Actual source code: dense.c
1: /*
2: Defines the basic matrix operations for sequential dense.
3: Portions of this code are under:
4: Copyright (c) 2022 Advanced Micro Devices, Inc. All rights reserved.
5: */
7: #include <../src/mat/impls/dense/seq/dense.h>
8: #include <../src/mat/impls/dense/mpi/mpidense.h>
9: #include <petscblaslapack.h>
10: #include <../src/mat/impls/aij/seq/aij.h>
11: #include <petsc/private/vecimpl.h>
13: PetscErrorCode MatSeqDenseSymmetrize_Private(Mat A, PetscBool hermitian)
14: {
15: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
16: PetscInt j, k, n = A->rmap->n;
17: PetscScalar *v;
19: PetscFunctionBegin;
20: PetscCheck(A->rmap->n == A->cmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Cannot symmetrize a rectangular matrix");
21: PetscCall(MatDenseGetArray(A, &v));
22: if (!hermitian) {
23: for (k = 0; k < n; k++) {
24: for (j = k; j < n; j++) v[j * mat->lda + k] = v[k * mat->lda + j];
25: }
26: } else {
27: for (k = 0; k < n; k++) {
28: for (j = k; j < n; j++) v[j * mat->lda + k] = PetscConj(v[k * mat->lda + j]);
29: }
30: }
31: PetscCall(MatDenseRestoreArray(A, &v));
32: PetscFunctionReturn(PETSC_SUCCESS);
33: }
35: PetscErrorCode MatSeqDenseInvertFactors_Private(Mat A)
36: {
37: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
38: PetscBLASInt info, n;
40: PetscFunctionBegin;
41: if (!A->rmap->n || !A->cmap->n) PetscFunctionReturn(PETSC_SUCCESS);
42: PetscCall(PetscBLASIntCast(A->cmap->n, &n));
43: if (A->factortype == MAT_FACTOR_LU) {
44: PetscCheck(mat->pivots, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Pivots not present");
45: if (!mat->fwork) {
46: mat->lfwork = n;
47: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
48: }
49: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
50: PetscCallBLAS("LAPACKgetri", LAPACKgetri_(&n, mat->v, &mat->lda, mat->pivots, mat->fwork, &mat->lfwork, &info));
51: PetscCall(PetscFPTrapPop());
52: PetscCall(PetscLogFlops((1.0 * A->cmap->n * A->cmap->n * A->cmap->n) / 3.0));
53: } else if (A->factortype == MAT_FACTOR_CHOLESKY) {
54: if (A->spd == PETSC_BOOL3_TRUE) {
55: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
56: PetscCallBLAS("LAPACKpotri", LAPACKpotri_("L", &n, mat->v, &mat->lda, &info));
57: PetscCall(PetscFPTrapPop());
58: PetscCall(MatSeqDenseSymmetrize_Private(A, PETSC_TRUE));
59: #if defined(PETSC_USE_COMPLEX)
60: } else if (A->hermitian == PETSC_BOOL3_TRUE) {
61: PetscCheck(mat->pivots, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Pivots not present");
62: PetscCheck(mat->fwork, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Fwork not present");
63: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
64: PetscCallBLAS("LAPACKhetri", LAPACKhetri_("L", &n, mat->v, &mat->lda, mat->pivots, mat->fwork, &info));
65: PetscCall(PetscFPTrapPop());
66: PetscCall(MatSeqDenseSymmetrize_Private(A, PETSC_TRUE));
67: #endif
68: } else { /* symmetric case */
69: PetscCheck(mat->pivots, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Pivots not present");
70: PetscCheck(mat->fwork, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Fwork not present");
71: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
72: PetscCallBLAS("LAPACKsytri", LAPACKsytri_("L", &n, mat->v, &mat->lda, mat->pivots, mat->fwork, &info));
73: PetscCall(PetscFPTrapPop());
74: PetscCall(MatSeqDenseSymmetrize_Private(A, PETSC_FALSE));
75: }
76: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_MAT_CH_ZRPVT, "Bad Inversion: zero pivot in row %" PetscBLASInt_FMT, info - 1);
77: PetscCall(PetscLogFlops((1.0 * A->cmap->n * A->cmap->n * A->cmap->n) / 3.0));
78: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix must be factored to solve");
80: A->ops->solve = NULL;
81: A->ops->matsolve = NULL;
82: A->ops->solvetranspose = NULL;
83: A->ops->matsolvetranspose = NULL;
84: A->ops->solveadd = NULL;
85: A->ops->solvetransposeadd = NULL;
86: A->factortype = MAT_FACTOR_NONE;
87: PetscCall(PetscFree(A->solvertype));
88: PetscFunctionReturn(PETSC_SUCCESS);
89: }
91: static PetscErrorCode MatZeroRowsColumns_SeqDense(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
92: {
93: Mat_SeqDense *l = (Mat_SeqDense *)A->data;
94: PetscInt m = l->lda, n = A->cmap->n, r = A->rmap->n, i, j;
95: PetscScalar *slot, *bb, *v;
96: const PetscScalar *xx;
98: PetscFunctionBegin;
99: if (PetscDefined(USE_DEBUG)) {
100: for (i = 0; i < N; i++) {
101: PetscCheck(rows[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative row requested to be zeroed");
102: PetscCheck(rows[i] < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %" PetscInt_FMT " requested to be zeroed greater than or equal number of rows %" PetscInt_FMT, rows[i], A->rmap->n);
103: PetscCheck(rows[i] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Col %" PetscInt_FMT " requested to be zeroed greater than or equal number of cols %" PetscInt_FMT, rows[i], A->cmap->n);
104: }
105: }
106: if (!N) PetscFunctionReturn(PETSC_SUCCESS);
108: /* fix right-hand side if needed */
109: if (x && b) {
110: Vec xt;
112: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only coded for square matrices");
113: PetscCall(VecDuplicate(x, &xt));
114: PetscCall(VecCopy(x, xt));
115: PetscCall(VecScale(xt, -1.0));
116: PetscCall(MatMultAdd(A, xt, b, b));
117: PetscCall(VecDestroy(&xt));
118: PetscCall(VecGetArrayRead(x, &xx));
119: PetscCall(VecGetArray(b, &bb));
120: for (i = 0; i < N; i++) bb[rows[i]] = diag * xx[rows[i]];
121: PetscCall(VecRestoreArrayRead(x, &xx));
122: PetscCall(VecRestoreArray(b, &bb));
123: }
125: PetscCall(MatDenseGetArray(A, &v));
126: for (i = 0; i < N; i++) {
127: slot = v + rows[i] * m;
128: PetscCall(PetscArrayzero(slot, r));
129: }
130: for (i = 0; i < N; i++) {
131: slot = v + rows[i];
132: for (j = 0; j < n; j++) {
133: *slot = 0.0;
134: slot += m;
135: }
136: }
137: if (diag != 0.0) {
138: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only coded for square matrices");
139: for (i = 0; i < N; i++) {
140: slot = v + (m + 1) * rows[i];
141: *slot = diag;
142: }
143: }
144: PetscCall(MatDenseRestoreArray(A, &v));
145: PetscFunctionReturn(PETSC_SUCCESS);
146: }
148: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
149: {
150: Mat B = NULL;
151: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
152: Mat_SeqDense *b;
153: PetscInt *ai = a->i, *aj = a->j, m = A->rmap->N, n = A->cmap->N, i;
154: const MatScalar *av;
155: PetscBool isseqdense;
157: PetscFunctionBegin;
158: if (reuse == MAT_REUSE_MATRIX) {
159: PetscCall(PetscObjectTypeCompare((PetscObject)*newmat, MATSEQDENSE, &isseqdense));
160: PetscCheck(isseqdense, PetscObjectComm((PetscObject)*newmat), PETSC_ERR_USER, "Cannot reuse matrix of type %s", ((PetscObject)*newmat)->type_name);
161: }
162: if (reuse != MAT_REUSE_MATRIX) {
163: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
164: PetscCall(MatSetSizes(B, m, n, m, n));
165: PetscCall(MatSetType(B, MATSEQDENSE));
166: PetscCall(MatSeqDenseSetPreallocation(B, NULL));
167: b = (Mat_SeqDense *)B->data;
168: } else {
169: b = (Mat_SeqDense *)((*newmat)->data);
170: for (i = 0; i < n; i++) PetscCall(PetscArrayzero(b->v + i * b->lda, m));
171: }
172: PetscCall(MatSeqAIJGetArrayRead(A, &av));
173: for (i = 0; i < m; i++) {
174: PetscInt j;
175: for (j = 0; j < ai[1] - ai[0]; j++) {
176: b->v[*aj * b->lda + i] = *av;
177: aj++;
178: av++;
179: }
180: ai++;
181: }
182: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
184: if (reuse == MAT_INPLACE_MATRIX) {
185: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
186: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
187: PetscCall(MatHeaderReplace(A, &B));
188: } else {
189: if (B) *newmat = B;
190: PetscCall(MatAssemblyBegin(*newmat, MAT_FINAL_ASSEMBLY));
191: PetscCall(MatAssemblyEnd(*newmat, MAT_FINAL_ASSEMBLY));
192: }
193: PetscFunctionReturn(PETSC_SUCCESS);
194: }
196: PETSC_INTERN PetscErrorCode MatConvert_SeqDense_SeqAIJ(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
197: {
198: Mat B = NULL;
199: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
200: PetscInt i, j;
201: PetscInt *rows, *nnz;
202: MatScalar *aa = a->v, *vals;
204: PetscFunctionBegin;
205: PetscCall(PetscCalloc3(A->rmap->n, &rows, A->rmap->n, &nnz, A->rmap->n, &vals));
206: if (reuse != MAT_REUSE_MATRIX) {
207: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
208: PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
209: PetscCall(MatSetType(B, MATSEQAIJ));
210: for (j = 0; j < A->cmap->n; j++) {
211: for (i = 0; i < A->rmap->n; i++)
212: if (aa[i] != 0.0 || (i == j && A->cmap->n == A->rmap->n)) ++nnz[i];
213: aa += a->lda;
214: }
215: PetscCall(MatSeqAIJSetPreallocation(B, PETSC_DETERMINE, nnz));
216: } else B = *newmat;
217: aa = a->v;
218: for (j = 0; j < A->cmap->n; j++) {
219: PetscInt numRows = 0;
220: for (i = 0; i < A->rmap->n; i++)
221: if (aa[i] != 0.0 || (i == j && A->cmap->n == A->rmap->n)) {
222: rows[numRows] = i;
223: vals[numRows++] = aa[i];
224: }
225: PetscCall(MatSetValues(B, numRows, rows, 1, &j, vals, INSERT_VALUES));
226: aa += a->lda;
227: }
228: PetscCall(PetscFree3(rows, nnz, vals));
229: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
230: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
232: if (reuse == MAT_INPLACE_MATRIX) {
233: PetscCall(MatHeaderReplace(A, &B));
234: } else if (reuse != MAT_REUSE_MATRIX) *newmat = B;
235: PetscFunctionReturn(PETSC_SUCCESS);
236: }
238: PetscErrorCode MatAXPY_SeqDense(Mat Y, PetscScalar alpha, Mat X, MatStructure str)
239: {
240: Mat_SeqDense *x = (Mat_SeqDense *)X->data, *y = (Mat_SeqDense *)Y->data;
241: const PetscScalar *xv;
242: PetscScalar *yv;
243: PetscBLASInt N, m, ldax = 0, lday = 0, one = 1;
245: PetscFunctionBegin;
246: PetscCall(MatDenseGetArrayRead(X, &xv));
247: PetscCall(MatDenseGetArray(Y, &yv));
248: PetscCall(PetscBLASIntCast(X->rmap->n * X->cmap->n, &N));
249: PetscCall(PetscBLASIntCast(X->rmap->n, &m));
250: PetscCall(PetscBLASIntCast(x->lda, &ldax));
251: PetscCall(PetscBLASIntCast(y->lda, &lday));
252: if (ldax > m || lday > m) {
253: for (PetscInt j = 0; j < X->cmap->n; j++) PetscCallBLAS("BLASaxpy", BLASaxpy_(&m, &alpha, PetscSafePointerPlusOffset(xv, j * ldax), &one, PetscSafePointerPlusOffset(yv, j * lday), &one));
254: } else {
255: PetscCallBLAS("BLASaxpy", BLASaxpy_(&N, &alpha, xv, &one, yv, &one));
256: }
257: PetscCall(MatDenseRestoreArrayRead(X, &xv));
258: PetscCall(MatDenseRestoreArray(Y, &yv));
259: PetscCall(PetscLogFlops(PetscMax(2.0 * N - 1, 0)));
260: PetscFunctionReturn(PETSC_SUCCESS);
261: }
263: static PetscErrorCode MatGetInfo_SeqDense(Mat A, MatInfoType flag, MatInfo *info)
264: {
265: PetscLogDouble N = A->rmap->n * A->cmap->n;
267: PetscFunctionBegin;
268: info->block_size = 1.0;
269: info->nz_allocated = N;
270: info->nz_used = N;
271: info->nz_unneeded = 0;
272: info->assemblies = A->num_ass;
273: info->mallocs = 0;
274: info->memory = 0; /* REVIEW ME */
275: info->fill_ratio_given = 0;
276: info->fill_ratio_needed = 0;
277: info->factor_mallocs = 0;
278: PetscFunctionReturn(PETSC_SUCCESS);
279: }
281: PetscErrorCode MatScale_SeqDense(Mat A, PetscScalar alpha)
282: {
283: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
284: PetscScalar *v;
285: PetscBLASInt one = 1, j, nz, lda = 0;
287: PetscFunctionBegin;
288: PetscCall(MatDenseGetArray(A, &v));
289: PetscCall(PetscBLASIntCast(a->lda, &lda));
290: if (lda > A->rmap->n) {
291: PetscCall(PetscBLASIntCast(A->rmap->n, &nz));
292: for (j = 0; j < A->cmap->n; j++) PetscCallBLAS("BLASscal", BLASscal_(&nz, &alpha, v + j * lda, &one));
293: } else {
294: PetscCall(PetscBLASIntCast(A->rmap->n * A->cmap->n, &nz));
295: PetscCallBLAS("BLASscal", BLASscal_(&nz, &alpha, v, &one));
296: }
297: PetscCall(PetscLogFlops(A->rmap->n * A->cmap->n));
298: PetscCall(MatDenseRestoreArray(A, &v));
299: PetscFunctionReturn(PETSC_SUCCESS);
300: }
302: PetscErrorCode MatShift_SeqDense(Mat A, PetscScalar alpha)
303: {
304: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
305: PetscScalar *v;
306: PetscInt j, k;
308: PetscFunctionBegin;
309: PetscCall(MatDenseGetArray(A, &v));
310: k = PetscMin(A->rmap->n, A->cmap->n);
311: for (j = 0; j < k; j++) v[j + j * a->lda] += alpha;
312: PetscCall(PetscLogFlops(k));
313: PetscCall(MatDenseRestoreArray(A, &v));
314: PetscFunctionReturn(PETSC_SUCCESS);
315: }
317: static PetscErrorCode MatIsHermitian_SeqDense(Mat A, PetscReal rtol, PetscBool *fl)
318: {
319: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
320: PetscInt i, j, m = A->rmap->n, N = a->lda;
321: const PetscScalar *v;
323: PetscFunctionBegin;
324: *fl = PETSC_FALSE;
325: if (A->rmap->n != A->cmap->n) PetscFunctionReturn(PETSC_SUCCESS);
326: PetscCall(MatDenseGetArrayRead(A, &v));
327: for (i = 0; i < m; i++) {
328: for (j = i; j < m; j++) {
329: if (PetscAbsScalar(v[i + j * N] - PetscConj(v[j + i * N])) > rtol) goto restore;
330: }
331: }
332: *fl = PETSC_TRUE;
333: restore:
334: PetscCall(MatDenseRestoreArrayRead(A, &v));
335: PetscFunctionReturn(PETSC_SUCCESS);
336: }
338: static PetscErrorCode MatIsSymmetric_SeqDense(Mat A, PetscReal rtol, PetscBool *fl)
339: {
340: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
341: PetscInt i, j, m = A->rmap->n, N = a->lda;
342: const PetscScalar *v;
344: PetscFunctionBegin;
345: *fl = PETSC_FALSE;
346: if (A->rmap->n != A->cmap->n) PetscFunctionReturn(PETSC_SUCCESS);
347: PetscCall(MatDenseGetArrayRead(A, &v));
348: for (i = 0; i < m; i++) {
349: for (j = i; j < m; j++) {
350: if (PetscAbsScalar(v[i + j * N] - v[j + i * N]) > rtol) goto restore;
351: }
352: }
353: *fl = PETSC_TRUE;
354: restore:
355: PetscCall(MatDenseRestoreArrayRead(A, &v));
356: PetscFunctionReturn(PETSC_SUCCESS);
357: }
359: PetscErrorCode MatDuplicateNoCreate_SeqDense(Mat newi, Mat A, MatDuplicateOption cpvalues)
360: {
361: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
362: PetscInt lda = mat->lda, j, m, nlda = lda;
363: PetscBool isdensecpu;
365: PetscFunctionBegin;
366: PetscCall(PetscLayoutReference(A->rmap, &newi->rmap));
367: PetscCall(PetscLayoutReference(A->cmap, &newi->cmap));
368: if (cpvalues == MAT_SHARE_NONZERO_PATTERN) { /* propagate LDA */
369: PetscCall(MatDenseSetLDA(newi, lda));
370: }
371: PetscCall(PetscObjectTypeCompare((PetscObject)newi, MATSEQDENSE, &isdensecpu));
372: if (isdensecpu) PetscCall(MatSeqDenseSetPreallocation(newi, NULL));
373: if (cpvalues == MAT_COPY_VALUES) {
374: const PetscScalar *av;
375: PetscScalar *v;
377: PetscCall(MatDenseGetArrayRead(A, &av));
378: PetscCall(MatDenseGetArrayWrite(newi, &v));
379: PetscCall(MatDenseGetLDA(newi, &nlda));
380: m = A->rmap->n;
381: if (lda > m || nlda > m) {
382: for (j = 0; j < A->cmap->n; j++) PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(v, j * nlda), PetscSafePointerPlusOffset(av, j * lda), m));
383: } else {
384: PetscCall(PetscArraycpy(v, av, A->rmap->n * A->cmap->n));
385: }
386: PetscCall(MatDenseRestoreArrayWrite(newi, &v));
387: PetscCall(MatDenseRestoreArrayRead(A, &av));
388: PetscCall(MatPropagateSymmetryOptions(A, newi));
389: }
390: PetscFunctionReturn(PETSC_SUCCESS);
391: }
393: PetscErrorCode MatDuplicate_SeqDense(Mat A, MatDuplicateOption cpvalues, Mat *newmat)
394: {
395: PetscFunctionBegin;
396: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), newmat));
397: PetscCall(MatSetSizes(*newmat, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n));
398: PetscCall(MatSetType(*newmat, ((PetscObject)A)->type_name));
399: PetscCall(MatDuplicateNoCreate_SeqDense(*newmat, A, cpvalues));
400: PetscFunctionReturn(PETSC_SUCCESS);
401: }
403: static PetscErrorCode MatSolve_SeqDense_Internal_LU(Mat A, PetscScalar *x, PetscBLASInt ldx, PetscBLASInt m, PetscBLASInt nrhs, PetscBLASInt k, PetscBool T)
404: {
405: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
406: PetscBLASInt info;
408: PetscFunctionBegin;
409: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
410: PetscCallBLAS("LAPACKgetrs", LAPACKgetrs_(T ? "T" : "N", &m, &nrhs, mat->v, &mat->lda, mat->pivots, x, &m, &info));
411: PetscCall(PetscFPTrapPop());
412: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_LIB, "GETRS - Bad solve %" PetscBLASInt_FMT, info);
413: PetscCall(PetscLogFlops(nrhs * (2.0 * m * m - m)));
414: PetscFunctionReturn(PETSC_SUCCESS);
415: }
417: static PetscErrorCode MatConjugate_SeqDense(Mat);
419: static PetscErrorCode MatSolve_SeqDense_Internal_Cholesky(Mat A, PetscScalar *x, PetscBLASInt ldx, PetscBLASInt m, PetscBLASInt nrhs, PetscBLASInt k, PetscBool T)
420: {
421: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
422: PetscBLASInt info;
424: PetscFunctionBegin;
425: if (A->spd == PETSC_BOOL3_TRUE) {
426: if (PetscDefined(USE_COMPLEX) && T) PetscCall(MatConjugate_SeqDense(A));
427: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
428: PetscCallBLAS("LAPACKpotrs", LAPACKpotrs_("L", &m, &nrhs, mat->v, &mat->lda, x, &m, &info));
429: PetscCall(PetscFPTrapPop());
430: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_LIB, "POTRS Bad solve %" PetscBLASInt_FMT, info);
431: if (PetscDefined(USE_COMPLEX) && T) PetscCall(MatConjugate_SeqDense(A));
432: #if defined(PETSC_USE_COMPLEX)
433: } else if (A->hermitian == PETSC_BOOL3_TRUE) {
434: if (T) PetscCall(MatConjugate_SeqDense(A));
435: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
436: PetscCallBLAS("LAPACKhetrs", LAPACKhetrs_("L", &m, &nrhs, mat->v, &mat->lda, mat->pivots, x, &m, &info));
437: PetscCall(PetscFPTrapPop());
438: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_LIB, "HETRS Bad solve %" PetscBLASInt_FMT, info);
439: if (T) PetscCall(MatConjugate_SeqDense(A));
440: #endif
441: } else { /* symmetric case */
442: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
443: PetscCallBLAS("LAPACKsytrs", LAPACKsytrs_("L", &m, &nrhs, mat->v, &mat->lda, mat->pivots, x, &m, &info));
444: PetscCall(PetscFPTrapPop());
445: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_LIB, "SYTRS Bad solve %" PetscBLASInt_FMT, info);
446: }
447: PetscCall(PetscLogFlops(nrhs * (2.0 * m * m - m)));
448: PetscFunctionReturn(PETSC_SUCCESS);
449: }
451: static PetscErrorCode MatSolve_SeqDense_Internal_QR(Mat A, PetscScalar *x, PetscBLASInt ldx, PetscBLASInt m, PetscBLASInt nrhs, PetscBLASInt k)
452: {
453: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
454: PetscBLASInt info;
455: char trans;
457: PetscFunctionBegin;
458: if (PetscDefined(USE_COMPLEX)) {
459: trans = 'C';
460: } else {
461: trans = 'T';
462: }
463: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
464: { /* lwork depends on the number of right-hand sides */
465: PetscBLASInt nlfwork, lfwork = -1;
466: PetscScalar fwork;
468: PetscCallBLAS("LAPACKormqr", LAPACKormqr_("L", &trans, &m, &nrhs, &mat->rank, mat->v, &mat->lda, mat->tau, x, &ldx, &fwork, &lfwork, &info));
469: nlfwork = (PetscBLASInt)PetscRealPart(fwork);
470: if (nlfwork > mat->lfwork) {
471: mat->lfwork = nlfwork;
472: PetscCall(PetscFree(mat->fwork));
473: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
474: }
475: }
476: PetscCallBLAS("LAPACKormqr", LAPACKormqr_("L", &trans, &m, &nrhs, &mat->rank, mat->v, &mat->lda, mat->tau, x, &ldx, mat->fwork, &mat->lfwork, &info));
477: PetscCall(PetscFPTrapPop());
478: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_LIB, "ORMQR - Bad orthogonal transform %" PetscBLASInt_FMT, info);
479: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
480: PetscCallBLAS("LAPACKtrtrs", LAPACKtrtrs_("U", "N", "N", &mat->rank, &nrhs, mat->v, &mat->lda, x, &ldx, &info));
