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 >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Error in LAPACK argument %" PetscBLASInt_FMT, -info);
77: PetscCheck(info <= 0, PETSC_COMM_SELF, PETSC_ERR_MAT_CH_ZRPVT, "Bad Inversion: zero pivot in row %" PetscBLASInt_FMT, info - 1);
78: PetscCall(PetscLogFlops((1.0 * A->cmap->n * A->cmap->n * A->cmap->n) / 3.0));
79: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix must be factored to solve");
81: A->ops->solve = NULL;
82: A->ops->matsolve = NULL;
83: A->ops->solvetranspose = NULL;
84: A->ops->matsolvetranspose = NULL;
85: A->ops->solveadd = NULL;
86: A->ops->solvetransposeadd = NULL;
87: A->factortype = MAT_FACTOR_NONE;
88: PetscCall(PetscFree(A->solvertype));
89: PetscFunctionReturn(PETSC_SUCCESS);
90: }
92: static PetscErrorCode MatZeroRowsColumns_SeqDense(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
93: {
94: Mat_SeqDense *l = (Mat_SeqDense *)A->data;
95: PetscInt m = l->lda, n = A->cmap->n, r = A->rmap->n, i, j;
96: PetscScalar *slot, *bb, *v;
97: const PetscScalar *xx;
99: PetscFunctionBegin;
100: if (PetscDefined(USE_DEBUG)) {
101: for (i = 0; i < N; i++) {
102: PetscCheck(rows[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative row requested to be zeroed");
103: 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);
104: 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);
105: }
106: }
107: if (!N) PetscFunctionReturn(PETSC_SUCCESS);
109: /* fix right-hand side if needed */
110: if (x && b) {
111: Vec xt;
113: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only coded for square matrices");
114: PetscCall(VecDuplicate(x, &xt));
115: PetscCall(VecCopy(x, xt));
116: PetscCall(VecScale(xt, -1.0));
117: PetscCall(MatMultAdd(A, xt, b, b));
118: PetscCall(VecDestroy(&xt));
119: PetscCall(VecGetArrayRead(x, &xx));
120: PetscCall(VecGetArray(b, &bb));
121: for (i = 0; i < N; i++) bb[rows[i]] = diag * xx[rows[i]];
122: PetscCall(VecRestoreArrayRead(x, &xx));
123: PetscCall(VecRestoreArray(b, &bb));
124: }
126: PetscCall(MatDenseGetArray(A, &v));
127: for (i = 0; i < N; i++) {
128: slot = v + rows[i] * m;
129: PetscCall(PetscArrayzero(slot, r));
130: }
131: for (i = 0; i < N; i++) {
132: slot = v + rows[i];
133: for (j = 0; j < n; j++) {
134: *slot = 0.0;
135: slot += m;
136: }
137: }
138: if (diag != 0.0) {
139: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only coded for square matrices");
140: for (i = 0; i < N; i++) {
141: slot = v + (m + 1) * rows[i];
142: *slot = diag;
143: }
144: }
145: PetscCall(MatDenseRestoreArray(A, &v));
146: PetscFunctionReturn(PETSC_SUCCESS);
147: }
149: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
150: {
151: Mat B = NULL;
152: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
153: Mat_SeqDense *b;
154: PetscInt *ai = a->i, *aj = a->j, m = A->rmap->N, n = A->cmap->N, i;
155: const MatScalar *av;
156: PetscBool isseqdense;
158: PetscFunctionBegin;
159: if (reuse == MAT_REUSE_MATRIX) {
160: PetscCall(PetscObjectTypeCompare((PetscObject)*newmat, MATSEQDENSE, &isseqdense));
161: PetscCheck(isseqdense, PetscObjectComm((PetscObject)*newmat), PETSC_ERR_USER, "Cannot reuse matrix of type %s", ((PetscObject)*newmat)->type_name);
162: }
163: if (reuse != MAT_REUSE_MATRIX) {
164: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
165: PetscCall(MatSetSizes(B, m, n, m, n));
166: PetscCall(MatSetType(B, MATSEQDENSE));
167: PetscCall(MatSeqDenseSetPreallocation(B, NULL));
168: b = (Mat_SeqDense *)B->data;
169: } else {
170: b = (Mat_SeqDense *)((*newmat)->data);
171: for (i = 0; i < n; i++) PetscCall(PetscArrayzero(b->v + i * b->lda, m));
172: }
173: PetscCall(MatSeqAIJGetArrayRead(A, &av));
174: for (i = 0; i < m; i++) {
175: PetscInt j;
176: for (j = 0; j < ai[1] - ai[0]; j++) {
177: b->v[*aj * b->lda + i] = *av;
178: aj++;
179: av++;
180: }
181: ai++;
182: }
183: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
185: if (reuse == MAT_INPLACE_MATRIX) {
186: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
187: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
188: PetscCall(MatHeaderReplace(A, &B));
189: } else {
190: if (B) *newmat = B;
191: PetscCall(MatAssemblyBegin(*newmat, MAT_FINAL_ASSEMBLY));
192: PetscCall(MatAssemblyEnd(*newmat, MAT_FINAL_ASSEMBLY));
193: }
194: PetscFunctionReturn(PETSC_SUCCESS);
195: }
197: PETSC_INTERN PetscErrorCode MatConvert_SeqDense_SeqAIJ(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
198: {
199: Mat B = NULL;
200: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
201: PetscInt i, j;
202: PetscInt *rows, *nnz;
203: MatScalar *aa = a->v, *vals;
205: PetscFunctionBegin;
206: PetscCall(PetscCalloc3(A->rmap->n, &rows, A->rmap->n, &nnz, A->rmap->n, &vals));
207: if (reuse != MAT_REUSE_MATRIX) {
208: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
209: PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
210: PetscCall(MatSetType(B, MATSEQAIJ));
211: for (j = 0; j < A->cmap->n; j++) {
212: for (i = 0; i < A->rmap->n; i++)
213: if (aa[i] != 0.0 || (i == j && A->cmap->n == A->rmap->n)) ++nnz[i];
214: aa += a->lda;
215: }
216: PetscCall(MatSeqAIJSetPreallocation(B, PETSC_DETERMINE, nnz));
217: } else B = *newmat;
218: aa = a->v;
219: for (j = 0; j < A->cmap->n; j++) {
220: PetscInt numRows = 0;
221: for (i = 0; i < A->rmap->n; i++)
222: if (aa[i] != 0.0 || (i == j && A->cmap->n == A->rmap->n)) {
223: rows[numRows] = i;
224: vals[numRows++] = aa[i];
225: }
226: PetscCall(MatSetValues(B, numRows, rows, 1, &j, vals, INSERT_VALUES));
227: aa += a->lda;
228: }
229: PetscCall(PetscFree3(rows, nnz, vals));
230: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
231: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
233: if (reuse == MAT_INPLACE_MATRIX) 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;
407: PetscFunctionBegin;
408: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
409: PetscCallLAPACKInfo("LAPACKgetrs", LAPACKgetrs_(T ? "T" : "N", &m, &nrhs, mat->v, &mat->lda, mat->pivots, x, &m, &info));
410: PetscCall(PetscFPTrapPop());
411: PetscCall(PetscLogFlops(nrhs * (2.0 * m * m - m)));
412: PetscFunctionReturn(PETSC_SUCCESS);
413: }
415: static PetscErrorCode MatSolve_SeqDense_Internal_Cholesky(Mat A, PetscScalar *x, PetscBLASInt ldx, PetscBLASInt m, PetscBLASInt nrhs, PetscBLASInt k, PetscBool T)
416: {
417: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
419: PetscFunctionBegin;
420: if (A->spd == PETSC_BOOL3_TRUE) {
421: if (PetscDefined(USE_COMPLEX) && T) PetscCall(MatConjugate_SeqDense(A));
422: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
423: PetscCallLAPACKInfo("LAPACKpotrs", LAPACKpotrs_("L", &m, &nrhs, mat->v, &mat->lda, x, &m, &info));
424: PetscCall(PetscFPTrapPop());
425: if (PetscDefined(USE_COMPLEX) && T) PetscCall(MatConjugate_SeqDense(A));
426: #if defined(PETSC_USE_COMPLEX)
427: } else if (A->hermitian == PETSC_BOOL3_TRUE) {
428: if (T) PetscCall(MatConjugate_SeqDense(A));
429: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
430: PetscCallLAPACKInfo("LAPACKhetrs", LAPACKhetrs_("L", &m, &nrhs, mat->v, &mat->lda, mat->pivots, x, &m, &info));
431: PetscCall(PetscFPTrapPop());
432: if (T) PetscCall(MatConjugate_SeqDense(A));
433: #endif
434: } else { /* symmetric case */
435: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
436: PetscCallLAPACKInfo("LAPACKsytrs", LAPACKsytrs_("L", &m, &nrhs, mat->v, &mat->lda, mat->pivots, x, &m, &info));
437: PetscCall(PetscFPTrapPop());
438: }
439: PetscCall(PetscLogFlops(nrhs * (2.0 * m * m - m)));
440: PetscFunctionReturn(PETSC_SUCCESS);
441: }
443: static PetscErrorCode MatSolve_SeqDense_Internal_QR(Mat A, PetscScalar *x, PetscBLASInt ldx, PetscBLASInt m, PetscBLASInt nrhs, PetscBLASInt k)
444: {
445: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
446: char trans;
448: PetscFunctionBegin;
449: if (PetscDefined(USE_COMPLEX)) {
450: trans = 'C';
451: } else {
452: trans = 'T';
453: }
454: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
455: { /* lwork depends on the number of right-hand sides */
456: PetscBLASInt nlfwork, lfwork = -1;
457: PetscScalar fwork;
459: PetscCallLAPACKInfo("LAPACKormqr", LAPACKormqr_("L", &trans, &m, &nrhs, &mat->rank, mat->v, &mat->lda, mat->tau, x, &ldx, &fwork, &lfwork, &info));
460: nlfwork = (PetscBLASInt)PetscRealPart(fwork);
461: if (nlfwork > mat->lfwork) {
462: mat->lfwork = nlfwork;
463: PetscCall(PetscFree(mat->fwork));
464: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
465: }
466: }
467: PetscCallLAPACKInfo("LAPACKormqr", LAPACKormqr_("L", &trans, &m, &nrhs, &mat->rank, mat->v, &mat->lda, mat->tau, x, &ldx, mat->fwork, &mat->lfwork, &info));
468: PetscCall(PetscFPTrapPop());
469: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
470: PetscCallLAPACKInfo("LAPACKtrtrs", LAPACKtrtrs_("U", "N", "N", &mat->rank, &nrhs, mat->v, &mat->lda, x, &ldx, &info));
471: PetscCall(PetscFPTrapPop());
472: for (PetscInt j = 0; j < nrhs; j++) {
473: for (PetscInt i = mat->rank; i < k; i++) x[j * ldx + i] = 0.;
474: }
475: PetscCall(PetscLogFlops(nrhs * (4.0 * m * mat->rank - PetscSqr(mat->rank))));
476: PetscFunctionReturn(PETSC_SUCCESS);
477: }
479: static PetscErrorCode MatSolveTranspose_SeqDense_Internal_QR(Mat A, PetscScalar *x, PetscBLASInt ldx, PetscBLASInt m, PetscBLASInt nrhs, PetscBLASInt k)
480: {
481: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
483: PetscFunctionBegin;
484: if (A->rmap->n == A->cmap->n && mat->rank == A->rmap->n) {
485: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
486: PetscCallLAPACKInfo("LAPACKtrtrs", LAPACKtrtrs_("U", "T", "N", &m, &nrhs, mat->v, &mat->lda, x, &ldx, &info));
487: PetscCall(PetscFPTrapPop());
488: if (PetscDefined(USE_COMPLEX)) PetscCall(MatConjugate_SeqDense(A));
489: { /* lwork depends on the number of right-hand sides */
490: PetscBLASInt nlfwork, lfwork = -1;
491: PetscScalar fwork;
493: PetscCallLAPACKInfo("LAPACKormqr", LAPACKormqr_("L", "N", &m, &nrhs, &mat->rank, mat->v, &mat->lda, mat->tau, x, &ldx, &fwork, &lfwork, &info));
494: nlfwork = (PetscBLASInt)PetscRealPart(fwork);
495: if (nlfwork > mat->lfwork) {
496: mat->lfwork = nlfwork;
497: PetscCall(PetscFree(mat->fwork));
498: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
499: }
500: }
501: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
502: PetscCallLAPACKInfo("LAPACKormqr", LAPACKormqr_("L", "N", &m, &nrhs, &mat->rank, mat->v, &mat->lda, mat->tau, x, &ldx, mat->fwork, &mat->lfwork, &info));
503: PetscCall(PetscFPTrapPop());
504: if (PetscDefined(USE_COMPLEX)) PetscCall(MatConjugate_SeqDense(A));
505: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "QR factored matrix cannot be used for transpose solve");
506: PetscCall(PetscLogFlops(nrhs * (4.0 * m * mat->rank - PetscSqr(mat->rank))));
507: PetscFunctionReturn(PETSC_SUCCESS);
508: }
510: static PetscErrorCode MatSolve_SeqDense_SetUp(Mat A, Vec xx, Vec yy, PetscScalar **_y, PetscBLASInt *_m, PetscBLASInt *_k)
511: {
512: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
513: PetscScalar *y;
514: PetscBLASInt m = 0, k = 0;
516: PetscFunctionBegin;
517: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
518: PetscCall(PetscBLASIntCast(A->cmap->n, &k));
519: if (k < m) {
520: PetscCall(VecCopy(xx, mat->qrrhs));
521: PetscCall(VecGetArray(mat->qrrhs, &y));
522: } else {
523: PetscCall(VecCopy(xx, yy));
524: PetscCall(VecGetArray(yy, &y));
525: }
526: *_y = y;
527: *_k = k;
528: *_m = m;
529: PetscFunctionReturn(PETSC_SUCCESS);
530: }
532: static PetscErrorCode MatSolve_SeqDense_TearDown(Mat A, Vec xx, Vec yy, PetscScalar **_y, PetscBLASInt *_m, PetscBLASInt *_k)
533: {
534: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
535: PetscScalar *y = NULL;
536: PetscBLASInt m, k;
538: PetscFunctionBegin;
539: y = *_y;
540: *_y = NULL;
541: k = *_k;
542: m = *_m;
543: if (k < m) {
544: PetscScalar *yv;
545: PetscCall(VecGetArray(yy, &yv));
546: PetscCall(PetscArraycpy(yv, y, k));
547: PetscCall(VecRestoreArray(yy, &yv));
548: PetscCall(VecRestoreArray(mat->qrrhs, &y));
549: } else {
550: PetscCall(VecRestoreArray(yy, &y));
551: }
552: PetscFunctionReturn(PETSC_SUCCESS);
553: }
555: static PetscErrorCode MatSolve_SeqDense_LU(Mat A, Vec xx, Vec yy)
556: {
557: PetscScalar *y = NULL;
558: PetscBLASInt m = 0, k = 0;
560: PetscFunctionBegin;
561: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
562: PetscCall(MatSolve_SeqDense_Internal_LU(A, y, m, m, 1, k, PETSC_FALSE));
563: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
564: PetscFunctionReturn(PETSC_SUCCESS);
565: }
567: static PetscErrorCode MatSolveTranspose_SeqDense_LU(Mat A, Vec xx, Vec yy)
568: {
569: PetscScalar *y = NULL;
570: PetscBLASInt m = 0, k = 0;
572: PetscFunctionBegin;
573: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
574: PetscCall(MatSolve_SeqDense_Internal_LU(A, y, m, m, 1, k, PETSC_TRUE));
575: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
576: PetscFunctionReturn(PETSC_SUCCESS);
577: }
579: static PetscErrorCode MatSolve_SeqDense_Cholesky(Mat A, Vec xx, Vec yy)
580: {
581: PetscScalar *y = NULL;
582: PetscBLASInt m = 0, k = 0;
584: PetscFunctionBegin;
585: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
586: PetscCall(MatSolve_SeqDense_Internal_Cholesky(A, y, m, m, 1, k, PETSC_FALSE));
587: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
588: PetscFunctionReturn(PETSC_SUCCESS);
589: }
591: static PetscErrorCode MatSolveTranspose_SeqDense_Cholesky(Mat A, Vec xx, Vec yy)
592: {
593: PetscScalar *y = NULL;
594: PetscBLASInt m = 0, k = 0;
596: PetscFunctionBegin;
597: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
598: PetscCall(MatSolve_SeqDense_Internal_Cholesky(A, y, m, m, 1, k, PETSC_TRUE));
599: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
600: PetscFunctionReturn(PETSC_SUCCESS);
601: }
603: static PetscErrorCode MatSolve_SeqDense_QR(Mat A, Vec xx, Vec yy)
604: {
605: PetscScalar *y = NULL;
606: PetscBLASInt m = 0, k = 0;
608: PetscFunctionBegin;
609: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
610: PetscCall(MatSolve_SeqDense_Internal_QR(A, y, PetscMax(m, k), m, 1, k));
611: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
612: PetscFunctionReturn(PETSC_SUCCESS);
613: }
615: static PetscErrorCode MatSolveTranspose_SeqDense_QR(Mat A, Vec xx, Vec yy)
616: {
617: PetscScalar *y = NULL;
618: PetscBLASInt m = 0, k = 0;
620: PetscFunctionBegin;
621: PetscCall(MatSolve_SeqDense_SetUp(A, xx, yy, &y, &m, &k));
622: PetscCall(MatSolveTranspose_SeqDense_Internal_QR(A, y, PetscMax(m, k), m, 1, k));
623: PetscCall(MatSolve_SeqDense_TearDown(A, xx, yy, &y, &m, &k));
624: PetscFunctionReturn(PETSC_SUCCESS);
625: }
627: static PetscErrorCode MatMatSolve_SeqDense_SetUp(Mat A, Mat B, Mat X, PetscScalar **_y, PetscBLASInt *_ldy, PetscBLASInt *_m, PetscBLASInt *_nrhs, PetscBLASInt *_k)
628: {
629: const PetscScalar *b;
630: PetscScalar *y;
631: PetscInt n, _ldb, _ldx;
632: PetscBLASInt nrhs = 0, m = 0, k = 0, ldb = 0, ldx = 0, ldy = 0;
634: PetscFunctionBegin;
635: *_ldy = 0;
636: *_m = 0;
637: *_nrhs = 0;
638: *_k = 0;
639: *_y = NULL;
640: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
641: PetscCall(PetscBLASIntCast(A->cmap->n, &k));
642: PetscCall(MatGetSize(B, NULL, &n));
643: PetscCall(PetscBLASIntCast(n, &nrhs));
644: PetscCall(MatDenseGetLDA(B, &_ldb));
645: PetscCall(PetscBLASIntCast(_ldb, &ldb));
646: PetscCall(MatDenseGetLDA(X, &_ldx));
647: PetscCall(PetscBLASIntCast(_ldx, &ldx));
648: if (ldx < m) {
649: PetscCall(MatDenseGetArrayRead(B, &b));
650: PetscCall(PetscMalloc1(nrhs * m, &y));
651: if (ldb == m) {
652: PetscCall(PetscArraycpy(y, b, ldb * nrhs));
653: } else {
654: for (PetscInt j = 0; j < nrhs; j++) PetscCall(PetscArraycpy(&y[j * m], &b[j * ldb], m));
655: }
656: ldy = m;
657: PetscCall(MatDenseRestoreArrayRead(B, &b));
658: } else {
659: if (ldb == ldx) {
660: PetscCall(MatCopy(B, X, SAME_NONZERO_PATTERN));
661: PetscCall(MatDenseGetArray(X, &y));
662: } else {
663: PetscCall(MatDenseGetArray(X, &y));
664: PetscCall(MatDenseGetArrayRead(B, &b));
665: for (PetscInt j = 0; j < nrhs; j++) PetscCall(PetscArraycpy(&y[j * ldx], &b[j * ldb], m));
666: PetscCall(MatDenseRestoreArrayRead(B, &b));
667: }
668: ldy = ldx;
669: }
670: *_y = y;
671: *_ldy = ldy;
672: *_k = k;
673: *_m = m;
674: *_nrhs = nrhs;
675: PetscFunctionReturn(PETSC_SUCCESS);
676: }
678: static PetscErrorCode MatMatSolve_SeqDense_TearDown(Mat A, Mat B, Mat X, PetscScalar **_y, PetscBLASInt *_ldy, PetscBLASInt *_m, PetscBLASInt *_nrhs, PetscBLASInt *_k)
679: {
680: PetscScalar *y;
681: PetscInt _ldx;
682: PetscBLASInt k, ldy, nrhs, ldx = 0;
684: PetscFunctionBegin;
685: y = *_y;
686: *_y = NULL;
687: k = *_k;
688: ldy = *_ldy;
689: nrhs = *_nrhs;
690: PetscCall(MatDenseGetLDA(X, &_ldx));
691: PetscCall(PetscBLASIntCast(_ldx, &ldx));
692: if (ldx != ldy) {
693: PetscScalar *xv;
694: PetscCall(MatDenseGetArray(X, &xv));
695: for (PetscInt j = 0; j < nrhs; j++) PetscCall(PetscArraycpy(&xv[j * ldx], &y[j * ldy], k));
696: PetscCall(MatDenseRestoreArray(X, &xv));
697: PetscCall(PetscFree(y));
698: } else {
699: PetscCall(MatDenseRestoreArray(X, &y));
700: }
701: PetscFunctionReturn(PETSC_SUCCESS);
702: }
704: static PetscErrorCode MatMatSolve_SeqDense_LU(Mat A, Mat B, Mat X)
705: {
706: PetscScalar *y;
