Actual source code: sbaij.c
1: /*
2: Defines the basic matrix operations for the SBAIJ (compressed row)
3: matrix storage format.
4: */
5: #include <../src/mat/impls/baij/seq/baij.h>
6: #include <../src/mat/impls/sbaij/seq/sbaij.h>
7: #include <petsc/private/kernels/blocktranspose.h>
8: #include <petscblaslapack.h>
10: #include <../src/mat/impls/sbaij/seq/relax.h>
11: #define USESHORT
12: #include <../src/mat/impls/sbaij/seq/relax.h>
14: /* defines MatSetValues_Seq_Hash(), MatAssemblyEnd_Seq_Hash(), MatSetUp_Seq_Hash() */
15: #define TYPE SBAIJ
16: #define TYPE_SBAIJ
17: #define TYPE_BS
18: #include "../src/mat/impls/aij/seq/seqhashmatsetvalues.h"
19: #undef TYPE_BS
20: #define TYPE_BS _BS
21: #define TYPE_BS_ON
22: #include "../src/mat/impls/aij/seq/seqhashmatsetvalues.h"
23: #undef TYPE_BS
24: #undef TYPE_SBAIJ
25: #include "../src/mat/impls/aij/seq/seqhashmat.h"
26: #undef TYPE
27: #undef TYPE_BS_ON
29: #if PetscDefined(HAVE_ELEMENTAL)
30: PETSC_INTERN PetscErrorCode MatConvert_SeqSBAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
31: #endif
32: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
33: PETSC_INTERN PetscErrorCode MatConvert_SBAIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
34: #endif
35: PETSC_INTERN PetscErrorCode MatConvert_MPISBAIJ_Basic(Mat, MatType, MatReuse, Mat *);
37: MatGetDiagonalMarkers(SeqSBAIJ, A->rmap->bs)
39: static PetscErrorCode MatGetColumnReductions_SeqSBAIJ(Mat A, PetscInt type, PetscReal *reductions)
40: {
41: Mat_SeqSBAIJ *aij = (Mat_SeqSBAIJ *)A->data;
42: PetscInt m, n, bs = A->rmap->bs, bs2 = aij->bs2;
43: PetscBool implicit, hermitian;
44: const PetscInt *ai = aij->i, *aj = aij->j;
45: const MatScalar *aa = aij->a;
47: PetscFunctionBegin;
48: PetscCall(MatGetSize(A, &m, &n));
49: PetscCheck(type == NORM_2 || type == NORM_1 || type == NORM_INFINITY || type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART || type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
50: PetscCall(PetscArrayzero(reductions, n));
51: implicit = (PetscBool)(m == n && (A->symmetric == PETSC_BOOL3_TRUE || A->hermitian == PETSC_BOOL3_TRUE));
52: hermitian = (PetscBool)(implicit && PetscDefined(USE_COMPLEX) && A->hermitian == PETSC_BOOL3_TRUE);
53: for (PetscInt i = 0; i < aij->mbs; i++) {
54: for (PetscInt k = ai[i]; k < ai[i + 1]; k++) {
55: const PetscBool offdiag = (PetscBool)(implicit && aj[k] != i);
57: for (PetscInt jb = 0; jb < bs; jb++) {
58: for (PetscInt ib = 0; ib < bs; ib++) {
59: const MatScalar value = aa[k * bs2 + jb * bs + ib];
60: const PetscInt col = aj[k] * bs + jb, row = i * bs + ib;
61: PetscReal reduction;
63: if (type == NORM_2) reduction = PetscAbsScalar(value * value);
64: else if (type == NORM_1 || type == NORM_INFINITY) reduction = PetscAbsScalar(value);
65: else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) reduction = PetscRealPart(value);
66: else reduction = PetscImaginaryPart(value);
67: if (type == NORM_INFINITY) reductions[col] = PetscMax(reduction, reductions[col]);
68: else reductions[col] += reduction;
69: if (offdiag) {
70: if (hermitian && (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART)) reduction = -reduction;
71: if (type == NORM_INFINITY) reductions[row] = PetscMax(reduction, reductions[row]);
72: else reductions[row] += reduction;
73: }
74: }
75: }
76: }
77: }
78: if (type == NORM_2) {
79: for (PetscInt i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
80: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
81: for (PetscInt i = 0; i < n; i++) reductions[i] /= m;
82: }
83: PetscFunctionReturn(PETSC_SUCCESS);
84: }
86: static PetscErrorCode MatGetRowIJ_SeqSBAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool blockcompressed, PetscInt *nn, const PetscInt *inia[], const PetscInt *inja[], PetscBool *done)
87: {
88: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
89: PetscInt i, j, n = a->mbs, nz = a->i[n], *tia, *tja, bs = A->rmap->bs, k, l, cnt;
90: PetscInt **ia = (PetscInt **)inia, **ja = (PetscInt **)inja;
92: PetscFunctionBegin;
93: *nn = n;
94: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
95: if (symmetric) {
96: PetscCall(MatToSymmetricIJ_SeqAIJ(n, a->i, a->j, PETSC_FALSE, 0, 0, &tia, &tja));
97: nz = tia[n];
98: } else {
99: tia = a->i;
100: tja = a->j;
101: }
103: if (!blockcompressed && bs > 1) {
104: (*nn) *= bs;
105: /* malloc & create the natural set of indices */
106: PetscCall(PetscMalloc1((n + 1) * bs, ia));
107: if (n) {
108: (*ia)[0] = oshift;
109: for (j = 1; j < bs; j++) (*ia)[j] = (tia[1] - tia[0]) * bs + (*ia)[j - 1];
110: }
112: for (i = 1; i < n; i++) {
113: (*ia)[i * bs] = (tia[i] - tia[i - 1]) * bs + (*ia)[i * bs - 1];
114: for (j = 1; j < bs; j++) (*ia)[i * bs + j] = (tia[i + 1] - tia[i]) * bs + (*ia)[i * bs + j - 1];
115: }
116: if (n) (*ia)[n * bs] = (tia[n] - tia[n - 1]) * bs + (*ia)[n * bs - 1];
118: if (inja) {
119: PetscCall(PetscMalloc1(nz * bs * bs, ja));
120: cnt = 0;
121: for (i = 0; i < n; i++) {
122: for (j = 0; j < bs; j++) {
123: for (k = tia[i]; k < tia[i + 1]; k++) {
124: for (l = 0; l < bs; l++) (*ja)[cnt++] = bs * tja[k] + l;
125: }
126: }
127: }
128: }
130: if (symmetric) { /* deallocate memory allocated in MatToSymmetricIJ_SeqAIJ() */
131: PetscCall(PetscFree(tia));
132: PetscCall(PetscFree(tja));
133: }
134: } else if (oshift == 1) {
135: if (symmetric) {
136: nz = tia[A->rmap->n / bs];
137: /* add 1 to i and j indices */
138: for (i = 0; i < A->rmap->n / bs + 1; i++) tia[i] = tia[i] + 1;
139: *ia = tia;
140: if (ja) {
141: for (i = 0; i < nz; i++) tja[i] = tja[i] + 1;
142: *ja = tja;
143: }
144: } else {
145: nz = a->i[A->rmap->n / bs];
146: /* malloc space and add 1 to i and j indices */
147: PetscCall(PetscMalloc1(A->rmap->n / bs + 1, ia));
148: for (i = 0; i < A->rmap->n / bs + 1; i++) (*ia)[i] = a->i[i] + 1;
149: if (ja) {
150: PetscCall(PetscMalloc1(nz, ja));
151: for (i = 0; i < nz; i++) (*ja)[i] = a->j[i] + 1;
152: }
153: }
154: } else {
155: *ia = tia;
156: if (ja) *ja = tja;
157: }
158: PetscFunctionReturn(PETSC_SUCCESS);
159: }
161: static PetscErrorCode MatRestoreRowIJ_SeqSBAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool blockcompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
162: {
163: PetscFunctionBegin;
164: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
165: if ((!blockcompressed && A->rmap->bs > 1) || (symmetric || oshift == 1)) {
166: PetscCall(PetscFree(*ia));
167: if (ja) PetscCall(PetscFree(*ja));
168: }
169: PetscFunctionReturn(PETSC_SUCCESS);
170: }
172: PetscErrorCode MatDestroy_SeqSBAIJ(Mat A)
173: {
174: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
176: PetscFunctionBegin;
177: if (A->hash_active) {
178: PetscInt bs;
179: A->ops[0] = a->cops;
180: PetscCall(PetscHMapIJVDestroy(&a->ht));
181: PetscCall(MatGetBlockSize(A, &bs));
182: if (bs > 1) PetscCall(PetscHSetIJDestroy(&a->bht));
183: PetscCall(PetscFree(a->dnz));
184: PetscCall(PetscFree(a->bdnz));
185: A->hash_active = PETSC_FALSE;
186: }
187: PetscCall(PetscLogObjectState((PetscObject)A, "Rows=%" PetscInt_FMT ", NZ=%" PetscInt_FMT, A->rmap->N, a->nz));
188: PetscCall(MatSeqXAIJFreeAIJ(A, &a->a, &a->j, &a->i));
189: PetscCall(PetscFree(a->diag));
190: PetscCall(ISDestroy(&a->row));
191: PetscCall(ISDestroy(&a->col));
192: PetscCall(ISDestroy(&a->icol));
193: PetscCall(PetscFree(a->idiag));
194: PetscCall(PetscFree(a->inode.size_csr));
195: if (a->free_imax_ilen) PetscCall(PetscFree2(a->imax, a->ilen));
196: PetscCall(PetscFree(a->solve_work));
197: PetscCall(PetscFree(a->sor_work));
198: PetscCall(PetscFree(a->solves_work));
199: PetscCall(PetscFree(a->mult_work));
200: PetscCall(PetscFree(a->saved_values));
201: if (a->free_jshort) PetscCall(PetscFree(a->jshort));
202: PetscCall(PetscFree(a->inew));
203: PetscCall(MatDestroy(&a->parent));
204: PetscCall(PetscFree(A->data));
206: PetscCall(PetscObjectChangeTypeName((PetscObject)A, NULL));
207: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqSBAIJGetArray_C", NULL));
208: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqSBAIJRestoreArray_C", NULL));
209: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatStoreValues_C", NULL));
210: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatRetrieveValues_C", NULL));
211: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqSBAIJSetColumnIndices_C", NULL));
212: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqsbaij_seqaij_C", NULL));
213: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqsbaij_seqbaij_C", NULL));
214: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqSBAIJSetPreallocation_C", NULL));
215: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqSBAIJSetPreallocationCSR_C", NULL));
216: #if PetscDefined(HAVE_ELEMENTAL)
217: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqsbaij_elemental_C", NULL));
218: #endif
219: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
220: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqsbaij_scalapack_C", NULL));
221: #endif
222: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
223: PetscFunctionReturn(PETSC_SUCCESS);
224: }
226: static PetscErrorCode MatSetOption_SeqSBAIJ(Mat A, MatOption op, PetscBool flg)
227: {
228: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
230: PetscFunctionBegin;
231: switch (op) {
232: case MAT_ROW_ORIENTED:
233: a->roworiented = flg;
234: break;
235: case MAT_KEEP_NONZERO_PATTERN:
236: a->keepnonzeropattern = flg;
237: break;
238: case MAT_NEW_NONZERO_LOCATIONS:
239: a->nonew = (flg ? 0 : 1);
240: break;
241: case MAT_NEW_NONZERO_LOCATION_ERR:
242: a->nonew = (flg ? -1 : 0);
243: break;
244: case MAT_NEW_NONZERO_ALLOCATION_ERR:
245: a->nonew = (flg ? -2 : 0);
246: break;
247: case MAT_UNUSED_NONZERO_LOCATION_ERR:
248: a->nounused = (flg ? -1 : 0);
249: break;
250: case MAT_HERMITIAN:
251: if (PetscDefined(USE_COMPLEX) && flg) { /* disable transpose ops */
252: PetscInt bs;
254: PetscCall(MatGetBlockSize(A, &bs));
255: PetscCheck(bs <= 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for Hermitian with block size greater than 1");
256: A->ops->multtranspose = NULL;
257: A->ops->multtransposeadd = NULL;
258: }
259: break;
260: case MAT_SYMMETRIC:
261: case MAT_SPD:
262: if (PetscDefined(USE_COMPLEX) && flg) { /* An Hermitian and symmetric matrix has zero imaginary part (restore back transpose ops) */
263: A->ops->multtranspose = A->ops->mult;
264: A->ops->multtransposeadd = A->ops->multadd;
265: }
266: break;
267: case MAT_IGNORE_LOWER_TRIANGULAR:
268: case MAT_ERROR_LOWER_TRIANGULAR:
269: a->ignore_ltriangular = flg;
270: break;
271: case MAT_GETROW_UPPERTRIANGULAR:
272: a->getrow_utriangular = flg;
273: break;
274: default:
275: break;
276: }
277: PetscFunctionReturn(PETSC_SUCCESS);
278: }
280: PetscErrorCode MatGetRow_SeqSBAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
281: {
282: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
284: PetscFunctionBegin;
285: PetscCheck(!A || a->getrow_utriangular, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatGetRow is not supported for SBAIJ matrix format. Getting the upper triangular part of row, run with -mat_getrow_uppertriangular, call MatSetOption(mat,MAT_GETROW_UPPERTRIANGULAR,PETSC_TRUE) or MatGetRowUpperTriangular()");
287: /* Get the upper triangular part of the row */
288: PetscCall(MatGetRow_SeqBAIJ_private(A, row, nz, idx, v, a->i, a->j, a->a));
289: PetscFunctionReturn(PETSC_SUCCESS);
290: }
292: PetscErrorCode MatRestoreRow_SeqSBAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
