { /* loop over added rows */
430: row = im[k];
431: if (row < 0) continue;
432: PetscCheck(row < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, row, A->rmap->n - 1);
433: rp = PetscSafePointerPlusOffset(aj, ai[row]);
434: if (!A->structure_only) ap = PetscSafePointerPlusOffset(aa, ai[row]);
435: rmax = imax[row];
436: nrow = ailen[row];
437: low = 0;
438: high = nrow;
439: for (l = 0; l < n; l++) { /* loop over added columns */
440: if (in[l] < 0) continue;
441: PetscCheck(in[l] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[l], A->cmap->n - 1);
442: col = in[l];
443: if (v && !A->structure_only) value = roworiented ? v[l + k * n] : v[k + l * m];
444: if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue;
446: if (col <= lastcol) low = 0;
447: else high = nrow;
448: lastcol = col;
449: while (high - low > 5) {
450: t = (low + high) / 2;
451: if (rp[t] > col) high = t;
452: else low = t;
453: }
454: for (i = low; i < high; i++) {
455: if (rp[i] > col) break;
456: if (rp[i] == col) {
457: if (!A->structure_only) {
458: if (is == ADD_VALUES) {
459: ap[i] += value;
460: (void)PetscLogFlops(1.0);
461: } else ap[i] = value;
462: }
463: low = i + 1;
464: goto noinsert;
465: }
466: }
467: if ((!A->structure_only && value == 0.0 && ignorezeroentries && row != col) || nonew == 1) goto noinsert;
468: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") in the matrix", row, col);
469: if (A->structure_only) {
470: MatSeqXAIJReallocateAIJ_structure_only(A, A->rmap->n, 1, nrow, row, col, rmax, ai, aj, rp, imax, nonew, MatScalar);
471: } else {
472: MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
473: }
474: N = nrow++ - 1;
475: a->nz++;
476: high++;
477: /* shift up all the later entries in this row */
478: PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1));
479: rp[i] = col;
480: if (!A->structure_only) {
481: PetscCall(PetscArraymove(ap + i + 1, ap + i, N - i + 1));
482: ap[i] = value;
483: }
484: low = i + 1;
485: noinsert:;
486: }
487: ailen[row] = nrow;
488: }
489: PetscCall(MatSeqAIJRestoreArray(A, &aa));
490: PetscFunctionReturn(PETSC_SUCCESS);
491: }
493: static PetscErrorCode MatSetValues_SeqAIJ_SortedFullNoPreallocation(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
494: {
495: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
496: PetscInt *rp, k, row;
497: PetscInt *ai = a->i;
498: PetscInt *aj = a->j;
499: MatScalar *aa, *ap;
501: PetscFunctionBegin;
502: PetscCheck(!A->was_assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot call on assembled matrix.");
503: 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);
505: PetscCall(MatSeqAIJGetArray(A, &aa));
506: for (k = 0; k < m; k++) { /* loop over added rows */
507: row = im[k];
508: rp = aj + ai[row];
509: ap = PetscSafePointerPlusOffset(aa, ai[row]);
511: PetscCall(PetscArraycpy(rp, in, n));
512: if (!A->structure_only) {
513: if (v) {
514: PetscCall(PetscArraycpy(ap, v, n));
515: v += n;
516: } else {
517: PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar)));
518: }
519: }
520: a->ilen[row] = n;
521: a->imax[row] = n;
522: a->i[row + 1] = a->i[row] + n;
523: a->nz += n;
524: }
525: PetscCall(MatSeqAIJRestoreArray(A, &aa));
526: PetscFunctionReturn(PETSC_SUCCESS);
527: }
529: /*@
530: MatSeqAIJSetTotalPreallocation - Sets an upper bound on the total number of expected nonzeros in the matrix.
532: Input Parameters:
533: + A - the `MATSEQAIJ` matrix
534: - nztotal - bound on the number of nonzeros
536: Level: advanced
538: Notes:
539: This can be called if you will be provided the matrix row by row (from row zero) with sorted column indices for each row.
540: Simply call `MatSetValues()` after this call to provide the matrix entries in the usual manner. This matrix may be used
541: as always with multiple matrix assemblies.
543: .seealso: [](ch_matrices), `Mat`, `MatSetOption()`, `MAT_SORTED_FULL`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`
544: @*/
545: PetscErrorCode MatSeqAIJSetTotalPreallocation(Mat A, PetscInt nztotal)
546: {
547: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
549: PetscFunctionBegin;
550: PetscCall(PetscLayoutSetUp(A->rmap));
551: PetscCall(PetscLayoutSetUp(A->cmap));
552: a->maxnz = nztotal;
553: if (!a->imax) PetscCall(PetscMalloc1(A->rmap->n, &a->imax));
554: if (!a->ilen) {
555: PetscCall(PetscMalloc1(A->rmap->n, &a->ilen));
556: } else {
557: PetscCall(PetscMemzero(a->ilen, A->rmap->n * sizeof(PetscInt)));
558: }
560: /* allocate the matrix space */
561: PetscCall(PetscShmgetAllocateArray(A->rmap->n + 1, sizeof(PetscInt), (void **)&a->i));
562: PetscCall(PetscShmgetAllocateArray(nztotal, sizeof(PetscInt), (void **)&a->j));
563: a->free_ij = PETSC_TRUE;
564: if (A->structure_only) {
565: a->free_a = PETSC_FALSE;
566: } else {
567: PetscCall(PetscShmgetAllocateArray(nztotal, sizeof(PetscScalar), (void **)&a->a));
568: a->free_a = PETSC_TRUE;
569: }
570: a->i[0] = 0;
571: A->ops->setvalues = MatSetValues_SeqAIJ_SortedFullNoPreallocation;
572: A->preallocated = PETSC_TRUE;
573: PetscFunctionReturn(PETSC_SUCCESS);
574: }
576: static PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
577: {
578: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
579: PetscInt *rp, k, row;
580: PetscInt *ai = a->i, *ailen = a->ilen;
581: PetscInt *aj = a->j;
582: MatScalar *aa, *ap;
584: PetscFunctionBegin;
585: PetscCall(MatSeqAIJGetArray(A, &aa));
586: for (k = 0; k < m; k++) { /* loop over added rows */
587: row = im[k];
588: 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);
589: rp = aj + ai[row];
590: ap = aa + ai[row];
591: if (!A->was_assembled) PetscCall(PetscArraycpy(rp, in, n));
592: if (!A->structure_only) {
593: if (v) {
594: PetscCall(PetscArraycpy(ap, v, n));
595: v += n;
596: } else {
597: PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar)));
598: }
599: }
600: ailen[row] = n;
601: a->nz += n;
602: }
603: PetscCall(MatSeqAIJRestoreArray(A, &aa));
604: PetscFunctionReturn(PETSC_SUCCESS);
605: }
607: static PetscErrorCode MatGetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], PetscScalar v[])
608: {
609: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
610: PetscInt *rp, k, low, high, t, row, nrow, i, col, l, *aj = a->j;
611: PetscInt *ai = a->i, *ailen = a->ilen;
612: const MatScalar *ap, *aa;
613: PetscBool hyprecoo;
614: PetscBool roworiented = a->roworiented;
615: PetscScalar *value;
617: PetscFunctionBegin;
618: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)A)->name, &hyprecoo));
620: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
621: for (k = 0; k < m; k++) { /* loop over rows */
622: row = im[k];
623: if (row < 0) continue; /* negative row */
624: 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);
625: rp = PetscSafePointerPlusOffset(aj, ai[row]);
626: ap = PetscSafePointerPlusOffset(aa, ai[row]);
627: nrow = ailen[row];
628: for (l = 0; l < n; l++) { /* loop over columns */
629: if (in[l] < 0) continue; /* negative column */
630: value = roworiented ? &v[l + k * n] : &v[k + l * m];
631: 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);
632: col = in[l];
633: /* hypre coo mat stores its diagonal at the front, out of sort */
634: if (hyprecoo) {
635: if (col == rp[0]) {
636: *value = ap[0];
637: goto finished;
638: }
639: low = 1;
640: } else low = 0;
641: high = nrow;
642: /* assume sorted */
643: while (high - low > 5) {
644: t = (low + high) / 2;
645: if (rp[t] > col) high = t;
646: else low = t;
647: }
648: for (i = low; i < high; i++) {
649: if (rp[i] > col) break;
650: if (rp[i] == col) {
651: *value = ap[i];
652: goto finished;
653: }
654: }
655: *value = 0.0;
656: finished:;
657: }
658: }
659: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
660: PetscFunctionReturn(PETSC_SUCCESS);
661: }
663: static PetscErrorCode MatView_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
664: {
665: Mat_SeqAIJ *A = (Mat_SeqAIJ *)mat->data;
666: const PetscScalar *av;
667: PetscInt header[4], M, N, m, nz, i;
668: PetscInt *rowlens;
670: PetscFunctionBegin;
671: PetscCall(PetscViewerSetUp(viewer));
673: M = mat->rmap->N;
674: N = mat->cmap->N;
675: m = mat->rmap->n;
676: nz = A->nz;
678: /* write matrix header */
679: header[0] = MAT_FILE_CLASSID;
680: header[1] = M;
681: header[2] = N;
682: header[3] = nz;
683: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
685: /* fill in and store row lengths */
686: PetscCall(PetscMalloc1(m, &rowlens));
687: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i];
688: if (PetscDefined(USE_DEBUG)) {
689: PetscInt mnz = 0;
691: for (i = 0; i < m; i++) mnz += rowlens[i];
692: PetscCheck(nz == mnz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Row lens %" PetscInt_FMT " do not sum to nz %" PetscInt_FMT, mnz, nz);
693: }
694: PetscCall(PetscViewerBinaryWrite(viewer, rowlens, m, PETSC_INT));
695: PetscCall(PetscFree(rowlens));
696: /* store column indices */
697: PetscCall(PetscViewerBinaryWrite(viewer, A->j, nz, PETSC_INT));
698: /* store nonzero values */
699: PetscCall(MatSeqAIJGetArrayRead(mat, &av));
700: PetscCall(PetscViewerBinaryWrite(viewer, av, nz, PETSC_SCALAR));
701: PetscCall(MatSeqAIJRestoreArrayRead(mat, &av));
703: /* write block size option to the viewer's .info file */
704: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
705: PetscFunctionReturn(PETSC_SUCCESS);
706: }
708: static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A, PetscViewer viewer)
709: {
710: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
711: PetscInt i, k, m = A->rmap->N;
713: PetscFunctionBegin;
714: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
715: for (i = 0; i < m; i++) {
716: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
717: for (k = a->i[i]; k < a->i[i + 1]; k++) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ") ", a->j[k]));
718: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
719: }
720: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
721: PetscFunctionReturn(PETSC_SUCCESS);
722: }
724: static PetscErrorCode MatView_SeqAIJ_ASCII(Mat A, PetscViewer viewer)
725: {
726: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
727: const PetscScalar *av;
728: PetscInt i, j, m = A->rmap->n;
729: const char *name;
730: PetscViewerFormat format;
732: PetscFunctionBegin;
733: if (A->structure_only) {
734: PetscCall(MatView_SeqAIJ_ASCII_structonly(A, viewer));
735: PetscFunctionReturn(PETSC_SUCCESS);
736: }
738: PetscCall(PetscViewerGetFormat(viewer, &format));
739: // By petsc's rule, even PETSC_VIEWER_ASCII_INFO_DETAIL doesn't print matrix entries
740: if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) PetscFunctionReturn(PETSC_SUCCESS);
742: /* trigger copy to CPU if needed */
743: PetscCall(MatSeqAIJGetArrayRead(A, &av));
744: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
745: if (format == PETSC_VIEWER_ASCII_MATLAB) {
746: PetscInt nofinalvalue = 0;
747: if (m && ((a->i[m] == a->i[m - 1]) || (a->j[a->nz - 1] != A->cmap->n - 1))) {
748: /* Need a dummy value to ensure the dimension of the matrix. */
749: nofinalvalue = 1;
750: }
751: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
752: PetscCall(PetscViewerASCIIPrintf(viewer, "%% Size = %" PetscInt_FMT " %" PetscInt_FMT " \n", m, A->cmap->n));
753: PetscCall(PetscViewerASCIIPrintf(viewer, "%% Nonzeros = %" PetscInt_FMT " \n", a->nz));
754: PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",%d);\n", a->nz + nofinalvalue, PetscDefined(USE_COMPLEX) ? 4 : 3));
755: PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = [\n"));
757: for (i = 0; i < m; i++) {
758: for (j = a->i[i]; j < a->i[i + 1]; j++) {
759: #if PetscDefined(USE_COMPLEX)
760: 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])));
761: #else
762: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", i + 1, a->j[j] + 1, (double)a->a[j]));
763: #endif
764: }
765: }
766: if (nofinalvalue) {
767: if (PetscDefined(USE_COMPLEX)) PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e %18.16e\n", m, A->cmap->n, 0., 0.));
768: else PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", m, A->cmap->n, 0.0));
769: }
770: PetscCall(PetscObjectGetName((PetscObject)A, &name));
771: PetscCall(PetscViewerASCIIPrintf(viewer, "];\n %s = spconvert(zzz);\n", name));
772: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
773: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
774: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
775: for (i = 0; i < m; i++) {
776: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
777: for (j = a->i[i]; j < a->i[i + 1]; j++) {
778: #if PetscDefined(USE_COMPLEX)
779: if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) {
780: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
781: } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) {
782: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j])));
783: } else if (PetscRealPart(a->a[j]) != 0.0) {
784: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
785: }
786: #else
787: if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]));
788: #endif
789: }
790: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
791: }
792: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
793: } else if (format == PETSC_VIEWER_ASCII_SYMMODU) {
794: PetscInt nzd = 0, fshift = 1, *sptr;
795: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
796: PetscCall(PetscMalloc1(m + 1, &sptr));
797: for (i = 0; i < m; i++) {
798: sptr[i] = nzd + 1;
799: for (j = a->i[i]; j < a->i[i + 1]; j++) {
800: if (a->j[j] >= i) {
801: if (PetscDefined(USE_COMPLEX) ? (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) : a->a[j] != 0.0) nzd++;
802: }
803: }
804: }
805: sptr[m] = nzd + 1;
806: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n\n", m, nzd));
807: for (i = 0; i < m + 1; i += 6) {
808: if (i + 4 < m) {
809: 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]));
810: } else if (i + 3 < m) {
811: 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]));
812: } else if (i + 2 < m) {
813: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3]));
814: } else if (i + 1 < m) {
815: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2]));
816: } else if (i < m) {
817: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1]));
818: } else {
819: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT "\n", sptr[i]));
820: }
821: }
822: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
823: PetscCall(PetscFree(sptr));
824: for (i = 0; i < m; i++) {
825: for (j = a->i[i]; j < a->i[i + 1]; j++) {
826: if (a->j[j] >= i) PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " ", a->j[j] + fshift));
827: }
828: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
829: }
830: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
831: for (i = 0; i < m; i++) {
832: for (j = a->i[i]; j < a->i[i + 1]; j++) {
833: if (a->j[j] >= i) {
834: #if PetscDefined(USE_COMPLEX)
835: 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])));
836: #else
837: if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " %18.16e ", (double)a->a[j]));
838: #endif
839: }
840: }
841: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
842: }
843: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
844: } else if (format == PETSC_VIEWER_ASCII_DENSE) {
845: PetscInt cnt = 0, jcnt;
846: PetscScalar value;
847: PetscBool realonly = PETSC_TRUE;
849: if (PetscDefined(USE_COMPLEX)) {
850: for (i = 0; i < a->i[m]; i++) {
851: if (PetscImaginaryPart(a->a[i]) != 0.0) {
852: realonly = PETSC_FALSE;
853: break;
854: }
855: }
856: }
858: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
859: for (i = 0; i < m; i++) {
860: jcnt = 0;
861: for (j = 0; j < A->cmap->n; j++) {
862: if (jcnt < a->i[i + 1] - a->i[i] && j == a->j[cnt]) {
863: value = a->a[cnt++];
864: jcnt++;
865: } else {
866: value = 0.0;
867: }
868: if (!PetscDefined(USE_COMPLEX) || realonly) {
869: PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)PetscRealPart(value)));
870: } else {
871: PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e+%7.5e i ", (double)PetscRealPart(value), (double)PetscImaginaryPart(value)));
872: }
873: }
874: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
875: }
876: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
877: } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) {
878: PetscInt fshift = 1;
879: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
880: PetscCall(PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate %s general\n", PetscDefined(USE_COMPLEX) ? "complex" : "real"));
881: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", m, A->cmap->n, a->nz));
882: for (i = 0; i < m; i++) {
883: for (j = a->i[i]; j < a->i[i + 1]; j++) {
884: #if PetscDefined(USE_COMPLEX)
885: 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])));
886: #else
887: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %g\n", i + fshift, a->j[j] + fshift, (double)a->a[j]));
888: #endif
889: }
890: }
891: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
892: } else {
893: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
894: if (A->factortype) {
895: const PetscInt *adiag;
897: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &adiag, NULL));
898: for (i = 0; i < m; i++) {
899: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
900: /* L part */
901: for (j = a->i[i]; j < a->i[i + 1]; j++) {
902: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) > 0.0) {
903: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
904: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) < 0.0) {
905: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j]))));
906: } else {
907: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
908: }
909: }
910: /* diagonal */
911: j = adiag[i];
912: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) > 0.0) {
913: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(1 / a->a[j]), (double)PetscImaginaryPart(1 / a->a[j])));
914: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) < 0.0) {
915: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(1 / a->a[j]), (double)(-PetscImaginaryPart(1 / a->a[j]))));
916: } else {
917: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(1 / a->a[j])));
918: }
920: /* U part */
921: for (j = adiag[i + 1] + 1; j < adiag[i]; j++) {
922: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) > 0.0) {
923: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
924: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) < 0.0) {
925: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j]))));
926: } else {
927: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
928: }
929: }
930: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
931: }
932: } else {
933: for (i = 0; i < m; i++) {
934: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
935: for (j = a->i[i]; j < a->i[i + 1]; j++) {
936: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) > 0.0) {
937: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
938: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) < 0.0) {
939: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j])));
940: } else {
941: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
942: }
943: }
944: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
945: }
946: }
947: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
948: }
949: PetscCall(PetscViewerFlush(viewer));
950: PetscFunctionReturn(PETSC_SUCCESS);
951: }
953: #include <petscdraw.h>
954: #if defined(__GNUC__) && !defined(__clang__)
955: #pragma GCC diagnostic push
956: #pragma GCC diagnostic ignored "-Wclobbered"
957: #endif
958: static PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw, void *Aa)
959: {
960: Mat A = (Mat)Aa;
961: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
962: PetscInt i, j, m = A->rmap->n;
963: int color;
964: PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r;
965: PetscViewer viewer;
966: PetscViewerFormat format;
967: const PetscScalar *aa;
969: PetscFunctionBegin;
970: PetscCall(PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer));
971: PetscCall(PetscViewerGetFormat(viewer, &format));
972: PetscCall(PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr));
974: /* loop over matrix elements drawing boxes */
975: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
976: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
977: PetscDrawCollectiveBegin(draw);
978: /* Blue for negative, Cyan for zero and Red for positive */
979: color = PETSC_DRAW_BLUE;
980: for (i = 0; i < m; i++) {
981: y_l = m - i - 1.0;
982: y_r = y_l + 1.0;
983: for (j = a->i[i]; j < a->i[i + 1]; j++) {
984: x_l = a->j[j];
985: x_r = x_l + 1.0;
986: if (PetscRealPart(aa[j]) >= 0.) continue;
987: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
988: }
989: }
990: color = PETSC_DRAW_CYAN;
991: for (i = 0; i < m; i++) {
992: y_l = m - i - 1.0;
993: y_r = y_l + 1.0;
994: for (j = a->i[i]; j < a->i[i + 1]; j++) {
995: x_l = a->j[j];
996: x_r = x_l + 1.0;
997: if (aa[j] != 0.) continue;
998: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
999: }
1000: }
1001: color = PETSC_DRAW_RED;
1002: for (i = 0; i < m; i++) {
1003: y_l = m - i - 1.0;
1004: y_r = y_l + 1.0;
1005: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1006: x_l = a->j[j];
1007: x_r = x_l + 1.0;
1008: if (PetscRealPart(aa[j]) <= 0.) continue;
1009: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1010: }
1011: }
1012: PetscDrawCollectiveEnd(draw);
1013: } else {
1014: /* use contour shading to indicate magnitude of values */
1015: /* first determine max of all nonzero values */
1016: PetscReal minv = 0.0, maxv = 0.0;
1017: PetscInt nz = a->nz, count = 0;
1018: PetscDraw popup;
1020: for (i = 0; i < nz; i++) {
1021: if (PetscAbsScalar(aa[i]) > maxv) maxv = PetscAbsScalar(aa[i]);
1022: }
1023: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1024: PetscCall(PetscDrawGetPopup(draw, &popup));
1025: PetscCall(PetscDrawScalePopup(popup, minv, maxv));
1027: PetscDrawCollectiveBegin(draw);
1028: for (i = 0; i < m; i++) {
1029: y_l = m - i - 1.0;
1030: y_r = y_l + 1.0;
1031: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1032: x_l = a->j[j];
1033: x_r = x_l + 1.0;
1034: color = PetscDrawRealToColor(PetscAbsScalar(aa[count]), minv, maxv);
1035: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1036: count++;
1037: }
1038: }
1039: PetscDrawCollectiveEnd(draw);
1040: }
1041: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1042: PetscFunctionReturn(PETSC_SUCCESS);
1043: }
1044: #if defined(__GNUC__) && !defined(__clang__)
1045: #pragma GCC diagnostic pop
1046: #endif
1048: #include <petscdraw.h>
1049: static PetscErrorCode MatView_SeqAIJ_Draw(Mat A, PetscViewer viewer)
1050: {
1051: PetscDraw draw;
1052: PetscReal xr, yr, xl, yl, h, w;
1053: PetscBool isnull;
1055: PetscFunctionBegin;
1056: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1057: PetscCall(PetscDrawIsNull(draw, &isnull));
1058: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1060: xr = A->cmap->n;
1061: yr = A->rmap->n;
1062: h = yr / 10.0;
1063: w = xr / 10.0;
1064: xr += w;
1065: yr += h;
1066: xl = -w;
1067: yl = -h;
1068: PetscCall(PetscDrawSetCoordinates(draw, xl, yl, xr, yr));
1069: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer));
1070: PetscCall(PetscDrawZoom(draw, MatView_SeqAIJ_Draw_Zoom, A));
1071: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL));
1072: PetscCall(PetscDrawSave(draw));
1073: PetscFunctionReturn(PETSC_SUCCESS);
1074: }
1076: PetscErrorCode MatView_SeqAIJ(Mat A, PetscViewer viewer)
1077: {
1078: PetscBool isascii, isbinary, isdraw;
1080: PetscFunctionBegin;
1081: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1082: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1083: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1084: if (isascii) PetscCall(MatView_SeqAIJ_ASCII(A, viewer));
1085: else if (isbinary) PetscCall(MatView_SeqAIJ_Binary(A, viewer));
1086: else if (isdraw) PetscCall(MatView_SeqAIJ_Draw(A, viewer));
1087: PetscCall(MatView_SeqAIJ_Inode(A, viewer));
1088: PetscFunctionReturn(PETSC_SUCCESS);
1089: }
1091: PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A, MatAssemblyType mode)
1092: {
1093: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1094: PetscInt fshift = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax;
1095: PetscInt m = A->rmap->n, *ip, N, *ailen = a->ilen, rmax = 0;
1096: MatScalar *aa = a->a, *ap;
1097: PetscReal ratio = 0.6;
1099: PetscFunctionBegin;
1100: if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(PETSC_SUCCESS);
1101: if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) {
1102: /* 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) */
1103: PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode)); /* read the sparsity pattern */
1104: PetscFunctionReturn(PETSC_SUCCESS);
1105: }
1107: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
1108: for (i = 1; i < m; i++) {
1109: /* move each row back by the amount of empty slots (fshift) before it*/
1110: fshift += imax[i - 1] - ailen[i - 1];
1111: rmax = PetscMax(rmax, ailen[i]);
1112: if (fshift) {
1113: ip = aj + ai[i];
1114: N = ailen[i];
1115: PetscCall(PetscArraymove(ip - fshift, ip, N));
1116: if (!A->structure_only) {
1117: ap = aa + ai[i];
1118: PetscCall(PetscArraymove(ap - fshift, ap, N));
1119: }
1120: }
1121: ai[i] = ai[i - 1] + ailen[i - 1];
1122: }
1123: if (m) {
1124: fshift += imax[m - 1] - ailen[m - 1];
1125: ai[m] = ai[m - 1] + ailen[m - 1];
1126: }
1127: /* reset ilen and imax for each row */
1128: a->nonzerorowcnt = 0;
1129: for (i = 0; i < m; i++) {
1130: ailen[i] = imax[i] = ai[i + 1] - ai[i];
1131: a->nonzerorowcnt += ((ai[i + 1] - ai[i]) > 0);
1132: }
1133: a->nz = ai[m];
1134: 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);
1135: 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));
1136: PetscCall(PetscInfo(A, "Number of mallocs during MatSetValues() is %" PetscInt_FMT "\n", a->reallocs));
1137: PetscCall(PetscInfo(A, "Maximum nonzeros in any row is %" PetscInt_FMT "\n", rmax));
1139: A->info.mallocs += a->reallocs;
1140: a->reallocs = 0;
1141: A->info.nz_unneeded = (PetscReal)fshift;
1142: a->rmax = rmax;
1144: if (!A->structure_only) PetscCall(MatCheckCompressedRow(A, a->nonzerorowcnt, &a->compressedrow, a->i, m, ratio));
1145: PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode));
1146: PetscFunctionReturn(PETSC_SUCCESS);
1147: }
1149: static PetscErrorCode MatRealPart_SeqAIJ(Mat A)
1150: {
1151: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1152: PetscInt i, nz = a->nz;
1153: MatScalar *aa;
1155: PetscFunctionBegin;
1156: PetscCall(MatSeqAIJGetArray(A, &aa));
1157: for (i = 0; i < nz; i++) aa[i] = PetscRealPart(aa[i]);
1158: PetscCall(MatSeqAIJRestoreArray(A, &aa));
1159: PetscFunctionReturn(PETSC_SUCCESS);
1160: }
1162: static PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A)
1163: {
1164: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1165: PetscInt i, nz = a->nz;
1166: MatScalar *aa;
1168: PetscFunctionBegin;
1169: PetscCall(MatSeqAIJGetArray(A, &aa));
1170: for (i = 0; i < nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
1171: PetscCall(MatSeqAIJRestoreArray(A, &aa));
1172: PetscFunctionReturn(PETSC_SUCCESS);
1173: }
1175: PetscErrorCode MatZeroEntries_SeqAIJ(Mat A)
1176: {
1177: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1178: MatScalar *aa;
1180: PetscFunctionBegin;
1181: PetscCall(MatSeqAIJGetArrayWrite(A, &aa));
1182: PetscCall(PetscArrayzero(aa, a->i[A->rmap->n]));
1183: PetscCall(MatSeqAIJRestoreArrayWrite(A, &aa));
1184: PetscFunctionReturn(PETSC_SUCCESS);
1185: }
1187: static PetscErrorCode MatReset_SeqAIJ(Mat A)
1188: {
1189: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1191: PetscFunctionBegin;
1192: if (A->hash_active) {
1193: A->ops[0] = a->cops;
1194: PetscCall(PetscHMapIJVDestroy(&a->ht));
1195: PetscCall(PetscFree(a->dnz));
1196: A->hash_active = PETSC_FALSE;
1197: }
1199: PetscCall(PetscLogObjectState((PetscObject)A, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT ", NZ=%" PetscInt_FMT, A->rmap->n, A->cmap->n, a->nz));
1200: PetscCall(MatSeqXAIJFreeAIJ(A, &a->a, &a->j, &a->i));
1201: PetscCall(ISDestroy(&a->row));
1202: PetscCall(ISDestroy(&a->col));
1203: PetscCall(PetscFree(a->diag));
1204: PetscCall(PetscFree(a->ibdiag));
1205: a->ibdiagsize = 0;
1206: PetscCall(PetscFree(a->imax));
1207: PetscCall(PetscFree(a->ilen));
1208: PetscCall(PetscFree(a->ipre));
1209: PetscCall(PetscFree3(a->idiag, a->mdiag, a->ssor_work));
1210: PetscCall(PetscFree(a->solve_work));
1211: PetscCall(ISDestroy(&a->icol));
1212: PetscCall(PetscFree(a->saved_values));
1213: a->compressedrow.use = PETSC_FALSE;
1214: PetscCall(PetscFree2(a->compressedrow.i, a->compressedrow.rindex));
1215: PetscCall(MatDestroy_SeqAIJ_Inode(A));
1216: PetscFunctionReturn(PETSC_SUCCESS);
1217: }
1219: static PetscErrorCode MatResetHash_SeqAIJ(Mat A)
1220: {
1221: PetscFunctionBegin;
1222: PetscCall(MatReset_SeqAIJ(A));
1223: PetscCall(MatCreate_SeqAIJ_Inode(A));
1224: PetscCall(MatSetUp_Seq_Hash(A));
1225: A->nonzerostate++;
1226: PetscFunctionReturn(PETSC_SUCCESS);
1227: }
1229: PetscErrorCode MatDestroy_SeqAIJ(Mat A)
1230: {
1231: PetscFunctionBegin;
1232: PetscCall(MatReset_SeqAIJ(A));
1233: PetscCall(PetscFree(A->data));
1235: /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this.
