Actual source code: mpidense.c
1: /*
2: Basic functions for basic parallel dense matrices.
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/mpi/mpidense.h>
8: #include <../src/mat/impls/aij/mpi/mpiaij.h>
9: #include <petscblaslapack.h>
10: #include <petsc/private/vecimpl.h>
11: #include <petsc/private/deviceimpl.h>
12: #include <petsc/private/sfimpl.h>
14: /*@
15: MatDenseGetLocalMatrix - For a `MATMPIDENSE` or `MATSEQDENSE` matrix returns the sequential
16: matrix that represents the operator. For sequential matrices it returns itself.
18: Input Parameter:
19: . A - the sequential or MPI `MATDENSE` matrix
21: Output Parameter:
22: . B - the inner matrix
24: Level: intermediate
26: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATMPIDENSE`, `MATSEQDENSE`
27: @*/
28: PetscErrorCode MatDenseGetLocalMatrix(Mat A, Mat *B)
29: {
30: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
31: PetscBool flg;
33: PetscFunctionBegin;
35: PetscAssertPointer(B, 2);
36: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &flg));
37: if (flg) *B = mat->A;
38: else {
39: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQDENSE, &flg));
40: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for matrix type %s", ((PetscObject)A)->type_name);
41: *B = A;
42: }
43: PetscFunctionReturn(PETSC_SUCCESS);
44: }
46: static PetscErrorCode MatCopy_MPIDense(Mat A, Mat B, MatStructure s)
47: {
48: Mat_MPIDense *Amat = (Mat_MPIDense *)A->data;
49: Mat_MPIDense *Bmat = (Mat_MPIDense *)B->data;
51: PetscFunctionBegin;
52: PetscCall(MatCopy(Amat->A, Bmat->A, s));
53: PetscFunctionReturn(PETSC_SUCCESS);
54: }
56: PetscErrorCode MatShift_MPIDense(Mat A, PetscScalar alpha)
57: {
58: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
59: PetscInt j, lda, rstart = A->rmap->rstart, rend = A->rmap->rend, rend2;
60: PetscScalar *v;
62: PetscFunctionBegin;
63: PetscCall(MatDenseGetArray(mat->A, &v));
64: PetscCall(MatDenseGetLDA(mat->A, &lda));
65: rend2 = PetscMin(rend, A->cmap->N);
66: if (rend2 > rstart) {
67: for (j = rstart; j < rend2; j++) v[j - rstart + j * lda] += alpha;
68: PetscCall(PetscLogFlops(rend2 - rstart));
69: }
70: PetscCall(MatDenseRestoreArray(mat->A, &v));
71: PetscFunctionReturn(PETSC_SUCCESS);
72: }
74: static PetscErrorCode MatGetRow_MPIDense(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
75: {
76: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
77: PetscInt lrow, rstart = A->rmap->rstart, rend = A->rmap->rend;
79: PetscFunctionBegin;
80: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "only local rows");
81: lrow = row - rstart;
82: PetscCall(MatGetRow(mat->A, lrow, nz, (const PetscInt **)idx, (const PetscScalar **)v));
83: PetscFunctionReturn(PETSC_SUCCESS);
84: }
86: static PetscErrorCode MatRestoreRow_MPIDense(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
87: {
88: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
89: PetscInt lrow, rstart = A->rmap->rstart, rend = A->rmap->rend;
91: PetscFunctionBegin;
92: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "only local rows");
93: lrow = row - rstart;
94: PetscCall(MatRestoreRow(mat->A, lrow, nz, (const PetscInt **)idx, (const PetscScalar **)v));
95: PetscFunctionReturn(PETSC_SUCCESS);
96: }
98: static PetscErrorCode MatGetDiagonalBlock_MPIDense(Mat A, Mat *a)
99: {
100: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
101: PetscInt m = A->rmap->n, rstart = A->rmap->rstart;
102: PetscScalar *array;
103: MPI_Comm comm;
104: PetscBool flg;
105: Mat B;
107: PetscFunctionBegin;
108: PetscCall(MatHasCongruentLayouts(A, &flg));
109: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only square matrices supported.");
110: PetscCall(PetscObjectQuery((PetscObject)A, "DiagonalBlock", (PetscObject *)&B));
111: if (!B) { /* This should use MatDenseGetSubMatrix (not create), but we would need a call like MatRestoreDiagonalBlock */
112: #if PetscDefined(HAVE_CUDA)
113: PetscCall(PetscObjectTypeCompare((PetscObject)mdn->A, MATSEQDENSECUDA, &flg));
114: PetscCheck(!flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not coded for %s. Send an email to petsc-dev@mcs.anl.gov to request this feature", MATSEQDENSECUDA);
115: #elif PetscDefined(HAVE_HIP)
116: PetscCall(PetscObjectTypeCompare((PetscObject)mdn->A, MATSEQDENSEHIP, &flg));
117: PetscCheck(!flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not coded for %s. Send an email to petsc-dev@mcs.anl.gov to request this feature", MATSEQDENSEHIP);
118: #endif
119: PetscCall(PetscObjectGetComm((PetscObject)mdn->A, &comm));
120: PetscCall(MatCreate(comm, &B));
121: PetscCall(MatSetSizes(B, m, m, m, m));
122: PetscCall(MatSetType(B, ((PetscObject)mdn->A)->type_name));
123: PetscCall(MatDenseGetArrayRead(mdn->A, (const PetscScalar **)&array));
124: PetscCall(MatSeqDenseSetPreallocation(B, array + m * rstart));
125: PetscCall(MatDenseRestoreArrayRead(mdn->A, (const PetscScalar **)&array));
126: PetscCall(PetscObjectCompose((PetscObject)A, "DiagonalBlock", (PetscObject)B));
127: *a = B;
128: PetscCall(MatDestroy(&B));
129: } else *a = B;
130: PetscFunctionReturn(PETSC_SUCCESS);
131: }
133: static PetscErrorCode MatSetValues_MPIDense(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], const PetscScalar v[], InsertMode addv)
134: {
135: Mat_MPIDense *A = (Mat_MPIDense *)mat->data;
136: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend, row;
137: PetscBool roworiented = A->roworiented;
139: PetscFunctionBegin;
140: for (i = 0; i < m; i++) {
141: if (idxm[i] < 0) continue;
142: PetscCheck(idxm[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large");
143: if (idxm[i] >= rstart && idxm[i] < rend) {
144: row = idxm[i] - rstart;
145: if (roworiented) {
146: PetscCall(MatSetValues(A->A, 1, &row, n, idxn, PetscSafePointerPlusOffset(v, i * n), addv));
147: } else {
148: for (j = 0; j < n; j++) {
149: if (idxn[j] < 0) continue;
150: PetscCheck(idxn[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large");
151: PetscCall(MatSetValues(A->A, 1, &row, 1, &idxn[j], PetscSafePointerPlusOffset(v, i + j * m), addv));
152: }
153: }
154: } else if (!A->donotstash) {
155: mat->assembled = PETSC_FALSE;
156: if (roworiented) {
157: PetscCall(MatStashValuesRow_Private(&mat->stash, idxm[i], n, idxn, PetscSafePointerPlusOffset(v, i * n), PETSC_FALSE));
158: } else {
159: PetscCall(MatStashValuesCol_Private(&mat->stash, idxm[i], n, idxn, PetscSafePointerPlusOffset(v, i), m, PETSC_FALSE));
160: }
161: }
162: }
163: PetscFunctionReturn(PETSC_SUCCESS);
164: }
166: static PetscErrorCode MatGetValues_MPIDense(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
167: {
168: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
169: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend, row;
170: PetscBool roworiented = mdn->roworiented;
171: PetscScalar *value;
173: PetscFunctionBegin;
174: for (i = 0; i < m; i++) {
175: if (idxm[i] < 0) continue; /* negative row */
176: PetscCheck(idxm[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large");
177: PetscCheck(idxm[i] >= rstart && idxm[i] < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only local values currently supported");
178: row = idxm[i] - rstart;
179: for (j = 0; j < n; j++) {
180: if (idxn[j] < 0) continue; /* negative column */
181: PetscCheck(idxn[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large");
182: value = roworiented ? &v[j + i * n] : &v[i + j * m];
183: PetscCall(MatGetValues(mdn->A, 1, &row, 1, &idxn[j], value));
184: }
185: }
186: PetscFunctionReturn(PETSC_SUCCESS);
187: }
189: static PetscErrorCode MatDenseGetLDA_MPIDense(Mat A, PetscInt *lda)
190: {
191: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
193: PetscFunctionBegin;
194: PetscCall(MatDenseGetLDA(a->A, lda));
195: PetscFunctionReturn(PETSC_SUCCESS);
196: }
198: static PetscErrorCode MatDenseSetLDA_MPIDense(Mat A, PetscInt lda)
199: {
200: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
201: MatType mtype = MATSEQDENSE;
203: PetscFunctionBegin;
204: if (!a->A) {
205: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
206: PetscCall(PetscLayoutSetUp(A->rmap));
207: PetscCall(PetscLayoutSetUp(A->cmap));
208: PetscCall(MatCreate(PETSC_COMM_SELF, &a->A));
209: PetscCall(MatSetSizes(a->A, A->rmap->n, A->cmap->N, A->rmap->n, A->cmap->N));
210: #if PetscDefined(HAVE_CUDA)
211: PetscBool iscuda;
212: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIDENSECUDA, &iscuda));
213: if (iscuda) mtype = MATSEQDENSECUDA;
214: #elif PetscDefined(HAVE_HIP)
215: PetscBool iship;
216: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIDENSEHIP, &iship));
217: if (iship) mtype = MATSEQDENSEHIP;
218: #endif
219: PetscCall(MatSetType(a->A, mtype));
220: }
221: PetscCall(MatDenseSetLDA(a->A, lda));
222: PetscFunctionReturn(PETSC_SUCCESS);
223: }
225: static PetscErrorCode MatDenseGetArray_MPIDense(Mat A, PetscScalar **array)
226: {
227: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
229: PetscFunctionBegin;
230: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
231: PetscCall(MatDenseGetArray(a->A, array));
232: PetscFunctionReturn(PETSC_SUCCESS);
233: }
235: static PetscErrorCode MatDenseGetArrayRead_MPIDense(Mat A, PetscScalar **array)
236: {
237: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
239: PetscFunctionBegin;
240: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
241: PetscCall(MatDenseGetArrayRead(a->A, (const PetscScalar **)array));
242: PetscFunctionReturn(PETSC_SUCCESS);
243: }
245: static PetscErrorCode MatDenseGetArrayWrite_MPIDense(Mat A, PetscScalar **array)
246: {
247: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
249: PetscFunctionBegin;
250: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
251: PetscCall(MatDenseGetArrayWrite(a->A, array));
252: PetscFunctionReturn(PETSC_SUCCESS);