293: {
294: PetscFunctionBegin;
295: if (idx) PetscCall(PetscFree(*idx));
296: if (v) PetscCall(PetscFree(*v));
297: PetscFunctionReturn(PETSC_SUCCESS);
298: }
300: static PetscErrorCode MatGetRowUpperTriangular_SeqSBAIJ(Mat A)
301: {
302: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
304: PetscFunctionBegin;
305: a->getrow_utriangular = PETSC_TRUE;
306: PetscFunctionReturn(PETSC_SUCCESS);
307: }
309: static PetscErrorCode MatRestoreRowUpperTriangular_SeqSBAIJ(Mat A)
310: {
311: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
313: PetscFunctionBegin;
314: a->getrow_utriangular = PETSC_FALSE;
315: PetscFunctionReturn(PETSC_SUCCESS);
316: }
318: static PetscErrorCode MatTranspose_SeqSBAIJ(Mat A, MatReuse reuse, Mat *B)
319: {
320: PetscFunctionBegin;
321: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *B));
322: if (reuse == MAT_INITIAL_MATRIX) {
323: PetscCall(MatDuplicate(A, MAT_COPY_VALUES, B));
324: } else if (reuse == MAT_REUSE_MATRIX) {
325: PetscCall(MatCopy(A, *B, SAME_NONZERO_PATTERN));
326: }
327: PetscFunctionReturn(PETSC_SUCCESS);
328: }
330: static PetscErrorCode MatView_SeqSBAIJ_ASCII(Mat A, PetscViewer viewer)
331: {
332: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
333: PetscInt i, j, bs = A->rmap->bs, k, l, bs2 = a->bs2;
334: PetscViewerFormat format;
335: const PetscInt *diag;
336: const char *matname;
338: PetscFunctionBegin;
339: PetscCall(PetscViewerGetFormat(viewer, &format));
340: if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
341: } else if (format == PETSC_VIEWER_ASCII_MATLAB) {
342: Mat aij;
344: if (A->factortype && bs > 1) {
345: PetscCall(PetscPrintf(PETSC_COMM_SELF, "Warning: matrix is factored with bs>1. MatView() with PETSC_VIEWER_ASCII_MATLAB is not supported and ignored!\n"));
346: PetscFunctionReturn(PETSC_SUCCESS);
347: }
348: PetscCall(MatConvert(A, MATSEQAIJ, MAT_INITIAL_MATRIX, &aij));
349: if (((PetscObject)A)->name) PetscCall(PetscObjectGetName((PetscObject)A, &matname));
350: if (((PetscObject)A)->name) PetscCall(PetscObjectSetName((PetscObject)aij, matname));
351: PetscCall(MatView_SeqAIJ(aij, viewer));
352: PetscCall(MatDestroy(&aij));
353: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
354: Mat B;
356: PetscCall(MatConvert(A, MATSEQAIJ, MAT_INITIAL_MATRIX, &B));
357: if (((PetscObject)A)->name) PetscCall(PetscObjectGetName((PetscObject)A, &matname));
358: if (((PetscObject)A)->name) PetscCall(PetscObjectSetName((PetscObject)B, matname));
359: PetscCall(MatView_SeqAIJ(B, viewer));
360: PetscCall(MatDestroy(&B));
361: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
362: PetscFunctionReturn(PETSC_SUCCESS);
363: } else {
364: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
365: if (A->factortype) { /* for factored matrix */
366: PetscCheck(bs <= 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "matrix is factored with bs>1. Not implemented yet");
367: PetscCall(MatGetDiagonalMarkers_SeqSBAIJ(A, &diag, NULL));
368: for (i = 0; i < a->mbs; i++) { /* for row block i */
369: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
370: /* diagonal entry */
371: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[diag[i]]) > 0.0) {
372: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i) ", a->j[diag[i]], (double)PetscRealPart(1.0 / a->a[diag[i]]), (double)PetscImaginaryPart(1.0 / a->a[diag[i]])));
373: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[diag[i]]) < 0.0) {
374: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i) ", a->j[diag[i]], (double)PetscRealPart(1.0 / a->a[diag[i]]), -(double)PetscImaginaryPart(1.0 / a->a[diag[i]])));
375: } else {
376: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[diag[i]], (double)PetscRealPart(1.0 / a->a[diag[i]])));
377: }
378: /* off-diagonal entries */
379: for (k = a->i[i]; k < a->i[i + 1] - 1; k++) {
380: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[k]) > 0.0) {
381: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i) ", bs * a->j[k], (double)PetscRealPart(a->a[k]), (double)PetscImaginaryPart(a->a[k])));
382: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[k]) < 0.0) {
383: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i) ", bs * a->j[k], (double)PetscRealPart(a->a[k]), -(double)PetscImaginaryPart(a->a[k])));
384: } else {
385: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", bs * a->j[k], (double)PetscRealPart(a->a[k])));
386: }
387: }
388: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
389: }
391: } else { /* for non-factored matrix */
392: for (i = 0; i < a->mbs; i++) { /* for row block i */
393: for (j = 0; j < bs; j++) { /* for row bs*i + j */
394: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i * bs + j));
395: for (k = a->i[i]; k < a->i[i + 1]; k++) { /* for column block */
396: for (l = 0; l < bs; l++) { /* for column */
397: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[bs2 * k + l * bs + j]) > 0.0) {
398: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i) ", bs * a->j[k] + l, (double)PetscRealPart(a->a[bs2 * k + l * bs + j]), (double)PetscImaginaryPart(a->a[bs2 * k + l * bs + j])));
399: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[bs2 * k + l * bs + j]) < 0.0) {
400: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i) ", bs * a->j[k] + l, (double)PetscRealPart(a->a[bs2 * k + l * bs + j]), -(double)PetscImaginaryPart(a->a[bs2 * k + l * bs + j])));
401: } else {
402: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", bs * a->j[k] + l, (double)PetscRealPart(a->a[bs2 * k + l * bs + j])));
403: }
404: }
405: }
406: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
407: }
408: }
409: }
410: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
411: }
412: PetscCall(PetscViewerFlush(viewer));
413: PetscFunctionReturn(PETSC_SUCCESS);
414: }
416: #include <petscdraw.h>
417: static PetscErrorCode MatView_SeqSBAIJ_Draw_Zoom(PetscDraw draw, void *Aa)
418: {
419: Mat A = (Mat)Aa;
420: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
421: PetscInt row, i, j, k, l, mbs = a->mbs, bs = A->rmap->bs, bs2 = a->bs2;
422: PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r;
423: MatScalar *aa;
424: PetscViewer viewer;
425: int color;
427: PetscFunctionBegin;
428: PetscCall(PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer));
429: PetscCall(PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr));
431: /* loop over matrix elements drawing boxes */
433: PetscDrawCollectiveBegin(draw);
434: PetscCall(PetscDrawString(draw, .3 * (xl + xr), .3 * (yl + yr), PETSC_DRAW_BLACK, "symmetric"));
435: /* Blue for negative, Cyan for zero and Red for positive */
436: color = PETSC_DRAW_BLUE;
437: for (i = 0, row = 0; i < mbs; i++, row += bs) {
438: for (j = a->i[i]; j < a->i[i + 1]; j++) {
439: y_l = A->rmap->N - row - 1.0;
440: y_r = y_l + 1.0;
441: x_l = a->j[j] * bs;
442: x_r = x_l + 1.0;
443: aa = a->a + j * bs2;
444: for (k = 0; k < bs; k++) {
445: for (l = 0; l < bs; l++) {
446: if (PetscRealPart(*aa++) >= 0.) continue;
447: PetscCall(PetscDrawRectangle(draw, x_l + k, y_l - l, x_r + k, y_r - l, color, color, color, color));
448: }
449: }
450: }
451: }
452: color = PETSC_DRAW_CYAN;
453: for (i = 0, row = 0; i < mbs; i++, row += bs) {
454: for (j = a->i[i]; j < a->i[i + 1]; j++) {
455: y_l = A->rmap->N - row - 1.0;
456: y_r = y_l + 1.0;
457: x_l = a->j[j] * bs;
458: x_r = x_l + 1.0;
459: aa = a->a + j * bs2;
460: for (k = 0; k < bs; k++) {
461: for (l = 0; l < bs; l++) {
462: if (PetscRealPart(*aa++) != 0.) continue;
463: PetscCall(PetscDrawRectangle(draw, x_l + k, y_l - l, x_r + k, y_r - l, color, color, color, color));
464: }
465: }
466: }
467: }
468: color = PETSC_DRAW_RED;
469: for (i = 0, row = 0; i < mbs; i++, row += bs) {
470: for (j = a->i[i]; j < a->i[i + 1]; j++) {
471: y_l = A->rmap->N - row - 1.0;
472: y_r = y_l + 1.0;
473: x_l = a->j[j] * bs;
474: x_r = x_l + 1.0;
475: aa = a->a + j * bs2;
476: for (k = 0; k < bs; k++) {
477: for (l = 0; l < bs; l++) {
478: if (PetscRealPart(*aa++) <= 0.) continue;
479: PetscCall(PetscDrawRectangle(draw, x_l + k, y_l - l, x_r + k, y_r - l, color, color, color, color));
480: }
481: }
482: }
483: }
484: PetscDrawCollectiveEnd(draw);
485: PetscFunctionReturn(PETSC_SUCCESS);
486: }
488: static PetscErrorCode MatView_SeqSBAIJ_Draw(Mat A, PetscViewer viewer)
489: {
490: PetscReal xl, yl, xr, yr, w, h;
491: PetscDraw draw;
492: PetscBool isnull;
494: PetscFunctionBegin;
495: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
496: PetscCall(PetscDrawIsNull(draw, &isnull));
497: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
499: xr = A->rmap->N;
500: yr = A->rmap->N;
501: h = yr / 10.0;
502: w = xr / 10.0;
503: xr += w;
504: yr += h;
505: xl = -w;
506: yl = -h;
507: PetscCall(PetscDrawSetCoordinates(draw, xl, yl, xr, yr));
508: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer));
509: PetscCall(PetscDrawZoom(draw, MatView_SeqSBAIJ_Draw_Zoom, A));
510: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL));
511: PetscCall(PetscDrawSave(draw));
512: PetscFunctionReturn(PETSC_SUCCESS);
513: }
515: /* Used for both MPIBAIJ and MPISBAIJ matrices */
516: #define MatView_SeqSBAIJ_Binary MatView_SeqBAIJ_Binary
518: PetscErrorCode MatView_SeqSBAIJ(Mat A, PetscViewer viewer)
519: {
520: PetscBool isascii, isbinary, isdraw;
522: PetscFunctionBegin;
523: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
524: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
525: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
526: if (isascii) {
527: PetscCall(MatView_SeqSBAIJ_ASCII(A, viewer));
528: } else if (isbinary) {
529: PetscCall(MatView_SeqSBAIJ_Binary(A, viewer));
530: } else if (isdraw) {
531: PetscCall(MatView_SeqSBAIJ_Draw(A, viewer));
532: } else {
533: Mat B;
534: const char *matname;
535: PetscCall(MatConvert(A, MATSEQAIJ, MAT_INITIAL_MATRIX, &B));
536: if (((PetscObject)A)->name) PetscCall(PetscObjectGetName((PetscObject)A, &matname));
537: if (((PetscObject)A)->name) PetscCall(PetscObjectSetName((PetscObject)B, matname));
538: PetscCall(MatView(B, viewer));
539: PetscCall(MatDestroy(&B));
540: }
541: PetscFunctionReturn(PETSC_SUCCESS);
542: }
544: PetscErrorCode MatGetValues_SeqSBAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], PetscScalar v[])
545: {
546: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
547: PetscInt *rp, k, low, high, t, row, nrow, i, col, l, *aj = a->j;
548: PetscInt *ai = a->i, *ailen = a->ilen;
549: PetscInt brow, bcol, ridx, cidx, bs = A->rmap->bs, bs2 = a->bs2;
550: MatScalar *ap, *aa = a->a;
551: PetscBool roworiented = a->roworiented;
552: PetscScalar *value;
554: PetscFunctionBegin;
555: for (k = 0; k < m; k++) { /* loop over rows */
556: row = im[k];
557: if (row < 0) continue; /* negative row */
558: 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);
559: brow = row / bs;
560: rp = aj + ai[brow];
561: ap = aa + bs2 * ai[brow];
562: nrow = ailen[brow];
563: for (l = 0; l < n; l++) { /* loop over columns */
564: if (in[l] < 0) continue; /* negative column */
565: 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);
566: value = roworiented ? &v[l + k * n] : &v[k + l * m];
567: col = in[l];
568: bcol = col / bs;
569: cidx = col % bs;
570: ridx = row % bs;
571: high = nrow;
572: low = 0; /* assume unsorted */
573: while (high - low > 5) {
574: t = (low + high) / 2;
575: if (rp[t] > bcol) high = t;
576: else low = t;
577: }
578: for (i = low; i < high; i++) {
579: if (rp[i] > bcol) break;