1236: That function is so heavily used (sometimes in an hidden way through multnumeric function pointers)
1237: that is hard to properly add this data to the MatProduct data. We free it here to avoid
1238: users reusing the matrix object with different data to incur in obscure segmentation faults
1239: due to different matrix sizes */
1240: PetscCall(PetscObjectCompose((PetscObject)A, "__PETSc__ab_dense", NULL));
1242: PetscCall(PetscObjectChangeTypeName((PetscObject)A, NULL));
1243: PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEnginePut_C", NULL));
1244: PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEngineGet_C", NULL));
1245: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetColumnIndices_C", NULL));
1246: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatStoreValues_C", NULL));
1247: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatRetrieveValues_C", NULL));
1248: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsbaij_C", NULL));
1249: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqbaij_C", NULL));
1250: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijperm_C", NULL));
1251: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijsell_C", NULL));
1252: #if PetscDefined(HAVE_MKL_SPARSE)
1253: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijmkl_C", NULL));
1254: #endif
1255: #if PetscDefined(HAVE_CUDA)
1256: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcusparse_C", NULL));
1257: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", NULL));
1258: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", NULL));
1259: #endif
1260: #if PetscDefined(HAVE_HIP)
1261: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijhipsparse_C", NULL));
1262: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", NULL));
1263: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", NULL));
1264: #endif
1265: #if PetscDefined(HAVE_KOKKOS_KERNELS)
1266: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijkokkos_C", NULL));
1267: #endif
1268: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcrl_C", NULL));
1269: #if PetscDefined(HAVE_ELEMENTAL)
1270: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_elemental_C", NULL));
1271: #endif
1272: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
1273: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_scalapack_C", NULL));
1274: #endif
1275: #if PetscDefined(HAVE_HYPRE)
1276: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_hypre_C", NULL));
1277: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", NULL));
1278: #endif
1279: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqdense_C", NULL));
1280: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsell_C", NULL));
1281: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_is_C", NULL));
1282: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsTranspose_C", NULL));
1283: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsHermitianTranspose_C", NULL));
1284: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocation_C", NULL));
1285: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetPreallocation_C", NULL));
1286: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetHash_C", NULL));
1287: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocationCSR_C", NULL));
1288: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatReorderForNonzeroDiagonal_C", NULL));
1289: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_is_seqaij_C", NULL));
1290: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqdense_seqaij_C", NULL));
1291: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaij_C", NULL));
1292: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJKron_C", NULL));
1293: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
1294: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
1295: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
1296: /* these calls do not belong here: the subclasses Duplicate/Destroy are wrong */
1297: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijsell_seqaij_C", NULL));
1298: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijperm_seqaij_C", NULL));
1299: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijviennacl_C", NULL));
1300: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqdense_C", NULL));
1301: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqaij_C", NULL));
1302: PetscFunctionReturn(PETSC_SUCCESS);
1303: }
1305: PetscErrorCode MatSetOption_SeqAIJ(Mat A, MatOption op, PetscBool flg)
1306: {
1307: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1309: PetscFunctionBegin;
1310: switch (op) {
1311: case MAT_ROW_ORIENTED:
1312: a->roworiented = flg;
1313: break;
1314: case MAT_KEEP_NONZERO_PATTERN:
1315: a->keepnonzeropattern = flg;
1316: break;
1317: case MAT_NEW_NONZERO_LOCATIONS:
1318: a->nonew = (flg ? 0 : 1);
1319: break;
1320: case MAT_NEW_NONZERO_LOCATION_ERR:
1321: a->nonew = (flg ? -1 : 0);
1322: break;
1323: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1324: a->nonew = (flg ? -2 : 0);
1325: break;
1326: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1327: a->nounused = (flg ? -1 : 0);
1328: break;
1329: case MAT_IGNORE_ZERO_ENTRIES:
1330: a->ignorezeroentries = flg;
1331: break;
1332: case MAT_USE_INODES:
1333: PetscCall(MatSetOption_SeqAIJ_Inode(A, MAT_USE_INODES, flg));
1334: break;
1335: case MAT_SUBMAT_SINGLEIS:
1336: A->submat_singleis = flg;
1337: break;
1338: case MAT_SORTED_FULL:
1339: if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
1340: else A->ops->setvalues = MatSetValues_SeqAIJ;
1341: break;
1342: case MAT_FORM_EXPLICIT_TRANSPOSE:
1343: A->form_explicit_transpose = flg;
1344: break;
1345: case MAT_STRUCTURE_ONLY:
1346: if (flg) {
1347: PetscCall(MatXAIJDeallocatea(A, &a->a));
1348: a->a = NULL;
1349: }
1350: break;
1351: default:
1352: break;
1353: }
1354: PetscFunctionReturn(PETSC_SUCCESS);
1355: }
1357: PETSC_INTERN PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A, Vec v)
1358: {
1359: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1360: PetscInt n, *ai = a->i;
1361: PetscScalar *x;
1362: const PetscScalar *aa;
1363: const PetscInt *diag;
1364: PetscBool diagDense;
1366: PetscFunctionBegin;
1367: PetscCall(VecGetLocalSize(v, &n));
1368: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
1369: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1370: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1371: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
1372: PetscCall(VecGetArrayWrite(v, &x));
1373: for (PetscInt i = 0; i < n; i++) x[i] = 1.0 / aa[diag[i]];
1374: PetscCall(VecRestoreArrayWrite(v, &x));
1375: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1376: PetscFunctionReturn(PETSC_SUCCESS);
1377: }
1379: PetscCheck(A->factortype == MAT_FACTOR_NONE, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Not for factor matrices that are not ILU or LU");
1380: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
1381: PetscCall(VecGetArrayWrite(v, &x));
1382: if (diagDense) {
1383: for (PetscInt i = 0; i < n; i++) x[i] = aa[diag[i]];
1384: } else {
1385: for (PetscInt i = 0; i < n; i++) x[i] = (diag[i] == ai[i + 1]) ? 0.0 : aa[diag[i]];
1386: }
1387: PetscCall(VecRestoreArrayWrite(v, &x));
1388: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1389: PetscFunctionReturn(PETSC_SUCCESS);
1390: }
1392: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1393: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A, Vec xx, Vec zz, Vec yy)
1394: {
1395: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1396: const MatScalar *aa;
1397: PetscScalar *y;
1398: const PetscScalar *x;
1399: PetscInt m = A->rmap->n;
1400: #if !PetscDefined(USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1401: const MatScalar *v;
1402: PetscScalar alpha;
1403: PetscInt n, i, j;
1404: const PetscInt *idx, *ii, *ridx = NULL;
1405: Mat_CompressedRow cprow = a->compressedrow;
1406: PetscBool usecprow = cprow.use;
1407: #endif
1409: PetscFunctionBegin;
1410: if (zz != yy) PetscCall(VecCopy(zz, yy));
1411: PetscCall(VecGetArrayRead(xx, &x));
1412: PetscCall(VecGetArray(yy, &y));
1413: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1415: #if PetscDefined(USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1416: fortranmulttransposeaddaij_(&m, x, a->i, a->j, aa, y);
1417: #else
1418: if (usecprow) {
1419: m = cprow.nrows;
1420: ii = cprow.i;
1421: ridx = cprow.rindex;
1422: } else {
1423: ii = a->i;
1424: }
1425: for (i = 0; i < m; i++) {
1426: idx = a->j + ii[i];
1427: v = aa + ii[i];
1428: n = ii[i + 1] - ii[i];
1429: if (usecprow) {
1430: alpha = x[ridx[i]];
1431: } else {
1432: alpha = x[i];
1433: }
1434: for (j = 0; j < n; j++) y[idx[j]] += alpha * v[j];
1435: }
1436: #endif
1437: PetscCall(PetscLogFlops(2.0 * a->nz));
1438: PetscCall(VecRestoreArrayRead(xx, &x));
1439: PetscCall(VecRestoreArray(yy, &y));
1440: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1441: PetscFunctionReturn(PETSC_SUCCESS);
1442: }
1444: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A, Vec xx, Vec yy)
1445: {
1446: PetscFunctionBegin;
1447: PetscCall(VecSet(yy, 0.0));
1448: PetscCall(MatMultTransposeAdd_SeqAIJ(A, xx, yy, yy));
1449: PetscFunctionReturn(PETSC_SUCCESS);
1450: }
1452: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1454: PetscErrorCode MatMult_SeqAIJ(Mat A, Vec xx, Vec yy)
1455: {
1456: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1457: PetscScalar *y;
1458: const PetscScalar *x;
1459: const MatScalar *a_a;
1460: PetscInt m = A->rmap->n;
1461: const PetscInt *ii, *ridx = NULL;
1462: PetscBool usecprow = a->compressedrow.use;
1464: #if PetscDefined(HAVE_PRAGMA_DISJOINT)
1465: #pragma disjoint(*x, *y, *aa)
1466: #endif
1468: PetscFunctionBegin;
1469: if (a->inode.use && a->inode.checked) {
1470: PetscCall(MatMult_SeqAIJ_Inode(A, xx, yy));
1471: PetscFunctionReturn(PETSC_SUCCESS);
1472: }
1473: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1474: PetscCall(VecGetArrayRead(xx, &x));
1475: PetscCall(VecGetArray(yy, &y));
1476: ii = a->i;
1477: if (usecprow) { /* use compressed row format */
1478: PetscCall(PetscArrayzero(y, m));
1479: m = a->compressedrow.nrows;
1480: ii = a->compressedrow.i;
1481: ridx = a->compressedrow.rindex;
1482: PetscPragmaUseOMPKernels(parallel for)
1483: for (PetscInt i = 0; i < m; i++) {
1484: PetscInt n = ii[i + 1] - ii[i];
1485: const PetscInt *aj = a->j + ii[i];
1486: const PetscScalar *aa = a_a + ii[i];
1487: PetscScalar sum = 0.0;
1488: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1489: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1490: y[ridx[i]] = sum;
1491: }
1492: } else { /* do not use compressed row format */
1493: #if PetscDefined(USE_FORTRAN_KERNEL_MULTAIJ)
1494: fortranmultaij_(&m, x, ii, a->j, a_a, y);
1495: #else
1496: PetscPragmaUseOMPKernels(parallel for)
1497: for (PetscInt i = 0; i < m; i++) {
1498: PetscInt n = ii[i + 1] - ii[i];
1499: const PetscInt *aj = a->j + ii[i];
1500: const PetscScalar *aa = a_a + ii[i];
1501: PetscScalar sum = 0.0;
1502: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1503: y[i] = sum;
1504: }
1505: #endif
1506: }
1507: PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt));
1508: PetscCall(VecRestoreArrayRead(xx, &x));
1509: PetscCall(VecRestoreArray(yy, &y));
1510: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1511: PetscFunctionReturn(PETSC_SUCCESS);
1512: }
1514: // HACK!!!!! Used by src/mat/tests/ex170.c
1515: PETSC_EXTERN PetscErrorCode MatMultMax_SeqAIJ(Mat A, Vec xx, Vec yy)
1516: {
1517: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1518: PetscScalar *y;
1519: const PetscScalar *x;
1520: const MatScalar *aa, *a_a;
1521: PetscInt m = A->rmap->n;
1522: const PetscInt *aj, *ii, *ridx = NULL;
1523: PetscInt n, i, nonzerorow = 0;
1524: PetscScalar sum;
1525: PetscBool usecprow = a->compressedrow.use;
1527: #if PetscDefined(HAVE_PRAGMA_DISJOINT)
1528: #pragma disjoint(*x, *y, *aa)
1529: #endif
1531: PetscFunctionBegin;
1532: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1533: PetscCall(VecGetArrayRead(xx, &x));
1534: PetscCall(VecGetArray(yy, &y));
1535: if (usecprow) { /* use compressed row format */
1536: m = a->compressedrow.nrows;
1537: ii = a->compressedrow.i;
1538: ridx = a->compressedrow.rindex;
1539: for (i = 0; i < m; i++) {
1540: n = ii[i + 1] - ii[i];
1541: aj = a->j + ii[i];
1542: aa = a_a + ii[i];
1543: sum = 0.0;
1544: nonzerorow += (n > 0);
1545: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1546: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1547: y[*ridx++] = sum;
1548: }
1549: } else { /* do not use compressed row format */
1550: ii = a->i;
1551: for (i = 0; i < m; i++) {
1552: n = ii[i + 1] - ii[i];
1553: aj = a->j + ii[i];
1554: aa = a_a + ii[i];
1555: sum = 0.0;
1556: nonzerorow += (n > 0);
1557: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1558: y[i] = sum;
1559: }
1560: }
1561: PetscCall(PetscLogFlops(2.0 * a->nz - nonzerorow));
1562: PetscCall(VecRestoreArrayRead(xx, &x));
1563: PetscCall(VecRestoreArray(yy, &y));
1564: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1565: PetscFunctionReturn(PETSC_SUCCESS);
1566: }
1568: // HACK!!!!! Used by src/mat/tests/ex170.c
1569: PETSC_EXTERN PetscErrorCode MatMultAddMax_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1570: {
1571: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1572: PetscScalar *y, *z;
1573: const PetscScalar *x;
1574: const MatScalar *aa, *a_a;
1575: PetscInt m = A->rmap->n, *aj, *ii;
1576: PetscInt n, i, *ridx = NULL;
1577: PetscScalar sum;
1578: PetscBool usecprow = a->compressedrow.use;
1580: PetscFunctionBegin;
1581: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1582: PetscCall(VecGetArrayRead(xx, &x));
1583: PetscCall(VecGetArrayPair(yy, zz, &y, &z));
1584: if (usecprow) { /* use compressed row format */
1585: if (zz != yy) PetscCall(PetscArraycpy(z, y, m));
1586: m = a->compressedrow.nrows;
1587: ii = a->compressedrow.i;
1588: ridx = a->compressedrow.rindex;
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[*ridx];
1594: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1595: z[*ridx++] = sum;
1596: }
1597: } else { /* do not use compressed row format */
1598: ii = a->i;
1599: for (i = 0; i < m; i++) {
1600: n = ii[i + 1] - ii[i];
1601: aj = a->j + ii[i];
1602: aa = a_a + ii[i];
1603: sum = y[i];
1604: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1605: z[i] = sum;
1606: }
1607: }
1608: PetscCall(PetscLogFlops(2.0 * a->nz));
1609: PetscCall(VecRestoreArrayRead(xx, &x));
1610: PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
1611: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1612: PetscFunctionReturn(PETSC_SUCCESS);
1613: }
1615: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1616: PetscErrorCode MatMultAdd_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1617: {
1618: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1619: PetscScalar *y, *z;
1620: const PetscScalar *x;
1621: const MatScalar *a_a;
1622: const PetscInt *ii, *ridx = NULL;
1623: PetscInt m = A->rmap->n;
1624: PetscBool usecprow = a->compressedrow.use;
1626: PetscFunctionBegin;
1627: if (a->inode.use && a->inode.checked) {
1628: PetscCall(MatMultAdd_SeqAIJ_Inode(A, xx, yy, zz));
1629: PetscFunctionReturn(PETSC_SUCCESS);
1630: }
1631: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1632: PetscCall(VecGetArrayRead(xx, &x));
1633: PetscCall(VecGetArrayPair(yy, zz, &y, &z));
1634: if (usecprow) { /* use compressed row format */
1635: if (zz != yy) PetscCall(PetscArraycpy(z, y, m));
1636: m = a->compressedrow.nrows;
1637: ii = a->compressedrow.i;
1638: ridx = a->compressedrow.rindex;
1639: for (PetscInt i = 0; i < m; i++) {
1640: PetscInt n = ii[i + 1] - ii[i];
1641: const PetscInt *aj = a->j + ii[i];
1642: const PetscScalar *aa = a_a + ii[i];
1643: PetscScalar sum = y[*ridx];
1644: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1645: z[*ridx++] = sum;
1646: }
1647: } else { /* do not use compressed row format */
1648: ii = a->i;
1649: #if PetscDefined(USE_FORTRAN_KERNEL_MULTADDAIJ)
1650: fortranmultaddaij_(&m, x, ii, a->j, a_a, y, z);
1651: #else
1652: PetscPragmaUseOMPKernels(parallel for)
1653: for (PetscInt i = 0; i < m; i++) {
1654: PetscInt n = ii[i + 1] - ii[i];
1655: const PetscInt *aj = a->j + ii[i];
1656: const PetscScalar *aa = a_a + ii[i];
1657: PetscScalar sum = y[i];
1658: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1659: z[i] = sum;
1660: }
1661: #endif
1662: }
1663: PetscCall(PetscLogFlops(2.0 * a->nz));
1664: PetscCall(VecRestoreArrayRead(xx, &x));
1665: PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
1666: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1667: PetscFunctionReturn(PETSC_SUCCESS);
1668: }
1670: static PetscErrorCode MatShift_SeqAIJ(Mat A, PetscScalar v)
1671: {
1672: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1673: const PetscInt *diag;
1674: const PetscInt *ii = (const PetscInt *)a->i;
1675: const PetscInt ndiag = PetscMin(A->rmap->n, A->cmap->n);
1676: PetscBool diagDense, hasDiag;
1678: PetscFunctionBegin;
1679: /* If A is preallocated but has no nonzeros, preallocate it again only if it is not 0 x n or m x 0, otherwise the nonzero state
1680: would increase even though the shift creates no new nonzeros */
1681: if (!A->preallocated || (!a->nz && ndiag > 0)) {
1682: PetscCall(MatSeqAIJSetPreallocation(A, 1, NULL));
1683: PetscCall(MatShift_Basic(A, v));
1684: PetscFunctionReturn(PETSC_SUCCESS);
1685: }
1687: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
1688: /* Reallocate A only if a diagonal entry is missing, otherwise the nonzero state would increase even though the shift creates no
1689: new nonzeros. diagDense alone cannot tell, as it also counts the rows past the last column, which cannot have a diagonal entry */
1690: hasDiag = diagDense;
1691: if (!hasDiag) {
1692: hasDiag = PETSC_TRUE;
1693: for (PetscInt i = 0; i < ndiag && hasDiag; i++) hasDiag = (PetscBool)(diag[i] < ii[i + 1]);
1694: }
1695: if (hasDiag) {
1696: PetscScalar *Aa;
1698: PetscCall(MatSeqAIJGetArray(A, &Aa));
1699: for (PetscInt i = 0; i < ndiag; i++) Aa[diag[i]] += v;
1700: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
1701: } else {
1702: PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */
1703: PetscInt *oldj = a->j, *oldi = a->i;
1704: PetscBool free_a = a->free_a, free_ij = a->free_ij;
1705: const PetscScalar *Aa;
1706: PetscInt *mdiag = NULL;
1708: PetscCall(PetscCalloc1(A->rmap->n, &mdiag));
1709: for (PetscInt i = 0; i < A->rmap->n; i++) {
1710: if (i < A->cmap->n && diag[i] >= ii[i + 1]) { /* 'out of range' rows never have diagonals */
1711: mdiag[i] = 1;
1712: }
1713: }
1714: PetscCall(MatSeqAIJGetArrayRead(A, &Aa)); // sync the host
1715: PetscCall(MatSeqAIJRestoreArrayRead(A, &Aa));
1717: a->a = NULL;
1718: a->j = NULL;
1719: a->i = NULL;
1720: /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */
1721: for (PetscInt i = 0; i < PetscMin(A->rmap->n, A->cmap->n); i++) a->imax[i] += mdiag[i];
1722: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(A, 0, a->imax));
1724: /* copy old values into new matrix data structure */
1725: for (PetscInt i = 0; i < A->rmap->n; i++) {
1726: PetscCall(MatSetValues(A, 1, &i, a->imax[i] - mdiag[i], &oldj[oldi[i]], &olda[oldi[i]], ADD_VALUES));
1727: if (i < A->cmap->n) PetscCall(MatSetValue(A, i, i, v, ADD_VALUES));
1728: }
1729: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
1730: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
1731: if (free_a) PetscCall(PetscShmgetDeallocateArray((void **)&olda));
1732: if (free_ij) PetscCall(PetscShmgetDeallocateArray((void **)&oldj));
1733: if (free_ij) PetscCall(PetscShmgetDeallocateArray((void **)&oldi));
1734: PetscCall(PetscFree(mdiag));
1735: }
1736: PetscFunctionReturn(PETSC_SUCCESS);
1737: }
1739: #include <petscblaslapack.h>
1740: #include <petsc/private/kernels/blockinvert.h>
1742: /*
1743: Note that values is allocated externally by the PC and then passed into this routine
1744: */
1745: static PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
1746: {
1747: PetscInt n = A->rmap->n, i, ncnt = 0, *indx, j, bsizemax = 0, *v_pivots;
1748: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