707: PetscBLASInt m, k, ldy, nrhs;
709: PetscFunctionBegin;
710: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
711: PetscCall(MatSolve_SeqDense_Internal_LU(A, y, ldy, m, nrhs, k, PETSC_FALSE));
712: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
713: PetscFunctionReturn(PETSC_SUCCESS);
714: }
716: static PetscErrorCode MatMatSolveTranspose_SeqDense_LU(Mat A, Mat B, Mat X)
717: {
718: PetscScalar *y;
719: PetscBLASInt m, k, ldy, nrhs;
721: PetscFunctionBegin;
722: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
723: PetscCall(MatSolve_SeqDense_Internal_LU(A, y, ldy, m, nrhs, k, PETSC_TRUE));
724: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
725: PetscFunctionReturn(PETSC_SUCCESS);
726: }
728: static PetscErrorCode MatMatSolve_SeqDense_Cholesky(Mat A, Mat B, Mat X)
729: {
730: PetscScalar *y;
731: PetscBLASInt m, k, ldy, nrhs;
733: PetscFunctionBegin;
734: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
735: PetscCall(MatSolve_SeqDense_Internal_Cholesky(A, y, ldy, m, nrhs, k, PETSC_FALSE));
736: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
737: PetscFunctionReturn(PETSC_SUCCESS);
738: }
740: static PetscErrorCode MatMatSolveTranspose_SeqDense_Cholesky(Mat A, Mat B, Mat X)
741: {
742: PetscScalar *y;
743: PetscBLASInt m, k, ldy, nrhs;
745: PetscFunctionBegin;
746: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
747: PetscCall(MatSolve_SeqDense_Internal_Cholesky(A, y, ldy, m, nrhs, k, PETSC_TRUE));
748: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
749: PetscFunctionReturn(PETSC_SUCCESS);
750: }
752: static PetscErrorCode MatMatSolve_SeqDense_QR(Mat A, Mat B, Mat X)
753: {
754: PetscScalar *y;
755: PetscBLASInt m, k, ldy, nrhs;
757: PetscFunctionBegin;
758: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
759: PetscCall(MatSolve_SeqDense_Internal_QR(A, y, ldy, m, nrhs, k));
760: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
761: PetscFunctionReturn(PETSC_SUCCESS);
762: }
764: static PetscErrorCode MatMatSolveTranspose_SeqDense_QR(Mat A, Mat B, Mat X)
765: {
766: PetscScalar *y;
767: PetscBLASInt m, k, ldy, nrhs;
769: PetscFunctionBegin;
770: PetscCall(MatMatSolve_SeqDense_SetUp(A, B, X, &y, &ldy, &m, &nrhs, &k));
771: PetscCall(MatSolveTranspose_SeqDense_Internal_QR(A, y, ldy, m, nrhs, k));
772: PetscCall(MatMatSolve_SeqDense_TearDown(A, B, X, &y, &ldy, &m, &nrhs, &k));
773: PetscFunctionReturn(PETSC_SUCCESS);
774: }
776: /* COMMENT: I have chosen to hide row permutation in the pivots,
777: rather than put it in the Mat->row slot.*/
778: PetscErrorCode MatLUFactor_SeqDense(Mat A, IS row, IS col, PETSC_UNUSED const MatFactorInfo *minfo)
779: {
780: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
781: PetscBLASInt n, m, info;
783: PetscFunctionBegin;
784: PetscCall(PetscBLASIntCast(A->cmap->n, &n));
785: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
786: if (!mat->pivots) PetscCall(PetscMalloc1(A->rmap->n, &mat->pivots));
787: if (!A->rmap->n || !A->cmap->n) PetscFunctionReturn(PETSC_SUCCESS);
788: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
789: PetscCallBLAS("LAPACKgetrf", LAPACKgetrf_(&m, &n, mat->v, &mat->lda, mat->pivots, &info));
790: PetscCheck(info >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Error in LAPACK argument %" PetscBLASInt_FMT, -info);
791: PetscCheck(info <= 0, PETSC_COMM_SELF, PETSC_ERR_MAT_LU_ZRPVT, "Bad factorization: zero pivot in row %" PetscBLASInt_FMT, info - 1);
792: PetscCall(PetscFPTrapPop());
794: A->ops->solve = MatSolve_SeqDense_LU;
795: A->ops->matsolve = MatMatSolve_SeqDense_LU;
796: A->ops->solvetranspose = MatSolveTranspose_SeqDense_LU;
797: A->ops->matsolvetranspose = MatMatSolveTranspose_SeqDense_LU;
798: A->factortype = MAT_FACTOR_LU;
800: PetscCall(PetscFree(A->solvertype));
801: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &A->solvertype));
803: PetscCall(PetscLogFlops((2.0 * A->cmap->n * A->cmap->n * A->cmap->n) / 3));
804: PetscFunctionReturn(PETSC_SUCCESS);
805: }
807: static PetscErrorCode MatLUFactorNumeric_SeqDense(Mat fact, Mat A, const MatFactorInfo *info)
808: {
809: PetscFunctionBegin;
810: PetscCall(MatDuplicateNoCreate_SeqDense(fact, A, MAT_COPY_VALUES));
811: PetscUseTypeMethod(fact, lufactor, NULL, NULL, info);
812: PetscFunctionReturn(PETSC_SUCCESS);
813: }
815: PetscErrorCode MatLUFactorSymbolic_SeqDense(Mat fact, Mat A, IS row, IS col, PETSC_UNUSED const MatFactorInfo *info)
816: {
817: PetscFunctionBegin;
818: fact->preallocated = PETSC_TRUE;
819: fact->assembled = PETSC_TRUE;
820: fact->ops->lufactornumeric = MatLUFactorNumeric_SeqDense;
821: PetscFunctionReturn(PETSC_SUCCESS);
822: }
824: /* Cholesky as L*L^T or L*D*L^T and the symmetric/hermitian complex variants */
825: PetscErrorCode MatCholeskyFactor_SeqDense(Mat A, IS perm, PETSC_UNUSED const MatFactorInfo *minfo)
826: {
827: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
828: PetscBLASInt info, n;
830: PetscFunctionBegin;
831: PetscCall(PetscBLASIntCast(A->cmap->n, &n));
832: if (!A->rmap->n || !A->cmap->n) PetscFunctionReturn(PETSC_SUCCESS);
833: if (A->spd == PETSC_BOOL3_TRUE) {
834: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
835: PetscCallBLAS("LAPACKpotrf", LAPACKpotrf_("L", &n, mat->v, &mat->lda, &info));
836: PetscCall(PetscFPTrapPop());
837: #if defined(PETSC_USE_COMPLEX)
838: } else if (A->hermitian == PETSC_BOOL3_TRUE) {
839: if (!mat->pivots) PetscCall(PetscMalloc1(A->rmap->n, &mat->pivots));
840: if (!mat->fwork) {
841: PetscScalar dummy;
843: mat->lfwork = -1;
844: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
845: PetscCallBLAS("LAPACKhetrf", LAPACKhetrf_("L", &n, mat->v, &mat->lda, mat->pivots, &dummy, &mat->lfwork, &info));
846: PetscCall(PetscFPTrapPop());
847: PetscCall(PetscBLASIntCast((PetscCount)(PetscRealPart(dummy)), &mat->lfwork));
848: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
849: }
850: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
851: PetscCallBLAS("LAPACKhetrf", LAPACKhetrf_("L", &n, mat->v, &mat->lda, mat->pivots, mat->fwork, &mat->lfwork, &info));
852: PetscCall(PetscFPTrapPop());
853: #endif
854: } else { /* symmetric case */
855: if (!mat->pivots) PetscCall(PetscMalloc1(A->rmap->n, &mat->pivots));
856: if (!mat->fwork) {
857: PetscScalar dummy;
859: mat->lfwork = -1;
860: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
861: PetscCallBLAS("LAPACKsytrf", LAPACKsytrf_("L", &n, mat->v, &mat->lda, mat->pivots, &dummy, &mat->lfwork, &info));
862: PetscCall(PetscFPTrapPop());
863: PetscCall(PetscBLASIntCast((PetscCount)(PetscRealPart(dummy)), &mat->lfwork));
864: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
865: }
866: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
867: PetscCallBLAS("LAPACKsytrf", LAPACKsytrf_("L", &n, mat->v, &mat->lda, mat->pivots, mat->fwork, &mat->lfwork, &info));
868: PetscCall(PetscFPTrapPop());
869: }
870: PetscCheck(info >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Error in LAPACK argument %" PetscBLASInt_FMT, -info);
871: PetscCheck(info <= 0, PETSC_COMM_SELF, PETSC_ERR_MAT_CH_ZRPVT, "Bad factorization: zero pivot in row %" PetscBLASInt_FMT, info - 1);
873: A->ops->solve = MatSolve_SeqDense_Cholesky;
874: A->ops->matsolve = MatMatSolve_SeqDense_Cholesky;
875: A->ops->solvetranspose = MatSolveTranspose_SeqDense_Cholesky;
876: A->ops->matsolvetranspose = MatMatSolveTranspose_SeqDense_Cholesky;
877: A->factortype = MAT_FACTOR_CHOLESKY;
879: PetscCall(PetscFree(A->solvertype));
880: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &A->solvertype));
882: PetscCall(PetscLogFlops((1.0 * A->cmap->n * A->cmap->n * A->cmap->n) / 3.0));
883: PetscFunctionReturn(PETSC_SUCCESS);
884: }
886: static PetscErrorCode MatCholeskyFactorNumeric_SeqDense(Mat fact, Mat A, const MatFactorInfo *info)
887: {
888: PetscFunctionBegin;
889: PetscCall(MatDuplicateNoCreate_SeqDense(fact, A, MAT_COPY_VALUES));
890: PetscUseTypeMethod(fact, choleskyfactor, NULL, info);
891: PetscFunctionReturn(PETSC_SUCCESS);
892: }
894: PetscErrorCode MatCholeskyFactorSymbolic_SeqDense(Mat fact, Mat A, IS row, const MatFactorInfo *info)
895: {
896: PetscFunctionBegin;
897: fact->assembled = PETSC_TRUE;
898: fact->preallocated = PETSC_TRUE;
899: fact->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqDense;
900: PetscFunctionReturn(PETSC_SUCCESS);
901: }
903: PetscErrorCode MatQRFactor_SeqDense(Mat A, IS col, PETSC_UNUSED const MatFactorInfo *minfo)
904: {
905: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
906: PetscBLASInt n, m, min, max;
908: PetscFunctionBegin;
909: PetscCall(PetscBLASIntCast(A->cmap->n, &n));
910: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
911: max = PetscMax(m, n);
912: min = PetscMin(m, n);
913: if (!mat->tau) PetscCall(PetscMalloc1(min, &mat->tau));
914: if (!mat->pivots) PetscCall(PetscMalloc1(n, &mat->pivots));
915: if (!mat->qrrhs) PetscCall(MatCreateVecs(A, NULL, &mat->qrrhs));
916: if (!A->rmap->n || !A->cmap->n) PetscFunctionReturn(PETSC_SUCCESS);
917: if (!mat->fwork) {
918: PetscScalar dummy;
920: mat->lfwork = -1;
921: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
922: PetscCallLAPACKInfo("LAPACKgeqrf", LAPACKgeqrf_(&m, &n, mat->v, &mat->lda, mat->tau, &dummy, &mat->lfwork, &info));
923: PetscCall(PetscFPTrapPop());
924: PetscCall(PetscBLASIntCast((PetscCount)(PetscRealPart(dummy)), &mat->lfwork));
925: PetscCall(PetscMalloc1(mat->lfwork, &mat->fwork));
926: }
927: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
928: PetscCallLAPACKInfo("LAPACKgeqrf", LAPACKgeqrf_(&m, &n, mat->v, &mat->lda, mat->tau, mat->fwork, &mat->lfwork, &info));
929: PetscCall(PetscFPTrapPop());
930: // 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
931: mat->rank = min;
933: A->ops->solve = MatSolve_SeqDense_QR;
934: A->ops->matsolve = MatMatSolve_SeqDense_QR;
935: A->factortype = MAT_FACTOR_QR;
936: if (m == n) {
937: A->ops->solvetranspose = MatSolveTranspose_SeqDense_QR;
938: A->ops->matsolvetranspose = MatMatSolveTranspose_SeqDense_QR;
939: }
941: PetscCall(PetscFree(A->solvertype));
942: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &A->solvertype));
944: PetscCall(PetscLogFlops(2.0 * min * min * (max - min / 3.0)));
945: PetscFunctionReturn(PETSC_SUCCESS);
946: }
948: static PetscErrorCode MatQRFactorNumeric_SeqDense(Mat fact, Mat A, const MatFactorInfo *info)
949: {
950: PetscFunctionBegin;
951: PetscCall(MatDuplicateNoCreate_SeqDense(fact, A, MAT_COPY_VALUES));
952: PetscUseMethod(fact, "MatQRFactor_C", (Mat, IS, const MatFactorInfo *), (fact, NULL, info));
953: PetscFunctionReturn(PETSC_SUCCESS);
954: }
956: PetscErrorCode MatQRFactorSymbolic_SeqDense(Mat fact, Mat A, IS row, const MatFactorInfo *info)
957: {
958: PetscFunctionBegin;
959: fact->assembled = PETSC_TRUE;
960: fact->preallocated = PETSC_TRUE;
961: PetscCall(PetscObjectComposeFunction((PetscObject)fact, "MatQRFactorNumeric_C", MatQRFactorNumeric_SeqDense));
962: PetscFunctionReturn(PETSC_SUCCESS);
963: }
965: /* uses LAPACK */
966: PETSC_INTERN PetscErrorCode MatGetFactor_seqdense_petsc(Mat A, MatFactorType ftype, Mat *fact)
967: {
968: PetscFunctionBegin;
969: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), fact));
970: PetscCall(MatSetSizes(*fact, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n));
971: PetscCall(MatSetType(*fact, MATDENSE));
972: (*fact)->trivialsymbolic = PETSC_TRUE;
973: if (ftype == MAT_FACTOR_LU || ftype == MAT_FACTOR_ILU) {
974: (*fact)->ops->lufactorsymbolic = MatLUFactorSymbolic_SeqDense;
975: (*fact)->ops->ilufactorsymbolic = MatLUFactorSymbolic_SeqDense;
976: } else if (ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC) {
977: (*fact)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqDense;
978: } else if (ftype == MAT_FACTOR_QR) {
979: PetscCall(PetscObjectComposeFunction((PetscObject)*fact, "MatQRFactorSymbolic_C", MatQRFactorSymbolic_SeqDense));
980: }
981: (*fact)->factortype = ftype;
983: PetscCall(PetscFree((*fact)->solvertype));
984: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &(*fact)->solvertype));
985: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&(*fact)->preferredordering[MAT_FACTOR_LU]));
986: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&(*fact)->preferredordering[MAT_FACTOR_ILU]));
987: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&(*fact)->preferredordering[MAT_FACTOR_CHOLESKY]));
988: PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&(*fact)->preferredordering[MAT_FACTOR_ICC]));
989: PetscFunctionReturn(PETSC_SUCCESS);
990: }
992: static PetscErrorCode MatSOR_SeqDense(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal shift, PetscInt its, PetscInt lits, Vec xx)
993: {
994: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
995: PetscScalar *x, *v = mat->v, zero = 0.0, xt;
996: const PetscScalar *b;
997: PetscInt m = A->rmap->n, i;
998: PetscBLASInt o = 1, bm = 0;
1000: PetscFunctionBegin;
1001: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1002: PetscCheck(A->offloadmask != PETSC_OFFLOAD_GPU, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
1003: #endif
1004: if (shift == -1) shift = 0.0; /* negative shift indicates do not error on zero diagonal; this code never zeros on zero diagonal */
1005: PetscCall(PetscBLASIntCast(m, &bm));
1006: if (flag & SOR_ZERO_INITIAL_GUESS) {
1007: /* this is a hack fix, should have another version without the second BLASdotu */
1008: PetscCall(VecSet(xx, zero));
1009: }
1010: PetscCall(VecGetArray(xx, &x));
1011: PetscCall(VecGetArrayRead(bb, &b));
1012: its = its * lits;
1013: 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);
1014: while (its--) {
1015: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1016: for (i = 0; i < m; i++) {
1017: PetscCallBLAS("BLASdotu", xt = b[i] - BLASdotu_(&bm, v + i, &bm, x, &o));
1018: x[i] = (1. - omega) * x[i] + (xt + v[i + i * m] * x[i]) * omega / (v[i + i * m] + shift);
1019: }
1020: }
1021: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1022: for (i = m - 1; i >= 0; i--) {
1023: PetscCallBLAS("BLASdotu", xt = b[i] - BLASdotu_(&bm, v + i, &bm, x, &o));
1024: x[i] = (1. - omega) * x[i] + (xt + v[i + i * m] * x[i]) * omega / (v[i + i * m] + shift);
1025: }
1026: }
1027: }
1028: PetscCall(VecRestoreArrayRead(bb, &b));
1029: PetscCall(VecRestoreArray(xx, &x));
1030: PetscFunctionReturn(PETSC_SUCCESS);
1031: }
1033: PETSC_INTERN PetscErrorCode MatMultColumnRangeKernel_SeqDense(Mat A, Vec xx, Vec yy, PetscInt c_start, PetscInt c_end, PetscBool trans, PetscBool herm)
1034: {
1035: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1036: PetscScalar *y, _DOne = 1.0, _DZero = 0.0;
1037: PetscBLASInt m, n, _One = 1;
1038: const PetscScalar *v = mat->v, *x;
1040: PetscFunctionBegin;
1041: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
1042: PetscCall(PetscBLASIntCast(c_end - c_start, &n));
1043: PetscCall(VecGetArrayRead(xx, &x));
1044: PetscCall(VecGetArrayWrite(yy, &y));
1045: if (!m || !n) {
1046: PetscBLASInt i;
1047: if (trans)
1048: for (i = 0; i < n; i++) y[i] = 0.0;
1049: else
1050: for (i = 0; i < m; i++) y[i] = 0.0;
1051: } else {
1052: if (trans) {
1053: if (herm) PetscCallBLAS("BLASgemv", BLASgemv_("C", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x, &_One, &_DZero, y + c_start, &_One));
1054: else PetscCallBLAS("BLASgemv", BLASgemv_("T", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x, &_One, &_DZero, y + c_start, &_One));
1055: } else {
1056: PetscCallBLAS("BLASgemv", BLASgemv_("N", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x + c_start, &_One, &_DZero, y, &_One));
1057: }
1058: PetscCall(PetscLogFlops(2.0 * m * n - n));
1059: }
1060: PetscCall(VecRestoreArrayRead(xx, &x));
1061: PetscCall(VecRestoreArrayWrite(yy, &y));
1062: PetscFunctionReturn(PETSC_SUCCESS);
1063: }
1065: PetscErrorCode MatMultHermitianTransposeColumnRange_SeqDense(Mat A, Vec xx, Vec yy, PetscInt c_start, PetscInt c_end)
1066: {
1067: PetscFunctionBegin;
1068: PetscCall(MatMultColumnRangeKernel_SeqDense(A, xx, yy, c_start, c_end, PETSC_TRUE, PETSC_TRUE));
1069: PetscFunctionReturn(PETSC_SUCCESS);
1070: }
1072: PetscErrorCode MatMult_SeqDense(Mat A, Vec xx, Vec yy)
1073: {
1074: PetscFunctionBegin;
1075: PetscCall(MatMultColumnRangeKernel_SeqDense(A, xx, yy, 0, A->cmap->n, PETSC_FALSE, PETSC_FALSE));
1076: PetscFunctionReturn(PETSC_SUCCESS);
1077: }
1079: PetscErrorCode MatMultTranspose_SeqDense(Mat A, Vec xx, Vec yy)
1080: {
1081: PetscFunctionBegin;
1082: PetscCall(MatMultColumnRangeKernel_SeqDense(A, xx, yy, 0, A->cmap->n, PETSC_TRUE, PETSC_FALSE));
1083: PetscFunctionReturn(PETSC_SUCCESS);
1084: }
1086: PetscErrorCode MatMultHermitianTranspose_SeqDense(Mat A, Vec xx, Vec yy)
1087: {
1088: PetscFunctionBegin;
1089: PetscCall(MatMultColumnRangeKernel_SeqDense(A, xx, yy, 0, A->cmap->n, PETSC_TRUE, PETSC_TRUE));
1090: PetscFunctionReturn(PETSC_SUCCESS);
1091: }
1093: PETSC_INTERN PetscErrorCode MatMultAddColumnRangeKernel_SeqDense(Mat A, Vec xx, Vec zz, Vec yy, PetscInt c_start, PetscInt c_end, PetscBool trans, PetscBool herm)
1094: {
1095: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1096: const PetscScalar *v = mat->v, *x;
1097: PetscScalar *y, _DOne = 1.0;
1098: PetscBLASInt m, n, _One = 1;
1100: PetscFunctionBegin;
1101: PetscCall(PetscBLASIntCast(A->rmap->n, &m));
1102: PetscCall(PetscBLASIntCast(c_end - c_start, &n));