580: if (rp[i] == bcol) {
581: *value = ap[bs2 * i + bs * cidx + ridx];
582: goto finished;
583: }
584: }
585: *value = 0.0;
586: finished:;
587: }
588: }
589: PetscFunctionReturn(PETSC_SUCCESS);
590: }
592: static PetscErrorCode MatPermute_SeqSBAIJ(Mat A, IS rowp, IS colp, Mat *B)
593: {
594: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data, *b = NULL;
595: Mat_SeqBAIJ *c = NULL;
596: Mat C;
597: IS browp, bcolp;
598: const PetscInt *row, *col;
599: PetscInt *irow, *icol, *lens, *bi, *bj, *bilen;
600: MatScalar *ba, *work;
601: PetscInt mbs, nbs, bs = A->rmap->bs, bs2 = a->bs2;
602: PetscBool same = (PetscBool)(rowp == colp), implicit = (PetscBool)(A->rmap->N == A->cmap->N && (A->symmetric == PETSC_BOOL3_TRUE || A->hermitian == PETSC_BOOL3_TRUE)), hermitian;
604: PetscFunctionBegin;
605: if (same == PETSC_FALSE) PetscCall(ISEqualUnsorted(rowp, colp, &same));
606: hermitian = (PetscBool)(implicit == PETSC_TRUE && PetscDefined(USE_COMPLEX) && A->hermitian == PETSC_BOOL3_TRUE);
607: same = (PetscBool)(implicit == PETSC_TRUE && same == PETSC_TRUE);
608: PetscCall(ISCompressIndicesGeneral(A->rmap->N, A->rmap->n, bs, 1, &rowp, &browp));
609: if (rowp == colp) {
610: bcolp = browp;
611: PetscCall(PetscObjectReference((PetscObject)bcolp));
612: } else PetscCall(ISCompressIndicesGeneral(A->cmap->N, A->cmap->n, bs, 1, &colp, &bcolp));
613: PetscCall(ISGetLocalSize(browp, &mbs));
614: PetscCall(ISGetLocalSize(bcolp, &nbs));
615: PetscCheck(mbs == a->mbs && nbs == a->nbs, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Row and column index sets must be block permutations");
616: PetscCall(ISGetIndices(browp, &row));
617: PetscCall(ISGetIndices(bcolp, &col));
618: PetscCall(PetscMalloc2(mbs, &irow, nbs, &icol));
619: for (PetscInt i = 0; i < mbs; i++) {
620: PetscCheck(row[i] >= 0 && row[i] < mbs, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Row index set is not a block permutation");
621: irow[row[i]] = i;
622: }
623: for (PetscInt i = 0; i < nbs; i++) {
624: PetscCheck(col[i] >= 0 && col[i] < nbs, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Column index set is not a block permutation");
625: icol[col[i]] = i;
626: }
627: PetscCall(PetscCalloc1(mbs, &lens));
628: for (PetscInt i = 0; i < a->mbs; i++) {
629: for (PetscInt k = a->i[i]; k < a->i[i + 1]; k++) {
630: const PetscInt j = a->j[k];
632: if (same == PETSC_TRUE) lens[PetscMin(irow[i], icol[j])]++;
633: else {
634: lens[irow[i]]++;
635: if (implicit == PETSC_TRUE && i != j) lens[irow[j]]++;
636: }
637: }
638: }
639: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
640: PetscCall(MatSetSizes(C, mbs * bs, nbs * bs, mbs * bs, nbs * bs));
641: PetscCall(MatSetType(C, same == PETSC_TRUE ? ((PetscObject)A)->type_name : MATSEQBAIJ));
642: if (same == PETSC_TRUE) {
643: PetscCall(MatSeqSBAIJSetPreallocation(C, bs, 0, lens));
644: b = (Mat_SeqSBAIJ *)C->data;
645: bi = b->i;
646: bj = b->j;
647: ba = b->a;
648: bilen = b->ilen;
649: } else {
650: PetscCall(MatSeqBAIJSetPreallocation(C, bs, 0, lens));
651: c = (Mat_SeqBAIJ *)C->data;
652: bi = c->i;
653: bj = c->j;
654: ba = c->a;
655: bilen = c->ilen;
656: }
657: PetscCall(PetscFree(lens));
658: for (PetscInt i = 0; i < a->mbs; i++) {
659: for (PetscInt k = a->i[i]; k < a->i[i + 1]; k++) {
660: const PetscInt j = a->j[k];
661: PetscInt r = irow[i], colidx = icol[j], pos;
662: MatScalar *block;
664: if (same == PETSC_TRUE && r > colidx) {
665: pos = r;
666: r = colidx;
667: colidx = pos;
668: }
669: pos = bi[r] + bilen[r]++;
670: bj[pos] = colidx;
671: block = ba + pos * bs2;
672: PetscCall(PetscArraycpy(block, a->a + k * bs2, bs2));
673: if (same == PETSC_TRUE && irow[i] > icol[j]) {
674: PetscCall(PetscKernel_A_gets_transpose_A_N(block, bs));
675: if (hermitian == PETSC_TRUE) {
676: for (PetscInt q = 0; q < bs2; q++) block[q] = PetscConj(block[q]);
677: }
678: }
679: if (same == PETSC_FALSE && implicit == PETSC_TRUE && i != j) {
680: r = irow[j];
681: colidx = icol[i];
682: pos = bi[r] + bilen[r]++;
683: bj[pos] = colidx;
684: block = ba + pos * bs2;
685: PetscCall(PetscArraycpy(block, a->a + k * bs2, bs2));
686: PetscCall(PetscKernel_A_gets_transpose_A_N(block, bs));
687: if (hermitian == PETSC_TRUE) {
688: for (PetscInt q = 0; q < bs2; q++) block[q] = PetscConj(block[q]);
689: }
690: }
691: }
692: }
693: PetscCall(PetscMalloc1(bs2, &work));
694: for (PetscInt i = 0; i < mbs; i++) PetscCall(PetscSortIntWithDataArray(bilen[i], PetscSafePointerPlusOffset(bj, bi[i]), PetscSafePointerPlusOffset(ba, bi[i] * bs2), bs2 * sizeof(MatScalar), work));
695: PetscCall(PetscFree(work));
696: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
697: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
698: if (same == PETSC_TRUE) PetscCall(MatPropagateSymmetryOptions(A, C));
699: PetscCall(PetscFree2(irow, icol));
700: PetscCall(ISRestoreIndices(browp, &row));
701: PetscCall(ISRestoreIndices(bcolp, &col));
702: PetscCall(ISDestroy(&browp));
703: PetscCall(ISDestroy(&bcolp));
704: *B = C;
705: PetscFunctionReturn(PETSC_SUCCESS);
706: }
708: PetscErrorCode MatSetValuesBlocked_SeqSBAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
709: {
710: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
711: PetscInt *rp, k, low, high, t, ii, jj, row, nrow, i, col, l, rmax, N, lastcol = -1;
712: PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen;
713: PetscInt *aj = a->j, nonew = a->nonew, bs2 = a->bs2, bs = A->rmap->bs, stepval;
714: PetscBool roworiented = a->roworiented;
715: const PetscScalar *value = v;
716: MatScalar *ap, *aa = a->a, *bap;
718: PetscFunctionBegin;
719: if (roworiented) stepval = (n - 1) * bs;
720: else stepval = (m - 1) * bs;
721: for (k = 0; k < m; k++) { /* loop over added rows */
722: row = im[k];
723: if (row < 0) continue;
724: PetscCheck(row < a->mbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Block index row too large %" PetscInt_FMT " max %" PetscInt_FMT, row, a->mbs - 1);
725: rp = aj + ai[row];
726: ap = aa + bs2 * ai[row];
727: rmax = imax[row];
728: nrow = ailen[row];
729: low = 0;
730: high = nrow;
731: for (l = 0; l < n; l++) { /* loop over added columns */
732: if (in[l] < 0) continue;
733: col = in[l];
734: PetscCheck(col < a->nbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Block index column too large %" PetscInt_FMT " max %" PetscInt_FMT, col, a->nbs - 1);
735: if (col < row) {
736: PetscCheck(a->ignore_ltriangular, PETSC_COMM_SELF, PETSC_ERR_USER, "Lower triangular value cannot be set for sbaij format. Ignoring these values, run with -mat_ignore_lower_triangular or call MatSetOption(mat,MAT_IGNORE_LOWER_TRIANGULAR,PETSC_TRUE)");
737: continue; /* ignore lower triangular block */
738: }
739: if (roworiented) value = v + k * (stepval + bs) * bs + l * bs;
740: else value = v + l * (stepval + bs) * bs + k * bs;
742: if (col <= lastcol) low = 0;
743: else high = nrow;
745: lastcol = col;
746: while (high - low > 7) {
747: t = (low + high) / 2;
748: if (rp[t] > col) high = t;
749: else low = t;
750: }
751: for (i = low; i < high; i++) {
752: if (rp[i] > col) break;
753: if (rp[i] == col) {
754: bap = ap + bs2 * i;
755: if (roworiented) {
756: if (is == ADD_VALUES) {
757: for (ii = 0; ii < bs; ii++, value += stepval) {
758: for (jj = ii; jj < bs2; jj += bs) bap[jj] += *value++;
759: }
760: } else {
761: for (ii = 0; ii < bs; ii++, value += stepval) {
762: for (jj = ii; jj < bs2; jj += bs) bap[jj] = *value++;
763: }
764: }
765: } else {
766: if (is == ADD_VALUES) {
767: for (ii = 0; ii < bs; ii++, value += stepval) {
768: for (jj = 0; jj < bs; jj++) *bap++ += *value++;
769: }
770: } else {
771: for (ii = 0; ii < bs; ii++, value += stepval) {
772: for (jj = 0; jj < bs; jj++) *bap++ = *value++;
773: }
774: }
775: }
776: goto noinsert2;
777: }
778: }
779: if (nonew == 1) goto noinsert2;
780: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new block index nonzero block (%" PetscInt_FMT ", %" PetscInt_FMT ") in the matrix", row, col);
781: MatSeqXAIJReallocateAIJ(A, a->mbs, bs2, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
782: N = nrow++ - 1;
783: high++;
784: /* shift up all the later entries in this row */
785: PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1));
786: PetscCall(PetscArraymove(ap + bs2 * (i + 1), ap + bs2 * i, bs2 * (N - i + 1)));
787: PetscCall(PetscArrayzero(ap + bs2 * i, bs2));
788: rp[i] = col;
789: bap = ap + bs2 * i;
790: if (roworiented) {
791: for (ii = 0; ii < bs; ii++, value += stepval) {
792: for (jj = ii; jj < bs2; jj += bs) bap[jj] = *value++;
793: }
794: } else {
795: for (ii = 0; ii < bs; ii++, value += stepval) {
796: for (jj = 0; jj < bs; jj++) *bap++ = *value++;
797: }
798: }
799: noinsert2:;
800: low = i;
801: }
802: ailen[row] = nrow;
803: }
804: PetscFunctionReturn(PETSC_SUCCESS);
805: }
807: static PetscErrorCode MatAssemblyEnd_SeqSBAIJ(Mat A, MatAssemblyType mode)
808: {
809: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
810: PetscInt fshift = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax;
811: PetscInt m = A->rmap->N, *ip, N, *ailen = a->ilen;
812: PetscInt mbs = a->mbs, bs2 = a->bs2, rmax = 0;
813: MatScalar *aa = a->a, *ap;
815: PetscFunctionBegin;
816: if (mode == MAT_FLUSH_ASSEMBLY || (A->was_assembled && A->ass_nonzerostate == A->nonzerostate)) PetscFunctionReturn(PETSC_SUCCESS);
818: if (m) rmax = ailen[0];
819: for (i = 1; i < mbs; i++) {
820: /* move each row back by the amount of empty slots (fshift) before it*/
821: fshift += imax[i - 1] - ailen[i - 1];
822: rmax = PetscMax(rmax, ailen[i]);
823: if (fshift) {
824: ip = aj + ai[i];
825: ap = aa + bs2 * ai[i];
826: N = ailen[i];
827: PetscCall(PetscArraymove(ip - fshift, ip, N));
828: PetscCall(PetscArraymove(ap - bs2 * fshift, ap, bs2 * N));
829: }
830: ai[i] = ai[i - 1] + ailen[i - 1];
831: }
832: if (mbs) {
833: fshift += imax[mbs - 1] - ailen[mbs - 1];
834: ai[mbs] = ai[mbs - 1] + ailen[mbs - 1];
835: }
836: /* reset ilen and imax for each row */
837: for (i = 0; i < mbs; i++) ailen[i] = imax[i] = ai[i + 1] - ai[i];
838: a->nz = ai[mbs];
840: PetscCheck(!fshift || a->nounused != -1, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unused space detected in matrix: %" PetscInt_FMT " X %" PetscInt_FMT " block size %" PetscInt_FMT ", %" PetscInt_FMT " unneeded", m, A->cmap->n, A->rmap->bs, fshift * bs2);
842: PetscCall(PetscInfo(A, "Matrix size: %" PetscInt_FMT " X %" PetscInt_FMT ", block size %" PetscInt_FMT "; storage space: %" PetscInt_FMT " unneeded, %" PetscInt_FMT " used\n", m, A->rmap->N, A->rmap->bs, fshift * bs2, a->nz * bs2));
843: PetscCall(PetscInfo(A, "Number of mallocs during MatSetValues is %" PetscInt_FMT "\n", a->reallocs));
844: PetscCall(PetscInfo(A, "Most nonzeros blocks in any row is %" PetscInt_FMT "\n", rmax));
846: A->info.mallocs += a->reallocs;
847: a->reallocs = 0;
848: A->info.nz_unneeded = (PetscReal)fshift * bs2;
849: a->rmax = rmax;
851: if (A->cmap->n < 65536 && A->cmap->bs == 1) {
852: if (a->jshort && a->free_jshort) {
853: /* when matrix data structure is changed, previous jshort must be replaced */
854: PetscCall(PetscFree(a->jshort));
855: }
856: PetscCall(PetscMalloc1(a->i[A->rmap->n], &a->jshort));
857: for (i = 0; i < a->i[A->rmap->n]; i++) a->jshort[i] = (short)a->j[i];
858: A->ops->mult = MatMult_SeqSBAIJ_1_ushort;
859: A->ops->sor = MatSOR_SeqSBAIJ_ushort;
860: a->free_jshort = PETSC_TRUE;
861: }
862: PetscFunctionReturn(PETSC_SUCCESS);
863: }
865: /* Only add/insert a(i,j) with i<=j (blocks).