481: PetscCall(PetscFPTrapPop());
482: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_LIB, "TRTRS - Bad triangular solve %" PetscBLASInt_FMT, info);
483: for (PetscInt j = 0; j < nrhs; j++) {
484: for (PetscInt i = mat->rank; i < k; i++) x[j * ldx + i] = 0.;
485: }
486: PetscCall(PetscLogFlops(nrhs * (4.0 * m * mat->rank - PetscSqr(mat->rank))));
487: PetscFunctionReturn(PETSC_SUCCESS);
488: }
490: static PetscErrorCode MatSolveTranspose_SeqDense_Internal_QR(Mat A, PetscScalar *x, PetscBLASInt ldx, PetscBLASInt m, PetscBLASInt nrhs, PetscBLASInt k)
491: {
492: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
493: PetscBLASInt info;
495: PetscFunctionBegin;
496: if (A->rmap->n == A->cmap->n && mat->rank == A->rmap->n) {
497: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
498: PetscCallBLAS("LAPACKtrtrs", LAPACKtrtrs_("U", "T", "N", &m, &nrhs, mat->v, &mat->lda, x, &ldx, &info));
499: PetscCall(PetscFPTrapPop());
500: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_LIB, "TRTRS - Bad triangular solve %" PetscBLASInt_FMT, info);
501: if (PetscDefined(USE_COMPLEX)) PetscCall(MatConjugate_SeqDense(A));
502: { /* lwork depends on the number of right-hand sides */
503: PetscBLASInt nlfwork, lfwork = -1;
504: PetscScalar fwork;
506: PetscCallBLAS("LAPACKormqr", LAPACKormqr_("L", "N", &m, &nrhs, &mat->rank, mat->v, &mat->lda, mat->tau, x, &ldx, &fwork, &lfwork, &info));
507: nlfwork = (PetscBLASInt)PetscRealPart(fwork);
508: if (nlfwork > mat->lfwork) {
509: mat->lfwork = nlfwork;
510: PetscCall(PetscFree(mat->fwork));
511: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
512: }
513: }
514: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
515: PetscCallBLAS("LAPACKormqr", LAPACKormqr_("L", "N", &m, &nrhs, &mat->rank, mat->v, &mat->lda, mat->tau, x, &ldx, mat->fwork, &mat->lfwork, &info));
516: PetscCall(PetscFPTrapPop());
517: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_LIB, "ORMQR - Bad orthogonal transform %" PetscBLASInt_FMT, info);
518: if (PetscDefined(USE_COMPLEX)) PetscCall(MatConjugate_SeqDense(A));
519: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "QR factored matrix cannot be used for transpose solve");
520: PetscCall(PetscLogFlops(nrhs * (4.0 * m * mat->rank - PetscSqr(mat->rank))));
521: PetscFunctionReturn(PETSC_SUCCESS);
522: }
524: static PetscErrorCode MatSolve_SeqDense_SetUp(Mat A, Vec xx, Vec yy, PetscScalar **_y, PetscBLASInt *_m, PetscBLASInt *_k)
525: {
526: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
527: PetscScalar *y;
528: PetscBLASInt m = 0, k = 0;
530: PetscFunctionBegin;
531: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
532: PetscCall(PetscBLASIntCast(A->cmap->n, &k));
533: if (k < m) {
534: PetscCall(VecCopy(xx, mat->qrrhs));
535: PetscCall(VecGetArray(mat->qrrhs, &y));
536: } else {
537: PetscCall(VecCopy(xx, yy));
538: PetscCall(VecGetArray(yy, &y));
539: }
540: *_y = y;
541: *_k = k;
542: *_m = m;
543: PetscFunctionReturn(PETSC_SUCCESS);
544: }
546: static PetscErrorCode MatSolve_SeqDense_TearDown(Mat A, Vec xx, Vec yy, PetscScalar **_y, PetscBLASInt *_m, PetscBLASInt *_k)
547: {
548: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
549: PetscScalar *y = NULL;
550: PetscBLASInt m, k;
552: PetscFunctionBegin;
553: y = *_y;
554: *_y = NULL;
555: k = *_k;
556: m = *_m;
557: if (k < m) {
558: PetscScalar *yv;
559: PetscCall(VecGetArray(yy, &yv));
560: PetscCall(PetscArraycpy(yv, y, k));
561: PetscCall(VecRestoreArray(yy, &yv));
562: PetscCall(VecRestoreArray(mat->qrrhs, &y));
563: } else {
564: PetscCall(VecRestoreArray(yy, &y));
565: }
566: PetscFunctionReturn(PETSC_SUCCESS);
567: }
569: static PetscErrorCode MatSolve_SeqDense_LU(Mat A, Vec xx, Vec yy)
570: {
571: PetscScalar *y = NULL;
572: PetscBLASInt m = 0, k = 0;
574: PetscFunctionBegin;
575: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
576: PetscCall(MatSolve_SeqDense_Internal_LU(A, y, m, m, 1, k, PETSC_FALSE));
577: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
578: PetscFunctionReturn(PETSC_SUCCESS);
579: }
581: static PetscErrorCode MatSolveTranspose_SeqDense_LU(Mat A, Vec xx, Vec yy)
582: {
583: PetscScalar *y = NULL;
584: PetscBLASInt m = 0, k = 0;
586: PetscFunctionBegin;
587: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
588: PetscCall(MatSolve_SeqDense_Internal_LU(A, y, m, m, 1, k, PETSC_TRUE));
589: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
590: PetscFunctionReturn(PETSC_SUCCESS);
591: }
593: static PetscErrorCode MatSolve_SeqDense_Cholesky(Mat A, Vec xx, Vec yy)
594: {
595: PetscScalar *y = NULL;
596: PetscBLASInt m = 0, k = 0;
598: PetscFunctionBegin;
599: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
600: PetscCall(MatSolve_SeqDense_Internal_Cholesky(A, y, m, m, 1, k, PETSC_FALSE));
601: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
602: PetscFunctionReturn(PETSC_SUCCESS);
603: }
605: static PetscErrorCode MatSolveTranspose_SeqDense_Cholesky(Mat A, Vec xx, Vec yy)
606: {
607: PetscScalar *y = NULL;
608: PetscBLASInt m = 0, k = 0;
610: PetscFunctionBegin;
611: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
612: PetscCall(MatSolve_SeqDense_Internal_Cholesky(A, y, m, m, 1, k, PETSC_TRUE));
613: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
614: PetscFunctionReturn(PETSC_SUCCESS);
615: }
617: static PetscErrorCode MatSolve_SeqDense_QR(Mat A, Vec xx, Vec yy)
618: {
619: PetscScalar *y = NULL;
620: PetscBLASInt m = 0, k = 0;
622: PetscFunctionBegin;
623: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
624: PetscCall(MatSolve_SeqDense_Internal_QR(A, y, PetscMax(m, k), m, 1, k));
625: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
626: PetscFunctionReturn(PETSC_SUCCESS);
627: }
629: static PetscErrorCode MatSolveTranspose_SeqDense_QR(Mat A, Vec xx, Vec yy)
630: {
631: PetscScalar *y = NULL;
632: PetscBLASInt m = 0, k = 0;
634: PetscFunctionBegin;
635: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
636: PetscCall(MatSolveTranspose_SeqDense_Internal_QR(A, y, PetscMax(m, k), m, 1, k));
637: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
638: PetscFunctionReturn(PETSC_SUCCESS);
639: }
641: static PetscErrorCode MatMatSolve_SeqDense_SetUp(Mat A, Mat B, Mat X, PetscScalar **_y, PetscBLASInt *_ldy, PetscBLASInt *_m, PetscBLASInt *_nrhs, PetscBLASInt *_k)
642: {
643: const PetscScalar *b;
644: PetscScalar *y;
645: PetscInt n, _ldb, _ldx;
646: PetscBLASInt nrhs = 0, m = 0, k = 0, ldb = 0, ldx = 0, ldy = 0;
648: PetscFunctionBegin;
649: *_ldy = 0;
650: *_m = 0;
651: *_nrhs = 0;
652: *_k = 0;
653: *_y = NULL;
654: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
655: PetscCall(PetscBLASIntCast(A->cmap->n, &k));
656: PetscCall(MatGetSize(B, NULL, &n));
657: PetscCall(PetscBLASIntCast(n, &nrhs));
658: PetscCall(MatDenseGetLDA(B, &_ldb));
659: PetscCall(PetscBLASIntCast(_ldb, &ldb));
660: PetscCall(MatDenseGetLDA(X, &_ldx));
661: PetscCall(PetscBLASIntCast(_ldx, &ldx));
662: if (ldx < m) {
663: PetscCall(MatDenseGetArrayRead(B, &b));
664: PetscCall(PetscMalloc1(nrhs * m, &y));
665: if (ldb == m) {
666: PetscCall(PetscArraycpy(y, b, ldb * nrhs));
667: } else {
668: for (PetscInt j = 0; j < nrhs; j++) PetscCall(PetscArraycpy(&y[j * m], &b[j * ldb], m));
669: }
670: ldy = m;
671: PetscCall(MatDenseRestoreArrayRead(B, &b));
672: } else {
673: if (ldb == ldx) {
674: PetscCall(MatCopy(B, X, SAME_NONZERO_PATTERN));
675: PetscCall(MatDenseGetArray(X, &y));
676: } else {
677: PetscCall(MatDenseGetArray(X, &y));
678: PetscCall(MatDenseGetArrayRead(B, &b));
679: for (PetscInt j = 0; j < nrhs; j++) PetscCall(PetscArraycpy(&y[j * ldx], &b[j * ldb], m));
680: PetscCall(MatDenseRestoreArrayRead(B, &b));
681: }
682: ldy = ldx;
683: }
684: *_y = y;
685: *_ldy = ldy;
686: *_k = k;
687: *_m = m;
688: *_nrhs = nrhs;
689: PetscFunctionReturn(PETSC_SUCCESS);
690: }
692: static PetscErrorCode MatMatSolve_SeqDense_TearDown(Mat A, Mat B, Mat X, PetscScalar **_y, PetscBLASInt *_ldy, PetscBLASInt *_m, PetscBLASInt *_nrhs, PetscBLASInt *_k)
693: {
694: PetscScalar *y;
695: PetscInt _ldx;
696: PetscBLASInt k, ldy, nrhs, ldx = 0;
698: PetscFunctionBegin;
699: y = *_y;
700: *_y = NULL;
701: k = *_k;
702: ldy = *_ldy;
703: nrhs = *_nrhs;
704: PetscCall(MatDenseGetLDA(X, &_ldx));
705: PetscCall(PetscBLASIntCast(_ldx, &ldx));
706: if (ldx != ldy) {
707: PetscScalar *xv;
708: PetscCall(MatDenseGetArray(X, &xv));
709: for (PetscInt j = 0; j < nrhs; j++) PetscCall(PetscArraycpy(&xv[j * ldx], &y[j * ldy], k));
710: PetscCall(MatDenseRestoreArray(X, &xv));
711: PetscCall(PetscFree(y));
712: } else {
713: PetscCall(MatDenseRestoreArray(X, &y));
714: }
715: PetscFunctionReturn(PETSC_SUCCESS);
716: }
718: static PetscErrorCode MatMatSolve_SeqDense_LU(Mat A, Mat B, Mat X)
719: {
720: PetscScalar *y;
721: PetscBLASInt m, k, ldy, nrhs;
723: PetscFunctionBegin;
724: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
725: PetscCall(MatSolve_SeqDense_Internal_LU(A, y, ldy, m, nrhs, k, PETSC_FALSE));
726: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
727: PetscFunctionReturn(PETSC_SUCCESS);
728: }
730: static PetscErrorCode MatMatSolveTranspose_SeqDense_LU(Mat A, Mat B, Mat X)
731: {
732: PetscScalar *y;
733: PetscBLASInt m, k, ldy, nrhs;
735: PetscFunctionBegin;
736: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
737: PetscCall(MatSolve_SeqDense_Internal_LU(A, y, ldy, m, nrhs, k, PETSC_TRUE));
738: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
739: PetscFunctionReturn(PETSC_SUCCESS);
740: }
742: static PetscErrorCode MatMatSolve_SeqDense_Cholesky(Mat A, Mat B, Mat X)
743: {
744: PetscScalar *y;
745: PetscBLASInt m, k, ldy, nrhs;
747: PetscFunctionBegin;
748: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
749: PetscCall(MatSolve_SeqDense_Internal_Cholesky(A, y, ldy, m, nrhs, k, PETSC_FALSE));
750: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
751: PetscFunctionReturn(PETSC_SUCCESS);
752: }
754: static PetscErrorCode MatMatSolveTranspose_SeqDense_Cholesky(Mat A, Mat B, Mat X)
755: {
756: PetscScalar *y;
757: PetscBLASInt m, k, ldy, nrhs;
759: PetscFunctionBegin;
760: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
761: PetscCall(MatSolve_SeqDense_Internal_Cholesky(A, y, ldy, m, nrhs, k, PETSC_TRUE));
762: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
763: PetscFunctionReturn(PETSC_SUCCESS);
764: }
766: static PetscErrorCode MatMatSolve_SeqDense_QR(Mat A, Mat B, Mat X)
767: {
768: PetscScalar *y;
769: PetscBLASInt m, k, ldy, nrhs;
771: PetscFunctionBegin;
772: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
773: PetscCall(MatSolve_SeqDense_Internal_QR(A, y, ldy, m, nrhs, k));
774: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
775: PetscFunctionReturn(PETSC_SUCCESS);
776: }
778: static PetscErrorCode MatMatSolveTranspose_SeqDense_QR(Mat A, Mat B, Mat X)
779: {
780: PetscScalar *y;
781: PetscBLASInt m, k, ldy, nrhs;
783: PetscFunctionBegin;
784: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
785: PetscCall(MatSolveTranspose_SeqDense_Internal_QR(A, y, ldy, m, nrhs, k));
786: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
787: PetscFunctionReturn(PETSC_SUCCESS);
788: }
790: /* COMMENT: I have chosen to hide row permutation in the pivots,
791: rather than put it in the Mat->row slot.*/
792: PetscErrorCode MatLUFactor_SeqDense(Mat A, IS row, IS col, const MatFactorInfo *minfo)
793: {
794: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
795: PetscBLASInt n, m, info;
797: PetscFunctionBegin;
798: PetscCall(PetscBLASIntCast(A->cmap->n, &n));
799: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
800: if (!mat->pivots) { PetscCall(PetscMalloc1(A->rmap->n, &mat->pivots)); }
801: if (!A->rmap->n || !A->cmap->n) PetscFunctionReturn(PETSC_SUCCESS);
802: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
803: PetscCallBLAS("LAPACKgetrf", LAPACKgetrf_(&m, &n, mat->v, &mat->lda, mat->pivots, &info));
804: PetscCall(PetscFPTrapPop());
806: PetscCheck(info >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Bad argument to LU factorization %" PetscBLASInt_FMT, info);
807: PetscCheck(info <= 0, PETSC_COMM_SELF, PETSC_ERR_MAT_LU_ZRPVT, "Bad LU factorization %" PetscBLASInt_FMT, info);
809: A->ops->solve = MatSolve_SeqDense_LU;
810: A->ops->matsolve = MatMatSolve_SeqDense_LU;
811: A->ops->solvetranspose = MatSolveTranspose_SeqDense_LU;
812: A->ops->matsolvetranspose = MatMatSolveTranspose_SeqDense_LU;
813: A->factortype = MAT_FACTOR_LU;
815: PetscCall(PetscFree(A->solvertype));
816: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &A->solvertype));
818: PetscCall(PetscLogFlops((2.0 * A->cmap->n * A->cmap->n * A->cmap->n) / 3));
819: PetscFunctionReturn(PETSC_SUCCESS);
820: }
822: static PetscErrorCode MatLUFactorNumeric_SeqDense(Mat fact, Mat A, const MatFactorInfo *info_dummy)
823: {
824: MatFactorInfo info;
826: PetscFunctionBegin;
827: PetscCall(MatDuplicateNoCreate_SeqDense(fact, A, MAT_COPY_VALUES));
828: PetscUseTypeMethod(fact, lufactor, NULL, NULL, &info);
829: PetscFunctionReturn(PETSC_SUCCESS);
830: }
832: PetscErrorCode MatLUFactorSymbolic_SeqDense(Mat fact, Mat A, IS row, IS col, const MatFactorInfo *info)
833: {
834: PetscFunctionBegin;
835: fact->preallocated = PETSC_TRUE;
836: fact->assembled = PETSC_TRUE;
837: fact->ops->lufactornumeric = MatLUFactorNumeric_SeqDense;
838: PetscFunctionReturn(PETSC_SUCCESS);
839: }
841: /* Cholesky as L*L^T or L*D*L^T and the symmetric/hermitian complex variants */
842: PetscErrorCode MatCholeskyFactor_SeqDense(Mat A, IS perm, const MatFactorInfo *factinfo)
843: {
844: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
845: PetscBLASInt info, n;
847: PetscFunctionBegin;
848: PetscCall(PetscBLASIntCast(A->cmap->n, &n));
849: if (!A->rmap->n || !A->cmap->n) PetscFunctionReturn(PETSC_SUCCESS);
850: if (A->spd == PETSC_BOOL3_TRUE) {
851: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
852: PetscCallBLAS("LAPACKpotrf", LAPACKpotrf_("L", &n, mat->v, &mat->lda, &info));
853: PetscCall(PetscFPTrapPop());
854: #if defined(PETSC_USE_COMPLEX)
855: } else if (A->hermitian == PETSC_BOOL3_TRUE) {
856: if (!mat->pivots) { PetscCall(PetscMalloc1(A->rmap->n, &mat->pivots)); }
857: if (!mat->fwork) {
858: PetscScalar dummy;
860: mat->lfwork = -1;
861: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
862: PetscCallBLAS("LAPACKhetrf", LAPACKhetrf_("L", &n, mat->v, &mat->lda, mat->pivots, &dummy, &mat->lfwork, &info));
863: PetscCall(PetscFPTrapPop());
864: PetscCall(PetscBLASIntCast((PetscCount)(PetscRealPart(dummy)), &mat->lfwork));
865: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
866: }
867: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
868: PetscCallBLAS("LAPACKhetrf", LAPACKhetrf_("L", &n, mat->v, &mat->lda, mat->pivots, mat->fwork, &mat->lfwork, &info));
869: PetscCall(PetscFPTrapPop());
870: #endif
871: } else { /* symmetric case */
872: if (!mat->pivots) { PetscCall(PetscMalloc1(A->rmap->n, &mat->pivots)); }
873: if (!mat->fwork) {
874: PetscScalar dummy;
876: mat->lfwork = -1;
877: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
878: PetscCallBLAS("LAPACKsytrf", LAPACKsytrf_("L", &n, mat->v, &mat->lda, mat->pivots, &dummy, &mat->lfwork, &info));
879: PetscCall(PetscFPTrapPop());
880: PetscCall(PetscBLASIntCast((PetscCount)(PetscRealPart(dummy)), &mat->lfwork));
881: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
882: }
883: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
884: PetscCallBLAS("LAPACKsytrf", LAPACKsytrf_("L", &n, mat->v, &mat->lda, mat->pivots, mat->fwork, &mat->lfwork, &info));
885: PetscCall(PetscFPTrapPop());
886: }
887: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_MAT_CH_ZRPVT, "Bad factorization: zero pivot in row %" PetscBLASInt_FMT, info - 1);
889: A->ops->solve = MatSolve_SeqDense_Cholesky;
890: A->ops->matsolve = MatMatSolve_SeqDense_Cholesky;
891: A->ops->solvetranspose = MatSolveTranspose_SeqDense_Cholesky;
892: A->ops->matsolvetranspose = MatMatSolveTranspose_SeqDense_Cholesky;
893: A->factortype = MAT_FACTOR_CHOLESKY;
895: PetscCall(PetscFree(A->solvertype));
896: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &A->solvertype));
898: PetscCall(PetscLogFlops((1.0 * A->cmap->n * A->cmap->n * A->cmap->n) / 3.0));
899: PetscFunctionReturn(PETSC_SUCCESS);
900: }
902: static PetscErrorCode MatCholeskyFactorNumeric_SeqDense(Mat fact, Mat A, const MatFactorInfo *info_dummy)
903: {
904: MatFactorInfo info;
906: PetscFunctionBegin;
907: info.fill = 1.0;
909: PetscCall(MatDuplicateNoCreate_SeqDense(fact, A, MAT_COPY_VALUES));
910: PetscUseTypeMethod(fact, choleskyfactor, NULL, &info);
911: PetscFunctionReturn(PETSC_SUCCESS);
912: }
914: PetscErrorCode MatCholeskyFactorSymbolic_SeqDense(Mat fact, Mat A, IS row, const MatFactorInfo *info)
915: {
916: PetscFunctionBegin;
917: fact->assembled = PETSC_TRUE;
918: fact->preallocated = PETSC_TRUE;
919: fact->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqDense;
920: PetscFunctionReturn(PETSC_SUCCESS);
921: }
923: PetscErrorCode MatQRFactor_SeqDense(Mat A, IS col, const MatFactorInfo *minfo)
924: {
925: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
926: PetscBLASInt n, m, info, min, max;
928: PetscFunctionBegin;
929: PetscCall(PetscBLASIntCast(A->cmap->n, &n));
930: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
931: max = PetscMax(m, n);
932: min = PetscMin(m, n);
933: if (!mat->tau) { PetscCall(PetscMalloc1(min, &mat->tau)); }
934: if (!mat->pivots) { PetscCall(PetscMalloc1(n, &mat->pivots)); }
935: if (!mat->qrrhs) PetscCall(MatCreateVecs(A, NULL, &mat->qrrhs));
936: if (!A->rmap->n || !A->cmap->n) PetscFunctionReturn(PETSC_SUCCESS);
937: if (!mat->fwork) {
938: PetscScalar dummy;
940: mat->lfwork = -1;
941: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
942: PetscCallBLAS("LAPACKgeqrf", LAPACKgeqrf_(&m, &n, mat->v, &mat->lda, mat->tau, &dummy, &mat->lfwork, &info));
943: PetscCall(PetscFPTrapPop());
944: PetscCall(PetscBLASIntCast((PetscCount)(PetscRealPart(dummy)), &mat->lfwork));
945: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
946: }
947: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
948: PetscCallBLAS("LAPACKgeqrf", LAPACKgeqrf_(&m, &n, mat->v, &mat->lda, mat->tau, mat->fwork, &mat->lfwork, &info));
949: PetscCall(PetscFPTrapPop());
950: PetscCheck(!info, PETSC_COMM_SELF, PETSC_ERR_LIB, "Bad argument to QR factorization %" PetscBLASInt_FMT, info);
951: // TODO: try to estimate rank or test for and use geqp3 for rank revealing QR. For now just say rank is min of m and n
952: mat->rank = min;
954: A->ops->solve = MatSolve_SeqDense_QR;
955: A->ops->matsolve = MatMatSolve_SeqDense_QR;
956: A->factortype = MAT_FACTOR_QR;
957: if (m == n) {
958: A->ops->solvetranspose = MatSolveTranspose_SeqDense_QR;
959: A->ops->matsolvetranspose = MatMatSolveTranspose_SeqDense_QR;
960: }
962: PetscCall(PetscFree(A->solvertype));
963: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &A->solvertype));
965: PetscCall(PetscLogFlops(2.0 * min * min * (max - min / 3.0)));
966: PetscFunctionReturn(PETSC_SUCCESS);
967: }
969: static PetscErrorCode MatQRFactorNumeric_SeqDense(Mat fact, Mat A, const MatFactorInfo *info_dummy)
970: {
971: MatFactorInfo info;
973: PetscFunctionBegin;
974: info.fill = 1.0;
976: PetscCall(MatDuplicateNoCreate_SeqDense(fact, A, MAT_COPY_VALUES));
977: PetscUseMethod(fact, "MatQRFactor_C", (Mat, IS, const MatFactorInfo *), (fact, NULL, &info));
978: PetscFunctionReturn(PETSC_SUCCESS);
979: }