1749: const PetscReal shift = 0.0;
1750: PetscInt ipvt[5];
1751: PetscCount flops = 0;
1752: PetscScalar work[25], *v_work;
1754: PetscFunctionBegin;
1755: allowzeropivot = PetscNot(A->erroriffailure);
1756: for (i = 0; i < nblocks; i++) ncnt += bsizes[i];
1757: PetscCheck(ncnt == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Total blocksizes %" PetscInt_FMT " doesn't match number matrix rows %" PetscInt_FMT, ncnt, n);
1758: for (i = 0; i < nblocks; i++) bsizemax = PetscMax(bsizemax, bsizes[i]);
1759: PetscCall(PetscMalloc1(bsizemax, &indx));
1760: if (bsizemax > 7) PetscCall(PetscMalloc2(bsizemax, &v_work, bsizemax, &v_pivots));
1761: ncnt = 0;
1762: for (i = 0; i < nblocks; i++) {
1763: for (j = 0; j < bsizes[i]; j++) indx[j] = ncnt + j;
1764: PetscCall(MatGetValues(A, bsizes[i], indx, bsizes[i], indx, diag));
1765: switch (bsizes[i]) {
1766: case 1:
1767: *diag = 1.0 / (*diag);
1768: break;
1769: case 2:
1770: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
1771: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1772: PetscCall(PetscKernel_A_gets_transpose_A_2(diag));
1773: break;
1774: case 3:
1775: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
1776: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1777: PetscCall(PetscKernel_A_gets_transpose_A_3(diag));
1778: break;
1779: case 4:
1780: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
1781: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1782: PetscCall(PetscKernel_A_gets_transpose_A_4(diag));
1783: break;
1784: case 5:
1785: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
1786: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1787: PetscCall(PetscKernel_A_gets_transpose_A_5(diag));
1788: break;
1789: case 6:
1790: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
1791: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1792: PetscCall(PetscKernel_A_gets_transpose_A_6(diag));
1793: break;
1794: case 7:
1795: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
1796: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1797: PetscCall(PetscKernel_A_gets_transpose_A_7(diag));
1798: break;
1799: default:
1800: PetscCall(PetscKernel_A_gets_inverse_A(bsizes[i], diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
1801: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1802: PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bsizes[i]));
1803: }
1804: ncnt += bsizes[i];
1805: diag += bsizes[i] * bsizes[i];
1806: flops += 2 * PetscPowInt64(bsizes[i], 3) / 3;
1807: }
1808: PetscCall(PetscLogFlops(flops));
1809: if (bsizemax > 7) PetscCall(PetscFree2(v_work, v_pivots));
1810: PetscCall(PetscFree(indx));
1811: PetscFunctionReturn(PETSC_SUCCESS);
1812: }
1814: /*
1815: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1816: */
1817: static PetscErrorCode MatInvertDiagonalForSOR_SeqAIJ(Mat A, PetscScalar omega, PetscScalar fshift)
1818: {
1819: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1820: PetscInt i, m = A->rmap->n;
1821: const MatScalar *v;
1822: PetscScalar *idiag, *mdiag;
1823: PetscBool diagDense;
1824: const PetscInt *diag;
1826: PetscFunctionBegin;
1827: if (a->idiagState == ((PetscObject)A)->state && a->omega == omega && a->fshift == fshift) PetscFunctionReturn(PETSC_SUCCESS);
1828: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
1829: PetscCheck(diagDense, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix must have all diagonal locations to invert them");
1830: if (!a->idiag) PetscCall(PetscMalloc3(m, &a->idiag, m, &a->mdiag, m, &a->ssor_work));
1832: mdiag = a->mdiag;
1833: idiag = a->idiag;
1834: PetscCall(MatSeqAIJGetArrayRead(A, &v));
1835: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1836: for (i = 0; i < m; i++) {
1837: mdiag[i] = v[diag[i]];
1838: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1839: PetscCheck(PetscRealPart(fshift), PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Zero diagonal on row %" PetscInt_FMT, i);
1840: PetscCall(PetscInfo(A, "Zero diagonal on row %" PetscInt_FMT "\n", i));
1841: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1842: A->factorerror_zeropivot_value = 0.0;
1843: A->factorerror_zeropivot_row = i;
1844: }
1845: idiag[i] = 1.0 / v[diag[i]];
1846: }
1847: PetscCall(PetscLogFlops(m));
1848: } else {
1849: for (i = 0; i < m; i++) {
1850: mdiag[i] = v[diag[i]];
1851: idiag[i] = omega / (fshift + v[diag[i]]);
1852: }
1853: PetscCall(PetscLogFlops(2.0 * m));
1854: }
1855: PetscCall(MatSeqAIJRestoreArrayRead(A, &v));
1856: a->idiagState = ((PetscObject)A)->state;
1857: a->omega = omega;
1858: a->fshift = fshift;
1859: PetscFunctionReturn(PETSC_SUCCESS);
1860: }
1862: PetscErrorCode MatSOR_SeqAIJ(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1863: {
1864: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1865: PetscScalar *x, d, sum, *t, scale;
1866: const MatScalar *v, *idiag = NULL, *mdiag, *aa;
1867: const PetscScalar *b, *bs, *xb, *ts;
1868: PetscInt n, m = A->rmap->n, i;
1869: const PetscInt *idx, *diag;
1871: PetscFunctionBegin;
1872: if (a->inode.use && a->inode.checked && omega == 1.0 && fshift == 0.0) {
1873: PetscCall(MatSOR_SeqAIJ_Inode(A, bb, omega, flag, fshift, its, lits, xx));
1874: PetscFunctionReturn(PETSC_SUCCESS);
1875: }
1876: its = its * lits;
1877: PetscCall(MatInvertDiagonalForSOR_SeqAIJ(A, omega, fshift));
1878: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
1879: t = a->ssor_work;
1880: idiag = a->idiag;
1881: mdiag = a->mdiag;
1883: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1884: PetscCall(VecGetArray(xx, &x));
1885: PetscCall(VecGetArrayRead(bb, &b));
1886: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1887: if (flag == SOR_APPLY_UPPER) {
1888: /* apply (U + D/omega) to the vector */
1889: bs = b;
1890: for (i = 0; i < m; i++) {
1891: d = fshift + mdiag[i];
1892: n = a->i[i + 1] - diag[i] - 1;
1893: idx = a->j + diag[i] + 1;
1894: v = aa + diag[i] + 1;
1895: sum = b[i] * d / omega;
1896: PetscSparseDensePlusDot(sum, bs, v, idx, n);
1897: x[i] = sum;
1898: }
1899: PetscCall(VecRestoreArray(xx, &x));
1900: PetscCall(VecRestoreArrayRead(bb, &b));
1901: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1902: PetscCall(PetscLogFlops(a->nz));
1903: PetscFunctionReturn(PETSC_SUCCESS);
1904: }
1906: PetscCheck(flag != SOR_APPLY_LOWER, PETSC_COMM_SELF, PETSC_ERR_SUP, "SOR_APPLY_LOWER is not implemented");
1907: if (flag & SOR_EISENSTAT) {
1908: /* Let A = L + U + D; where L is lower triangular,
1909: U is upper triangular, E = D/omega; This routine applies
1911: (L + E)^{-1} A (U + E)^{-1}
1913: to a vector efficiently using Eisenstat's trick.
1914: */
1915: scale = (2.0 / omega) - 1.0;
1917: /* x = (E + U)^{-1} b */
1918: for (i = m - 1; i >= 0; i--) {
1919: n = a->i[i + 1] - diag[i] - 1;
1920: idx = a->j + diag[i] + 1;
1921: v = aa + diag[i] + 1;
1922: sum = b[i];
1923: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1924: x[i] = sum * idiag[i];
1925: }
1927: /* t = b - (2*E - D)x */
1928: v = aa;
1929: for (i = 0; i < m; i++) t[i] = b[i] - scale * v[*diag++] * x[i];
1931: /* t = (E + L)^{-1}t */
1932: ts = t;
1933: diag = a->diag;
1934: for (i = 0; i < m; i++) {
1935: n = diag[i] - a->i[i];
1936: idx = a->j + a->i[i];
1937: v = aa + a->i[i];
1938: sum = t[i];
1939: PetscSparseDenseMinusDot(sum, ts, v, idx, n);
1940: t[i] = sum * idiag[i];
1941: /* x = x + t */
1942: x[i] += t[i];
1943: }
1945: PetscCall(PetscLogFlops(6.0 * m - 1 + 2.0 * a->nz));
1946: PetscCall(VecRestoreArray(xx, &x));
1947: PetscCall(VecRestoreArrayRead(bb, &b));
1948: PetscFunctionReturn(PETSC_SUCCESS);
1949: }
1950: if (flag & SOR_ZERO_INITIAL_GUESS) {
1951: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1952: for (i = 0; i < m; i++) {
1953: n = diag[i] - a->i[i];
1954: idx = a->j + a->i[i];
1955: v = aa + a->i[i];
1956: sum = b[i];
1957: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1958: t[i] = sum;
1959: x[i] = sum * idiag[i];
1960: }
1961: xb = t;
1962: PetscCall(PetscLogFlops(a->nz));
1963: } else xb = b;
1964: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1965: for (i = m - 1; i >= 0; i--) {
1966: n = a->i[i + 1] - diag[i] - 1;
1967: idx = a->j + diag[i] + 1;
1968: v = aa + diag[i] + 1;
1969: sum = xb[i];
1970: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1971: if (xb == b) {
1972: x[i] = sum * idiag[i];
1973: } else {
1974: x[i] = (1 - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
1975: }
1976: }
1977: PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */
1978: }
1979: its--;
1980: }
1981: while (its--) {
1982: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1983: for (i = 0; i < m; i++) {
1984: /* lower */
1985: n = diag[i] - a->i[i];
1986: idx = a->j + a->i[i];
1987: v = aa + a->i[i];
1988: sum = b[i];
1989: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1990: t[i] = sum; /* save application of the lower-triangular part */
1991: /* upper */
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: xb = t;
1999: PetscCall(PetscLogFlops(2.0 * a->nz));
2000: } else xb = b;
2001: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
2002: for (i = m - 1; i >= 0; i--) {
2003: sum = xb[i];
2004: if (xb == b) {
2005: /* whole matrix (no checkpointing available) */
2006: n = a->i[i + 1] - a->i[i];
2007: idx = a->j + a->i[i];
2008: v = aa + a->i[i];
2009: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2010: x[i] = (1. - omega) * x[i] + (sum + mdiag[i] * x[i]) * idiag[i];
2011: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
2012: n = a->i[i + 1] - diag[i] - 1;
2013: idx = a->j + diag[i] + 1;
2014: v = aa + diag[i] + 1;
2015: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2016: x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
2017: }
2018: }
2019: if (xb == b) PetscCall(PetscLogFlops(2.0 * a->nz));
2020: else PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */
2021: }
2022: }
2023: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2024: PetscCall(VecRestoreArray(xx, &x));
2025: PetscCall(VecRestoreArrayRead(bb, &b));
2026: PetscFunctionReturn(PETSC_SUCCESS);
2027: }
2029: static PetscErrorCode MatGetInfo_SeqAIJ(Mat A, MatInfoType flag, MatInfo *info)
2030: {
2031: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2033: PetscFunctionBegin;
2034: info->block_size = 1.0;
2035: info->nz_allocated = a->maxnz;
2036: info->nz_used = a->nz;
2037: info->nz_unneeded = (a->maxnz - a->nz);
2038: info->assemblies = A->num_ass;
2039: info->mallocs = A->info.mallocs;
2040: info->memory = 0; /* REVIEW ME */
2041: if (A->factortype) {
2042: info->fill_ratio_given = A->info.fill_ratio_given;
2043: info->fill_ratio_needed = A->info.fill_ratio_needed;
2044: info->factor_mallocs = A->info.factor_mallocs;
2045: } else {
2046: info->fill_ratio_given = 0;
2047: info->fill_ratio_needed = 0;
2048: info->factor_mallocs = 0;
2049: }
2050: PetscFunctionReturn(PETSC_SUCCESS);
2051: }
2053: static PetscErrorCode MatZeroRows_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diagv, Vec x, Vec b)
2054: {
2055: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2056: PetscInt i, m = A->rmap->n - 1;
2057: const PetscScalar *xx;
2058: PetscScalar *bb, *aa;
2059: PetscInt d = 0;
2060: const PetscInt *diag;
2062: PetscFunctionBegin;
2063: if (x && b) {
2064: PetscCall(VecGetArrayRead(x, &xx));
2065: PetscCall(VecGetArray(b, &bb));
2066: for (i = 0; i < N; i++) {
2067: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2068: if (rows[i] >= A->cmap->n) continue;
2069: bb[rows[i]] = diagv * xx[rows[i]];
2070: }
2071: PetscCall(VecRestoreArrayRead(x, &xx));
2072: PetscCall(VecRestoreArray(b, &bb));
2073: }
2075: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
2076: PetscCall(MatSeqAIJGetArray(A, &aa));
2077: if (a->keepnonzeropattern) {
2078: for (i = 0; i < N; i++) {
2079: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2080: PetscCall(PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]]));
2081: }
2082: if (diagv != 0.0) {
2083: for (i = 0; i < N; i++) {
2084: d = rows[i];
2085: if (d >= A->cmap->n) continue;
2086: 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);
2087: aa[diag[d]] = diagv;
2088: }
2089: }
2090: } else {
2091: if (diagv != 0.0) {
2092: for (i = 0; i < N; i++) {
2093: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2094: if (a->ilen[rows[i]] > 0) {
2095: if (rows[i] >= A->cmap->n) {
2096: a->ilen[rows[i]] = 0;
2097: } else {
2098: a->ilen[rows[i]] = 1;
2099: aa[a->i[rows[i]]] = diagv;
2100: a->j[a->i[rows[i]]] = rows[i];
2101: }
2102: } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */
2103: PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diagv, INSERT_VALUES));
2104: }
2105: }
2106: } else {
2107: for (i = 0; i < N; i++) {
2108: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2109: a->ilen[rows[i]] = 0;
2110: }
2111: }
2112: if (N) A->nonzerostate++;
2113: }
2114: PetscCall(MatSeqAIJRestoreArray(A, &aa));
2115: PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2116: PetscFunctionReturn(PETSC_SUCCESS);
2117: }
2119: static PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diagv, Vec x, Vec b)
2120: {
2121: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2122: PetscInt i, j, m = A->rmap->n - 1;
2123: PetscBool *zeroed, vecs = PETSC_FALSE;
2124: const PetscScalar *xx;
2125: PetscScalar *bb, *aa;
2126: const PetscInt *diag;
2127: PetscBool diagDense;
2129: PetscFunctionBegin;
2130: if (!N) PetscFunctionReturn(PETSC_SUCCESS);
2131: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
2132: PetscCall(MatSeqAIJGetArray(A, &aa));
2133: if (x && b) {
2134: PetscCall(VecGetArrayRead(x, &xx));
2135: PetscCall(VecGetArray(b, &bb));
2136: vecs = PETSC_TRUE;
2137: }
2138: PetscCall(PetscCalloc1(A->rmap->n, &zeroed));
2139: for (i = 0; i < N; i++) {
2140: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2141: PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aa, a->i[rows[i]]), a->ilen[rows[i]]));
2143: zeroed[rows[i]] = PETSC_TRUE;
2144: }
2145: for (i = 0; i < A->rmap->n; i++) {
2146: if (!zeroed[i]) {
2147: for (j = a->i[i]; j < a->i[i + 1]; j++) {
2148: if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) {
2149: if (vecs) bb[i] -= aa[j] * xx[a->j[j]];
2150: aa[j] = 0.0;
2151: }
2152: }
2153: } else if (vecs && i < A->cmap->N) bb[i] = diagv * xx[i];
2154: }
2155: if (x && b) {
2156: PetscCall(VecRestoreArrayRead(x, &xx));
2157: PetscCall(VecRestoreArray(b, &bb));
2158: }
2159: PetscCall(PetscFree(zeroed));
2160: if (diagv != 0.0) {
2161: if (!diagDense) {
2162: for (i = 0; i < N; i++) {
2163: if (rows[i] >= A->cmap->N || rows[i] < 0) continue;
2164: PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diagv, INSERT_VALUES));
2165: }
2166: } else {
2167: for (i = 0; i < N; i++) aa[diag[rows[i]]] = diagv;
2168: }
2169: }
2170: PetscCall(MatSeqAIJRestoreArray(A, &aa));
2171: if (!diagDense) PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2172: PetscFunctionReturn(PETSC_SUCCESS);
2173: }
2175: PetscErrorCode MatGetRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2176: {
2177: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2178: const PetscScalar *aa;
2180: PetscFunctionBegin;
2181: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2182: *nz = a->i[row + 1] - a->i[row];
2183: if (v) *v = PetscSafePointerPlusOffset((PetscScalar *)aa, a->i[row]);
2184: if (idx) {
2185: if (*nz && a->j) *idx = a->j + a->i[row];
2186: else *idx = NULL;
2187: }
2188: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2189: PetscFunctionReturn(PETSC_SUCCESS);
2190: }
2192: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2193: {
2194: PetscFunctionBegin;
2195: PetscFunctionReturn(PETSC_SUCCESS);
2196: }
2198: static PetscErrorCode MatNorm_SeqAIJ(Mat A, NormType type, PetscReal *nrm)
2199: {
2200: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2201: const MatScalar *v;
2202: PetscReal sum = 0.0;
2203: PetscInt i, j;
2205: PetscFunctionBegin;
2206: PetscCall(MatSeqAIJGetArrayRead(A, &v));
2207: if (type == NORM_FROBENIUS) {
2208: #if PetscDefined(USE_REAL___FP16)
2209: PetscBLASInt one = 1, nz = a->nz;
2210: PetscCallBLAS("BLASnrm2", *nrm = BLASnrm2_(&nz, v, &one));
2211: #else
2212: for (i = 0; i < a->nz; i++) {
2213: sum += PetscRealPart(PetscConj(*v) * (*v));
2214: v++;
2215: }
2216: *nrm = PetscSqrtReal(sum);
2217: #endif
2218: PetscCall(PetscLogFlops(2.0 * a->nz));
2219: } else if (type == NORM_1) {
2220: PetscReal *tmp;
2221: PetscInt *jj = a->j;
2222: PetscCall(PetscCalloc1(A->cmap->n, &tmp));
2223: *nrm = 0.0;
2224: for (j = 0; j < a->nz; j++) {
2225: tmp[*jj++] += PetscAbsScalar(*v);
2226: v++;
2227: }
2228: for (j = 0; j < A->cmap->n; j++) {
2229: if (tmp[j] > *nrm) *nrm = tmp[j];
2230: }
2231: PetscCall(PetscFree(tmp));
2232: PetscCall(PetscLogFlops(a->nz));
2233: } else if (type == NORM_INFINITY) {
2234: *nrm = 0.0;
2235: for (j = 0; j < A->rmap->n; j++) {
2236: const PetscScalar *v2 = PetscSafePointerPlusOffset(v, a->i[j]);
2237: sum = 0.0;
2238: for (i = 0; i < a->i[j + 1] - a->i[j]; i++) {
2239: sum += PetscAbsScalar(*v2);
2240: v2++;
2241: }
2242: if (sum > *nrm) *nrm = sum;
2243: }
2244: PetscCall(PetscLogFlops(a->nz));
2245: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for two norm");
2246: PetscCall(MatSeqAIJRestoreArrayRead(A, &v));
2247: PetscFunctionReturn(PETSC_SUCCESS);
2248: }
2250: static PetscErrorCode MatIsTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2251: {
2252: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2253: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2254: const MatScalar *va, *vb;
2255: PetscInt ma, na, mb, nb, i;
2257: PetscFunctionBegin;
2258: PetscCall(MatGetSize(A, &ma, &na));
2259: PetscCall(MatGetSize(B, &mb, &nb));
2260: if (ma != nb || na != mb) {
2261: *f = PETSC_FALSE;
2262: PetscFunctionReturn(PETSC_SUCCESS);
2263: }
2264: PetscCall(MatSeqAIJGetArrayRead(A, &va));
2265: PetscCall(MatSeqAIJGetArrayRead(B, &vb));
2266: aii = aij->i;
2267: bii = bij->i;
2268: adx = aij->j;
2269: bdx = bij->j;
2270: PetscCall(PetscMalloc1(ma, &aptr));
2271: PetscCall(PetscMalloc1(mb, &bptr));
2272: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2273: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2275: *f = PETSC_TRUE;
2276: for (i = 0; i < ma; i++) {
2277: while (aptr[i] < aii[i + 1]) {
2278: PetscInt idc, idr;
2279: PetscScalar vc, vr;
2280: /* column/row index/value */
2281: idc = adx[aptr[i]];
2282: idr = bdx[bptr[idc]];
2283: vc = va[aptr[i]];
2284: vr = vb[bptr[idc]];
2285: if (i != idr || PetscAbsScalar(vc - vr) > tol) {
2286: *f = PETSC_FALSE;
2287: goto done;
2288: } else {
2289: aptr[i]++;
2290: if (B || i != idc) bptr[idc]++;
2291: }
2292: }
2293: }
2294: done:
2295: PetscCall(PetscFree(aptr));
2296: PetscCall(PetscFree(bptr));
2297: PetscCall(MatSeqAIJRestoreArrayRead(A, &va));
2298: PetscCall(MatSeqAIJRestoreArrayRead(B, &vb));
2299: PetscFunctionReturn(PETSC_SUCCESS);
2300: }
2302: static PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2303: {
2304: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2305: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2306: MatScalar *va, *vb;
2307: PetscInt ma, na, mb, nb, i;
2309: PetscFunctionBegin;
2310: PetscCall(MatGetSize(A, &ma, &na));
2311: PetscCall(MatGetSize(B, &mb, &nb));
2312: if (ma != nb || na != mb) {
2313: *f = PETSC_FALSE;
2314: PetscFunctionReturn(PETSC_SUCCESS);
2315: }
2316: aii = aij->i;
2317: bii = bij->i;
2318: adx = aij->j;
2319: bdx = bij->j;
2320: va = aij->a;
2321: vb = bij->a;
2322: PetscCall(PetscMalloc1(ma, &aptr));
2323: PetscCall(PetscMalloc1(mb, &bptr));
2324: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2325: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2327: *f = PETSC_TRUE;
2328: for (i = 0; i < ma; i++) {
2329: while (aptr[i] < aii[i + 1]) {
2330: PetscInt idc, idr;
2331: PetscScalar vc, vr;
2332: /* column/row index/value */
2333: idc = adx[aptr[i]];
2334: idr = bdx[bptr[idc]];
2335: vc = va[aptr[i]];
2336: vr = vb[bptr[idc]];
2337: if (i != idr || PetscAbsScalar(vc - PetscConj(vr)) > tol) {
2338: *f = PETSC_FALSE;
2339: goto done;
2340: } else {
2341: aptr[i]++;
2342: if (B || i != idc) bptr[idc]++;
2343: }
2344: }
2345: }
2346: done:
2347: PetscCall(PetscFree(aptr));
2348: PetscCall(PetscFree(bptr));