866: Any a(i,j) with i>j input by user is ignored.
867: */
869: PetscErrorCode MatSetValues_SeqSBAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
870: {
871: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
872: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N, lastcol = -1;
873: PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen, roworiented = a->roworiented;
874: PetscInt *aj = a->j, nonew = a->nonew, bs = A->rmap->bs, brow, bcol;
875: PetscInt ridx, cidx, bs2 = a->bs2;
876: MatScalar *ap, value, *aa = a->a, *bap;
878: PetscFunctionBegin;
879: for (k = 0; k < m; k++) { /* loop over added rows */
880: row = im[k]; /* row number */
881: brow = row / bs; /* block row number */
882: if (row < 0) continue;
883: 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);
884: rp = aj + ai[brow]; /*ptr to beginning of column value of the row block*/
885: ap = aa + bs2 * ai[brow]; /*ptr to beginning of element value of the row block*/
886: rmax = imax[brow]; /* maximum space allocated for this row */
887: nrow = ailen[brow]; /* actual length of this row */
888: low = 0;
889: high = nrow;
890: for (l = 0; l < n; l++) { /* loop over added columns */
891: if (in[l] < 0) continue;
892: 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);
893: col = in[l];
894: bcol = col / bs; /* block col number */
896: if (brow > bcol) {
897: PetscCheck(a->ignore_ltriangular, PETSC_COMM_SELF, PETSC_ERR_USER, "Lower triangular value cannot be set for sbaij format. Ignoring these values, run with -mat_ignore_lower_triangular or call MatSetOption(mat,MAT_IGNORE_LOWER_TRIANGULAR,PETSC_TRUE)");
898: continue; /* ignore lower triangular values */
899: }
901: ridx = row % bs;
902: cidx = col % bs; /*row and col index inside the block */
903: if ((brow == bcol && ridx <= cidx) || (brow < bcol)) {
904: /* element value a(k,l) */
905: if (roworiented) value = v[l + k * n];
906: else value = v[k + l * m];
908: /* move pointer bap to a(k,l) quickly and add/insert value */
909: if (col <= lastcol) low = 0;
910: else high = nrow;
912: lastcol = col;
913: while (high - low > 7) {
914: t = (low + high) / 2;
915: if (rp[t] > bcol) high = t;
916: else low = t;
917: }
918: for (i = low; i < high; i++) {
919: if (rp[i] > bcol) break;
920: if (rp[i] == bcol) {
921: bap = ap + bs2 * i + bs * cidx + ridx;
922: if (is == ADD_VALUES) *bap += value;
923: else *bap = value;
924: /* for diag block, add/insert its symmetric element a(cidx,ridx) */
925: if (brow == bcol && ridx < cidx) {
926: bap = ap + bs2 * i + bs * ridx + cidx;
927: if (is == ADD_VALUES) *bap += value;
928: else *bap = value;
929: }
930: goto noinsert1;
931: }
932: }
934: if (nonew == 1) goto noinsert1;
935: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero (%" PetscInt_FMT ", %" PetscInt_FMT ") in the matrix", row, col);
936: MatSeqXAIJReallocateAIJ(A, a->mbs, bs2, nrow, brow, bcol, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
938: N = nrow++ - 1;
939: high++;
940: /* shift up all the later entries in this row */
941: PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1));
942: PetscCall(PetscArraymove(ap + bs2 * (i + 1), ap + bs2 * i, bs2 * (N - i + 1)));
943: PetscCall(PetscArrayzero(ap + bs2 * i, bs2));
944: rp[i] = bcol;
945: ap[bs2 * i + bs * cidx + ridx] = value;
946: /* for diag block, add/insert its symmetric element a(cidx,ridx) */
947: if (brow == bcol && ridx < cidx) ap[bs2 * i + bs * ridx + cidx] = value;
948: noinsert1:;
949: low = i;
950: }
951: } /* end of loop over added columns */
952: ailen[brow] = nrow;
953: } /* end of loop over added rows */
954: PetscFunctionReturn(PETSC_SUCCESS);
955: }
957: static PetscErrorCode MatICCFactor_SeqSBAIJ(Mat inA, IS row, const MatFactorInfo *info)
958: {
959: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)inA->data;
960: Mat outA;
961: PetscBool row_identity;
963: PetscFunctionBegin;
964: PetscCheck(info->levels == 0, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only levels=0 is supported for in-place icc");
965: PetscCall(ISIdentity(row, &row_identity));
966: PetscCheck(row_identity, PETSC_COMM_SELF, PETSC_ERR_SUP, "Matrix reordering is not supported");
967: PetscCheck(inA->rmap->bs == 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "Matrix block size %" PetscInt_FMT " is not supported", inA->rmap->bs); /* Need to replace MatCholeskyFactorSymbolic_SeqSBAIJ_MSR()! */
969: outA = inA;
970: PetscCall(PetscFree(inA->solvertype));
971: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &inA->solvertype));
973: inA->factortype = MAT_FACTOR_ICC;
974: PetscCall(MatSeqSBAIJSetNumericFactorization_inplace(inA, row_identity));
976: PetscCall(PetscObjectReference((PetscObject)row));
977: PetscCall(ISDestroy(&a->row));
978: a->row = row;
979: PetscCall(PetscObjectReference((PetscObject)row));
980: PetscCall(ISDestroy(&a->col));
981: a->col = row;
983: /* Create the invert permutation so that it can be used in MatCholeskyFactorNumeric() */
984: if (a->icol) PetscCall(ISInvertPermutation(row, PETSC_DECIDE, &a->icol));
986: if (!a->solve_work) PetscCall(PetscMalloc1(inA->rmap->N + inA->rmap->bs, &a->solve_work));
988: PetscCall(MatCholeskyFactorNumeric(outA, inA, info));
989: PetscFunctionReturn(PETSC_SUCCESS);
990: }
992: static PetscErrorCode MatSeqSBAIJSetColumnIndices_SeqSBAIJ(Mat mat, PetscInt *indices)
993: {
994: Mat_SeqSBAIJ *baij = (Mat_SeqSBAIJ *)mat->data;
995: PetscInt i, nz, n;
997: PetscFunctionBegin;
998: nz = baij->maxnz;
999: n = mat->cmap->n;
1000: for (i = 0; i < nz; i++) baij->j[i] = indices[i];
1002: baij->nz = nz;
1003: for (i = 0; i < n; i++) baij->ilen[i] = baij->imax[i];
1005: PetscCall(MatSetOption(mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
1006: PetscFunctionReturn(PETSC_SUCCESS);
1007: }
1009: /*@
1010: MatSeqSBAIJSetColumnIndices - Set the column indices for all the rows
1011: in a `MATSEQSBAIJ` matrix.
1013: Input Parameters:
1014: + mat - the `MATSEQSBAIJ` matrix
1015: - indices - the column indices
1017: Level: advanced
1019: Notes:
1020: This can be called if you have precomputed the nonzero structure of the
1021: matrix and want to provide it to the matrix object to improve the performance
1022: of the `MatSetValues()` operation.
1024: You MUST have set the correct numbers of nonzeros per row in the call to
1025: `MatCreateSeqSBAIJ()`, and the columns indices MUST be sorted.
1027: MUST be called before any calls to `MatSetValues()`
1029: .seealso: [](ch_matrices), `Mat`, `MATSEQSBAIJ`, `MatCreateSeqSBAIJ`
1030: @*/
1031: PetscErrorCode MatSeqSBAIJSetColumnIndices(Mat mat, PetscInt *indices)
1032: {
1033: PetscFunctionBegin;
1035: PetscAssertPointer(indices, 2);
1036: PetscUseMethod(mat, "MatSeqSBAIJSetColumnIndices_C", (Mat, PetscInt *), (mat, indices));
1037: PetscFunctionReturn(PETSC_SUCCESS);
1038: }
1040: static PetscErrorCode MatCopy_SeqSBAIJ(Mat A, Mat B, MatStructure str)
1041: {
1042: PetscBool isbaij;
1044: PetscFunctionBegin;
1045: PetscCall(PetscObjectTypeCompareAny((PetscObject)B, &isbaij, MATSEQSBAIJ, MATMPISBAIJ, ""));
1046: PetscCheck(isbaij, PetscObjectComm((PetscObject)B), PETSC_ERR_SUP, "Not for matrix type %s", ((PetscObject)B)->type_name);
1047: /* If the two matrices have the same copy implementation and nonzero pattern, use fast copy. */
1048: if (str == SAME_NONZERO_PATTERN && A->ops->copy == B->ops->copy) {
1049: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
1050: Mat_SeqSBAIJ *b = (Mat_SeqSBAIJ *)B->data;
1052: PetscCheck(a->i[a->mbs] == b->i[b->mbs], PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Number of nonzeros in two matrices are different");
1053: PetscCheck(a->mbs == b->mbs, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Number of rows in two matrices are different");
1054: PetscCheck(a->bs2 == b->bs2, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Different block size");
1055: PetscCall(PetscArraycpy(b->a, a->a, a->bs2 * a->i[a->mbs]));
1056: PetscCall(PetscObjectStateIncrease((PetscObject)B));
1057: } else {
1058: PetscCall(MatGetRowUpperTriangular(A));
1059: PetscCall(MatCopy_Basic(A, B, str));
1060: PetscCall(MatRestoreRowUpperTriangular(A));
1061: }
1062: PetscFunctionReturn(PETSC_SUCCESS);
1063: }
1065: static PetscErrorCode MatSeqSBAIJGetArray_SeqSBAIJ(Mat A, PetscScalar *array[])
1066: {
1067: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
1069: PetscFunctionBegin;
1070: *array = a->a;
1071: PetscFunctionReturn(PETSC_SUCCESS);
1072: }
1074: static PetscErrorCode MatSeqSBAIJRestoreArray_SeqSBAIJ(Mat A, PetscScalar *array[])
1075: {
1076: PetscFunctionBegin;
1077: *array = NULL;
1078: PetscFunctionReturn(PETSC_SUCCESS);
1079: }
1081: PetscErrorCode MatAXPYGetPreallocation_SeqSBAIJ(Mat Y, Mat X, PetscInt *nnz)
1082: {
1083: PetscInt bs = Y->rmap->bs, mbs = Y->rmap->N / bs;
1084: Mat_SeqSBAIJ *x = (Mat_SeqSBAIJ *)X->data;
1085: Mat_SeqSBAIJ *y = (Mat_SeqSBAIJ *)Y->data;
1087: PetscFunctionBegin;
1088: /* Set the number of nonzeros in the new matrix */
1089: PetscCall(MatAXPYGetPreallocation_SeqX_private(mbs, x->i, x->j, y->i, y->j, nnz));
1090: PetscFunctionReturn(PETSC_SUCCESS);
1091: }
1093: static PetscErrorCode MatAXPY_SeqSBAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
1094: {
1095: Mat_SeqSBAIJ *x = (Mat_SeqSBAIJ *)X->data, *y = (Mat_SeqSBAIJ *)Y->data;
1096: PetscInt bs = Y->rmap->bs, bs2 = bs * bs;
1097: PetscBLASInt one = 1;
1099: PetscFunctionBegin;
1100: if (str == UNKNOWN_NONZERO_PATTERN || (PetscDefined(USE_DEBUG) && str == SAME_NONZERO_PATTERN)) {
1101: PetscBool e = x->nz == y->nz && x->mbs == y->mbs ? PETSC_TRUE : PETSC_FALSE;
1102: if (e) {
1103: PetscCall(PetscArraycmp(x->i, y->i, x->mbs + 1, &e));
1104: if (e) {
1105: PetscCall(PetscArraycmp(x->j, y->j, x->i[x->mbs], &e));
1106: if (e) str = SAME_NONZERO_PATTERN;
1107: }
1108: }
1109: if (!e) PetscCheck(str != SAME_NONZERO_PATTERN, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "MatStructure is not SAME_NONZERO_PATTERN");
1110: }
1111: if (str == SAME_NONZERO_PATTERN) {
1112: PetscScalar alpha = a;
1113: PetscBLASInt bnz;
1114: PetscCall(PetscBLASIntCast(x->nz * bs2, &bnz));
1115: PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, x->a, &one, y->a, &one));
1116: PetscCall(PetscObjectStateIncrease((PetscObject)Y));
1117: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
1118: PetscCall(MatSetOption(X, MAT_GETROW_UPPERTRIANGULAR, PETSC_TRUE));
1119: PetscCall(MatAXPY_Basic(Y, a, X, str));
1120: PetscCall(MatSetOption(X, MAT_GETROW_UPPERTRIANGULAR, PETSC_FALSE));
1121: } else {
1122: Mat B;