981: PetscErrorCode MatQRFactorSymbolic_SeqDense(Mat fact, Mat A, IS row, const MatFactorInfo *info)
982: {
983: PetscFunctionBegin;
984: fact->assembled = PETSC_TRUE;
985: fact->preallocated = PETSC_TRUE;
986: PetscCall(PetscObjectComposeFunction((PetscObject)fact, "MatQRFactorNumeric_C", MatQRFactorNumeric_SeqDense));
987: PetscFunctionReturn(PETSC_SUCCESS);
988: }
990: /* uses LAPACK */
991: PETSC_INTERN PetscErrorCode MatGetFactor_seqdense_petsc(Mat A, MatFactorType ftype, Mat *fact)
992: {
993: PetscFunctionBegin;
994: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), fact));
995: PetscCall(MatSetSizes(*fact, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n));
996: PetscCall(MatSetType(*fact, MATDENSE));
997: (*fact)->trivialsymbolic = PETSC_TRUE;
998: if (ftype == MAT_FACTOR_LU || ftype == MAT_FACTOR_ILU) {
999: (*fact)->ops->lufactorsymbolic = MatLUFactorSymbolic_SeqDense;
1000: (*fact)->ops->ilufactorsymbolic = MatLUFactorSymbolic_SeqDense;
1001: } else if (ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC) {
1002: (*fact)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqDense;
1003: } else if (ftype == MAT_FACTOR_QR) {
1004: PetscCall(PetscObjectComposeFunction((PetscObject)*fact, "MatQRFactorSymbolic_C", MatQRFactorSymbolic_SeqDense));
1005: }
1006: (*fact)->factortype = ftype;
1008: PetscCall(PetscFree((*fact)->solvertype));
1009: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &(*fact)->solvertype));
1010: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&(*fact)->preferredordering[MAT_FACTOR_LU]));
1011: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&(*fact)->preferredordering[MAT_FACTOR_ILU]));
1012: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&(*fact)->preferredordering[MAT_FACTOR_CHOLESKY]));
1013: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&(*fact)->preferredordering[MAT_FACTOR_ICC]));
1014: PetscFunctionReturn(PETSC_SUCCESS);
1015: }
1017: static PetscErrorCode MatSOR_SeqDense(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal shift, PetscInt its, PetscInt lits, Vec xx)
1018: {
1019: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1020: PetscScalar *x, *v = mat->v, zero = 0.0, xt;
1021: const PetscScalar *b;
1022: PetscInt m = A->rmap->n, i;
1023: PetscBLASInt o = 1, bm = 0;
1025: PetscFunctionBegin;
1026: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1027: PetscCheck(A->offloadmask != PETSC_OFFLOAD_GPU, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
1028: #endif
1029: if (shift == -1) shift = 0.0; /* negative shift indicates do not error on zero diagonal; this code never zeros on zero diagonal */
1030: PetscCall(PetscBLASIntCast(m, &bm));
1031: if (flag & SOR_ZERO_INITIAL_GUESS) {
1032: /* this is a hack fix, should have another version without the second BLASdotu */
1033: PetscCall(VecSet(xx, zero));
1034: }
1035: PetscCall(VecGetArray(xx, &x));
1036: PetscCall(VecGetArrayRead(bb, &b));
1037: its = its * lits;
1038: PetscCheck(its > 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Relaxation requires global its %" PetscInt_FMT " and local its %" PetscInt_FMT " both positive", its, lits);
1039: while (its--) {
1040: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1041: for (i = 0; i < m; i++) {
1042: PetscCallBLAS("BLASdotu", xt = b[i] - BLASdotu_(&bm, v + i, &bm, x, &o));
1043: x[i] = (1. - omega) * x[i] + omega * (xt + v[i + i * m] * x[i]) / (v[i + i * m] + shift);
1044: }
1045: }
1046: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1047: for (i = m - 1; i >= 0; i--) {
1048: PetscCallBLAS("BLASdotu", xt = b[i] - BLASdotu_(&bm, v + i, &bm, x, &o));
1049: x[i] = (1. - omega) * x[i] + omega * (xt + v[i + i * m] * x[i]) / (v[i + i * m] + shift);
1050: }
1051: }
1052: }
1053: PetscCall(VecRestoreArrayRead(bb, &b));
1054: PetscCall(VecRestoreArray(xx, &x));
1055: PetscFunctionReturn(PETSC_SUCCESS);
1056: }
1058: static PetscErrorCode MatMultColumnRangeKernel_SeqDense(Mat A, Vec xx, Vec yy, PetscInt c_start, PetscInt c_end, PetscBool trans, PetscBool herm)
1059: {
1060: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1061: PetscScalar *y, _DOne = 1.0, _DZero = 0.0;
1062: PetscBLASInt m, n, _One = 1;
1063: const PetscScalar *v = mat->v, *x;
1065: PetscFunctionBegin;
1066: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
1067: PetscCall(PetscBLASIntCast(c_end - c_start, &n));
1068: PetscCall(VecGetArrayRead(xx, &x));
1069: PetscCall(VecGetArrayWrite(yy, &y));
1070: if (!m || !n) {
1071: PetscBLASInt i;
1072: if (trans)
1073: for (i = 0; i < n; i++) y[i] = 0.0;
1074: else
1075: for (i = 0; i < m; i++) y[i] = 0.0;
1076: } else {
1077: if (trans) {
1078: if (herm) PetscCallBLAS("BLASgemv", BLASgemv_("C", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x, &_One, &_DZero, y + c_start, &_One));
1079: else PetscCallBLAS("BLASgemv", BLASgemv_("T", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x, &_One, &_DZero, y + c_start, &_One));
1080: } else {
1081: PetscCallBLAS("BLASgemv", BLASgemv_("N", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x + c_start, &_One, &_DZero, y, &_One));
1082: }
1083: PetscCall(PetscLogFlops(2.0 * m * n - n));
1084: }
1085: PetscCall(VecRestoreArrayRead(xx, &x));
1086: PetscCall(VecRestoreArrayWrite(yy, &y));
1087: PetscFunctionReturn(PETSC_SUCCESS);
1088: }
1090: PetscErrorCode MatMultHermitianTransposeColumnRange_SeqDense(Mat A, Vec xx, Vec yy, PetscInt c_start, PetscInt c_end)
1091: {
1092: PetscFunctionBegin;
1093: PetscCall(MatMultColumnRangeKernel_SeqDense(A, xx, yy, c_start, c_end, PETSC_TRUE, PETSC_TRUE));
1094: PetscFunctionReturn(PETSC_SUCCESS);
1095: }
1097: PetscErrorCode MatMult_SeqDense(Mat A, Vec xx, Vec yy)
1098: {
1099: PetscFunctionBegin;
1100: PetscCall(MatMultColumnRangeKernel_SeqDense(A, xx, yy, 0, A->cmap->n, PETSC_FALSE, PETSC_FALSE));
1101: PetscFunctionReturn(PETSC_SUCCESS);
1102: }
1104: PetscErrorCode MatMultTranspose_SeqDense(Mat A, Vec xx, Vec yy)
1105: {
1106: PetscFunctionBegin;
1107: PetscCall(MatMultColumnRangeKernel_SeqDense(A, xx, yy, 0, A->cmap->n, PETSC_TRUE, PETSC_FALSE));
1108: PetscFunctionReturn(PETSC_SUCCESS);
1109: }
1111: PetscErrorCode MatMultHermitianTranspose_SeqDense(Mat A, Vec xx, Vec yy)
1112: {
1113: PetscFunctionBegin;
1114: PetscCall(MatMultColumnRangeKernel_SeqDense(A, xx, yy, 0, A->cmap->n, PETSC_TRUE, PETSC_TRUE));
1115: PetscFunctionReturn(PETSC_SUCCESS);
1116: }
1118: static PetscErrorCode MatMultAddColumnRangeKernel_SeqDense(Mat A, Vec xx, Vec zz, Vec yy, PetscInt c_start, PetscInt c_end, PetscBool trans, PetscBool herm)
1119: {
1120: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1121: const PetscScalar *v = mat->v, *x;
1122: PetscScalar *y, _DOne = 1.0;
1123: PetscBLASInt m, n, _One = 1;
1125: PetscFunctionBegin;
1126: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
1127: PetscCall(PetscBLASIntCast(c_end - c_start, &n));
1128: PetscCall(VecCopy(zz, yy));
1129: if (!m || !n) PetscFunctionReturn(PETSC_SUCCESS);
1130: PetscCall(VecGetArray(yy, &y));
1131: PetscCall(VecGetArrayRead(xx, &x));
1132: if (trans) {
1133: if (herm) PetscCallBLAS("BLASgemv", BLASgemv_("C", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x, &_One, &_DOne, y + c_start, &_One));
1134: else PetscCallBLAS("BLASgemv", BLASgemv_("T", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x, &_One, &_DOne, y + c_start, &_One));
1135: } else {
1136: PetscCallBLAS("BLASgemv", BLASgemv_("N", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x + c_start, &_One, &_DOne, y, &_One));
1137: }
1138: PetscCall(VecRestoreArrayRead(xx, &x));
1139: PetscCall(VecRestoreArray(yy, &y));
1140: PetscCall(PetscLogFlops(2.0 * m * n));
1141: PetscFunctionReturn(PETSC_SUCCESS);
1142: }
1144: PetscErrorCode MatMultAddColumnRange_SeqDense(Mat A, Vec xx, Vec zz, Vec yy, PetscInt c_start, PetscInt c_end)
1145: {
1146: PetscFunctionBegin;
1147: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, c_start, c_end, PETSC_FALSE, PETSC_FALSE));
1148: PetscFunctionReturn(PETSC_SUCCESS);
1149: }
1151: PetscErrorCode MatMultHermitianTransposeAddColumnRange_SeqDense(Mat A, Vec xx, Vec zz, Vec yy, PetscInt c_start, PetscInt c_end)
1152: {
1153: PetscFunctionBegin;
1154: PetscMPIInt rank;
1155: PetscCallMPI(MPI_Comm_rank(MPI_COMM_WORLD, &rank));
1156: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, c_start, c_end, PETSC_TRUE, PETSC_TRUE));
1157: PetscFunctionReturn(PETSC_SUCCESS);
1158: }
1160: PetscErrorCode MatMultAdd_SeqDense(Mat A, Vec xx, Vec zz, Vec yy)
1161: {
1162: PetscFunctionBegin;
1163: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, 0, A->cmap->n, PETSC_FALSE, PETSC_FALSE));
1164: PetscFunctionReturn(PETSC_SUCCESS);
1165: }
1167: PetscErrorCode MatMultTransposeAdd_SeqDense(Mat A, Vec xx, Vec zz, Vec yy)
1168: {
1169: PetscFunctionBegin;
1170: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, 0, A->cmap->n, PETSC_TRUE, PETSC_FALSE));
1171: PetscFunctionReturn(PETSC_SUCCESS);
1172: }
1174: PetscErrorCode MatMultHermitianTransposeAdd_SeqDense(Mat A, Vec xx, Vec zz, Vec yy)
1175: {
1176: PetscFunctionBegin;
1177: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, 0, A->cmap->n, PETSC_TRUE, PETSC_TRUE));
1178: PetscFunctionReturn(PETSC_SUCCESS);
1179: }
1181: static PetscErrorCode MatGetRow_SeqDense(Mat A, PetscInt row, PetscInt *ncols, PetscInt **cols, PetscScalar **vals)
1182: {
1183: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1184: PetscInt i;
1186: PetscFunctionBegin;
1187: if (ncols) *ncols = A->cmap->n;
1188: if (cols) {
1189: PetscCall(PetscMalloc1(A->cmap->n, cols));
1190: for (i = 0; i < A->cmap->n; i++) (*cols)[i] = i;
1191: }
1192: if (vals) {
1193: const PetscScalar *v;
1195: PetscCall(MatDenseGetArrayRead(A, &v));
1196: PetscCall(PetscMalloc1(A->cmap->n, vals));
1197: v += row;
1198: for (i = 0; i < A->cmap->n; i++) {
1199: (*vals)[i] = *v;
1200: v += mat->lda;
1201: }
1202: PetscCall(MatDenseRestoreArrayRead(A, &v));
1203: }
1204: PetscFunctionReturn(PETSC_SUCCESS);
1205: }
1207: static PetscErrorCode MatRestoreRow_SeqDense(Mat A, PetscInt row, PetscInt *ncols, PetscInt **cols, PetscScalar **vals)
1208: {
1209: PetscFunctionBegin;
1210: if (cols) PetscCall(PetscFree(*cols));
1211: if (vals) PetscCall(PetscFree(*vals));
1212: PetscFunctionReturn(PETSC_SUCCESS);
1213: }
1215: static PetscErrorCode MatSetValues_SeqDense(Mat A, PetscInt m, const PetscInt indexm[], PetscInt n, const PetscInt indexn[], const PetscScalar v[], InsertMode addv)
1216: {
1217: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1218: PetscScalar *av;
1219: PetscInt i, j, idx = 0;
1220: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1221: PetscOffloadMask oldf;
1222: #endif
1224: PetscFunctionBegin;
1225: PetscCall(MatDenseGetArray(A, &av));
1226: if (!mat->roworiented) {
1227: if (addv == INSERT_VALUES) {
1228: for (j = 0; j < n; j++) {
1229: if (indexn[j] < 0) {
1230: idx += m;
1231: continue;
1232: }
1233: PetscCheck(indexn[j] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, indexn[j], A->cmap->n - 1);
1234: for (i = 0; i < m; i++) {
1235: if (indexm[i] < 0) {
1236: idx++;
1237: continue;
1238: }
1239: PetscCheck(indexm[i] < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, indexm[i], A->rmap->n - 1);
1240: av[indexn[j] * mat->lda + indexm[i]] = v ? v[idx++] : (idx++, 0.0);
1241: }
1242: }
1243: } else {
1244: for (j = 0; j < n; j++) {
1245: if (indexn[j] < 0) {
1246: idx += m;
1247: continue;
1248: }
1249: PetscCheck(indexn[j] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, indexn[j], A->cmap->n - 1);
1250: for (i = 0; i < m; i++) {
1251: if (indexm[i] < 0) {
1252: idx++;
1253: continue;
1254: }
1255: PetscCheck(indexm[i] < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, indexm[i], A->rmap->n - 1);
1256: av[indexn[j] * mat->lda + indexm[i]] += v ? v[idx++] : (idx++, 0.0);
1257: }
1258: }
1259: }
1260: } else {
1261: if (addv == INSERT_VALUES) {
1262: for (i = 0; i < m; i++) {
1263: if (indexm[i] < 0) {
1264: idx += n;
1265: continue;
1266: }
1267: PetscCheck(indexm[i] < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, indexm[i], A->rmap->n - 1);
1268: for (j = 0; j < n; j++) {
1269: if (indexn[j] < 0) {
1270: idx++;
1271: continue;
1272: }
1273: PetscCheck(indexn[j] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, indexn[j], A->cmap->n - 1);
1274: av[indexn[j] * mat->lda + indexm[i]] = v ? v[idx++] : (idx++, 0.0);
1275: }
1276: }
1277: } else {
1278: for (i = 0; i < m; i++) {
1279: if (indexm[i] < 0) {
1280: idx += n;
1281: continue;
1282: }
1283: PetscCheck(indexm[i] < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, indexm[i], A->rmap->n - 1);
1284: for (j = 0; j < n; j++) {
1285: if (indexn[j] < 0) {
1286: idx++;
1287: continue;
1288: }
1289: PetscCheck(indexn[j] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, indexn[j], A->cmap->n - 1);
1290: av[indexn[j] * mat->lda + indexm[i]] += v ? v[idx++] : (idx++, 0.0);
1291: }
1292: }
1293: }
1294: }
1295: /* hack to prevent unneeded copy to the GPU while returning the array */
1296: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1297: oldf = A->offloadmask;
1298: A->offloadmask = PETSC_OFFLOAD_GPU;
1299: #endif
1300: PetscCall(MatDenseRestoreArray(A, &av));
1301: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1302: A->offloadmask = (oldf == PETSC_OFFLOAD_UNALLOCATED ? PETSC_OFFLOAD_UNALLOCATED : PETSC_OFFLOAD_CPU);
1303: #endif
1304: PetscFunctionReturn(PETSC_SUCCESS);
1305: }
1307: static PetscErrorCode MatGetValues_SeqDense(Mat A, PetscInt m, const PetscInt indexm[], PetscInt n, const PetscInt indexn[], PetscScalar v[])
1308: {
1309: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1310: const PetscScalar *vv;
1311: PetscInt i, j;
1313: PetscFunctionBegin;
1314: PetscCall(MatDenseGetArrayRead(A, &vv));
1315: /* row-oriented output */
1316: for (i = 0; i < m; i++) {
1317: if (indexm[i] < 0) {
1318: v += n;
1319: continue;
1320: }
1321: PetscCheck(indexm[i] < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %" PetscInt_FMT " requested larger than number rows %" PetscInt_FMT, indexm[i], A->rmap->n);
1322: for (j = 0; j < n; j++) {
1323: if (indexn[j] < 0) {
1324: v++;
1325: continue;
1326: }
1327: PetscCheck(indexn[j] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column %" PetscInt_FMT " requested larger than number columns %" PetscInt_FMT, indexn[j], A->cmap->n);
1328: *v++ = vv[indexn[j] * mat->lda + indexm[i]];
1329: }
1330: }
1331: PetscCall(MatDenseRestoreArrayRead(A, &vv));
1332: PetscFunctionReturn(PETSC_SUCCESS);
1333: }
1335: PetscErrorCode MatView_Dense_Binary(Mat mat, PetscViewer viewer)
1336: {
1337: PetscBool skipHeader;
1338: PetscViewerFormat format;
1339: PetscInt header[4], M, N, m, lda, i, j;
1340: PetscCount k;
1341: const PetscScalar *v;
1342: PetscScalar *vwork;
1344: PetscFunctionBegin;
1345: PetscCall(PetscViewerSetUp(viewer));
1346: PetscCall(PetscViewerBinaryGetSkipHeader(viewer, &skipHeader));
1347: PetscCall(PetscViewerGetFormat(viewer, &format));
1348: if (skipHeader) format = PETSC_VIEWER_NATIVE;
1350: PetscCall(MatGetSize(mat, &M, &N));
1352: /* write matrix header */
1353: header[0] = MAT_FILE_CLASSID;
1354: header[1] = M;
1355: header[2] = N;
1356: header[3] = (format == PETSC_VIEWER_NATIVE) ? MATRIX_BINARY_FORMAT_DENSE : M * N;
1357: if (!skipHeader) PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1359: PetscCall(MatGetLocalSize(mat, &m, NULL));
1360: if (format != PETSC_VIEWER_NATIVE) {
1361: PetscInt nnz = m * N, *iwork;
1362: /* store row lengths for each row */
1363: PetscCall(PetscMalloc1(nnz, &iwork));
1364: for (i = 0; i < m; i++) iwork[i] = N;
1365: PetscCall(PetscViewerBinaryWriteAll(viewer, iwork, m, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1366: /* store column indices (zero start index) */
1367: for (k = 0, i = 0; i < m; i++)
1368: for (j = 0; j < N; j++, k++) iwork[k] = j;
1369: PetscCall(PetscViewerBinaryWriteAll(viewer, iwork, nnz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1370: PetscCall(PetscFree(iwork));
1371: }
1372: /* store matrix values as a dense matrix in row major order */
1373: PetscCall(PetscMalloc1(m * N, &vwork));
1374: PetscCall(MatDenseGetArrayRead(mat, &v));
1375: PetscCall(MatDenseGetLDA(mat, &lda));
1376: for (k = 0, i = 0; i < m; i++)
1377: for (j = 0; j < N; j++, k++) vwork[k] = v[i + (size_t)lda * j];
1378: PetscCall(MatDenseRestoreArrayRead(mat, &v));
1379: PetscCall(PetscViewerBinaryWriteAll(viewer, vwork, m * N, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1380: PetscCall(PetscFree(vwork));
1381: PetscFunctionReturn(PETSC_SUCCESS);
1382: }
1384: PetscErrorCode MatLoad_Dense_Binary(Mat mat, PetscViewer viewer)
1385: {
1386: PetscBool skipHeader;
1387: PetscInt header[4], M, N, m, nz, lda, i, j, k;
1388: PetscInt rows, cols;
1389: PetscScalar *v, *vwork;
1391: PetscFunctionBegin;
1392: PetscCall(PetscViewerSetUp(viewer));
1393: PetscCall(PetscViewerBinaryGetSkipHeader(viewer, &skipHeader));
1395: if (!skipHeader) {
1396: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
1397: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
1398: M = header[1];
1399: N = header[2];
1400: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
1401: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
1402: nz = header[3];
1403: PetscCheck(nz == MATRIX_BINARY_FORMAT_DENSE || nz >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Unknown matrix format %" PetscInt_FMT " in file", nz);
1404: } else {
1405: PetscCall(MatGetSize(mat, &M, &N));
1406: PetscCheck(M >= 0 && N >= 0, PETSC_COMM_SELF, PETSC_ERR_USER, "Matrix binary file header was skipped, thus the user must specify the global sizes of input matrix");
1407: nz = MATRIX_BINARY_FORMAT_DENSE;
1408: }
1410: /* setup global sizes if not set */
1411: if (mat->rmap->N < 0) mat->rmap->N = M;
1412: if (mat->cmap->N < 0) mat->cmap->N = N;
1413: PetscCall(MatSetUp(mat));
1414: /* check if global sizes are correct */
1415: PetscCall(MatGetSize(mat, &rows, &cols));
1416: PetscCheck(M == rows && N == cols, PetscObjectComm((PetscObject)viewer), 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);
1418: PetscCall(MatGetSize(mat, NULL, &N));
1419: PetscCall(MatGetLocalSize(mat, &m, NULL));
1420: PetscCall(MatDenseGetArray(mat, &v));
1421: PetscCall(MatDenseGetLDA(mat, &lda));
1422: if (nz == MATRIX_BINARY_FORMAT_DENSE) { /* matrix in file is dense format */
1423: PetscCount nnz = (size_t)m * N;
1424: /* read in matrix values */
1425: PetscCall(PetscMalloc1(nnz, &vwork));
1426: PetscCall(PetscViewerBinaryReadAll(viewer, vwork, nnz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1427: /* store values in column major order */
1428: for (j = 0; j < N; j++)
1429: for (i = 0; i < m; i++) v[i + (size_t)lda * j] = vwork[(size_t)i * N + j];
1430: PetscCall(PetscFree(vwork));
1431: } else { /* matrix in file is sparse format */
1432: PetscInt nnz = 0, *rlens, *icols;
1433: /* read in row lengths */
1434: PetscCall(PetscMalloc1(m, &rlens));
1435: PetscCall(PetscViewerBinaryReadAll(viewer, rlens, m, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1436: for (i = 0; i < m; i++) nnz += rlens[i];
1437: /* read in column indices and values */
1438: PetscCall(PetscMalloc2(nnz, &icols, nnz, &vwork));
1439: PetscCall(PetscViewerBinaryReadAll(viewer, icols, nnz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1440: PetscCall(PetscViewerBinaryReadAll(viewer, vwork, nnz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1441: /* store values in column major order */