2349: PetscFunctionReturn(PETSC_SUCCESS);
2350: }
2352: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A, Vec ll, Vec rr)
2353: {
2354: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2355: const PetscScalar *l, *r;
2356: PetscScalar x;
2357: MatScalar *v;
2358: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, M, nz = a->nz;
2359: const PetscInt *jj;
2361: PetscFunctionBegin;
2362: if (ll) {
2363: /* The local size is used so that VecMPI can be passed to this routine
2364: by MatDiagonalScale_MPIAIJ */
2365: PetscCall(VecGetLocalSize(ll, &m));
2366: PetscCheck(m == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vector wrong length");
2367: PetscCall(VecGetArrayRead(ll, &l));
2368: PetscCall(MatSeqAIJGetArray(A, &v));
2369: for (i = 0; i < m; i++) {
2370: x = l[i];
2371: M = a->i[i + 1] - a->i[i];
2372: for (j = 0; j < M; j++) (*v++) *= x;
2373: }
2374: PetscCall(VecRestoreArrayRead(ll, &l));
2375: PetscCall(PetscLogFlops(nz));
2376: PetscCall(MatSeqAIJRestoreArray(A, &v));
2377: }
2378: if (rr) {
2379: PetscCall(VecGetLocalSize(rr, &n));
2380: PetscCheck(n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vector wrong length");
2381: PetscCall(VecGetArrayRead(rr, &r));
2382: PetscCall(MatSeqAIJGetArray(A, &v));
2383: jj = a->j;
2384: for (i = 0; i < nz; i++) (*v++) *= r[*jj++];
2385: PetscCall(MatSeqAIJRestoreArray(A, &v));
2386: PetscCall(VecRestoreArrayRead(rr, &r));
2387: PetscCall(PetscLogFlops(nz));
2388: }
2389: PetscFunctionReturn(PETSC_SUCCESS);
2390: }
2392: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A, IS isrow, IS iscol, PetscInt csize, MatReuse scall, Mat *B)
2393: {
2394: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *c;
2395: PetscInt *smap, i, k, kstart, kend, oldcols = A->cmap->n, *lens;
2396: PetscInt row, mat_i, *mat_j, tcol, first, step, *mat_ilen, sum, lensi;
2397: const PetscInt *irow, *icol;
2398: const PetscScalar *aa;
2399: PetscInt nrows, ncols;
2400: PetscInt *starts, *j_new, *i_new, *aj = a->j, *ai = a->i, ii, *ailen = a->ilen;
2401: MatScalar *a_new, *mat_a, *c_a;
2402: Mat C;
2403: PetscBool stride;
2405: PetscFunctionBegin;
2406: PetscCall(ISGetIndices(isrow, &irow));
2407: PetscCall(ISGetLocalSize(isrow, &nrows));
2408: PetscCall(ISGetLocalSize(iscol, &ncols));
2410: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &stride));
2411: if (stride) {
2412: PetscCall(ISStrideGetInfo(iscol, &first, &step));
2413: } else {
2414: first = 0;
2415: step = 0;
2416: }
2417: if (stride && step == 1) {
2418: /* special case of contiguous rows */
2419: PetscCall(PetscMalloc2(nrows, &lens, nrows, &starts));
2420: /* loop over new rows determining lens and starting points */
2421: for (i = 0; i < nrows; i++) {
2422: kstart = ai[irow[i]];
2423: kend = kstart + ailen[irow[i]];
2424: starts[i] = kstart;
2425: for (k = kstart; k < kend; k++) {
2426: if (aj[k] >= first) {
2427: starts[i] = k;
2428: break;
2429: }
2430: }
2431: sum = 0;
2432: while (k < kend) {
2433: if (aj[k++] >= first + ncols) break;
2434: sum++;
2435: }
2436: lens[i] = sum;
2437: }
2438: /* create submatrix */
2439: if (scall == MAT_REUSE_MATRIX) {
2440: PetscInt n_cols, n_rows;
2441: PetscCall(MatGetSize(*B, &n_rows, &n_cols));
2442: PetscCheck(n_rows == nrows && n_cols == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Reused submatrix wrong size");
2443: PetscCall(MatZeroEntries(*B));
2444: C = *B;
2445: } else {
2446: PetscInt rbs, cbs;
2447: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
2448: PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE));
2449: PetscCall(ISGetBlockSize(isrow, &rbs));
2450: PetscCall(ISGetBlockSize(iscol, &cbs));
2451: PetscCall(MatSetBlockSizes(C, rbs, cbs));
2452: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2453: PetscCall(MatSetOption(C, MAT_STRUCTURE_ONLY, A->structure_only));
2454: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens));
2455: }
2456: c = (Mat_SeqAIJ *)C->data;
2458: /* loop over rows inserting into submatrix */
2459: j_new = c->j;
2460: i_new = c->i;
2461: PetscCall(MatSeqAIJGetArrayWrite(C, &a_new)); // Not 'a_new = c->a-new', since that raw usage ignores offload state of C
2462: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2463: for (i = 0; i < nrows; i++) {
2464: ii = starts[i];
2465: lensi = lens[i];
2466: if (lensi) {
2467: for (k = 0; k < lensi; k++) *j_new++ = aj[ii + k] - first;
2468: if (!A->structure_only) {
2469: PetscCall(PetscArraycpy(a_new, aa + starts[i], lensi));
2470: a_new += lensi;
2471: }
2472: }
2473: i_new[i + 1] = i_new[i] + lensi;
2474: c->ilen[i] = lensi;
2475: }
2476: PetscCall(MatSeqAIJRestoreArrayWrite(C, &a_new)); // Set C's offload state properly
2477: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2478: PetscCall(PetscFree2(lens, starts));
2479: } else {
2480: PetscCall(ISGetIndices(iscol, &icol));
2481: PetscCall(PetscCalloc1(oldcols, &smap));
2482: PetscCall(PetscMalloc1(1 + nrows, &lens));
2483: for (i = 0; i < ncols; i++) {
2484: 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);
2485: smap[icol[i]] = i + 1;
2486: }
2488: /* determine lens of each row */
2489: for (i = 0; i < nrows; i++) {
2490: kstart = ai[irow[i]];
2491: kend = kstart + a->ilen[irow[i]];
2492: lens[i] = 0;
2493: for (k = kstart; k < kend; k++) {
2494: if (smap[aj[k]]) lens[i]++;
2495: }
2496: }
2497: /* Create and fill new matrix */
2498: if (scall == MAT_REUSE_MATRIX) {
2499: PetscBool equal;
2501: c = (Mat_SeqAIJ *)(*B)->data;
2502: PetscCheck((*B)->rmap->n == nrows && (*B)->cmap->n == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong size");
2503: PetscCall(PetscArraycmp(c->ilen, lens, (*B)->rmap->n, &equal));
2504: PetscCheck(equal, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong number of nonzeros");
2505: PetscCall(PetscArrayzero(c->ilen, (*B)->rmap->n));
2506: C = *B;
2507: } else {
2508: PetscInt rbs, cbs;
2509: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
2510: PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE));
2511: PetscCall(ISGetBlockSize(isrow, &rbs));
2512: PetscCall(ISGetBlockSize(iscol, &cbs));
2513: if (rbs > 1 || cbs > 1) PetscCall(MatSetBlockSizes(C, rbs, cbs));
2514: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2515: PetscCall(MatSetOption(C, MAT_STRUCTURE_ONLY, A->structure_only));
2516: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens));
2517: }
2518: c = (Mat_SeqAIJ *)C->data;
2519: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2520: PetscCall(MatSeqAIJGetArrayWrite(C, &c_a)); // Not 'c->a', since that raw usage ignores offload state of C
2521: for (i = 0; i < nrows; i++) {
2522: row = irow[i];
2523: kstart = ai[row];
2524: kend = kstart + a->ilen[row];
2525: mat_i = c->i[i];
2526: mat_j = PetscSafePointerPlusOffset(c->j, mat_i);
2527: mat_a = PetscSafePointerPlusOffset(c_a, mat_i);
2528: mat_ilen = c->ilen + i;
2529: for (k = kstart; k < kend; k++) {
2530: if ((tcol = smap[a->j[k]])) {
2531: *mat_j++ = tcol - 1;
2532: if (!A->structure_only) *mat_a++ = aa[k];
2533: (*mat_ilen)++;
2534: }
2535: }
2536: }
2537: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2538: /* Free work space */
2539: PetscCall(ISRestoreIndices(iscol, &icol));
2540: PetscCall(PetscFree(smap));
2541: PetscCall(PetscFree(lens));
2542: /* sort */
2543: for (i = 0; i < nrows; i++) {
2544: PetscInt ilen;
2546: mat_i = c->i[i];
2547: mat_j = PetscSafePointerPlusOffset(c->j, mat_i);
2548: mat_a = PetscSafePointerPlusOffset(c_a, mat_i);
2549: ilen = c->ilen[i];
2550: if (A->structure_only) PetscCall(PetscSortInt(ilen, mat_j));
2551: else PetscCall(PetscSortIntWithScalarArray(ilen, mat_j, mat_a));
2552: }
2553: PetscCall(MatSeqAIJRestoreArrayWrite(C, &c_a));
2554: }
2555: #if PetscDefined(HAVE_DEVICE)
2556: PetscCall(MatBindToCPU(C, A->boundtocpu));
2557: #endif
2558: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2559: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2561: PetscCall(ISRestoreIndices(isrow, &irow));
2562: *B = C;
2563: PetscFunctionReturn(PETSC_SUCCESS);
2564: }
2566: static PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat, MPI_Comm subComm, MatReuse scall, Mat *subMat)
2567: {
2568: Mat B;
2570: PetscFunctionBegin;
2571: if (scall == MAT_INITIAL_MATRIX) {
2572: PetscCall(MatCreate(subComm, &B));
2573: PetscCall(MatSetSizes(B, mat->rmap->n, mat->cmap->n, mat->rmap->n, mat->cmap->n));
2574: PetscCall(MatSetBlockSizesFromMats(B, mat, mat));
2575: PetscCall(MatSetType(B, MATSEQAIJ));
2576: PetscCall(MatDuplicateNoCreate_SeqAIJ(B, mat, MAT_COPY_VALUES, PETSC_TRUE));
2577: *subMat = B;
2578: } else {
2579: PetscCall(MatCopy_SeqAIJ(mat, *subMat, SAME_NONZERO_PATTERN));
2580: }
2581: PetscFunctionReturn(PETSC_SUCCESS);
2582: }
2584: static PetscErrorCode MatILUFactor_SeqAIJ(Mat inA, IS row, IS col, const MatFactorInfo *info)
2585: {
2586: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2587: Mat outA;
2588: PetscBool row_identity, col_identity;
2590: PetscFunctionBegin;
2591: PetscCheck(info->levels == 0, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only levels=0 supported for in-place ilu");
2593: PetscCall(ISIdentity(row, &row_identity));
2594: PetscCall(ISIdentity(col, &col_identity));
2596: outA = inA;
2597: PetscCall(PetscFree(inA->solvertype));
2598: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &inA->solvertype));
2600: PetscCall(PetscObjectReference((PetscObject)row));
2601: PetscCall(ISDestroy(&a->row));
2603: a->row = row;
2605: PetscCall(PetscObjectReference((PetscObject)col));
2606: PetscCall(ISDestroy(&a->col));
2608: a->col = col;
2610: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2611: PetscCall(ISDestroy(&a->icol));
2612: PetscCall(ISInvertPermutation(col, PETSC_DECIDE, &a->icol));
2614: if (!a->solve_work) { /* this matrix may have been factored before */
2615: PetscCall(PetscMalloc1(inA->rmap->n, &a->solve_work));
2616: }
2618: if (row_identity && col_identity) {
2619: PetscCall(MatLUFactorNumeric_SeqAIJ_inplace(outA, inA, info));
2620: } else {
2621: PetscCall(MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA, inA, info));
2622: }
2623: outA->factortype = MAT_FACTOR_LU;
2624: PetscFunctionReturn(PETSC_SUCCESS);
2625: }
2627: PetscErrorCode MatScale_SeqAIJ(Mat inA, PetscScalar alpha)
2628: {
2629: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2630: PetscScalar *v;
2631: PetscBLASInt one = 1, bnz;
2633: PetscFunctionBegin;
2634: PetscCall(MatSeqAIJGetArray(inA, &v));
2635: PetscCall(PetscBLASIntCast(a->nz, &bnz));
2636: PetscCallBLAS("BLASscal", BLASscal_(&bnz, &alpha, v, &one));
2637: PetscCall(PetscLogFlops(a->nz));
2638: PetscCall(MatSeqAIJRestoreArray(inA, &v));
2639: PetscFunctionReturn(PETSC_SUCCESS);
2640: }
2642: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2643: {
2644: PetscInt i;
2646: PetscFunctionBegin;
2647: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2648: PetscCall(PetscFree4(submatj->sbuf1, submatj->ptr, submatj->tmp, submatj->ctr));
2650: for (i = 0; i < submatj->nrqr; ++i) PetscCall(PetscFree(submatj->sbuf2[i]));
2651: PetscCall(PetscFree3(submatj->sbuf2, submatj->req_size, submatj->req_source1));
2653: if (submatj->rbuf1) {
2654: PetscCall(PetscFree(submatj->rbuf1[0]));
2655: PetscCall(PetscFree(submatj->rbuf1));
2656: }
2658: for (i = 0; i < submatj->nrqs; ++i) PetscCall(PetscFree(submatj->rbuf3[i]));
2659: PetscCall(PetscFree3(submatj->req_source2, submatj->rbuf2, submatj->rbuf3));
2660: PetscCall(PetscFree(submatj->pa));
2661: PetscCall(PetscFree2(submatj->local_a_parent, submatj->local_a_sub));
2662: PetscCall(PetscFree2(submatj->local_b_parent, submatj->local_b_sub));
2663: }
2665: #if PetscDefined(USE_CTABLE)
2666: PetscCall(PetscHMapIDestroy(&submatj->rmap));
2667: PetscCall(PetscFree(submatj->cmap_loc));
2668: PetscCall(PetscFree(submatj->rmap_loc));
2669: #else
2670: PetscCall(PetscFree(submatj->rmap));
2671: #endif
2673: if (!submatj->allcolumns) {
2674: #if PetscDefined(USE_CTABLE)
2675: PetscCall(PetscHMapIDestroy(&submatj->cmap));
2676: #else
2677: PetscCall(PetscFree(submatj->cmap));
2678: #endif
2679: }
2680: PetscCall(PetscFree(submatj->row2proc));
2682: PetscCall(PetscFree(submatj));
2683: PetscFunctionReturn(PETSC_SUCCESS);
2684: }
2686: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2687: {
2688: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data;
2689: Mat_SubSppt *submatj = c->submatis1;
2691: PetscFunctionBegin;
2692: PetscCall((*submatj->destroy)(C));
2693: PetscCall(MatDestroySubMatrix_Private(submatj));
2694: PetscFunctionReturn(PETSC_SUCCESS);
2695: }
2697: /* Note this has code duplication with MatDestroySubMatrices_SeqBAIJ() */
2698: static PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n, Mat *mat[])
2699: {
2700: PetscInt i;
2701: Mat C;
2702: Mat_SeqAIJ *c;
2703: Mat_SubSppt *submatj;
2705: PetscFunctionBegin;
2706: for (i = 0; i < n; i++) {
2707: C = (*mat)[i];
2708: c = (Mat_SeqAIJ *)C->data;
2709: submatj = c->submatis1;
2710: if (submatj) {
2711: if (--((PetscObject)C)->refct <= 0) {
2712: PetscCall(PetscFree(C->factorprefix));
2713: PetscCall((*submatj->destroy)(C));
2714: PetscCall(MatDestroySubMatrix_Private(submatj));
2715: PetscCall(PetscFree(C->defaultvectype));
2716: PetscCall(PetscFree(C->defaultrandtype));
2717: PetscCall(PetscFree(C->solvertype));
2718: PetscCall(PetscLayoutDestroy(&C->rmap));
2719: PetscCall(PetscLayoutDestroy(&C->cmap));
2720: PetscCall(PetscHeaderDestroy(&C));
2721: }
2722: } else {
2723: PetscCall(MatDestroy(&C));
2724: }
2725: }
2727: /* Destroy Dummy submatrices created for reuse */
2728: PetscCall(MatDestroySubMatrices_Dummy(n, mat));
2730: PetscCall(PetscFree(*mat));
2731: PetscFunctionReturn(PETSC_SUCCESS);
2732: }
2734: static PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A, PetscInt n, const IS irow[], const IS icol[], MatReuse scall, Mat *B[])
2735: {
2736: PetscInt i;
2738: PetscFunctionBegin;
2739: if (scall == MAT_INITIAL_MATRIX) PetscCall(PetscCalloc1(n + 1, B));
2741: for (i = 0; i < n; i++) PetscCall(MatCreateSubMatrix_SeqAIJ(A, irow[i], icol[i], PETSC_DECIDE, scall, &(*B)[i]));
2742: PetscFunctionReturn(PETSC_SUCCESS);
2743: }
2745: static PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A, PetscInt is_max, IS is[], PetscInt ov)
2746: {
2747: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2748: PetscInt row, i, j, k, l, ll, m, n, *nidx, isz, val;
2749: const PetscInt *idx;
2750: PetscInt start, end, *ai, *aj, bs = A->rmap->bs == A->cmap->bs ? A->rmap->bs : 1;
2751: PetscBT table;
2753: PetscFunctionBegin;
2754: m = A->rmap->n / bs;
2755: ai = a->i;
2756: aj = a->j;
2758: PetscCheck(ov >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "illegal negative overlap value used");
2760: PetscCall(PetscMalloc1(m + 1, &nidx));
2761: PetscCall(PetscBTCreate(m, &table));
2763: for (i = 0; i < is_max; i++) {
2764: /* Initialize the two local arrays */
2765: isz = 0;
2766: PetscCall(PetscBTMemzero(m, table));
2768: /* Extract the indices, assume there can be duplicate entries */
2769: PetscCall(ISGetIndices(is[i], &idx));
2770: PetscCall(ISGetLocalSize(is[i], &n));
2772: if (bs > 1) {
2773: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2774: for (j = 0; j < n; ++j) {
2775: if (!PetscBTLookupSet(table, idx[j] / bs)) nidx[isz++] = idx[j] / bs;
2776: }
2777: PetscCall(ISRestoreIndices(is[i], &idx));
2778: PetscCall(ISDestroy(&is[i]));
2780: k = 0;
2781: for (j = 0; j < ov; j++) { /* for each overlap */
2782: n = isz;
2783: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2784: for (ll = 0; ll < bs; ll++) {
2785: row = bs * nidx[k] + ll;
2786: start = ai[row];
2787: end = ai[row + 1];
2788: for (l = start; l < end; l++) {
2789: val = aj[l] / bs;
2790: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2791: }
2792: }
2793: }
2794: }
2795: PetscCall(ISCreateBlock(PETSC_COMM_SELF, bs, isz, nidx, PETSC_COPY_VALUES, is + i));
2796: } else {
2797: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2798: for (j = 0; j < n; ++j) {
2799: if (!PetscBTLookupSet(table, idx[j])) nidx[isz++] = idx[j];
2800: }
2801: PetscCall(ISRestoreIndices(is[i], &idx));
2802: PetscCall(ISDestroy(&is[i]));
2804: k = 0;
2805: for (j = 0; j < ov; j++) { /* for each overlap */
2806: n = isz;
2807: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2808: row = nidx[k];
2809: start = ai[row];
2810: end = ai[row + 1];
2811: for (l = start; l < end; l++) {
2812: val = aj[l];
2813: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2814: }
2815: }
2816: }
2817: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, isz, nidx, PETSC_COPY_VALUES, is + i));
2818: }
2819: }
2820: PetscCall(PetscBTDestroy(&table));
2821: PetscCall(PetscFree(nidx));
2822: PetscFunctionReturn(PETSC_SUCCESS);
2823: }
2825: static PetscErrorCode MatPermute_SeqAIJ(Mat A, IS rowp, IS colp, Mat *B)
2826: {
2827: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2828: PetscInt i, nz = 0, m = A->rmap->n, n = A->cmap->n;
2829: const PetscInt *row, *col;
2830: PetscInt *cnew, j, *lens;
2831: IS icolp, irowp;
2832: PetscInt *cwork = NULL;
2833: PetscScalar *vwork = NULL;
2835: PetscFunctionBegin;
2836: PetscCall(ISInvertPermutation(rowp, PETSC_DECIDE, &irowp));
2837: PetscCall(ISGetIndices(irowp, &row));
2838: PetscCall(ISInvertPermutation(colp, PETSC_DECIDE, &icolp));
2839: PetscCall(ISGetIndices(icolp, &col));
2841: /* determine lengths of permuted rows */
2842: PetscCall(PetscMalloc1(m + 1, &lens));
2843: for (i = 0; i < m; i++) lens[row[i]] = a->i[i + 1] - a->i[i];
2844: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
2845: PetscCall(MatSetSizes(*B, m, n, m, n));
2846: PetscCall(MatSetBlockSizesFromMats(*B, A, A));
2847: PetscCall(MatSetType(*B, ((PetscObject)A)->type_name));
2848: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*B, 0, lens));
2849: PetscCall(PetscFree(lens));
2851: PetscCall(PetscMalloc1(n, &cnew));
2852: for (i = 0; i < m; i++) {
2853: PetscCall(MatGetRow_SeqAIJ(A, i, &nz, &cwork, &vwork));
2854: for (j = 0; j < nz; j++) cnew[j] = col[cwork[j]];
2855: PetscCall(MatSetValues_SeqAIJ(*B, 1, &row[i], nz, cnew, vwork, INSERT_VALUES));
2856: PetscCall(MatRestoreRow_SeqAIJ(A, i, &nz, &cwork, &vwork));
2857: }
2858: PetscCall(PetscFree(cnew));
2860: (*B)->assembled = PETSC_FALSE;
2862: #if PetscDefined(HAVE_DEVICE)
2863: PetscCall(MatBindToCPU(*B, A->boundtocpu));
2864: #endif
2865: PetscCall(MatAssemblyBegin(*B, MAT_FINAL_ASSEMBLY));
2866: PetscCall(MatAssemblyEnd(*B, MAT_FINAL_ASSEMBLY));
2867: PetscCall(ISRestoreIndices(irowp, &row));
2868: PetscCall(ISRestoreIndices(icolp, &col));
2869: PetscCall(ISDestroy(&irowp));
2870: PetscCall(ISDestroy(&icolp));
2871: if (rowp == colp) PetscCall(MatPropagateSymmetryOptions(A, *B));
2872: PetscFunctionReturn(PETSC_SUCCESS);
2873: }
2875: PetscErrorCode MatCopy_SeqAIJ(Mat A, Mat B, MatStructure str)
2876: {
2877: PetscFunctionBegin;
2878: /* If the two matrices have the same copy implementation, use fast copy. */
2879: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2880: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2881: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2882: const PetscScalar *aa;
2883: PetscScalar *bb;
2885: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2886: PetscCall(MatSeqAIJGetArrayWrite(B, &bb));
2888: 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]);
2889: PetscCall(PetscArraycpy(bb, aa, a->i[A->rmap->n]));
2890: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2891: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2892: PetscCall(MatSeqAIJRestoreArrayWrite(B, &bb));
2893: } else {
2894: PetscCall(MatCopy_Basic(A, B, str));
2895: }
2896: PetscFunctionReturn(PETSC_SUCCESS);
2897: }
2899: PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A, PetscScalar *array[])
2900: {
2901: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2903: PetscFunctionBegin;
2904: *array = a->a;
2905: PetscFunctionReturn(PETSC_SUCCESS);
2906: }
2908: PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A, PetscScalar *array[])
2909: {
2910: PetscFunctionBegin;
2911: *array = NULL;
2912: PetscFunctionReturn(PETSC_SUCCESS);
2913: }
2915: /*
2916: Computes the number of nonzeros per row needed for preallocation when X and Y
2917: have different nonzero structure.