1103: PetscCall(VecCopy(zz, yy));
1104: if (!m || !n) PetscFunctionReturn(PETSC_SUCCESS);
1105: PetscCall(VecGetArray(yy, &y));
1106: PetscCall(VecGetArrayRead(xx, &x));
1107: if (trans) {
1108: if (herm) PetscCallBLAS("BLASgemv", BLASgemv_("C", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x, &_One, &_DOne, y + c_start, &_One));
1109: else PetscCallBLAS("BLASgemv", BLASgemv_("T", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x, &_One, &_DOne, y + c_start, &_One));
1110: } else {
1111: PetscCallBLAS("BLASgemv", BLASgemv_("N", &m, &n, &_DOne, v + c_start * mat->lda, &mat->lda, x + c_start, &_One, &_DOne, y, &_One));
1112: }
1113: PetscCall(VecRestoreArrayRead(xx, &x));
1114: PetscCall(VecRestoreArray(yy, &y));
1115: PetscCall(PetscLogFlops(2.0 * m * n));
1116: PetscFunctionReturn(PETSC_SUCCESS);
1117: }
1119: PetscErrorCode MatMultColumnRange_SeqDense(Mat A, Vec xx, Vec yy, PetscInt c_start, PetscInt c_end)
1120: {
1121: PetscFunctionBegin;
1122: PetscCall(MatMultColumnRangeKernel_SeqDense(A, xx, yy, c_start, c_end, PETSC_FALSE, PETSC_FALSE));
1123: PetscFunctionReturn(PETSC_SUCCESS);
1124: }
1126: PetscErrorCode MatMultAddColumnRange_SeqDense(Mat A, Vec xx, Vec zz, Vec yy, PetscInt c_start, PetscInt c_end)
1127: {
1128: PetscFunctionBegin;
1129: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, c_start, c_end, PETSC_FALSE, PETSC_FALSE));
1130: PetscFunctionReturn(PETSC_SUCCESS);
1131: }
1133: PetscErrorCode MatMultHermitianTransposeAddColumnRange_SeqDense(Mat A, Vec xx, Vec zz, Vec yy, PetscInt c_start, PetscInt c_end)
1134: {
1135: PetscFunctionBegin;
1136: PetscMPIInt rank;
1137: PetscCallMPI(MPI_Comm_rank(MPI_COMM_WORLD, &rank));
1138: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, c_start, c_end, PETSC_TRUE, PETSC_TRUE));
1139: PetscFunctionReturn(PETSC_SUCCESS);
1140: }
1142: PetscErrorCode MatMultAdd_SeqDense(Mat A, Vec xx, Vec zz, Vec yy)
1143: {
1144: PetscFunctionBegin;
1145: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, 0, A->cmap->n, PETSC_FALSE, PETSC_FALSE));
1146: PetscFunctionReturn(PETSC_SUCCESS);
1147: }
1149: PetscErrorCode MatMultTransposeAdd_SeqDense(Mat A, Vec xx, Vec zz, Vec yy)
1150: {
1151: PetscFunctionBegin;
1152: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, 0, A->cmap->n, PETSC_TRUE, PETSC_FALSE));
1153: PetscFunctionReturn(PETSC_SUCCESS);
1154: }
1156: PetscErrorCode MatMultHermitianTransposeAdd_SeqDense(Mat A, Vec xx, Vec zz, Vec yy)
1157: {
1158: PetscFunctionBegin;
1159: PetscCall(MatMultAddColumnRangeKernel_SeqDense(A, xx, zz, yy, 0, A->cmap->n, PETSC_TRUE, PETSC_TRUE));
1160: PetscFunctionReturn(PETSC_SUCCESS);
1161: }
1163: static PetscErrorCode MatGetRow_SeqDense(Mat A, PetscInt row, PetscInt *ncols, PetscInt **cols, PetscScalar **vals)
1164: {
1165: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1166: PetscInt i;
1168: PetscFunctionBegin;
1169: if (ncols) *ncols = A->cmap->n;
1170: if (cols) {
1171: PetscCall(PetscMalloc1(A->cmap->n, cols));
1172: for (i = 0; i < A->cmap->n; i++) (*cols)[i] = i;
1173: }
1174: if (vals) {
1175: const PetscScalar *v;
1177: PetscCall(MatDenseGetArrayRead(A, &v));
1178: PetscCall(PetscMalloc1(A->cmap->n, vals));
1179: v += row;
1180: for (i = 0; i < A->cmap->n; i++) {
1181: (*vals)[i] = *v;
1182: v += mat->lda;
1183: }
1184: PetscCall(MatDenseRestoreArrayRead(A, &v));
1185: }
1186: PetscFunctionReturn(PETSC_SUCCESS);
1187: }
1189: static PetscErrorCode MatRestoreRow_SeqDense(Mat A, PetscInt row, PetscInt *ncols, PetscInt **cols, PetscScalar **vals)
1190: {
1191: PetscFunctionBegin;
1192: if (cols) PetscCall(PetscFree(*cols));
1193: if (vals) PetscCall(PetscFree(*vals));
1194: PetscFunctionReturn(PETSC_SUCCESS);
1195: }
1197: static PetscErrorCode MatSetValues_SeqDense(Mat A, PetscInt m, const PetscInt indexm[], PetscInt n, const PetscInt indexn[], const PetscScalar v[], InsertMode addv)
1198: {
1199: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1200: PetscScalar *av;
1201: PetscInt i, j, idx = 0;
1202: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1203: PetscOffloadMask oldf;
1204: #endif
1206: PetscFunctionBegin;
1207: PetscCall(MatDenseGetArray(A, &av));
1208: if (!mat->roworiented) {
1209: if (addv == INSERT_VALUES) {
1210: for (j = 0; j < n; j++) {
1211: if (indexn[j] < 0) {
1212: idx += m;
1213: continue;
1214: }
1215: 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);
1216: for (i = 0; i < m; i++) {
1217: if (indexm[i] < 0) {
1218: idx++;
1219: continue;
1220: }
1221: 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);
1222: av[indexn[j] * mat->lda + indexm[i]] = v ? v[idx++] : (idx++, 0.0);
1223: }
1224: }
1225: } else {
1226: for (j = 0; j < n; j++) {
1227: if (indexn[j] < 0) {
1228: idx += m;
1229: continue;
1230: }
1231: 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);
1232: for (i = 0; i < m; i++) {
1233: if (indexm[i] < 0) {
1234: idx++;
1235: continue;
1236: }
1237: 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);
1238: av[indexn[j] * mat->lda + indexm[i]] += v ? v[idx++] : (idx++, 0.0);
1239: }
1240: }
1241: }
1242: } else {
1243: if (addv == INSERT_VALUES) {
1244: for (i = 0; i < m; i++) {
1245: if (indexm[i] < 0) {
1246: idx += n;
1247: continue;
1248: }
1249: 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);
1250: for (j = 0; j < n; j++) {
1251: if (indexn[j] < 0) {
1252: idx++;
1253: continue;
1254: }
1255: 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);
1256: av[indexn[j] * mat->lda + indexm[i]] = v ? v[idx++] : (idx++, 0.0);
1257: }
1258: }
1259: } else {
1260: for (i = 0; i < m; i++) {
1261: if (indexm[i] < 0) {
1262: idx += n;
1263: continue;
1264: }
1265: 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);
1266: for (j = 0; j < n; j++) {
1267: if (indexn[j] < 0) {
1268: idx++;
1269: continue;
1270: }
1271: 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);
1272: av[indexn[j] * mat->lda + indexm[i]] += v ? v[idx++] : (idx++, 0.0);
1273: }
1274: }
1275: }
1276: }
1277: /* hack to prevent unneeded copy to the GPU while returning the array */
1278: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1279: oldf = A->offloadmask;
1280: A->offloadmask = PETSC_OFFLOAD_GPU;
1281: #endif
1282: PetscCall(MatDenseRestoreArray(A, &av));
1283: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1284: A->offloadmask = (oldf == PETSC_OFFLOAD_UNALLOCATED ? PETSC_OFFLOAD_UNALLOCATED : PETSC_OFFLOAD_CPU);
1285: #endif
1286: PetscFunctionReturn(PETSC_SUCCESS);
1287: }
1289: static PetscErrorCode MatGetValues_SeqDense(Mat A, PetscInt m, const PetscInt indexm[], PetscInt n, const PetscInt indexn[], PetscScalar v[])
1290: {
1291: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1292: const PetscScalar *vv;
1293: PetscInt i, j;
1295: PetscFunctionBegin;
1296: PetscCall(MatDenseGetArrayRead(A, &vv));
1297: /* row-oriented output */
1298: for (i = 0; i < m; i++) {
1299: if (indexm[i] < 0) {
1300: v += n;
1301: continue;
1302: }
1303: 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);
1304: for (j = 0; j < n; j++) {
1305: if (indexn[j] < 0) {
1306: v++;
1307: continue;
1308: }
1309: 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);
1310: *v++ = vv[indexn[j] * mat->lda + indexm[i]];
1311: }
1312: }
1313: PetscCall(MatDenseRestoreArrayRead(A, &vv));
1314: PetscFunctionReturn(PETSC_SUCCESS);
1315: }
1317: PetscErrorCode MatView_Dense_Binary(Mat mat, PetscViewer viewer)
1318: {
1319: PetscBool skipHeader;
1320: PetscViewerFormat format;
1321: PetscInt header[4], M, N, m, lda, i, j;
1322: PetscCount k;
1323: const PetscScalar *v;
1324: PetscScalar *vwork;
1326: PetscFunctionBegin;
1327: PetscCall(PetscViewerSetUp(viewer));
1328: PetscCall(PetscViewerBinaryGetSkipHeader(viewer, &skipHeader));
1329: PetscCall(PetscViewerGetFormat(viewer, &format));
1330: if (skipHeader) format = PETSC_VIEWER_NATIVE;
1332: PetscCall(MatGetSize(mat, &M, &N));
1334: /* write matrix header */
1335: header[0] = MAT_FILE_CLASSID;
1336: header[1] = M;
1337: header[2] = N;
1338: header[3] = (format == PETSC_VIEWER_NATIVE) ? MATRIX_BINARY_FORMAT_DENSE : M * N;
1339: if (!skipHeader) PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1341: PetscCall(MatGetLocalSize(mat, &m, NULL));
1342: if (format != PETSC_VIEWER_NATIVE) {
1343: PetscInt nnz = m * N, *iwork;
1344: /* store row lengths for each row */
1345: PetscCall(PetscMalloc1(nnz, &iwork));
1346: for (i = 0; i < m; i++) iwork[i] = N;
1347: PetscCall(PetscViewerBinaryWriteAll(viewer, iwork, m, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1348: /* store column indices (zero start index) */
1349: for (k = 0, i = 0; i < m; i++)
1350: for (j = 0; j < N; j++, k++) iwork[k] = j;
1351: PetscCall(PetscViewerBinaryWriteAll(viewer, iwork, nnz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1352: PetscCall(PetscFree(iwork));
1353: }
1354: /* store matrix values as a dense matrix in row major order */
1355: PetscCall(PetscMalloc1(m * N, &vwork));
1356: PetscCall(MatDenseGetArrayRead(mat, &v));
1357: PetscCall(MatDenseGetLDA(mat, &lda));
1358: for (k = 0, i = 0; i < m; i++)
1359: for (j = 0; j < N; j++, k++) vwork[k] = v[i + (size_t)lda * j];
1360: PetscCall(MatDenseRestoreArrayRead(mat, &v));
1361: PetscCall(PetscViewerBinaryWriteAll(viewer, vwork, m * N, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1362: PetscCall(PetscFree(vwork));
1363: PetscFunctionReturn(PETSC_SUCCESS);
1364: }
1366: PetscErrorCode MatLoad_Dense_Binary(Mat mat, PetscViewer viewer)
1367: {
1368: PetscBool skipHeader;
1369: PetscInt header[4], M, N, m, nz, lda, i, j, k;
1370: PetscInt rows, cols;
1371: PetscScalar *v, *vwork;
1373: PetscFunctionBegin;
1374: PetscCall(PetscViewerSetUp(viewer));
1375: PetscCall(PetscViewerBinaryGetSkipHeader(viewer, &skipHeader));
1377: if (!skipHeader) {
1378: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
1379: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
1380: M = header[1];
1381: N = header[2];
1382: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
1383: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
1384: nz = header[3];
1385: PetscCheck(nz == MATRIX_BINARY_FORMAT_DENSE || nz >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Unknown matrix format %" PetscInt_FMT " in file", nz);
1386: } else {
1387: PetscCall(MatGetSize(mat, &M, &N));
1388: 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");
1389: nz = MATRIX_BINARY_FORMAT_DENSE;
1390: }
1392: /* setup global sizes if not set */
1393: if (mat->rmap->N < 0) mat->rmap->N = M;
1394: if (mat->cmap->N < 0) mat->cmap->N = N;
1395: PetscCall(MatSetUp(mat));
1396: /* check if global sizes are correct */
1397: PetscCall(MatGetSize(mat, &rows, &cols));
1398: 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);
1400: PetscCall(MatGetSize(mat, NULL, &N));
1401: PetscCall(MatGetLocalSize(mat, &m, NULL));
1402: PetscCall(MatDenseGetArray(mat, &v));
1403: PetscCall(MatDenseGetLDA(mat, &lda));
1404: if (nz == MATRIX_BINARY_FORMAT_DENSE) { /* matrix in file is dense format */
1405: PetscCount nnz = (size_t)m * N;
1406: /* read in matrix values */
1407: PetscCall(PetscMalloc1(nnz, &vwork));
1408: PetscCall(PetscViewerBinaryReadAll(viewer, vwork, nnz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1409: /* store values in column major order */
1410: for (j = 0; j < N; j++)
1411: for (i = 0; i < m; i++) v[i + (size_t)lda * j] = vwork[(size_t)i * N + j];
1412: PetscCall(PetscFree(vwork));
1413: } else { /* matrix in file is sparse format */
1414: PetscInt nnz = 0, *rlens, *icols;
1415: /* read in row lengths */
1416: PetscCall(PetscMalloc1(m, &rlens));
1417: PetscCall(PetscViewerBinaryReadAll(viewer, rlens, m, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1418: for (i = 0; i < m; i++) nnz += rlens[i];
1419: /* read in column indices and values */
1420: PetscCall(PetscMalloc2(nnz, &icols, nnz, &vwork));
1421: PetscCall(PetscViewerBinaryReadAll(viewer, icols, nnz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1422: PetscCall(PetscViewerBinaryReadAll(viewer, vwork, nnz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1423: /* store values in column major order */
1424: for (k = 0, i = 0; i < m; i++)
1425: for (j = 0; j < rlens[i]; j++, k++) v[i + lda * icols[k]] = vwork[k];
1426: PetscCall(PetscFree(rlens));
1427: PetscCall(PetscFree2(icols, vwork));
1428: }
1429: PetscCall(MatDenseRestoreArray(mat, &v));
1430: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
1431: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
1432: PetscFunctionReturn(PETSC_SUCCESS);
1433: }
1435: static PetscErrorCode MatLoad_SeqDense(Mat newMat, PetscViewer viewer)
1436: {
1437: PetscBool isbinary, ishdf5;
1439: PetscFunctionBegin;
1442: /* force binary viewer to load .info file if it has not yet done so */
1443: PetscCall(PetscViewerSetUp(viewer));
1444: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1445: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
1446: if (isbinary) {
1447: PetscCall(MatLoad_Dense_Binary(newMat, viewer));
1448: } else if (ishdf5) {
1449: #if defined(PETSC_HAVE_HDF5)
1450: PetscCall(MatLoad_Dense_HDF5(newMat, viewer));
1451: #else
1452: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
1453: #endif
1454: } else {
1455: 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);
1456: }
1457: PetscFunctionReturn(PETSC_SUCCESS);
1458: }
1460: static PetscErrorCode MatView_SeqDense_ASCII(Mat A, PetscViewer viewer)
1461: {
1462: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
1463: PetscInt i, j;
1464: const char *name;
1465: PetscScalar *v, *av;
1466: PetscViewerFormat format;
1467: #if defined(PETSC_USE_COMPLEX)
1468: PetscBool allreal = PETSC_TRUE;
1469: #endif
1471: PetscFunctionBegin;
1472: PetscCall(MatDenseGetArrayRead(A, (const PetscScalar **)&av));
1473: PetscCall(PetscViewerGetFormat(viewer, &format));
1474: if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1475: PetscFunctionReturn(PETSC_SUCCESS); /* do nothing for now */
1476: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
1477: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
1478: for (i = 0; i < A->rmap->n; i++) {
1479: v = av + i;
1480: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
1481: for (j = 0; j < A->cmap->n; j++) {
1482: #if defined(PETSC_USE_COMPLEX)
1483: if (PetscRealPart(*v) != 0.0 && PetscImaginaryPart(*v) != 0.0) {
1484: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i) ", j, (double)PetscRealPart(*v), (double)PetscImaginaryPart(*v)));
1485: } else if (PetscRealPart(*v)) {
1486: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", j, (double)PetscRealPart(*v)));
1487: }
1488: #else
1489: if (*v) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", j, (double)*v));
1490: #endif
1491: v += a->lda;
1492: }
1493: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
1494: }
1495: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
1496: } else {
1497: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
1498: #if defined(PETSC_USE_COMPLEX)
1499: /* determine if matrix has all real values */
1500: for (j = 0; j < A->cmap->n; j++) {
1501: v = av + j * a->lda;
1502: for (i = 0; i < A->rmap->n; i++) {
1503: if (PetscImaginaryPart(v[i])) {
1504: allreal = PETSC_FALSE;
1505: break;
1506: }
1507: }
1508: }
1509: #endif
1510: if (format == PETSC_VIEWER_ASCII_MATLAB) {
1511: PetscCall(PetscObjectGetName((PetscObject)A, &name));
1512: PetscCall(PetscViewerASCIIPrintf(viewer, "%% Size = %" PetscInt_FMT " %" PetscInt_FMT " \n", A->rmap->n, A->cmap->n));
1513: PetscCall(PetscViewerASCIIPrintf(viewer, "%s = zeros(%" PetscInt_FMT ",%" PetscInt_FMT ");\n", name, A->rmap->n, A->cmap->n));
1514: PetscCall(PetscViewerASCIIPrintf(viewer, "%s = [\n", name));
1515: }
1517: for (i = 0; i < A->rmap->n; i++) {
1518: v = av + i;
1519: for (j = 0; j < A->cmap->n; j++) {
1520: #if defined(PETSC_USE_COMPLEX)
1521: if (allreal) {
1522: PetscCall(PetscViewerASCIIPrintf(viewer, "%18.16e ", (double)PetscRealPart(*v)));
1523: } else {
1524: PetscCall(PetscViewerASCIIPrintf(viewer, "%18.16e + %18.16ei ", (double)PetscRealPart(*v), (double)PetscImaginaryPart(*v)));
1525: }
1526: #else
1527: PetscCall(PetscViewerASCIIPrintf(viewer, "%18.16e ", (double)*v));
1528: #endif
1529: v += a->lda;
1530: }
1531: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
1532: }
1533: if (format == PETSC_VIEWER_ASCII_MATLAB) PetscCall(PetscViewerASCIIPrintf(viewer, "];\n"));
1534: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
1535: }
1536: PetscCall(MatDenseRestoreArrayRead(A, (const PetscScalar **)&av));
1537: PetscCall(PetscViewerFlush(viewer));
1538: PetscFunctionReturn(PETSC_SUCCESS);
1539: }
1541: #include <petscdraw.h>
1542: static PetscErrorCode MatView_SeqDense_Draw_Zoom(PetscDraw draw, void *Aa)
1543: {
1544: Mat A = (Mat)Aa;
1545: PetscInt m = A->rmap->n, n = A->cmap->n, i, j;
1546: int color = PETSC_DRAW_WHITE;
1547: const PetscScalar *v;
1548: PetscViewer viewer;
1549: PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r;
1550: PetscViewerFormat format;
1552: PetscFunctionBegin;
1553: PetscCall(PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer));
1554: PetscCall(PetscViewerGetFormat(viewer, &format));
1555: PetscCall(PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr));