1442: for (k = 0, i = 0; i < m; i++)
1443: for (j = 0; j < rlens[i]; j++, k++) v[i + lda * icols[k]] = vwork[k];
1444: PetscCall(PetscFree(rlens));
1445: PetscCall(PetscFree2(icols, vwork));
1446: }
1447: PetscCall(MatDenseRestoreArray(mat, &v));
1448: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
1449: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
1450: PetscFunctionReturn(PETSC_SUCCESS);
1451: }
1453: static PetscErrorCode MatLoad_SeqDense(Mat newMat, PetscViewer viewer)
1454: {
1455: PetscBool isbinary, ishdf5;
1457: PetscFunctionBegin;
1460: /* force binary viewer to load .info file if it has not yet done so */
1461: PetscCall(PetscViewerSetUp(viewer));
1462: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1463: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
1464: if (isbinary) {
1465: PetscCall(MatLoad_Dense_Binary(newMat, viewer));
1466: } else if (ishdf5) {
1467: #if defined(PETSC_HAVE_HDF5)
1468: PetscCall(MatLoad_Dense_HDF5(newMat, viewer));
1469: #else
1470: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
1471: #endif
1472: } else {
1473: 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);
1474: }
1475: PetscFunctionReturn(PETSC_SUCCESS);
1476: }
1478: static PetscErrorCode MatView_SeqDense_ASCII(Mat A, PetscViewer viewer)
1479: {
1480: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
1481: PetscInt i, j;
1482: const char *name;
1483: PetscScalar *v, *av;
1484: PetscViewerFormat format;
1485: #if defined(PETSC_USE_COMPLEX)
1486: PetscBool allreal = PETSC_TRUE;
1487: #endif
1489: PetscFunctionBegin;
1490: PetscCall(MatDenseGetArrayRead(A, (const PetscScalar **)&av));
1491: PetscCall(PetscViewerGetFormat(viewer, &format));
1492: if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1493: PetscFunctionReturn(PETSC_SUCCESS); /* do nothing for now */
1494: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
1495: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
1496: for (i = 0; i < A->rmap->n; i++) {
1497: v = av + i;
1498: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
1499: for (j = 0; j < A->cmap->n; j++) {
1500: #if defined(PETSC_USE_COMPLEX)
1501: if (PetscRealPart(*v) != 0.0 && PetscImaginaryPart(*v) != 0.0) {
1502: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i) ", j, (double)PetscRealPart(*v), (double)PetscImaginaryPart(*v)));
1503: } else if (PetscRealPart(*v)) {
1504: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", j, (double)PetscRealPart(*v)));
1505: }
1506: #else
1507: if (*v) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", j, (double)*v));
1508: #endif
1509: v += a->lda;
1510: }
1511: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
1512: }
1513: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
1514: } else {
1515: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
1516: #if defined(PETSC_USE_COMPLEX)
1517: /* determine if matrix has all real values */
1518: for (j = 0; j < A->cmap->n; j++) {
1519: v = av + j * a->lda;
1520: for (i = 0; i < A->rmap->n; i++) {
1521: if (PetscImaginaryPart(v[i])) {
1522: allreal = PETSC_FALSE;
1523: break;
1524: }
1525: }
1526: }
1527: #endif
1528: if (format == PETSC_VIEWER_ASCII_MATLAB) {
1529: PetscCall(PetscObjectGetName((PetscObject)A, &name));
1530: PetscCall(PetscViewerASCIIPrintf(viewer, "%% Size = %" PetscInt_FMT " %" PetscInt_FMT " \n", A->rmap->n, A->cmap->n));
1531: PetscCall(PetscViewerASCIIPrintf(viewer, "%s = zeros(%" PetscInt_FMT ",%" PetscInt_FMT ");\n", name, A->rmap->n, A->cmap->n));
1532: PetscCall(PetscViewerASCIIPrintf(viewer, "%s = [\n", name));
1533: }
1535: for (i = 0; i < A->rmap->n; i++) {
1536: v = av + i;
1537: for (j = 0; j < A->cmap->n; j++) {
1538: #if defined(PETSC_USE_COMPLEX)
1539: if (allreal) {
1540: PetscCall(PetscViewerASCIIPrintf(viewer, "%18.16e ", (double)PetscRealPart(*v)));
1541: } else {
1542: PetscCall(PetscViewerASCIIPrintf(viewer, "%18.16e + %18.16ei ", (double)PetscRealPart(*v), (double)PetscImaginaryPart(*v)));
1543: }
1544: #else
1545: PetscCall(PetscViewerASCIIPrintf(viewer, "%18.16e ", (double)*v));
1546: #endif
1547: v += a->lda;
1548: }
1549: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
1550: }
1551: if (format == PETSC_VIEWER_ASCII_MATLAB) PetscCall(PetscViewerASCIIPrintf(viewer, "];\n"));
1552: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
1553: }
1554: PetscCall(MatDenseRestoreArrayRead(A, (const PetscScalar **)&av));
1555: PetscCall(PetscViewerFlush(viewer));
1556: PetscFunctionReturn(PETSC_SUCCESS);
1557: }
1559: #include <petscdraw.h>
1560: static PetscErrorCode MatView_SeqDense_Draw_Zoom(PetscDraw draw, void *Aa)
1561: {
1562: Mat A = (Mat)Aa;
1563: PetscInt m = A->rmap->n, n = A->cmap->n, i, j;
1564: int color = PETSC_DRAW_WHITE;
1565: const PetscScalar *v;
1566: PetscViewer viewer;
1567: PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r;
1568: PetscViewerFormat format;
1570: PetscFunctionBegin;
1571: PetscCall(PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer));
1572: PetscCall(PetscViewerGetFormat(viewer, &format));
1573: PetscCall(PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr));
1575: /* Loop over matrix elements drawing boxes */
1576: PetscCall(MatDenseGetArrayRead(A, &v));
1577: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
1578: PetscDrawCollectiveBegin(draw);
1579: /* Blue for negative and Red for positive */
1580: for (j = 0; j < n; j++) {
1581: x_l = j;
1582: x_r = x_l + 1.0;
1583: for (i = 0; i < m; i++) {
1584: y_l = m - i - 1.0;
1585: y_r = y_l + 1.0;
1586: if (PetscRealPart(v[j * m + i]) > 0.) color = PETSC_DRAW_RED;
1587: else if (PetscRealPart(v[j * m + i]) < 0.) color = PETSC_DRAW_BLUE;
1588: else continue;
1589: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1590: }
1591: }
1592: PetscDrawCollectiveEnd(draw);
1593: } else {
1594: /* use contour shading to indicate magnitude of values */
1595: /* first determine max of all nonzero values */
1596: PetscReal minv = 0.0, maxv = 0.0;
1597: PetscDraw popup;
1599: for (i = 0; i < m * n; i++) {
1600: if (PetscAbsScalar(v[i]) > maxv) maxv = PetscAbsScalar(v[i]);
1601: }
1602: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1603: PetscCall(PetscDrawGetPopup(draw, &popup));
1604: PetscCall(PetscDrawScalePopup(popup, minv, maxv));
1606: PetscDrawCollectiveBegin(draw);
1607: for (j = 0; j < n; j++) {
1608: x_l = j;
1609: x_r = x_l + 1.0;
1610: for (i = 0; i < m; i++) {
1611: y_l = m - i - 1.0;
1612: y_r = y_l + 1.0;
1613: color = PetscDrawRealToColor(PetscAbsScalar(v[j * m + i]), minv, maxv);
1614: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1615: }
1616: }
1617: PetscDrawCollectiveEnd(draw);
1618: }
1619: PetscCall(MatDenseRestoreArrayRead(A, &v));
1620: PetscFunctionReturn(PETSC_SUCCESS);
1621: }
1623: static PetscErrorCode MatView_SeqDense_Draw(Mat A, PetscViewer viewer)
1624: {
1625: PetscDraw draw;
1626: PetscBool isnull;
1627: PetscReal xr, yr, xl, yl, h, w;
1629: PetscFunctionBegin;
1630: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1631: PetscCall(PetscDrawIsNull(draw, &isnull));
1632: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1634: xr = A->cmap->n;
1635: yr = A->rmap->n;
1636: h = yr / 10.0;
1637: w = xr / 10.0;
1638: xr += w;
1639: yr += h;
1640: xl = -w;
1641: yl = -h;
1642: PetscCall(PetscDrawSetCoordinates(draw, xl, yl, xr, yr));
1643: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer));
1644: PetscCall(PetscDrawZoom(draw, MatView_SeqDense_Draw_Zoom, A));
1645: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL));
1646: PetscCall(PetscDrawSave(draw));
1647: PetscFunctionReturn(PETSC_SUCCESS);
1648: }
1650: PetscErrorCode MatView_SeqDense(Mat A, PetscViewer viewer)
1651: {
1652: PetscBool iascii, isbinary, isdraw;
1654: PetscFunctionBegin;
1655: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1656: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1657: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1658: if (iascii) PetscCall(MatView_SeqDense_ASCII(A, viewer));
1659: else if (isbinary) PetscCall(MatView_Dense_Binary(A, viewer));
1660: else if (isdraw) PetscCall(MatView_SeqDense_Draw(A, viewer));
1661: PetscFunctionReturn(PETSC_SUCCESS);
1662: }
1664: static PetscErrorCode MatDensePlaceArray_SeqDense(Mat A, const PetscScalar *array)
1665: {
1666: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
1668: PetscFunctionBegin;
1669: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1670: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1671: PetscCheck(!a->unplacedarray, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseResetArray() first");
1672: a->unplacedarray = a->v;
1673: a->unplaced_user_alloc = a->user_alloc;
1674: a->v = (PetscScalar *)array;
1675: a->user_alloc = PETSC_TRUE;
1676: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1677: A->offloadmask = PETSC_OFFLOAD_CPU;
1678: #endif
1679: PetscFunctionReturn(PETSC_SUCCESS);
1680: }
1682: static PetscErrorCode MatDenseResetArray_SeqDense(Mat A)
1683: {
1684: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
1686: PetscFunctionBegin;
1687: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1688: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1689: a->v = a->unplacedarray;
1690: a->user_alloc = a->unplaced_user_alloc;
1691: a->unplacedarray = NULL;
1692: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1693: A->offloadmask = PETSC_OFFLOAD_CPU;
1694: #endif
1695: PetscFunctionReturn(PETSC_SUCCESS);
1696: }
1698: static PetscErrorCode MatDenseReplaceArray_SeqDense(Mat A, const PetscScalar *array)
1699: {
1700: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
1702: PetscFunctionBegin;
1703: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1704: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1705: if (!a->user_alloc) PetscCall(PetscFree(a->v));
1706: a->v = (PetscScalar *)array;
1707: a->user_alloc = PETSC_FALSE;
1708: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1709: A->offloadmask = PETSC_OFFLOAD_CPU;
1710: #endif
1711: PetscFunctionReturn(PETSC_SUCCESS);
1712: }
1714: PetscErrorCode MatDestroy_SeqDense(Mat mat)
1715: {
1716: Mat_SeqDense *l = (Mat_SeqDense *)mat->data;
1718: PetscFunctionBegin;
1719: PetscCall(PetscLogObjectState((PetscObject)mat, "Rows %" PetscInt_FMT " Cols %" PetscInt_FMT, mat->rmap->n, mat->cmap->n));
1720: PetscCall(VecDestroy(&l->qrrhs));
1721: PetscCall(PetscFree(l->tau));
1722: PetscCall(PetscFree(l->pivots));
1723: PetscCall(PetscFree(l->fwork));
1724: if (!l->user_alloc) PetscCall(PetscFree(l->v));
1725: if (!l->unplaced_user_alloc) PetscCall(PetscFree(l->unplacedarray));
1726: PetscCheck(!l->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1727: PetscCheck(!l->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1728: PetscCall(VecDestroy(&l->cvec));
1729: PetscCall(MatDestroy(&l->cmat));
1730: PetscCall(PetscFree(mat->data));
1732: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, NULL));
1733: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatQRFactor_C", NULL));
1734: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatQRFactorSymbolic_C", NULL));
1735: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatQRFactorNumeric_C", NULL));
1736: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetLDA_C", NULL));
1737: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseSetLDA_C", NULL));
1738: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArray_C", NULL));
1739: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArray_C", NULL));
1740: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDensePlaceArray_C", NULL));
1741: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseResetArray_C", NULL));
1742: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseReplaceArray_C", NULL));
1743: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayRead_C", NULL));
1744: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayRead_C", NULL));
1745: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayWrite_C", NULL));
1746: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayWrite_C", NULL));
1747: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_seqaij_C", NULL));
1748: #if defined(PETSC_HAVE_ELEMENTAL)
1749: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_elemental_C", NULL));
1750: #endif
1751: #if defined(PETSC_HAVE_SCALAPACK)
1752: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_scalapack_C", NULL));
1753: #endif
1754: #if defined(PETSC_HAVE_CUDA)
1755: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_seqdensecuda_C", NULL));
1756: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdensecuda_seqdensecuda_C", NULL));
1757: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdensecuda_seqdense_C", NULL));
1758: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdense_seqdensecuda_C", NULL));
1759: #endif
1760: #if defined(PETSC_HAVE_HIP)
1761: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_seqdensehip_C", NULL));
1762: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdensehip_seqdensehip_C", NULL));
1763: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdensehip_seqdense_C", NULL));
1764: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdense_seqdensehip_C", NULL));
1765: #endif
1766: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSeqDenseSetPreallocation_C", NULL));
1767: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqaij_seqdense_C", NULL));
1768: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdense_seqdense_C", NULL));
1769: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqbaij_seqdense_C", NULL));
1770: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqsbaij_seqdense_C", NULL));
1772: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumn_C", NULL));
1773: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumn_C", NULL));
1774: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVec_C", NULL));
1775: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVec_C", NULL));
1776: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecRead_C", NULL));
1777: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecRead_C", NULL));
1778: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecWrite_C", NULL));
1779: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecWrite_C", NULL));
1780: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetSubMatrix_C", NULL));
1781: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreSubMatrix_C", NULL));
1782: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultAddColumnRange_C", NULL));
1783: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeColumnRange_C", NULL));
1784: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeAddColumnRange_C", NULL));
1785: PetscFunctionReturn(PETSC_SUCCESS);
1786: }
1788: static PetscErrorCode MatTranspose_SeqDense(Mat A, MatReuse reuse, Mat *matout)
1789: {
1790: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1791: PetscInt k, j, m = A->rmap->n, M = mat->lda, n = A->cmap->n;
1792: PetscScalar *v, tmp;
1794: PetscFunctionBegin;
1795: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1796: if (reuse == MAT_INPLACE_MATRIX) {
1797: if (m == n) { /* in place transpose */
1798: PetscCall(MatDenseGetArray(A, &v));
1799: for (j = 0; j < m; j++) {
1800: for (k = 0; k < j; k++) {
1801: tmp = v[j + k * M];
1802: v[j + k * M] = v[k + j * M];
1803: v[k + j * M] = tmp;
1804: }
1805: }
1806: PetscCall(MatDenseRestoreArray(A, &v));
1807: } else { /* reuse memory, temporary allocates new memory */
1808: PetscScalar *v2;
1809: PetscLayout tmplayout;
1811: PetscCall(PetscMalloc1((size_t)m * n, &v2));
1812: PetscCall(MatDenseGetArray(A, &v));
1813: for (j = 0; j < n; j++) {
1814: for (k = 0; k < m; k++) v2[j + (size_t)k * n] = v[k + (size_t)j * M];
1815: }
1816: PetscCall(PetscArraycpy(v, v2, (size_t)m * n));
1817: PetscCall(PetscFree(v2));
1818: PetscCall(MatDenseRestoreArray(A, &v));
1819: /* cleanup size dependent quantities */
1820: PetscCall(VecDestroy(&mat->cvec));
1821: PetscCall(MatDestroy(&mat->cmat));
1822: PetscCall(PetscFree(mat->pivots));
1823: PetscCall(PetscFree(mat->fwork));
1824: /* swap row/col layouts */
1825: PetscCall(PetscBLASIntCast(n, &mat->lda));
1826: tmplayout = A->rmap;
1827: A->rmap = A->cmap;
1828: A->cmap = tmplayout;
1829: }
1830: } else { /* out-of-place transpose */
1831: Mat tmat;
1832: Mat_SeqDense *tmatd;
1833: PetscScalar *v2;
1834: PetscInt M2;
1836: if (reuse == MAT_INITIAL_MATRIX) {
1837: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &tmat));
1838: PetscCall(MatSetSizes(tmat, A->cmap->n, A->rmap->n, A->cmap->n, A->rmap->n));
1839: PetscCall(MatSetType(tmat, ((PetscObject)A)->type_name));
1840: PetscCall(MatSeqDenseSetPreallocation(tmat, NULL));
1841: } else tmat = *matout;
1843: PetscCall(MatDenseGetArrayRead(A, (const PetscScalar **)&v));
1844: PetscCall(MatDenseGetArray(tmat, &v2));
1845: tmatd = (Mat_SeqDense *)tmat->data;
1846: M2 = tmatd->lda;
1847: for (j = 0; j < n; j++) {
1848: for (k = 0; k < m; k++) v2[j + k * M2] = v[k + j * M];
1849: }
1850: PetscCall(MatDenseRestoreArray(tmat, &v2));
1851: PetscCall(MatDenseRestoreArrayRead(A, (const PetscScalar **)&v));
1852: PetscCall(MatAssemblyBegin(tmat, MAT_FINAL_ASSEMBLY));
1853: PetscCall(MatAssemblyEnd(tmat, MAT_FINAL_ASSEMBLY));
1854: *matout = tmat;
1855: }
1856: PetscFunctionReturn(PETSC_SUCCESS);
1857: }
1859: static PetscErrorCode MatEqual_SeqDense(Mat A1, Mat A2, PetscBool *flg)
1860: {
1861: Mat_SeqDense *mat1 = (Mat_SeqDense *)A1->data;
1862: Mat_SeqDense *mat2 = (Mat_SeqDense *)A2->data;
1863: PetscInt i;
1864: const PetscScalar *v1, *v2;
1866: PetscFunctionBegin;
1867: if (A1->rmap->n != A2->rmap->n) {
1868: *flg = PETSC_FALSE;
1869: PetscFunctionReturn(PETSC_SUCCESS);
1870: }
1871: if (A1->cmap->n != A2->cmap->n) {
1872: *flg = PETSC_FALSE;
1873: PetscFunctionReturn(PETSC_SUCCESS);
1874: }
1875: PetscCall(MatDenseGetArrayRead(A1, &v1));
1876: PetscCall(MatDenseGetArrayRead(A2, &v2));
1877: for (i = 0; i < A1->cmap->n; i++) {
1878: PetscCall(PetscArraycmp(v1, v2, A1->rmap->n, flg));
1879: if (*flg == PETSC_FALSE) PetscFunctionReturn(PETSC_SUCCESS);
1880: v1 += mat1->lda;
1881: v2 += mat2->lda;
1882: }
1883: PetscCall(MatDenseRestoreArrayRead(A1, &v1));
1884: PetscCall(MatDenseRestoreArrayRead(A2, &v2));
1885: *flg = PETSC_TRUE;
1886: PetscFunctionReturn(PETSC_SUCCESS);
1887: }
1889: PetscErrorCode MatGetDiagonal_SeqDense(Mat A, Vec v)
1890: {
1891: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1892: PetscInt i, n, len;
1893: PetscScalar *x;
1894: const PetscScalar *vv;
1896: PetscFunctionBegin;
1897: PetscCall(VecGetSize(v, &n));
1898: PetscCall(VecGetArray(v, &x));
1899: len = PetscMin(A->rmap->n, A->cmap->n);
1900: PetscCall(MatDenseGetArrayRead(A, &vv));
1901: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming mat and vec");
1902: for (i = 0; i < len; i++) x[i] = vv[i * mat->lda + i];
1903: PetscCall(MatDenseRestoreArrayRead(A, &vv));
1904: PetscCall(VecRestoreArray(v, &x));
1905: PetscFunctionReturn(PETSC_SUCCESS);
1906: }
1908: static PetscErrorCode MatDiagonalScale_SeqDense(Mat A, Vec ll, Vec rr)
1909: {
1910: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1911: const PetscScalar *l, *r;
1912: PetscScalar x, *v, *vv;
1913: PetscInt i, j, m = A->rmap->n, n = A->cmap->n;
1915: PetscFunctionBegin;
1916: PetscCall(MatDenseGetArray(A, &vv));
1917: if (ll) {
1918: PetscCall(VecGetSize(ll, &m));
1919: PetscCall(VecGetArrayRead(ll, &l));
1920: PetscCheck(m == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vec wrong size");