2918: */
2919: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m, const PetscInt *xi, const PetscInt *xj, const PetscInt *yi, const PetscInt *yj, PetscInt *nnz)
2920: {
2921: PetscInt i, j, k, nzx, nzy;
2923: PetscFunctionBegin;
2924: /* Set the number of nonzeros in the new matrix */
2925: for (i = 0; i < m; i++) {
2926: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2927: nzx = xi[i + 1] - xi[i];
2928: nzy = yi[i + 1] - yi[i];
2929: nnz[i] = 0;
2930: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2931: for (; k < nzy && yjj[k] < xjj[j]; k++) nnz[i]++; /* Catch up to X */
2932: if (k < nzy && yjj[k] == xjj[j]) k++; /* Skip duplicate */
2933: nnz[i]++;
2934: }
2935: for (; k < nzy; k++) nnz[i]++;
2936: }
2937: PetscFunctionReturn(PETSC_SUCCESS);
2938: }
2940: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y, Mat X, PetscInt *nnz)
2941: {
2942: PetscInt m = Y->rmap->N;
2943: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2944: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2946: PetscFunctionBegin;
2947: /* Set the number of nonzeros in the new matrix */
2948: PetscCall(MatAXPYGetPreallocation_SeqX_private(m, x->i, x->j, y->i, y->j, nnz));
2949: PetscFunctionReturn(PETSC_SUCCESS);
2950: }
2952: PetscErrorCode MatAXPY_SeqAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2953: {
2954: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data;
2956: PetscFunctionBegin;
2957: if (str == UNKNOWN_NONZERO_PATTERN || (PetscDefined(USE_DEBUG) && str == SAME_NONZERO_PATTERN)) {
2958: PetscBool e = x->nz == y->nz ? PETSC_TRUE : PETSC_FALSE;
2959: if (e) {
2960: PetscCall(PetscArraycmp(x->i, y->i, Y->rmap->n + 1, &e));
2961: if (e) {
2962: PetscCall(PetscArraycmp(x->j, y->j, y->nz, &e));
2963: if (e) str = SAME_NONZERO_PATTERN;
2964: }
2965: }
2966: if (!e) PetscCheck(str != SAME_NONZERO_PATTERN, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "MatStructure is not SAME_NONZERO_PATTERN");
2967: }
2968: if (str == SAME_NONZERO_PATTERN) {
2969: const PetscScalar *xa;
2970: PetscScalar *ya, alpha = a;
2971: PetscBLASInt one = 1, bnz;
2973: PetscCall(PetscBLASIntCast(x->nz, &bnz));
2974: PetscCall(MatSeqAIJGetArray(Y, &ya));
2975: PetscCall(MatSeqAIJGetArrayRead(X, &xa));
2976: PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, xa, &one, ya, &one));
2977: PetscCall(MatSeqAIJRestoreArrayRead(X, &xa));
2978: PetscCall(MatSeqAIJRestoreArray(Y, &ya));
2979: PetscCall(PetscLogFlops(2.0 * bnz));
2980: PetscCall(PetscObjectStateIncrease((PetscObject)Y));
2981: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2982: PetscCall(MatAXPY_Basic(Y, a, X, str));
2983: } else {
2984: Mat B;
2985: PetscInt *nnz;
2986: PetscBool subset = PETSC_TRUE;
2988: PetscCall(PetscMalloc1(Y->rmap->N, &nnz));
2989: PetscCall(MatAXPYGetPreallocation_SeqAIJ(Y, X, nnz));
2990: /* nnz[i] is the size of the union of row i of X and Y, so the pattern of X is contained in that of Y when every nnz[i] equals
2991: the length of row i of Y. Adding in place then keeps the nonzero pattern of Y, and its nonzero state, unchanged. */
2992: for (PetscInt i = 0; i < Y->rmap->N && subset; i++) subset = (PetscBool)(nnz[i] == y->i[i + 1] - y->i[i]);
2993: if (subset) PetscCall(MatAXPY_Basic(Y, a, X, SUBSET_NONZERO_PATTERN));
2994: else {
2995: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2996: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2997: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2998: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2999: PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz));
3000: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
3001: PetscCall(MatHeaderMerge(Y, &B));
3002: PetscCall(MatSeqAIJCheckInode(Y));
3003: }
3004: PetscCall(PetscFree(nnz));
3005: }
3006: PetscFunctionReturn(PETSC_SUCCESS);
3007: }
3009: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
3010: {
3011: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3012: PetscInt i, nz = aij->nz;
3013: PetscScalar *a;
3015: PetscFunctionBegin;
3016: PetscCall(MatSeqAIJGetArray(mat, &a));
3017: for (i = 0; i < nz; i++) a[i] = PetscConj(a[i]);
3018: PetscCall(MatSeqAIJRestoreArray(mat, &a));
3019: PetscFunctionReturn(PETSC_SUCCESS);
3020: }
3022: static PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3023: {
3024: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3025: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3026: PetscReal atmp;
3027: PetscScalar *x;
3028: const MatScalar *aa, *av;
3030: PetscFunctionBegin;
3031: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3032: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3033: aa = av;
3034: ai = a->i;
3035: aj = a->j;
3037: PetscCall(VecGetArrayWrite(v, &x));
3038: PetscCall(VecGetLocalSize(v, &n));
3039: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3040: for (i = 0; i < m; i++) {
3041: ncols = ai[1] - ai[0];
3042: ai++;
3043: x[i] = 0;
3044: for (j = 0; j < ncols; j++) {
3045: atmp = PetscAbsScalar(*aa);
3046: if (PetscAbsScalar(x[i]) < atmp) {
3047: x[i] = atmp;
3048: if (idx) idx[i] = *aj;
3049: }
3050: aa++;
3051: aj++;
3052: }
3053: }
3054: PetscCall(VecRestoreArrayWrite(v, &x));
3055: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3056: PetscFunctionReturn(PETSC_SUCCESS);
3057: }
3059: PETSC_INTERN PetscErrorCode MatGetRowSumAbs_SeqAIJ(Mat A, Vec v)
3060: {
3061: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3062: PetscInt i, j, m = A->rmap->n, *ai, ncols, n;
3063: PetscScalar *x;
3064: const MatScalar *aa, *av;
3066: PetscFunctionBegin;
3067: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3068: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3069: aa = av;
3070: ai = a->i;
3072: PetscCall(VecGetArrayWrite(v, &x));
3073: PetscCall(VecGetLocalSize(v, &n));
3074: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3075: for (i = 0; i < m; i++) {
3076: ncols = ai[1] - ai[0];
3077: ai++;
3078: x[i] = 0;
3079: for (j = 0; j < ncols; j++) {
3080: x[i] += PetscAbsScalar(*aa);
3081: aa++;
3082: }
3083: }
3084: PetscCall(VecRestoreArrayWrite(v, &x));
3085: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3086: PetscFunctionReturn(PETSC_SUCCESS);
3087: }
3089: static PetscErrorCode MatGetRowMax_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3090: {
3091: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3092: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3093: PetscScalar *x;
3094: const MatScalar *aa, *av;
3096: PetscFunctionBegin;
3097: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3098: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3099: aa = av;
3100: ai = a->i;
3101: aj = a->j;
3103: PetscCall(VecGetArrayWrite(v, &x));
3104: PetscCall(VecGetLocalSize(v, &n));
3105: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3106: for (i = 0; i < m; i++) {
3107: ncols = ai[1] - ai[0];
3108: ai++;
3109: if (ncols == A->cmap->n) { /* row is dense */
3110: x[i] = *aa;
3111: if (idx) idx[i] = 0;
3112: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
3113: x[i] = 0.0;
3114: if (idx) {
3115: for (j = 0; j < ncols; j++) { /* find first implicit 0.0 in the row */
3116: if (aj[j] > j) {
3117: idx[i] = j;
3118: break;
3119: }
3120: }
3121: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3122: if (j == ncols && j < A->cmap->n) idx[i] = j;
3123: }
3124: }
3125: for (j = 0; j < ncols; j++) {
3126: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {
3127: x[i] = *aa;
3128: if (idx) idx[i] = *aj;
3129: }
3130: aa++;
3131: aj++;
3132: }
3133: }
3134: PetscCall(VecRestoreArrayWrite(v, &x));
3135: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3136: PetscFunctionReturn(PETSC_SUCCESS);
3137: }
3139: static PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3140: {
3141: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3142: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3143: PetscScalar *x;
3144: const MatScalar *aa, *av;
3146: PetscFunctionBegin;
3147: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3148: aa = av;
3149: ai = a->i;
3150: aj = a->j;
3152: PetscCall(VecGetArrayWrite(v, &x));
3153: PetscCall(VecGetLocalSize(v, &n));
3154: PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector, %" PetscInt_FMT " vs. %" PetscInt_FMT " rows", m, n);
3155: for (i = 0; i < m; i++) {
3156: ncols = ai[1] - ai[0];
3157: ai++;
3158: if (ncols == A->cmap->n) { /* row is dense */
3159: x[i] = *aa;
3160: if (idx) idx[i] = 0;
3161: } else { /* row is sparse so already KNOW minimum is 0.0 or higher */
3162: x[i] = 0.0;
3163: if (idx) { /* find first implicit 0.0 in the row */
3164: for (j = 0; j < ncols; j++) {
3165: if (aj[j] > j) {
3166: idx[i] = j;
3167: break;
3168: }
3169: }
3170: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3171: if (j == ncols && j < A->cmap->n) idx[i] = j;
3172: }
3173: }
3174: for (j = 0; j < ncols; j++) {
3175: if (PetscAbsScalar(x[i]) > PetscAbsScalar(*aa)) {
3176: x[i] = *aa;
3177: if (idx) idx[i] = *aj;
3178: }
3179: aa++;
3180: aj++;
3181: }
3182: }
3183: PetscCall(VecRestoreArrayWrite(v, &x));
3184: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3185: PetscFunctionReturn(PETSC_SUCCESS);
3186: }
3188: static PetscErrorCode MatGetRowMin_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3189: {
3190: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3191: PetscInt i, j, m = A->rmap->n, ncols, n;
3192: const PetscInt *ai, *aj;
3193: PetscScalar *x;
3194: const MatScalar *aa, *av;
3196: PetscFunctionBegin;
3197: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3198: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3199: aa = av;
3200: ai = a->i;
3201: aj = a->j;
3203: PetscCall(VecGetArrayWrite(v, &x));
3204: PetscCall(VecGetLocalSize(v, &n));
3205: PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3206: for (i = 0; i < m; i++) {
3207: ncols = ai[1] - ai[0];
3208: ai++;
3209: if (ncols == A->cmap->n) { /* row is dense */
3210: x[i] = *aa;
3211: if (idx) idx[i] = 0;
3212: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
3213: x[i] = 0.0;
3214: if (idx) { /* find first implicit 0.0 in the row */
3215: for (j = 0; j < ncols; j++) {
3216: if (aj[j] > j) {
3217: idx[i] = j;
3218: break;
3219: }
3220: }
3221: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3222: if (j == ncols && j < A->cmap->n) idx[i] = j;
3223: }
3224: }
3225: for (j = 0; j < ncols; j++) {
3226: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {
3227: x[i] = *aa;
3228: if (idx) idx[i] = *aj;
3229: }
3230: aa++;
3231: aj++;
3232: }
3233: }
3234: PetscCall(VecRestoreArrayWrite(v, &x));
3235: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3236: PetscFunctionReturn(PETSC_SUCCESS);
3237: }
3239: static PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A, const PetscScalar **values)
3240: {
3241: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3242: PetscInt i, bs = A->rmap->bs, mbs = A->rmap->n / bs, ipvt[5], bs2 = bs * bs, *v_pivots, ij[7], *IJ, j;
3243: MatScalar *diag, work[25], *v_work;
3244: const PetscReal shift = 0.0;
3245: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
3247: PetscFunctionBegin;
3248: allowzeropivot = PetscNot(A->erroriffailure);
3249: if (a->ibdiag && a->ibdiagsize == bs2 * mbs && a->ibdiagState == ((PetscObject)A)->state) {
3250: if (values) *values = a->ibdiag;
3251: PetscFunctionReturn(PETSC_SUCCESS);
3252: }
3253: /* reallocate only when the length changes, so that the pointer stays valid for callers that
3254: hold on to it, such as PCSetUp_PBJacobi_Host() */
3255: if (!a->ibdiag || a->ibdiagsize != bs2 * mbs) {
3256: PetscCall(PetscFree(a->ibdiag));
3257: PetscCall(PetscMalloc1(bs2 * mbs, &a->ibdiag));
3258: a->ibdiagsize = bs2 * mbs;
3259: }
3260: diag = a->ibdiag;
3261: if (values) *values = a->ibdiag;
3262: /* factor and invert each block */
3263: switch (bs) {
3264: case 1:
3265: for (i = 0; i < mbs; i++) {
3266: PetscCall(MatGetValues(A, 1, &i, 1, &i, diag + i));
3267: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
3268: 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);
3269: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3270: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3271: A->factorerror_zeropivot_row = i;
3272: PetscCall(PetscInfo(A, "Zero pivot, row %" PetscInt_FMT " pivot %g tolerance %g\n", i, (double)PetscAbsScalar(diag[i]), (double)PETSC_MACHINE_EPSILON));
3273: }
3274: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3275: }
3276: break;
3277: case 2:
3278: for (i = 0; i < mbs; i++) {
3279: ij[0] = 2 * i;
3280: ij[1] = 2 * i + 1;
3281: PetscCall(MatGetValues(A, 2, ij, 2, ij, diag));
3282: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
3283: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3284: PetscCall(PetscKernel_A_gets_transpose_A_2(diag));
3285: diag += 4;
3286: }
3287: break;
3288: case 3:
3289: for (i = 0; i < mbs; i++) {
3290: ij[0] = 3 * i;
3291: ij[1] = 3 * i + 1;
3292: ij[2] = 3 * i + 2;
3293: PetscCall(MatGetValues(A, 3, ij, 3, ij, diag));
3294: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
3295: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3296: PetscCall(PetscKernel_A_gets_transpose_A_3(diag));
3297: diag += 9;
3298: }
3299: break;
3300: case 4:
3301: for (i = 0; i < mbs; i++) {
3302: ij[0] = 4 * i;
3303: ij[1] = 4 * i + 1;
3304: ij[2] = 4 * i + 2;
3305: ij[3] = 4 * i + 3;
3306: PetscCall(MatGetValues(A, 4, ij, 4, ij, diag));
3307: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
3308: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3309: PetscCall(PetscKernel_A_gets_transpose_A_4(diag));
3310: diag += 16;
3311: }
3312: break;
3313: case 5:
3314: for (i = 0; i < mbs; i++) {
3315: ij[0] = 5 * i;
3316: ij[1] = 5 * i + 1;
3317: ij[2] = 5 * i + 2;
3318: ij[3] = 5 * i + 3;
3319: ij[4] = 5 * i + 4;
3320: PetscCall(MatGetValues(A, 5, ij, 5, ij, diag));
3321: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
3322: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3323: PetscCall(PetscKernel_A_gets_transpose_A_5(diag));
3324: diag += 25;
3325: }
3326: break;
3327: case 6:
3328: for (i = 0; i < mbs; i++) {
3329: ij[0] = 6 * i;
3330: ij[1] = 6 * i + 1;
3331: ij[2] = 6 * i + 2;
3332: ij[3] = 6 * i + 3;
3333: ij[4] = 6 * i + 4;
3334: ij[5] = 6 * i + 5;
3335: PetscCall(MatGetValues(A, 6, ij, 6, ij, diag));
3336: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
3337: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3338: PetscCall(PetscKernel_A_gets_transpose_A_6(diag));
3339: diag += 36;
3340: }
3341: break;
3342: case 7:
3343: for (i = 0; i < mbs; i++) {
3344: ij[0] = 7 * i;
3345: ij[1] = 7 * i + 1;
3346: ij[2] = 7 * i + 2;
3347: ij[3] = 7 * i + 3;
3348: ij[4] = 7 * i + 4;
3349: ij[5] = 7 * i + 5;
3350: ij[6] = 7 * i + 6;
3351: PetscCall(MatGetValues(A, 7, ij, 7, ij, diag));
3352: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
3353: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3354: PetscCall(PetscKernel_A_gets_transpose_A_7(diag));
3355: diag += 49;
3356: }
3357: break;
3358: default:
3359: PetscCall(PetscMalloc3(bs, &v_work, bs, &v_pivots, bs, &IJ));
3360: for (i = 0; i < mbs; i++) {
3361: for (j = 0; j < bs; j++) IJ[j] = bs * i + j;
3362: PetscCall(MatGetValues(A, bs, IJ, bs, IJ, diag));
3363: PetscCall(PetscKernel_A_gets_inverse_A(bs, diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
3364: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3365: PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bs));
3366: diag += bs2;
3367: }
3368: PetscCall(PetscFree3(v_work, v_pivots, IJ));
3369: }
3370: a->ibdiagState = ((PetscObject)A)->state;
3371: PetscFunctionReturn(PETSC_SUCCESS);
3372: }
3374: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x, PetscRandom rctx)
3375: {
3376: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3377: PetscScalar a, *aa;
3378: PetscInt m, n, i, j, col;
3380: PetscFunctionBegin;
3381: if (!x->assembled) {
3382: PetscCall(MatGetSize(x, &m, &n));
3383: for (i = 0; i < m; i++) {
3384: for (j = 0; j < aij->imax[i]; j++) {
3385: PetscCall(PetscRandomGetValue(rctx, &a));
3386: col = (PetscInt)(n * PetscRealPart(a));
3387: PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES));
3388: }
3389: }
3390: } else {
3391: PetscCall(MatSeqAIJGetArrayWrite(x, &aa));
3392: for (i = 0; i < aij->nz; i++) PetscCall(PetscRandomGetValue(rctx, aa + i));
3393: PetscCall(MatSeqAIJRestoreArrayWrite(x, &aa));
3394: }
3395: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
3396: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
3397: PetscFunctionReturn(PETSC_SUCCESS);
3398: }
3400: /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */
3401: PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x, PetscInt low, PetscInt high, PetscRandom rctx)
3402: {
3403: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3404: PetscScalar a;
3405: PetscInt m, n, i, j, col, nskip;
3407: PetscFunctionBegin;
3408: nskip = high - low;
3409: PetscCall(MatGetSize(x, &m, &n));
3410: n -= nskip; /* shrink number of columns where nonzeros can be set */
3411: for (i = 0; i < m; i++) {
3412: for (j = 0; j < aij->imax[i]; j++) {
3413: PetscCall(PetscRandomGetValue(rctx, &a));
3414: col = (PetscInt)(n * PetscRealPart(a));
3415: if (col >= low) col += nskip; /* shift col rightward to skip the hole */
3416: PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES));
3417: }
3418: }
3419: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
3420: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
3421: PetscFunctionReturn(PETSC_SUCCESS);
3422: }
3424: static struct _MatOps MatOps_Values = {MatSetValues_SeqAIJ,
3425: MatGetRow_SeqAIJ,
3426: MatRestoreRow_SeqAIJ,
3427: MatMult_SeqAIJ,
3428: /* 4*/ MatMultAdd_SeqAIJ,
3429: MatMultTranspose_SeqAIJ,
3430: MatMultTransposeAdd_SeqAIJ,
3431: NULL,
3432: NULL,
3433: NULL,
3434: /* 10*/ NULL,
3435: MatLUFactor_SeqAIJ,
3436: NULL,
3437: MatSOR_SeqAIJ,
3438: MatTranspose_SeqAIJ,
3439: /* 15*/ MatGetInfo_SeqAIJ,
3440: MatEqual_SeqAIJ,
3441: MatGetDiagonal_SeqAIJ,
3442: MatDiagonalScale_SeqAIJ,
3443: MatNorm_SeqAIJ,
3444: /* 20*/ NULL,
3445: MatAssemblyEnd_SeqAIJ,
3446: MatSetOption_SeqAIJ,
3447: MatZeroEntries_SeqAIJ,
3448: /* 24*/ MatZeroRows_SeqAIJ,
3449: NULL,
3450: NULL,
3451: NULL,
3452: NULL,
3453: /* 29*/ MatSetUp_Seq_Hash,
3454: NULL,
3455: NULL,
3456: NULL,
3457: NULL,
3458: /* 34*/ MatDuplicate_SeqAIJ,
3459: NULL,
3460: NULL,
3461: MatILUFactor_SeqAIJ,
3462: NULL,
3463: /* 39*/ MatAXPY_SeqAIJ,
3464: MatCreateSubMatrices_SeqAIJ,
3465: MatIncreaseOverlap_SeqAIJ,
3466: MatGetValues_SeqAIJ,
3467: MatCopy_SeqAIJ,
3468: /* 44*/ MatGetRowMax_SeqAIJ,
3469: MatScale_SeqAIJ,
3470: MatShift_SeqAIJ,
3471: MatDiagonalSet_SeqAIJ,
3472: MatZeroRowsColumns_SeqAIJ,
3473: /* 49*/ MatSetRandom_SeqAIJ,
3474: MatGetRowIJ_SeqAIJ,
3475: MatRestoreRowIJ_SeqAIJ,
3476: MatGetColumnIJ_SeqAIJ,
3477: MatRestoreColumnIJ_SeqAIJ,
3478: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3479: NULL,
3480: NULL,
3481: MatPermute_SeqAIJ,
3482: NULL,
3483: /* 59*/ NULL,
3484: MatDestroy_SeqAIJ,
3485: MatView_SeqAIJ,
3486: NULL,
3487: NULL,
3488: /* 64*/ MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3489: NULL,
3490: NULL,
3491: NULL,
3492: MatGetRowMaxAbs_SeqAIJ,
3493: /* 69*/ MatGetRowMinAbs_SeqAIJ,
3494: NULL,
3495: NULL,
3496: MatFDColoringApply_AIJ,
3497: NULL,
3498: /* 74*/ MatFindZeroDiagonals_SeqAIJ,
3499: NULL,
3500: NULL,
3501: NULL,
3502: MatLoad_SeqAIJ,
3503: /* 79*/ NULL,
3504: NULL,
3505: NULL,
3506: NULL,
3507: NULL,
3508: /* 84*/ NULL,
3509: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3510: MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy,
3511: NULL,
3512: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3513: /* 90*/ NULL,
3514: MatProductSetFromOptions_SeqAIJ,
3515: NULL,
3516: NULL,
3517: MatConjugate_SeqAIJ,
3518: /* 94*/ NULL,
3519: MatSetValuesRow_SeqAIJ,
3520: MatRealPart_SeqAIJ,
3521: MatImaginaryPart_SeqAIJ,
3522: NULL,
3523: /* 99*/ NULL,
3524: MatMatSolve_SeqAIJ,
3525: NULL,
3526: MatGetRowMin_SeqAIJ,
3527: NULL,
3528: /*104*/ NULL,
3529: NULL,
3530: NULL,
3531: NULL,
3532: NULL,
3533: /*109*/ NULL,
3534: NULL,
3535: NULL,
3536: NULL,
3537: MatGetMultiProcBlock_SeqAIJ,
3538: /*114*/ MatFindNonzeroRows_SeqAIJ,
3539: MatGetColumnReductions_SeqAIJ,
3540: MatInvertBlockDiagonal_SeqAIJ,
3541: MatInvertVariableBlockDiagonal_SeqAIJ,
3542: NULL,
3543: /*119*/ NULL,
3544: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3545: MatTransposeColoringCreate_SeqAIJ,
3546: MatTransColoringApplySpToDen_SeqAIJ,
3547: MatTransColoringApplyDenToSp_SeqAIJ,
3548: /*124*/ MatRARtNumeric_SeqAIJ_SeqAIJ,
3549: NULL,
3550: NULL,
3551: MatFDColoringSetUp_SeqXAIJ,
3552: MatFindOffBlockDiagonalEntries_SeqAIJ,
3553: /*129*/ MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3554: MatDestroySubMatrices_SeqAIJ,
3555: NULL,
3556: NULL,
3557: MatCreateGraph_Simple_AIJ,
3558: /*134*/ MatTransposeSymbolic_SeqAIJ,
3559: MatEliminateZeros_SeqAIJ,
3560: MatGetRowSumAbs_SeqAIJ,
3561: NULL,
3562: NULL,
3563: /*139*/ NULL,
3564: MatCopyHashToXAIJ_Seq_Hash,
3565: NULL,
3566: NULL,
3567: NULL,
3568: /*144*/ NULL,
3569: NULL,
3570: NULL,
3571: NULL};
3573: static PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat, PetscInt *indices)
3574: {
3575: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3576: PetscInt i, nz, n;
3578: PetscFunctionBegin;
3579: nz = aij->maxnz;
3580: n = mat->rmap->n;
3581: for (i = 0; i < nz; i++) aij->j[i] = indices[i];
3582: aij->nz = nz;
3583: for (i = 0; i < n; i++) aij->ilen[i] = aij->imax[i];
3584: PetscFunctionReturn(PETSC_SUCCESS);
3585: }
3587: /*
3588: * Given a sparse matrix with global column indices, compact it by using a local column space.