253: }
255: static PetscErrorCode MatDensePlaceArray_MPIDense(Mat A, const PetscScalar *array)
256: {
257: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
259: PetscFunctionBegin;
260: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
261: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
262: PetscCall(MatDensePlaceArray(a->A, array));
263: PetscFunctionReturn(PETSC_SUCCESS);
264: }
266: static PetscErrorCode MatDenseResetArray_MPIDense(Mat A)
267: {
268: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
270: PetscFunctionBegin;
271: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
272: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
273: PetscCall(MatDenseResetArray(a->A));
274: PetscFunctionReturn(PETSC_SUCCESS);
275: }
277: static PetscErrorCode MatDenseReplaceArray_MPIDense(Mat A, const PetscScalar *array)
278: {
279: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
281: PetscFunctionBegin;
282: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
283: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
284: PetscCall(MatDenseReplaceArray(a->A, array));
285: PetscFunctionReturn(PETSC_SUCCESS);
286: }
288: static PetscErrorCode MatCreateSubMatrix_MPIDense(Mat A, IS isrow, IS iscol, MatReuse scall, Mat *B)
289: {
290: Mat_MPIDense *mat = (Mat_MPIDense *)A->data, *newmatd;
291: PetscInt lda, i, j, rstart, rend, nrows, ncols, Ncols, nlrows, nlcols;
292: const PetscInt *irow, *icol;
293: const PetscScalar *v;
294: PetscScalar *bv;
295: Mat newmat;
296: IS iscol_local;
297: MPI_Comm comm_is, comm_mat;
299: PetscFunctionBegin;
300: PetscCall(PetscObjectGetComm((PetscObject)A, &comm_mat));
301: PetscCall(PetscObjectGetComm((PetscObject)iscol, &comm_is));
302: PetscCheck(comm_mat == comm_is, PETSC_COMM_SELF, PETSC_ERR_ARG_NOTSAMECOMM, "IS communicator must match matrix communicator");
304: PetscCall(ISAllGather(iscol, &iscol_local));
305: PetscCall(ISGetIndices(isrow, &irow));
306: PetscCall(ISGetIndices(iscol_local, &icol));
307: PetscCall(ISGetLocalSize(isrow, &nrows));
308: PetscCall(ISGetLocalSize(iscol, &ncols));
309: PetscCall(ISGetSize(iscol, &Ncols)); /* global number of columns, size of iscol_local */
311: /* No parallel redistribution currently supported! Should really check each index set
312: to confirm that it is OK. ... Currently supports only submatrix same partitioning as
313: original matrix! */
315: PetscCall(MatGetLocalSize(A, &nlrows, &nlcols));
316: PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
318: /* Check submatrix call */
319: if (scall == MAT_REUSE_MATRIX) {
320: /* SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size"); */
321: /* Really need to test rows and column sizes! */
322: newmat = *B;
323: } else {
324: /* Create and fill new matrix */
325: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &newmat));
326: PetscCall(MatSetSizes(newmat, nrows, ncols, PETSC_DECIDE, Ncols));
327: PetscCall(MatSetType(newmat, ((PetscObject)A)->type_name));
328: PetscCall(MatMPIDenseSetPreallocation(newmat, NULL));
329: }
331: /* Now extract the data pointers and do the copy, column at a time */
332: newmatd = (Mat_MPIDense *)newmat->data;
333: PetscCall(MatDenseGetArray(newmatd->A, &bv));
334: PetscCall(MatDenseGetArrayRead(mat->A, &v));
335: PetscCall(MatDenseGetLDA(mat->A, &lda));
336: for (i = 0; i < Ncols; i++) {
337: const PetscScalar *av = v + lda * icol[i];
338: for (j = 0; j < nrows; j++) *bv++ = av[irow[j] - rstart];
339: }
340: PetscCall(MatDenseRestoreArrayRead(mat->A, &v));
341: PetscCall(MatDenseRestoreArray(newmatd->A, &bv));
343: /* Assemble the matrices so that the correct flags are set */
344: PetscCall(MatAssemblyBegin(newmat, MAT_FINAL_ASSEMBLY));
345: PetscCall(MatAssemblyEnd(newmat, MAT_FINAL_ASSEMBLY));
347: /* Free work space */
348: PetscCall(ISRestoreIndices(isrow, &irow));
349: PetscCall(ISRestoreIndices(iscol_local, &icol));
350: PetscCall(ISDestroy(&iscol_local));
351: *B = newmat;
352: PetscFunctionReturn(PETSC_SUCCESS);
353: }
355: static PetscErrorCode MatDenseRestoreArray_MPIDense(Mat A, PetscScalar **array)
356: {
357: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
359: PetscFunctionBegin;
360: PetscCall(MatDenseRestoreArray(a->A, array));
361: PetscFunctionReturn(PETSC_SUCCESS);
362: }
364: static PetscErrorCode MatDenseRestoreArrayRead_MPIDense(Mat A, PetscScalar **array)
365: {
366: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
368: PetscFunctionBegin;
369: PetscCall(MatDenseRestoreArrayRead(a->A, (const PetscScalar **)array));
370: PetscFunctionReturn(PETSC_SUCCESS);
371: }
373: static PetscErrorCode MatDenseRestoreArrayWrite_MPIDense(Mat A, PetscScalar **array)
374: {
375: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
377: PetscFunctionBegin;
378: PetscCall(MatDenseRestoreArrayWrite(a->A, array));
379: PetscFunctionReturn(PETSC_SUCCESS);
380: }
382: static PetscErrorCode MatAssemblyBegin_MPIDense(Mat mat, MatAssemblyType mode)
383: {
384: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
385: PetscInt nstash, reallocs;
387: PetscFunctionBegin;
388: if (mdn->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
390: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
391: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
392: PetscCall(PetscInfo(mdn->A, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
393: PetscFunctionReturn(PETSC_SUCCESS);
394: }
396: static PetscErrorCode MatAssemblyEnd_MPIDense(Mat mat, MatAssemblyType mode)
397: {
398: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
399: PetscInt i, *row, *col, flg, j, rstart, ncols;
400: PetscMPIInt n;
401: PetscScalar *val;
403: PetscFunctionBegin;
404: if (!mdn->donotstash && !mat->nooffprocentries) {
405: /* wait on receives */
406: while (1) {
407: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
408: if (!flg) break;
410: for (i = 0; i < n;) {
411: /* Now identify the consecutive vals belonging to the same row */
412: for (j = i, rstart = row[j]; j < n; j++) {
413: if (row[j] != rstart) break;
414: }
415: if (j < n) ncols = j - i;
416: else ncols = n - i;
417: /* Now assemble all these values with a single function call */
418: PetscCall(MatSetValues_MPIDense(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
419: i = j;
420: }
421: }
422: PetscCall(MatStashScatterEnd_Private(&mat->stash));
423: }
425: PetscCall(MatAssemblyBegin(mdn->A, mode));
426: PetscCall(MatAssemblyEnd(mdn->A, mode));
427: PetscFunctionReturn(PETSC_SUCCESS);
428: }
430: static PetscErrorCode MatZeroEntries_MPIDense(Mat A)
431: {
432: Mat_MPIDense *l = (Mat_MPIDense *)A->data;
434: PetscFunctionBegin;
435: PetscCall(MatZeroEntries(l->A));
436: PetscFunctionReturn(PETSC_SUCCESS);
437: }
439: static PetscErrorCode MatZeroRows_MPIDense(Mat A, PetscInt n, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
440: {
441: Mat_MPIDense *l = (Mat_MPIDense *)A->data;
442: PetscInt i, len, *lrows;
444: PetscFunctionBegin;
445: /* get locally owned rows */
446: PetscCall(PetscLayoutMapLocal(A->rmap, n, rows, &len, &lrows, NULL));
447: /* fix right-hand side if needed */
448: if (x && b) {
449: const PetscScalar *xx;
450: PetscScalar *bb;
452: PetscCall(VecGetArrayRead(x, &xx));
453: PetscCall(VecGetArrayWrite(b, &bb));
454: for (i = 0; i < len; ++i) bb[lrows[i]] = diag * xx[lrows[i]];
455: PetscCall(VecRestoreArrayRead(x, &xx));
456: PetscCall(VecRestoreArrayWrite(b, &bb));
457: }
458: PetscCall(MatZeroRows(l->A, len, lrows, 0.0, NULL, NULL));
459: if (diag != 0.0) {
460: Vec d;
462: PetscCall(MatCreateVecs(A, NULL, &d));
463: PetscCall(VecSet(d, diag));
464: PetscCall(MatDiagonalSet(A, d, INSERT_VALUES));
465: PetscCall(VecDestroy(&d));
466: }
467: PetscCall(PetscFree(lrows));
468: PetscFunctionReturn(PETSC_SUCCESS);
469: }
471: PETSC_INTERN PetscErrorCode MatMult_SeqDense(Mat, Vec, Vec);
472: PETSC_INTERN PetscErrorCode MatMultAdd_SeqDense(Mat, Vec, Vec, Vec);
473: PETSC_INTERN PetscErrorCode MatMultTranspose_SeqDense(Mat, Vec, Vec);
474: PETSC_INTERN PetscErrorCode MatMultTransposeAdd_SeqDense(Mat, Vec, Vec, Vec);
476: static PetscErrorCode MatMultColumnRange_MPIDense(Mat mat, Vec xx, Vec yy, PetscInt c_start, PetscInt c_end)
477: {
478: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
479: const PetscScalar *ax;
480: PetscScalar *ay;
481: PetscMemType axmtype, aymtype;
483: PetscFunctionBegin;
484: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(mat));
485: PetscCall(VecGetArrayReadAndMemType(xx, &ax, &axmtype));
486: PetscCall(VecGetArrayWriteAndMemType(mdn->lvec, &ay, &aymtype));
487: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPI_REPLACE));
488: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, ax, ay, MPI_REPLACE));
489: PetscCall(VecRestoreArrayWriteAndMemType(mdn->lvec, &ay));
490: PetscCall(VecRestoreArrayReadAndMemType(xx, &ax));
491: PetscUseMethod(mdn->A, "MatMultColumnRange_C", (Mat, Vec, Vec, PetscInt, PetscInt), (mdn->A, mdn->lvec, yy, c_start, c_end));
492: PetscFunctionReturn(PETSC_SUCCESS);
493: }
495: static PetscErrorCode MatMult_MPIDense(Mat mat, Vec xx, Vec yy)
496: {
497: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
498: const PetscScalar *ax;
499: PetscScalar *ay;
500: PetscMemType axmtype, aymtype;
502: PetscFunctionBegin;
503: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(mat));
504: PetscCall(VecGetArrayReadAndMemType(xx, &ax, &axmtype));
505: PetscCall(VecGetArrayWriteAndMemType(mdn->lvec, &ay, &aymtype));
506: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPI_REPLACE));
507: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, ax, ay, MPI_REPLACE));
508: PetscCall(VecRestoreArrayWriteAndMemType(mdn->lvec, &ay));
509: PetscCall(VecRestoreArrayReadAndMemType(xx, &ax));
510: PetscUseTypeMethod(mdn->A, mult, mdn->lvec, yy);