1921: for (i = 0; i < m; i++) {
1922: x = l[i];
1923: v = vv + i;
1924: for (j = 0; j < n; j++) {
1925: (*v) *= x;
1926: v += mat->lda;
1927: }
1928: }
1929: PetscCall(VecRestoreArrayRead(ll, &l));
1930: PetscCall(PetscLogFlops(1.0 * n * m));
1931: }
1932: if (rr) {
1933: PetscCall(VecGetSize(rr, &n));
1934: PetscCall(VecGetArrayRead(rr, &r));
1935: PetscCheck(n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vec wrong size");
1936: for (i = 0; i < n; i++) {
1937: x = r[i];
1938: v = vv + i * mat->lda;
1939: for (j = 0; j < m; j++) (*v++) *= x;
1940: }
1941: PetscCall(VecRestoreArrayRead(rr, &r));
1942: PetscCall(PetscLogFlops(1.0 * n * m));
1943: }
1944: PetscCall(MatDenseRestoreArray(A, &vv));
1945: PetscFunctionReturn(PETSC_SUCCESS);
1946: }
1948: PetscErrorCode MatNorm_SeqDense(Mat A, NormType type, PetscReal *nrm)
1949: {
1950: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1951: PetscScalar *v, *vv;
1952: PetscReal sum = 0.0;
1953: PetscInt lda, m = A->rmap->n, i, j;
1955: PetscFunctionBegin;
1956: PetscCall(MatDenseGetArrayRead(A, (const PetscScalar **)&vv));
1957: PetscCall(MatDenseGetLDA(A, &lda));
1958: v = vv;
1959: if (type == NORM_FROBENIUS) {
1960: if (lda > m) {
1961: for (j = 0; j < A->cmap->n; j++) {
1962: v = vv + j * lda;
1963: for (i = 0; i < m; i++) {
1964: sum += PetscRealPart(PetscConj(*v) * (*v));
1965: v++;
1966: }
1967: }
1968: } else {
1969: #if defined(PETSC_USE_REAL___FP16)
1970: PetscBLASInt one = 1, cnt = A->cmap->n * A->rmap->n;
1971: PetscCallBLAS("BLASnrm2", *nrm = BLASnrm2_(&cnt, v, &one));
1972: }
1973: #else
1974: for (i = 0; i < A->cmap->n * A->rmap->n; i++) {
1975: sum += PetscRealPart(PetscConj(*v) * (*v));
1976: v++;
1977: }
1978: }
1979: *nrm = PetscSqrtReal(sum);
1980: #endif
1981: PetscCall(PetscLogFlops(2.0 * A->cmap->n * A->rmap->n));
1982: } else if (type == NORM_1) {
1983: *nrm = 0.0;
1984: for (j = 0; j < A->cmap->n; j++) {
1985: v = vv + j * mat->lda;
1986: sum = 0.0;
1987: for (i = 0; i < A->rmap->n; i++) {
1988: sum += PetscAbsScalar(*v);
1989: v++;
1990: }
1991: if (sum > *nrm) *nrm = sum;
1992: }
1993: PetscCall(PetscLogFlops(1.0 * A->cmap->n * A->rmap->n));
1994: } else if (type == NORM_INFINITY) {
1995: *nrm = 0.0;
1996: for (j = 0; j < A->rmap->n; j++) {
1997: v = vv + j;
1998: sum = 0.0;
1999: for (i = 0; i < A->cmap->n; i++) {
2000: sum += PetscAbsScalar(*v);
2001: v += mat->lda;
2002: }
2003: if (sum > *nrm) *nrm = sum;
2004: }
2005: PetscCall(PetscLogFlops(1.0 * A->cmap->n * A->rmap->n));
2006: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No two norm");
2007: PetscCall(MatDenseRestoreArrayRead(A, (const PetscScalar **)&vv));
2008: PetscFunctionReturn(PETSC_SUCCESS);
2009: }
2011: static PetscErrorCode MatSetOption_SeqDense(Mat A, MatOption op, PetscBool flg)
2012: {
2013: Mat_SeqDense *aij = (Mat_SeqDense *)A->data;
2015: PetscFunctionBegin;
2016: switch (op) {
2017: case MAT_ROW_ORIENTED:
2018: aij->roworiented = flg;
2019: break;
2020: case MAT_NEW_NONZERO_LOCATIONS:
2021: case MAT_NEW_NONZERO_LOCATION_ERR:
2022: case MAT_NEW_NONZERO_ALLOCATION_ERR:
2023: case MAT_FORCE_DIAGONAL_ENTRIES:
2024: case MAT_KEEP_NONZERO_PATTERN:
2025: case MAT_IGNORE_OFF_PROC_ENTRIES:
2026: case MAT_USE_HASH_TABLE:
2027: case MAT_IGNORE_ZERO_ENTRIES:
2028: case MAT_IGNORE_LOWER_TRIANGULAR:
2029: case MAT_SORTED_FULL:
2030: PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op]));
2031: break;
2032: case MAT_SPD:
2033: case MAT_SYMMETRIC:
2034: case MAT_STRUCTURALLY_SYMMETRIC:
2035: case MAT_HERMITIAN:
2036: case MAT_SYMMETRY_ETERNAL:
2037: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
2038: case MAT_SPD_ETERNAL:
2039: break;
2040: default:
2041: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %s", MatOptions[op]);
2042: }
2043: PetscFunctionReturn(PETSC_SUCCESS);
2044: }
2046: PetscErrorCode MatZeroEntries_SeqDense(Mat A)
2047: {
2048: Mat_SeqDense *l = (Mat_SeqDense *)A->data;
2049: PetscInt lda = l->lda, m = A->rmap->n, n = A->cmap->n, j;
2050: PetscScalar *v;
2052: PetscFunctionBegin;
2053: PetscCall(MatDenseGetArrayWrite(A, &v));
2054: if (lda > m) {
2055: for (j = 0; j < n; j++) PetscCall(PetscArrayzero(v + j * lda, m));
2056: } else {
2057: PetscCall(PetscArrayzero(v, PetscInt64Mult(m, n)));
2058: }
2059: PetscCall(MatDenseRestoreArrayWrite(A, &v));
2060: PetscFunctionReturn(PETSC_SUCCESS);
2061: }
2063: static PetscErrorCode MatZeroRows_SeqDense(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
2064: {
2065: Mat_SeqDense *l = (Mat_SeqDense *)A->data;
2066: PetscInt m = l->lda, n = A->cmap->n, i, j;
2067: PetscScalar *slot, *bb, *v;
2068: const PetscScalar *xx;
2070: PetscFunctionBegin;
2071: if (PetscDefined(USE_DEBUG)) {
2072: for (i = 0; i < N; i++) {
2073: PetscCheck(rows[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative row requested to be zeroed");
2074: PetscCheck(rows[i] < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %" PetscInt_FMT " requested to be zeroed greater than or equal number of rows %" PetscInt_FMT, rows[i], A->rmap->n);
2075: }
2076: }
2077: if (!N) PetscFunctionReturn(PETSC_SUCCESS);
2079: /* fix right-hand side if needed */
2080: if (x && b) {
2081: PetscCall(VecGetArrayRead(x, &xx));
2082: PetscCall(VecGetArray(b, &bb));
2083: for (i = 0; i < N; i++) bb[rows[i]] = diag * xx[rows[i]];
2084: PetscCall(VecRestoreArrayRead(x, &xx));
2085: PetscCall(VecRestoreArray(b, &bb));
2086: }
2088: PetscCall(MatDenseGetArray(A, &v));
2089: for (i = 0; i < N; i++) {
2090: slot = v + rows[i];
2091: for (j = 0; j < n; j++) {
2092: *slot = 0.0;
2093: slot += m;
2094: }
2095: }
2096: if (diag != 0.0) {
2097: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only coded for square matrices");
2098: for (i = 0; i < N; i++) {
2099: slot = v + (m + 1) * rows[i];
2100: *slot = diag;
2101: }
2102: }
2103: PetscCall(MatDenseRestoreArray(A, &v));
2104: PetscFunctionReturn(PETSC_SUCCESS);
2105: }
2107: static PetscErrorCode MatDenseGetLDA_SeqDense(Mat A, PetscInt *lda)
2108: {
2109: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2111: PetscFunctionBegin;
2112: *lda = mat->lda;
2113: PetscFunctionReturn(PETSC_SUCCESS);
2114: }
2116: PetscErrorCode MatDenseGetArray_SeqDense(Mat A, PetscScalar **array)
2117: {
2118: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2120: PetscFunctionBegin;
2121: PetscCheck(!mat->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
2122: *array = mat->v;
2123: PetscFunctionReturn(PETSC_SUCCESS);
2124: }
2126: PetscErrorCode MatDenseRestoreArray_SeqDense(Mat A, PetscScalar **array)
2127: {
2128: PetscFunctionBegin;
2129: if (array) *array = NULL;
2130: PetscFunctionReturn(PETSC_SUCCESS);
2131: }
2133: /*@
2134: MatDenseGetLDA - gets the leading dimension of the array returned from `MatDenseGetArray()`
2136: Not Collective
2138: Input Parameter:
2139: . A - a `MATDENSE` or `MATDENSECUDA` matrix
2141: Output Parameter:
2142: . lda - the leading dimension
2144: Level: intermediate
2146: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseSetLDA()`
2147: @*/
2148: PetscErrorCode MatDenseGetLDA(Mat A, PetscInt *lda)
2149: {
2150: PetscFunctionBegin;
2152: PetscAssertPointer(lda, 2);
2153: MatCheckPreallocated(A, 1);
2154: PetscUseMethod(A, "MatDenseGetLDA_C", (Mat, PetscInt *), (A, lda));
2155: PetscFunctionReturn(PETSC_SUCCESS);
2156: }
2158: /*@
2159: MatDenseSetLDA - Sets the leading dimension of the array used by the `MATDENSE` matrix
2161: Collective if the matrix layouts have not yet been setup
2163: Input Parameters:
2164: + A - a `MATDENSE` or `MATDENSECUDA` matrix
2165: - lda - the leading dimension
2167: Level: intermediate
2169: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetLDA()`
2170: @*/
2171: PetscErrorCode MatDenseSetLDA(Mat A, PetscInt lda)
2172: {
2173: PetscFunctionBegin;
2175: PetscTryMethod(A, "MatDenseSetLDA_C", (Mat, PetscInt), (A, lda));
2176: PetscFunctionReturn(PETSC_SUCCESS);
2177: }
2179: /*@C
2180: MatDenseGetArray - gives read-write access to the array where the data for a `MATDENSE` matrix is stored
2182: Logically Collective
2184: Input Parameter:
2185: . A - a dense matrix
2187: Output Parameter:
2188: . array - pointer to the data
2190: Level: intermediate
2192: Fortran Notes:
2193: `MatDenseGetArray()` Fortran binding is deprecated (since PETSc 3.19), use `MatDenseGetArrayF90()`
2195: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2196: @*/
2197: PetscErrorCode MatDenseGetArray(Mat A, PetscScalar *array[])
2198: {
2199: PetscFunctionBegin;
2201: PetscAssertPointer(array, 2);
2202: PetscUseMethod(A, "MatDenseGetArray_C", (Mat, PetscScalar **), (A, array));
2203: PetscFunctionReturn(PETSC_SUCCESS);
2204: }
2206: /*@C
2207: MatDenseRestoreArray - returns access to the array where the data for a `MATDENSE` matrix is stored obtained by `MatDenseGetArray()`
2209: Logically Collective
2211: Input Parameters:
2212: + A - a dense matrix
2213: - array - pointer to the data (may be `NULL`)
2215: Level: intermediate
2217: Fortran Notes:
2218: `MatDenseRestoreArray()` Fortran binding is deprecated (since PETSc 3.19), use `MatDenseRestoreArrayF90()`
2220: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2221: @*/
2222: PetscErrorCode MatDenseRestoreArray(Mat A, PetscScalar *array[])
2223: {
2224: PetscFunctionBegin;
2226: if (array) PetscAssertPointer(array, 2);
2227: PetscUseMethod(A, "MatDenseRestoreArray_C", (Mat, PetscScalar **), (A, array));
2228: PetscCall(PetscObjectStateIncrease((PetscObject)A));
2229: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
2230: A->offloadmask = PETSC_OFFLOAD_CPU;
2231: #endif
2232: PetscFunctionReturn(PETSC_SUCCESS);
2233: }
2235: /*@C
2236: MatDenseGetArrayRead - gives read-only access to the array where the data for a `MATDENSE` matrix is stored
2238: Not Collective
2240: Input Parameter:
2241: . A - a dense matrix
2243: Output Parameter:
2244: . array - pointer to the data
2246: Level: intermediate
2248: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayRead()`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2249: @*/
2250: PetscErrorCode MatDenseGetArrayRead(Mat A, const PetscScalar *array[])
2251: {
2252: PetscFunctionBegin;
2254: PetscAssertPointer(array, 2);
2255: PetscUseMethod(A, "MatDenseGetArrayRead_C", (Mat, PetscScalar **), (A, (PetscScalar **)array));
2256: PetscFunctionReturn(PETSC_SUCCESS);
2257: }
2259: /*@C
2260: MatDenseRestoreArrayRead - returns access to the array where the data for a `MATDENSE` matrix is stored obtained by `MatDenseGetArrayRead()`
2262: Not Collective
2264: Input Parameters:
2265: + A - a dense matrix
2266: - array - pointer to the data (may be `NULL`)
2268: Level: intermediate
2270: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayRead()`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2271: @*/
2272: PetscErrorCode MatDenseRestoreArrayRead(Mat A, const PetscScalar *array[])
2273: {
2274: PetscFunctionBegin;
2276: if (array) PetscAssertPointer(array, 2);
2277: PetscUseMethod(A, "MatDenseRestoreArrayRead_C", (Mat, PetscScalar **), (A, (PetscScalar **)array));
2278: PetscFunctionReturn(PETSC_SUCCESS);
2279: }
2281: /*@C
2282: MatDenseGetArrayWrite - gives write-only access to the array where the data for a `MATDENSE` matrix is stored
2284: Not Collective
2286: Input Parameter:
2287: . A - a dense matrix
2289: Output Parameter:
2290: . array - pointer to the data
2292: Level: intermediate
2294: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayWrite()`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`
2295: @*/
2296: PetscErrorCode MatDenseGetArrayWrite(Mat A, PetscScalar *array[])
2297: {
2298: PetscFunctionBegin;
2300: PetscAssertPointer(array, 2);
2301: PetscUseMethod(A, "MatDenseGetArrayWrite_C", (Mat, PetscScalar **), (A, array));
2302: PetscFunctionReturn(PETSC_SUCCESS);
2303: }
2305: /*@C
2306: MatDenseRestoreArrayWrite - returns access to the array where the data for a `MATDENSE` matrix is stored obtained by `MatDenseGetArrayWrite()`
2308: Not Collective
2310: Input Parameters:
2311: + A - a dense matrix
2312: - array - pointer to the data (may be `NULL`)
2314: Level: intermediate
2316: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayWrite()`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`
2317: @*/
2318: PetscErrorCode MatDenseRestoreArrayWrite(Mat A, PetscScalar *array[])
2319: {
2320: PetscFunctionBegin;
2322: if (array) PetscAssertPointer(array, 2);
2323: PetscUseMethod(A, "MatDenseRestoreArrayWrite_C", (Mat, PetscScalar **), (A, array));
2324: PetscCall(PetscObjectStateIncrease((PetscObject)A));
2325: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
2326: A->offloadmask = PETSC_OFFLOAD_CPU;
2327: #endif
2328: PetscFunctionReturn(PETSC_SUCCESS);
2329: }
2331: /*@C
2332: MatDenseGetArrayAndMemType - gives read-write access to the array where the data for a `MATDENSE` matrix is stored
2334: Logically Collective
2336: Input Parameter:
2337: . A - a dense matrix
2339: Output Parameters:
2340: + array - pointer to the data
2341: - mtype - memory type of the returned pointer
2343: Level: intermediate
2345: Note:
2346: If the matrix is of a device type such as `MATDENSECUDA`, `MATDENSEHIP`, etc.,
2347: an array on device is always returned and is guaranteed to contain the matrix's latest data.
2349: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayAndMemType()`, `MatDenseGetArrayReadAndMemType()`, `MatDenseGetArrayWriteAndMemType()`, `MatDenseGetArrayRead()`,
2350: `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`, `MatSeqAIJGetCSRAndMemType()`
2351: @*/
2352: PetscErrorCode MatDenseGetArrayAndMemType(Mat A, PetscScalar *array[], PetscMemType *mtype)
2353: {
2354: PetscBool isMPI;
2356: PetscFunctionBegin;
2358: PetscAssertPointer(array, 2);
2359: PetscCall(MatBindToCPU(A, PETSC_FALSE)); /* We want device matrices to always return device arrays, so we unbind the matrix if it is bound to CPU */
2360: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2361: if (isMPI) {
2362: /* Dispatch here so that the code can be reused for all subclasses of MATDENSE */
2363: PetscCall(MatDenseGetArrayAndMemType(((Mat_MPIDense *)A->data)->A, array, mtype));
2364: } else {
2365: PetscErrorCode (*fptr)(Mat, PetscScalar **, PetscMemType *);
2367: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseGetArrayAndMemType_C", &fptr));
2368: if (fptr) {
2369: PetscCall((*fptr)(A, array, mtype));
2370: } else {
2371: PetscUseMethod(A, "MatDenseGetArray_C", (Mat, PetscScalar **), (A, array));
2372: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
2373: }
2374: }
2375: PetscFunctionReturn(PETSC_SUCCESS);
2376: }
2378: /*@C
2379: MatDenseRestoreArrayAndMemType - returns access to the array that is obtained by `MatDenseGetArrayAndMemType()`
2381: Logically Collective
2383: Input Parameters:
2384: + A - a dense matrix
2385: - array - pointer to the data
2387: Level: intermediate
2389: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayAndMemType()`, `MatDenseGetArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2390: @*/
2391: PetscErrorCode MatDenseRestoreArrayAndMemType(Mat A, PetscScalar *array[])
2392: {
2393: PetscBool isMPI;
2395: PetscFunctionBegin;
2397: PetscAssertPointer(array, 2);
2398: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2399: if (isMPI) {
2400: PetscCall(MatDenseRestoreArrayAndMemType(((Mat_MPIDense *)A->data)->A, array));
2401: } else {
2402: PetscErrorCode (*fptr)(Mat, PetscScalar **);
2404: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseRestoreArrayAndMemType_C", &fptr));
2405: if (fptr) {
2406: PetscCall((*fptr)(A, array));
2407: } else {
2408: PetscUseMethod(A, "MatDenseRestoreArray_C", (Mat, PetscScalar **), (A, array));
2409: }
2410: *array = NULL;
2411: }
2412: PetscCall(PetscObjectStateIncrease((PetscObject)A));
2413: PetscFunctionReturn(PETSC_SUCCESS);
2414: }
2416: /*@C
2417: MatDenseGetArrayReadAndMemType - gives read-only access to the array where the data for a `MATDENSE` matrix is stored
2419: Logically Collective
2421: Input Parameter:
2422: . A - a dense matrix
2424: Output Parameters:
2425: + array - pointer to the data
2426: - mtype - memory type of the returned pointer
2428: Level: intermediate
2430: Note:
2431: If the matrix is of a device type such as `MATDENSECUDA`, `MATDENSEHIP`, etc.,
2432: an array on device is always returned and is guaranteed to contain the matrix's latest data.
2434: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayReadAndMemType()`, `MatDenseGetArrayWriteAndMemType()`,
2435: `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`, `MatSeqAIJGetCSRAndMemType()`
2436: @*/
2437: PetscErrorCode MatDenseGetArrayReadAndMemType(Mat A, const PetscScalar *array[], PetscMemType *mtype)
2438: {
2439: PetscBool isMPI;
2441: PetscFunctionBegin;
2443: PetscAssertPointer(array, 2);
2444: PetscCall(MatBindToCPU(A, PETSC_FALSE)); /* We want device matrices to always return device arrays, so we unbind the matrix if it is bound to CPU */
2445: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2446: if (isMPI) { /* Dispatch here so that the code can be reused for all subclasses of MATDENSE */
2447: PetscCall(MatDenseGetArrayReadAndMemType(((Mat_MPIDense *)A->data)->A, array, mtype));
2448: } else {
2449: PetscErrorCode (*fptr)(Mat, const PetscScalar **, PetscMemType *);
2451: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseGetArrayReadAndMemType_C", &fptr));
2452: if (fptr) {
2453: PetscCall((*fptr)(A, array, mtype));
2454: } else {
2455: PetscUseMethod(A, "MatDenseGetArrayRead_C", (Mat, PetscScalar **), (A, (PetscScalar **)array));
2456: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
2457: }
2458: }
2459: PetscFunctionReturn(PETSC_SUCCESS);
2460: }
2462: /*@C
2463: MatDenseRestoreArrayReadAndMemType - returns access to the array that is obtained by `MatDenseGetArrayReadAndMemType()`
2465: Logically Collective
2467: Input Parameters:
2468: + A - a dense matrix
2469: - array - pointer to the data
2471: Level: intermediate
2473: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayReadAndMemType()`, `MatDenseGetArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2474: @*/
2475: PetscErrorCode MatDenseRestoreArrayReadAndMemType(Mat A, const PetscScalar *array[])
2476: {
2477: PetscBool isMPI;
2479: PetscFunctionBegin;
2481: PetscAssertPointer(array, 2);
2482: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2483: if (isMPI) {
2484: PetscCall(MatDenseRestoreArrayReadAndMemType(((Mat_MPIDense *)A->data)->A, array));
2485: } else {
2486: PetscErrorCode (*fptr)(Mat, const PetscScalar **);
2488: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseRestoreArrayReadAndMemType_C", &fptr));
2489: if (fptr) {
2490: PetscCall((*fptr)(A, array));
2491: } else {
2492: PetscUseMethod(A, "MatDenseRestoreArrayRead_C", (Mat, PetscScalar **), (A, (PetscScalar **)array));
2493: }
2494: *array = NULL;
2495: }
2496: PetscFunctionReturn(PETSC_SUCCESS);
2497: }
2499: /*@C
2500: MatDenseGetArrayWriteAndMemType - gives write-only access to the array where the data for a `MATDENSE` matrix is stored
2502: Logically Collective
2504: Input Parameter:
2505: . A - a dense matrix
2507: Output Parameters:
2508: + array - pointer to the data
2509: - mtype - memory type of the returned pointer
2511: Level: intermediate
2513: Note:
2514: If the matrix is of a device type such as `MATDENSECUDA`, `MATDENSEHIP`, etc.,
2515: an array on device is always returned and is guaranteed to contain the matrix's latest data.