1123: PetscInt *nnz;
1124: PetscCheck(bs == X->rmap->bs, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrices must have same block size");
1125: PetscCall(MatGetRowUpperTriangular(X));
1126: PetscCall(MatGetRowUpperTriangular(Y));
1127: PetscCall(PetscMalloc1(Y->rmap->N, &nnz));
1128: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
1129: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
1130: PetscCall(MatSetSizes(B, Y->rmap->n, Y->cmap->n, Y->rmap->N, Y->cmap->N));
1131: PetscCall(MatSetBlockSizesFromMats(B, Y, Y));
1132: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
1133: PetscCall(MatAXPYGetPreallocation_SeqSBAIJ(Y, X, nnz));
1134: PetscCall(MatSeqSBAIJSetPreallocation(B, bs, 0, nnz));
1136: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
1138: PetscCall(MatHeaderMerge(Y, &B));
1139: PetscCall(PetscFree(nnz));
1140: PetscCall(MatRestoreRowUpperTriangular(X));
1141: PetscCall(MatRestoreRowUpperTriangular(Y));
1142: }
1143: PetscFunctionReturn(PETSC_SUCCESS);
1144: }
1146: static PetscErrorCode MatIsStructurallySymmetric_SeqSBAIJ(Mat A, PetscBool *flg)
1147: {
1148: PetscFunctionBegin;
1149: *flg = PETSC_TRUE;
1150: PetscFunctionReturn(PETSC_SUCCESS);
1151: }
1153: static PetscErrorCode MatConjugate_SeqSBAIJ(Mat A)
1154: {
1155: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
1156: PetscInt i, nz = a->bs2 * a->i[a->mbs];
1157: MatScalar *aa = a->a;
1159: PetscFunctionBegin;
1160: for (i = 0; i < nz; i++) aa[i] = PetscConj(aa[i]);
1161: PetscFunctionReturn(PETSC_SUCCESS);
1162: }
1164: static PetscErrorCode MatRealPart_SeqSBAIJ(Mat A)
1165: {
1166: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
1167: PetscInt i, nz = a->bs2 * a->i[a->mbs];
1168: MatScalar *aa = a->a;
1170: PetscFunctionBegin;
1171: for (i = 0; i < nz; i++) aa[i] = PetscRealPart(aa[i]);
1172: PetscFunctionReturn(PETSC_SUCCESS);
1173: }
1175: static PetscErrorCode MatImaginaryPart_SeqSBAIJ(Mat A)
1176: {
1177: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
1178: PetscInt i, nz = a->bs2 * a->i[a->mbs];
1179: MatScalar *aa = a->a;
1181: PetscFunctionBegin;
1182: for (i = 0; i < nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
1183: PetscFunctionReturn(PETSC_SUCCESS);
1184: }
1186: static PetscErrorCode MatZeroRowsColumns_SeqSBAIJ(Mat A, PetscInt is_n, const PetscInt is_idx[], PetscScalar diag, Vec x, Vec b)
1187: {
1188: Mat_SeqSBAIJ *baij = (Mat_SeqSBAIJ *)A->data;
1189: PetscInt i, j, k, count;
1190: PetscInt bs = A->rmap->bs, bs2 = baij->bs2, row, col;
1191: PetscScalar zero = 0.0;
1192: MatScalar *aa;
1193: const PetscScalar *xx;
1194: PetscScalar *bb;
1195: PetscBool *zeroed, vecs = PETSC_FALSE;
1197: PetscFunctionBegin;
1198: /* fix right-hand side if needed */
1199: if (x && b) {
1200: PetscCall(VecGetArrayRead(x, &xx));
1201: PetscCall(VecGetArray(b, &bb));
1202: vecs = PETSC_TRUE;
1203: }
1205: /* zero the columns */
1206: PetscCall(PetscCalloc1(A->rmap->n, &zeroed));
1207: for (i = 0; i < is_n; i++) {
1208: PetscCheck(is_idx[i] >= 0 && is_idx[i] < A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", is_idx[i]);
1209: zeroed[is_idx[i]] = PETSC_TRUE;
1210: }
1211: if (vecs) {
1212: for (i = 0; i < A->rmap->N; i++) {
1213: row = i / bs;
1214: for (j = baij->i[row]; j < baij->i[row + 1]; j++) {
1215: for (k = 0; k < bs; k++) {
1216: col = bs * baij->j[j] + k;
1217: if (col <= i) continue;
1218: aa = baij->a + j * bs2 + (i % bs) + bs * k;
1219: if (!zeroed[i] && zeroed[col]) bb[i] -= aa[0] * xx[col];
1220: if (zeroed[i] && !zeroed[col]) bb[col] -= aa[0] * xx[i];
1221: }
1222: }
1223: }
1224: for (i = 0; i < is_n; i++) bb[is_idx[i]] = diag * xx[is_idx[i]];
1225: }
1227: for (i = 0; i < A->rmap->N; i++) {
1228: if (!zeroed[i]) {
1229: row = i / bs;
1230: for (j = baij->i[row]; j < baij->i[row + 1]; j++) {
1231: for (k = 0; k < bs; k++) {
1232: col = bs * baij->j[j] + k;
1233: if (zeroed[col]) {
1234: aa = baij->a + j * bs2 + (i % bs) + bs * k;
1235: aa[0] = 0.0;
1236: }
1237: }
1238: }
1239: }
1240: }
1241: PetscCall(PetscFree(zeroed));
1242: if (vecs) {
1243: PetscCall(VecRestoreArrayRead(x, &xx));
1244: PetscCall(VecRestoreArray(b, &bb));
1245: }
1247: /* zero the rows */
1248: for (i = 0; i < is_n; i++) {
1249: row = is_idx[i];
1250: count = (baij->i[row / bs + 1] - baij->i[row / bs]) * bs;
1251: aa = PetscSafePointerPlusOffset(baij->a, baij->i[row / bs] * bs2 + (row % bs));
1252: for (k = 0; k < count; k++) {
1253: aa[0] = zero;
1254: aa += bs;
1255: }
1256: if (diag != 0.0) PetscUseTypeMethod(A, setvalues, 1, &row, 1, &row, &diag, INSERT_VALUES);
1257: }
1258: PetscCall(MatAssemblyEnd_SeqSBAIJ(A, MAT_FINAL_ASSEMBLY));
1259: PetscFunctionReturn(PETSC_SUCCESS);
1260: }
1262: static PetscErrorCode MatShift_SeqSBAIJ(Mat Y, PetscScalar a)
1263: {
1264: Mat_SeqSBAIJ *aij = (Mat_SeqSBAIJ *)Y->data;
1266: PetscFunctionBegin;
1267: if (!Y->preallocated || !aij->nz) PetscCall(MatSeqSBAIJSetPreallocation(Y, Y->rmap->bs, 1, NULL));
1268: PetscCall(MatShift_Basic(Y, a));
1269: PetscFunctionReturn(PETSC_SUCCESS);
1270: }
1272: PetscErrorCode MatEliminateZeros_SeqSBAIJ(Mat A, PetscBool keep)
1273: {
1274: Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data;
1275: PetscInt fshift = 0, fshift_prev = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax, j, k;
1276: PetscInt m = A->rmap->N, *ailen = a->ilen;
1277: PetscInt mbs = a->mbs, bs2 = a->bs2, rmax = 0;
1278: MatScalar *aa = a->a, *ap;
1279: PetscBool zero;
1281: PetscFunctionBegin;
1282: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot eliminate zeros for unassembled matrix");
1283: if (m) rmax = ailen[0];
1284: for (i = 1, a->nonzerorowcnt = 0; i <= mbs; i++) {
1285: for (k = ai[i - 1]; k < ai[i]; k++) {
1286: zero = PETSC_TRUE;
1287: ap = aa + bs2 * k;
1288: for (j = 0; j < bs2 && zero; j++) {
1289: if (ap[j] != 0.0) zero = PETSC_FALSE;
1290: }
1291: if (zero && (aj[k] != i - 1 || !keep)) fshift++;
1292: else {
1293: if (zero && aj[k] == i - 1) PetscCall(PetscInfo(A, "Keep the diagonal block at row %" PetscInt_FMT "\n", i - 1));
1294: aj[k - fshift] = aj[k];
1295: PetscCall(PetscArraymove(ap - bs2 * fshift, ap, bs2));
1296: }
1297: }
1298: ai[i - 1] -= fshift_prev;
1299: fshift_prev = fshift;
1300: ailen[i - 1] = imax[i - 1] = ai[i] - fshift - ai[i - 1];
1301: a->nonzerorowcnt += ((ai[i] - fshift - ai[i - 1]) > 0);
1302: rmax = PetscMax(rmax, ailen[i - 1]);
1303: }
1304: if (fshift) {
1305: if (mbs) {
1306: ai[mbs] -= fshift;
1307: a->nz = ai[mbs];
1308: }
1309: 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));
1310: A->nonzerostate++;
1311: A->info.nz_unneeded += (PetscReal)fshift;
1312: a->rmax = rmax;
1313: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
1314: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
1315: }
1316: PetscFunctionReturn(PETSC_SUCCESS);
1317: }
1319: static struct _MatOps MatOps_Values = {MatSetValues_SeqSBAIJ,
1320: MatGetRow_SeqSBAIJ,
1321: MatRestoreRow_SeqSBAIJ,
1322: MatMult_SeqSBAIJ_N,
1323: /* 4*/ MatMultAdd_SeqSBAIJ_N,
1324: MatMult_SeqSBAIJ_N, /* transpose versions are same as non-transpose versions */
1325: MatMultAdd_SeqSBAIJ_N,
1326: NULL,
1327: NULL,
1328: NULL,
1329: /* 10*/ NULL,
1330: NULL,
1331: MatCholeskyFactor_SeqSBAIJ,
1332: MatSOR_SeqSBAIJ,
1333: MatTranspose_SeqSBAIJ,
1334: /* 15*/ MatGetInfo_SeqSBAIJ,
1335: MatEqual_SeqSBAIJ,
1336: MatGetDiagonal_SeqSBAIJ,
1337: MatDiagonalScale_SeqSBAIJ,
1338: MatNorm_SeqSBAIJ,
1339: /* 20*/ NULL,
1340: MatAssemblyEnd_SeqSBAIJ,
1341: MatSetOption_SeqSBAIJ,
1342: MatZeroEntries_SeqSBAIJ,
1343: /* 24*/ NULL,
1344: NULL,
1345: NULL,
1346: NULL,
1347: NULL,
1348: /* 29*/ MatSetUp_Seq_Hash,
1349: NULL,
1350: NULL,
1351: NULL,
1352: NULL,
1353: /* 34*/ MatDuplicate_SeqSBAIJ,
1354: NULL,
1355: NULL,
1356: NULL,
1357: MatICCFactor_SeqSBAIJ,
1358: /* 39*/ MatAXPY_SeqSBAIJ,
1359: MatCreateSubMatrices_SeqSBAIJ,
1360: MatIncreaseOverlap_SeqSBAIJ,
1361: MatGetValues_SeqSBAIJ,
1362: MatCopy_SeqSBAIJ,
1363: /* 44*/ NULL,
1364: MatScale_SeqSBAIJ,
1365: MatShift_SeqSBAIJ,
1366: NULL,
1367: MatZeroRowsColumns_SeqSBAIJ,
1368: /* 49*/ NULL,
1369: MatGetRowIJ_SeqSBAIJ,
1370: MatRestoreRowIJ_SeqSBAIJ,
1371: NULL,
1372: NULL,
1373: /* 54*/ NULL,
1374: NULL,
1375: NULL,
1376: MatPermute_SeqSBAIJ,
1377: MatSetValuesBlocked_SeqSBAIJ,
1378: /* 59*/ MatCreateSubMatrix_SeqSBAIJ,
1379: NULL,
1380: NULL,
1381: NULL,
1382: NULL,
1383: /* 64*/ NULL,
1384: NULL,
1385: NULL,
1386: NULL,
1387: MatGetRowMaxAbs_SeqSBAIJ,
1388: /* 69*/ NULL,
1389: MatConvert_MPISBAIJ_Basic,
1390: NULL,
1391: NULL,
1392: NULL,
1393: /* 74*/ NULL,
1394: NULL,
1395: NULL,
1396: MatGetInertia_SeqSBAIJ,
1397: MatLoad_SeqSBAIJ,
1398: /* 79*/ NULL,
1399: NULL,
1400: MatIsStructurallySymmetric_SeqSBAIJ,
1401: NULL,
1402: NULL,
1403: /* 84*/ NULL,
1404: NULL,
1405: NULL,
1406: NULL,
1407: NULL,
1408: /* 89*/ NULL,
1409: NULL,
1410: NULL,
1411: NULL,
1412: MatConjugate_SeqSBAIJ,
1413: /* 94*/ NULL,
1414: NULL,
1415: MatRealPart_SeqSBAIJ,
1416: MatImaginaryPart_SeqSBAIJ,
1417: MatGetRowUpperTriangular_SeqSBAIJ,
1418: /* 99*/ MatRestoreRowUpperTriangular_SeqSBAIJ,
1419: NULL,
1420: NULL,
1421: NULL,
1422: NULL,
1423: /*104*/ NULL,
1424: NULL,
1425: NULL,
1426: NULL,
1427: NULL,
1428: /*109*/ NULL,
1429: NULL,
1430: NULL,
1431: NULL,
1432: NULL,
1433: /*114*/ NULL,
1434: MatGetColumnReductions_SeqSBAIJ,
1435: NULL,
1436: NULL,
1437: NULL,
1438: /*119*/ NULL,
1439: NULL,
1440: NULL,
1441: NULL,
1442: NULL,
1443: /*124*/ NULL,
1444: MatSetBlockSizes_Default,
1445: NULL,
1446: NULL,
1447: NULL,
1448: /*129*/ MatCreateMPIMatConcatenateSeqMat_SeqSBAIJ,
1449: NULL,
1450: NULL,
1451: NULL,
1452: NULL,
1453: /*134*/ NULL,
1454: MatEliminateZeros_SeqSBAIJ,
1455: NULL,
1456: NULL,
1457: NULL,
1458: /*139*/ NULL,
1459: MatCopyHashToXAIJ_Seq_Hash,
1460: NULL,
1461: NULL,
1462: NULL,
1463: /*144*/ NULL,
1464: NULL,
1465: NULL,
1466: NULL};
1468: static PetscErrorCode MatStoreValues_SeqSBAIJ(Mat mat)
1469: {
1470: Mat_SeqSBAIJ *aij = (Mat_SeqSBAIJ *)mat->data;
1471: PetscInt nz = aij->i[mat->rmap->N] * mat->rmap->bs * aij->bs2;
1473: PetscFunctionBegin;
1474: PetscCheck(aij->nonew == 1, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
1476: /* allocate space for values if not already there */
1477: if (!aij->saved_values) PetscCall(PetscMalloc1(nz + 1, &aij->saved_values));
1479: /* copy values over */
1480: PetscCall(PetscArraycpy(aij->saved_values, aij->a, nz));
1481: PetscFunctionReturn(PETSC_SUCCESS);
1482: }
1484: static PetscErrorCode MatRetrieveValues_SeqSBAIJ(Mat mat)
1485: {
1486: Mat_SeqSBAIJ *aij = (Mat_SeqSBAIJ *)mat->data;