3589: * The result matrix helps saving memory in other algorithms, such as MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable()
3590: */
3591: PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping)
3592: {
3593: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3594: PetscHMapI gid1_lid1;
3595: PetscHashIter tpos;
3596: PetscInt gid, lid, i, ec, nz = aij->nz;
3597: PetscInt *garray, *jj = aij->j;
3599: PetscFunctionBegin;
3601: PetscAssertPointer(mapping, 2);
3602: /* use a table */
3603: PetscCall(PetscHMapICreateWithSize(mat->rmap->n, &gid1_lid1));
3604: ec = 0;
3605: for (i = 0; i < nz; i++) {
3606: PetscInt data, gid1 = jj[i] + 1;
3607: PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &data));
3608: if (!data) {
3609: /* one based table */
3610: PetscCall(PetscHMapISet(gid1_lid1, gid1, ++ec));
3611: }
3612: }
3613: /* form array of columns we need */
3614: PetscCall(PetscMalloc1(ec, &garray));
3615: PetscHashIterBegin(gid1_lid1, tpos);
3616: while (!PetscHashIterAtEnd(gid1_lid1, tpos)) {
3617: PetscHashIterGetKey(gid1_lid1, tpos, gid);
3618: PetscHashIterGetVal(gid1_lid1, tpos, lid);
3619: PetscHashIterNext(gid1_lid1, tpos);
3620: gid--;
3621: lid--;
3622: garray[lid] = gid;
3623: }
3624: PetscCall(PetscSortInt(ec, garray)); /* sort, and rebuild */
3625: PetscCall(PetscHMapIClear(gid1_lid1));
3626: for (i = 0; i < ec; i++) PetscCall(PetscHMapISet(gid1_lid1, garray[i] + 1, i + 1));
3627: /* compact out the extra columns in B */
3628: for (i = 0; i < nz; i++) {
3629: PetscInt gid1 = jj[i] + 1;
3630: PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &lid));
3631: lid--;
3632: jj[i] = lid;
3633: }
3634: PetscCall(PetscLayoutDestroy(&mat->cmap));
3635: PetscCall(PetscHMapIDestroy(&gid1_lid1));
3636: PetscCall(PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat), ec, ec, 1, &mat->cmap));
3637: PetscCall(ISLocalToGlobalMappingCreate(PETSC_COMM_SELF, mat->cmap->bs, mat->cmap->n, garray, PETSC_OWN_POINTER, mapping));
3638: PetscCall(ISLocalToGlobalMappingSetType(*mapping, ISLOCALTOGLOBALMAPPINGHASH));
3639: PetscFunctionReturn(PETSC_SUCCESS);
3640: }
3642: /*@
3643: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3644: in the matrix.
3646: Input Parameters:
3647: + mat - the `MATSEQAIJ` matrix
3648: - indices - the column indices
3650: Level: advanced
3652: Notes:
3653: This can be called if you have precomputed the nonzero structure of the
3654: matrix and want to provide it to the matrix object to improve the performance
3655: of the `MatSetValues()` operation.
3657: You MUST have set the correct numbers of nonzeros per row in the call to
3658: `MatCreateSeqAIJ()`, and the columns indices MUST be sorted.
3660: MUST be called before any calls to `MatSetValues()`
3662: The indices should start with zero, not one.
3664: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`
3665: @*/
3666: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat, PetscInt *indices)
3667: {
3668: PetscFunctionBegin;
3670: PetscAssertPointer(indices, 2);
3671: PetscUseMethod(mat, "MatSeqAIJSetColumnIndices_C", (Mat, PetscInt *), (mat, indices));
3672: PetscFunctionReturn(PETSC_SUCCESS);
3673: }
3675: static PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3676: {
3677: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3678: size_t nz = aij->i[mat->rmap->n];
3680: PetscFunctionBegin;
3681: PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3683: /* allocate space for values if not already there */
3684: if (!aij->saved_values) PetscCall(PetscMalloc1(nz + 1, &aij->saved_values));
3686: /* copy values over */
3687: PetscCall(PetscArraycpy(aij->saved_values, aij->a, nz));
3688: PetscFunctionReturn(PETSC_SUCCESS);
3689: }
3691: /*@
3692: MatStoreValues - Stashes a copy of the matrix values; this allows reusing of the linear part of a Jacobian, while recomputing only the
3693: nonlinear portion.
3695: Logically Collect
3697: Input Parameter:
3698: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3700: Level: advanced
3702: Example Usage:
3703: .vb
3704: Using SNES
3705: Create Jacobian matrix
3706: Set linear terms into matrix
3707: Apply boundary conditions to matrix, at this time matrix must have
3708: final nonzero structure (i.e. setting the nonlinear terms and applying
3709: boundary conditions again will not change the nonzero structure
3710: MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE);
3711: MatStoreValues(mat);
3712: Call SNESSetJacobian() with matrix
3713: In your Jacobian routine
3714: MatRetrieveValues(mat);
3715: Set nonlinear terms in matrix
3717: Without `SNESSolve()`, i.e. when you handle nonlinear solve yourself:
3718: // build linear portion of Jacobian
3719: MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE);
3720: MatStoreValues(mat);
3721: loop over nonlinear iterations
3722: MatRetrieveValues(mat);
3723: // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3724: // call MatAssemblyBegin/End() on matrix
3725: Solve linear system with Jacobian
3726: endloop
3727: .ve
3729: Notes:
3730: Matrix must already be assembled before calling this routine
3731: Must set the matrix option `MatSetOption`(mat,`MAT_NEW_NONZERO_LOCATIONS`,`PETSC_FALSE`); before
3732: calling this routine.
3734: When this is called multiple times it overwrites the previous set of stored values
3735: and does not allocated additional space.
3737: .seealso: [](ch_matrices), `Mat`, `MatRetrieveValues()`
3738: @*/
3739: PetscErrorCode MatStoreValues(Mat mat)
3740: {
3741: PetscFunctionBegin;
3743: PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
3744: PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3745: PetscUseMethod(mat, "MatStoreValues_C", (Mat), (mat));
3746: PetscFunctionReturn(PETSC_SUCCESS);
3747: }
3749: static PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3750: {
3751: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3752: PetscInt nz = aij->i[mat->rmap->n];
3754: PetscFunctionBegin;
3755: PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3756: PetscCheck(aij->saved_values, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatStoreValues(A);first");
3757: /* copy values over */
3758: PetscCall(PetscArraycpy(aij->a, aij->saved_values, nz));
3759: PetscFunctionReturn(PETSC_SUCCESS);
3760: }
3762: /*@
3763: MatRetrieveValues - Retrieves the copy of the matrix values that was stored with `MatStoreValues()`
3765: Logically Collect
3767: Input Parameter:
3768: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3770: Level: advanced
3772: .seealso: [](ch_matrices), `Mat`, `MatStoreValues()`
3773: @*/
3774: PetscErrorCode MatRetrieveValues(Mat mat)
3775: {
3776: PetscFunctionBegin;
3778: PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
3779: PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3780: PetscUseMethod(mat, "MatRetrieveValues_C", (Mat), (mat));
3781: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
3782: PetscFunctionReturn(PETSC_SUCCESS);
3783: }
3785: /*@
3786: MatCreateSeqAIJ - Creates a sparse matrix in `MATSEQAIJ` (compressed row) format
3787: (the default parallel PETSc format). For good matrix assembly performance
3788: the user should preallocate the matrix storage by setting the parameter `nz`
3789: (or the array `nnz`).
3791: Collective
3793: Input Parameters:
3794: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3795: . m - number of rows
3796: . n - number of columns
3797: . nz - number of nonzeros per row (same for all rows)
3798: - nnz - array containing the number of nonzeros in the various rows
3799: (possibly different for each row) or NULL
3801: Output Parameter:
3802: . A - the matrix
3804: Options Database Keys:
3805: + -mat_no_inode - Do not use inodes
3806: - -mat_inode_limit limit - Sets inode limit (max limit=5)
3808: Level: intermediate
3810: Notes:
3811: It is recommend to use `MatCreateFromOptions()` instead of this routine
3813: If `nnz` is given then `nz` is ignored
3815: The `MATSEQAIJ` format, also called
3816: compressed row storage, is fully compatible with standard Fortran
3817: storage. That is, the stored row and column indices can begin at
3818: either one (as in Fortran) or zero.
3820: Specify the preallocated storage with either `nz` or `nnz` (not both).
3821: Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3822: allocation.
3824: By default, this format uses inodes (identical nodes) when possible, to
3825: improve numerical efficiency of matrix-vector products and solves. We
3826: search for consecutive rows with the same nonzero structure, thereby
3827: reusing matrix information to achieve increased efficiency.
3829: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`
3830: @*/
3831: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
3832: {
3833: PetscFunctionBegin;
3834: PetscCall(MatCreate(comm, A));
3835: PetscCall(MatSetSizes(*A, m, n, m, n));
3836: PetscCall(MatSetType(*A, MATSEQAIJ));
3837: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, nnz));
3838: PetscFunctionReturn(PETSC_SUCCESS);
3839: }
3841: /*@
3842: MatSeqAIJSetPreallocation - For good matrix assembly performance
3843: the user should preallocate the matrix storage by setting the parameter nz
3844: (or the array nnz). By setting these parameters accurately, performance
3845: during matrix assembly can be increased by more than a factor of 50.
3847: Collective
3849: Input Parameters:
3850: + B - The matrix
3851: . nz - number of nonzeros per row (same for all rows)
3852: - nnz - array containing the number of nonzeros in the various rows
3853: (possibly different for each row) or NULL
3855: Options Database Keys:
3856: + -mat_no_inode - Do not use inodes
3857: - -mat_inode_limit limit - Sets inode limit (max limit=5)
3859: Level: intermediate
3861: Notes:
3862: If `nnz` is given then `nz` is ignored
3864: The `MATSEQAIJ` format also called
3865: compressed row storage, is fully compatible with standard Fortran
3866: storage. That is, the stored row and column indices can begin at
3867: either one (as in Fortran) or zero. See the users' manual for details.
3869: Specify the preallocated storage with either `nz` or `nnz` (not both).
3870: Set nz = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3871: allocation.
3873: You can call `MatGetInfo()` to get information on how effective the preallocation was;
3874: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3875: You can also run with the option -info and look for messages with the string
3876: malloc in them to see if additional memory allocation was needed.
3878: Developer Notes:
3879: Use nz of `MAT_SKIP_ALLOCATION` to not allocate any space for the matrix
3880: entries or columns indices
3882: By default, this format uses inodes (identical nodes) when possible, to
3883: improve numerical efficiency of matrix-vector products and solves. We
3884: search for consecutive rows with the same nonzero structure, thereby
3885: reusing matrix information to achieve increased efficiency.
3887: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MatGetInfo()`,
3888: `MatSeqAIJSetTotalPreallocation()`
3889: @*/
3890: PetscErrorCode MatSeqAIJSetPreallocation(Mat B, PetscInt nz, const PetscInt nnz[])
3891: {
3892: PetscFunctionBegin;
3895: PetscTryMethod(B, "MatSeqAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[]), (B, nz, nnz));
3896: PetscFunctionReturn(PETSC_SUCCESS);
3897: }
3899: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B, PetscInt nz, const PetscInt *nnz)
3900: {
3901: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
3902: PetscBool skipallocation = PETSC_FALSE, realalloc = PETSC_FALSE;
3903: PetscInt i;
3905: PetscFunctionBegin;
3906: if (B->hash_active) {
3907: B->ops[0] = b->cops;
3908: PetscCall(PetscHMapIJVDestroy(&b->ht));
3909: PetscCall(PetscFree(b->dnz));
3910: B->hash_active = PETSC_FALSE;
3911: }
3912: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3913: if (nz == MAT_SKIP_ALLOCATION) {
3914: skipallocation = PETSC_TRUE;
3915: nz = 0;
3916: }
3917: PetscCall(PetscLayoutSetUp(B->rmap));
3918: PetscCall(PetscLayoutSetUp(B->cmap));
3920: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3921: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nz cannot be less than 0: value %" PetscInt_FMT, nz);
3922: if (nnz) {
3923: for (i = 0; i < B->rmap->n; i++) {
3924: 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]);
3925: 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);
3926: }
3927: }
3929: B->preallocated = PETSC_TRUE;
3930: if (!skipallocation) {
3931: if (!b->imax) PetscCall(PetscMalloc1(B->rmap->n, &b->imax));
3932: if (!b->ilen) {
3933: /* b->ilen will count nonzeros in each row so far. */
3934: PetscCall(PetscCalloc1(B->rmap->n, &b->ilen));
3935: } else {
3936: PetscCall(PetscMemzero(b->ilen, B->rmap->n * sizeof(PetscInt)));
3937: }
3938: if (!b->ipre) PetscCall(PetscMalloc1(B->rmap->n, &b->ipre));
3939: if (!nnz) {
3940: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
3941: else if (nz < 0) nz = 1;
3942: nz = PetscMin(nz, B->cmap->n);
3943: for (i = 0; i < B->rmap->n; i++) b->imax[i] = nz;
3944: PetscCall(PetscIntMultError(nz, B->rmap->n, &nz));
3945: } else {
3946: PetscInt64 nz64 = 0;
3947: for (i = 0; i < B->rmap->n; i++) {
3948: b->imax[i] = nnz[i];
3949: nz64 += nnz[i];
3950: }
3951: PetscCall(PetscIntCast(nz64, &nz));
3952: }
3954: /* allocate the matrix space */
3955: PetscCall(MatSeqXAIJFreeAIJ(B, &b->a, &b->j, &b->i));
3956: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscInt), (void **)&b->j));
3957: PetscCall(PetscShmgetAllocateArray(B->rmap->n + 1, sizeof(PetscInt), (void **)&b->i));
3958: b->free_ij = PETSC_TRUE;
3959: if (B->structure_only) {
3960: b->free_a = PETSC_FALSE;
3961: } else {
3962: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscScalar), (void **)&b->a));
3963: b->free_a = PETSC_TRUE;
3964: }
3965: b->i[0] = 0;
3966: for (i = 1; i < B->rmap->n + 1; i++) b->i[i] = b->i[i - 1] + b->imax[i - 1];
3967: } else {
3968: b->free_a = PETSC_FALSE;
3969: b->free_ij = PETSC_FALSE;
3970: }
3972: if (b->ipre && nnz != b->ipre && b->imax) {
3973: /* reserve user-requested sparsity */
3974: PetscCall(PetscArraycpy(b->ipre, b->imax, B->rmap->n));
3975: }
3977: b->nz = 0;
3978: b->maxnz = nz;
3979: B->info.nz_unneeded = (double)b->maxnz;
3980: if (realalloc) PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
3981: B->was_assembled = PETSC_FALSE;
3982: B->assembled = PETSC_FALSE;
3983: /* We simply deem preallocation has changed nonzero state. Updating the state
3984: will give clients (like AIJKokkos) a chance to know something has happened.
3985: */
3986: B->nonzerostate++;
3987: PetscFunctionReturn(PETSC_SUCCESS);
3988: }
3990: PetscErrorCode MatResetPreallocation_SeqAIJ_Private(Mat A, PetscBool *memoryreset)
3991: {
3992: Mat_SeqAIJ *a;
3993: PetscInt i;
3994: PetscBool skipreset;
3996: PetscFunctionBegin;
3999: 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()");
4000: if (A->num_ass == 0) PetscFunctionReturn(PETSC_SUCCESS);
4002: /* Check local size. If zero, then return */
4003: if (!A->rmap->n) PetscFunctionReturn(PETSC_SUCCESS);
4005: a = (Mat_SeqAIJ *)A->data;
4006: /* if no saved info, we error out */
4007: PetscCheck(a->ipre, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "No saved preallocation info ");
4009: PetscCheck(a->i && a->imax && a->ilen, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "Memory info is incomplete, and cannot reset preallocation ");
4011: PetscCall(PetscArraycmp(a->ipre, a->ilen, A->rmap->n, &skipreset));
4012: if (skipreset) PetscCall(MatZeroEntries(A));
4013: else {
4014: PetscCall(PetscArraycpy(a->imax, a->ipre, A->rmap->n));
4015: PetscCall(PetscArrayzero(a->ilen, A->rmap->n));
4016: a->i[0] = 0;
4017: for (i = 1; i < A->rmap->n + 1; i++) a->i[i] = a->i[i - 1] + a->imax[i - 1];
4018: A->preallocated = PETSC_TRUE;
4019: a->nz = 0;
4020: a->maxnz = a->i[A->rmap->n];
4021: A->info.nz_unneeded = (double)a->maxnz;
4022: A->was_assembled = PETSC_FALSE;
4023: A->assembled = PETSC_FALSE;
4024: A->nonzerostate++;
4025: /* Log that the state of this object has changed; this will help guarantee that preconditioners get re-setup */
4026: PetscCall(PetscObjectStateIncrease((PetscObject)A));
4027: }
4028: if (memoryreset) *memoryreset = (PetscBool)!skipreset;
4029: PetscFunctionReturn(PETSC_SUCCESS);
4030: }
4032: static PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A)
4033: {
4034: PetscFunctionBegin;
4035: PetscCall(MatResetPreallocation_SeqAIJ_Private(A, NULL));
4036: PetscFunctionReturn(PETSC_SUCCESS);
4037: }
4039: /*@
4040: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in `MATSEQAIJ` format.
4042: Input Parameters:
4043: + B - the matrix
4044: . i - the indices into `j` for the start of each row (indices start with zero)
4045: . j - the column indices for each row (indices start with zero) these must be sorted for each row
4046: - v - optional values in the matrix, use `NULL` if not provided
4048: Level: developer
4050: Notes:
4051: The `i`,`j`,`v` values are COPIED with this routine; to avoid the copy use `MatCreateSeqAIJWithArrays()`
4053: This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero
4054: structure will be the union of all the previous nonzero structures.
4056: Developer Notes:
4057: 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
4058: then just copies the `v` values directly with `PetscMemcpy()`.
4060: This routine could also take a `PetscCopyMode` argument to allow sharing the values instead of always copying them.
4062: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`, `MATSEQAIJ`, `MatResetPreallocation()`
4063: @*/
4064: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4065: {
4066: PetscFunctionBegin;
4069: PetscTryMethod(B, "MatSeqAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4070: PetscFunctionReturn(PETSC_SUCCESS);
4071: }
4073: static PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4074: {
4075: PetscInt i;
4076: PetscInt m, n;
4077: PetscInt nz;
4078: PetscInt *nnz;
4080: PetscFunctionBegin;
4081: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]);
4083: PetscCall(PetscLayoutSetUp(B->rmap));
4084: PetscCall(PetscLayoutSetUp(B->cmap));
4086: PetscCall(MatGetSize(B, &m, &n));
4087: PetscCall(PetscMalloc1(m + 1, &nnz));
4088: for (i = 0; i < m; i++) {
4089: nz = Ii[i + 1] - Ii[i];
4090: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative number of columns %" PetscInt_FMT, i, nz);
4091: nnz[i] = nz;
4092: }
4093: PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz));
4094: PetscCall(PetscFree(nnz));
4096: 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));
4098: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
4099: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
4101: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
4102: PetscFunctionReturn(PETSC_SUCCESS);
4103: }
4105: /*@
4106: MatSeqAIJKron - Computes `C`, the Kronecker product of `A` and `B`.
4108: Input Parameters:
4109: + A - left-hand side matrix
4110: . B - right-hand side matrix
4111: - reuse - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
4113: Output Parameter:
4114: . C - Kronecker product of `A` and `B`
4116: Level: intermediate
4118: Note:
4119: `MAT_REUSE_MATRIX` can only be used when the nonzero structure of the product matrix has not changed from that last call to `MatSeqAIJKron()`.
4121: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATKAIJ`, `MatReuse`
4122: @*/
4123: PetscErrorCode MatSeqAIJKron(Mat A, Mat B, MatReuse reuse, Mat *C)
4124: {
4125: PetscFunctionBegin;
4130: PetscAssertPointer(C, 4);
4131: if (reuse == MAT_REUSE_MATRIX) {
4134: }
4135: PetscTryMethod(A, "MatSeqAIJKron_C", (Mat, Mat, MatReuse, Mat *), (A, B, reuse, C));
4136: PetscFunctionReturn(PETSC_SUCCESS);
4137: }
4139: static PetscErrorCode MatSeqAIJKron_SeqAIJ(Mat A, Mat B, MatReuse reuse, Mat *C)
4140: {
4141: Mat newmat;
4142: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
4143: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
4144: PetscScalar *v;
4145: const PetscScalar *aa, *ba;
4146: 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;
4147: PetscBool flg;
4149: PetscFunctionBegin;
4150: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
4151: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
4152: PetscCheck(!B->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
4153: PetscCheck(B->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
4154: PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJ, &flg));
4155: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatType %s", ((PetscObject)B)->type_name);
4156: PetscCheck(reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatReuse %d", (int)reuse);
4157: if (reuse == MAT_INITIAL_MATRIX) {
4158: PetscCall(PetscMalloc2(am * bm + 1, &i, a->i[am] * b->i[bm], &j));
4159: PetscCall(MatCreate(PETSC_COMM_SELF, &newmat));
4160: PetscCall(MatSetSizes(newmat, am * bm, an * bn, am * bm, an * bn));
4161: PetscCall(MatSetType(newmat, MATAIJ));
4162: i[0] = 0;
4163: for (m = 0; m < am; ++m) {
4164: for (p = 0; p < bm; ++p) {
4165: i[m * bm + p + 1] = i[m * bm + p] + (a->i[m + 1] - a->i[m]) * (b->i[p + 1] - b->i[p]);
4166: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4167: for (q = b->i[p]; q < b->i[p + 1]; ++q) j[nnz++] = a->j[n] * bn + b->j[q];
4168: }
4169: }
4170: }
4171: PetscCall(MatSeqAIJSetPreallocationCSR(newmat, i, j, NULL));
4172: *C = newmat;
4173: PetscCall(PetscFree2(i, j));
4174: nnz = 0;
4175: }
4176: PetscCall(MatSeqAIJGetArray(*C, &v));
4177: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
4178: PetscCall(MatSeqAIJGetArrayRead(B, &ba));
4179: for (m = 0; m < am; ++m) {
4180: for (p = 0; p < bm; ++p) {
4181: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4182: for (q = b->i[p]; q < b->i[p + 1]; ++q) v[nnz++] = aa[n] * ba[q];
4183: }
4184: }
4185: }
4186: PetscCall(MatSeqAIJRestoreArray(*C, &v));
4187: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
4188: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
4189: PetscFunctionReturn(PETSC_SUCCESS);
4190: }
4192: #include <../src/mat/impls/dense/seq/dense.h>
4193: #include <petsc/private/kernels/petscaxpy.h>
4195: /*
4196: Computes (B'*A')' since computing B*A directly is untenable
4198: n p p
4199: [ ] [ ] [ ]
4200: m [ A ] * n [ B ] = m [ C ]
4201: [ ] [ ] [ ]
4203: */
4204: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A, Mat B, Mat C)
4205: {
4206: Mat_SeqDense *sub_a = (Mat_SeqDense *)A->data;
4207: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ *)B->data;
4208: Mat_SeqDense *sub_c = (Mat_SeqDense *)C->data;
4209: PetscInt i, j, n, m, q, p;
4210: const PetscInt *ii, *idx;
4211: const PetscScalar *b, *a, *a_q;
4212: PetscScalar *c, *c_q;
4213: PetscInt clda = sub_c->lda;
4214: PetscInt alda = sub_a->lda;
4216: PetscFunctionBegin;
4217: m = A->rmap->n;
4218: n = A->cmap->n;
4219: p = B->cmap->n;
4220: a = sub_a->v;
4221: b = sub_b->a;
4222: c = sub_c->v;
4223: if (clda == m) {
4224: PetscCall(PetscArrayzero(c, m * p));
4225: } else {
4226: for (j = 0; j < p; j++)
4227: for (i = 0; i < m; i++) c[j * clda + i] = 0.0;
4228: }
4229: ii = sub_b->i;
4230: idx = sub_b->j;
4231: for (i = 0; i < n; i++) {
4232: q = ii[i + 1] - ii[i];
4233: while (q-- > 0) {
4234: c_q = c + clda * (*idx);
4235: a_q = a + alda * i;
4236: PetscKernelAXPY(c_q, *b, a_q, m);
4237: idx++;
4238: b++;
4239: }
4240: }
4241: PetscFunctionReturn(PETSC_SUCCESS);
4242: }
4244: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A, Mat B, PetscReal fill, Mat C)
4245: {
4246: PetscInt m = A->rmap->n, n = B->cmap->n;
4247: PetscBool cisdense;
4249: PetscFunctionBegin;
4250: 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);
4251: PetscCall(MatSetSizes(C, m, n, m, n));
4252: PetscCall(MatSetBlockSizesFromMats(C, A, B));
4253: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, MATSEQDENSEHIP, ""));
4254: if (!cisdense) {
4255: PetscCall(MatSetType(C, MATDENSE));
4256: PetscCall(MatSetVecType(C, A->defaultvectype));
4257: }
4258: PetscCall(MatSetUp(C));
4260: C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
4261: PetscFunctionReturn(PETSC_SUCCESS);
4262: }
4264: /*MC
4265: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
4266: based on compressed sparse row format.
4268: Options Database Key:
4269: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
4271: Level: beginner
4273: Notes:
4274: `MatSetValues()` may be called with a `NULL` argument for the numerical values to insert zeros at the supplied row and column indices.
4276: Call `MatSetOption(A, MAT_STRUCTURE_ONLY, PETSC_TRUE)` before preallocation or `MatSetUp()` to store only the nonzero pattern.