511: PetscFunctionReturn(PETSC_SUCCESS);
512: }
514: static PetscErrorCode MatGetMultPetscSF_MPIDense(Mat A, PetscSF *sf)
515: {
516: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
518: PetscFunctionBegin;
519: *sf = mdn->Mvctx;
520: PetscFunctionReturn(PETSC_SUCCESS);
521: }
523: static PetscErrorCode MatMultAddColumnRange_MPIDense(Mat mat, Vec xx, Vec yy, Vec zz, PetscInt c_start, PetscInt c_end)
524: {
525: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
526: const PetscScalar *ax;
527: PetscScalar *ay;
528: PetscMemType axmtype, aymtype;
530: PetscFunctionBegin;
531: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(mat));
532: PetscCall(VecGetArrayReadAndMemType(xx, &ax, &axmtype));
533: PetscCall(VecGetArrayAndMemType(mdn->lvec, &ay, &aymtype));
534: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPI_REPLACE));
535: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, ax, ay, MPI_REPLACE));
536: PetscCall(VecRestoreArrayAndMemType(mdn->lvec, &ay));
537: PetscCall(VecRestoreArrayReadAndMemType(xx, &ax));
538: PetscUseMethod(mdn->A, "MatMultAddColumnRange_C", (Mat, Vec, Vec, Vec, PetscInt, PetscInt), (mdn->A, mdn->lvec, yy, zz, c_start, c_end));
539: PetscFunctionReturn(PETSC_SUCCESS);
540: }
542: static PetscErrorCode MatMultAdd_MPIDense(Mat mat, Vec xx, Vec yy, Vec zz)
543: {
544: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
545: const PetscScalar *ax;
546: PetscScalar *ay;
547: PetscMemType axmtype, aymtype;
549: PetscFunctionBegin;
550: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(mat));
551: PetscCall(VecGetArrayReadAndMemType(xx, &ax, &axmtype));
552: PetscCall(VecGetArrayAndMemType(mdn->lvec, &ay, &aymtype));
553: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPI_REPLACE));
554: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, ax, ay, MPI_REPLACE));
555: PetscCall(VecRestoreArrayAndMemType(mdn->lvec, &ay));
556: PetscCall(VecRestoreArrayReadAndMemType(xx, &ax));
557: PetscUseTypeMethod(mdn->A, multadd, mdn->lvec, yy, zz);
558: PetscFunctionReturn(PETSC_SUCCESS);
559: }
561: static PetscErrorCode MatMultHermitianTransposeColumnRange_MPIDense(Mat A, Vec xx, Vec yy, PetscInt c_start, PetscInt c_end)
562: {
563: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
564: const PetscScalar *ax;
565: PetscScalar *ay;
566: PetscMemType axmtype, aymtype;
567: PetscInt r_start, r_end;
568: PetscInt c_start_local, c_end_local;
570: PetscFunctionBegin;
571: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
572: PetscCall(VecZeroEntries(a->lvec));
573: PetscCall(VecGetOwnershipRange(yy, &r_start, &r_end));
574: c_start_local = PetscMax(c_start, r_start);
575: c_end_local = PetscMin(c_end, r_end);
576: PetscCall(VecGetArrayAndMemType(yy, &ay, &aymtype));
577: if (c_end_local > c_start_local) {
578: if (PetscMemTypeHost(aymtype)) {
579: PetscCall(PetscArrayzero(&ay[c_start_local], (size_t)(c_end_local - c_start_local)));
580: } else {
581: PetscCall(PetscDeviceRegisterMemory(ay, aymtype, sizeof(*ay) * ((size_t)(r_end - r_start))));
582: PetscCall(PetscDeviceArrayZero(NULL, &ay[c_start_local], (size_t)(c_end_local - c_start_local)));
583: }
584: }
585: PetscUseMethod(a->A, "MatMultHermitianTransposeColumnRange_C", (Mat, Vec, Vec, PetscInt, PetscInt), (a->A, xx, a->lvec, c_start, c_end));
586: PetscCall(VecGetArrayReadAndMemType(a->lvec, &ax, &axmtype));
587: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPIU_SUM));
588: PetscCall(PetscSFReduceEnd(a->Mvctx, MPIU_SCALAR, ax, ay, MPIU_SUM));
589: PetscCall(VecRestoreArrayReadAndMemType(a->lvec, &ax));
590: PetscCall(VecRestoreArrayAndMemType(yy, &ay));
591: PetscFunctionReturn(PETSC_SUCCESS);
592: }
594: static PetscErrorCode MatMultTransposeKernel_MPIDense(Mat A, Vec xx, Vec yy, PetscBool herm)
595: {
596: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
597: const PetscScalar *ax;
598: PetscScalar *ay;
599: PetscMemType axmtype, aymtype;
601: PetscFunctionBegin;
602: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
603: PetscCall(VecSet(yy, 0.0));
604: if (herm) PetscUseTypeMethod(a->A, multhermitiantranspose, xx, a->lvec);
605: else PetscUseTypeMethod(a->A, multtranspose, xx, a->lvec);
606: PetscCall(VecGetArrayReadAndMemType(a->lvec, &ax, &axmtype));
607: PetscCall(VecGetArrayAndMemType(yy, &ay, &aymtype));
608: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPIU_SUM));
609: PetscCall(PetscSFReduceEnd(a->Mvctx, MPIU_SCALAR, ax, ay, MPIU_SUM));
610: PetscCall(VecRestoreArrayReadAndMemType(a->lvec, &ax));
611: PetscCall(VecRestoreArrayAndMemType(yy, &ay));
612: PetscFunctionReturn(PETSC_SUCCESS);
613: }
615: static PetscErrorCode MatMultHermitianTransposeAddColumnRange_MPIDense(Mat A, Vec xx, Vec yy, Vec zz, PetscInt c_start, PetscInt c_end)
616: {
617: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
618: const PetscScalar *ax;
619: PetscScalar *ay;
620: PetscMemType axmtype, aymtype;
622: PetscFunctionBegin;
623: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
624: PetscCall(VecCopy(yy, zz));
625: PetscCall(VecZeroEntries(a->lvec));
626: PetscUseMethod(a->A, "MatMultHermitianTransposeColumnRange_C", (Mat, Vec, Vec, PetscInt, PetscInt), (a->A, xx, a->lvec, c_start, c_end));
627: PetscCall(VecGetArrayReadAndMemType(a->lvec, &ax, &axmtype));
628: PetscCall(VecGetArrayAndMemType(zz, &ay, &aymtype));
629: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPIU_SUM));
630: PetscCall(PetscSFReduceEnd(a->Mvctx, MPIU_SCALAR, ax, ay, MPIU_SUM));
631: PetscCall(VecRestoreArrayReadAndMemType(a->lvec, &ax));
632: PetscCall(VecRestoreArrayAndMemType(zz, &ay));
633: PetscFunctionReturn(PETSC_SUCCESS);
634: }
636: static PetscErrorCode MatMultTransposeAddKernel_MPIDense(Mat A, Vec xx, Vec yy, Vec zz, PetscBool herm)
637: {
638: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
639: const PetscScalar *ax;
640: PetscScalar *ay;
641: PetscMemType axmtype, aymtype;
643: PetscFunctionBegin;
644: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
645: PetscCall(VecCopy(yy, zz));
646: if (herm) PetscUseTypeMethod(a->A, multhermitiantranspose, xx, a->lvec);
647: else PetscUseTypeMethod(a->A, multtranspose, xx, a->lvec);
648: PetscCall(VecGetArrayReadAndMemType(a->lvec, &ax, &axmtype));
649: PetscCall(VecGetArrayAndMemType(zz, &ay, &aymtype));
650: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPIU_SUM));
651: PetscCall(PetscSFReduceEnd(a->Mvctx, MPIU_SCALAR, ax, ay, MPIU_SUM));
652: PetscCall(VecRestoreArrayReadAndMemType(a->lvec, &ax));
653: PetscCall(VecRestoreArrayAndMemType(zz, &ay));
654: PetscFunctionReturn(PETSC_SUCCESS);
655: }
657: static PetscErrorCode MatMultTranspose_MPIDense(Mat A, Vec xx, Vec yy)
658: {
659: PetscFunctionBegin;
660: PetscCall(MatMultTransposeKernel_MPIDense(A, xx, yy, PETSC_FALSE));
661: PetscFunctionReturn(PETSC_SUCCESS);
662: }
664: static PetscErrorCode MatMultTransposeAdd_MPIDense(Mat A, Vec xx, Vec yy, Vec zz)
665: {
666: PetscFunctionBegin;
667: PetscCall(MatMultTransposeAddKernel_MPIDense(A, xx, yy, zz, PETSC_FALSE));
668: PetscFunctionReturn(PETSC_SUCCESS);
669: }
671: static PetscErrorCode MatMultHermitianTranspose_MPIDense(Mat A, Vec xx, Vec yy)
672: {
673: PetscFunctionBegin;
674: PetscCall(MatMultTransposeKernel_MPIDense(A, xx, yy, PETSC_TRUE));
675: PetscFunctionReturn(PETSC_SUCCESS);
676: }
678: static PetscErrorCode MatMultHermitianTransposeAdd_MPIDense(Mat A, Vec xx, Vec yy, Vec zz)
679: {
680: PetscFunctionBegin;
681: PetscCall(MatMultTransposeAddKernel_MPIDense(A, xx, yy, zz, PETSC_TRUE));
682: PetscFunctionReturn(PETSC_SUCCESS);
683: }
685: PetscErrorCode MatGetDiagonal_MPIDense(Mat A, Vec v)
686: {
687: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
688: PetscInt lda, len, i, nl, ng, m = A->rmap->n, radd;
689: PetscScalar *x;
690: const PetscScalar *av;
692: PetscFunctionBegin;
693: PetscCall(VecGetArray(v, &x));
694: PetscCall(VecGetSize(v, &ng));
695: PetscCheck(ng == A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming mat and vec");
696: PetscCall(VecGetLocalSize(v, &nl));
697: len = PetscMin(a->A->rmap->n, a->A->cmap->n);
698: radd = A->rmap->rstart * m;
699: PetscCall(MatDenseGetArrayRead(a->A, &av));
700: PetscCall(MatDenseGetLDA(a->A, &lda));
701: for (i = 0; i < len; i++) x[i] = av[radd + i * lda + i];
702: PetscCall(MatDenseRestoreArrayRead(a->A, &av));
703: if (nl - i > 0) PetscCall(PetscArrayzero(x + i, nl - i));
704: PetscCall(VecRestoreArray(v, &x));
705: PetscFunctionReturn(PETSC_SUCCESS);
706: }
708: static PetscErrorCode MatDestroy_MPIDense(Mat mat)
709: {
710: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
712: PetscFunctionBegin;
713: PetscCall(PetscLogObjectState((PetscObject)mat, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT, mat->rmap->N, mat->cmap->N));
714: PetscCall(MatStashDestroy_Private(&mat->stash));
715: PetscCheck(!mdn->vecinuse, PetscObjectComm((PetscObject)mat), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
716: PetscCheck(!mdn->matinuse, PetscObjectComm((PetscObject)mat), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
717: PetscCall(MatDestroy(&mdn->A));
718: PetscCall(VecDestroy(&mdn->lvec));
719: PetscCall(PetscSFDestroy(&mdn->Mvctx));
720: PetscCall(VecDestroy(&mdn->cvec));
721: PetscCall(MatDestroy(&mdn->cmat));
723: PetscCall(PetscFree(mat->data));
724: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, NULL));
726: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetLDA_C", NULL));
727: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseSetLDA_C", NULL));
728: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArray_C", NULL));
729: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArray_C", NULL));
730: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayRead_C", NULL));