1557: /* Loop over matrix elements drawing boxes */
1558: PetscCall(MatDenseGetArrayRead(A, &v));
1559: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
1560: PetscDrawCollectiveBegin(draw);
1561: /* Blue for negative and Red for positive */
1562: for (j = 0; j < n; j++) {
1563: x_l = j;
1564: x_r = x_l + 1.0;
1565: for (i = 0; i < m; i++) {
1566: y_l = m - i - 1.0;
1567: y_r = y_l + 1.0;
1568: if (PetscRealPart(v[j * m + i]) > 0.) color = PETSC_DRAW_RED;
1569: else if (PetscRealPart(v[j * m + i]) < 0.) color = PETSC_DRAW_BLUE;
1570: else continue;
1571: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1572: }
1573: }
1574: PetscDrawCollectiveEnd(draw);
1575: } else {
1576: /* use contour shading to indicate magnitude of values */
1577: /* first determine max of all nonzero values */
1578: PetscReal minv = 0.0, maxv = 0.0;
1579: PetscDraw popup;
1581: for (i = 0; i < m * n; i++) {
1582: if (PetscAbsScalar(v[i]) > maxv) maxv = PetscAbsScalar(v[i]);
1583: }
1584: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1585: PetscCall(PetscDrawGetPopup(draw, &popup));
1586: PetscCall(PetscDrawScalePopup(popup, minv, maxv));
1588: PetscDrawCollectiveBegin(draw);
1589: for (j = 0; j < n; j++) {
1590: x_l = j;
1591: x_r = x_l + 1.0;
1592: for (i = 0; i < m; i++) {
1593: y_l = m - i - 1.0;
1594: y_r = y_l + 1.0;
1595: color = PetscDrawRealToColor(PetscAbsScalar(v[j * m + i]), minv, maxv);
1596: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1597: }
1598: }
1599: PetscDrawCollectiveEnd(draw);
1600: }
1601: PetscCall(MatDenseRestoreArrayRead(A, &v));
1602: PetscFunctionReturn(PETSC_SUCCESS);
1603: }
1605: static PetscErrorCode MatView_SeqDense_Draw(Mat A, PetscViewer viewer)
1606: {
1607: PetscDraw draw;
1608: PetscBool isnull;
1609: PetscReal xr, yr, xl, yl, h, w;
1611: PetscFunctionBegin;
1612: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1613: PetscCall(PetscDrawIsNull(draw, &isnull));
1614: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1616: xr = A->cmap->n;
1617: yr = A->rmap->n;
1618: h = yr / 10.0;
1619: w = xr / 10.0;
1620: xr += w;
1621: yr += h;
1622: xl = -w;
1623: yl = -h;
1624: PetscCall(PetscDrawSetCoordinates(draw, xl, yl, xr, yr));
1625: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer));
1626: PetscCall(PetscDrawZoom(draw, MatView_SeqDense_Draw_Zoom, A));
1627: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL));
1628: PetscCall(PetscDrawSave(draw));
1629: PetscFunctionReturn(PETSC_SUCCESS);
1630: }
1632: PetscErrorCode MatView_SeqDense(Mat A, PetscViewer viewer)
1633: {
1634: PetscBool isascii, isbinary, isdraw;
1636: PetscFunctionBegin;
1637: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1638: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1639: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1640: if (isascii) PetscCall(MatView_SeqDense_ASCII(A, viewer));
1641: else if (isbinary) PetscCall(MatView_Dense_Binary(A, viewer));
1642: else if (isdraw) PetscCall(MatView_SeqDense_Draw(A, viewer));
1643: PetscFunctionReturn(PETSC_SUCCESS);
1644: }
1646: static PetscErrorCode MatDensePlaceArray_SeqDense(Mat A, const PetscScalar *array)
1647: {
1648: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
1650: PetscFunctionBegin;
1651: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1652: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1653: PetscCheck(!a->unplacedarray, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseResetArray() first");
1654: a->unplacedarray = a->v;
1655: a->unplaced_user_alloc = a->user_alloc;
1656: a->v = (PetscScalar *)array;
1657: a->user_alloc = PETSC_TRUE;
1658: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1659: A->offloadmask = PETSC_OFFLOAD_CPU;
1660: #endif
1661: PetscFunctionReturn(PETSC_SUCCESS);
1662: }
1664: static PetscErrorCode MatDenseResetArray_SeqDense(Mat A)
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: a->v = a->unplacedarray;
1672: a->user_alloc = a->unplaced_user_alloc;
1673: a->unplacedarray = NULL;
1674: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1675: A->offloadmask = PETSC_OFFLOAD_CPU;
1676: #endif
1677: PetscFunctionReturn(PETSC_SUCCESS);
1678: }
1680: static PetscErrorCode MatDenseReplaceArray_SeqDense(Mat A, const PetscScalar *array)
1681: {
1682: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
1684: PetscFunctionBegin;
1685: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1686: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1687: if (!a->user_alloc) PetscCall(PetscFree(a->v));
1688: a->v = (PetscScalar *)array;
1689: a->user_alloc = PETSC_FALSE;
1690: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
1691: A->offloadmask = PETSC_OFFLOAD_CPU;
1692: #endif
1693: PetscFunctionReturn(PETSC_SUCCESS);
1694: }
1696: PetscErrorCode MatDestroy_SeqDense(Mat mat)
1697: {
1698: Mat_SeqDense *l = (Mat_SeqDense *)mat->data;
1700: PetscFunctionBegin;
1701: PetscCall(PetscLogObjectState((PetscObject)mat, "Rows %" PetscInt_FMT " Cols %" PetscInt_FMT, mat->rmap->n, mat->cmap->n));
1702: PetscCall(VecDestroy(&l->qrrhs));
1703: PetscCall(PetscFree(l->tau));
1704: PetscCall(PetscFree(l->pivots));
1705: PetscCall(PetscFree(l->fwork));
1706: if (!l->user_alloc) PetscCall(PetscFree(l->v));
1707: if (!l->unplaced_user_alloc) PetscCall(PetscFree(l->unplacedarray));
1708: PetscCheck(!l->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1709: PetscCheck(!l->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1710: PetscCall(VecDestroy(&l->cvec));
1711: PetscCall(MatDestroy(&l->cmat));
1712: PetscCall(PetscFree(mat->data));
1714: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, NULL));
1715: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatQRFactor_C", NULL));
1716: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatQRFactorSymbolic_C", NULL));
1717: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatQRFactorNumeric_C", NULL));
1718: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetLDA_C", NULL));
1719: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseSetLDA_C", NULL));
1720: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArray_C", NULL));
1721: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArray_C", NULL));
1722: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDensePlaceArray_C", NULL));
1723: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseResetArray_C", NULL));
1724: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseReplaceArray_C", NULL));
1725: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayRead_C", NULL));
1726: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayRead_C", NULL));
1727: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayWrite_C", NULL));
1728: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayWrite_C", NULL));
1729: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_seqaij_C", NULL));
1730: #if defined(PETSC_HAVE_ELEMENTAL)
1731: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_elemental_C", NULL));
1732: #endif
1733: #if defined(PETSC_HAVE_SCALAPACK) && (defined(PETSC_USE_REAL_SINGLE) || defined(PETSC_USE_REAL_DOUBLE))
1734: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_scalapack_C", NULL));
1735: #endif
1736: #if defined(PETSC_HAVE_CUDA)
1737: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_seqdensecuda_C", NULL));
1738: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdensecuda_seqdensecuda_C", NULL));
1739: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdensecuda_seqdense_C", NULL));
1740: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdense_seqdensecuda_C", NULL));
1741: #endif
1742: #if defined(PETSC_HAVE_HIP)
1743: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_seqdense_seqdensehip_C", NULL));
1744: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdensehip_seqdensehip_C", NULL));
1745: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdensehip_seqdense_C", NULL));
1746: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdense_seqdensehip_C", NULL));
1747: #endif
1748: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSeqDenseSetPreallocation_C", NULL));
1749: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqaij_seqdense_C", NULL));
1750: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqdense_seqdense_C", NULL));
1751: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqbaij_seqdense_C", NULL));
1752: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_seqsbaij_seqdense_C", NULL));
1754: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumn_C", NULL));
1755: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumn_C", NULL));
1756: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVec_C", NULL));
1757: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVec_C", NULL));
1758: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecRead_C", NULL));
1759: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecRead_C", NULL));
1760: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecWrite_C", NULL));
1761: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecWrite_C", NULL));
1762: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetSubMatrix_C", NULL));
1763: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreSubMatrix_C", NULL));
1764: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultColumnRange_C", NULL));
1765: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultAddColumnRange_C", NULL));
1766: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeColumnRange_C", NULL));
1767: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeAddColumnRange_C", NULL));
1768: PetscFunctionReturn(PETSC_SUCCESS);
1769: }
1771: static PetscErrorCode MatTranspose_SeqDense(Mat A, MatReuse reuse, Mat *matout)
1772: {
1773: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1774: PetscInt k, j, m = A->rmap->n, M = mat->lda, n = A->cmap->n;
1775: PetscScalar *v, tmp;
1777: PetscFunctionBegin;
1778: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1779: if (reuse == MAT_INPLACE_MATRIX) {
1780: if (m == n) { /* in place transpose */
1781: PetscCall(MatDenseGetArray(A, &v));
1782: for (j = 0; j < m; j++) {
1783: for (k = 0; k < j; k++) {
1784: tmp = v[j + k * M];
1785: v[j + k * M] = v[k + j * M];
1786: v[k + j * M] = tmp;
1787: }
1788: }
1789: PetscCall(MatDenseRestoreArray(A, &v));
1790: } else { /* reuse memory, temporary allocates new memory */
1791: PetscScalar *v2;
1792: PetscLayout tmplayout;
1794: PetscCall(PetscMalloc1((size_t)m * n, &v2));
1795: PetscCall(MatDenseGetArray(A, &v));
1796: for (j = 0; j < n; j++) {
1797: for (k = 0; k < m; k++) v2[j + (size_t)k * n] = v[k + (size_t)j * M];
1798: }
1799: PetscCall(PetscArraycpy(v, v2, (size_t)m * n));
1800: PetscCall(PetscFree(v2));
1801: PetscCall(MatDenseRestoreArray(A, &v));
1802: /* cleanup size dependent quantities */
1803: PetscCall(VecDestroy(&mat->cvec));
1804: PetscCall(MatDestroy(&mat->cmat));
1805: PetscCall(PetscFree(mat->pivots));
1806: PetscCall(PetscFree(mat->fwork));
1807: /* swap row/col layouts */
1808: PetscCall(PetscBLASIntCast(n, &mat->lda));
1809: tmplayout = A->rmap;
1810: A->rmap = A->cmap;
1811: A->cmap = tmplayout;
1812: }
1813: } else { /* out-of-place transpose */
1814: Mat tmat;
1815: Mat_SeqDense *tmatd;
1816: PetscScalar *v2;
1817: PetscInt M2;
1819: if (reuse == MAT_INITIAL_MATRIX) {
1820: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &tmat));
1821: PetscCall(MatSetSizes(tmat, A->cmap->n, A->rmap->n, A->cmap->n, A->rmap->n));
1822: PetscCall(MatSetType(tmat, ((PetscObject)A)->type_name));
1823: PetscCall(MatSeqDenseSetPreallocation(tmat, NULL));
1824: } else tmat = *matout;
1826: PetscCall(MatDenseGetArrayRead(A, (const PetscScalar **)&v));
1827: PetscCall(MatDenseGetArray(tmat, &v2));
1828: tmatd = (Mat_SeqDense *)tmat->data;
1829: M2 = tmatd->lda;
1830: for (j = 0; j < n; j++) {
1831: for (k = 0; k < m; k++) v2[j + k * M2] = v[k + j * M];
1832: }
1833: PetscCall(MatDenseRestoreArray(tmat, &v2));
1834: PetscCall(MatDenseRestoreArrayRead(A, (const PetscScalar **)&v));
1835: PetscCall(MatAssemblyBegin(tmat, MAT_FINAL_ASSEMBLY));
1836: PetscCall(MatAssemblyEnd(tmat, MAT_FINAL_ASSEMBLY));
1837: *matout = tmat;
1838: }
1839: PetscFunctionReturn(PETSC_SUCCESS);
1840: }
1842: static PetscErrorCode MatEqual_SeqDense(Mat A1, Mat A2, PetscBool *flg)
1843: {
1844: Mat_SeqDense *mat1 = (Mat_SeqDense *)A1->data;
1845: Mat_SeqDense *mat2 = (Mat_SeqDense *)A2->data;
1846: PetscInt i;
1847: const PetscScalar *v1, *v2;
1849: PetscFunctionBegin;
1850: if (A1->rmap->n != A2->rmap->n) {
1851: *flg = PETSC_FALSE;
1852: PetscFunctionReturn(PETSC_SUCCESS);
1853: }
1854: if (A1->cmap->n != A2->cmap->n) {
1855: *flg = PETSC_FALSE;
1856: PetscFunctionReturn(PETSC_SUCCESS);
1857: }
1858: PetscCall(MatDenseGetArrayRead(A1, &v1));
1859: PetscCall(MatDenseGetArrayRead(A2, &v2));
1860: for (i = 0; i < A1->cmap->n; i++) {
1861: PetscCall(PetscArraycmp(v1, v2, A1->rmap->n, flg));
1862: if (*flg == PETSC_FALSE) PetscFunctionReturn(PETSC_SUCCESS);
1863: v1 += mat1->lda;
1864: v2 += mat2->lda;
1865: }
1866: PetscCall(MatDenseRestoreArrayRead(A1, &v1));
1867: PetscCall(MatDenseRestoreArrayRead(A2, &v2));
1868: *flg = PETSC_TRUE;
1869: PetscFunctionReturn(PETSC_SUCCESS);
1870: }
1872: PetscErrorCode MatGetDiagonal_SeqDense(Mat A, Vec v)
1873: {
1874: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1875: PetscInt i, n, len;
1876: PetscScalar *x;
1877: const PetscScalar *vv;
1879: PetscFunctionBegin;
1880: PetscCall(VecGetSize(v, &n));
1881: PetscCall(VecGetArray(v, &x));
1882: len = PetscMin(A->rmap->n, A->cmap->n);
1883: PetscCall(MatDenseGetArrayRead(A, &vv));
1884: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming mat and vec");
1885: for (i = 0; i < len; i++) x[i] = vv[i * mat->lda + i];
1886: PetscCall(MatDenseRestoreArrayRead(A, &vv));
1887: PetscCall(VecRestoreArray(v, &x));
1888: PetscFunctionReturn(PETSC_SUCCESS);
1889: }
1891: PetscErrorCode MatDiagonalScale_SeqDense(Mat A, Vec ll, Vec rr)
1892: {
1893: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1894: const PetscScalar *l, *r;
1895: PetscScalar x, *v, *vv;
1896: PetscInt i, j, m = A->rmap->n, n = A->cmap->n;
1898: PetscFunctionBegin;
1899: PetscCall(MatDenseGetArray(A, &vv));
1900: if (ll) {
1901: PetscCall(VecGetSize(ll, &m));
1902: PetscCall(VecGetArrayRead(ll, &l));
1903: PetscCheck(m == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vec wrong size");
1904: for (i = 0; i < m; i++) {
1905: x = l[i];
1906: v = vv + i;
1907: for (j = 0; j < n; j++) {
1908: (*v) *= x;
1909: v += mat->lda;
1910: }
1911: }
1912: PetscCall(VecRestoreArrayRead(ll, &l));
1913: PetscCall(PetscLogFlops(1.0 * n * m));
1914: }
1915: if (rr) {
1916: PetscCall(VecGetSize(rr, &n));
1917: PetscCall(VecGetArrayRead(rr, &r));
1918: PetscCheck(n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vec wrong size");
1919: for (i = 0; i < n; i++) {
1920: x = r[i];
1921: v = vv + i * mat->lda;
1922: for (j = 0; j < m; j++) (*v++) *= x;
1923: }
1924: PetscCall(VecRestoreArrayRead(rr, &r));
1925: PetscCall(PetscLogFlops(1.0 * n * m));
1926: }
1927: PetscCall(MatDenseRestoreArray(A, &vv));
1928: PetscFunctionReturn(PETSC_SUCCESS);
1929: }
1931: PetscErrorCode MatNorm_SeqDense(Mat A, NormType type, PetscReal *nrm)
1932: {
1933: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
1934: PetscScalar *v, *vv, *work, *av = NULL;
1935: PetscReal sum = 0.0;
1936: PetscInt lda, m = A->rmap->n, i, j;
1938: PetscFunctionBegin;
1939: PetscCall(MatDenseGetArrayRead(A, (const PetscScalar **)&vv));
1940: PetscCall(MatDenseGetLDA(A, &lda));
1941: v = vv;
1942: if (type == NORM_FROBENIUS) {
1943: if (lda > m) {
1944: for (j = 0; j < A->cmap->n; j++) {
1945: v = vv + j * lda;
1946: for (i = 0; i < m; i++) {
1947: sum += PetscRealPart(PetscConj(*v) * (*v));
1948: v++;
1949: }
1950: }
1951: } else {
1952: #if defined(PETSC_USE_REAL___FP16)
1953: PetscBLASInt one = 1, cnt = A->cmap->n * A->rmap->n;
1954: PetscCallBLAS("BLASnrm2", *nrm = BLASnrm2_(&cnt, v, &one));
1955: }
1956: #else
1957: for (i = 0; i < A->cmap->n * A->rmap->n; i++) {
1958: sum += PetscRealPart(PetscConj(*v) * (*v));
1959: v++;
1960: }
1961: }
1962: *nrm = PetscSqrtReal(sum);
1963: #endif
1964: PetscCall(PetscLogFlops(2.0 * A->cmap->n * A->rmap->n));
1965: } else if (type == NORM_1) {
1966: *nrm = 0.0;
1967: for (j = 0; j < A->cmap->n; j++) {
1968: v = vv + j * mat->lda;
1969: sum = 0.0;
1970: for (i = 0; i < A->rmap->n; i++) {
1971: sum += PetscAbsScalar(*v);
1972: v++;
1973: }
1974: if (sum > *nrm) *nrm = sum;
1975: }
1976: PetscCall(PetscLogFlops(1.0 * A->cmap->n * A->rmap->n));
1977: } else if (type == NORM_INFINITY) {
1978: *nrm = 0.0;
1979: for (j = 0; j < A->rmap->n; j++) {
1980: v = vv + j;
1981: sum = 0.0;
1982: for (i = 0; i < A->cmap->n; i++) {
1983: sum += PetscAbsScalar(*v);
1984: v += mat->lda;
1985: }
1986: if (sum > *nrm) *nrm = sum;
1987: }
1988: PetscCall(PetscLogFlops(1.0 * A->cmap->n * A->rmap->n));
1989: } else if (type == NORM_2) {
1990: PetscReal *s;
1991: PetscBLASInt bm, bn, blda, min, lwork;
1993: PetscCall(PetscBLASIntCast(A->rmap->n, &bm));
1994: PetscCall(PetscBLASIntCast(A->cmap->n, &bn));
1995: PetscCall(PetscBLASIntCast(PetscMax(A->rmap->n, 1), &blda));
1996: min = PetscMin(bm, bn);
1997: if (!min) {
1998: *nrm = 0.0;
1999: PetscCall(MatDenseRestoreArrayRead(A, (const PetscScalar **)&vv));