4277: The assembled matrix has no numerical value array. Row and column indices supplied during insertion are retained, while numerical values are ignored.
4278: Such matrices can be used for structural operations, but not for numerical operations.
4280: Developer Note:
4281: It would be nice if all matrix formats supported passing `NULL` in for the numerical values
4283: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MatSetFromOptions()`, `MatSetType()`, `MatCreate()`, `MatType`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4284: M*/
4286: /*MC
4287: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
4289: This matrix type is identical to `MATSEQAIJ` when constructed with a single process communicator,
4290: and `MATMPIAIJ` otherwise. As a result, for single process communicators,
4291: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
4292: for communicators controlling multiple processes. It is recommended that you call both of
4293: the above preallocation routines for simplicity.
4295: Options Database Key:
4296: . -mat_type aij - sets the matrix type to "aij" during a call to `MatSetFromOptions()`
4298: Level: beginner
4300: Notes:
4301: Call `MatSetOption(A, MAT_STRUCTURE_ONLY, PETSC_TRUE)` before preallocation or `MatSetUp()` to store only the nonzero pattern.
4302: The assembled matrix has no numerical value array. Row and column indices supplied during insertion are retained, while numerical values are ignored.
4303: Such matrices can be used for structural operations, but not for numerical operations.
4305: Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, and also automatically switches over to use inodes when
4306: enough exist.
4308: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4309: M*/
4311: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *);
4312: #if PetscDefined(HAVE_ELEMENTAL)
4313: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
4314: #endif
4315: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
4316: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
4317: #endif
4318: #if PetscDefined(HAVE_HYPRE)
4319: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A, MatType, MatReuse, Mat *);
4320: #endif
4322: PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat, MatType, MatReuse, Mat *);
4323: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
4324: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
4326: /*@
4327: MatSeqAIJGetArray - gives read/write access to the array where the data for a `MATSEQAIJ` matrix is stored
4329: Not Collective
4331: Input Parameter:
4332: . A - a `MATSEQAIJ` matrix
4334: Output Parameter:
4335: . array - pointer to the data
4337: Level: intermediate
4339: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`
4340: @*/
4341: PetscErrorCode MatSeqAIJGetArray(Mat A, PetscScalar *array[])
4342: {
4343: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4345: PetscFunctionBegin;
4346: if (A->structure_only || aij->ops->getarray == NULL) *array = aij->a;
4347: else PetscCall((*aij->ops->getarray)(A, array));
4348: PetscFunctionReturn(PETSC_SUCCESS);
4349: }
4351: /*@
4352: MatSeqAIJRestoreArray - returns access to the array where the data for a `MATSEQAIJ` matrix is stored obtained by `MatSeqAIJGetArray()`
4354: Not Collective
4356: Input Parameters:
4357: + A - a `MATSEQAIJ` matrix
4358: - array - pointer to the data
4360: Level: intermediate
4362: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`
4363: @*/
4364: PetscErrorCode MatSeqAIJRestoreArray(Mat A, PetscScalar *array[])
4365: {
4366: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4368: PetscFunctionBegin;
4369: if (A->structure_only || aij->ops->restorearray == NULL) *array = NULL;
4370: else PetscCall((*aij->ops->restorearray)(A, array));
4371: PetscCall(PetscObjectStateIncrease((PetscObject)A));
4372: PetscFunctionReturn(PETSC_SUCCESS);
4373: }
4375: /*@
4376: MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4378: Not Collective
4380: Input Parameter:
4381: . A - a `MATSEQAIJ` matrix
4383: Output Parameter:
4384: . array - pointer to the data
4386: Level: intermediate
4388: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()`
4389: @*/
4390: PetscErrorCode MatSeqAIJGetArrayRead(Mat A, const PetscScalar *array[])
4391: {
4392: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4394: PetscFunctionBegin;
4395: if (A->structure_only || aij->ops->getarrayread == NULL) *array = aij->a;
4396: else PetscCall((*aij->ops->getarrayread)(A, array));
4397: PetscFunctionReturn(PETSC_SUCCESS);
4398: }
4400: /*@
4401: MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJGetArrayRead()`
4403: Not Collective
4405: Input Parameter:
4406: . A - a `MATSEQAIJ` matrix
4408: Output Parameter:
4409: . array - pointer to the data
4411: Level: intermediate
4413: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4414: @*/
4415: PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A, const PetscScalar *array[])
4416: {
4417: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4419: PetscFunctionBegin;
4420: if (A->structure_only || aij->ops->restorearrayread == NULL) *array = NULL;
4421: else PetscCall((*aij->ops->restorearrayread)(A, array));
4422: PetscFunctionReturn(PETSC_SUCCESS);
4423: }
4425: /*@
4426: MatSeqAIJGetArrayWrite - gives write-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4428: Not Collective
4430: Input Parameter:
4431: . A - a `MATSEQAIJ` matrix
4433: Output Parameter:
4434: . array - pointer to the data
4436: Level: intermediate
4438: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayWrite()`
4439: @*/
4440: PetscErrorCode MatSeqAIJGetArrayWrite(Mat A, PetscScalar *array[])
4441: {
4442: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4444: PetscFunctionBegin;
4445: if (A->structure_only || aij->ops->getarraywrite == NULL) *array = aij->a;
4446: else PetscCall((*aij->ops->getarraywrite)(A, array));
4447: PetscCall(PetscObjectStateIncrease((PetscObject)A));
4448: PetscFunctionReturn(PETSC_SUCCESS);
4449: }
4451: /*@
4452: MatSeqAIJRestoreArrayWrite - restore the write-only access array obtained from `MatSeqAIJGetArrayWrite()`
4454: Not Collective
4456: Input Parameter:
4457: . A - a `MATSEQAIJ` matrix
4459: Output Parameter:
4460: . array - pointer to the data
4462: Level: intermediate
4464: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayWrite()`
4465: @*/
4466: PetscErrorCode MatSeqAIJRestoreArrayWrite(Mat A, PetscScalar *array[])
4467: {
4468: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4470: PetscFunctionBegin;
4471: if (A->structure_only || aij->ops->restorearraywrite == NULL) *array = NULL;
4472: else PetscCall((*aij->ops->restorearraywrite)(A, array));
4473: PetscFunctionReturn(PETSC_SUCCESS);
4474: }
4476: /*@
4477: MatSeqAIJGetCSRAndMemType - Get the CSR arrays and the memory type of the `MATSEQAIJ` matrix
4479: Not Collective; No Fortran Support
4481: Input Parameter:
4482: . mat - a matrix of type `MATSEQAIJ` or its subclasses
4484: Output Parameters:
4485: + i - row map array of the matrix
4486: . j - column index array of the matrix
4487: . a - data array of the matrix
4488: - mtype - memory type of the arrays
4490: Level: developer
4492: Notes:
4493: Any of the output parameters can be `NULL`, in which case the corresponding value is not returned.
4494: If mat is a device matrix, the arrays are on the device. Otherwise, they are on the host.
4496: One can call this routine on a preallocated but not assembled matrix to just get the memory of the CSR underneath the matrix.
4497: If the matrix is assembled, the data array `a` is guaranteed to have the latest values of the matrix.
4499: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4500: @*/
4501: PetscErrorCode MatSeqAIJGetCSRAndMemType(Mat mat, const PetscInt *i[], const PetscInt *j[], PetscScalar *a[], PetscMemType *mtype)
4502: {
4503: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
4505: PetscFunctionBegin;
4506: PetscCheck(mat->preallocated, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "matrix is not preallocated");
4507: if (aij->ops->getcsrandmemtype) {
4508: PetscCall((*aij->ops->getcsrandmemtype)(mat, i, j, a, mtype));
4509: } else {
4510: if (i) *i = aij->i;
4511: if (j) *j = aij->j;
4512: if (a) *a = aij->a;
4513: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
4514: }
4515: PetscFunctionReturn(PETSC_SUCCESS);
4516: }
4518: /*@
4519: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
4521: Not Collective
4523: Input Parameter:
4524: . A - a `MATSEQAIJ` matrix
4526: Output Parameter:
4527: . nz - the maximum number of nonzeros in any row
4529: Level: intermediate
4531: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`
4532: @*/
4533: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A, PetscInt *nz)
4534: {
4535: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4537: PetscFunctionBegin;
4538: *nz = aij->rmax;
4539: PetscFunctionReturn(PETSC_SUCCESS);
4540: }
4542: static PetscErrorCode MatCOOStructDestroy_SeqAIJ(PetscCtxRt data)
4543: {
4544: MatCOOStruct_SeqAIJ *coo = *(MatCOOStruct_SeqAIJ **)data;
4546: PetscFunctionBegin;
4547: PetscCall(PetscFree(coo->perm));
4548: PetscCall(PetscFree(coo->jmap));
4549: PetscCall(PetscFree(coo));
4550: PetscFunctionReturn(PETSC_SUCCESS);
4551: }
4553: PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
4554: {
4555: MPI_Comm comm;
4556: PetscInt *i, *j;
4557: PetscInt M, N, row, iprev;
4558: PetscCount k, p, q, nneg, nnz, start, end; /* Index the coo array, so use PetscCount as their type */
4559: PetscInt *Ai; /* Change to PetscCount once we use it for row pointers */
4560: PetscInt *Aj;
4561: PetscScalar *Aa = NULL;
4562: Mat_SeqAIJ *seqaij = (Mat_SeqAIJ *)mat->data;
4563: MatType rtype;
4564: PetscCount *perm, *jmap;
4565: MatCOOStruct_SeqAIJ *coo;
4566: PetscBool isorted;
4567: PetscBool hypre;
4569: PetscFunctionBegin;
4570: // The CSR storage built below replaces the hash table, so MatSetValues() must use the regular SeqAIJ operations
4571: if (mat->hash_active) {
4572: mat->ops[0] = seqaij->cops;
4573: PetscCall(PetscHMapIJVDestroy(&seqaij->ht));
4574: PetscCall(PetscFree(seqaij->dnz));
4575: mat->hash_active = PETSC_FALSE;
4576: }
4577: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
4578: PetscCall(MatGetSize(mat, &M, &N));
4579: i = coo_i;
4580: j = coo_j;
4581: PetscCall(PetscMalloc1(coo_n, &perm));
4583: /* 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) */
4584: isorted = PETSC_TRUE;
4585: iprev = PETSC_INT_MIN;
4586: for (k = 0; k < coo_n; k++) {
4587: if (j[k] < 0) i[k] = -1;
4588: if (isorted) {
4589: if (i[k] < iprev) isorted = PETSC_FALSE;
4590: else iprev = i[k];
4591: }
4592: perm[k] = k;
4593: }
4595: /* Sort by row if not already */
4596: if (!isorted) PetscCall(PetscSortIntWithIntCountArrayPair(coo_n, i, j, perm));
4597: 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);
4599: /* Advance k to the first row with a non-negative index */
4600: for (k = 0; k < coo_n; k++)
4601: if (i[k] >= 0) break;
4602: nneg = k;
4603: 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 */
4604: nnz = 0; /* Total number of unique nonzeros to be counted */
4605: jmap++; /* Inc jmap by 1 for convenience */
4607: PetscCall(PetscShmgetAllocateArray(M + 1, sizeof(PetscInt), (void **)&Ai)); /* CSR of A */
4608: PetscCall(PetscArrayzero(Ai, M + 1));
4609: PetscCall(PetscShmgetAllocateArray(coo_n - nneg, sizeof(PetscInt), (void **)&Aj)); /* We have at most coo_n-nneg unique nonzeros */
4611: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
4613: /* In each row, sort by column, then unique column indices to get row length */
4614: Ai++; /* Inc by 1 for convenience */
4615: q = 0; /* q-th unique nonzero, with q starting from 0 */
4616: while (k < coo_n) {
4617: PetscBool strictly_sorted; // this row is strictly sorted?
4618: PetscInt jprev;
4620: /* get [start,end) indices for this row; also check if cols in this row are strictly sorted */
4621: row = i[k];
4622: start = k;
4623: jprev = PETSC_INT_MIN;
4624: strictly_sorted = PETSC_TRUE;
4625: while (k < coo_n && i[k] == row) {
4626: if (strictly_sorted) {
4627: if (j[k] <= jprev) strictly_sorted = PETSC_FALSE;
4628: else jprev = j[k];
4629: }
4630: k++;
4631: }
4632: end = k;
4634: /* hack for HYPRE: swap min column to diag so that diagonal values will go first */
4635: if (hypre) {
4636: PetscInt minj = PETSC_INT_MAX;
4637: PetscBool hasdiag = PETSC_FALSE;
4639: if (strictly_sorted) { // fast path to swap the first and the diag
4640: PetscCount tmp;
4641: for (p = start; p < end; p++) {
4642: if (j[p] == row && p != start) {
4643: j[p] = j[start]; // swap j[], so that the diagonal value will go first (manipulated by perm[])
4644: j[start] = row;
4645: tmp = perm[start];
4646: perm[start] = perm[p]; // also swap perm[] so we can save the call to PetscSortIntWithCountArray() below
4647: perm[p] = tmp;
4648: break;
4649: }
4650: }
4651: } else {
4652: for (p = start; p < end; p++) {
4653: hasdiag = (PetscBool)(hasdiag || (j[p] == row));
4654: minj = PetscMin(minj, j[p]);
4655: }
4657: if (hasdiag) {
4658: for (p = start; p < end; p++) {
4659: if (j[p] == minj) j[p] = row;
4660: else if (j[p] == row) j[p] = minj;
4661: }
4662: }
4663: }
4664: }
4665: // sort by columns in a row. perm[] indicates their original order
4666: if (!strictly_sorted) PetscCall(PetscSortIntWithCountArray(end - start, j + start, perm + start));
4667: 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);
4669: if (strictly_sorted) { // fast path to set Aj[], jmap[], Ai[], nnz, q
4670: for (p = start; p < end; p++, q++) {
4671: Aj[q] = j[p];
4672: jmap[q] = 1;
4673: }
4674: PetscCall(PetscIntCast(end - start, Ai + row));
4675: nnz += Ai[row]; // q is already advanced
4676: } else {
4677: /* Find number of unique col entries in this row */
4678: Aj[q] = j[start]; /* Log the first nonzero in this row */
4679: jmap[q] = 1; /* Number of repeats of this nonzero entry */
4680: Ai[row] = 1;
4681: nnz++;
4683: for (p = start + 1; p < end; p++) { /* Scan remaining nonzero in this row */
4684: if (j[p] != j[p - 1]) { /* Meet a new nonzero */
4685: q++;
4686: jmap[q] = 1;
4687: Aj[q] = j[p];
4688: Ai[row]++;
4689: nnz++;
4690: } else {
4691: jmap[q]++;
4692: }
4693: }
4694: q++; /* Move to next row and thus next unique nonzero */
4695: }
4696: }
4698: Ai--; /* Back to the beginning of Ai[] */
4699: for (k = 0; k < M; k++) Ai[k + 1] += Ai[k];
4700: jmap--; // Back to the beginning of jmap[]
4701: jmap[0] = 0;
4702: for (k = 0; k < nnz; k++) jmap[k + 1] += jmap[k];
4704: if (nnz < coo_n - nneg) { /* Reallocate with actual number of unique nonzeros */
4705: PetscCount *jmap_new;
4706: PetscInt *Aj_new;
4708: PetscCall(PetscMalloc1(nnz + 1, &jmap_new));
4709: PetscCall(PetscArraycpy(jmap_new, jmap, nnz + 1));
4710: PetscCall(PetscFree(jmap));
4711: jmap = jmap_new;
4713: PetscCall(PetscShmgetAllocateArray(nnz, sizeof(PetscInt), (void **)&Aj_new));
4714: PetscCall(PetscArraycpy(Aj_new, Aj, nnz));
4715: PetscCall(PetscShmgetDeallocateArray((void **)&Aj));
4716: Aj = Aj_new;
4717: }
4719: if (nneg) { /* Discard heading entries with negative indices in perm[], as we'll access it from index 0 in MatSetValuesCOO */
4720: PetscCount *perm_new;
4722: PetscCall(PetscMalloc1(coo_n - nneg, &perm_new));
4723: PetscCall(PetscArraycpy(perm_new, perm + nneg, coo_n - nneg));
4724: PetscCall(PetscFree(perm));
4725: perm = perm_new;
4726: }
4728: PetscCall(MatGetRootType_Private(mat, &rtype));
4729: if (!mat->structure_only) {
4730: PetscCall(PetscShmgetAllocateArray(nnz, sizeof(PetscScalar), (void **)&Aa));
4731: PetscCall(PetscArrayzero(Aa, nnz));
4732: }
4733: PetscCall(MatSetSeqAIJWithArrays_private(PETSC_COMM_SELF, M, N, Ai, Aj, Aa, rtype, mat));
4735: seqaij->free_a = (PetscBool)!mat->structure_only;
4736: seqaij->free_ij = PETSC_TRUE; /* Let mat own Ai, Aj and any allocated Aa */
4738: // Put the COO struct in a container and then attach that to the matrix
4739: PetscCall(PetscMalloc1(1, &coo));
4740: PetscCall(PetscIntCast(nnz, &coo->nz));
4741: coo->n = coo_n;
4742: coo->Atot = coo_n - nneg; // Annz is seqaij->nz, so no need to record that again
4743: coo->jmap = jmap; // of length nnz+1
4744: coo->perm = perm;
4745: // MatSetPreallocationCOO() increments the nonzero state once more after this routine returns; if that changes, every
4746: // MatSetValuesCOO() fails the check in MatSeqAIJCheckCOONonzeroState(), so a mismatch cannot go unnoticed
4747: coo->nonzerostate = mat->nonzerostate + 1;
4748: PetscCall(PetscObjectContainerCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", coo, MatCOOStructDestroy_SeqAIJ));
4749: PetscFunctionReturn(PETSC_SUCCESS);
4750: }
4752: // PetscClangLinter pragma disable: -fdoc-internal-linkage
4753: /*@
4754: MatSeqAIJCheckCOONonzeroState - Checks that the COO maps built by `MatSetPreallocationCOO()` can still be used by
4755: `MatSetValuesCOO()` on a `MATSEQAIJ` matrix
4757: Not Collective
4759: Input Parameter:
4760: . A - the `MATSEQAIJ` matrix
4762: Level: developer
4764: Notes:
4765: The maps address the nonzeros of the matrix by position, so they are only valid while its nonzero pattern is the one built by
4766: `MatSetPreallocationCOO()`. Any change to the pattern, for example a new nonzero inserted with `MatSetValues()`, would make
4767: `MatSetValuesCOO()` write values into the wrong locations, so this routine generates an error instead. It also generates an
4768: error when `MatSetPreallocationCOO()` has not been called.
4770: The `MatSetValuesCOO()` implementations of `MATSEQAIJ` and its device subtypes call this routine before using the maps.