731: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayRead_C", NULL));
732: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayWrite_C", NULL));
733: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayWrite_C", NULL));
734: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDensePlaceArray_C", NULL));
735: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseResetArray_C", NULL));
736: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseReplaceArray_C", NULL));
737: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", NULL));
738: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpiaij_C", NULL));
739: #if PetscDefined(HAVE_ELEMENTAL)
740: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_elemental_C", NULL));
741: #endif
742: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
743: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_scalapack_C", NULL));
744: #endif
745: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIDenseSetPreallocation_C", NULL));
746: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpidense_C", NULL));
747: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaij_C", NULL));
748: #if PetscDefined(HAVE_CUDA)
749: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijcusparse_mpidense_C", NULL));
750: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijcusparse_C", NULL));
751: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensecuda_C", NULL));
752: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidensecuda_mpidense_C", NULL));
753: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpidensecuda_C", NULL));
754: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijcusparse_mpidensecuda_C", NULL));
755: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidensecuda_mpiaij_C", NULL));
756: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidensecuda_mpiaijcusparse_C", NULL));
757: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAGetArray_C", NULL));
758: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAGetArrayRead_C", NULL));
759: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAGetArrayWrite_C", NULL));
760: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDARestoreArray_C", NULL));
761: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDARestoreArrayRead_C", NULL));
762: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDARestoreArrayWrite_C", NULL));
763: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAPlaceArray_C", NULL));
764: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAResetArray_C", NULL));
765: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAReplaceArray_C", NULL));
766: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDASetPreallocation_C", NULL));
767: #endif
768: #if PetscDefined(HAVE_HIP)
769: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijhipsparse_mpidense_C", NULL));
770: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijhipsparse_C", NULL));
771: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensehip_C", NULL));
772: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidensehip_mpidense_C", NULL));
773: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpidensehip_C", NULL));
774: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijhipsparse_mpidensehip_C", NULL));
775: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidensehip_mpiaij_C", NULL));
776: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidensehip_mpiaijhipsparse_C", NULL));
777: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPGetArray_C", NULL));
778: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPGetArrayRead_C", NULL));
779: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPGetArrayWrite_C", NULL));
780: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPRestoreArray_C", NULL));
781: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPRestoreArrayRead_C", NULL));
782: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPRestoreArrayWrite_C", NULL));
783: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPPlaceArray_C", NULL));
784: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPResetArray_C", NULL));
785: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPReplaceArray_C", NULL));
786: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPSetPreallocation_C", NULL));
787: #endif
788: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumn_C", NULL));
789: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumn_C", NULL));
790: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVec_C", NULL));
791: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVec_C", NULL));
792: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecRead_C", NULL));
793: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecRead_C", NULL));
794: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecWrite_C", NULL));
795: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecWrite_C", NULL));
796: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetSubMatrix_C", NULL));
797: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreSubMatrix_C", NULL));
798: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultColumnRange_C", NULL));
799: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultAddColumnRange_C", NULL));
800: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeColumnRange_C", NULL));
801: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeAddColumnRange_C", NULL));
802: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatGetMultPetscSF_C", NULL));
804: PetscCall(PetscObjectCompose((PetscObject)mat, "DiagonalBlock", NULL));
805: PetscFunctionReturn(PETSC_SUCCESS);
806: }
808: #include <petscdraw.h>
809: static PetscErrorCode MatView_MPIDense_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
810: {
811: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
812: PetscMPIInt rank;
813: PetscViewerType vtype;
814: PetscBool isascii, isdraw;
815: PetscViewer sviewer;
816: PetscViewerFormat format;
818: PetscFunctionBegin;
819: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
820: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
821: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
822: if (isascii) {
823: PetscCall(PetscViewerGetType(viewer, &vtype));
824: PetscCall(PetscViewerGetFormat(viewer, &format));
825: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
826: MatInfo info;
827: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
828: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
829: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, " [%d] local rows %" PetscInt_FMT " nz %" PetscInt_FMT " nz alloced %" PetscInt_FMT " mem %" PetscInt_FMT " \n", rank, mat->rmap->n, (PetscInt)info.nz_used, (PetscInt)info.nz_allocated,
830: (PetscInt)info.memory));
831: PetscCall(PetscViewerFlush(viewer));
832: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
833: if (mdn->Mvctx) PetscCall(PetscSFView(mdn->Mvctx, viewer));
834: PetscFunctionReturn(PETSC_SUCCESS);
835: } else if (format == PETSC_VIEWER_ASCII_INFO) {
836: PetscFunctionReturn(PETSC_SUCCESS);
837: }
838: } else if (isdraw) {
839: PetscDraw draw;
840: PetscBool isnull;
842: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
843: PetscCall(PetscDrawIsNull(draw, &isnull));
844: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
845: }
847: {
848: /* assemble the entire matrix onto first processor. */
849: Mat A;
850: PetscInt M = mat->rmap->N, N = mat->cmap->N, m, row, i, nz;
851: PetscInt *cols;
852: PetscScalar *vals;
854: PetscCall(MatCreate(PetscObjectComm((PetscObject)mat), &A));
855: if (rank == 0) {
856: PetscCall(MatSetSizes(A, M, N, M, N));
857: } else {
858: PetscCall(MatSetSizes(A, 0, 0, M, N));
859: }
860: /* Since this is a temporary matrix, MATMPIDENSE instead of ((PetscObject)A)->type_name here is probably acceptable. */
861: PetscCall(MatSetType(A, MATMPIDENSE));
862: PetscCall(MatMPIDenseSetPreallocation(A, NULL));
864: /* Copy the matrix ... This isn't the most efficient means,
865: but it's quick for now */
866: A->insertmode = INSERT_VALUES;
868: row = mat->rmap->rstart;
869: m = mdn->A->rmap->n;
870: for (i = 0; i < m; i++) {
871: PetscCall(MatGetRow_MPIDense(mat, row, &nz, &cols, &vals));
872: PetscCall(MatSetValues_MPIDense(A, 1, &row, nz, cols, vals, INSERT_VALUES));
873: PetscCall(MatRestoreRow_MPIDense(mat, row, &nz, &cols, &vals));
874: row++;
875: }
877: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
878: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
879: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
880: if (rank == 0) {
881: PetscCall(PetscObjectSetName((PetscObject)((Mat_MPIDense *)A->data)->A, ((PetscObject)mat)->name));
882: PetscCall(MatView_SeqDense(((Mat_MPIDense *)A->data)->A, sviewer));
883: }
884: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
885: PetscCall(MatDestroy(&A));
886: }
887: PetscFunctionReturn(PETSC_SUCCESS);
888: }
890: static PetscErrorCode MatView_MPIDense(Mat mat, PetscViewer viewer)
891: {
892: PetscBool isascii, isbinary, isdraw, issocket;
894: PetscFunctionBegin;
895: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
896: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
897: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
898: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
900: if (isascii || issocket || isdraw) PetscCall(MatView_MPIDense_ASCIIorDraworSocket(mat, viewer));
901: else if (isbinary) PetscCall(MatView_Dense_Binary(mat, viewer));
902: PetscFunctionReturn(PETSC_SUCCESS);
903: }
905: static PetscErrorCode MatGetInfo_MPIDense(Mat A, MatInfoType flag, MatInfo *info)
906: {
907: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
908: Mat mdn = mat->A;
909: PetscLogDouble irecv[5];
911: PetscFunctionBegin;
912: info->block_size = 1.0;
914: PetscCall(MatGetInfo(mdn, MAT_LOCAL, info));