2000: PetscFunctionReturn(PETSC_SUCCESS);
2001: }
2002: PetscCall(PetscMalloc2(A->rmap->n * A->cmap->n, &av, min, &s));
2003: for (j = 0; j < A->cmap->n; j++) PetscCall(PetscArraycpy(av + j * A->rmap->n, vv + j * lda, A->rmap->n));
2005: lwork = -1;
2006: {
2007: PetscScalar workquery;
2008: #if defined(PETSC_USE_COMPLEX)
2009: PetscReal *rwork;
2011: PetscCall(PetscMalloc1(5 * min, &rwork));
2012: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
2013: PetscCallLAPACKInfo("LAPACKgesvd", LAPACKgesvd_("N", "N", &bm, &bn, av, &blda, s, NULL, &bm, NULL, &min, &workquery, &lwork, rwork, &info));
2014: lwork = (PetscBLASInt)PetscRealPart(workquery);
2015: PetscCall(PetscMalloc1(lwork, &work));
2016: PetscCallLAPACKInfo("LAPACKgesvd", LAPACKgesvd_("N", "N", &bm, &bn, av, &blda, s, NULL, &bm, NULL, &min, work, &lwork, rwork, &info));
2017: PetscCall(PetscFPTrapPop());
2018: PetscCall(PetscFree(rwork));
2019: #else
2020: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
2021: PetscCallLAPACKInfo("LAPACKgesvd", LAPACKgesvd_("N", "N", &bm, &bn, av, &blda, s, NULL, &bm, NULL, &min, &workquery, &lwork, &info));
2022: lwork = (PetscBLASInt)PetscRealPart(workquery);
2023: PetscCall(PetscMalloc1(lwork, &work));
2024: PetscCallLAPACKInfo("LAPACKgesvd", LAPACKgesvd_("N", "N", &bm, &bn, av, &blda, s, NULL, &bm, NULL, &min, work, &lwork, &info));
2025: PetscCall(PetscFPTrapPop());
2026: #endif
2027: }
2028: *nrm = s[0];
2029: PetscCall(PetscFree(work));
2030: PetscCall(PetscFree2(av, s));
2031: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Unsupported norm type %s", NormTypes[type]);
2032: PetscCall(MatDenseRestoreArrayRead(A, (const PetscScalar **)&vv));
2033: PetscFunctionReturn(PETSC_SUCCESS);
2034: }
2036: static PetscErrorCode MatSetOption_SeqDense(Mat A, MatOption op, PetscBool flg)
2037: {
2038: Mat_SeqDense *aij = (Mat_SeqDense *)A->data;
2040: PetscFunctionBegin;
2041: switch (op) {
2042: case MAT_ROW_ORIENTED:
2043: aij->roworiented = flg;
2044: break;
2045: default:
2046: break;
2047: }
2048: PetscFunctionReturn(PETSC_SUCCESS);
2049: }
2051: PetscErrorCode MatZeroEntries_SeqDense(Mat A)
2052: {
2053: Mat_SeqDense *l = (Mat_SeqDense *)A->data;
2054: PetscInt lda = l->lda, m = A->rmap->n, n = A->cmap->n, j;
2055: PetscScalar *v;
2057: PetscFunctionBegin;
2058: PetscCall(MatDenseGetArrayWrite(A, &v));
2059: if (lda > m) {
2060: for (j = 0; j < n; j++) PetscCall(PetscArrayzero(v + j * lda, m));
2061: } else {
2062: PetscCall(PetscArrayzero(v, PetscInt64Mult(m, n)));
2063: }
2064: PetscCall(MatDenseRestoreArrayWrite(A, &v));
2065: PetscFunctionReturn(PETSC_SUCCESS);
2066: }
2068: static PetscErrorCode MatZeroRows_SeqDense(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
2069: {
2070: Mat_SeqDense *l = (Mat_SeqDense *)A->data;
2071: PetscInt m = l->lda, n = A->cmap->n, i, j;
2072: PetscScalar *slot, *bb, *v;
2073: const PetscScalar *xx;
2075: PetscFunctionBegin;
2076: if (PetscDefined(USE_DEBUG)) {
2077: for (i = 0; i < N; i++) {
2078: PetscCheck(rows[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative row requested to be zeroed");
2079: 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);
2080: }
2081: }
2082: if (!N) PetscFunctionReturn(PETSC_SUCCESS);
2084: /* fix right-hand side if needed */
2085: if (x && b) {
2086: PetscCall(VecGetArrayRead(x, &xx));
2087: PetscCall(VecGetArray(b, &bb));
2088: for (i = 0; i < N; i++) bb[rows[i]] = diag * xx[rows[i]];
2089: PetscCall(VecRestoreArrayRead(x, &xx));
2090: PetscCall(VecRestoreArray(b, &bb));
2091: }
2093: PetscCall(MatDenseGetArray(A, &v));
2094: for (i = 0; i < N; i++) {
2095: slot = v + rows[i];
2096: for (j = 0; j < n; j++) {
2097: *slot = 0.0;
2098: slot += m;
2099: }
2100: }
2101: if (diag != 0.0) {
2102: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only coded for square matrices");
2103: for (i = 0; i < N; i++) {
2104: slot = v + (m + 1) * rows[i];
2105: *slot = diag;
2106: }
2107: }
2108: PetscCall(MatDenseRestoreArray(A, &v));
2109: PetscFunctionReturn(PETSC_SUCCESS);
2110: }
2112: static PetscErrorCode MatDenseGetLDA_SeqDense(Mat A, PetscInt *lda)
2113: {
2114: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2116: PetscFunctionBegin;
2117: *lda = mat->lda;
2118: PetscFunctionReturn(PETSC_SUCCESS);
2119: }
2121: PetscErrorCode MatDenseGetArray_SeqDense(Mat A, PetscScalar **array)
2122: {
2123: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2125: PetscFunctionBegin;
2126: PetscCheck(!mat->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
2127: *array = mat->v;
2128: PetscFunctionReturn(PETSC_SUCCESS);
2129: }
2131: PetscErrorCode MatDenseRestoreArray_SeqDense(Mat A, PetscScalar **array)
2132: {
2133: PetscFunctionBegin;
2134: if (array) *array = NULL;
2135: PetscFunctionReturn(PETSC_SUCCESS);
2136: }
2138: /*@
2139: MatDenseGetLDA - gets the leading dimension of the array returned from `MatDenseGetArray()`
2141: Not Collective
2143: Input Parameter:
2144: . A - a `MATDENSE` or `MATDENSECUDA` matrix
2146: Output Parameter:
2147: . lda - the leading dimension
2149: Level: intermediate
2151: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseSetLDA()`
2152: @*/
2153: PetscErrorCode MatDenseGetLDA(Mat A, PetscInt *lda)
2154: {
2155: PetscFunctionBegin;
2157: PetscAssertPointer(lda, 2);
2158: MatCheckPreallocated(A, 1);
2159: PetscUseMethod(A, "MatDenseGetLDA_C", (Mat, PetscInt *), (A, lda));
2160: PetscFunctionReturn(PETSC_SUCCESS);
2161: }
2163: /*@
2164: MatDenseSetLDA - Sets the leading dimension of the array used by the `MATDENSE` matrix
2166: Collective if the matrix layouts have not yet been setup
2168: Input Parameters:
2169: + A - a `MATDENSE` or `MATDENSECUDA` matrix
2170: - lda - the leading dimension
2172: Level: intermediate
2174: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetLDA()`
2175: @*/
2176: PetscErrorCode MatDenseSetLDA(Mat A, PetscInt lda)
2177: {
2178: PetscFunctionBegin;
2180: PetscTryMethod(A, "MatDenseSetLDA_C", (Mat, PetscInt), (A, lda));
2181: PetscFunctionReturn(PETSC_SUCCESS);
2182: }
2184: /*@C
2185: MatDenseGetArray - gives read-write access to the array where the data for a `MATDENSE` matrix is stored
2187: Logically Collective
2189: Input Parameter:
2190: . A - a dense matrix
2192: Output Parameter:
2193: . array - pointer to the data
2195: Level: intermediate
2197: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2198: @*/
2199: PetscErrorCode MatDenseGetArray(Mat A, PetscScalar *array[]) PeNS
2200: {
2201: PetscFunctionBegin;
2203: PetscAssertPointer(array, 2);
2204: PetscUseMethod(A, "MatDenseGetArray_C", (Mat, PetscScalar **), (A, array));
2205: PetscFunctionReturn(PETSC_SUCCESS);
2206: }
2208: /*@C
2209: MatDenseRestoreArray - returns access to the array where the data for a `MATDENSE` matrix is stored obtained by `MatDenseGetArray()`
2211: Logically Collective
2213: Input Parameters:
2214: + A - a dense matrix
2215: - array - pointer to the data (may be `NULL`)
2217: Level: intermediate
2219: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2220: @*/
2221: PetscErrorCode MatDenseRestoreArray(Mat A, PetscScalar *array[]) PeNS
2222: {
2223: PetscFunctionBegin;
2225: if (array) PetscAssertPointer(array, 2);
2226: PetscUseMethod(A, "MatDenseRestoreArray_C", (Mat, PetscScalar **), (A, array));
2227: PetscCall(PetscObjectStateIncrease((PetscObject)A));
2228: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
2229: A->offloadmask = PETSC_OFFLOAD_CPU;
2230: #endif
2231: PetscFunctionReturn(PETSC_SUCCESS);
2232: }
2234: /*@C
2235: MatDenseGetArrayRead - gives read-only access to the array where the data for a `MATDENSE` matrix is stored
2237: Not Collective
2239: Input Parameter:
2240: . A - a dense matrix
2242: Output Parameter:
2243: . array - pointer to the data
2245: Level: intermediate
2247: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayRead()`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2248: @*/
2249: PetscErrorCode MatDenseGetArrayRead(Mat A, const PetscScalar *array[]) PeNS
2250: {
2251: PetscFunctionBegin;
2253: PetscAssertPointer(array, 2);
2254: PetscUseMethod(A, "MatDenseGetArrayRead_C", (Mat, PetscScalar **), (A, (PetscScalar **)array));
2255: PetscFunctionReturn(PETSC_SUCCESS);
2256: }
2258: /*@C
2259: MatDenseRestoreArrayRead - returns access to the array where the data for a `MATDENSE` matrix is stored obtained by `MatDenseGetArrayRead()`
2261: Not Collective
2263: Input Parameters:
2264: + A - a dense matrix
2265: - array - pointer to the data (may be `NULL`)
2267: Level: intermediate
2269: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayRead()`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2270: @*/
2271: PetscErrorCode MatDenseRestoreArrayRead(Mat A, const PetscScalar *array[]) PeNS
2272: {
2273: PetscFunctionBegin;
2275: if (array) PetscAssertPointer(array, 2);
2276: PetscUseMethod(A, "MatDenseRestoreArrayRead_C", (Mat, PetscScalar **), (A, (PetscScalar **)array));
2277: PetscFunctionReturn(PETSC_SUCCESS);
2278: }
2280: /*@C
2281: MatDenseGetArrayWrite - gives write-only access to the array where the data for a `MATDENSE` matrix is stored
2283: Not Collective
2285: Input Parameter:
2286: . A - a dense matrix
2288: Output Parameter:
2289: . array - pointer to the data
2291: Level: intermediate
2293: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayWrite()`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`
2294: @*/
2295: PetscErrorCode MatDenseGetArrayWrite(Mat A, PetscScalar *array[]) PeNS
2296: {
2297: PetscFunctionBegin;
2299: PetscAssertPointer(array, 2);
2300: PetscUseMethod(A, "MatDenseGetArrayWrite_C", (Mat, PetscScalar **), (A, array));
2301: PetscFunctionReturn(PETSC_SUCCESS);
2302: }
2304: /*@C
2305: MatDenseRestoreArrayWrite - returns access to the array where the data for a `MATDENSE` matrix is stored obtained by `MatDenseGetArrayWrite()`
2307: Not Collective
2309: Input Parameters:
2310: + A - a dense matrix
2311: - array - pointer to the data (may be `NULL`)
2313: Level: intermediate
2315: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayWrite()`, `MatDenseGetArray()`, `MatDenseRestoreArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`
2316: @*/
2317: PetscErrorCode MatDenseRestoreArrayWrite(Mat A, PetscScalar *array[]) PeNS
2318: {
2319: PetscFunctionBegin;
2321: if (array) PetscAssertPointer(array, 2);
2322: PetscUseMethod(A, "MatDenseRestoreArrayWrite_C", (Mat, PetscScalar **), (A, array));
2323: PetscCall(PetscObjectStateIncrease((PetscObject)A));
2324: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
2325: A->offloadmask = PETSC_OFFLOAD_CPU;
2326: #endif
2327: PetscFunctionReturn(PETSC_SUCCESS);
2328: }
2330: /*@C
2331: MatDenseGetArrayAndMemType - gives read-write access to the array where the data for a `MATDENSE` matrix is stored
2333: Logically Collective
2335: Input Parameter:
2336: . A - a dense matrix
2338: Output Parameters:
2339: + array - pointer to the data
2340: - mtype - memory type of the returned pointer
2342: Level: intermediate
2344: Note:
2345: If the matrix is of a device type such as `MATDENSECUDA`, `MATDENSEHIP`, etc.,
2346: an array on device is always returned and is guaranteed to contain the matrix's latest data.
2348: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayAndMemType()`, `MatDenseGetArrayReadAndMemType()`, `MatDenseGetArrayWriteAndMemType()`, `MatDenseGetArrayRead()`,
2349: `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`, `MatSeqAIJGetCSRAndMemType()`
2350: @*/
2351: PetscErrorCode MatDenseGetArrayAndMemType(Mat A, PetscScalar *array[], PetscMemType *mtype)
2352: {
2353: PetscBool isMPI;
2355: PetscFunctionBegin;
2357: PetscAssertPointer(array, 2);
2358: 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 */
2359: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2360: if (isMPI) {
2361: /* Dispatch here so that the code can be reused for all subclasses of MATDENSE */
2362: PetscCall(MatDenseGetArrayAndMemType(((Mat_MPIDense *)A->data)->A, array, mtype));
2363: } else {
2364: PetscErrorCode (*fptr)(Mat, PetscScalar **, PetscMemType *);
2366: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseGetArrayAndMemType_C", &fptr));
2367: if (fptr) {
2368: PetscCall((*fptr)(A, array, mtype));
2369: } else {
2370: PetscUseMethod(A, "MatDenseGetArray_C", (Mat, PetscScalar **), (A, array));
2371: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
2372: }
2373: }
2374: PetscFunctionReturn(PETSC_SUCCESS);
2375: }
2377: /*@C
2378: MatDenseRestoreArrayAndMemType - returns access to the array that is obtained by `MatDenseGetArrayAndMemType()`
2380: Logically Collective
2382: Input Parameters:
2383: + A - a dense matrix
2384: - array - pointer to the data
2386: Level: intermediate
2388: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayAndMemType()`, `MatDenseGetArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2389: @*/
2390: PetscErrorCode MatDenseRestoreArrayAndMemType(Mat A, PetscScalar *array[])
2391: {
2392: PetscBool isMPI;
2394: PetscFunctionBegin;
2396: if (array) PetscAssertPointer(array, 2);
2397: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2398: if (isMPI) {
2399: PetscCall(MatDenseRestoreArrayAndMemType(((Mat_MPIDense *)A->data)->A, array));
2400: } else {
2401: PetscErrorCode (*fptr)(Mat, PetscScalar **);
2403: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseRestoreArrayAndMemType_C", &fptr));
2404: if (fptr) {
2405: PetscCall((*fptr)(A, array));
2406: } else {
2407: PetscUseMethod(A, "MatDenseRestoreArray_C", (Mat, PetscScalar **), (A, array));
2408: }
2409: if (array) *array = NULL;
2410: }
2411: PetscCall(PetscObjectStateIncrease((PetscObject)A));
2412: PetscFunctionReturn(PETSC_SUCCESS);
2413: }
2415: /*@C
2416: MatDenseGetArrayReadAndMemType - gives read-only access to the array where the data for a `MATDENSE` matrix is stored
2418: Logically Collective
2420: Input Parameter:
2421: . A - a dense matrix
2423: Output Parameters:
2424: + array - pointer to the data
2425: - mtype - memory type of the returned pointer
2427: Level: intermediate
2429: Note:
2430: If the matrix is of a device type such as `MATDENSECUDA`, `MATDENSEHIP`, etc.,
2431: an array on device is always returned and is guaranteed to contain the matrix's latest data.
2433: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayReadAndMemType()`, `MatDenseGetArrayWriteAndMemType()`,
2434: `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`, `MatSeqAIJGetCSRAndMemType()`
2435: @*/
2436: PetscErrorCode MatDenseGetArrayReadAndMemType(Mat A, const PetscScalar *array[], PetscMemType *mtype)
2437: {
2438: PetscBool isMPI;
2440: PetscFunctionBegin;
2442: PetscAssertPointer(array, 2);
2443: 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 */
2444: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2445: if (isMPI) { /* Dispatch here so that the code can be reused for all subclasses of MATDENSE */
2446: PetscCall(MatDenseGetArrayReadAndMemType(((Mat_MPIDense *)A->data)->A, array, mtype));
2447: } else {
2448: PetscErrorCode (*fptr)(Mat, const PetscScalar **, PetscMemType *);
2450: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseGetArrayReadAndMemType_C", &fptr));
2451: if (fptr) {
2452: PetscCall((*fptr)(A, array, mtype));
2453: } else {
2454: PetscUseMethod(A, "MatDenseGetArrayRead_C", (Mat, PetscScalar **), (A, (PetscScalar **)array));
2455: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
2456: }
2457: }
2458: PetscFunctionReturn(PETSC_SUCCESS);
2459: }
2461: /*@C
2462: MatDenseRestoreArrayReadAndMemType - returns access to the array that is obtained by `MatDenseGetArrayReadAndMemType()`
2464: Logically Collective
2466: Input Parameters:
2467: + A - a dense matrix
2468: - array - pointer to the data
2470: Level: intermediate
2472: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayReadAndMemType()`, `MatDenseGetArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2473: @*/
2474: PetscErrorCode MatDenseRestoreArrayReadAndMemType(Mat A, const PetscScalar *array[])
2475: {
2476: PetscBool isMPI;
2478: PetscFunctionBegin;
2480: if (array) PetscAssertPointer(array, 2);
2481: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2482: if (isMPI) {
2483: PetscCall(MatDenseRestoreArrayReadAndMemType(((Mat_MPIDense *)A->data)->A, array));
2484: } else {
2485: PetscErrorCode (*fptr)(Mat, const PetscScalar **);
2487: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseRestoreArrayReadAndMemType_C", &fptr));
2488: if (fptr) {
2489: PetscCall((*fptr)(A, array));
2490: } else {
2491: PetscUseMethod(A, "MatDenseRestoreArrayRead_C", (Mat, PetscScalar **), (A, (PetscScalar **)array));
2492: }
2493: if (array) *array = NULL;
2494: }
2495: PetscFunctionReturn(PETSC_SUCCESS);
2496: }
2498: /*@C
2499: MatDenseGetArrayWriteAndMemType - gives write-only access to the array where the data for a `MATDENSE` matrix is stored
2501: Logically Collective
2503: Input Parameter:
2504: . A - a dense matrix
2506: Output Parameters:
2507: + array - pointer to the data
2508: - mtype - memory type of the returned pointer
2510: Level: intermediate
2512: Note:
2513: If the matrix is of a device type such as `MATDENSECUDA`, `MATDENSEHIP`, etc.,
2514: an array on device is always returned and is guaranteed to contain the matrix's latest data.