1487: PetscInt nz = aij->i[mat->rmap->N] * mat->rmap->bs * aij->bs2;
1489: PetscFunctionBegin;
1490: PetscCheck(aij->nonew == 1, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
1491: PetscCheck(aij->saved_values, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatStoreValues(A);first");
1493: /* copy values over */
1494: PetscCall(PetscArraycpy(aij->a, aij->saved_values, nz));
1495: PetscFunctionReturn(PETSC_SUCCESS);
1496: }
1498: static PetscErrorCode MatSeqSBAIJSetPreallocation_SeqSBAIJ(Mat B, PetscInt bs, PetscInt nz, const PetscInt nnz[])
1499: {
1500: Mat_SeqSBAIJ *b = (Mat_SeqSBAIJ *)B->data;
1501: PetscInt i, mbs, nbs, bs2;
1502: PetscBool skipallocation = PETSC_FALSE, flg = PETSC_FALSE, realalloc = PETSC_FALSE;
1504: PetscFunctionBegin;
1505: if (B->hash_active) {
1506: PetscInt bs;
1507: B->ops[0] = b->cops;
1508: PetscCall(PetscHMapIJVDestroy(&b->ht));
1509: PetscCall(MatGetBlockSize(B, &bs));
1510: if (bs > 1) PetscCall(PetscHSetIJDestroy(&b->bht));
1511: PetscCall(PetscFree(b->dnz));
1512: PetscCall(PetscFree(b->bdnz));
1513: B->hash_active = PETSC_FALSE;
1514: }
1515: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
1517: PetscCall(MatSetBlockSize(B, bs));
1518: PetscCall(PetscLayoutSetUp(B->rmap));
1519: PetscCall(PetscLayoutSetUp(B->cmap));
1520: PetscCheck(B->rmap->N <= B->cmap->N, PETSC_COMM_SELF, PETSC_ERR_SUP, "SEQSBAIJ matrix cannot have more rows %" PetscInt_FMT " than columns %" PetscInt_FMT, B->rmap->N, B->cmap->N);
1521: PetscCall(PetscLayoutGetBlockSize(B->rmap, &bs));
1523: B->preallocated = PETSC_TRUE;
1525: mbs = B->rmap->N / bs;
1526: nbs = B->cmap->n / bs;
1527: bs2 = bs * bs;
1529: PetscCheck(mbs * bs == B->rmap->N && nbs * bs == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Number rows, cols must be divisible by blocksize");
1531: if (nz == MAT_SKIP_ALLOCATION) {
1532: skipallocation = PETSC_TRUE;
1533: nz = 0;
1534: }
1536: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 3;
1537: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nz cannot be less than 0: value %" PetscInt_FMT, nz);
1538: if (nnz) {
1539: for (i = 0; i < mbs; i++) {
1540: 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]);
1541: PetscCheck(nnz[i] <= nbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nnz cannot be greater than block row length: local row %" PetscInt_FMT " value %" PetscInt_FMT " block rowlength %" PetscInt_FMT, i, nnz[i], nbs);
1542: }
1543: }
1545: B->ops->mult = MatMult_SeqSBAIJ_N;
1546: B->ops->multadd = MatMultAdd_SeqSBAIJ_N;
1547: B->ops->multtranspose = MatMult_SeqSBAIJ_N;
1548: B->ops->multtransposeadd = MatMultAdd_SeqSBAIJ_N;
1550: PetscCall(PetscOptionsGetBool(((PetscObject)B)->options, ((PetscObject)B)->prefix, "-mat_no_unroll", &flg, NULL));
1551: if (!flg) {
1552: switch (bs) {
1553: case 1:
1554: B->ops->mult = MatMult_SeqSBAIJ_1;
1555: B->ops->multadd = MatMultAdd_SeqSBAIJ_1;
1556: B->ops->multtranspose = MatMult_SeqSBAIJ_1;
1557: B->ops->multtransposeadd = MatMultAdd_SeqSBAIJ_1;
1558: break;
1559: case 2:
1560: B->ops->mult = MatMult_SeqSBAIJ_2;
1561: B->ops->multadd = MatMultAdd_SeqSBAIJ_2;
1562: B->ops->multtranspose = MatMult_SeqSBAIJ_2;
1563: B->ops->multtransposeadd = MatMultAdd_SeqSBAIJ_2;
1564: break;
1565: case 3:
1566: B->ops->mult = MatMult_SeqSBAIJ_3;
1567: B->ops->multadd = MatMultAdd_SeqSBAIJ_3;
1568: B->ops->multtranspose = MatMult_SeqSBAIJ_3;
1569: B->ops->multtransposeadd = MatMultAdd_SeqSBAIJ_3;
1570: break;
1571: case 4:
1572: B->ops->mult = MatMult_SeqSBAIJ_4;
1573: B->ops->multadd = MatMultAdd_SeqSBAIJ_4;
1574: B->ops->multtranspose = MatMult_SeqSBAIJ_4;
1575: B->ops->multtransposeadd = MatMultAdd_SeqSBAIJ_4;
1576: break;
1577: case 5:
1578: B->ops->mult = MatMult_SeqSBAIJ_5;
1579: B->ops->multadd = MatMultAdd_SeqSBAIJ_5;
1580: B->ops->multtranspose = MatMult_SeqSBAIJ_5;
1581: B->ops->multtransposeadd = MatMultAdd_SeqSBAIJ_5;
1582: break;
1583: case 6:
1584: B->ops->mult = MatMult_SeqSBAIJ_6;
1585: B->ops->multadd = MatMultAdd_SeqSBAIJ_6;
1586: B->ops->multtranspose = MatMult_SeqSBAIJ_6;
1587: B->ops->multtransposeadd = MatMultAdd_SeqSBAIJ_6;
1588: break;
1589: case 7:
1590: B->ops->mult = MatMult_SeqSBAIJ_7;
1591: B->ops->multadd = MatMultAdd_SeqSBAIJ_7;
1592: B->ops->multtranspose = MatMult_SeqSBAIJ_7;
1593: B->ops->multtransposeadd = MatMultAdd_SeqSBAIJ_7;
1594: break;
1595: }
1596: }
1598: b->mbs = mbs;
1599: b->nbs = nbs;
1600: if (!skipallocation) {
1601: if (!b->imax) {
1602: PetscCall(PetscMalloc2(mbs, &b->imax, mbs, &b->ilen));
1603: b->free_imax_ilen = PETSC_TRUE;
1604: }
1605: if (!nnz) {
1606: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
1607: else if (nz <= 0) nz = 1;
1608: nz = PetscMin(nbs, nz);
1609: for (i = 0; i < mbs; i++) b->imax[i] = nz;
1610: PetscCall(PetscIntMultError(nz, mbs, &nz));
1611: } else {
1612: PetscInt64 nz64 = 0;
1613: for (i = 0; i < mbs; i++) {
1614: b->imax[i] = nnz[i];
1615: nz64 += nnz[i];
1616: }
1617: PetscCall(PetscIntCast(nz64, &nz));
1618: }
1619: /* b->ilen will count nonzeros in each block row so far. */
1620: for (i = 0; i < mbs; i++) b->ilen[i] = 0;
1621: /* nz=(nz+mbs)/2; */ /* total diagonal and superdiagonal nonzero blocks */
1623: /* allocate the matrix space */
1624: PetscCall(MatSeqXAIJFreeAIJ(B, &b->a, &b->j, &b->i));
1625: PetscCall(PetscShmgetAllocateArray(bs2 * nz, sizeof(PetscScalar), (void **)&b->a));
1626: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscInt), (void **)&b->j));
1627: PetscCall(PetscShmgetAllocateArray(B->rmap->n + 1, sizeof(PetscInt), (void **)&b->i));
1628: PetscCall(PetscArrayzero(b->a, nz * bs2));
1629: PetscCall(PetscArrayzero(b->j, nz));
1630: b->free_a = PETSC_TRUE;
1631: b->free_ij = PETSC_TRUE;
1633: /* pointer to beginning of each row */
1634: b->i[0] = 0;
1635: for (i = 1; i < mbs + 1; i++) b->i[i] = b->i[i - 1] + b->imax[i - 1];
1636: } else {
1637: b->free_a = PETSC_FALSE;
1638: b->free_ij = PETSC_FALSE;
1639: }
1641: b->bs2 = bs2;
1642: b->nz = 0;
1643: b->maxnz = nz;
1644: b->inew = NULL;
1645: b->jnew = NULL;
1646: b->anew = NULL;
1647: b->a2anew = NULL;
1648: b->permute = PETSC_FALSE;
1650: B->was_assembled = PETSC_FALSE;
1651: B->assembled = PETSC_FALSE;
1652: if (realalloc) PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1653: PetscFunctionReturn(PETSC_SUCCESS);
1654: }
1656: static PetscErrorCode MatSeqSBAIJSetPreallocationCSR_SeqSBAIJ(Mat B, PetscInt bs, const PetscInt ii[], const PetscInt jj[], const PetscScalar V[])
1657: {
1658: PetscInt i, j, m, nz, anz, nz_max = 0, *nnz;
1659: PetscScalar *values = NULL;
1660: Mat_SeqSBAIJ *b = (Mat_SeqSBAIJ *)B->data;
1661: PetscBool roworiented = b->roworiented;
1662: PetscBool ilw = b->ignore_ltriangular;
1664: PetscFunctionBegin;
1665: PetscCheck(bs >= 1, PetscObjectComm((PetscObject)B), PETSC_ERR_ARG_OUTOFRANGE, "Invalid block size specified, must be positive but it is %" PetscInt_FMT, bs);
1666: PetscCall(PetscLayoutSetBlockSize(B->rmap, bs));
1667: PetscCall(PetscLayoutSetBlockSize(B->cmap, bs));
1668: PetscCall(PetscLayoutSetUp(B->rmap));
1669: PetscCall(PetscLayoutSetUp(B->cmap));
1670: PetscCall(PetscLayoutGetBlockSize(B->rmap, &bs));
1671: m = B->rmap->n / bs;
1673: PetscCheck(!ii[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "ii[0] must be 0 but it is %" PetscInt_FMT, ii[0]);
1674: PetscCall(PetscMalloc1(m + 1, &nnz));
1675: for (i = 0; i < m; i++) {
1676: nz = ii[i + 1] - ii[i];
1677: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %" PetscInt_FMT " has a negative number of columns %" PetscInt_FMT, i, nz);
1678: PetscCheckSorted(nz, jj + ii[i]);
1679: anz = 0;
1680: for (j = 0; j < nz; j++) {
1681: /* count only values on the diagonal or above */
1682: if (jj[ii[i] + j] >= i) {
1683: anz = nz - j;
1684: break;
1685: }
1686: }
1687: nz_max = PetscMax(nz_max, nz);
1688: nnz[i] = anz;
1689: }
1690: PetscCall(MatSeqSBAIJSetPreallocation(B, bs, 0, nnz));
1691: PetscCall(PetscFree(nnz));
1693: values = (PetscScalar *)V;
1694: if (!values) PetscCall(PetscCalloc1(bs * bs * nz_max, &values));
1695: b->ignore_ltriangular = PETSC_TRUE;
1696: for (i = 0; i < m; i++) {
1697: PetscInt ncols = ii[i + 1] - ii[i];
1698: const PetscInt *icols = jj + ii[i];
1700: if (!roworiented || bs == 1) {
1701: const PetscScalar *svals = values + (V ? (bs * bs * ii[i]) : 0);
1702: PetscCall(MatSetValuesBlocked_SeqSBAIJ(B, 1, &i, ncols, icols, svals, INSERT_VALUES));
1703: } else {
1704: for (j = 0; j < ncols; j++) {
1705: const PetscScalar *svals = values + (V ? (bs * bs * (ii[i] + j)) : 0);
1706: PetscCall(MatSetValuesBlocked_SeqSBAIJ(B, 1, &i, 1, &icols[j], svals, INSERT_VALUES));
1707: }
1708: }
1709: }
1710: if (!V) PetscCall(PetscFree(values));
1711: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1712: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1713: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
1714: b->ignore_ltriangular = ilw;
1715: PetscFunctionReturn(PETSC_SUCCESS);
1716: }
1718: /*
1719: This is used to set the numeric factorization for both Cholesky and ICC symbolic factorization
1720: */
1721: PetscErrorCode MatSeqSBAIJSetNumericFactorization_inplace(Mat B, PetscBool natural)
1722: {
1723: PetscBool flg = PETSC_FALSE;
1724: PetscInt bs = B->rmap->bs;
1726: PetscFunctionBegin;
1727: PetscCall(PetscOptionsGetBool(((PetscObject)B)->options, ((PetscObject)B)->prefix, "-mat_no_unroll", &flg, NULL));
1728: if (flg) bs = 8;
1730: if (!natural) {
1731: switch (bs) {
1732: case 1:
1733: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_1_inplace;
1734: break;
1735: case 2:
1736: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_2;
1737: break;
1738: case 3:
1739: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_3;
1740: break;
1741: case 4:
1742: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_4;
1743: break;
1744: case 5:
1745: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_5;
1746: break;
1747: case 6:
1748: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_6;
1749: break;
1750: case 7:
1751: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_7;
1752: break;
1753: default:
1754: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_N;
1755: break;
1756: }
1757: } else {
1758: switch (bs) {
1759: case 1:
1760: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_1_NaturalOrdering_inplace;
1761: break;
1762: case 2:
1763: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_2_NaturalOrdering;
1764: break;
1765: case 3:
1766: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_3_NaturalOrdering;
1767: break;