916: irecv[0] = info->nz_used;
917: irecv[1] = info->nz_allocated;
918: irecv[2] = info->nz_unneeded;
919: irecv[3] = info->memory;
920: irecv[4] = info->mallocs;
921: if (flag == MAT_LOCAL) {
922: info->nz_used = irecv[0];
923: info->nz_allocated = irecv[1];
924: info->nz_unneeded = irecv[2];
925: info->memory = irecv[3];
926: info->mallocs = irecv[4];
927: } else if (flag == MAT_GLOBAL_MAX) {
928: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)A)));
930: info->nz_used = irecv[0];
931: info->nz_allocated = irecv[1];
932: info->nz_unneeded = irecv[2];
933: info->memory = irecv[3];
934: info->mallocs = irecv[4];
935: } else if (flag == MAT_GLOBAL_SUM) {
936: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)A)));
938: info->nz_used = irecv[0];
939: info->nz_allocated = irecv[1];
940: info->nz_unneeded = irecv[2];
941: info->memory = irecv[3];
942: info->mallocs = irecv[4];
943: }
944: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
945: info->fill_ratio_needed = 0;
946: info->factor_mallocs = 0;
947: PetscFunctionReturn(PETSC_SUCCESS);
948: }
950: static PetscErrorCode MatSetOption_MPIDense(Mat A, MatOption op, PetscBool flg)
951: {
952: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
954: PetscFunctionBegin;
955: switch (op) {
956: case MAT_NEW_NONZERO_LOCATIONS:
957: case MAT_NEW_NONZERO_LOCATION_ERR:
958: case MAT_NEW_NONZERO_ALLOCATION_ERR:
959: MatCheckPreallocated(A, 1);
960: PetscCall(MatSetOption(a->A, op, flg));
961: break;
962: case MAT_ROW_ORIENTED:
963: MatCheckPreallocated(A, 1);
964: a->roworiented = flg;
965: PetscCall(MatSetOption(a->A, op, flg));
966: break;
967: case MAT_IGNORE_OFF_PROC_ENTRIES:
968: a->donotstash = flg;
969: break;
970: case MAT_SYMMETRIC:
971: case MAT_STRUCTURALLY_SYMMETRIC:
972: case MAT_HERMITIAN:
973: case MAT_SYMMETRY_ETERNAL:
974: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
975: case MAT_SPD:
976: case MAT_SPD_ETERNAL:
977: /* if the diagonal matrix is square it inherits some of the properties above */
978: if (a->A && A->rmap->n == A->cmap->n) PetscCall(MatSetOption(a->A, op, flg));
979: break;
980: default:
981: break;
982: }
983: PetscFunctionReturn(PETSC_SUCCESS);
984: }
986: static PetscErrorCode MatDiagonalScale_MPIDense(Mat A, Vec ll, Vec rr)
987: {
988: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
989: const PetscScalar *l;
990: PetscScalar x, *v, *vv, *r;
991: PetscInt i, j, s2a, s3a, s2, s3, m = mdn->A->rmap->n, n = mdn->A->cmap->n, lda;
993: PetscFunctionBegin;
994: PetscCall(MatDenseGetArray(mdn->A, &vv));
995: PetscCall(MatDenseGetLDA(mdn->A, &lda));
996: PetscCall(MatGetLocalSize(A, &s2, &s3));
997: if (ll) {
998: PetscCall(VecGetLocalSize(ll, &s2a));
999: PetscCheck(s2a == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vector non-conforming local size, %" PetscInt_FMT " != %" PetscInt_FMT, s2a, s2);
1000: PetscCall(VecGetArrayRead(ll, &l));
1001: for (i = 0; i < m; i++) {
1002: x = l[i];
1003: v = vv + i;
1004: for (j = 0; j < n; j++) {
1005: (*v) *= x;
1006: v += lda;
1007: }
1008: }
1009: PetscCall(VecRestoreArrayRead(ll, &l));
1010: PetscCall(PetscLogFlops(1.0 * n * m));
1011: }
1012: if (rr) {
1013: const PetscScalar *ar;
1015: PetscCall(VecGetLocalSize(rr, &s3a));
1016: PetscCheck(s3a == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vec non-conforming local size, %" PetscInt_FMT " != %" PetscInt_FMT ".", s3a, s3);
1017: PetscCall(VecGetArrayRead(rr, &ar));
1018: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
1019: PetscCall(VecGetArray(mdn->lvec, &r));
1020: PetscCall(PetscSFBcastBegin(mdn->Mvctx, MPIU_SCALAR, ar, r, MPI_REPLACE));
1021: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, ar, r, MPI_REPLACE));
1022: PetscCall(VecRestoreArrayRead(rr, &ar));
1023: for (i = 0; i < n; i++) {
1024: x = r[i];
1025: v = vv + i * lda;
1026: for (j = 0; j < m; j++) (*v++) *= x;
1027: }
1028: PetscCall(VecRestoreArray(mdn->lvec, &r));
1029: PetscCall(PetscLogFlops(1.0 * n * m));
1030: }
1031: PetscCall(MatDenseRestoreArray(mdn->A, &vv));
1032: PetscFunctionReturn(PETSC_SUCCESS);
1033: }
1035: static PetscErrorCode MatNorm_MPIDense(Mat A, NormType type, PetscReal *nrm)
1036: {
1037: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
1038: PetscInt i, j;
1039: PetscMPIInt size;
1040: const PetscScalar *av, *v;
1042: PetscFunctionBegin;
1043: PetscCall(MatDenseGetArrayRead(mdn->A, &av));
1044: v = av;
1045: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1046: if (size == 1) {
1047: PetscCall(MatNorm(mdn->A, type, nrm));
1048: } else {
1049: if (type == NORM_FROBENIUS) {
1050: *nrm = 0.0;
1051: for (i = 0; i < mdn->A->cmap->n * mdn->A->rmap->n; i++) {
1052: *nrm += PetscRealPart(PetscConj(*v) * (*v));
1053: v++;
1054: }
1055: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, nrm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
1056: *nrm = PetscSqrtReal(*nrm);
1057: PetscCall(PetscLogFlops(2.0 * mdn->A->cmap->n * mdn->A->rmap->n));
1058: } else if (type == NORM_1) {
1059: PetscReal *tmp;
1061: PetscCall(PetscCalloc1(A->cmap->N, &tmp));
1062: *nrm = 0.0;
1063: v = av;
1064: for (j = 0; j < mdn->A->cmap->n; j++) {
1065: for (i = 0; i < mdn->A->rmap->n; i++) {
1066: tmp[j] += PetscAbsScalar(*v);
1067: v++;
1068: }
1069: }
1070: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, tmp, A->cmap->N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
1071: for (j = 0; j < A->cmap->N; j++) {
1072: if (tmp[j] > *nrm) *nrm = tmp[j];
1073: }
1074: PetscCall(PetscFree(tmp));
1075: PetscCall(PetscLogFlops(A->cmap->n * A->rmap->n));
1076: } else if (type == NORM_INFINITY) { /* max row norm */
1077: PetscCall(MatNorm(mdn->A, type, nrm));
1078: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, nrm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)A)));
1079: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Unsupported norm type %s", NormTypes[type]);
1080: }
1081: PetscCall(MatDenseRestoreArrayRead(mdn->A, &av));
1082: PetscFunctionReturn(PETSC_SUCCESS);
1083: }
1085: static PetscErrorCode MatTranspose_MPIDense(Mat A, MatReuse reuse, Mat *matout)
1086: {
1087: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1088: Mat B;
1089: PetscInt M = A->rmap->N, N = A->cmap->N, m, n, *rwork, rstart = A->rmap->rstart;
1090: PetscInt j, i, lda;
1091: PetscScalar *v;
1093: PetscFunctionBegin;
1094: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1095: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_INPLACE_MATRIX) {
1096: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1097: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1098: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1099: PetscCall(MatMPIDenseSetPreallocation(B, NULL));
1100: } else B = *matout;
1102: m = a->A->rmap->n;
1103: n = a->A->cmap->n;
1104: PetscCall(MatDenseGetArrayRead(a->A, (const PetscScalar **)&v));
1105: PetscCall(MatDenseGetLDA(a->A, &lda));
1106: PetscCall(PetscMalloc1(m, &rwork));
1107: for (i = 0; i < m; i++) rwork[i] = rstart + i;
1108: for (j = 0; j < n; j++) {
1109: PetscCall(MatSetValues(B, 1, &j, m, rwork, v, INSERT_VALUES));
1110: v = PetscSafePointerPlusOffset(v, lda);
1111: }
1112: PetscCall(MatDenseRestoreArrayRead(a->A, (const PetscScalar **)&v));
1113: PetscCall(PetscFree(rwork));
1114: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1115: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1116: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1117: *matout = B;
1118: } else {
1119: PetscCall(MatHeaderMerge(A, &B));
1120: }
1121: PetscFunctionReturn(PETSC_SUCCESS);
1122: }
1124: static PetscErrorCode MatDuplicate_MPIDense(Mat, MatDuplicateOption, Mat *);
1125: PETSC_INTERN PetscErrorCode MatScale_MPIDense(Mat, PetscScalar);
1127: static PetscErrorCode MatSetUp_MPIDense(Mat A)
1128: {
1129: PetscFunctionBegin;
1130: PetscCall(PetscLayoutSetUp(A->rmap));
1131: PetscCall(PetscLayoutSetUp(A->cmap));
1132: if (!A->preallocated) PetscCall(MatMPIDenseSetPreallocation(A, NULL));
1133: PetscFunctionReturn(PETSC_SUCCESS);
1134: }
1136: static PetscErrorCode MatAXPY_MPIDense(Mat Y, PetscScalar alpha, Mat X, MatStructure str)
1137: {
1138: Mat_MPIDense *A = (Mat_MPIDense *)Y->data, *B = (Mat_MPIDense *)X->data;
1140: PetscFunctionBegin;
1141: PetscCall(MatAXPY(A->A, alpha, B->A, str));
1142: PetscFunctionReturn(PETSC_SUCCESS);
1143: }
1145: static PetscErrorCode MatConjugate_MPIDense(Mat mat)
1146: {
1147: Mat_MPIDense *a = (Mat_MPIDense *)mat->data;
1149: PetscFunctionBegin;
1150: PetscCall(MatConjugate(a->A));
1151: PetscFunctionReturn(PETSC_SUCCESS);
1152: }
1154: static PetscErrorCode MatRealPart_MPIDense(Mat A)
1155: {
1156: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1158: PetscFunctionBegin;
1159: PetscCall(MatRealPart(a->A));
1160: PetscFunctionReturn(PETSC_SUCCESS);
1161: }
1163: static PetscErrorCode MatImaginaryPart_MPIDense(Mat A)
1164: {
1165: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1167: PetscFunctionBegin;
1168: PetscCall(MatImaginaryPart(a->A));
1169: PetscFunctionReturn(PETSC_SUCCESS);
1170: }
1172: static PetscErrorCode MatGetColumnVector_MPIDense(Mat A, Vec v, PetscInt col)
1173: {
1174: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1176: PetscFunctionBegin;
1177: PetscCheck(a->A, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Missing local matrix");
1178: PetscCheck(a->A->ops->getcolumnvector, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Missing get column operation");
1179: PetscUseTypeMethod(a->A, getcolumnvector, v, col);
1180: PetscFunctionReturn(PETSC_SUCCESS);
1181: }
1183: PETSC_INTERN PetscErrorCode MatGetColumnReductions_SeqDense(Mat, PetscInt, PetscReal *);
1185: static PetscErrorCode MatGetColumnReductions_MPIDense(Mat A, PetscInt type, PetscReal *reductions)
1186: {