2516: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreArrayWriteAndMemType()`, `MatDenseGetArrayReadAndMemType()`, `MatDenseGetArrayRead()`,
2517: `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`, `MatSeqAIJGetCSRAndMemType()`
2518: @*/
2519: PetscErrorCode MatDenseGetArrayWriteAndMemType(Mat A, PetscScalar *array[], PetscMemType *mtype)
2520: {
2521: PetscBool isMPI;
2523: PetscFunctionBegin;
2525: PetscAssertPointer(array, 2);
2526: 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 */
2527: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2528: if (isMPI) {
2529: PetscCall(MatDenseGetArrayWriteAndMemType(((Mat_MPIDense *)A->data)->A, array, mtype));
2530: } else {
2531: PetscErrorCode (*fptr)(Mat, PetscScalar **, PetscMemType *);
2533: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseGetArrayWriteAndMemType_C", &fptr));
2534: if (fptr) {
2535: PetscCall((*fptr)(A, array, mtype));
2536: } else {
2537: PetscUseMethod(A, "MatDenseGetArrayWrite_C", (Mat, PetscScalar **), (A, array));
2538: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
2539: }
2540: }
2541: PetscFunctionReturn(PETSC_SUCCESS);
2542: }
2544: /*@C
2545: MatDenseRestoreArrayWriteAndMemType - returns access to the array that is obtained by `MatDenseGetArrayReadAndMemType()`
2547: Logically Collective
2549: Input Parameters:
2550: + A - a dense matrix
2551: - array - pointer to the data
2553: Level: intermediate
2555: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArrayWriteAndMemType()`, `MatDenseGetArray()`, `MatDenseGetArrayRead()`, `MatDenseRestoreArrayRead()`, `MatDenseGetArrayWrite()`, `MatDenseRestoreArrayWrite()`
2556: @*/
2557: PetscErrorCode MatDenseRestoreArrayWriteAndMemType(Mat A, PetscScalar *array[])
2558: {
2559: PetscBool isMPI;
2561: PetscFunctionBegin;
2563: if (array) PetscAssertPointer(array, 2);
2564: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &isMPI));
2565: if (isMPI) {
2566: PetscCall(MatDenseRestoreArrayWriteAndMemType(((Mat_MPIDense *)A->data)->A, array));
2567: } else {
2568: PetscErrorCode (*fptr)(Mat, PetscScalar **);
2570: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatDenseRestoreArrayWriteAndMemType_C", &fptr));
2571: if (fptr) {
2572: PetscCall((*fptr)(A, array));
2573: } else {
2574: PetscUseMethod(A, "MatDenseRestoreArrayWrite_C", (Mat, PetscScalar **), (A, array));
2575: }
2576: if (array) *array = NULL;
2577: }
2578: PetscCall(PetscObjectStateIncrease((PetscObject)A));
2579: PetscFunctionReturn(PETSC_SUCCESS);
2580: }
2582: static PetscErrorCode MatCreateSubMatrix_SeqDense(Mat A, IS isrow, IS iscol, MatReuse scall, Mat *B)
2583: {
2584: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2585: PetscInt i, j, nrows, ncols, ldb;
2586: const PetscInt *irow, *icol;
2587: PetscScalar *av, *bv, *v = mat->v;
2588: Mat newmat;
2590: PetscFunctionBegin;
2591: PetscCall(ISGetIndices(isrow, &irow));
2592: PetscCall(ISGetIndices(iscol, &icol));
2593: PetscCall(ISGetLocalSize(isrow, &nrows));
2594: PetscCall(ISGetLocalSize(iscol, &ncols));
2596: /* Check submatrixcall */
2597: if (scall == MAT_REUSE_MATRIX) {
2598: PetscInt n_cols, n_rows;
2599: PetscCall(MatGetSize(*B, &n_rows, &n_cols));
2600: if (n_rows != nrows || n_cols != ncols) {
2601: /* resize the result matrix to match number of requested rows/columns */
2602: PetscCall(MatSetSizes(*B, nrows, ncols, nrows, ncols));
2603: }
2604: newmat = *B;
2605: } else {
2606: /* Create and fill new matrix */
2607: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &newmat));
2608: PetscCall(MatSetSizes(newmat, nrows, ncols, nrows, ncols));
2609: PetscCall(MatSetType(newmat, ((PetscObject)A)->type_name));
2610: PetscCall(MatSeqDenseSetPreallocation(newmat, NULL));
2611: }
2613: /* Now extract the data pointers and do the copy,column at a time */
2614: PetscCall(MatDenseGetArray(newmat, &bv));
2615: PetscCall(MatDenseGetLDA(newmat, &ldb));
2616: for (i = 0; i < ncols; i++) {
2617: av = v + mat->lda * icol[i];
2618: for (j = 0; j < nrows; j++) bv[j] = av[irow[j]];
2619: bv += ldb;
2620: }
2621: PetscCall(MatDenseRestoreArray(newmat, &bv));
2623: /* Assemble the matrices so that the correct flags are set */
2624: PetscCall(MatAssemblyBegin(newmat, MAT_FINAL_ASSEMBLY));
2625: PetscCall(MatAssemblyEnd(newmat, MAT_FINAL_ASSEMBLY));
2627: /* Free work space */
2628: PetscCall(ISRestoreIndices(isrow, &irow));
2629: PetscCall(ISRestoreIndices(iscol, &icol));
2630: *B = newmat;
2631: PetscFunctionReturn(PETSC_SUCCESS);
2632: }
2634: static PetscErrorCode MatCreateSubMatrices_SeqDense(Mat A, PetscInt n, const IS irow[], const IS icol[], MatReuse scall, Mat *B[])
2635: {
2636: PetscInt i;
2638: PetscFunctionBegin;
2639: if (scall == MAT_INITIAL_MATRIX) PetscCall(PetscCalloc1(n, B));
2641: for (i = 0; i < n; i++) PetscCall(MatCreateSubMatrix_SeqDense(A, irow[i], icol[i], scall, &(*B)[i]));
2642: PetscFunctionReturn(PETSC_SUCCESS);
2643: }
2645: PetscErrorCode MatCopy_SeqDense(Mat A, Mat B, MatStructure str)
2646: {
2647: Mat_SeqDense *a = (Mat_SeqDense *)A->data, *b = (Mat_SeqDense *)B->data;
2648: const PetscScalar *va;
2649: PetscScalar *vb;
2650: PetscInt lda1 = a->lda, lda2 = b->lda, m = A->rmap->n, n = A->cmap->n, j;
2652: PetscFunctionBegin;
2653: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2654: if (A->ops->copy != B->ops->copy) {
2655: PetscCall(MatCopy_Basic(A, B, str));
2656: PetscFunctionReturn(PETSC_SUCCESS);
2657: }
2658: PetscCheck(m == B->rmap->n && n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "size(B) != size(A)");
2659: PetscCall(MatDenseGetArrayRead(A, &va));
2660: PetscCall(MatDenseGetArray(B, &vb));
2661: if (lda1 > m || lda2 > m) {
2662: for (j = 0; j < n; j++) PetscCall(PetscArraycpy(vb + j * lda2, va + j * lda1, m));
2663: } else {
2664: PetscCall(PetscArraycpy(vb, va, A->rmap->n * A->cmap->n));
2665: }
2666: PetscCall(MatDenseRestoreArray(B, &vb));
2667: PetscCall(MatDenseRestoreArrayRead(A, &va));
2668: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
2669: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
2670: PetscFunctionReturn(PETSC_SUCCESS);
2671: }
2673: PetscErrorCode MatSetUp_SeqDense(Mat A)
2674: {
2675: PetscFunctionBegin;
2676: PetscCall(PetscLayoutSetUp(A->rmap));
2677: PetscCall(PetscLayoutSetUp(A->cmap));
2678: if (!A->preallocated) PetscCall(MatSeqDenseSetPreallocation(A, NULL));
2679: PetscFunctionReturn(PETSC_SUCCESS);
2680: }
2682: PetscErrorCode MatConjugate_SeqDense(Mat A)
2683: {
2684: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2685: PetscInt i, j;
2686: PetscInt min = PetscMin(A->rmap->n, A->cmap->n);
2687: PetscScalar *aa;
2689: PetscFunctionBegin;
2690: PetscCall(MatDenseGetArray(A, &aa));
2691: for (j = 0; j < A->cmap->n; j++)
2692: for (i = 0; i < A->rmap->n; i++) aa[i + j * mat->lda] = PetscConj(aa[i + j * mat->lda]);
2693: PetscCall(MatDenseRestoreArray(A, &aa));
2694: if (mat->tau)
2695: for (i = 0; i < min; i++) mat->tau[i] = PetscConj(mat->tau[i]);
2696: PetscFunctionReturn(PETSC_SUCCESS);
2697: }
2699: static PetscErrorCode MatRealPart_SeqDense(Mat A)
2700: {
2701: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2702: PetscInt i, j;
2703: PetscScalar *aa;
2705: PetscFunctionBegin;
2706: PetscCall(MatDenseGetArray(A, &aa));
2707: for (j = 0; j < A->cmap->n; j++) {
2708: for (i = 0; i < A->rmap->n; i++) aa[i + j * mat->lda] = PetscRealPart(aa[i + j * mat->lda]);
2709: }
2710: PetscCall(MatDenseRestoreArray(A, &aa));
2711: PetscFunctionReturn(PETSC_SUCCESS);
2712: }
2714: static PetscErrorCode MatImaginaryPart_SeqDense(Mat A)
2715: {
2716: Mat_SeqDense *mat = (Mat_SeqDense *)A->data;
2717: PetscInt i, j;
2718: PetscScalar *aa;
2720: PetscFunctionBegin;
2721: PetscCall(MatDenseGetArray(A, &aa));
2722: for (j = 0; j < A->cmap->n; j++) {
2723: for (i = 0; i < A->rmap->n; i++) aa[i + j * mat->lda] = PetscImaginaryPart(aa[i + j * mat->lda]);
2724: }
2725: PetscCall(MatDenseRestoreArray(A, &aa));
2726: PetscFunctionReturn(PETSC_SUCCESS);
2727: }
2729: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqDense(Mat A, Mat B, PetscReal fill, Mat C)
2730: {
2731: PetscInt m = A->rmap->n, n = B->cmap->n;
2732: PetscBool cisdense = PETSC_FALSE;
2734: PetscFunctionBegin;
2735: PetscCall(MatSetSizes(C, m, n, m, n));
2736: #if defined(PETSC_HAVE_CUDA)
2737: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, ""));
2738: #endif
2739: #if defined(PETSC_HAVE_HIP)
2740: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSEHIP, ""));
2741: #endif
2742: if (!cisdense) {
2743: PetscBool flg;
2745: PetscCall(PetscObjectTypeCompare((PetscObject)B, ((PetscObject)A)->type_name, &flg));
2746: PetscCall(MatSetType(C, flg ? ((PetscObject)A)->type_name : MATDENSE));
2747: }
2748: PetscCall(MatSetUp(C));
2749: PetscFunctionReturn(PETSC_SUCCESS);
2750: }
2752: PetscErrorCode MatMatMultNumeric_SeqDense_SeqDense(Mat A, Mat B, Mat C)
2753: {
2754: Mat_SeqDense *a = (Mat_SeqDense *)A->data, *b = (Mat_SeqDense *)B->data, *c = (Mat_SeqDense *)C->data;
2755: const PetscScalar *av, *bv;
2756: PetscScalar *cv;
2757: PetscBLASInt m, n, k;
2758: PetscScalar _DOne = 1.0, _DZero = 0.0;
2760: PetscFunctionBegin;
2761: PetscCall(PetscBLASIntCast(C->rmap->n, &m));
2762: PetscCall(PetscBLASIntCast(C->cmap->n, &n));
2763: PetscCall(PetscBLASIntCast(A->cmap->n, &k));
2764: if (!m || !n || !k) {
2765: PetscCall(MatZeroEntries(C));
2766: PetscFunctionReturn(PETSC_SUCCESS);
2767: }
2768: PetscCall(MatDenseGetArrayRead(A, &av));
2769: PetscCall(MatDenseGetArrayRead(B, &bv));
2770: PetscCall(MatDenseGetArrayWrite(C, &cv));
2771: PetscCallBLAS("BLASgemm", BLASgemm_("N", "N", &m, &n, &k, &_DOne, av, &a->lda, bv, &b->lda, &_DZero, cv, &c->lda));
2772: PetscCall(MatDenseRestoreArrayRead(A, &av));
2773: PetscCall(MatDenseRestoreArrayRead(B, &bv));
2774: PetscCall(MatDenseRestoreArrayWrite(C, &cv));
2775: PetscCall(PetscLogFlops(1.0 * m * n * k + 1.0 * m * n * (k - 1)));
2776: PetscFunctionReturn(PETSC_SUCCESS);
2777: }
2779: PetscErrorCode MatMatTransposeMultSymbolic_SeqDense_SeqDense(Mat A, Mat B, PetscReal fill, Mat C)
2780: {
2781: PetscInt m = A->rmap->n, n = B->rmap->n;
2782: PetscBool cisdense = PETSC_FALSE;
2784: PetscFunctionBegin;
2785: PetscCall(MatSetSizes(C, m, n, m, n));
2786: #if defined(PETSC_HAVE_CUDA)
2787: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, ""));
2788: #endif
2789: #if defined(PETSC_HAVE_HIP)
2790: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSEHIP, ""));
2791: #endif
2792: if (!cisdense) {
2793: PetscBool flg;
2795: PetscCall(PetscObjectTypeCompare((PetscObject)B, ((PetscObject)A)->type_name, &flg));
2796: PetscCall(MatSetType(C, flg ? ((PetscObject)A)->type_name : MATDENSE));
2797: }
2798: PetscCall(MatSetUp(C));
2799: PetscFunctionReturn(PETSC_SUCCESS);
2800: }
2802: PetscErrorCode MatMatTransposeMultNumeric_SeqDense_SeqDense(Mat A, Mat B, Mat C)
2803: {
2804: Mat_SeqDense *a = (Mat_SeqDense *)A->data, *b = (Mat_SeqDense *)B->data, *c = (Mat_SeqDense *)C->data;
2805: const PetscScalar *av, *bv;
2806: PetscScalar *cv;
2807: PetscBLASInt m, n, k;
2808: PetscScalar _DOne = 1.0, _DZero = 0.0;
2810: PetscFunctionBegin;
2811: PetscCall(PetscBLASIntCast(C->rmap->n, &m));
2812: PetscCall(PetscBLASIntCast(C->cmap->n, &n));
2813: PetscCall(PetscBLASIntCast(A->cmap->n, &k));
2814: if (!m || !n || !k) {
2815: PetscCall(MatZeroEntries(C));
2816: PetscFunctionReturn(PETSC_SUCCESS);
2817: }
2818: PetscCall(MatDenseGetArrayRead(A, &av));
2819: PetscCall(MatDenseGetArrayRead(B, &bv));
2820: PetscCall(MatDenseGetArrayWrite(C, &cv));
2821: PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &m, &n, &k, &_DOne, av, &a->lda, bv, &b->lda, &_DZero, cv, &c->lda));
2822: PetscCall(MatDenseRestoreArrayRead(A, &av));
2823: PetscCall(MatDenseRestoreArrayRead(B, &bv));
2824: PetscCall(MatDenseRestoreArrayWrite(C, &cv));
2825: PetscCall(PetscLogFlops(1.0 * m * n * k + 1.0 * m * n * (k - 1)));
2826: PetscFunctionReturn(PETSC_SUCCESS);
2827: }
2829: PetscErrorCode MatTransposeMatMultSymbolic_SeqDense_SeqDense(Mat A, Mat B, PetscReal fill, Mat C)
2830: {
2831: PetscInt m = A->cmap->n, n = B->cmap->n;
2832: PetscBool cisdense = PETSC_FALSE;
2834: PetscFunctionBegin;
2835: PetscCall(MatSetSizes(C, m, n, m, n));
2836: #if defined(PETSC_HAVE_CUDA)
2837: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, ""));
2838: #endif
2839: #if defined(PETSC_HAVE_HIP)
2840: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSEHIP, ""));
2841: #endif
2842: if (!cisdense) {
2843: PetscBool flg;
2845: PetscCall(PetscObjectTypeCompare((PetscObject)B, ((PetscObject)A)->type_name, &flg));
2846: PetscCall(MatSetType(C, flg ? ((PetscObject)A)->type_name : MATDENSE));
2847: }
2848: PetscCall(MatSetUp(C));
2849: PetscFunctionReturn(PETSC_SUCCESS);
2850: }
2852: PetscErrorCode MatTransposeMatMultNumeric_SeqDense_SeqDense(Mat A, Mat B, Mat C)
2853: {
2854: Mat_SeqDense *a = (Mat_SeqDense *)A->data, *b = (Mat_SeqDense *)B->data, *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) {
2865: PetscCall(MatZeroEntries(C));
2866: PetscFunctionReturn(PETSC_SUCCESS);
2867: }
2868: PetscCall(MatDenseGetArrayRead(A, &av));
2869: PetscCall(MatDenseGetArrayRead(B, &bv));
2870: PetscCall(MatDenseGetArrayWrite(C, &cv));
2871: PetscCallBLAS("BLASgemm", BLASgemm_("T", "N", &m, &n, &k, &_DOne, av, &a->lda, bv, &b->lda, &_DZero, cv, &c->lda));
2872: PetscCall(MatDenseRestoreArrayRead(A, &av));
2873: PetscCall(MatDenseRestoreArrayRead(B, &bv));
2874: PetscCall(MatDenseRestoreArrayWrite(C, &cv));
2875: PetscCall(PetscLogFlops(1.0 * m * n * k + 1.0 * m * n * (k - 1)));
2876: PetscFunctionReturn(PETSC_SUCCESS);
2877: }
2879: static PetscErrorCode MatProductSetFromOptions_SeqDense_AB(Mat C)
2880: {
2881: PetscFunctionBegin;
2882: C->ops->matmultsymbolic = MatMatMultSymbolic_SeqDense_SeqDense;
2883: C->ops->productsymbolic = MatProductSymbolic_AB;
2884: PetscFunctionReturn(PETSC_SUCCESS);
2885: }
2887: static PetscErrorCode MatProductSetFromOptions_SeqDense_AtB(Mat C)
2888: {
2889: PetscFunctionBegin;
2890: C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_SeqDense_SeqDense;
2891: C->ops->productsymbolic = MatProductSymbolic_AtB;
2892: PetscFunctionReturn(PETSC_SUCCESS);
2893: }
2895: static PetscErrorCode MatProductSetFromOptions_SeqDense_ABt(Mat C)
2896: {
2897: PetscFunctionBegin;
2898: C->ops->mattransposemultsymbolic = MatMatTransposeMultSymbolic_SeqDense_SeqDense;
2899: C->ops->productsymbolic = MatProductSymbolic_ABt;
2900: PetscFunctionReturn(PETSC_SUCCESS);
2901: }
2903: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqDense(Mat C)
2904: {
2905: Mat_Product *product = C->product;
2907: PetscFunctionBegin;
2908: switch (product->type) {
2909: case MATPRODUCT_AB:
2910: PetscCall(MatProductSetFromOptions_SeqDense_AB(C));
2911: break;
2912: case MATPRODUCT_AtB:
2913: PetscCall(MatProductSetFromOptions_SeqDense_AtB(C));
2914: break;
2915: case MATPRODUCT_ABt:
2916: PetscCall(MatProductSetFromOptions_SeqDense_ABt(C));
2917: break;
2918: default:
2919: break;
2920: }
2921: PetscFunctionReturn(PETSC_SUCCESS);
2922: }
2924: static PetscErrorCode MatGetRowMax_SeqDense(Mat A, Vec v, PetscInt idx[])
2925: {
2926: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
2927: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, p;
2928: PetscScalar *x;
2929: const PetscScalar *aa;
2931: PetscFunctionBegin;
2932: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2933: PetscCall(VecGetArray(v, &x));
2934: PetscCall(VecGetLocalSize(v, &p));
2935: PetscCall(MatDenseGetArrayRead(A, &aa));
2936: PetscCheck(p == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
2937: for (i = 0; i < m; i++) {
2938: x[i] = aa[i];
2939: if (idx) idx[i] = 0;
2940: for (j = 1; j < n; j++) {
2941: if (PetscRealPart(x[i]) < PetscRealPart(aa[i + a->lda * j])) {
2942: x[i] = aa[i + a->lda * j];
2943: if (idx) idx[i] = j;
2944: }
2945: }
2946: }
2947: PetscCall(MatDenseRestoreArrayRead(A, &aa));
2948: PetscCall(VecRestoreArray(v, &x));
2949: PetscFunctionReturn(PETSC_SUCCESS);
2950: }
2952: static PetscErrorCode MatGetRowMaxAbs_SeqDense(Mat A, Vec v, PetscInt idx[])
2953: {
2954: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
2955: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, p;
2956: PetscScalar *x;
2957: PetscReal atmp;
2958: const PetscScalar *aa;
2960: PetscFunctionBegin;
2961: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2962: PetscCall(VecGetArray(v, &x));
2963: PetscCall(VecGetLocalSize(v, &p));
2964: PetscCall(MatDenseGetArrayRead(A, &aa));
2965: PetscCheck(p == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
2966: for (i = 0; i < m; i++) {
2967: x[i] = PetscAbsScalar(aa[i]);
2968: for (j = 1; j < n; j++) {
2969: atmp = PetscAbsScalar(aa[i + a->lda * j]);
2970: if (PetscAbsScalar(x[i]) < atmp) {
2971: x[i] = atmp;
2972: if (idx) idx[i] = j;
2973: }
2974: }
2975: }
2976: PetscCall(MatDenseRestoreArrayRead(A, &aa));
2977: PetscCall(VecRestoreArray(v, &x));
2978: PetscFunctionReturn(PETSC_SUCCESS);
2979: }
2981: static PetscErrorCode MatGetRowMin_SeqDense(Mat A, Vec v, PetscInt idx[])
2982: {
2983: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
2984: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, p;
2985: PetscScalar *x;
2986: const PetscScalar *aa;
2988: PetscFunctionBegin;
2989: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2990: PetscCall(MatDenseGetArrayRead(A, &aa));
2991: PetscCall(VecGetArray(v, &x));
2992: PetscCall(VecGetLocalSize(v, &p));
2993: PetscCheck(p == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
2994: for (i = 0; i < m; i++) {
2995: x[i] = aa[i];
2996: if (idx) idx[i] = 0;
2997: for (j = 1; j < n; j++) {
2998: if (PetscRealPart(x[i]) > PetscRealPart(aa[i + a->lda * j])) {
2999: x[i] = aa[i + a->lda * j];
3000: if (idx) idx[i] = j;
3001: }
3002: }
3003: }
3004: PetscCall(VecRestoreArray(v, &x));
3005: PetscCall(MatDenseRestoreArrayRead(A, &aa));
3006: PetscFunctionReturn(PETSC_SUCCESS);
3007: }
3009: PetscErrorCode MatGetColumnVector_SeqDense(Mat A, Vec v, PetscInt col)
3010: {
3011: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3012: PetscScalar *x;
3013: const PetscScalar *aa;
3015: PetscFunctionBegin;
3016: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3017: PetscCall(MatDenseGetArrayRead(A, &aa));
3018: PetscCall(VecGetArray(v, &x));
3019: PetscCall(PetscArraycpy(x, aa + col * a->lda, A->rmap->n));
3020: PetscCall(VecRestoreArray(v, &x));
3021: PetscCall(MatDenseRestoreArrayRead(A, &aa));
3022: PetscFunctionReturn(PETSC_SUCCESS);
3023: }
3025: PETSC_INTERN PetscErrorCode MatGetColumnReductions_SeqDense(Mat A, PetscInt type, PetscReal *reductions)
3026: {
3027: PetscInt i, j, m, n;
3028: const PetscScalar *a;
3030: PetscFunctionBegin;
3031: PetscCall(MatGetSize(A, &m, &n));
3032: PetscCall(PetscArrayzero(reductions, n));
3033: PetscCall(MatDenseGetArrayRead(A, &a));
3034: if (type == NORM_2) {
3035: for (i = 0; i < n; i++) {
3036: for (j = 0; j < m; j++) reductions[i] += PetscAbsScalar(a[j] * a[j]);
3037: a = PetscSafePointerPlusOffset(a, m);
3038: }
3039: } else if (type == NORM_1) {
3040: for (i = 0; i < n; i++) {
3041: for (j = 0; j < m; j++) reductions[i] += PetscAbsScalar(a[j]);
3042: a = PetscSafePointerPlusOffset(a, m);
3043: }
3044: } else if (type == NORM_INFINITY) {