1768: case 4:
1769: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering;
1770: break;
1771: case 5:
1772: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering;
1773: break;
1774: case 6:
1775: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering;
1776: break;
1777: case 7:
1778: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_7_NaturalOrdering;
1779: break;
1780: default:
1781: B->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqSBAIJ_N_NaturalOrdering;
1782: break;
1783: }
1784: }
1785: PetscFunctionReturn(PETSC_SUCCESS);
1786: }
1788: PETSC_INTERN PetscErrorCode MatConvert_SeqSBAIJ_SeqAIJ(Mat, MatType, MatReuse, Mat *);
1789: PETSC_INTERN PetscErrorCode MatConvert_SeqSBAIJ_SeqBAIJ(Mat, MatType, MatReuse, Mat *);
1790: static PetscErrorCode MatFactorGetSolverType_petsc(Mat A, MatSolverType *type)
1791: {
1792: PetscFunctionBegin;
1793: *type = MATSOLVERPETSC;
1794: PetscFunctionReturn(PETSC_SUCCESS);
1795: }
1797: PETSC_INTERN PetscErrorCode MatGetFactor_seqsbaij_petsc(Mat A, MatFactorType ftype, Mat *B)
1798: {
1799: PetscInt n = A->rmap->n;
1801: PetscFunctionBegin;
1802: if (PetscDefined(USE_COMPLEX) && (ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC) && A->hermitian == PETSC_BOOL3_TRUE && A->symmetric != PETSC_BOOL3_TRUE) {
1803: PetscCall(PetscInfo(A, "Hermitian MAT_FACTOR_CHOLESKY or MAT_FACTOR_ICC are not supported. Use MAT_FACTOR_LU instead.\n"));
1804: *B = NULL;
1805: PetscFunctionReturn(PETSC_SUCCESS);
1806: }
1808: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
1809: PetscCall(MatSetSizes(*B, n, n, n, n));
1810: PetscCheck(ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC, PETSC_COMM_SELF, PETSC_ERR_SUP, "Factor type not supported");
1811: PetscCall(MatSetType(*B, MATSEQSBAIJ));
1812: PetscCall(MatSeqSBAIJSetPreallocation(*B, A->rmap->bs, MAT_SKIP_ALLOCATION, NULL));
1814: (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqSBAIJ;
1815: (*B)->ops->iccfactorsymbolic = MatICCFactorSymbolic_SeqSBAIJ;
1816: PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_CHOLESKY]));
1817: PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ICC]));
1819: (*B)->factortype = ftype;
1820: (*B)->canuseordering = PETSC_TRUE;
1821: PetscCall(PetscFree((*B)->solvertype));
1822: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &(*B)->solvertype));
1823: PetscCall(PetscObjectComposeFunction((PetscObject)*B, "MatFactorGetSolverType_C", MatFactorGetSolverType_petsc));
1824: PetscFunctionReturn(PETSC_SUCCESS);
1825: }
1827: /*@
1828: MatSeqSBAIJGetArray - gives access to the array where the numerical data for a `MATSEQSBAIJ` matrix is stored
1830: Not Collective
1832: Input Parameter:
1833: . A - a `MATSEQSBAIJ` matrix
1835: Output Parameter:
1836: . array - pointer to the data
1838: Level: intermediate
1840: .seealso: [](ch_matrices), `Mat`, `MATSEQSBAIJ`, `MatSeqSBAIJRestoreArray()`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArray()`
1841: @*/
1842: PetscErrorCode MatSeqSBAIJGetArray(Mat A, PetscScalar *array[])
1843: {
1844: PetscFunctionBegin;
1845: PetscUseMethod(A, "MatSeqSBAIJGetArray_C", (Mat, PetscScalar **), (A, array));
1846: PetscFunctionReturn(PETSC_SUCCESS);
1847: }
1849: /*@
1850: MatSeqSBAIJRestoreArray - returns access to the array where the numerical data for a `MATSEQSBAIJ` matrix is stored obtained by `MatSeqSBAIJGetArray()`
1852: Not Collective
1854: Input Parameters:
1855: + A - a `MATSEQSBAIJ` matrix
1856: - array - pointer to the data
1858: Level: intermediate
1860: .seealso: [](ch_matrices), `Mat`, `MATSEQSBAIJ`, `MatSeqSBAIJGetArray()`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArray()`
1861: @*/
1862: PetscErrorCode MatSeqSBAIJRestoreArray(Mat A, PetscScalar *array[])
1863: {
1864: PetscFunctionBegin;
1865: PetscUseMethod(A, "MatSeqSBAIJRestoreArray_C", (Mat, PetscScalar **), (A, array));
1866: PetscCall(PetscObjectStateIncrease((PetscObject)A));
1867: PetscFunctionReturn(PETSC_SUCCESS);
1868: }
1870: /*MC
1871: MATSEQSBAIJ - MATSEQSBAIJ = "seqsbaij" - A matrix type to be used for sequential symmetric block sparse matrices,
1872: based on block compressed sparse row format. Only the upper triangular portion of the matrix is stored.
1874: For complex numbers by default this matrix is symmetric, NOT Hermitian symmetric. To make it Hermitian symmetric you
1875: can call `MatSetOption`(`Mat`, `MAT_HERMITIAN`).
1877: Options Database Key:
1878: . -mat_type seqsbaij - sets the matrix type to "seqsbaij" during a call to `MatSetFromOptions()`
1880: Level: beginner
1882: Notes:
1883: By default if you insert values into the lower triangular part of the matrix they are simply ignored (since they are not
1884: stored and it is assumed they symmetric to the upper triangular). If you call `MatSetOption`(`Mat`,`MAT_IGNORE_LOWER_TRIANGULAR`,`PETSC_FALSE`) or use
1885: the options database `-mat_ignore_lower_triangular` false it will generate an error if you try to set a value in the lower triangular portion.
1887: The number of rows in the matrix must be less than or equal to the number of columns
1889: .seealso: [](ch_matrices), `Mat`, `MATSEQSBAIJ`, `MatCreateSeqSBAIJ()`, `MatType`, `MATMPISBAIJ`
1890: M*/
1891: PETSC_EXTERN PetscErrorCode MatCreate_SeqSBAIJ(Mat B)
1892: {
1893: Mat_SeqSBAIJ *b;
1894: PetscMPIInt size;
1895: PetscBool no_unroll = PETSC_FALSE, no_inode = PETSC_FALSE;
1897: PetscFunctionBegin;
1898: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
1899: PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Comm must be of size 1");
1901: PetscCall(PetscNew(&b));
1902: B->data = (void *)b;
1903: B->ops[0] = MatOps_Values;
1905: B->ops->destroy = MatDestroy_SeqSBAIJ;
1906: B->ops->view = MatView_SeqSBAIJ;
1907: b->row = NULL;
1908: b->icol = NULL;
1909: b->reallocs = 0;
1910: b->saved_values = NULL;
1911: b->inode.limit = 5;
1912: b->inode.max_limit = 5;
1914: b->roworiented = PETSC_TRUE;
1915: b->nonew = 0;
1916: b->diag = NULL;
1917: b->solve_work = NULL;
1918: b->mult_work = NULL;
1919: B->spptr = NULL;
1920: B->info.nz_unneeded = (PetscReal)b->maxnz * b->bs2;
1921: b->keepnonzeropattern = PETSC_FALSE;
1923: b->inew = NULL;
1924: b->jnew = NULL;
1925: b->anew = NULL;
1926: b->a2anew = NULL;
1927: b->permute = PETSC_FALSE;
1929: b->ignore_ltriangular = PETSC_TRUE;
1931: PetscCall(PetscOptionsGetBool(((PetscObject)B)->options, ((PetscObject)B)->prefix, "-mat_ignore_lower_triangular", &b->ignore_ltriangular, NULL));
1933: b->getrow_utriangular = PETSC_FALSE;
1935: PetscCall(PetscOptionsGetBool(((PetscObject)B)->options, ((PetscObject)B)->prefix, "-mat_getrow_uppertriangular", &b->getrow_utriangular, NULL));
1937: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqSBAIJGetArray_C", MatSeqSBAIJGetArray_SeqSBAIJ));
1938: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqSBAIJRestoreArray_C", MatSeqSBAIJRestoreArray_SeqSBAIJ));
1939: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_SeqSBAIJ));
1940: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_SeqSBAIJ));
1941: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqSBAIJSetColumnIndices_C", MatSeqSBAIJSetColumnIndices_SeqSBAIJ));
1942: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqsbaij_seqaij_C", MatConvert_SeqSBAIJ_SeqAIJ));
1943: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqsbaij_seqbaij_C", MatConvert_SeqSBAIJ_SeqBAIJ));
1944: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqSBAIJSetPreallocation_C", MatSeqSBAIJSetPreallocation_SeqSBAIJ));
1945: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqSBAIJSetPreallocationCSR_C", MatSeqSBAIJSetPreallocationCSR_SeqSBAIJ));
1946: #if PetscDefined(HAVE_ELEMENTAL)
1947: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqsbaij_elemental_C", MatConvert_SeqSBAIJ_Elemental));
1948: #endif
1949: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
1950: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqsbaij_scalapack_C", MatConvert_SBAIJ_ScaLAPACK));
1951: #endif
1953: B->symmetry_eternal = PETSC_TRUE;
1954: B->structural_symmetry_eternal = PETSC_TRUE;
1955: B->symmetric = PETSC_BOOL3_TRUE;
1956: B->structurally_symmetric = PETSC_BOOL3_TRUE;
1957: #if !PetscDefined(USE_COMPLEX)
1958: B->hermitian = PETSC_BOOL3_TRUE;
1959: #endif
1961: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQSBAIJ));
1963: PetscOptionsBegin(PetscObjectComm((PetscObject)B), ((PetscObject)B)->prefix, "Options for SEQSBAIJ matrix", "Mat");
1964: PetscCall(PetscOptionsBool("-mat_no_unroll", "Do not optimize for inodes (slower)", NULL, no_unroll, &no_unroll, NULL));
1965: if (no_unroll) PetscCall(PetscInfo(B, "Not using Inode routines due to -mat_no_unroll\n"));
1966: PetscCall(PetscOptionsBool("-mat_no_inode", "Do not optimize for inodes (slower)", NULL, no_inode, &no_inode, NULL));
1967: if (no_inode) PetscCall(PetscInfo(B, "Not using Inode routines due to -mat_no_inode\n"));
1968: PetscCall(PetscOptionsInt("-mat_inode_limit", "Do not use inodes larger than this value", NULL, b->inode.limit, &b->inode.limit, NULL));
1969: PetscOptionsEnd();
1970: b->inode.use = (PetscBool)(!(no_unroll || no_inode));
1971: if (b->inode.limit > b->inode.max_limit) b->inode.limit = b->inode.max_limit;
1972: PetscFunctionReturn(PETSC_SUCCESS);
1973: }
1975: /*@
1976: MatSeqSBAIJSetPreallocation - Creates a sparse symmetric matrix in block AIJ (block
1977: compressed row) `MATSEQSBAIJ` format. For good matrix assembly performance the
1978: user should preallocate the matrix storage by setting the parameter `nz`
1979: (or the array `nnz`).
1981: Collective
1983: Input Parameters:
1984: + B - the symmetric matrix
1985: . bs - size of block, the blocks are ALWAYS square. One can use `MatSetBlockSizes()` to set a different row and column blocksize but the row
1986: blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with `MatCreateVecs()`
1987: . nz - number of block nonzeros per block row (same for all rows)
1988: - nnz - array containing the number of block nonzeros in the upper triangular plus
1989: diagonal portion of each block (possibly different for each block row) or `NULL`
1991: Options Database Keys:
1992: + -mat_no_unroll - uses code that does not unroll the loops in the block calculations (much slower)
1993: - -mat_block_size - size of the blocks to use (only works if a negative bs is passed in
1995: Level: intermediate
1997: Notes:
1998: Specify the preallocated storage with either `nz` or `nnz` (not both).
1999: Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
2000: allocation. See [Sparse Matrices](sec_matsparse) for details.