2517: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayWriteAndMemType()`, `MatDenseGetArrayReadAndMemType()`, `MatDenseGetArrayRead()`,
2518: `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`, `MatSeqAIJGetCSRAndMemType()`
2519: @*/
2520: PetscErrorCode MatDenseGetArrayWriteAndMemType(Mat A, PetscScalar *array[], PetscMemType *mtype)
2521: {
2522: PetscBool isMPI;
2524: PetscFunctionBegin;
2526: PetscAssertPointer(array, 2);
2527: PetscCall(MatBindToCPU(A, PETSC_FALSE)); /* We want device matrices to always return device arrays, so we unbind the matrix if it is bound to CPU */
2528: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2529: if (isMPI) {
2530: PetscCall(MatDenseGetArrayWriteAndMemType(((Mat_MPIDense *)A->data)->A, array, mtype));
2531: } else {
2532: PetscErrorCode (*fptr)(Mat, PetscScalar **, PetscMemType *);
2534: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseGetArrayWriteAndMemType_C", &fptr));
2535: if (fptr) {
2536: PetscCall((*fptr)(A, array, mtype));
2537: } else {
2538: PetscUseMethod(A, "MatDenseGetArrayWrite_C", (Mat, PetscScalar **), (A, array));
2539: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
2540: }
2541: }
2542: PetscFunctionReturn(PETSC_SUCCESS);
2543: }
2545: /*@C
2546: MatDenseRestoreArrayWriteAndMemType - returns access to the array that is obtained by `MatDenseGetArrayReadAndMemType()`
2548: Logically Collective
2550: Input Parameters:
2551: + A - a dense matrix
2552: - array - pointer to the data
2554: Level: intermediate
2556: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayWriteAndMemType()`, `MatDenseGetArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2557: @*/
2558: PetscErrorCode MatDenseRestoreArrayWriteAndMemType(Mat A, PetscScalar *array[])
2559: {
2560: PetscBool isMPI;
2562: PetscFunctionBegin;
2564: PetscAssertPointer(array, 2);
2565: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2566: if (isMPI) {
2567: PetscCall(MatDenseRestoreArrayWriteAndMemType(((Mat_MPIDense *)A->data)->A, array));
2568: } else {
2569: PetscErrorCode (*fptr)(Mat, PetscScalar **);
2571: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseRestoreArrayWriteAndMemType_C", &fptr));
2572: if (fptr) {
2573: PetscCall((*fptr)(A, array));
2574: } else {
2575: PetscUseMethod(A, "MatDenseRestoreArrayWrite_C", (Mat, PetscScalar **), (A, array));
2576: }
2577: *array = NULL;
2578: }
2579: PetscCall(PetscObjectStateIncrease((PetscObject)A));
2580: PetscFunctionReturn(PETSC_SUCCESS);
2581: }
2583: static PetscErrorCode MatCreateSubMatrix_SeqDense(Mat A, IS isrow, IS iscol, MatReuse scall, Mat *B)
2584: {
2585: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2586: PetscInt i, j, nrows, ncols, ldb;
2587: const PetscInt *irow, *icol;
2588: PetscScalar *av, *bv, *v = mat->v;
2589: Mat newmat;
2591: PetscFunctionBegin;
2592: PetscCall(ISGetIndices(isrow, &irow));
2593: PetscCall(ISGetIndices(iscol, &icol));
2594: PetscCall(ISGetLocalSize(isrow, &nrows));
2595: PetscCall(ISGetLocalSize(iscol, &ncols));
2597: /* Check submatrixcall */
2598: if (scall == MAT_REUSE_MATRIX) {
2599: PetscInt n_cols, n_rows;
2600: PetscCall(MatGetSize(*B, &n_rows, &n_cols));
2601: if (n_rows != nrows || n_cols != ncols) {
2602: /* resize the result matrix to match number of requested rows/columns */
2603: PetscCall(MatSetSizes(*B, nrows, ncols, nrows, ncols));
2604: }
2605: newmat = *B;
2606: } else {
2607: /* Create and fill new matrix */
2608: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &newmat));
2609: PetscCall(MatSetSizes(newmat, nrows, ncols, nrows, ncols));
2610: PetscCall(MatSetType(newmat, ((PetscObject)A)->type_name));
2611: PetscCall(MatSeqDenseSetPreallocation(newmat, NULL));
2612: }
2614: /* Now extract the data pointers and do the copy,column at a time */
2615: PetscCall(MatDenseGetArray(newmat, &bv));
2616: PetscCall(MatDenseGetLDA(newmat, &ldb));
2617: for (i = 0; i < ncols; i++) {
2618: av = v + mat->lda * icol[i];
2619: for (j = 0; j < nrows; j++) bv[j] = av[irow[j]];
2620: bv += ldb;
2621: }
2622: PetscCall(MatDenseRestoreArray(newmat, &bv));
2624: /* Assemble the matrices so that the correct flags are set */
2625: PetscCall(MatAssemblyBegin(newmat, MAT_FINAL_ASSEMBLY));
2626: PetscCall(MatAssemblyEnd(newmat, MAT_FINAL_ASSEMBLY));
2628: /* Free work space */
2629: PetscCall(ISRestoreIndices(isrow, &irow));
2630: PetscCall(ISRestoreIndices(iscol, &icol));
2631: *B = newmat;
2632: PetscFunctionReturn(PETSC_SUCCESS);
2633: }
2635: static PetscErrorCode MatCreateSubMatrices_SeqDense(Mat A, PetscInt n, const IS irow[], const IS icol[], MatReuse scall, Mat *B[])
2636: {
2637: PetscInt i;
2639: PetscFunctionBegin;
2640: if (scall == MAT_INITIAL_MATRIX) PetscCall(PetscCalloc1(n, B));
2642: for (i = 0; i < n; i++) PetscCall(MatCreateSubMatrix_SeqDense(A, irow[i], icol[i], scall, &(*B)[i]));
2643: PetscFunctionReturn(PETSC_SUCCESS);
2644: }
2646: PetscErrorCode MatCopy_SeqDense(Mat A, Mat B, MatStructure str)
2647: {
2648: Mat_SeqDense *a = (Mat_SeqDense *)A->data, *b = (Mat_SeqDense *)B->data;
2649: const PetscScalar *va;
2650: PetscScalar *vb;
2651: PetscInt lda1 = a->lda, lda2 = b->lda, m = A->rmap->n, n = A->cmap->n, j;
2653: PetscFunctionBegin;
2654: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2655: if (A->ops->copy != B->ops->copy) {
2656: PetscCall(MatCopy_Basic(A, B, str));
2657: PetscFunctionReturn(PETSC_SUCCESS);
2658: }
2659: PetscCheck(m == B->rmap->n && n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "size(B) != size(A)");
2660: PetscCall(MatDenseGetArrayRead(A, &va));
2661: PetscCall(MatDenseGetArray(B, &vb));
2662: if (lda1 > m || lda2 > m) {
2663: for (j = 0; j < n; j++) PetscCall(PetscArraycpy(vb + j * lda2, va + j * lda1, m));
2664: } else {
2665: PetscCall(PetscArraycpy(vb, va, A->rmap->n * A->cmap->n));
2666: }
2667: PetscCall(MatDenseRestoreArray(B, &vb));
2668: PetscCall(MatDenseRestoreArrayRead(A, &va));
2669: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
2670: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
2671: PetscFunctionReturn(PETSC_SUCCESS);
2672: }
2674: PetscErrorCode MatSetUp_SeqDense(Mat A)
2675: {
2676: PetscFunctionBegin;
2677: PetscCall(PetscLayoutSetUp(A->rmap));
2678: PetscCall(PetscLayoutSetUp(A->cmap));
2679: if (!A->preallocated) PetscCall(MatSeqDenseSetPreallocation(A, NULL));
2680: PetscFunctionReturn(PETSC_SUCCESS);
2681: }
2683: static PetscErrorCode MatConjugate_SeqDense(Mat A)
2684: {
2685: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2686: PetscInt i, j;
2687: PetscInt min = PetscMin(A->rmap->n, A->cmap->n);
2688: PetscScalar *aa;
2690: PetscFunctionBegin;
2691: PetscCall(MatDenseGetArray(A, &aa));
2692: for (j = 0; j < A->cmap->n; j++) {
2693: for (i = 0; i < A->rmap->n; i++) aa[i + j * mat->lda] = PetscConj(aa[i + j * mat->lda]);
2694: }
2695: PetscCall(MatDenseRestoreArray(A, &aa));
2696: if (mat->tau)
2697: for (i = 0; i < min; i++) mat->tau[i] = PetscConj(mat->tau[i]);
2698: PetscFunctionReturn(PETSC_SUCCESS);
2699: }
2701: static PetscErrorCode MatRealPart_SeqDense(Mat A)
2702: {
2703: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2704: PetscInt i, j;
2705: PetscScalar *aa;
2707: PetscFunctionBegin;
2708: PetscCall(MatDenseGetArray(A, &aa));
2709: for (j = 0; j < A->cmap->n; j++) {
2710: for (i = 0; i < A->rmap->n; i++) aa[i + j * mat->lda] = PetscRealPart(aa[i + j * mat->lda]);
2711: }
2712: PetscCall(MatDenseRestoreArray(A, &aa));
2713: PetscFunctionReturn(PETSC_SUCCESS);
2714: }
2716: static PetscErrorCode MatImaginaryPart_SeqDense(Mat A)
2717: {
2718: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2719: PetscInt i, j;
2720: PetscScalar *aa;
2722: PetscFunctionBegin;
2723: PetscCall(MatDenseGetArray(A, &aa));
2724: for (j = 0; j < A->cmap->n; j++) {
2725: for (i = 0; i < A->rmap->n; i++) aa[i + j * mat->lda] = PetscImaginaryPart(aa[i + j * mat->lda]);
2726: }
2727: PetscCall(MatDenseRestoreArray(A, &aa));
2728: PetscFunctionReturn(PETSC_SUCCESS);
2729: }
2731: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqDense(Mat A, Mat B, PetscReal fill, Mat C)
2732: {
2733: PetscInt m = A->rmap->n, n = B->cmap->n;
2734: PetscBool cisdense = PETSC_FALSE;
2736: PetscFunctionBegin;
2737: PetscCall(MatSetSizes(C, m, n, m, n));
2738: #if defined(PETSC_HAVE_CUDA)
2739: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, ""));
2740: #endif
2741: #if defined(PETSC_HAVE_HIP)
2742: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSEHIP, ""));
2743: #endif
2744: if (!cisdense) {
2745: PetscBool flg;
2747: PetscCall(PetscObjectTypeCompare((PetscObject)B, ((PetscObject)A)->type_name, &flg));
2748: PetscCall(MatSetType(C, flg ? ((PetscObject)A)->type_name : MATDENSE));
2749: }
2750: PetscCall(MatSetUp(C));
2751: PetscFunctionReturn(PETSC_SUCCESS);
2752: }
2754: PetscErrorCode MatMatMultNumeric_SeqDense_SeqDense(Mat A, Mat B, Mat C)
2755: {
2756: Mat_SeqDense *a = (Mat_SeqDense *)A->data, *b = (Mat_SeqDense *)B->data, *c = (Mat_SeqDense *)C->data;
2757: PetscBLASInt m, n, k;
2758: const PetscScalar *av, *bv;
2759: PetscScalar *cv;
2760: PetscScalar _DOne = 1.0, _DZero = 0.0;
2762: PetscFunctionBegin;
2763: PetscCall(PetscBLASIntCast(C->rmap->n, &m));
2764: PetscCall(PetscBLASIntCast(C->cmap->n, &n));
2765: PetscCall(PetscBLASIntCast(A->cmap->n, &k));
2766: if (!m || !n || !k) PetscFunctionReturn(PETSC_SUCCESS);
2767: PetscCall(MatDenseGetArrayRead(A, &av));
2768: PetscCall(MatDenseGetArrayRead(B, &bv));
2769: PetscCall(MatDenseGetArrayWrite(C, &cv));
2770: PetscCallBLAS("BLASgemm", BLASgemm_("N", "N", &m, &n, &k, &_DOne, av, &a->lda, bv, &b->lda, &_DZero, cv, &c->lda));
2771: PetscCall(PetscLogFlops(1.0 * m * n * k + 1.0 * m * n * (k - 1)));
2772: PetscCall(MatDenseRestoreArrayRead(A, &av));
2773: PetscCall(MatDenseRestoreArrayRead(B, &bv));
2774: PetscCall(MatDenseRestoreArrayWrite(C, &cv));
2775: PetscFunctionReturn(PETSC_SUCCESS);
2776: }
2778: PetscErrorCode MatMatTransposeMultSymbolic_SeqDense_SeqDense(Mat A, Mat B, PetscReal fill, Mat C)
2779: {
2780: PetscInt m = A->rmap->n, n = B->rmap->n;
2781: PetscBool cisdense = PETSC_FALSE;
2783: PetscFunctionBegin;
2784: PetscCall(MatSetSizes(C, m, n, m, n));
2785: #if defined(PETSC_HAVE_CUDA)
2786: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, ""));
2787: #endif
2788: #if defined(PETSC_HAVE_HIP)
2789: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSEHIP, ""));
2790: #endif
2791: if (!cisdense) {
2792: PetscBool flg;
2794: PetscCall(PetscObjectTypeCompare((PetscObject)B, ((PetscObject)A)->type_name, &flg));
2795: PetscCall(MatSetType(C, flg ? ((PetscObject)A)->type_name : MATDENSE));
2796: }
2797: PetscCall(MatSetUp(C));
2798: PetscFunctionReturn(PETSC_SUCCESS);
2799: }
2801: PetscErrorCode MatMatTransposeMultNumeric_SeqDense_SeqDense(Mat A, Mat B, Mat C)
2802: {
2803: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
2804: Mat_SeqDense *b = (Mat_SeqDense *)B->data;
2805: Mat_SeqDense *c = (Mat_SeqDense *)C->data;
2806: const PetscScalar *av, *bv;
2807: PetscScalar *cv;
2808: PetscBLASInt m, n, k;
2809: PetscScalar _DOne = 1.0, _DZero = 0.0;
2811: PetscFunctionBegin;
2812: PetscCall(PetscBLASIntCast(C->rmap->n, &m));
2813: PetscCall(PetscBLASIntCast(C->cmap->n, &n));
2814: PetscCall(PetscBLASIntCast(A->cmap->n, &k));
2815: if (!m || !n || !k) PetscFunctionReturn(PETSC_SUCCESS);
2816: PetscCall(MatDenseGetArrayRead(A, &av));
2817: PetscCall(MatDenseGetArrayRead(B, &bv));
2818: PetscCall(MatDenseGetArrayWrite(C, &cv));
2819: PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &m, &n, &k, &_DOne, av, &a->lda, bv, &b->lda, &_DZero, cv, &c->lda));
2820: PetscCall(MatDenseRestoreArrayRead(A, &av));
2821: PetscCall(MatDenseRestoreArrayRead(B, &bv));
2822: PetscCall(MatDenseRestoreArrayWrite(C, &cv));
2823: PetscCall(PetscLogFlops(1.0 * m * n * k + 1.0 * m * n * (k - 1)));
2824: PetscFunctionReturn(PETSC_SUCCESS);
2825: }
2827: PetscErrorCode MatTransposeMatMultSymbolic_SeqDense_SeqDense(Mat A, Mat B, PetscReal fill, Mat C)
2828: {
2829: PetscInt m = A->cmap->n, n = B->cmap->n;
2830: PetscBool cisdense = PETSC_FALSE;
2832: PetscFunctionBegin;
2833: PetscCall(MatSetSizes(C, m, n, m, n));
2834: #if defined(PETSC_HAVE_CUDA)
2835: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, ""));
2836: #endif
2837: #if defined(PETSC_HAVE_HIP)
2838: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSEHIP, ""));
2839: #endif
2840: if (!cisdense) {
2841: PetscBool flg;
2843: PetscCall(PetscObjectTypeCompare((PetscObject)B, ((PetscObject)A)->type_name, &flg));
2844: PetscCall(MatSetType(C, flg ? ((PetscObject)A)->type_name : MATDENSE));
2845: }
2846: PetscCall(MatSetUp(C));
2847: PetscFunctionReturn(PETSC_SUCCESS);
2848: }
2850: PetscErrorCode MatTransposeMatMultNumeric_SeqDense_SeqDense(Mat A, Mat B, Mat C)
2851: {
2852: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
2853: Mat_SeqDense *b = (Mat_SeqDense *)B->data;
2854: Mat_SeqDense *c = (Mat_SeqDense *)C->data;
2855: const PetscScalar *av, *bv;
2856: PetscScalar *cv;
2857: PetscBLASInt m, n, k;
2858: PetscScalar _DOne = 1.0, _DZero = 0.0;
2860: PetscFunctionBegin;
2861: PetscCall(PetscBLASIntCast(C->rmap->n, &m));
2862: PetscCall(PetscBLASIntCast(C->cmap->n, &n));
2863: PetscCall(PetscBLASIntCast(A->rmap->n, &k));
2864: if (!m || !n || !k) PetscFunctionReturn(PETSC_SUCCESS);
2865: PetscCall(MatDenseGetArrayRead(A, &av));
2866: PetscCall(MatDenseGetArrayRead(B, &bv));
2867: PetscCall(MatDenseGetArrayWrite(C, &cv));
2868: PetscCallBLAS("BLASgemm", BLASgemm_("T", "N", &m, &n, &k, &_DOne, av, &a->lda, bv, &b->lda, &_DZero, cv, &c->lda));
2869: PetscCall(MatDenseRestoreArrayRead(A, &av));
2870: PetscCall(MatDenseRestoreArrayRead(B, &bv));
2871: PetscCall(MatDenseRestoreArrayWrite(C, &cv));
2872: PetscCall(PetscLogFlops(1.0 * m * n * k + 1.0 * m * n * (k - 1)));
2873: PetscFunctionReturn(PETSC_SUCCESS);
2874: }
2876: static PetscErrorCode MatProductSetFromOptions_SeqDense_AB(Mat C)
2877: {
2878: PetscFunctionBegin;
2879: C->ops->matmultsymbolic = MatMatMultSymbolic_SeqDense_SeqDense;
2880: C->ops->productsymbolic = MatProductSymbolic_AB;
2881: PetscFunctionReturn(PETSC_SUCCESS);
2882: }
2884: static PetscErrorCode MatProductSetFromOptions_SeqDense_AtB(Mat C)
2885: {
2886: PetscFunctionBegin;
2887: C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_SeqDense_SeqDense;
2888: C->ops->productsymbolic = MatProductSymbolic_AtB;
2889: PetscFunctionReturn(PETSC_SUCCESS);
2890: }
2892: static PetscErrorCode MatProductSetFromOptions_SeqDense_ABt(Mat C)
2893: {
2894: PetscFunctionBegin;
2895: C->ops->mattransposemultsymbolic = MatMatTransposeMultSymbolic_SeqDense_SeqDense;
2896: C->ops->productsymbolic = MatProductSymbolic_ABt;
2897: PetscFunctionReturn(PETSC_SUCCESS);
2898: }
2900: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqDense(Mat C)
2901: {
2902: Mat_Product *product = C->product;
2904: PetscFunctionBegin;
2905: switch (product->type) {
2906: case MATPRODUCT_AB:
2907: PetscCall(MatProductSetFromOptions_SeqDense_AB(C));
2908: break;
2909: case MATPRODUCT_AtB:
2910: PetscCall(MatProductSetFromOptions_SeqDense_AtB(C));
2911: break;
2912: case MATPRODUCT_ABt:
2913: PetscCall(MatProductSetFromOptions_SeqDense_ABt(C));
2914: break;
2915: default:
2916: break;
2917: }
2918: PetscFunctionReturn(PETSC_SUCCESS);
2919: }
2921: static PetscErrorCode MatGetRowMax_SeqDense(Mat A, Vec v, PetscInt idx[])
2922: {
2923: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
2924: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, p;
2925: PetscScalar *x;
2926: const PetscScalar *aa;
2928: PetscFunctionBegin;
2929: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2930: PetscCall(VecGetArray(v, &x));
2931: PetscCall(VecGetLocalSize(v, &p));
2932: PetscCall(MatDenseGetArrayRead(A, &aa));
2933: PetscCheck(p == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
2934: for (i = 0; i < m; i++) {
2935: x[i] = aa[i];
2936: if (idx) idx[i] = 0;
2937: for (j = 1; j < n; j++) {
2938: if (PetscRealPart(x[i]) < PetscRealPart(aa[i + a->lda * j])) {
2939: x[i] = aa[i + a->lda * j];
2940: if (idx) idx[i] = j;
2941: }
2942: }
2943: }
2944: PetscCall(MatDenseRestoreArrayRead(A, &aa));
2945: PetscCall(VecRestoreArray(v, &x));
2946: PetscFunctionReturn(PETSC_SUCCESS);
2947: }
2949: static PetscErrorCode MatGetRowMaxAbs_SeqDense(Mat A, Vec v, PetscInt idx[])
2950: {
2951: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
2952: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, p;
2953: PetscScalar *x;
2954: PetscReal atmp;
2955: const PetscScalar *aa;
2957: PetscFunctionBegin;
2958: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2959: PetscCall(VecGetArray(v, &x));
2960: PetscCall(VecGetLocalSize(v, &p));
2961: PetscCall(MatDenseGetArrayRead(A, &aa));
2962: PetscCheck(p == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
2963: for (i = 0; i < m; i++) {
2964: x[i] = PetscAbsScalar(aa[i]);
2965: for (j = 1; j < n; j++) {
2966: atmp = PetscAbsScalar(aa[i + a->lda * j]);
2967: if (PetscAbsScalar(x[i]) < atmp) {
2968: x[i] = atmp;
2969: if (idx) idx[i] = j;
2970: }
2971: }
2972: }
2973: PetscCall(MatDenseRestoreArrayRead(A, &aa));
2974: PetscCall(VecRestoreArray(v, &x));
2975: PetscFunctionReturn(PETSC_SUCCESS);
2976: }
2978: static PetscErrorCode MatGetRowMin_SeqDense(Mat A, Vec v, PetscInt idx[])
2979: {
2980: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
2981: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, p;
2982: PetscScalar *x;
2983: const PetscScalar *aa;
2985: PetscFunctionBegin;
2986: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2987: PetscCall(MatDenseGetArrayRead(A, &aa));
2988: PetscCall(VecGetArray(v, &x));
2989: PetscCall(VecGetLocalSize(v, &p));
2990: PetscCheck(p == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
2991: for (i = 0; i < m; i++) {
2992: x[i] = aa[i];
2993: if (idx) idx[i] = 0;
2994: for (j = 1; j < n; j++) {
2995: if (PetscRealPart(x[i]) > PetscRealPart(aa[i + a->lda * j])) {
2996: x[i] = aa[i + a->lda * j];
2997: if (idx) idx[i] = j;
2998: }
2999: }
3000: }
3001: PetscCall(VecRestoreArray(v, &x));
3002: PetscCall(MatDenseRestoreArrayRead(A, &aa));
3003: PetscFunctionReturn(PETSC_SUCCESS);
3004: }
3006: PetscErrorCode MatGetColumnVector_SeqDense(Mat A, Vec v, PetscInt col)
3007: {
3008: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3009: PetscScalar *x;
3010: const PetscScalar *aa;
3012: PetscFunctionBegin;
3013: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3014: PetscCall(MatDenseGetArrayRead(A, &aa));
3015: PetscCall(VecGetArray(v, &x));
3016: PetscCall(PetscArraycpy(x, aa + col * a->lda, A->rmap->n));
3017: PetscCall(VecRestoreArray(v, &x));
3018: PetscCall(MatDenseRestoreArrayRead(A, &aa));
3019: PetscFunctionReturn(PETSC_SUCCESS);
3020: }
3022: PETSC_INTERN PetscErrorCode MatGetColumnReductions_SeqDense(Mat A, PetscInt type, PetscReal *reductions)
3023: {
3024: PetscInt i, j, m, n;
3025: const PetscScalar *a;
3027: PetscFunctionBegin;
3028: PetscCall(MatGetSize(A, &m, &n));
3029: PetscCall(PetscArrayzero(reductions, n));
3030: PetscCall(MatDenseGetArrayRead(A, &a));
3031: if (type == NORM_2) {
3032: for (i = 0; i < n; i++) {
3033: for (j = 0; j < m; j++) reductions[i] += PetscAbsScalar(a[j] * a[j]);
3034: a = PetscSafePointerPlusOffset(a, m);
3035: }
3036: } else if (type == NORM_1) {
3037: for (i = 0; i < n; i++) {
3038: for (j = 0; j < m; j++) reductions[i] += PetscAbsScalar(a[j]);
3039: a = PetscSafePointerPlusOffset(a, m);
3040: }
3041: } else if (type == NORM_INFINITY) {
3042: for (i = 0; i < n; i++) {