3045: for (i = 0; i < n; i++) {
3046: for (j = 0; j < m; j++) reductions[i] = PetscMax(PetscAbsScalar(a[j]), reductions[i]);
3047: a = PetscSafePointerPlusOffset(a, m);
3048: }
3049: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
3050: for (i = 0; i < n; i++) {
3051: for (j = 0; j < m; j++) reductions[i] += PetscRealPart(a[j]);
3052: a = PetscSafePointerPlusOffset(a, m);
3053: }
3054: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
3055: for (i = 0; i < n; i++) {
3056: for (j = 0; j < m; j++) reductions[i] += PetscImaginaryPart(a[j]);
3057: a = PetscSafePointerPlusOffset(a, m);
3058: }
3059: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
3060: PetscCall(MatDenseRestoreArrayRead(A, &a));
3061: if (type == NORM_2) {
3062: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
3063: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
3064: for (i = 0; i < n; i++) reductions[i] /= m;
3065: }
3066: PetscFunctionReturn(PETSC_SUCCESS);
3067: }
3069: PetscErrorCode MatSetRandom_SeqDense(Mat x, PetscRandom rctx)
3070: {
3071: PetscScalar *a;
3072: PetscInt lda, m, n, i, j;
3074: PetscFunctionBegin;
3075: PetscCall(MatGetSize(x, &m, &n));
3076: PetscCall(MatDenseGetLDA(x, &lda));
3077: PetscCall(MatDenseGetArrayWrite(x, &a));
3078: for (j = 0; j < n; j++) {
3079: for (i = 0; i < m; i++) PetscCall(PetscRandomGetValue(rctx, a + j * lda + i));
3080: }
3081: PetscCall(MatDenseRestoreArrayWrite(x, &a));
3082: PetscFunctionReturn(PETSC_SUCCESS);
3083: }
3085: /* vals is not const */
3086: static PetscErrorCode MatDenseGetColumn_SeqDense(Mat A, PetscInt col, PetscScalar **vals)
3087: {
3088: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3089: PetscScalar *v;
3091: PetscFunctionBegin;
3092: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3093: PetscCall(MatDenseGetArray(A, &v));
3094: *vals = v + col * a->lda;
3095: PetscCall(MatDenseRestoreArray(A, &v));
3096: PetscFunctionReturn(PETSC_SUCCESS);
3097: }
3099: static PetscErrorCode MatDenseRestoreColumn_SeqDense(Mat A, PetscScalar **vals)
3100: {
3101: PetscFunctionBegin;
3102: if (vals) *vals = NULL; /* user cannot accidentally use the array later */
3103: PetscFunctionReturn(PETSC_SUCCESS);
3104: }
3106: static struct _MatOps MatOps_Values = {MatSetValues_SeqDense,
3107: MatGetRow_SeqDense,
3108: MatRestoreRow_SeqDense,
3109: MatMult_SeqDense,
3110: /* 4*/ MatMultAdd_SeqDense,
3111: MatMultTranspose_SeqDense,
3112: MatMultTransposeAdd_SeqDense,
3113: NULL,
3114: NULL,
3115: NULL,
3116: /* 10*/ NULL,
3117: MatLUFactor_SeqDense,
3118: MatCholeskyFactor_SeqDense,
3119: MatSOR_SeqDense,
3120: MatTranspose_SeqDense,
3121: /* 15*/ MatGetInfo_SeqDense,
3122: MatEqual_SeqDense,
3123: MatGetDiagonal_SeqDense,
3124: MatDiagonalScale_SeqDense,
3125: MatNorm_SeqDense,
3126: /* 20*/ NULL,
3127: NULL,
3128: MatSetOption_SeqDense,
3129: MatZeroEntries_SeqDense,
3130: /* 24*/ MatZeroRows_SeqDense,
3131: NULL,
3132: NULL,
3133: NULL,
3134: NULL,
3135: /* 29*/ MatSetUp_SeqDense,
3136: NULL,
3137: NULL,
3138: NULL,
3139: NULL,
3140: /* 34*/ MatDuplicate_SeqDense,
3141: NULL,
3142: NULL,
3143: NULL,
3144: NULL,
3145: /* 39*/ MatAXPY_SeqDense,
3146: MatCreateSubMatrices_SeqDense,
3147: NULL,
3148: MatGetValues_SeqDense,
3149: MatCopy_SeqDense,
3150: /* 44*/ MatGetRowMax_SeqDense,
3151: MatScale_SeqDense,
3152: MatShift_SeqDense,
3153: NULL,
3154: MatZeroRowsColumns_SeqDense,
3155: /* 49*/ MatSetRandom_SeqDense,
3156: NULL,
3157: NULL,
3158: NULL,
3159: NULL,
3160: /* 54*/ NULL,
3161: NULL,
3162: NULL,
3163: NULL,
3164: NULL,
3165: /* 59*/ MatCreateSubMatrix_SeqDense,
3166: MatDestroy_SeqDense,
3167: MatView_SeqDense,
3168: NULL,
3169: NULL,
3170: /* 64*/ NULL,
3171: NULL,
3172: NULL,
3173: NULL,
3174: MatGetRowMaxAbs_SeqDense,
3175: /* 69*/ NULL,
3176: NULL,
3177: NULL,
3178: NULL,
3179: NULL,
3180: /* 74*/ NULL,
3181: NULL,
3182: NULL,
3183: NULL,
3184: MatLoad_SeqDense,
3185: /* 79*/ MatIsSymmetric_SeqDense,
3186: MatIsHermitian_SeqDense,
3187: NULL,
3188: NULL,
3189: NULL,
3190: /* 84*/ NULL,
3191: MatMatMultNumeric_SeqDense_SeqDense,
3192: NULL,
3193: NULL,
3194: MatMatTransposeMultNumeric_SeqDense_SeqDense,
3195: /* 89*/ NULL,
3196: MatProductSetFromOptions_SeqDense,
3197: NULL,
3198: NULL,
3199: MatConjugate_SeqDense,
3200: /* 94*/ NULL,
3201: NULL,
3202: MatRealPart_SeqDense,
3203: MatImaginaryPart_SeqDense,
3204: NULL,
3205: /* 99*/ NULL,
3206: NULL,
3207: NULL,
3208: MatGetRowMin_SeqDense,
3209: MatGetColumnVector_SeqDense,
3210: /*104*/ NULL,
3211: NULL,
3212: NULL,
3213: NULL,
3214: NULL,
3215: /*109*/ NULL,
3216: NULL,
3217: MatMultHermitianTranspose_SeqDense,
3218: MatMultHermitianTransposeAdd_SeqDense,
3219: NULL,
3220: /*114*/ NULL,
3221: MatGetColumnReductions_SeqDense,
3222: NULL,
3223: NULL,
3224: NULL,
3225: /*119*/ NULL,
3226: MatTransposeMatMultNumeric_SeqDense_SeqDense,
3227: NULL,
3228: NULL,
3229: NULL,
3230: /*124*/ NULL,
3231: NULL,
3232: NULL,
3233: NULL,
3234: NULL,
3235: /*129*/ MatCreateMPIMatConcatenateSeqMat_SeqDense,
3236: NULL,
3237: NULL,
3238: NULL,
3239: NULL,
3240: /*134*/ NULL,
3241: NULL,
3242: NULL,
3243: NULL,
3244: NULL,
3245: /*139*/ NULL,
3246: NULL,
3247: NULL,
3248: NULL,
3249: MatADot_Default,
3250: /*144*/ MatANorm_Default,
3251: NULL,
3252: NULL,
3253: NULL};
3255: /*@
3256: MatCreateSeqDense - Creates a `MATSEQDENSE` that
3257: is stored in column major order (the usual Fortran format).
3259: Collective
3261: Input Parameters:
3262: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3263: . m - number of rows
3264: . n - number of columns
3265: - data - optional location of matrix data in column major order. Use `NULL` for PETSc
3266: to control all matrix memory allocation.
3268: Output Parameter:
3269: . A - the matrix
3271: Level: intermediate
3273: Note:
3274: The data input variable is intended primarily for Fortran programmers
3275: who wish to allocate their own matrix memory space. Most users should
3276: set `data` = `NULL`.
3278: Developer Note:
3279: Many of the matrix operations for this variant use the BLAS and LAPACK routines.
3281: .seealso: [](ch_matrices), `Mat`, `MATSEQDENSE`, `MatCreate()`, `MatCreateDense()`, `MatSetValues()`
3282: @*/
3283: PetscErrorCode MatCreateSeqDense(MPI_Comm comm, PetscInt m, PetscInt n, PetscScalar data[], Mat *A)
3284: {
3285: PetscFunctionBegin;
3286: PetscCall(MatCreate(comm, A));
3287: PetscCall(MatSetSizes(*A, m, n, m, n));
3288: PetscCall(MatSetType(*A, MATSEQDENSE));
3289: PetscCall(MatSeqDenseSetPreallocation(*A, data));
3290: PetscFunctionReturn(PETSC_SUCCESS);
3291: }
3293: /*@
3294: MatSeqDenseSetPreallocation - Sets the array used for storing the matrix elements of a `MATSEQDENSE` matrix
3296: Collective
3298: Input Parameters:
3299: + B - the matrix
3300: - data - the array (or `NULL`)
3302: Level: intermediate
3304: Note:
3305: The data input variable is intended primarily for Fortran programmers
3306: who wish to allocate their own matrix memory space. Most users should
3307: need not call this routine.
3309: .seealso: [](ch_matrices), `Mat`, `MATSEQDENSE`, `MatCreate()`, `MatCreateDense()`, `MatSetValues()`, `MatDenseSetLDA()`
3310: @*/
3311: PetscErrorCode MatSeqDenseSetPreallocation(Mat B, PetscScalar data[])
3312: {
3313: PetscFunctionBegin;
3315: PetscTryMethod(B, "MatSeqDenseSetPreallocation_C", (Mat, PetscScalar[]), (B, data));
3316: PetscFunctionReturn(PETSC_SUCCESS);
3317: }
3319: PetscErrorCode MatSeqDenseSetPreallocation_SeqDense(Mat B, PetscScalar *data)
3320: {
3321: Mat_SeqDense *b = (Mat_SeqDense *)B->data;
3323: PetscFunctionBegin;
3324: PetscCheck(!b->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3325: B->preallocated = PETSC_TRUE;
3327: PetscCall(PetscLayoutSetUp(B->rmap));
3328: PetscCall(PetscLayoutSetUp(B->cmap));
3330: if (b->lda <= 0) PetscCall(PetscBLASIntCast(B->rmap->n, &b->lda));
3332: if (!data) { /* petsc-allocated storage */
3333: if (!b->user_alloc) PetscCall(PetscFree(b->v));
3334: PetscCall(PetscCalloc1((size_t)b->lda * B->cmap->n, &b->v));
3336: b->user_alloc = PETSC_FALSE;
3337: } else { /* user-allocated storage */
3338: if (!b->user_alloc) PetscCall(PetscFree(b->v));
3339: b->v = data;
3340: b->user_alloc = PETSC_TRUE;
3341: }
3342: B->assembled = PETSC_TRUE;
3343: PetscFunctionReturn(PETSC_SUCCESS);
3344: }
3346: #if defined(PETSC_HAVE_ELEMENTAL)
3347: PETSC_INTERN PetscErrorCode MatConvert_SeqDense_Elemental(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
3348: {
3349: Mat mat_elemental;
3350: const PetscScalar *array;
3351: PetscScalar *v_colwise;
3352: PetscInt M = A->rmap->N, N = A->cmap->N, i, j, k, *rows, *cols;
3354: PetscFunctionBegin;
3355: PetscCall(PetscMalloc3(M * N, &v_colwise, M, &rows, N, &cols));
3356: PetscCall(MatDenseGetArrayRead(A, &array));
3357: /* convert column-wise array into row-wise v_colwise, see MatSetValues_Elemental() */
3358: k = 0;
3359: for (j = 0; j < N; j++) {
3360: cols[j] = j;
3361: for (i = 0; i < M; i++) v_colwise[j * M + i] = array[k++];
3362: }
3363: for (i = 0; i < M; i++) rows[i] = i;
3364: PetscCall(MatDenseRestoreArrayRead(A, &array));
3366: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental));
3367: PetscCall(MatSetSizes(mat_elemental, PETSC_DECIDE, PETSC_DECIDE, M, N));
3368: PetscCall(MatSetType(mat_elemental, MATELEMENTAL));
3369: PetscCall(MatSetUp(mat_elemental));
3371: /* PETSc-Elemental interaface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
3372: PetscCall(MatSetValues(mat_elemental, M, rows, N, cols, v_colwise, ADD_VALUES));
3373: PetscCall(MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY));
3374: PetscCall(MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY));
3375: PetscCall(PetscFree3(v_colwise, rows, cols));
3377: if (reuse == MAT_INPLACE_MATRIX) {
3378: PetscCall(MatHeaderReplace(A, &mat_elemental));
3379: } else {
3380: *newmat = mat_elemental;
3381: }
3382: PetscFunctionReturn(PETSC_SUCCESS);
3383: }
3384: #endif
3386: PetscErrorCode MatDenseSetLDA_SeqDense(Mat B, PetscInt lda)
3387: {
3388: Mat_SeqDense *b = (Mat_SeqDense *)B->data;
3389: PetscBool data;
3391: PetscFunctionBegin;
3392: data = (B->rmap->n > 0 && B->cmap->n > 0) ? (b->v ? PETSC_TRUE : PETSC_FALSE) : PETSC_FALSE;
3393: PetscCheck(b->user_alloc || !data || b->lda == lda, PETSC_COMM_SELF, PETSC_ERR_ORDER, "LDA cannot be changed after allocation of internal storage");
3394: 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);
3395: PetscCall(PetscBLASIntCast(lda, &b->lda));
3396: PetscFunctionReturn(PETSC_SUCCESS);
3397: }
3399: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqDense(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
3400: {
3401: PetscFunctionBegin;
3402: PetscCall(MatCreateMPIMatConcatenateSeqMat_MPIDense(comm, inmat, n, scall, outmat));
3403: PetscFunctionReturn(PETSC_SUCCESS);
3404: }
3406: PetscErrorCode MatDenseCreateColumnVec_Private(Mat A, Vec *v)
3407: {
3408: PetscBool isstd, iskok, iscuda, iship;
3409: PetscMPIInt size;
3410: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
3411: /* we pass the data of A, to prevent allocating needless GPU memory the first time VecCUPMPlaceArray is called. */
3412: const PetscScalar *a;
3413: #endif
3415: PetscFunctionBegin;
3416: *v = NULL;
3417: PetscCall(PetscStrcmpAny(A->defaultvectype, &isstd, VECSTANDARD, VECSEQ, VECMPI, ""));
3418: PetscCall(PetscStrcmpAny(A->defaultvectype, &iskok, VECKOKKOS, VECSEQKOKKOS, VECMPIKOKKOS, ""));
3419: PetscCall(PetscStrcmpAny(A->defaultvectype, &iscuda, VECCUDA, VECSEQCUDA, VECMPICUDA, ""));
3420: PetscCall(PetscStrcmpAny(A->defaultvectype, &iship, VECHIP, VECSEQHIP, VECMPIHIP, ""));
3421: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
3422: if (isstd) {
3423: if (size > 1) PetscCall(VecCreateMPIWithArray(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, A->rmap->N, NULL, v));
3424: else PetscCall(VecCreateSeqWithArray(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, NULL, v));
3425: } else if (iskok) {
3426: PetscCheck(PetscDefined(HAVE_KOKKOS_KERNELS), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Reconfigure using KOKKOS kernels support");
3427: #if PetscDefined(HAVE_KOKKOS_KERNELS)
3428: if (size > 1) PetscCall(VecCreateMPIKokkosWithArray(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, A->rmap->N, NULL, v));
3429: else PetscCall(VecCreateSeqKokkosWithArray(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, NULL, v));
3430: #endif
3431: } else if (iscuda) {
3432: PetscCheck(PetscDefined(HAVE_CUDA), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Reconfigure using CUDA support");
3433: #if PetscDefined(HAVE_CUDA)
3434: PetscCall(MatDenseCUDAGetArrayRead(A, &a));
3435: if (size > 1) PetscCall(VecCreateMPICUDAWithArrays(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, A->rmap->N, NULL, a, v));
3436: else PetscCall(VecCreateSeqCUDAWithArrays(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, NULL, a, v));
3437: #endif
3438: } else if (iship) {
3439: PetscCheck(PetscDefined(HAVE_HIP), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Reconfigure using HIP support");
3440: #if PetscDefined(HAVE_HIP)
3441: PetscCall(MatDenseHIPGetArrayRead(A, &a));
3442: if (size > 1) PetscCall(VecCreateMPIHIPWithArrays(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, A->rmap->N, NULL, a, v));
3443: else PetscCall(VecCreateSeqHIPWithArrays(PetscObjectComm((PetscObject)A), A->rmap->bs, A->rmap->n, NULL, a, v));
3444: #endif
3445: }
3446: PetscCheck(*v, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not coded for type %s", A->defaultvectype);
3447: PetscFunctionReturn(PETSC_SUCCESS);
3448: }
3450: PetscErrorCode MatDenseGetColumnVec_SeqDense(Mat A, PetscInt col, Vec *v)
3451: {
3452: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3454: PetscFunctionBegin;
3455: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
3456: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3457: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
3458: a->vecinuse = col + 1;
3459: PetscCall(MatDenseGetArray(A, (PetscScalar **)&a->ptrinuse));
3460: PetscCall(VecPlaceArray(a->cvec, a->ptrinuse + (size_t)col * (size_t)a->lda));
3461: *v = a->cvec;
3462: PetscFunctionReturn(PETSC_SUCCESS);
3463: }
3465: PetscErrorCode MatDenseRestoreColumnVec_SeqDense(Mat A, PetscInt col, Vec *v)
3466: {
3467: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3469: PetscFunctionBegin;
3470: PetscCheck(a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
3471: PetscCheck(a->cvec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column vector");
3472: VecCheckAssembled(a->cvec);
3473: a->vecinuse = 0;
3474: PetscCall(MatDenseRestoreArray(A, (PetscScalar **)&a->ptrinuse));
3475: PetscCall(VecResetArray(a->cvec));
3476: if (v) *v = NULL;
3477: PetscFunctionReturn(PETSC_SUCCESS);
3478: }
3480: PetscErrorCode MatDenseGetColumnVecRead_SeqDense(Mat A, PetscInt col, Vec *v)
3481: {
3482: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3484: PetscFunctionBegin;
3485: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
3486: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3487: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
3488: a->vecinuse = col + 1;
3489: PetscCall(MatDenseGetArrayRead(A, &a->ptrinuse));
3490: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)a->lda)));
3491: PetscCall(VecLockReadPush(a->cvec));
3492: *v = a->cvec;
3493: PetscFunctionReturn(PETSC_SUCCESS);
3494: }
3496: PetscErrorCode MatDenseRestoreColumnVecRead_SeqDense(Mat A, PetscInt col, Vec *v)
3497: {
3498: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3500: PetscFunctionBegin;
3501: PetscCheck(a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
3502: PetscCheck(a->cvec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column vector");
3503: VecCheckAssembled(a->cvec);
3504: a->vecinuse = 0;
3505: PetscCall(MatDenseRestoreArrayRead(A, &a->ptrinuse));
3506: PetscCall(VecLockReadPop(a->cvec));
3507: PetscCall(VecResetArray(a->cvec));
3508: if (v) *v = NULL;
3509: PetscFunctionReturn(PETSC_SUCCESS);
3510: }
3512: PetscErrorCode MatDenseGetColumnVecWrite_SeqDense(Mat A, PetscInt col, Vec *v)
3513: {
3514: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3516: PetscFunctionBegin;
3517: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
3518: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3519: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
3520: a->vecinuse = col + 1;
3521: PetscCall(MatDenseGetArrayWrite(A, (PetscScalar **)&a->ptrinuse));
3522: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)a->lda)));
3523: *v = a->cvec;
3524: PetscFunctionReturn(PETSC_SUCCESS);
3525: }
3527: PetscErrorCode MatDenseRestoreColumnVecWrite_SeqDense(Mat A, PetscInt col, Vec *v)
3528: {
3529: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3531: PetscFunctionBegin;
3532: PetscCheck(a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
3533: PetscCheck(a->cvec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column vector");
3534: VecCheckAssembled(a->cvec);
3535: a->vecinuse = 0;
3536: PetscCall(MatDenseRestoreArrayWrite(A, (PetscScalar **)&a->ptrinuse));
3537: PetscCall(VecResetArray(a->cvec));
3538: if (v) *v = NULL;
3539: PetscFunctionReturn(PETSC_SUCCESS);
3540: }
3542: PetscErrorCode MatDenseGetSubMatrix_SeqDense(Mat A, PetscInt rbegin, PetscInt rend, PetscInt cbegin, PetscInt cend, Mat *v)
3543: {
3544: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3546: PetscFunctionBegin;
3547: PetscCheck(!a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
3548: PetscCheck(!a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
3549: if (a->cmat && (cend - cbegin != a->cmat->cmap->N || rend - rbegin != a->cmat->rmap->N)) PetscCall(MatDestroy(&a->cmat));
3550: if (!a->cmat) {
3551: PetscCall(MatCreateDense(PetscObjectComm((PetscObject)A), rend - rbegin, PETSC_DECIDE, rend - rbegin, cend - cbegin, PetscSafePointerPlusOffset(a->v, rbegin + (size_t)cbegin * a->lda), &a->cmat));
3552: } else {
3553: PetscCall(MatDensePlaceArray(a->cmat, PetscSafePointerPlusOffset(a->v, rbegin + (size_t)cbegin * a->lda)));
3554: }
3555: PetscCall(MatDenseSetLDA(a->cmat, a->lda));
3556: a->matinuse = cbegin + 1;
3557: *v = a->cmat;
3558: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
3559: A->offloadmask = PETSC_OFFLOAD_CPU;
3560: #endif
3561: PetscFunctionReturn(PETSC_SUCCESS);
3562: }
3564: PetscErrorCode MatDenseRestoreSubMatrix_SeqDense(Mat A, Mat *v)
3565: {
3566: Mat_SeqDense *a = (Mat_SeqDense *)A->data;
3568: PetscFunctionBegin;
3569: PetscCheck(a->matinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetSubMatrix() first");
3570: PetscCheck(a->cmat, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column matrix");
3571: PetscCheck(*v == a->cmat, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Not the matrix obtained from MatDenseGetSubMatrix()");
3572: a->matinuse = 0;
3573: PetscCall(MatDenseResetArray(a->cmat));
3574: *v = NULL;
3575: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
3576: A->offloadmask = PETSC_OFFLOAD_CPU;
3577: #endif
3578: PetscFunctionReturn(PETSC_SUCCESS);
3579: }
3581: /*MC
3582: MATSEQDENSE - MATSEQDENSE = "seqdense" - A matrix type to be used for sequential dense matrices.