4772: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`, `MatMPIAIJCheckCOONonzeroState()`
4773: @*/
4774: PetscErrorCode MatSeqAIJCheckCOONonzeroState(Mat A)
4775: {
4776: PetscContainer container;
4777: MatCOOStruct_SeqAIJ *coo;
4779: PetscFunctionBegin;
4780: PetscCall(PetscObjectQuery((PetscObject)A, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
4781: PetscCheck(container, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatSetPreallocationCOO() must be called before MatSetValuesCOO()");
4782: PetscCall(PetscContainerGetPointer(container, &coo));
4783: PetscCheck(coo->nonzerostate == A->nonzerostate, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "The nonzero pattern of the matrix has changed since MatSetPreallocationCOO(), for example because MatSetValues() inserted a new nonzero; call MatSetPreallocationCOO() again before MatSetValuesCOO(), or stop inserting new nonzeros with MatSetValues()");
4784: PetscFunctionReturn(PETSC_SUCCESS);
4785: }
4787: static PetscErrorCode MatSetValuesCOO_SeqAIJ(Mat A, const PetscScalar v[], InsertMode imode)
4788: {
4789: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)A->data;
4790: PetscCount i, j, Annz = aseq->nz;
4791: PetscCount *perm, *jmap;
4792: PetscScalar *Aa;
4793: PetscContainer container;
4794: MatCOOStruct_SeqAIJ *coo;
4796: PetscFunctionBegin;
4797: PetscCall(MatSeqAIJCheckCOONonzeroState(A));
4798: PetscCall(PetscObjectQuery((PetscObject)A, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
4799: PetscCall(PetscContainerGetPointer(container, &coo));
4800: perm = coo->perm;
4801: jmap = coo->jmap;
4802: PetscCall(MatSeqAIJGetArray(A, &Aa));
4803: for (i = 0; i < Annz; i++) {
4804: PetscScalar sum = 0.0;
4805: for (j = jmap[i]; j < jmap[i + 1]; j++) sum += v[perm[j]];
4806: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
4807: }
4808: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
4809: PetscFunctionReturn(PETSC_SUCCESS);
4810: }
4812: #if PetscDefined(HAVE_CUDA)
4813: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
4814: #endif
4815: #if PetscDefined(HAVE_HIP)
4816: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
4817: #endif
4818: #if PetscDefined(HAVE_KOKKOS_KERNELS)
4819: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJKokkos(Mat, MatType, MatReuse, Mat *);
4820: #endif
4822: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
4823: {
4824: Mat_SeqAIJ *b;
4825: PetscMPIInt size;
4827: PetscFunctionBegin;
4828: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
4829: PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Comm must be of size 1");
4831: PetscCall(PetscNew(&b));
4833: B->data = (void *)b;
4834: B->ops[0] = MatOps_Values;
4835: if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
4837: b->row = NULL;
4838: b->col = NULL;
4839: b->icol = NULL;
4840: b->reallocs = 0;
4841: b->ignorezeroentries = PETSC_FALSE;
4842: b->roworiented = PETSC_TRUE;
4843: b->nonew = 0;
4844: b->diag = NULL;
4845: b->solve_work = NULL;
4846: B->spptr = NULL;
4847: b->saved_values = NULL;
4848: b->idiag = NULL;
4849: b->mdiag = NULL;
4850: b->ssor_work = NULL;
4851: b->omega = 1.0;
4852: b->fshift = 0.0;
4853: b->ibdiag = NULL;
4854: b->keepnonzeropattern = PETSC_FALSE;
4856: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ));
4857: #if PetscDefined(HAVE_MATLAB)
4858: PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEnginePut_C", MatlabEnginePut_SeqAIJ));
4859: PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEngineGet_C", MatlabEngineGet_SeqAIJ));
4860: #endif
4861: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetColumnIndices_C", MatSeqAIJSetColumnIndices_SeqAIJ));
4862: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_SeqAIJ));
4863: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_SeqAIJ));
4864: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsbaij_C", MatConvert_SeqAIJ_SeqSBAIJ));
4865: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqbaij_C", MatConvert_SeqAIJ_SeqBAIJ));
4866: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijperm_C", MatConvert_SeqAIJ_SeqAIJPERM));
4867: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijsell_C", MatConvert_SeqAIJ_SeqAIJSELL));
4868: #if PetscDefined(HAVE_MKL_SPARSE)
4869: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijmkl_C", MatConvert_SeqAIJ_SeqAIJMKL));
4870: #endif
4871: #if PetscDefined(HAVE_CUDA)
4872: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcusparse_C", MatConvert_SeqAIJ_SeqAIJCUSPARSE));
4873: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4874: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJ));
4875: #endif
4876: #if PetscDefined(HAVE_HIP)
4877: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijhipsparse_C", MatConvert_SeqAIJ_SeqAIJHIPSPARSE));
4878: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4879: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", MatProductSetFromOptions_SeqAIJ));
4880: #endif
4881: #if PetscDefined(HAVE_KOKKOS_KERNELS)
4882: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijkokkos_C", MatConvert_SeqAIJ_SeqAIJKokkos));
4883: #endif
4884: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcrl_C", MatConvert_SeqAIJ_SeqAIJCRL));
4885: #if PetscDefined(HAVE_ELEMENTAL)
4886: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_elemental_C", MatConvert_SeqAIJ_Elemental));
4887: #endif
4888: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
4889: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
4890: #endif
4891: #if PetscDefined(HAVE_HYPRE)
4892: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_hypre_C", MatConvert_AIJ_HYPRE));
4893: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
4894: #endif
4895: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqdense_C", MatConvert_SeqAIJ_SeqDense));
4896: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsell_C", MatConvert_SeqAIJ_SeqSELL));
4897: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_is_C", MatConvert_XAIJ_IS));
4898: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_SeqAIJ));
4899: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsHermitianTranspose_C", MatIsHermitianTranspose_SeqAIJ));
4900: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocation_C", MatSeqAIJSetPreallocation_SeqAIJ));
4901: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_SeqAIJ));
4902: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetHash_C", MatResetHash_SeqAIJ));
4903: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocationCSR_C", MatSeqAIJSetPreallocationCSR_SeqAIJ));
4904: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatReorderForNonzeroDiagonal_C", MatReorderForNonzeroDiagonal_SeqAIJ));
4905: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_seqaij_C", MatProductSetFromOptions_IS_XAIJ));
4906: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqaij_C", MatProductSetFromOptions_SeqDense_SeqAIJ));
4907: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4908: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJKron_C", MatSeqAIJKron_SeqAIJ));
4909: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJ));
4910: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJ));
4911: PetscCall(MatCreate_SeqAIJ_Inode(B));
4912: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ));
4913: PetscCall(MatSeqAIJSetTypeFromOptions(B)); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4914: PetscFunctionReturn(PETSC_SUCCESS);
4915: }
4917: /*
4918: Given a matrix generated with MatGetFactor() duplicates all the information in A into C
4919: */
4920: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C, Mat A, MatDuplicateOption cpvalues, PetscBool mallocmatspace)
4921: {
4922: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data, *a = (Mat_SeqAIJ *)A->data;
4923: PetscInt m = A->rmap->n, i;
4925: PetscFunctionBegin;
4926: PetscCheck(A->assembled || cpvalues == MAT_DO_NOT_COPY_VALUES, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot duplicate unassembled matrix");
4927: PetscCall(MatSetOption(C, MAT_STRUCTURE_ONLY, A->structure_only));
4929: C->factortype = A->factortype;
4930: c->row = NULL;
4931: c->col = NULL;
4932: c->icol = NULL;
4933: c->reallocs = 0;
4934: C->assembled = A->assembled;
4936: if (A->preallocated) {
4937: PetscCall(PetscLayoutReference(A->rmap, &C->rmap));
4938: PetscCall(PetscLayoutReference(A->cmap, &C->cmap));
4940: if (!A->hash_active) {
4941: PetscCall(PetscMalloc1(m, &c->imax));
4942: PetscCall(PetscArraycpy(c->imax, a->imax, m));
4943: PetscCall(PetscMalloc1(m, &c->ilen));
4944: PetscCall(PetscArraycpy(c->ilen, a->ilen, m));
4946: /* allocate the matrix space */
4947: if (mallocmatspace) {
4948: if (!A->structure_only) PetscCall(PetscShmgetAllocateArray(a->i[m], sizeof(PetscScalar), (void **)&c->a));
4949: PetscCall(PetscShmgetAllocateArray(a->i[m], sizeof(PetscInt), (void **)&c->j));
4950: PetscCall(PetscShmgetAllocateArray(m + 1, sizeof(PetscInt), (void **)&c->i));
4951: PetscCall(PetscArraycpy(c->i, a->i, m + 1));
4952: c->free_a = PETSC_TRUE;
4953: c->free_ij = PETSC_TRUE;
4954: if (m > 0) {
4955: PetscCall(PetscArraycpy(c->j, a->j, a->i[m]));
4956: if (!A->structure_only) {
4957: if (cpvalues == MAT_COPY_VALUES) {
4958: const PetscScalar *aa;
4960: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
4961: PetscCall(PetscArraycpy(c->a, aa, a->i[m]));
4962: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
4963: } else PetscCall(PetscArrayzero(c->a, a->i[m]));
4964: }
4965: }
4966: }
4967: C->preallocated = PETSC_TRUE;
4968: } else {
4969: PetscCheck(mallocmatspace, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Cannot malloc matrix memory from a non-preallocated matrix");
4970: PetscCall(MatSetUp(C));
4971: }
4973: c->ignorezeroentries = a->ignorezeroentries;
4974: c->roworiented = a->roworiented;
4975: c->nonew = a->nonew;
4976: c->solve_work = NULL;
4977: c->saved_values = NULL;
4978: c->idiag = NULL;
4979: c->ssor_work = NULL;
4980: c->keepnonzeropattern = a->keepnonzeropattern;
4982: c->rmax = a->rmax;
4983: c->nz = a->nz;
4984: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4986: c->compressedrow.use = a->compressedrow.use;
4987: c->compressedrow.nrows = a->compressedrow.nrows;
4988: if (a->compressedrow.use) {
4989: i = a->compressedrow.nrows;
4990: PetscCall(PetscMalloc2(i + 1, &c->compressedrow.i, i, &c->compressedrow.rindex));
4991: PetscCall(PetscArraycpy(c->compressedrow.i, a->compressedrow.i, i + 1));
4992: PetscCall(PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, i));
4993: } else {
4994: c->compressedrow.use = PETSC_FALSE;
4995: c->compressedrow.i = NULL;
4996: c->compressedrow.rindex = NULL;
4997: }
4998: c->nonzerorowcnt = a->nonzerorowcnt;
4999: C->nonzerostate = A->nonzerostate;
5001: PetscCall(MatDuplicate_SeqAIJ_Inode(A, cpvalues, &C));
5002: }
5003: PetscCall(PetscFunctionListDuplicate(((PetscObject)A)->qlist, &((PetscObject)C)->qlist));
5004: PetscFunctionReturn(PETSC_SUCCESS);
5005: }
5007: PetscErrorCode MatDuplicate_SeqAIJ(Mat A, MatDuplicateOption cpvalues, Mat *B)
5008: {
5009: PetscFunctionBegin;
5010: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
5011: PetscCall(MatSetSizes(*B, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n));
5012: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) PetscCall(MatSetBlockSizesFromMats(*B, A, A));
5013: PetscCall(MatSetType(*B, ((PetscObject)A)->type_name));
5014: PetscCall(MatDuplicateNoCreate_SeqAIJ(*B, A, cpvalues, PETSC_TRUE));
5015: PetscFunctionReturn(PETSC_SUCCESS);
5016: }
5018: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
5019: {
5020: PetscBool isbinary, ishdf5;
5022: PetscFunctionBegin;
5025: /* force binary viewer to load .info file if it has not yet done so */
5026: PetscCall(PetscViewerSetUp(viewer));
5027: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
5028: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
5029: if (isbinary) {
5030: PetscCall(MatLoad_SeqAIJ_Binary(newMat, viewer));
5031: } else if (ishdf5) {
5032: #if PetscDefined(HAVE_HDF5)
5033: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
5034: #else
5035: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
5036: #endif
5037: } else {
5038: 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);
5039: }
5040: PetscFunctionReturn(PETSC_SUCCESS);
5041: }
5043: PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
5044: {
5045: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->data;
5046: PetscInt header[4], *rowlens, M, N, nz, sum, rows, cols, i;
5048: PetscFunctionBegin;
5049: PetscCall(PetscViewerSetUp(viewer));
5051: /* read in matrix header */
5052: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
5053: PetscCheck(header[0] == MAT_FILE_CLASSID, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
5054: M = header[1];
5055: N = header[2];
5056: nz = header[3];
5057: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
5058: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
5059: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as SeqAIJ");
5061: /* set block sizes from the viewer's .info file */
5062: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
5063: /* set local and global sizes if not set already */
5064: if (mat->rmap->n < 0) mat->rmap->n = M;
5065: if (mat->cmap->n < 0) mat->cmap->n = N;
5066: if (mat->rmap->N < 0) mat->rmap->N = M;
5067: if (mat->cmap->N < 0) mat->cmap->N = N;
5068: PetscCall(PetscLayoutSetUp(mat->rmap));
5069: PetscCall(PetscLayoutSetUp(mat->cmap));
5071: /* check if the matrix sizes are correct */
5072: PetscCall(MatGetSize(mat, &rows, &cols));
5073: 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);
5075: /* read in row lengths */
5076: PetscCall(PetscMalloc1(M, &rowlens));
5077: PetscCall(PetscViewerBinaryRead(viewer, rowlens, M, NULL, PETSC_INT));
5078: /* check if sum(rowlens) is same as nz */
5079: sum = 0;
5080: for (i = 0; i < M; i++) sum += rowlens[i];
5081: 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);
5082: /* preallocate and check sizes */
5083: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(mat, 0, rowlens));
5084: PetscCall(MatGetSize(mat, &rows, &cols));
5085: 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);
5086: /* store row lengths */
5087: PetscCall(PetscArraycpy(a->ilen, rowlens, M));
5088: PetscCall(PetscFree(rowlens));
5090: /* fill in "i" row pointers */
5091: a->i[0] = 0;
5092: for (i = 0; i < M; i++) a->i[i + 1] = a->i[i] + a->ilen[i];
5093: /* read in "j" column indices */
5094: PetscCall(PetscViewerBinaryRead(viewer, a->j, nz, NULL, PETSC_INT));
5095: /* read in "a" nonzero values */
5096: PetscCall(PetscViewerBinaryRead(viewer, a->a, nz, NULL, PETSC_SCALAR));
5098: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
5099: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
5100: PetscFunctionReturn(PETSC_SUCCESS);
5101: }
5103: PetscErrorCode MatEqual_SeqAIJ(Mat A, Mat B, PetscBool *flg)
5104: {
5105: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data;
5106: const PetscScalar *aa, *ba;
5108: PetscFunctionBegin;
5109: /* If the matrix dimensions are not equal,or no of nonzeros */
5110: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) || (a->nz != b->nz)) {
5111: *flg = PETSC_FALSE;
5112: PetscFunctionReturn(PETSC_SUCCESS);
5113: }
5115: /* if the a->i are the same */
5116: PetscCall(PetscArraycmp(a->i, b->i, A->rmap->n + 1, flg));
5117: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
5119: /* if a->j are the same */
5120: PetscCall(PetscArraycmp(a->j, b->j, a->nz, flg));
5121: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
5123: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
5124: PetscCall(MatSeqAIJGetArrayRead(B, &ba));
5125: /* if a->a are the same */
5126: PetscCall(PetscArraycmp(aa, ba, a->nz, flg));
5127: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
5128: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
5129: PetscFunctionReturn(PETSC_SUCCESS);
5130: }
5132: /*@
5133: MatCreateSeqAIJWithArrays - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in CSR format)
5134: provided by the user.
5136: Collective
5138: Input Parameters:
5139: + comm - must be an MPI communicator of size 1
5140: . m - number of rows
5141: . n - number of columns
5142: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
5143: . j - column indices
5144: - a - matrix values
5146: Output Parameter:
5147: . mat - the matrix
5149: Level: intermediate
5151: Notes:
5152: The `i`, `j`, and `a` arrays are not copied by this routine, the user must free these arrays
5153: once the matrix is destroyed and not before
5155: You cannot set new nonzero locations into this matrix, that will generate an error.
5157: The `i` and `j` indices are 0 based
5159: The format which is used for the sparse matrix input, is equivalent to a
5160: row-major ordering.. i.e for the following matrix, the input data expected is
5161: as shown
5162: .vb
5163: 1 0 0
5164: 2 0 3
5165: 4 5 6
5167: i = {0,1,3,6} [size = nrow+1 = 3+1]
5168: j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
5169: v = {1,2,3,4,5,6} [size = 6]
5170: .ve
5172: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateMPIAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`
5173: @*/
5174: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat)
5175: {
5176: PetscInt ii;
5177: Mat_SeqAIJ *aij;
5179: PetscFunctionBegin;
5180: PetscCheck(m <= 0 || i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
5181: PetscCall(MatCreate(comm, mat));
5182: PetscCall(MatSetSizes(*mat, m, n, m, n));
5183: /* PetscCall(MatSetBlockSizes(*mat,,)); */
5184: PetscCall(MatSetType(*mat, MATSEQAIJ));
5185: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, MAT_SKIP_ALLOCATION, NULL));
5186: aij = (Mat_SeqAIJ *)(*mat)->data;
5187: PetscCall(PetscMalloc1(m, &aij->imax));
5188: PetscCall(PetscMalloc1(m, &aij->ilen));
5190: aij->i = i;
5191: aij->j = j;
5192: aij->a = a;
5193: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
5194: aij->free_a = PETSC_FALSE;
5195: aij->free_ij = PETSC_FALSE;
5197: for (ii = 0, aij->nonzerorowcnt = 0, aij->rmax = 0; ii < m; ii++) {
5198: aij->ilen[ii] = aij->imax[ii] = i[ii + 1] - i[ii];
5199: if (PetscDefined(USE_DEBUG)) {
5200: 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]);
5201: for (PetscInt jj = i[ii] + 1; jj < i[ii + 1]; jj++) {
5202: 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);
5203: 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);
5204: }
5205: }
5206: }
5207: if (PetscDefined(USE_DEBUG)) {
5208: for (ii = 0; ii < aij->i[m]; ii++) {
5209: PetscCheck(j[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative column index at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]);
5210: 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);
5211: }
5212: }
5214: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
5215: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
5216: PetscFunctionReturn(PETSC_SUCCESS);
5217: }
5219: /*@
5220: MatCreateSeqAIJFromTriple - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in COO format)
5221: provided by the user.
5223: Collective
5225: Input Parameters:
5226: + comm - must be an MPI communicator of size 1
5227: . m - number of rows
5228: . n - number of columns
5229: . i - row indices
5230: . j - column indices
5231: . a - matrix values
5232: . nz - number of nonzeros
5233: - idx - if the `i` and `j` indices start with 1 use `PETSC_TRUE` otherwise use `PETSC_FALSE`
5235: Output Parameter:
5236: . mat - the matrix
5238: Level: intermediate
5240: Example:
5241: For the following matrix, the input data expected is as shown (using 0 based indexing)
5242: .vb
5243: 1 0 0
5244: 2 0 3
5245: 4 5 6
5247: i = {0,1,1,2,2,2}
5248: j = {0,0,2,0,1,2}
5249: v = {1,2,3,4,5,6}
5250: .ve
5252: Note:
5253: Instead of using this function, users should also consider `MatSetPreallocationCOO()` and `MatSetValuesCOO()`, which allow repeated or remote entries,
5254: and are particularly useful in iterative applications.
5256: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateSeqAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`, `MatSetValuesCOO()`, `MatSetPreallocationCOO()`
5257: @*/
5258: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat, PetscCount nz, PetscBool idx)
5259: {
5260: PetscInt ii, *nnz, one = 1, row, col;
5262: PetscFunctionBegin;
5263: PetscCall(PetscCalloc1(m, &nnz));
5264: for (ii = 0; ii < nz; ii++) nnz[i[ii] - !!idx] += 1;
5265: PetscCall(MatCreate(comm, mat));
5266: PetscCall(MatSetSizes(*mat, m, n, m, n));
5267: PetscCall(MatSetType(*mat, MATSEQAIJ));
5268: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, 0, nnz));
5269: for (ii = 0; ii < nz; ii++) {
5270: if (idx) {
5271: row = i[ii] - 1;
5272: col = j[ii] - 1;
5273: } else {
5274: row = i[ii];
5275: col = j[ii];
5276: }
5277: PetscCall(MatSetValues(*mat, one, &row, one, &col, &a[ii], ADD_VALUES));
5278: }
5279: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
5280: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
5281: PetscCall(PetscFree(nnz));
5282: PetscFunctionReturn(PETSC_SUCCESS);
5283: }
5285: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
5286: {
5287: PetscFunctionBegin;
5288: PetscCall(MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm, inmat, n, scall, outmat));
5289: PetscFunctionReturn(PETSC_SUCCESS);
5290: }
5292: /*
5293: Permute A into C's *local* index space using rowemb,colemb.
5294: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
5295: of [0,m), colemb is in [0,n).
5296: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
5297: */
5298: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C, IS rowemb, IS colemb, MatStructure pattern, Mat B)
5299: {
5300: /* If making this function public, change the error returned in this function away from _PLIB. */
5301: Mat_SeqAIJ *Baij;
5302: PetscBool seqaij;
5303: PetscInt m, n, *nz, i, j, count;
5304: PetscScalar v;
5305: const PetscInt *rowindices, *colindices;
5307: PetscFunctionBegin;
5308: if (!B) PetscFunctionReturn(PETSC_SUCCESS);
5309: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
5310: PetscCall(PetscObjectBaseTypeCompare((PetscObject)B, MATSEQAIJ, &seqaij));
5311: PetscCheck(seqaij, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is of wrong type");
5312: if (rowemb) {
5313: PetscCall(ISGetLocalSize(rowemb, &m));
5314: 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);
5315: } else PetscCheck(C->rmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is row-incompatible with the target matrix");
5316: if (colemb) {
5317: PetscCall(ISGetLocalSize(colemb, &n));
5318: 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);
5319: } else PetscCheck(C->cmap->n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is col-incompatible with the target matrix");
5321: Baij = (Mat_SeqAIJ *)B->data;
5322: rowindices = NULL;
5323: if (rowemb) PetscCall(ISGetIndices(rowemb, &rowindices));
5324: if (pattern == DIFFERENT_NONZERO_PATTERN) {
5325: PetscCall(PetscMalloc1(C->rmap->n, &nz));
5326: if (rowemb) {
5327: PetscCall(PetscArrayzero(nz, C->rmap->n));
5328: for (i = 0; i < B->rmap->n; i++) nz[rowindices[i]] = Baij->i[i + 1] - Baij->i[i];
5329: } else {
5330: for (i = 0; i < B->rmap->n; i++) nz[i] = Baij->i[i + 1] - Baij->i[i];
5331: }
5332: PetscCall(MatSeqAIJSetPreallocation(C, 0, nz));
5333: PetscCall(PetscFree(nz));
5334: }
5335: if (pattern == SUBSET_NONZERO_PATTERN) PetscCall(MatZeroEntries(C));
5336: count = 0;
5337: colindices = NULL;
5338: if (colemb) PetscCall(ISGetIndices(colemb, &colindices));
5339: for (i = 0; i < B->rmap->n; i++) {
5340: PetscInt row;
5341: row = i;
5342: if (rowindices) row = rowindices[i];
5343: for (j = Baij->i[i]; j < Baij->i[i + 1]; j++) {
5344: PetscInt col;
5345: col = Baij->j[count];
5346: if (colindices) col = colindices[col];
5347: v = Baij->a[count];
5348: PetscCall(MatSetValues(C, 1, &row, 1, &col, &v, INSERT_VALUES));
5349: ++count;
5350: }
5351: }
5352: if (colemb) PetscCall(ISRestoreIndices(colemb, &colindices));
5353: if (rowemb) PetscCall(ISRestoreIndices(rowemb, &rowindices));
5354: /* FIXME: set C's nonzerostate correctly. */
5355: /* Assembly for C is necessary. */
5356: C->preallocated = PETSC_TRUE;
5357: C->assembled = PETSC_TRUE;
5358: C->was_assembled = PETSC_FALSE;
5359: PetscFunctionReturn(PETSC_SUCCESS);
5360: }
5362: PetscErrorCode MatEliminateZeros_SeqAIJ(Mat A, PetscBool keep)
5363: {
5364: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5365: MatScalar *aa = a->a;
5366: PetscInt m = A->rmap->n, fshift = 0, fshift_prev = 0, i, k;
5367: PetscInt *ailen = a->ilen, *imax = a->imax, *ai = a->i, *aj = a->j, rmax = 0;
5369: PetscFunctionBegin;
5370: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot eliminate zeros for unassembled matrix");
5371: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
5372: for (i = 1, a->nonzerorowcnt = 0; i <= m; i++) {
5373: /* move each nonzero entry back by the amount of zero slots (fshift) before it*/
5374: for (k = ai[i - 1]; k < ai[i]; k++) {
5375: if (aa[k] == 0 && (aj[k] != i - 1 || !keep)) fshift++;
5376: else {
5377: if (aa[k] == 0 && aj[k] == i - 1) PetscCall(PetscInfo(A, "Keep the diagonal zero at row %" PetscInt_FMT "\n", i - 1));
5378: aa[k - fshift] = aa[k];
5379: aj[k - fshift] = aj[k];
5380: }
5381: }
5382: ai[i - 1] -= fshift_prev; // safe to update ai[i-1] now since it will not be used in the next iteration
5383: fshift_prev = fshift;
5384: /* reset ilen and imax for each row */
5385: ailen[i - 1] = imax[i - 1] = ai[i] - fshift - ai[i - 1];
5386: a->nonzerorowcnt += ((ai[i] - fshift - ai[i - 1]) > 0);
5387: rmax = PetscMax(rmax, ailen[i - 1]);
5388: }
5389: if (fshift) {
5390: if (m) {
5391: ai[m] -= fshift;
5392: a->nz = ai[m];
5393: }
5394: 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));
5395: A->nonzerostate++;
5396: A->info.nz_unneeded += (PetscReal)fshift;
5397: a->rmax = rmax;
5398: if (a->inode.use && a->inode.checked) PetscCall(MatSeqAIJCheckInode(A));
5399: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
5400: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
5401: }
5402: PetscFunctionReturn(PETSC_SUCCESS);
5403: }
5405: PetscFunctionList MatSeqAIJList = NULL;
5407: /*@
5408: MatSeqAIJSetType - Converts a `MATSEQAIJ` matrix to a subtype
5410: Collective
5412: Input Parameters:
5413: + mat - the matrix object
5414: - matype - matrix type
5416: Options Database Key:
5417: . -mat_seqaij_type method - for example seqaijcrl
5419: Level: intermediate
5421: .seealso: [](ch_matrices), `Mat`, `PCSetType()`, `VecSetType()`, `MatCreate()`, `MatType`
5422: @*/
5423: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
5424: {
5425: PetscBool sametype;
5426: PetscErrorCode (*r)(Mat, MatType, MatReuse, Mat *);
5428: PetscFunctionBegin;
5430: PetscCall(PetscObjectTypeCompare((PetscObject)mat, matype, &sametype));
5431: if (sametype) PetscFunctionReturn(PETSC_SUCCESS);
5433: PetscCall(PetscFunctionListFind(MatSeqAIJList, matype, &r));
5434: PetscCheck(r, PetscObjectComm((PetscObject)mat), PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown Mat type given: %s", matype);
5435: PetscCall((*r)(mat, matype, MAT_INPLACE_MATRIX, &mat));
5436: PetscFunctionReturn(PETSC_SUCCESS);
5437: }
5439: /*@
5440: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential `MATSEQAIJ` matrices
5442: Not Collective, No Fortran Support
5444: Input Parameters:
5445: + sname - name of a new user-defined matrix type, for example `MATSEQAIJCRL`
5446: - function - routine to convert to subtype
5448: Level: advanced
5450: Notes:
5451: `MatSeqAIJRegister()` may be called multiple times to add several user-defined solvers.
5453: Then, your matrix can be chosen with the procedural interface at runtime via the option
5454: .vb
5455: -mat_seqaij_type my_mat
5456: .ve
5458: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRegisterAll()`
5459: @*/
5460: PetscErrorCode MatSeqAIJRegister(const char sname[], PetscErrorCode (*function)(Mat, MatType, MatReuse, Mat *))
5461: {
5462: PetscFunctionBegin;
5463: PetscCall(MatInitializePackage());
5464: PetscCall(PetscFunctionListAdd(&MatSeqAIJList, sname, function));
5465: PetscFunctionReturn(PETSC_SUCCESS);
5466: }
5468: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
5470: /*@
5471: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of `MATSSEQAIJ`
5473: Not Collective
5475: Level: advanced
5477: Note:
5478: This registers the versions of `MATSEQAIJ` for GPUs
5480: .seealso: [](ch_matrices), `Mat`, `MatRegisterAll()`, `MatSeqAIJRegister()`
5481: @*/
5482: PetscErrorCode MatSeqAIJRegisterAll(void)
5483: {
5484: PetscFunctionBegin;
5485: if (MatSeqAIJRegisterAllCalled) PetscFunctionReturn(PETSC_SUCCESS);
5486: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
5488: PetscCall(MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL));
5489: PetscCall(MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM));
5490: PetscCall(MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL));
5491: #if PetscDefined(HAVE_MKL_SPARSE)
5492: PetscCall(MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL));
5493: #endif
5494: #if PetscDefined(HAVE_CUDA)
5495: PetscCall(MatSeqAIJRegister(MATSEQAIJCUSPARSE, MatConvert_SeqAIJ_SeqAIJCUSPARSE));
5496: #endif
5497: #if PetscDefined(HAVE_HIP)
5498: PetscCall(MatSeqAIJRegister(MATSEQAIJHIPSPARSE, MatConvert_SeqAIJ_SeqAIJHIPSPARSE));
5499: #endif
5500: #if PetscDefined(HAVE_KOKKOS_KERNELS)
5501: PetscCall(MatSeqAIJRegister(MATSEQAIJKOKKOS, MatConvert_SeqAIJ_SeqAIJKokkos));
5502: #endif
5503: #if PetscDefined(HAVE_VIENNACL) && PetscDefined(HAVE_VIENNACL_NO_CUDA)
5504: PetscCall(MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL));
5505: #endif
5506: PetscFunctionReturn(PETSC_SUCCESS);
5507: }