1187: PetscInt i, m, n;
1188: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1190: PetscFunctionBegin;
1191: PetscCall(MatGetSize(A, &m, &n));
1192: if (type == REDUCTION_MEAN_REALPART) {
1193: PetscCall(MatGetColumnReductions_SeqDense(a->A, (PetscInt)REDUCTION_SUM_REALPART, reductions));
1194: } else if (type == REDUCTION_MEAN_IMAGINARYPART) {
1195: PetscCall(MatGetColumnReductions_SeqDense(a->A, (PetscInt)REDUCTION_SUM_IMAGINARYPART, reductions));
1196: } else {
1197: PetscCall(MatGetColumnReductions_SeqDense(a->A, type, reductions));
1198: }
1199: if (type == NORM_2) {
1200: for (i = 0; i < n; i++) reductions[i] *= reductions[i];
1201: }
1202: if (type == NORM_INFINITY) {
1203: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, reductions, n, MPIU_REAL, MPIU_MAX, A->hdr.comm));
1204: } else {
1205: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, reductions, n, MPIU_REAL, MPIU_SUM, A->hdr.comm));
1206: }
1207: if (type == NORM_2) {
1208: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
1209: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
1210: for (i = 0; i < n; i++) reductions[i] /= m;
1211: }
1212: PetscFunctionReturn(PETSC_SUCCESS);
1213: }
1215: static PetscErrorCode MatSetRandom_MPIDense(Mat x, PetscRandom rctx)
1216: {
1217: Mat_MPIDense *d = (Mat_MPIDense *)x->data;
1219: PetscFunctionBegin;
1220: PetscCall(MatSetRandom(d->A, rctx));
1221: #if PetscDefined(HAVE_DEVICE)
1222: x->offloadmask = d->A->offloadmask;
1223: #endif
1224: PetscFunctionReturn(PETSC_SUCCESS);
1225: }
1227: static PetscErrorCode MatMatTransposeMultSymbolic_MPIDense_MPIDense(Mat, Mat, PetscReal, Mat);
1228: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense(Mat, Mat, Mat);
1229: static PetscErrorCode MatTransposeMatMultSymbolic_MPIDense_MPIDense(Mat, Mat, PetscReal, Mat);
1230: static PetscErrorCode MatTransposeMatMultNumeric_MPIDense_MPIDense(Mat, Mat, Mat);
1231: static PetscErrorCode MatEqual_MPIDense(Mat, Mat, PetscBool *);
1232: static PetscErrorCode MatLoad_MPIDense(Mat, PetscViewer);
1233: static PetscErrorCode MatProductSetFromOptions_MPIDense(Mat);
1235: static struct _MatOps MatOps_Values = {MatSetValues_MPIDense,
1236: MatGetRow_MPIDense,
1237: MatRestoreRow_MPIDense,
1238: MatMult_MPIDense,
1239: /* 4*/ MatMultAdd_MPIDense,
1240: MatMultTranspose_MPIDense,
1241: MatMultTransposeAdd_MPIDense,
1242: NULL,
1243: NULL,
1244: NULL,
1245: /* 10*/ NULL,
1246: NULL,
1247: NULL,
1248: NULL,
1249: MatTranspose_MPIDense,
1250: /* 15*/ MatGetInfo_MPIDense,
1251: MatEqual_MPIDense,
1252: MatGetDiagonal_MPIDense,
1253: MatDiagonalScale_MPIDense,
1254: MatNorm_MPIDense,
1255: /* 20*/ MatAssemblyBegin_MPIDense,
1256: MatAssemblyEnd_MPIDense,
1257: MatSetOption_MPIDense,
1258: MatZeroEntries_MPIDense,
1259: /* 24*/ MatZeroRows_MPIDense,
1260: NULL,
1261: NULL,
1262: NULL,
1263: NULL,
1264: /* 29*/ MatSetUp_MPIDense,
1265: NULL,
1266: NULL,
1267: MatGetDiagonalBlock_MPIDense,
1268: NULL,
1269: /* 34*/ MatDuplicate_MPIDense,
1270: NULL,
1271: NULL,
1272: NULL,
1273: NULL,
1274: /* 39*/ MatAXPY_MPIDense,
1275: MatCreateSubMatrices_MPIDense,
1276: NULL,
1277: MatGetValues_MPIDense,
1278: MatCopy_MPIDense,
1279: /* 44*/ NULL,
1280: MatScale_MPIDense,
1281: MatShift_MPIDense,
1282: NULL,
1283: NULL,
1284: /* 49*/ MatSetRandom_MPIDense,
1285: NULL,
1286: NULL,
1287: NULL,
1288: NULL,
1289: /* 54*/ NULL,
1290: NULL,
1291: NULL,
1292: NULL,
1293: NULL,
1294: /* 59*/ MatCreateSubMatrix_MPIDense,
1295: MatDestroy_MPIDense,
1296: MatView_MPIDense,
1297: NULL,
1298: NULL,
1299: /* 64*/ NULL,
1300: NULL,
1301: NULL,
1302: NULL,
1303: NULL,
1304: /* 69*/ NULL,
1305: NULL,
1306: NULL,
1307: NULL,
1308: NULL,
1309: /* 74*/ NULL,
1310: NULL,
1311: NULL,
1312: NULL,
1313: MatLoad_MPIDense,
1314: /* 79*/ NULL,
1315: NULL,
1316: NULL,
1317: NULL,
1318: /* 83*/ NULL,
1319: NULL,
1320: NULL,
1321: NULL,
1322: MatMatTransposeMultSymbolic_MPIDense_MPIDense,
1323: MatMatTransposeMultNumeric_MPIDense_MPIDense,
1324: /* 89*/ NULL,
1325: MatProductSetFromOptions_MPIDense,
1326: NULL,
1327: NULL,
1328: MatConjugate_MPIDense,
1329: /* 94*/ NULL,
1330: NULL,
1331: MatRealPart_MPIDense,
1332: MatImaginaryPart_MPIDense,
1333: NULL,
1334: /*99*/ NULL,
1335: NULL,
1336: NULL,
1337: NULL,
1338: MatGetColumnVector_MPIDense,
1339: /*104*/ NULL,
1340: NULL,
1341: NULL,
1342: NULL,
1343: NULL,
1344: /*109*/ NULL,
1345: NULL,
1346: MatMultHermitianTranspose_MPIDense,
1347: MatMultHermitianTransposeAdd_MPIDense,
1348: NULL,
1349: /*114*/ NULL,
1350: MatGetColumnReductions_MPIDense,
1351: NULL,
1352: NULL,
1353: NULL,
1354: /*120*/ MatTransposeMatMultSymbolic_MPIDense_MPIDense,
1355: MatTransposeMatMultNumeric_MPIDense_MPIDense,
1356: NULL,
1357: NULL,
1358: /*124*/ NULL,
1359: NULL,
1360: NULL,
1361: NULL,
1362: NULL,
1363: /*129*/ NULL,
1364: MatCreateMPIMatConcatenateSeqMat_MPIDense,
1365: NULL,
1366: NULL,
1367: NULL,
1368: /*134*/ NULL,
1369: NULL,
1370: NULL,
1371: NULL,
1372: NULL,
1373: /*139*/ NULL,
1374: NULL,
1375: NULL,
1376: NULL,
1377: NULL,
1378: MatADot_Default,
1379: /*144*/ MatANorm_Default,
1380: NULL,
1381: NULL,
1382: NULL};
1384: static PetscErrorCode MatMPIDenseSetPreallocation_MPIDense(Mat mat, PetscScalar *data)
1385: {
1386: Mat_MPIDense *a = (Mat_MPIDense *)mat->data;
1387: MatType mtype = MATSEQDENSE;
1389: PetscFunctionBegin;
1390: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)mat), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1391: PetscCall(PetscLayoutSetUp(mat->rmap));
1392: PetscCall(PetscLayoutSetUp(mat->cmap));
1393: if (!a->A) {
1394: PetscCall(MatCreate(PETSC_COMM_SELF, &a->A));
1395: PetscCall(MatSetSizes(a->A, mat->rmap->n, mat->cmap->N, mat->rmap->n, mat->cmap->N));
1396: }
1397: #if PetscDefined(HAVE_CUDA)
1398: PetscBool iscuda;
1399: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATMPIDENSECUDA, &iscuda));
1400: if (iscuda) mtype = MATSEQDENSECUDA;
1401: #endif
1402: #if PetscDefined(HAVE_HIP)
1403: PetscBool iship;
1404: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATMPIDENSEHIP, &iship));
1405: if (iship) mtype = MATSEQDENSEHIP;
1406: #endif
1407: PetscCall(MatSetType(a->A, mtype));
1408: PetscCall(MatSeqDenseSetPreallocation(a->A, data));
1409: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
1410: mat->offloadmask = a->A->offloadmask;
1411: #endif
1412: mat->preallocated = PETSC_TRUE;
1413: mat->assembled = PETSC_TRUE;
1414: PetscFunctionReturn(PETSC_SUCCESS);
1415: }
1417: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIDense(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
1418: {
1419: Mat B, C;
1421: PetscFunctionBegin;
1422: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &C));
1423: PetscCall(MatConvert_SeqAIJ_SeqDense(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &B));
1424: PetscCall(MatDestroy(&C));
1425: if (reuse == MAT_REUSE_MATRIX) {
1426: C = *newmat;
1427: } else C = NULL;
1428: PetscCall(MatCreateMPIMatConcatenateSeqMat(PetscObjectComm((PetscObject)A), B, A->cmap->n, !C ? MAT_INITIAL_MATRIX : MAT_REUSE_MATRIX, &C));
1429: PetscCall(MatDestroy(&B));
1430: if (reuse == MAT_INPLACE_MATRIX) PetscCall(MatHeaderReplace(A, &C));
1431: else if (reuse == MAT_INITIAL_MATRIX) *newmat = C;
1432: PetscFunctionReturn(PETSC_SUCCESS);
1433: }
1435: static PetscErrorCode MatConvert_MPIDense_MPIAIJ(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
1436: {
1437: Mat B, C;
1439: PetscFunctionBegin;
1440: PetscCall(MatDenseGetLocalMatrix(A, &C));
1441: PetscCall(MatConvert_SeqDense_SeqAIJ(C, MATSEQAIJ, MAT_INITIAL_MATRIX, &B));
1442: if (reuse == MAT_REUSE_MATRIX) {
1443: C = *newmat;
1444: } else C = NULL;
1445: PetscCall(MatCreateMPIMatConcatenateSeqMat(PetscObjectComm((PetscObject)A), B, A->cmap->n, !C ? MAT_INITIAL_MATRIX : MAT_REUSE_MATRIX, &C));
1446: PetscCall(MatDestroy(&B));
1447: if (reuse == MAT_INPLACE_MATRIX) PetscCall(MatHeaderReplace(A, &C));
1448: else if (reuse == MAT_INITIAL_MATRIX) *newmat = C;
1449: PetscFunctionReturn(PETSC_SUCCESS);
1450: }
1452: #if PetscDefined(HAVE_ELEMENTAL)
1453: PETSC_INTERN PetscErrorCode MatConvert_MPIDense_Elemental(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
1454: {
1455: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1456: Mat mat_elemental;
1457: PetscScalar *v;
1458: PetscInt m = A->rmap->n, N = A->cmap->N, rstart = A->rmap->rstart, i, *rows, *cols, lda;
1460: PetscFunctionBegin;
1461: if (reuse == MAT_REUSE_MATRIX) {
1462: mat_elemental = *newmat;
1463: PetscCall(MatZeroEntries(*newmat));
1464: } else {
1465: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental));
1466: PetscCall(MatSetSizes(mat_elemental, PETSC_DECIDE, PETSC_DECIDE, A->rmap->N, A->cmap->N));
1467: PetscCall(MatSetType(mat_elemental, MATELEMENTAL));
1468: PetscCall(MatSetUp(mat_elemental));
1469: PetscCall(MatSetOption(mat_elemental, MAT_ROW_ORIENTED, PETSC_FALSE));
1470: }
1472: PetscCall(PetscMalloc2(m, &rows, N, &cols));
1473: for (i = 0; i < N; i++) cols[i] = i;
1474: for (i = 0; i < m; i++) rows[i] = rstart + i;
1476: /* PETSc-Elemental interface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
1477: PetscCall(MatDenseGetArray(A, &v));
1478: PetscCall(MatDenseGetLDA(a->A, &lda));
1479: if (lda == m) PetscCall(MatSetValues(mat_elemental, m, rows, N, cols, v, ADD_VALUES));
1480: else {
1481: for (i = 0; i < N; i++) PetscCall(MatSetValues(mat_elemental, m, rows, 1, &i, v + lda * i, ADD_VALUES));
1482: }
1483: PetscCall(MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY));
1484: PetscCall(MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY));
1485: PetscCall(MatDenseRestoreArray(A, &v));
1486: PetscCall(PetscFree2(rows, cols));
1488: if (reuse == MAT_INPLACE_MATRIX) {
1489: PetscCall(MatHeaderReplace(A, &mat_elemental));
1490: } else {