3584: Options Database Key:
3585: . -mat_type seqdense - sets the matrix type to `MATSEQDENSE` during a call to `MatSetFromOptions()`
3587: Level: beginner
3589: .seealso: [](ch_matrices), `Mat`, `MATSEQDENSE`, `MatCreateSeqDense()`
3590: M*/
3591: PetscErrorCode MatCreate_SeqDense(Mat B)
3592: {
3593: Mat_SeqDense *b;
3594: PetscMPIInt size;
3596: PetscFunctionBegin;
3597: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
3598: PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Comm must be of size 1");
3600: PetscCall(PetscNew(&b));
3601: B->data = (void *)b;
3602: B->ops[0] = MatOps_Values;
3604: b->roworiented = PETSC_TRUE;
3606: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatQRFactor_C", MatQRFactor_SeqDense));
3607: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetLDA_C", MatDenseGetLDA_SeqDense));
3608: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseSetLDA_C", MatDenseSetLDA_SeqDense));
3609: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetArray_C", MatDenseGetArray_SeqDense));
3610: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreArray_C", MatDenseRestoreArray_SeqDense));
3611: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDensePlaceArray_C", MatDensePlaceArray_SeqDense));
3612: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseResetArray_C", MatDenseResetArray_SeqDense));
3613: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseReplaceArray_C", MatDenseReplaceArray_SeqDense));
3614: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetArrayRead_C", MatDenseGetArray_SeqDense));
3615: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreArrayRead_C", MatDenseRestoreArray_SeqDense));
3616: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetArrayWrite_C", MatDenseGetArray_SeqDense));
3617: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreArrayWrite_C", MatDenseRestoreArray_SeqDense));
3618: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_seqaij_C", MatConvert_SeqDense_SeqAIJ));
3619: #if defined(PETSC_HAVE_ELEMENTAL)
3620: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_elemental_C", MatConvert_SeqDense_Elemental));
3621: #endif
3622: #if defined(PETSC_HAVE_SCALAPACK) && (defined(PETSC_USE_REAL_SINGLE) || defined(PETSC_USE_REAL_DOUBLE))
3623: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_scalapack_C", MatConvert_Dense_ScaLAPACK));
3624: #endif
3625: #if defined(PETSC_HAVE_CUDA)
3626: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_seqdensecuda_C", MatConvert_SeqDense_SeqDenseCUDA));
3627: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdensecuda_seqdensecuda_C", MatProductSetFromOptions_SeqDense));
3628: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdensecuda_seqdense_C", MatProductSetFromOptions_SeqDense));
3629: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqdensecuda_C", MatProductSetFromOptions_SeqDense));
3630: #endif
3631: #if defined(PETSC_HAVE_HIP)
3632: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqdense_seqdensehip_C", MatConvert_SeqDense_SeqDenseHIP));
3633: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdensehip_seqdensehip_C", MatProductSetFromOptions_SeqDense));
3634: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdensehip_seqdense_C", MatProductSetFromOptions_SeqDense));
3635: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqdensehip_C", MatProductSetFromOptions_SeqDense));
3636: #endif
3637: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqDenseSetPreallocation_C", MatSeqDenseSetPreallocation_SeqDense));
3638: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqdense_C", MatProductSetFromOptions_SeqAIJ_SeqDense));
3639: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqdense_C", MatProductSetFromOptions_SeqDense));
3640: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqbaij_seqdense_C", MatProductSetFromOptions_SeqXBAIJ_SeqDense));
3641: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqsbaij_seqdense_C", MatProductSetFromOptions_SeqXBAIJ_SeqDense));
3643: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetColumn_C", MatDenseGetColumn_SeqDense));
3644: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreColumn_C", MatDenseRestoreColumn_SeqDense));
3645: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetColumnVec_C", MatDenseGetColumnVec_SeqDense));
3646: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreColumnVec_C", MatDenseRestoreColumnVec_SeqDense));
3647: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetColumnVecRead_C", MatDenseGetColumnVecRead_SeqDense));
3648: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreColumnVecRead_C", MatDenseRestoreColumnVecRead_SeqDense));
3649: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetColumnVecWrite_C", MatDenseGetColumnVecWrite_SeqDense));
3650: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreColumnVecWrite_C", MatDenseRestoreColumnVecWrite_SeqDense));
3651: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseGetSubMatrix_C", MatDenseGetSubMatrix_SeqDense));
3652: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDenseRestoreSubMatrix_C", MatDenseRestoreSubMatrix_SeqDense));
3653: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMultColumnRange_C", MatMultColumnRange_SeqDense));
3654: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMultAddColumnRange_C", MatMultAddColumnRange_SeqDense));
3655: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMultHermitianTransposeColumnRange_C", MatMultHermitianTransposeColumnRange_SeqDense));
3656: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMultHermitianTransposeAddColumnRange_C", MatMultHermitianTransposeAddColumnRange_SeqDense));
3657: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQDENSE));
3658: PetscFunctionReturn(PETSC_SUCCESS);
3659: }
3661: /*@C
3662: 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.
3664: Not Collective
3666: Input Parameters:
3667: + A - a `MATSEQDENSE` or `MATMPIDENSE` matrix
3668: - col - column index
3670: Output Parameter:
3671: . vals - pointer to the data
3673: Level: intermediate
3675: Note:
3676: Use `MatDenseGetColumnVec()` to get access to a column of a `MATDENSE` treated as a `Vec`
3678: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseRestoreColumn()`, `MatDenseGetColumnVec()`
3679: @*/
3680: PetscErrorCode MatDenseGetColumn(Mat A, PetscInt col, PetscScalar *vals[])
3681: {
3682: PetscFunctionBegin;
3685: PetscAssertPointer(vals, 3);
3686: PetscUseMethod(A, "MatDenseGetColumn_C", (Mat, PetscInt, PetscScalar **), (A, col, vals));
3687: PetscFunctionReturn(PETSC_SUCCESS);
3688: }
3690: /*@C
3691: MatDenseRestoreColumn - returns access to a column of a `MATDENSE` matrix which is returned by `MatDenseGetColumn()`.
3693: Not Collective
3695: Input Parameters:
3696: + A - a `MATSEQDENSE` or `MATMPIDENSE` matrix
3697: - vals - pointer to the data (may be `NULL`)
3699: Level: intermediate
3701: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetColumn()`
3702: @*/
3703: PetscErrorCode MatDenseRestoreColumn(Mat A, PetscScalar *vals[])
3704: {
3705: PetscFunctionBegin;
3707: PetscAssertPointer(vals, 2);
3708: PetscUseMethod(A, "MatDenseRestoreColumn_C", (Mat, PetscScalar **), (A, vals));
3709: PetscFunctionReturn(PETSC_SUCCESS);
3710: }
3712: /*@
3713: MatDenseGetColumnVec - Gives read-write access to a column of a `MATDENSE` matrix, represented as a `Vec`.
3715: Collective
3717: Input Parameters:
3718: + A - the `Mat` object
3719: - col - the column index
3721: Output Parameter:
3722: . v - the vector
3724: Level: intermediate
3726: Notes:
3727: The vector is owned by PETSc. Users need to call `MatDenseRestoreColumnVec()` when the vector is no longer needed.
3729: Use `MatDenseGetColumnVecRead()` to obtain read-only access or `MatDenseGetColumnVecWrite()` for write-only access.
3731: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVecRead()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecRead()`, `MatDenseRestoreColumnVecWrite()`, `MatDenseGetColumn()`
3732: @*/
3733: PetscErrorCode MatDenseGetColumnVec(Mat A, PetscInt col, Vec *v)
3734: {
3735: PetscFunctionBegin;
3739: PetscAssertPointer(v, 3);
3740: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3741: 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);
3742: PetscUseMethod(A, "MatDenseGetColumnVec_C", (Mat, PetscInt, Vec *), (A, col, v));
3743: PetscFunctionReturn(PETSC_SUCCESS);
3744: }
3746: /*@
3747: MatDenseRestoreColumnVec - Returns access to a column of a dense matrix obtained from `MatDenseGetColumnVec()`.
3749: Collective
3751: Input Parameters:
3752: + A - the `Mat` object
3753: . col - the column index
3754: - v - the `Vec` object (may be `NULL`)
3756: Level: intermediate
3758: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecRead()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVecRead()`, `MatDenseRestoreColumnVecWrite()`
3759: @*/
3760: PetscErrorCode MatDenseRestoreColumnVec(Mat A, PetscInt col, Vec *v)
3761: {
3762: PetscFunctionBegin;
3767: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3768: 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);
3769: PetscUseMethod(A, "MatDenseRestoreColumnVec_C", (Mat, PetscInt, Vec *), (A, col, v));
3770: PetscFunctionReturn(PETSC_SUCCESS);
3771: }
3773: /*@
3774: MatDenseGetColumnVecRead - Gives read-only access to a column of a dense matrix, represented as a `Vec`.
3776: Collective
3778: Input Parameters:
3779: + A - the `Mat` object
3780: - col - the column index
3782: Output Parameter:
3783: . v - the vector
3785: Level: intermediate
3787: Notes:
3788: The vector is owned by PETSc and users cannot modify it.
3790: Users need to call `MatDenseRestoreColumnVecRead()` when the vector is no longer needed.
3792: Use `MatDenseGetColumnVec()` to obtain read-write access or `MatDenseGetColumnVecWrite()` for write-only access.
3794: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecRead()`, `MatDenseRestoreColumnVecWrite()`
3795: @*/
3796: PetscErrorCode MatDenseGetColumnVecRead(Mat A, PetscInt col, Vec *v)
3797: {
3798: PetscFunctionBegin;
3802: PetscAssertPointer(v, 3);
3803: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3804: 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);
3805: PetscUseMethod(A, "MatDenseGetColumnVecRead_C", (Mat, PetscInt, Vec *), (A, col, v));
3806: PetscFunctionReturn(PETSC_SUCCESS);
3807: }
3809: /*@
3810: MatDenseRestoreColumnVecRead - Returns access to a column of a dense matrix obtained from `MatDenseGetColumnVecRead()`.
3812: Collective
3814: Input Parameters:
3815: + A - the `Mat` object
3816: . col - the column index
3817: - v - the `Vec` object (may be `NULL`)
3819: Level: intermediate
3821: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecRead()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecWrite()`
3822: @*/
3823: PetscErrorCode MatDenseRestoreColumnVecRead(Mat A, PetscInt col, Vec *v)
3824: {
3825: PetscFunctionBegin;
3830: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3831: 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);
3832: PetscUseMethod(A, "MatDenseRestoreColumnVecRead_C", (Mat, PetscInt, Vec *), (A, col, v));
3833: PetscFunctionReturn(PETSC_SUCCESS);
3834: }
3836: /*@
3837: MatDenseGetColumnVecWrite - Gives write-only access to a column of a dense matrix, represented as a `Vec`.
3839: Collective
3841: Input Parameters:
3842: + A - the `Mat` object
3843: - col - the column index
3845: Output Parameter:
3846: . v - the vector
3848: Level: intermediate
3850: Notes:
3851: The vector is owned by PETSc. Users need to call `MatDenseRestoreColumnVecWrite()` when the vector is no longer needed.
3853: Use `MatDenseGetColumnVec()` to obtain read-write access or `MatDenseGetColumnVecRead()` for read-only access.
3855: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecRead()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecRead()`, `MatDenseRestoreColumnVecWrite()`
3856: @*/
3857: PetscErrorCode MatDenseGetColumnVecWrite(Mat A, PetscInt col, Vec *v)
3858: {
3859: PetscFunctionBegin;
3863: PetscAssertPointer(v, 3);
3864: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3865: 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);
3866: PetscUseMethod(A, "MatDenseGetColumnVecWrite_C", (Mat, PetscInt, Vec *), (A, col, v));
3867: PetscFunctionReturn(PETSC_SUCCESS);
3868: }
3870: /*@
3871: MatDenseRestoreColumnVecWrite - Returns access to a column of a dense matrix obtained from `MatDenseGetColumnVecWrite()`.
3873: Collective
3875: Input Parameters:
3876: + A - the `Mat` object
3877: . col - the column index
3878: - v - the `Vec` object (may be `NULL`)
3880: Level: intermediate
3882: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseGetColumnVecRead()`, `MatDenseGetColumnVecWrite()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreColumnVecRead()`
3883: @*/
3884: PetscErrorCode MatDenseRestoreColumnVecWrite(Mat A, PetscInt col, Vec *v)
3885: {
3886: PetscFunctionBegin;
3891: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3892: 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);
3893: PetscUseMethod(A, "MatDenseRestoreColumnVecWrite_C", (Mat, PetscInt, Vec *), (A, col, v));
3894: PetscFunctionReturn(PETSC_SUCCESS);
3895: }
3897: /*@
3898: MatDenseGetSubMatrix - Gives access to a block of rows and columns of a dense matrix, represented as a `Mat`.
3900: Collective
3902: Input Parameters:
3903: + A - the `Mat` object
3904: . rbegin - the first global row index in the block (if `PETSC_DECIDE`, is 0)
3905: . rend - the global row index past the last one in the block (if `PETSC_DECIDE`, is `M`)
3906: . cbegin - the first global column index in the block (if `PETSC_DECIDE`, is 0)
3907: - cend - the global column index past the last one in the block (if `PETSC_DECIDE`, is `N`)
3909: Output Parameter:
3910: . v - the matrix
3912: Level: intermediate
3914: Notes:
3915: The matrix is owned by PETSc. Users need to call `MatDenseRestoreSubMatrix()` when the matrix is no longer needed.
3917: The output matrix is not redistributed by PETSc, so depending on the values of `rbegin` and `rend`, some processes may have no local rows.
3919: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseRestoreColumnVec()`, `MatDenseRestoreSubMatrix()`
3920: @*/
3921: PetscErrorCode MatDenseGetSubMatrix(Mat A, PetscInt rbegin, PetscInt rend, PetscInt cbegin, PetscInt cend, Mat *v)
3922: {
3923: PetscFunctionBegin;
3930: PetscAssertPointer(v, 6);
3931: if (rbegin == PETSC_DECIDE) rbegin = 0;
3932: if (rend == PETSC_DECIDE) rend = A->rmap->N;
3933: if (cbegin == PETSC_DECIDE) cbegin = 0;
3934: if (cend == PETSC_DECIDE) cend = A->cmap->N;
3935: PetscCheck(A->preallocated, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Matrix not preallocated");
3936: 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);
3937: 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);
3938: 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);
3939: 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);
3940: PetscUseMethod(A, "MatDenseGetSubMatrix_C", (Mat, PetscInt, PetscInt, PetscInt, PetscInt, Mat *), (A, rbegin, rend, cbegin, cend, v));
3941: PetscFunctionReturn(PETSC_SUCCESS);
3942: }
3944: /*@
3945: MatDenseRestoreSubMatrix - Returns access to a block of columns of a dense matrix obtained from `MatDenseGetSubMatrix()`.
3947: Collective
3949: Input Parameters:
3950: + A - the `Mat` object
3951: - v - the `Mat` object (cannot be `NULL`)
3953: Level: intermediate
3955: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATDENSECUDA`, `MATDENSEHIP`, `MatDenseGetColumnVec()`, `MatDenseRestoreColumnVec()`, `MatDenseGetSubMatrix()`
3956: @*/
3957: PetscErrorCode MatDenseRestoreSubMatrix(Mat A, Mat *v)
3958: {
3959: PetscFunctionBegin;
3962: PetscAssertPointer(v, 2);
3964: PetscUseMethod(A, "MatDenseRestoreSubMatrix_C", (Mat, Mat *), (A, v));
3965: PetscFunctionReturn(PETSC_SUCCESS);
3966: }
3968: #include <petscblaslapack.h>
3969: #include <petsc/private/kernels/blockinvert.h>
3971: PetscErrorCode MatSeqDenseInvert(Mat A)
3972: {
3973: PetscInt m;
3974: const PetscReal shift = 0.0;
3975: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
3976: PetscScalar *values;
3978: PetscFunctionBegin;
3980: PetscCall(MatDenseGetArray(A, &values));
3981: PetscCall(MatGetLocalSize(A, &m, NULL));
3982: allowzeropivot = PetscNot(A->erroriffailure);
3983: /* factor and invert each block */
3984: switch (m) {
3985: case 1:
3986: values[0] = (PetscScalar)1.0 / (values[0] + shift);
3987: break;
3988: case 2:
3989: PetscCall(PetscKernel_A_gets_inverse_A_2(values, shift, allowzeropivot, &zeropivotdetected));
3990: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3991: break;
3992: case 3:
3993: PetscCall(PetscKernel_A_gets_inverse_A_3(values, shift, allowzeropivot, &zeropivotdetected));
3994: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3995: break;
3996: case 4:
3997: PetscCall(PetscKernel_A_gets_inverse_A_4(values, shift, allowzeropivot, &zeropivotdetected));
3998: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3999: break;
4000: case 5: {
4001: PetscScalar work[25];
4002: PetscInt ipvt[5];
4004: PetscCall(PetscKernel_A_gets_inverse_A_5(values, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
4005: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4006: } break;
4007: case 6:
4008: PetscCall(PetscKernel_A_gets_inverse_A_6(values, shift, allowzeropivot, &zeropivotdetected));
4009: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4010: break;
4011: case 7:
4012: PetscCall(PetscKernel_A_gets_inverse_A_7(values, shift, allowzeropivot, &zeropivotdetected));
4013: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4014: break;
4015: default: {
4016: PetscInt *v_pivots, *IJ, j;
4017: PetscScalar *v_work;
4019: PetscCall(PetscMalloc3(m, &v_work, m, &v_pivots, m, &IJ));
4020: for (j = 0; j < m; j++) IJ[j] = j;
4021: PetscCall(PetscKernel_A_gets_inverse_A(m, values, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
4022: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
4023: PetscCall(PetscFree3(v_work, v_pivots, IJ));
4024: }
4025: }
4026: PetscCall(MatDenseRestoreArray(A, &values));
4027: PetscFunctionReturn(PETSC_SUCCESS);
4028: }
4030: /*@
4031: MatDenseReplaceArrayWithMemType - Allows one to replace the array in a `MATDENSE`, `MATDENSECUDA`, or `MATDENSEHIP`
4032: with an array provided by the user and a matching `PetscMemType`. This is useful to avoid copying an array into a matrix.
4034: Not Collective
4036: Input Parameters:
4037: + mat - the matrix
4038: . mtype - the `PetscMemType` of the array
4039: - array - the array in column major order
4041: Level: developer
4043: Note:
4044: Adding `const` to `array` was an oversight, see notes in `VecPlaceArray()`.
4046: This permanently replaces the GPU array and frees the memory associated with the old GPU
4047: array. The memory passed in CANNOT be freed by the user. It will be freed when the matrix is
4048: destroyed. The array should respect the matrix leading dimension.
4050: .seealso: `MatDenseReplaceArray()`, `MatDenseCUDAReplaceArray()`, `MatDenseHIPReplaceArray()`
4051: @*/
4052: PetscErrorCode MatDenseReplaceArrayWithMemType(Mat mat, PetscMemType mtype, const PetscScalar array[])
4053: {
4054: const char *type = PetscMemTypeToString(mtype) + 14; /* skip "PETSC_MEMTYPE_" */
4055: char buffer[256];
4057: PetscFunctionBegin;
4059: PetscAssertPointer(array, 3);
4060: PetscCall(PetscSNPrintf(buffer, sizeof(buffer), "MatDense%sReplaceArray_C", PetscMemTypeHost(mtype) ? "" : type));
4061: PetscUseMethod(mat, buffer, (Mat, const PetscScalar[]), (mat, array));
4062: PetscFunctionReturn(PETSC_SUCCESS);
4063: }