3043: for (j = 0; j < m; j++) reductions[i] = PetscMax(PetscAbsScalar(a[j]), reductions[i]);
3044: a = PetscSafePointerPlusOffset(a, m);
3045: }
3046: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
3047: for (i = 0; i < n; i++) {
3048: for (j = 0; j < m; j++) reductions[i] += PetscRealPart(a[j]);
3049: a = PetscSafePointerPlusOffset(a, m);
3050: }
3051: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
3052: for (i = 0; i < n; i++) {
3053: for (j = 0; j < m; j++) reductions[i] += PetscImaginaryPart(a[j]);
3054: a = PetscSafePointerPlusOffset(a, m);
3055: }
3056: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
3057: PetscCall(MatDenseRestoreArrayRead(A, &a));
3058: if (type == NORM_2) {
3059: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
3060: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
3061: for (i = 0; i < n; i++) reductions[i] /= m;
3062: }
3063: PetscFunctionReturn(PETSC_SUCCESS);
3064: }
3066: PetscErrorCode MatSetRandom_SeqDense(Mat x, PetscRandom rctx)
3067: {
3068: PetscScalar *a;
3069: PetscInt lda, m, n, i, j;
3071: PetscFunctionBegin;
3072: PetscCall(MatGetSize(x, &m, &n));
3073: PetscCall(MatDenseGetLDA(x, &lda));
3074: PetscCall(MatDenseGetArrayWrite(x, &a));
3075: for (j = 0; j < n; j++) {
3076: for (i = 0; i < m; i++) PetscCall(PetscRandomGetValue(rctx, a + j * lda + i));
3077: }
3078: PetscCall(MatDenseRestoreArrayWrite(x, &a));
3079: PetscFunctionReturn(PETSC_SUCCESS);
3080: }
3082: static PetscErrorCode MatMissingDiagonal_SeqDense(Mat A, PetscBool *missing, PetscInt *d)
3083: {
3084: PetscFunctionBegin;
3085: *missing = PETSC_FALSE;
3086: PetscFunctionReturn(PETSC_SUCCESS);
3087: }
3089: /* vals is not const */
3090: static PetscErrorCode MatDenseGetColumn_SeqDense(Mat A, PetscInt col, PetscScalar **vals)
3091: {
3092: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3093: PetscScalar *v;
3095: PetscFunctionBegin;
3096: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3097: PetscCall(MatDenseGetArray(A, &v));
3098: *vals = v + col * a->lda;
3099: PetscCall(MatDenseRestoreArray(A, &v));
3100: PetscFunctionReturn(PETSC_SUCCESS);
3101: }
3103: static PetscErrorCode MatDenseRestoreColumn_SeqDense(Mat A, PetscScalar **vals)
3104: {
3105: PetscFunctionBegin;
3106: if (vals) *vals = NULL; /* user cannot accidentally use the array later */
3107: PetscFunctionReturn(PETSC_SUCCESS);
3108: }
3110: static struct _MatOps MatOps_Values = {MatSetValues_SeqDense,
3111: MatGetRow_SeqDense,
3112: MatRestoreRow_SeqDense,
3113: MatMult_SeqDense,
3114: /* 4*/ MatMultAdd_SeqDense,
3115: MatMultTranspose_SeqDense,
3116: MatMultTransposeAdd_SeqDense,
3117: NULL,
3118: NULL,
3119: NULL,
3120: /* 10*/ NULL,
3121: MatLUFactor_SeqDense,
3122: MatCholeskyFactor_SeqDense,
3123: MatSOR_SeqDense,
3124: MatTranspose_SeqDense,
3125: /* 15*/ MatGetInfo_SeqDense,
3126: MatEqual_SeqDense,
3127: MatGetDiagonal_SeqDense,
3128: MatDiagonalScale_SeqDense,
3129: MatNorm_SeqDense,
3130: /* 20*/ NULL,
3131: NULL,
3132: MatSetOption_SeqDense,
3133: MatZeroEntries_SeqDense,
3134: /* 24*/ MatZeroRows_SeqDense,
3135: NULL,
3136: NULL,
3137: NULL,
3138: NULL,
3139: /* 29*/ MatSetUp_SeqDense,
3140: NULL,
3141: NULL,
3142: NULL,
3143: NULL,
3144: /* 34*/ MatDuplicate_SeqDense,
3145: NULL,
3146: NULL,
3147: NULL,
3148: NULL,
3149: /* 39*/ MatAXPY_SeqDense,
3150: MatCreateSubMatrices_SeqDense,
3151: NULL,
3152: MatGetValues_SeqDense,
3153: MatCopy_SeqDense,
3154: /* 44*/ MatGetRowMax_SeqDense,
3155: MatScale_SeqDense,
3156: MatShift_SeqDense,
3157: NULL,
3158: MatZeroRowsColumns_SeqDense,
3159: /* 49*/ MatSetRandom_SeqDense,
3160: NULL,
3161: NULL,
3162: NULL,
3163: NULL,
3164: /* 54*/ NULL,
3165: NULL,
3166: NULL,
3167: NULL,
3168: NULL,
3169: /* 59*/ MatCreateSubMatrix_SeqDense,
3170: MatDestroy_SeqDense,
3171: MatView_SeqDense,
3172: NULL,
3173: NULL,
3174: /* 64*/ NULL,
3175: NULL,
3176: NULL,
3177: NULL,
3178: NULL,
3179: /* 69*/ MatGetRowMaxAbs_SeqDense,
3180: NULL,
3181: NULL,
3182: NULL,
3183: NULL,
3184: /* 74*/ NULL,
3185: NULL,
3186: NULL,
3187: NULL,
3188: NULL,
3189: /* 79*/ NULL,
3190: NULL,
3191: NULL,
3192: NULL,
3193: /* 83*/ MatLoad_SeqDense,
3194: MatIsSymmetric_SeqDense,
3195: MatIsHermitian_SeqDense,
3196: NULL,
3197: NULL,
3198: NULL,
3199: /* 89*/ NULL,
3200: NULL,
3201: MatMatMultNumeric_SeqDense_SeqDense,
3202: NULL,
3203: NULL,
3204: /* 94*/ NULL,
3205: NULL,
3206: NULL,
3207: MatMatTransposeMultNumeric_SeqDense_SeqDense,
3208: NULL,
3209: /* 99*/ MatProductSetFromOptions_SeqDense,
3210: NULL,
3211: NULL,
3212: MatConjugate_SeqDense,
3213: NULL,
3214: /*104*/ NULL,
3215: MatRealPart_SeqDense,
3216: MatImaginaryPart_SeqDense,
3217: NULL,
3218: NULL,
3219: /*109*/ NULL,
3220: NULL,
3221: MatGetRowMin_SeqDense,
3222: MatGetColumnVector_SeqDense,
3223: MatMissingDiagonal_SeqDense,
3224: /*114*/ NULL,
3225: NULL,
3226: NULL,
3227: NULL,
3228: NULL,
3229: /*119*/ NULL,
3230: NULL,
3231: MatMultHermitianTranspose_SeqDense,
3232: MatMultHermitianTransposeAdd_SeqDense,
3233: NULL,
3234: /*124*/ NULL,
3235: MatGetColumnReductions_SeqDense,
3236: NULL,
3237: NULL,
3238: NULL,
3239: /*129*/ NULL,
3240: NULL,
3241: NULL,
3242: MatTransposeMatMultNumeric_SeqDense_SeqDense,
3243: NULL,
3244: /*134*/ NULL,
3245: NULL,
3246: NULL,
3247: NULL,
3248: NULL,
3249: /*139*/ NULL,
3250: NULL,
3251: NULL,
3252: NULL,
3253: NULL,
3254: MatCreateMPIMatConcatenateSeqMat_SeqDense,
3255: /*145*/ NULL,
3256: NULL,
3257: NULL,
3258: NULL,
3259: NULL,
3260: /*150*/ NULL,
3261: NULL,
3262: NULL,
3263: NULL,
3264: NULL,
3265: /*155*/ NULL,
3266: NULL};
3268: /*@
3269: MatCreateSeqDense - Creates a `MATSEQDENSE` that
3270: is stored in column major order (the usual Fortran format).
3272: Collective
3274: Input Parameters:
3275: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3276: . m - number of rows
3277: . n - number of columns
3278: - data - optional location of matrix data in column major order. Use `NULL` for PETSc
3279: to control all matrix memory allocation.
3281: Output Parameter:
3282: . A - the matrix
3284: Level: intermediate
3286: Note:
3287: The data input variable is intended primarily for Fortran programmers
3288: who wish to allocate their own matrix memory space. Most users should
3289: set `data` = `NULL`.
3291: Developer Note:
3292: Many of the matrix operations for this variant use the BLAS and LAPACK routines.
3294: .seealso: [](ch_matrices), `Mat`, `MATSEQDENSE`, `MatCreate()`, `MatCreateDense()`, `MatSetValues()`
3295: @*/
3296: PetscErrorCode MatCreateSeqDense(MPI_Comm comm, PetscInt m, PetscInt n, PetscScalar data[], Mat *A)
3297: {
3298: PetscFunctionBegin;
3299: PetscCall(MatCreate(comm, A));
3300: PetscCall(MatSetSizes(*A, m, n, m, n));
3301: PetscCall(MatSetType(*A, MATSEQDENSE));
3302: PetscCall(MatSeqDenseSetPreallocation(*A, data));
3303: PetscFunctionReturn(PETSC_SUCCESS);
3304: }
3306: /*@
3307: MatSeqDenseSetPreallocation - Sets the array used for storing the matrix elements of a `MATSEQDENSE` matrix
3309: Collective
3311: Input Parameters:
3312: + B - the matrix
3313: - data - the array (or `NULL`)
3315: Level: intermediate
3317: Note:
3318: The data input variable is intended primarily for Fortran programmers
3319: who wish to allocate their own matrix memory space. Most users should
3320: need not call this routine.
3322: .seealso: [](ch_matrices), `Mat`, `MATSEQDENSE`, `MatCreate()`, `MatCreateDense()`, `MatSetValues()`, `MatDenseSetLDA()`
3323: @*/
3324: PetscErrorCode MatSeqDenseSetPreallocation(Mat B, PetscScalar data[])
3325: {
3326: PetscFunctionBegin;
3328: PetscTryMethod(B, "MatSeqDenseSetPreallocation_C", (Mat, PetscScalar[]), (B, data));
3329: PetscFunctionReturn(PETSC_SUCCESS);
3330: }
3332: PetscErrorCode MatSeqDenseSetPreallocation_SeqDense(Mat B, PetscScalar *data)
3333: {
3334: Mat_SeqDense *b = (Mat_SeqDense *)B->data;
3336: PetscFunctionBegin;
3337: PetscCheck(!b->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3338: B->preallocated = PETSC_TRUE;
3340: PetscCall(PetscLayoutSetUp(B->rmap));
3341: PetscCall(PetscLayoutSetUp(B->cmap));
3343: if (b->lda <= 0) PetscCall(PetscBLASIntCast(B->rmap->n, &b->lda));
3345: if (!data) { /* petsc-allocated storage */
3346: if (!b->user_alloc) PetscCall(PetscFree(b->v));
3347: PetscCall(PetscCalloc1((size_t)b->lda * B->cmap->n, &b->v));
3349: b->user_alloc = PETSC_FALSE;
3350: } else { /* user-allocated storage */
3351: if (!b->user_alloc) PetscCall(PetscFree(b->v));
3352: b->v = data;
3353: b->user_alloc = PETSC_TRUE;
3354: }
3355: B->assembled = PETSC_TRUE;
3356: PetscFunctionReturn(PETSC_SUCCESS);
3357: }
3359: #if defined(PETSC_HAVE_ELEMENTAL)
3360: PETSC_INTERN PetscErrorCode MatConvert_SeqDense_Elemental(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
3361: {
3362: Mat mat_elemental;
3363: const PetscScalar *array;
3364: PetscScalar *v_colwise;
3365: PetscInt M = A->rmap->N, N = A->cmap->N, i, j, k, *rows, *cols;
3367: PetscFunctionBegin;
3368: PetscCall(PetscMalloc3(M * N, &v_colwise, M, &rows, N, &cols));
3369: PetscCall(MatDenseGetArrayRead(A, &array));
3370: /* convert column-wise array into row-wise v_colwise, see MatSetValues_Elemental() */
3371: k = 0;
3372: for (j = 0; j < N; j++) {
3373: cols[j] = j;
3374: for (i = 0; i < M; i++) v_colwise[j * M + i] = array[k++];
3375: }
3376: for (i = 0; i < M; i++) rows[i] = i;
3377: PetscCall(MatDenseRestoreArrayRead(A, &array));
3379: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental));
3380: PetscCall(MatSetSizes(mat_elemental, PETSC_DECIDE, PETSC_DECIDE, M, N));
3381: PetscCall(MatSetType(mat_elemental, MATELEMENTAL));
3382: PetscCall(MatSetUp(mat_elemental));
3384: /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
3385: PetscCall(MatSetValues(mat_elemental, M, rows, N, cols, v_colwise, ADD_VALUES));
3386: PetscCall(MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY));
3387: PetscCall(MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY));
3388: PetscCall(PetscFree3(v_colwise, rows, cols));
3390: if (reuse == MAT_INPLACE_MATRIX) {
3391: PetscCall(MatHeaderReplace(A, &mat_elemental));
3392: } else {
3393: *newmat = mat_elemental;
3394: }
3395: PetscFunctionReturn(PETSC_SUCCESS);
3396: }
3397: #endif
3399: PetscErrorCode MatDenseSetLDA_SeqDense(Mat B, PetscInt lda)
3400: {
3401: Mat_SeqDense *b = (Mat_SeqDense *)B->data;
3402: PetscBool data;
3404: PetscFunctionBegin;
3405: data = (B->rmap->n > 0 && B->cmap->n > 0) ? (b->v ? PETSC_TRUE : PETSC_FALSE) : PETSC_FALSE;
3406: PetscCheck(b->user_alloc || !data || b->lda == lda, PETSC_COMM_SELF, PETSC_ERR_ORDER, "LDA cannot be changed after allocation of internal storage");
3407: PetscCheck(lda >= B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "LDA %" PetscInt_FMT " must be at least matrix dimension %" PetscInt_FMT, lda, B->rmap->n);
3408: PetscCall(PetscBLASIntCast(lda, &b->lda));
3409: PetscFunctionReturn(PETSC_SUCCESS);
3410: }
3412: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqDense(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
3413: {
3414: PetscFunctionBegin;
3415: PetscCall(MatCreateMPIMatConcatenateSeqMat_MPIDense(comm, inmat, n, scall, outmat));
3416: PetscFunctionReturn(PETSC_SUCCESS);
3417: }
3419: PetscErrorCode MatDenseCreateColumnVec_Private(Mat A, Vec *v)
3420: {
3421: PetscBool isstd, iskok, iscuda, iship;
3422: PetscMPIInt size;
3423: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
3424: /* we pass the data of A, to prevent allocating needless GPU memory the first time VecCUPMPlaceArray is called. */
3425: const PetscScalar *a;
3426: #endif
3428: PetscFunctionBegin;
3429: *v = NULL;
3430: PetscCall(PetscStrcmpAny(A->defaultvectype, &isstd, VECSTANDARD, VECSEQ, VECMPI, ""));
3431: PetscCall(PetscStrcmpAny(A->defaultvectype, &iskok, VECKOKKOS, VECSEQKOKKOS, VECMPIKOKKOS, ""));
3432: PetscCall(PetscStrcmpAny(A->defaultvectype, &iscuda, VECCUDA, VECSEQCUDA, VECMPICUDA, ""));
3433: PetscCall(PetscStrcmpAny(A->defaultvectype, &iship, VECHIP, VECSEQHIP, VECMPIHIP, ""));
3434: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
3435: if (isstd) {
3436: if (size > 1) PetscCall(VecCreateMPIWithArray(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, A->rmap->N, NULL, v));
3437: else PetscCall(VecCreateSeqWithArray(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, NULL, v));
3438: } else if (iskok) {
3439: PetscCheck(PetscDefined(HAVE_KOKKOS_KERNELS), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Reconfigure using KOKKOS kernels support");
3440: #if PetscDefined(HAVE_KOKKOS_KERNELS)
3441: if (size > 1) PetscCall(VecCreateMPIKokkosWithArray(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, A->rmap->N, NULL, v));
3442: else PetscCall(VecCreateSeqKokkosWithArray(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, NULL, v));
3443: #endif
3444: } else if (iscuda) {
3445: PetscCheck(PetscDefined(HAVE_CUDA), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Reconfigure using CUDA support");
3446: #if PetscDefined(HAVE_CUDA)
3447: PetscCall(MatDenseCUDAGetArrayRead(A, &a));
3448: if (size > 1) PetscCall(VecCreateMPICUDAWithArrays(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, A->rmap->N, NULL, a, v));
3449: else PetscCall(VecCreateSeqCUDAWithArrays(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, NULL, a, v));
3450: #endif
3451: } else if (iship) {
3452: PetscCheck(PetscDefined(HAVE_HIP), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Reconfigure using HIP support");
3453: #if PetscDefined(HAVE_HIP)
3454: PetscCall(MatDenseHIPGetArrayRead(A, &a));
3455: if (size > 1) PetscCall(VecCreateMPIHIPWithArrays(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, A->rmap->N, NULL, a, v));
3456: else PetscCall(VecCreateSeqHIPWithArrays(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, NULL, a, v));
3457: #endif
3458: }
3459: PetscCheck(*v, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not coded for type %s", A->defaultvectype);
3460: PetscFunctionReturn(PETSC_SUCCESS);
3461: }
3463: PetscErrorCode MatDenseGetColumnVec_SeqDense(Mat A, PetscInt col, Vec *v)
3464: {
3465: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3467: PetscFunctionBegin;
3468: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
3469: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3470: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
3471: a->vecinuse = col + 1;
3472: PetscCall(MatDenseGetArray(A, (PetscScalar **)&a->ptrinuse));
3473: PetscCall(VecPlaceArray(a->cvec, a->ptrinuse + (size_t)col * (size_t)a->lda));
3474: *v = a->cvec;
3475: PetscFunctionReturn(PETSC_SUCCESS);
3476: }
3478: PetscErrorCode MatDenseRestoreColumnVec_SeqDense(Mat A, PetscInt col, Vec *v)
3479: {
3480: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3482: PetscFunctionBegin;
3483: PetscCheck(a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
3484: PetscCheck(a->cvec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column vector");
3485: VecCheckAssembled(a->cvec);
3486: a->vecinuse = 0;
3487: PetscCall(MatDenseRestoreArray(A, (PetscScalar **)&a->ptrinuse));
3488: PetscCall(VecResetArray(a->cvec));
3489: if (v) *v = NULL;
3490: PetscFunctionReturn(PETSC_SUCCESS);
3491: }
3493: PetscErrorCode MatDenseGetColumnVecRead_SeqDense(Mat A, PetscInt col, Vec *v)
3494: {
3495: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3497: PetscFunctionBegin;
3498: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
3499: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3500: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
3501: a->vecinuse = col + 1;
3502: PetscCall(MatDenseGetArrayRead(A, &a->ptrinuse));
3503: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)a->lda)));
3504: PetscCall(VecLockReadPush(a->cvec));
3505: *v = a->cvec;
3506: PetscFunctionReturn(PETSC_SUCCESS);
3507: }
3509: PetscErrorCode MatDenseRestoreColumnVecRead_SeqDense(Mat A, PetscInt col, Vec *v)
3510: {
3511: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3513: PetscFunctionBegin;
3514: PetscCheck(a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
3515: PetscCheck(a->cvec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column vector");
3516: VecCheckAssembled(a->cvec);
3517: a->vecinuse = 0;
3518: PetscCall(MatDenseRestoreArrayRead(A, &a->ptrinuse));
3519: PetscCall(VecLockReadPop(a->cvec));
3520: PetscCall(VecResetArray(a->cvec));
3521: if (v) *v = NULL;
3522: PetscFunctionReturn(PETSC_SUCCESS);
3523: }
3525: PetscErrorCode MatDenseGetColumnVecWrite_SeqDense(Mat A, PetscInt col, Vec *v)
3526: {
3527: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3529: PetscFunctionBegin;
3530: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
3531: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3532: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
3533: a->vecinuse = col + 1;
3534: PetscCall(MatDenseGetArrayWrite(A, (PetscScalar **)&a->ptrinuse));
3535: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)a->lda)));
3536: *v = a->cvec;
3537: PetscFunctionReturn(PETSC_SUCCESS);
3538: }
3540: PetscErrorCode MatDenseRestoreColumnVecWrite_SeqDense(Mat A, PetscInt col, Vec *v)
3541: {
3542: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3544: PetscFunctionBegin;
3545: PetscCheck(a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
3546: PetscCheck(a->cvec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column vector");
3547: VecCheckAssembled(a->cvec);
3548: a->vecinuse = 0;
3549: PetscCall(MatDenseRestoreArrayWrite(A, (PetscScalar **)&a->ptrinuse));
3550: PetscCall(VecResetArray(a->cvec));
3551: if (v) *v = NULL;
3552: PetscFunctionReturn(PETSC_SUCCESS);
3553: }
3555: PetscErrorCode MatDenseGetSubMatrix_SeqDense(Mat A, PetscInt rbegin, PetscInt rend, PetscInt cbegin, PetscInt cend, Mat *v)
3556: {
3557: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3559: PetscFunctionBegin;
3560: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
3561: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3562: if (a->cmat && (cend - cbegin != a->cmat->cmap->N || rend - rbegin != a->cmat->rmap->N)) PetscCall(MatDestroy(&a->cmat));
3563: if (!a->cmat) {
3564: PetscCall(MatCreateDense(PetscObjectComm((PetscObject)A), rend - rbegin, PETSC_DECIDE, rend - rbegin, cend - cbegin, PetscSafePointerPlusOffset(a->v, rbegin + (size_t)cbegin * a->lda), &a->cmat));
3565: } else {
3566: PetscCall(MatDensePlaceArray(a->cmat, PetscSafePointerPlusOffset(a->v, rbegin + (size_t)cbegin * a->lda)));
3567: }
3568: PetscCall(MatDenseSetLDA(a->cmat, a->lda));
3569: a->matinuse = cbegin + 1;
3570: *v = a->cmat;
3571: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
3572: A->offloadmask = PETSC_OFFLOAD_CPU;
3573: #endif
3574: PetscFunctionReturn(PETSC_SUCCESS);
3575: }
3577: PetscErrorCode MatDenseRestoreSubMatrix_SeqDense(Mat A, Mat *v)
3578: {
3579: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3581: PetscFunctionBegin;
3582: PetscCheck(a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetSubMatrix() first");
3583: PetscCheck(a->cmat, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column matrix");
3584: PetscCheck(*v == a->cmat, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Not the matrix obtained from MatDenseGetSubMatrix()");
3585: a->matinuse = 0;
3586: PetscCall(MatDenseResetArray(a->cmat));
3587: if (v) *v = NULL;
3588: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
3589: A->offloadmask = PETSC_OFFLOAD_CPU;
3590: #endif
3591: PetscFunctionReturn(PETSC_SUCCESS);
3592: }
3594: /*MC
3595: MATSEQDENSE - MATSEQDENSE = "seqdense" - A matrix type to be used for sequential dense matrices.