1491: *newmat = mat_elemental;
1492: }
1493: PetscFunctionReturn(PETSC_SUCCESS);
1494: }
1495: #endif
1497: static PetscErrorCode MatDenseGetColumn_MPIDense(Mat A, PetscInt col, PetscScalar **vals)
1498: {
1499: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
1501: PetscFunctionBegin;
1502: PetscCall(MatDenseGetColumn(mat->A, col, vals));
1503: PetscFunctionReturn(PETSC_SUCCESS);
1504: }
1506: static PetscErrorCode MatDenseRestoreColumn_MPIDense(Mat A, PetscScalar **vals)
1507: {
1508: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
1510: PetscFunctionBegin;
1511: PetscCall(MatDenseRestoreColumn(mat->A, vals));
1512: PetscFunctionReturn(PETSC_SUCCESS);
1513: }
1515: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIDense(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
1516: {
1517: Mat_MPIDense *mat;
1518: PetscInt m, nloc, N;
1520: PetscFunctionBegin;
1521: PetscCall(MatGetSize(inmat, &m, &N));
1522: PetscCall(MatGetLocalSize(inmat, NULL, &nloc));
1523: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
1524: PetscInt sum;
1526: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
1527: /* Check sum(n) = N */
1528: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
1529: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
1531: PetscCall(MatCreateDense(comm, m, n, PETSC_DETERMINE, N, NULL, outmat));
1532: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
1533: }
1535: /* numeric phase */
1536: mat = (Mat_MPIDense *)(*outmat)->data;
1537: PetscCall(MatCopy(inmat, mat->A, SAME_NONZERO_PATTERN));
1538: PetscFunctionReturn(PETSC_SUCCESS);
1539: }
1541: PetscErrorCode MatDenseGetColumnVec_MPIDense(Mat A, PetscInt col, Vec *v)
1542: {
1543: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1544: PetscInt lda;
1546: PetscFunctionBegin;
1547: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1548: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1549: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
1550: a->vecinuse = col + 1;
1551: PetscCall(MatDenseGetLDA(a->A, &lda));
1552: PetscCall(MatDenseGetArray(a->A, (PetscScalar **)&a->ptrinuse));
1553: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)lda)));
1554: *v = a->cvec;
1555: PetscFunctionReturn(PETSC_SUCCESS);
1556: }
1558: PetscErrorCode MatDenseRestoreColumnVec_MPIDense(Mat A, PetscInt col, Vec *v)
1559: {
1560: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1562: PetscFunctionBegin;
1563: PetscCheck(a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
1564: PetscCheck(a->cvec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column vector");
1565: VecCheckAssembled(a->cvec);
1566: a->vecinuse = 0;
1567: PetscCall(MatDenseRestoreArray(a->A, (PetscScalar **)&a->ptrinuse));
1568: PetscCall(VecResetArray(a->cvec));
1569: if (v) *v = NULL;
1570: PetscFunctionReturn(PETSC_SUCCESS);
1571: }
1573: PetscErrorCode MatDenseGetColumnVecRead_MPIDense(Mat A, PetscInt col, Vec *v)
1574: {
1575: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1576: PetscInt lda;
1578: PetscFunctionBegin;
1579: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1580: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1581: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
1582: a->vecinuse = col + 1;
1583: PetscCall(MatDenseGetLDA(a->A, &lda));
1584: PetscCall(MatDenseGetArrayRead(a->A, &a->ptrinuse));
1585: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)lda)));
1586: PetscCall(VecLockReadPush(a->cvec));
1587: *v = a->cvec;
1588: PetscFunctionReturn(PETSC_SUCCESS);
1589: }
1591: PetscErrorCode MatDenseRestoreColumnVecRead_MPIDense(Mat A, PetscInt col, Vec *v)
1592: {
1593: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1595: PetscFunctionBegin;
1596: PetscCheck(a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
1597: PetscCheck(a->cvec, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing internal column vector");
1598: VecCheckAssembled(a->cvec);
1599: a->vecinuse = 0;
1600: PetscCall(MatDenseRestoreArrayRead(a->A, &a->ptrinuse));
1601: PetscCall(VecLockReadPop(a->cvec));
1602: PetscCall(VecResetArray(a->cvec));
1603: if (v) *v = NULL;
1604: PetscFunctionReturn(PETSC_SUCCESS);
1605: }
1607: PetscErrorCode MatDenseGetColumnVecWrite_MPIDense(Mat A, PetscInt col, Vec *v)
1608: {
1609: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1610: PetscInt lda;
1612: PetscFunctionBegin;
1613: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1614: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1615: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
1616: a->vecinuse = col + 1;
1617: PetscCall(MatDenseGetLDA(a->A, &lda));
1618: PetscCall(MatDenseGetArrayWrite(a->A, (PetscScalar **)&a->ptrinuse));
1619: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)lda)));
1620: *v = a->cvec;
1621: PetscFunctionReturn(PETSC_SUCCESS);
1622: }
1624: PetscErrorCode MatDenseRestoreColumnVecWrite_MPIDense(Mat A, PetscInt col, Vec *v)
1625: {
1626: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1628: PetscFunctionBegin;
1629: PetscCheck(a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
1630: PetscCheck(a->cvec, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing internal column vector");
1631: VecCheckAssembled(a->cvec);
1632: a->vecinuse = 0;
1633: PetscCall(MatDenseRestoreArrayWrite(a->A, (PetscScalar **)&a->ptrinuse));
1634: PetscCall(VecResetArray(a->cvec));
1635: if (v) *v = NULL;
1636: PetscFunctionReturn(PETSC_SUCCESS);
1637: }
1639: static PetscErrorCode MatDenseGetSubMatrix_MPIDense(Mat A, PetscInt rbegin, PetscInt rend, PetscInt cbegin, PetscInt cend, Mat *v)
1640: {
1641: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1642: Mat_MPIDense *c;
1643: MPI_Comm comm;
1644: PetscInt prbegin, prend, pcbegin, pcend;
1646: PetscFunctionBegin;
1647: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
1648: PetscCheck(!a->vecinuse, comm, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1649: PetscCheck(!a->matinuse, comm, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1650: prbegin = PetscMax(0, PetscMin(A->rmap->rend, rbegin) - A->rmap->rstart);
1651: prend = PetscMin(A->rmap->n, PetscMax(0, rend - A->rmap->rstart));
1652: pcbegin = PetscMax(0, PetscMin(A->cmap->rend, cbegin) - A->cmap->rstart);
1653: pcend = PetscMin(A->cmap->n, PetscMax(0, cend - A->cmap->rstart));
1654: if (!a->cmat) {
1655: PetscCall(MatCreate(comm, &a->cmat));
1656: PetscCall(MatSetType(a->cmat, ((PetscObject)A)->type_name));
1657: if (rend - rbegin == A->rmap->N) PetscCall(PetscLayoutReference(A->rmap, &a->cmat->rmap));
1658: else {
1659: PetscCall(PetscLayoutSetLocalSize(a->cmat->rmap, prend - prbegin));
1660: PetscCall(PetscLayoutSetSize(a->cmat->rmap, rend - rbegin));
1661: PetscCall(PetscLayoutSetUp(a->cmat->rmap));
1662: }
1663: if (cend - cbegin == A->cmap->N) PetscCall(PetscLayoutReference(A->cmap, &a->cmat->cmap));
1664: else {
1665: PetscCall(PetscLayoutSetLocalSize(a->cmat->cmap, pcend - pcbegin));
1666: PetscCall(PetscLayoutSetSize(a->cmat->cmap, cend - cbegin));
1667: PetscCall(PetscLayoutSetUp(a->cmat->cmap));
1668: }
1669: c = (Mat_MPIDense *)a->cmat->data;
1670: c->sub_rbegin = rbegin;
1671: c->sub_rend = rend;
1672: c->sub_cbegin = cbegin;
1673: c->sub_cend = cend;
1674: }
1675: c = (Mat_MPIDense *)a->cmat->data;
1676: if (c->sub_rbegin != rbegin || c->sub_rend != rend) {
1677: PetscCall(PetscLayoutDestroy(&a->cmat->rmap));
1678: PetscCall(PetscLayoutCreate(comm, &a->cmat->rmap));
1679: PetscCall(PetscLayoutSetLocalSize(a->cmat->rmap, prend - prbegin));
1680: PetscCall(PetscLayoutSetSize(a->cmat->rmap, rend - rbegin));
1681: PetscCall(PetscLayoutSetUp(a->cmat->rmap));
1682: c->sub_rbegin = rbegin;
1683: c->sub_rend = rend;
1684: }
1685: if (c->sub_cbegin != cbegin || c->sub_cend != cend) {
1686: // special optimization: check if all columns are owned by rank 0, in which case no communication is necessary
1687: if ((cend - cbegin != a->cmat->cmap->N) || (A->cmap->range[1] != A->cmap->N)) {
1688: PetscCall(PetscLayoutDestroy(&a->cmat->cmap));
1689: PetscCall(PetscLayoutCreate(comm, &a->cmat->cmap));
1690: PetscCall(PetscLayoutSetLocalSize(a->cmat->cmap, pcend - pcbegin));
1691: PetscCall(PetscLayoutSetSize(a->cmat->cmap, cend - cbegin));
1692: PetscCall(PetscLayoutSetUp(a->cmat->cmap));
1693: PetscCall(VecDestroy(&c->lvec));
1694: PetscCall(PetscSFDestroy(&c->Mvctx));
1695: }
1696: c->sub_cbegin = cbegin;
1697: c->sub_cend = cend;
1698: }
1699: PetscCheck(!c->A, comm, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1700: PetscCall(MatDenseGetSubMatrix(a->A, prbegin, prend, cbegin, cend, &c->A));
1702: a->cmat->preallocated = PETSC_TRUE;
1703: a->cmat->assembled = PETSC_TRUE;
1704: #if PetscDefined(HAVE_DEVICE)
1705: a->cmat->offloadmask = c->A->offloadmask;
1706: #endif
1707: a->matinuse = cbegin + 1;
1708: *v = a->cmat;
1709: PetscFunctionReturn(PETSC_SUCCESS);
1710: }
1712: static PetscErrorCode MatDenseRestoreSubMatrix_MPIDense(Mat A, Mat *v)
1713: {
1714: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1715: Mat_MPIDense *c;
1717: PetscFunctionBegin;
1718: PetscCheck(a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseGetSubMatrix() first");
1719: PetscCheck(a->cmat, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing internal matrix");
1720: PetscCheck(*v == a->cmat, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Not the matrix obtained from MatDenseGetSubMatrix()");
1721: a->matinuse = 0;
1722: c = (Mat_MPIDense *)a->cmat->data;
1723: PetscCall(MatDenseRestoreSubMatrix(a->A, &c->A));
1724: *v = NULL;
1725: #if PetscDefined(HAVE_DEVICE)
1726: A->offloadmask = a->A->offloadmask;
1727: #endif
1728: PetscFunctionReturn(PETSC_SUCCESS);
1729: }
1731: /*MC
1732: MATMPIDENSE - MATMPIDENSE = "mpidense" - A matrix type to be used for distributed dense matrices.