2002: You can call `MatGetInfo()` to get information on how effective the preallocation was;
2003: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
2004: You can also run with the option `-info` and look for messages with the string
2005: malloc in them to see if additional memory allocation was needed.
2007: If the `nnz` parameter is given then the `nz` parameter is ignored
2009: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATSEQSBAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateSBAIJ()`
2010: @*/
2011: PetscErrorCode MatSeqSBAIJSetPreallocation(Mat B, PetscInt bs, PetscInt nz, const PetscInt nnz[])
2012: {
2013: PetscFunctionBegin;
2017: PetscTryMethod(B, "MatSeqSBAIJSetPreallocation_C", (Mat, PetscInt, PetscInt, const PetscInt[]), (B, bs, nz, nnz));
2018: PetscFunctionReturn(PETSC_SUCCESS);
2019: }
2021: /*@
2022: MatSeqSBAIJSetPreallocationCSR - Creates a sparse parallel matrix in `MATSEQSBAIJ` format using the given nonzero structure and (optional) numerical values
2024: Input Parameters:
2025: + B - the matrix
2026: . bs - size of block, the blocks are ALWAYS square.
2027: . i - the indices into `j` for the start of each local row (indices start with zero)
2028: . j - the column indices for each local row (indices start with zero) these must be sorted for each row
2029: - v - optional values in the matrix, use `NULL` if not provided
2031: Level: advanced
2033: Notes:
2034: The `i`,`j`,`v` values are COPIED with this routine; to avoid the copy use `MatCreateSeqSBAIJWithArrays()`
2036: The order of the entries in values is specified by the `MatOption` `MAT_ROW_ORIENTED`. For example, C programs
2037: may want to use the default `MAT_ROW_ORIENTED` = `PETSC_TRUE` and use an array v[nnz][bs][bs] where the second index is
2038: over rows within a block and the last index is over columns within a block row. Fortran programs will likely set
2039: `MAT_ROW_ORIENTED` = `PETSC_FALSE` and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a
2040: block column and the second index is over columns within a block.
2042: Any entries provided that lie below the diagonal are ignored
2044: Though this routine has Preallocation() in the name it also sets the exact nonzero locations of the matrix entries
2045: and usually the numerical values as well
2047: .seealso: [](ch_matrices), `Mat`, `MATSEQSBAIJ`, `MatCreate()`, `MatCreateSeqSBAIJ()`, `MatSetValuesBlocked()`, `MatSeqSBAIJSetPreallocation()`
2048: @*/
2049: PetscErrorCode MatSeqSBAIJSetPreallocationCSR(Mat B, PetscInt bs, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
2050: {
2051: PetscFunctionBegin;
2055: PetscTryMethod(B, "MatSeqSBAIJSetPreallocationCSR_C", (Mat, PetscInt, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, bs, i, j, v));
2056: PetscFunctionReturn(PETSC_SUCCESS);
2057: }
2059: /*@
2060: MatCreateSeqSBAIJ - Creates a sparse symmetric matrix in (block
2061: compressed row) `MATSEQSBAIJ` format. For good matrix assembly performance the
2062: user should preallocate the matrix storage by setting the parameter `nz`
2063: (or the array `nnz`).
2065: Collective
2067: Input Parameters:
2068: + comm - MPI communicator, set to `PETSC_COMM_SELF`
2069: . bs - size of block, the blocks are ALWAYS square. One can use `MatSetBlockSizes()` to set a different row and column blocksize but the row
2070: blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs()
2071: . m - number of rows
2072: . n - number of columns
2073: . nz - number of block nonzeros per block row (same for all rows)
2074: - nnz - array containing the number of block nonzeros in the upper triangular plus
2075: diagonal portion of each block (possibly different for each block row) or `NULL`
2077: Output Parameter:
2078: . A - the symmetric matrix
2080: Options Database Keys:
2081: + -mat_no_unroll - uses code that does not unroll the loops in the block calculations (much slower)
2082: - -mat_block_size - size of the blocks to use
2084: Level: intermediate
2086: Notes:
2087: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
2088: MatXXXXSetPreallocation() paradigm instead of this routine directly.
2089: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
2091: The number of rows and columns must be divisible by blocksize.
2092: This matrix type does not support complex Hermitian operation.
2094: Specify the preallocated storage with either `nz` or `nnz` (not both).
2095: Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
2096: allocation. See [Sparse Matrices](sec_matsparse) for details.
2098: If the `nnz` parameter is given then the `nz` parameter is ignored
2100: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATSEQSBAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateSBAIJ()`
2101: @*/
2102: PetscErrorCode MatCreateSeqSBAIJ(MPI_Comm comm, PetscInt bs, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
2103: {
2104: PetscFunctionBegin;
2105: PetscCall(MatCreate(comm, A));
2106: PetscCall(MatSetSizes(*A, m, n, m, n));
2107: PetscCall(MatSetType(*A, MATSEQSBAIJ));
2108: PetscCall(MatSeqSBAIJSetPreallocation(*A, bs, nz, (PetscInt *)nnz));
2109: PetscFunctionReturn(PETSC_SUCCESS);
2110: }
2112: PetscErrorCode MatDuplicate_SeqSBAIJ(Mat A, MatDuplicateOption cpvalues, Mat *B)
2113: {
2114: Mat C;
2115: Mat_SeqSBAIJ *c, *a = (Mat_SeqSBAIJ *)A->data;
2116: PetscInt i, mbs = a->mbs, nz = a->nz, bs2 = a->bs2;
2118: PetscFunctionBegin;
2119: PetscCheck(A->assembled, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Cannot duplicate unassembled matrix");
2120: PetscCheck(a->i[mbs] == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Corrupt matrix");
2122: *B = NULL;
2123: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
2124: PetscCall(MatSetSizes(C, A->rmap->N, A->cmap->n, A->rmap->N, A->cmap->n));
2125: PetscCall(MatSetBlockSizesFromMats(C, A, A));
2126: PetscCall(MatSetType(C, MATSEQSBAIJ));
2127: c = (Mat_SeqSBAIJ *)C->data;
2129: C->preallocated = PETSC_TRUE;
2130: C->factortype = A->factortype;
2131: c->row = NULL;
2132: c->icol = NULL;
2133: c->saved_values = NULL;
2134: c->keepnonzeropattern = a->keepnonzeropattern;
2135: C->assembled = PETSC_TRUE;
2137: PetscCall(PetscLayoutReference(A->rmap, &C->rmap));
2138: PetscCall(PetscLayoutReference(A->cmap, &C->cmap));
2139: c->bs2 = a->bs2;
2140: c->mbs = a->mbs;
2141: c->nbs = a->nbs;
2143: if (cpvalues == MAT_SHARE_NONZERO_PATTERN) {
2144: c->imax = a->imax;
2145: c->ilen = a->ilen;
2146: c->free_imax_ilen = PETSC_FALSE;
2147: } else {
2148: PetscCall(PetscMalloc2(mbs + 1, &c->imax, mbs + 1, &c->ilen));
2149: for (i = 0; i < mbs; i++) {
2150: c->imax[i] = a->imax[i];
2151: c->ilen[i] = a->ilen[i];
2152: }
2153: c->free_imax_ilen = PETSC_TRUE;
2154: }
2156: /* allocate the matrix space */
2157: PetscCall(PetscShmgetAllocateArray(bs2 * nz, sizeof(PetscScalar), (void **)&c->a));
2158: c->free_a = PETSC_TRUE;
2159: if (cpvalues == MAT_SHARE_NONZERO_PATTERN) {
2160: PetscCall(PetscArrayzero(c->a, bs2 * nz));
2161: c->i = a->i;
2162: c->j = a->j;
2163: c->free_ij = PETSC_FALSE;
2164: c->parent = A;
2165: PetscCall(PetscObjectReference((PetscObject)A));
2166: PetscCall(MatSetOption(A, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
2167: PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
2168: } else {
2169: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscInt), (void **)&c->j));
2170: PetscCall(PetscShmgetAllocateArray(mbs + 1, sizeof(PetscInt), (void **)&c->i));
2171: PetscCall(PetscArraycpy(c->i, a->i, mbs + 1));
2172: c->free_ij = PETSC_TRUE;
2173: }
2174: if (mbs > 0) {
2175: if (cpvalues != MAT_SHARE_NONZERO_PATTERN) PetscCall(PetscArraycpy(c->j, a->j, nz));
2176: if (cpvalues == MAT_COPY_VALUES) {
2177: PetscCall(PetscArraycpy(c->a, a->a, bs2 * nz));
2178: } else {
2179: PetscCall(PetscArrayzero(c->a, bs2 * nz));
2180: }
2181: if (a->jshort) {
2182: /* cannot share jshort, it is reallocated in MatAssemblyEnd_SeqSBAIJ() */
2183: /* if the parent matrix is reassembled, this child matrix will never notice */
2184: PetscCall(PetscMalloc1(nz, &c->jshort));
2185: PetscCall(PetscArraycpy(c->jshort, a->jshort, nz));
2187: c->free_jshort = PETSC_TRUE;
2188: }
2189: }
2191: c->roworiented = a->roworiented;
2192: c->nonew = a->nonew;
2193: c->nz = a->nz;
2194: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
2195: c->solve_work = NULL;
2196: c->mult_work = NULL;
2198: *B = C;
2199: PetscCall(PetscFunctionListDuplicate(((PetscObject)A)->qlist, &((PetscObject)C)->qlist));
2200: PetscFunctionReturn(PETSC_SUCCESS);
2201: }
2203: /* Used for both SeqBAIJ and SeqSBAIJ matrices */
2204: #define MatLoad_SeqSBAIJ_Binary MatLoad_SeqBAIJ_Binary
2206: PetscErrorCode MatLoad_SeqSBAIJ(Mat mat, PetscViewer viewer)
2207: {
2208: PetscBool isbinary;
2210: PetscFunctionBegin;
2211: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
2212: PetscCheck(isbinary, PetscObjectComm((PetscObject)viewer), PETSC_ERR_SUP, "Viewer type %s not yet supported for reading %s matrices", ((PetscObject)viewer)->type_name, ((PetscObject)mat)->type_name);
2213: PetscCall(MatLoad_SeqSBAIJ_Binary(mat, viewer));
2214: PetscFunctionReturn(PETSC_SUCCESS);
2215: }
2217: /*@
2218: MatCreateSeqSBAIJWithArrays - Creates an sequential `MATSEQSBAIJ` matrix using matrix elements
2219: (upper triangular entries in CSR format) provided by the user.
2221: Collective
2223: Input Parameters:
2224: + comm - must be an MPI communicator of size 1
2225: . bs - size of block
2226: . m - number of rows
2227: . n - number of columns
2228: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of block elements in that row block row of the matrix
2229: . j - column indices
2230: - a - matrix values
2232: Output Parameter:
2233: . mat - the matrix
2235: Level: advanced
2237: Notes:
2238: The `i`, `j`, and `a` arrays are not copied by this routine, the user must free these arrays
2239: once the matrix is destroyed
2241: You cannot set new nonzero locations into this matrix, that will generate an error.
2243: The `i` and `j` indices are 0 based
2245: When block size is greater than 1 the matrix values must be stored using the `MATSBAIJ` storage format. For block size of 1
2246: it is the regular CSR format excluding the lower triangular elements.
2248: .seealso: [](ch_matrices), `Mat`, `MATSEQSBAIJ`, `MatCreate()`, `MatCreateSBAIJ()`, `MatCreateSeqSBAIJ()`
2249: @*/
2250: PetscErrorCode MatCreateSeqSBAIJWithArrays(MPI_Comm comm, PetscInt bs, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat)
2251: {
2252: Mat_SeqSBAIJ *sbaij;
2254: PetscFunctionBegin;
2255: PetscCheck(bs == 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "block size %" PetscInt_FMT " > 1 is not supported yet", bs);
2256: PetscCheck(m == 0 || i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
2258: PetscCall(MatCreate(comm, mat));
2259: PetscCall(MatSetSizes(*mat, m, n, m, n));
2260: PetscCall(MatSetType(*mat, MATSEQSBAIJ));
2261: PetscCall(MatSeqSBAIJSetPreallocation(*mat, bs, MAT_SKIP_ALLOCATION, NULL));
2262: sbaij = (Mat_SeqSBAIJ *)(*mat)->data;
2263: PetscCall(PetscMalloc2(m, &sbaij->imax, m, &sbaij->ilen));
2265: sbaij->i = i;
2266: sbaij->j = j;
2267: sbaij->a = a;
2269: sbaij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
2270: sbaij->free_a = PETSC_FALSE;
2271: sbaij->free_ij = PETSC_FALSE;
2272: sbaij->free_imax_ilen = PETSC_TRUE;
2274: for (PetscInt ii = 0; ii < m; ii++) {
2275: sbaij->ilen[ii] = sbaij->imax[ii] = i[ii + 1] - i[ii];
2276: PetscCheck(i[ii + 1] >= i[ii], 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]);
2277: }
2278: if (PetscDefined(USE_DEBUG)) {
2279: for (PetscInt ii = 0; ii < sbaij->i[m]; ii++) {
2280: PetscCheck(j[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative column index at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]);
2281: PetscCheck(j[ii] < n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column index too large at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]);
2282: }
2283: }
2285: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
2286: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
2287: PetscFunctionReturn(PETSC_SUCCESS);
2288: }
2290: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqSBAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
2291: {
2292: PetscFunctionBegin;
2293: PetscCall(MatCreateMPIMatConcatenateSeqMat_MPISBAIJ(comm, inmat, n, scall, outmat));
2294: PetscFunctionReturn(PETSC_SUCCESS);
2295: }