3597: Options Database Key:
3598: . -mat_type seqdense - sets the matrix type to `MATSEQDENSE` during a call to `MatSetFromOptions()`
3600: Level: beginner
3602: .seealso: [](ch_matrices), `Mat`, `MATSEQDENSE`, `MatCreateSeqDense()`
3603: M*/
3604: PetscErrorCode MatCreate_SeqDense(Mat B)
3605: {
3606: Mat_SeqDense *b;
3607: PetscMPIInt size;
3609: PetscFunctionBegin;
3610: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
3611: PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Comm must be of size 1");
3613: PetscCall(PetscNew(&b));
3614: B->data = (void *)b;
3615: B->ops[0] = MatOps_Values;
3617: b->roworiented = PETSC_TRUE;
3619: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatQRFactor_C", MatQRFactor_SeqDense));
3620: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetLDA_C", MatDenseGetLDA_SeqDense));
3621: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseSetLDA_C", MatDenseSetLDA_SeqDense));
3622: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetArray_C", MatDenseGetArray_SeqDense));
3623: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreArray_C", MatDenseRestoreArray_SeqDense));
3624: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDensePlaceArray_C", MatDensePlaceArray_SeqDense));
3625: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseResetArray_C", MatDenseResetArray_SeqDense));
3626: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseReplaceArray_C", MatDenseReplaceArray_SeqDense));
3627: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetArrayRead_C", MatDenseGetArray_SeqDense));
3628: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreArrayRead_C", MatDenseRestoreArray_SeqDense));
3629: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetArrayWrite_C", MatDenseGetArray_SeqDense));
3630: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreArrayWrite_C", MatDenseRestoreArray_SeqDense));
3631: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_seqaij_C", MatConvert_SeqDense_SeqAIJ));
3632: #if defined(PETSC_HAVE_ELEMENTAL)
3633: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_elemental_C", MatConvert_SeqDense_Elemental));
3634: #endif
3635: #if defined(PETSC_HAVE_SCALAPACK)
3636: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_scalapack_C", MatConvert_Dense_ScaLAPACK));
3637: #endif
3638: #if defined(PETSC_HAVE_CUDA)
3639: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_seqdensecuda_C", MatConvert_SeqDense_SeqDenseCUDA));
3640: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdensecuda_seqdensecuda_C", MatProductSetFromOptions_SeqDense));
3641: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdensecuda_seqdense_C", MatProductSetFromOptions_SeqDense));
3642: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqdensecuda_C", MatProductSetFromOptions_SeqDense));
3643: #endif
3644: #if defined(PETSC_HAVE_HIP)
3645: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_seqdensehip_C", MatConvert_SeqDense_SeqDenseHIP));
3646: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdensehip_seqdensehip_C", MatProductSetFromOptions_SeqDense));
3647: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdensehip_seqdense_C", MatProductSetFromOptions_SeqDense));
3648: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqdensehip_C", MatProductSetFromOptions_SeqDense));
3649: #endif
3650: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqDenseSetPreallocation_C", MatSeqDenseSetPreallocation_SeqDense));
3651: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqdense_C", MatProductSetFromOptions_SeqAIJ_SeqDense));
3652: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqdense_C", MatProductSetFromOptions_SeqDense));
3653: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqbaij_seqdense_C", MatProductSetFromOptions_SeqXBAIJ_SeqDense));
3654: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqsbaij_seqdense_C", MatProductSetFromOptions_SeqXBAIJ_SeqDense));
3656: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetColumn_C", MatDenseGetColumn_SeqDense));
3657: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreColumn_C", MatDenseRestoreColumn_SeqDense));
3658: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetColumnVec_C", MatDenseGetColumnVec_SeqDense));
3659: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreColumnVec_C", MatDenseRestoreColumnVec_SeqDense));
3660: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetColumnVecRead_C", MatDenseGetColumnVecRead_SeqDense));
3661: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreColumnVecRead_C", MatDenseRestoreColumnVecRead_SeqDense));
3662: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetColumnVecWrite_C", MatDenseGetColumnVecWrite_SeqDense));
3663: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreColumnVecWrite_C", MatDenseRestoreColumnVecWrite_SeqDense));
3664: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetSubMatrix_C", MatDenseGetSubMatrix_SeqDense));
3665: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreSubMatrix_C", MatDenseRestoreSubMatrix_SeqDense));
3666: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMultAddColumnRange_C", MatMultAddColumnRange_SeqDense));
3667: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMultHermitianTransposeColumnRange_C", MatMultHermitianTransposeColumnRange_SeqDense));
3668: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMultHermitianTransposeAddColumnRange_C", MatMultHermitianTransposeAddColumnRange_SeqDense));
3669: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQDENSE));
3670: PetscFunctionReturn(PETSC_SUCCESS);
3671: }
3673: /*@C
3674: MatDenseGetColumn - gives access to a column of a dense matrix. This is only the local part of the column. You MUST call `MatDenseRestoreColumn()` to avoid memory bleeding.
3676: Not Collective
3678: Input Parameters:
3679: + A - a `MATSEQDENSE` or `MATMPIDENSE` matrix
3680: - col - column index
3682: Output Parameter:
3683: . vals - pointer to the data
3685: Level: intermediate
3687: Note:
3688: Use `MatDenseGetColumnVec()` to get access to a column of a `MATDENSE` treated as a `Vec`
3690: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreColumn()`, `MatDenseGetColumnVec()`
3691: @*/
3692: PetscErrorCode MatDenseGetColumn(Mat A, PetscInt col, PetscScalar *vals[])
3693: {
3694: PetscFunctionBegin;
3697: PetscAssertPointer(vals, 3);
3698: PetscUseMethod(A, "MatDenseGetColumn_C", (Mat, PetscInt, PetscScalar **), (A, col, vals));
3699: PetscFunctionReturn(PETSC_SUCCESS);
3700: }
3702: /*@C
3703: MatDenseRestoreColumn - returns access to a column of a `MATDENSE` matrix which is returned by `MatDenseGetColumn()`.
3705: Not Collective
3707: Input Parameters:
3708: + A - a `MATSEQDENSE` or `MATMPIDENSE` matrix
3709: - vals - pointer to the data (may be `NULL`)
3711: Level: intermediate
3713: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetColumn()`
3714: @*/
3715: PetscErrorCode MatDenseRestoreColumn(Mat A, PetscScalar *vals[])
3716: {
3717: PetscFunctionBegin;
3719: PetscAssertPointer(vals, 2);
3720: PetscUseMethod(A, "MatDenseRestoreColumn_C", (Mat, PetscScalar **), (A, vals));
3721: PetscFunctionReturn(PETSC_SUCCESS);
3722: }
3724: /*@
3725: MatDenseGetColumnVec - Gives read-write access to a column of a `MATDENSE` matrix, represented as a `Vec`.
3727: Collective
3729: Input Parameters:
3730: + A - the `Mat` object
3731: - col - the column index
3733: Output Parameter:
3734: . v - the vector
3736: Level: intermediate
3738: Notes:
3739: The vector is owned by PETSc. Users need to call `MatDenseRestoreColumnVec()` when the vector is no longer needed.
3741: Use `MatDenseGetColumnVecRead()` to obtain read-only access or `MatDenseGetColumnVecWrite()` for write-only access.
3743: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVecRead()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecRead()`, `MatDenseRestoreColumnVecWrite()`, `MatDenseGetColumn()`
3744: @*/
3745: PetscErrorCode MatDenseGetColumnVec(Mat A, PetscInt col, Vec *v)
3746: {
3747: PetscFunctionBegin;
3751: PetscAssertPointer(v, 3);
3752: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3753: PetscCheck(col >= 0 && col < A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid col %" PetscInt_FMT ", should be in [0,%" PetscInt_FMT ")", col, A->cmap->N);
3754: PetscUseMethod(A, "MatDenseGetColumnVec_C", (Mat, PetscInt, Vec *), (A, col, v));
3755: PetscFunctionReturn(PETSC_SUCCESS);
3756: }
3758: /*@
3759: MatDenseRestoreColumnVec - Returns access to a column of a dense matrix obtained from `MatDenseGetColumnVec()`.
3761: Collective
3763: Input Parameters:
3764: + A - the `Mat` object
3765: . col - the column index
3766: - v - the `Vec` object (may be `NULL`)
3768: Level: intermediate
3770: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecRead()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVecRead()`, `MatDenseRestoreColumnVecWrite()`
3771: @*/
3772: PetscErrorCode MatDenseRestoreColumnVec(Mat A, PetscInt col, Vec *v)
3773: {
3774: PetscFunctionBegin;
3778: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3779: PetscCheck(col >= 0 && col < A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid col %" PetscInt_FMT ", should be in [0,%" PetscInt_FMT ")", col, A->cmap->N);
3780: PetscUseMethod(A, "MatDenseRestoreColumnVec_C", (Mat, PetscInt, Vec *), (A, col, v));
3781: PetscFunctionReturn(PETSC_SUCCESS);
3782: }
3784: /*@
3785: MatDenseGetColumnVecRead - Gives read-only access to a column of a dense matrix, represented as a `Vec`.
3787: Collective
3789: Input Parameters:
3790: + A - the `Mat` object
3791: - col - the column index
3793: Output Parameter:
3794: . v - the vector
3796: Level: intermediate
3798: Notes:
3799: The vector is owned by PETSc and users cannot modify it.
3801: Users need to call `MatDenseRestoreColumnVecRead()` when the vector is no longer needed.
3803: Use `MatDenseGetColumnVec()` to obtain read-write access or `MatDenseGetColumnVecWrite()` for write-only access.
3805: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecRead()`, `MatDenseRestoreColumnVecWrite()`
3806: @*/
3807: PetscErrorCode MatDenseGetColumnVecRead(Mat A, PetscInt col, Vec *v)
3808: {
3809: PetscFunctionBegin;
3813: PetscAssertPointer(v, 3);
3814: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3815: PetscCheck(col >= 0 && col < A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid col %" PetscInt_FMT ", should be in [0,%" PetscInt_FMT ")", col, A->cmap->N);
3816: PetscUseMethod(A, "MatDenseGetColumnVecRead_C", (Mat, PetscInt, Vec *), (A, col, v));
3817: PetscFunctionReturn(PETSC_SUCCESS);
3818: }
3820: /*@
3821: MatDenseRestoreColumnVecRead - Returns access to a column of a dense matrix obtained from `MatDenseGetColumnVecRead()`.
3823: Collective
3825: Input Parameters:
3826: + A - the `Mat` object
3827: . col - the column index
3828: - v - the `Vec` object (may be `NULL`)
3830: Level: intermediate
3832: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecRead()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecWrite()`
3833: @*/
3834: PetscErrorCode MatDenseRestoreColumnVecRead(Mat A, PetscInt col, Vec *v)
3835: {
3836: PetscFunctionBegin;
3840: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3841: PetscCheck(col >= 0 && col < A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid col %" PetscInt_FMT ", should be in [0,%" PetscInt_FMT ")", col, A->cmap->N);
3842: PetscUseMethod(A, "MatDenseRestoreColumnVecRead_C", (Mat, PetscInt, Vec *), (A, col, v));
3843: PetscFunctionReturn(PETSC_SUCCESS);
3844: }
3846: /*@
3847: MatDenseGetColumnVecWrite - Gives write-only access to a column of a dense matrix, represented as a `Vec`.
3849: Collective
3851: Input Parameters:
3852: + A - the `Mat` object
3853: - col - the column index
3855: Output Parameter:
3856: . v - the vector
3858: Level: intermediate
3860: Notes:
3861: The vector is owned by PETSc. Users need to call `MatDenseRestoreColumnVecWrite()` when the vector is no longer needed.
3863: Use `MatDenseGetColumnVec()` to obtain read-write access or `MatDenseGetColumnVecRead()` for read-only access.
3865: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecRead()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecRead()`, `MatDenseRestoreColumnVecWrite()`
3866: @*/
3867: PetscErrorCode MatDenseGetColumnVecWrite(Mat A, PetscInt col, Vec *v)
3868: {
3869: PetscFunctionBegin;
3873: PetscAssertPointer(v, 3);
3874: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3875: PetscCheck(col >= 0 && col < A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid col %" PetscInt_FMT ", should be in [0,%" PetscInt_FMT ")", col, A->cmap->N);
3876: PetscUseMethod(A, "MatDenseGetColumnVecWrite_C", (Mat, PetscInt, Vec *), (A, col, v));
3877: PetscFunctionReturn(PETSC_SUCCESS);
3878: }
3880: /*@
3881: MatDenseRestoreColumnVecWrite - Returns access to a column of a dense matrix obtained from `MatDenseGetColumnVecWrite()`.
3883: Collective
3885: Input Parameters:
3886: + A - the `Mat` object
3887: . col - the column index
3888: - v - the `Vec` object (may be `NULL`)
3890: Level: intermediate
3892: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecRead()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecRead()`
3893: @*/
3894: PetscErrorCode MatDenseRestoreColumnVecWrite(Mat A, PetscInt col, Vec *v)
3895: {
3896: PetscFunctionBegin;
3900: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3901: PetscCheck(col >= 0 && col < A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid col %" PetscInt_FMT ", should be in [0,%" PetscInt_FMT ")", col, A->cmap->N);
3902: PetscUseMethod(A, "MatDenseRestoreColumnVecWrite_C", (Mat, PetscInt, Vec *), (A, col, v));
3903: PetscFunctionReturn(PETSC_SUCCESS);
3904: }
3906: /*@
3907: MatDenseGetSubMatrix - Gives access to a block of rows and columns of a dense matrix, represented as a `Mat`.
3909: Collective
3911: Input Parameters:
3912: + A - the `Mat` object
3913: . rbegin - the first global row index in the block (if `PETSC_DECIDE`, is 0)
3914: . rend - the global row index past the last one in the block (if `PETSC_DECIDE`, is `M`)
3915: . cbegin - the first global column index in the block (if `PETSC_DECIDE`, is 0)
3916: - cend - the global column index past the last one in the block (if `PETSC_DECIDE`, is `N`)
3918: Output Parameter:
3919: . v - the matrix
3921: Level: intermediate
3923: Notes:
3924: The matrix is owned by PETSc. Users need to call `MatDenseRestoreSubMatrix()` when the matrix is no longer needed.
3926: The output matrix is not redistributed by PETSc, so depending on the values of `rbegin` and `rend`, some processes may have no local rows.
3928: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreSubMatrix()`
3929: @*/
3930: PetscErrorCode MatDenseGetSubMatrix(Mat A, PetscInt rbegin, PetscInt rend, PetscInt cbegin, PetscInt cend, Mat *v)
3931: {
3932: PetscFunctionBegin;
3939: PetscAssertPointer(v, 6);
3940: if (rbegin == PETSC_DECIDE) rbegin = 0;
3941: if (rend == PETSC_DECIDE) rend = A->rmap->N;
3942: if (cbegin == PETSC_DECIDE) cbegin = 0;
3943: if (cend == PETSC_DECIDE) cend = A->cmap->N;
3944: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3945: PetscCheck(rbegin >= 0 && rbegin <= A->rmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid rbegin %" PetscInt_FMT ", should be in [0,%" PetscInt_FMT "]", rbegin, A->rmap->N);
3946: PetscCheck(rend >= rbegin && rend <= A->rmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid rend %" PetscInt_FMT ", should be in [%" PetscInt_FMT ",%" PetscInt_FMT "]", rend, rbegin, A->rmap->N);
3947: PetscCheck(cbegin >= 0 && cbegin <= A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid cbegin %" PetscInt_FMT ", should be in [0,%" PetscInt_FMT "]", cbegin, A->cmap->N);
3948: PetscCheck(cend >= cbegin && cend <= A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid cend %" PetscInt_FMT ", should be in [%" PetscInt_FMT ",%" PetscInt_FMT "]", cend, cbegin, A->cmap->N);
3949: PetscUseMethod(A, "MatDenseGetSubMatrix_C", (Mat, PetscInt, PetscInt, PetscInt, PetscInt, Mat *), (A, rbegin, rend, cbegin, cend, v));
3950: PetscFunctionReturn(PETSC_SUCCESS);
3951: }
3953: /*@
3954: MatDenseRestoreSubMatrix - Returns access to a block of columns of a dense matrix obtained from `MatDenseGetSubMatrix()`.
3956: Collective
3958: Input Parameters:
3959: + A - the `Mat` object
3960: - v - the `Mat` object (may be `NULL`)
3962: Level: intermediate
3964: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseRestoreColumnVec()`, `MatDenseGetSubMatrix()`
3965: @*/
3966: PetscErrorCode MatDenseRestoreSubMatrix(Mat A, Mat *v)
3967: {
3968: PetscFunctionBegin;
3971: PetscAssertPointer(v, 2);
3972: PetscUseMethod(A, "MatDenseRestoreSubMatrix_C", (Mat, Mat *), (A, v));
3973: PetscFunctionReturn(PETSC_SUCCESS);
3974: }
3976: #include <petscblaslapack.h>
3977: #include <petsc/private/kernels/blockinvert.h>
3979: PetscErrorCode MatSeqDenseInvert(Mat A)
3980: {
3981: PetscInt m;
3982: const PetscReal shift = 0.0;
3983: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
3984: PetscScalar *values;
3986: PetscFunctionBegin;
3988: PetscCall(MatDenseGetArray(A, &values));
3989: PetscCall(MatGetLocalSize(A, &m, NULL));
3990: allowzeropivot = PetscNot(A->erroriffailure);
3991: /* factor and invert each block */
3992: switch (m) {
3993: case 1:
3994: values[0] = (PetscScalar)1.0 / (values[0] + shift);
3995: break;
3996: case 2:
3997: PetscCall(PetscKernel_A_gets_inverse_A_2(values, shift, allowzeropivot, &zeropivotdetected));
3998: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3999: break;
4000: case 3:
4001: PetscCall(PetscKernel_A_gets_inverse_A_3(values, shift, allowzeropivot, &zeropivotdetected));
4002: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4003: break;
4004: case 4:
4005: PetscCall(PetscKernel_A_gets_inverse_A_4(values, shift, allowzeropivot, &zeropivotdetected));
4006: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4007: break;
4008: case 5: {
4009: PetscScalar work[25];
4010: PetscInt ipvt[5];
4012: PetscCall(PetscKernel_A_gets_inverse_A_5(values, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
4013: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4014: } break;
4015: case 6:
4016: PetscCall(PetscKernel_A_gets_inverse_A_6(values, shift, allowzeropivot, &zeropivotdetected));
4017: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4018: break;
4019: case 7:
4020: PetscCall(PetscKernel_A_gets_inverse_A_7(values, shift, allowzeropivot, &zeropivotdetected));
4021: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4022: break;
4023: default: {
4024: PetscInt *v_pivots, *IJ, j;
4025: PetscScalar *v_work;
4027: PetscCall(PetscMalloc3(m, &v_work, m, &v_pivots, m, &IJ));
4028: for (j = 0; j < m; j++) IJ[j] = j;
4029: PetscCall(PetscKernel_A_gets_inverse_A(m, values, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
4030: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4031: PetscCall(PetscFree3(v_work, v_pivots, IJ));
4032: }
4033: }
4034: PetscCall(MatDenseRestoreArray(A, &values));
4035: PetscFunctionReturn(PETSC_SUCCESS);
4036: }