1734: Options Database Key:
1735: . -mat_type mpidense - sets the matrix type to `MATMPIDENSE` during a call to `MatSetFromOptions()`
1737: Level: beginner
1739: .seealso: [](ch_matrices), `Mat`, `MatCreateDense()`, `MATSEQDENSE`, `MATDENSE`
1740: M*/
1741: PetscErrorCode MatCreate_MPIDense(Mat mat)
1742: {
1743: Mat_MPIDense *a;
1745: PetscFunctionBegin;
1746: PetscCall(PetscNew(&a));
1747: mat->data = (void *)a;
1748: mat->ops[0] = MatOps_Values;
1750: mat->insertmode = NOT_SET_VALUES;
1752: /* build cache for off array entries formed */
1753: a->donotstash = PETSC_FALSE;
1755: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)mat), 1, &mat->stash));
1757: /* stuff used for matrix vector multiply */
1758: a->lvec = NULL;
1759: a->Mvctx = NULL;
1760: a->roworiented = PETSC_TRUE;
1762: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetLDA_C", MatDenseGetLDA_MPIDense));
1763: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseSetLDA_C", MatDenseSetLDA_MPIDense));
1764: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArray_C", MatDenseGetArray_MPIDense));
1765: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArray_C", MatDenseRestoreArray_MPIDense));
1766: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayRead_C", MatDenseGetArrayRead_MPIDense));
1767: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayRead_C", MatDenseRestoreArrayRead_MPIDense));
1768: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayWrite_C", MatDenseGetArrayWrite_MPIDense));
1769: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayWrite_C", MatDenseRestoreArrayWrite_MPIDense));
1770: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDensePlaceArray_C", MatDensePlaceArray_MPIDense));
1771: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseResetArray_C", MatDenseResetArray_MPIDense));
1772: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseReplaceArray_C", MatDenseReplaceArray_MPIDense));
1773: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVec_C", MatDenseGetColumnVec_MPIDense));
1774: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVec_C", MatDenseRestoreColumnVec_MPIDense));
1775: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecRead_C", MatDenseGetColumnVecRead_MPIDense));
1776: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecRead_C", MatDenseRestoreColumnVecRead_MPIDense));
1777: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecWrite_C", MatDenseGetColumnVecWrite_MPIDense));
1778: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecWrite_C", MatDenseRestoreColumnVecWrite_MPIDense));
1779: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetSubMatrix_C", MatDenseGetSubMatrix_MPIDense));
1780: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreSubMatrix_C", MatDenseRestoreSubMatrix_MPIDense));
1781: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
1782: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpiaij_C", MatConvert_MPIDense_MPIAIJ));
1783: #if PetscDefined(HAVE_ELEMENTAL)
1784: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_elemental_C", MatConvert_MPIDense_Elemental));
1785: #endif
1786: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
1787: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_scalapack_C", MatConvert_Dense_ScaLAPACK));
1788: #endif
1789: #if PetscDefined(HAVE_CUDA)
1790: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensecuda_C", MatConvert_MPIDense_MPIDenseCUDA));
1791: #endif
1792: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIDenseSetPreallocation_C", MatMPIDenseSetPreallocation_MPIDense));
1793: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpidense_C", MatProductSetFromOptions_MPIAIJ_MPIDense));
1794: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaij_C", MatProductSetFromOptions_MPIDense_MPIAIJ));
1795: #if PetscDefined(HAVE_CUDA)
1796: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijcusparse_mpidense_C", MatProductSetFromOptions_MPIAIJ_MPIDense));
1797: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijcusparse_C", MatProductSetFromOptions_MPIDense_MPIAIJ));
1798: #endif
1799: #if PetscDefined(HAVE_HIP)
1800: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensehip_C", MatConvert_MPIDense_MPIDenseHIP));
1801: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijhipsparse_mpidense_C", MatProductSetFromOptions_MPIAIJ_MPIDense));
1802: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijhipsparse_C", MatProductSetFromOptions_MPIDense_MPIAIJ));
1803: #endif
1804: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumn_C", MatDenseGetColumn_MPIDense));
1805: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumn_C", MatDenseRestoreColumn_MPIDense));
1806: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultColumnRange_C", MatMultColumnRange_MPIDense));
1807: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultAddColumnRange_C", MatMultAddColumnRange_MPIDense));
1808: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeColumnRange_C", MatMultHermitianTransposeColumnRange_MPIDense));
1809: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeAddColumnRange_C", MatMultHermitianTransposeAddColumnRange_MPIDense));
1810: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatGetMultPetscSF_C", MatGetMultPetscSF_MPIDense));
1811: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, MATMPIDENSE));
1812: PetscFunctionReturn(PETSC_SUCCESS);
1813: }
1815: /*MC
1816: MATDENSE - MATDENSE = "dense" - A matrix type to be used for dense matrices.
1818: This matrix type is identical to `MATSEQDENSE` when constructed with a single process communicator,
1819: and `MATMPIDENSE` otherwise.
1821: Options Database Key:
1822: . -mat_type dense - sets the matrix type to `MATDENSE` during a call to `MatSetFromOptions()`
1824: Level: beginner
1826: .seealso: [](ch_matrices), `Mat`, `MATSEQDENSE`, `MATMPIDENSE`, `MATDENSECUDA`, `MATDENSEHIP`
1827: M*/
1829: /*@
1830: MatMPIDenseSetPreallocation - Sets the array used to store the matrix entries
1832: Collective
1834: Input Parameters:
1835: + B - the matrix
1836: - data - optional location of matrix data. Set to `NULL` for PETSc
1837: to control all matrix memory allocation.
1839: Level: intermediate
1841: Notes:
1842: The dense format is fully compatible with standard Fortran
1843: storage by columns.
1845: The data input variable is intended primarily for Fortran programmers
1846: who wish to allocate their own matrix memory space. Most users should
1847: set `data` to `NULL`.
1849: .seealso: [](ch_matrices), `Mat`, `MATMPIDENSE`, `MatCreate()`, `MatCreateSeqDense()`, `MatSetValues()`
1850: @*/
1851: PetscErrorCode MatMPIDenseSetPreallocation(Mat B, PetscScalar *data)
1852: {
1853: PetscFunctionBegin;
1855: PetscTryMethod(B, "MatMPIDenseSetPreallocation_C", (Mat, PetscScalar *), (B, data));
1856: PetscFunctionReturn(PETSC_SUCCESS);
1857: }
1859: /*@
1860: MatDensePlaceArray - Allows one to replace the array in a `MATDENSE` matrix with an
1861: array provided by the user. This is useful to avoid copying an array
1862: into a matrix
1864: Not Collective
1866: Input Parameters:
1867: + mat - the matrix
1868: - array - the array in column major order
1870: Level: developer
1872: Note:
1873: Adding `const` to `array` was an oversight, see notes in `VecPlaceArray()`.
1875: You can return to the original array with a call to `MatDenseResetArray()`. The user is responsible for freeing this array; it will not be
1876: freed when the matrix is destroyed.
1878: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArray()`, `MatDenseResetArray()`, `VecPlaceArray()`, `VecGetArray()`, `VecRestoreArray()`, `VecReplaceArray()`, `VecResetArray()`,
1879: `MatDenseReplaceArray()`
1880: @*/
1881: PetscErrorCode MatDensePlaceArray(Mat mat, const PetscScalar *array)
1882: {
1883: PetscFunctionBegin;
1885: PetscUseMethod(mat, "MatDensePlaceArray_C", (Mat, const PetscScalar *), (mat, array));
1886: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
1887: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
1888: mat->offloadmask = PETSC_OFFLOAD_CPU;
1889: #endif
1890: PetscFunctionReturn(PETSC_SUCCESS);
1891: }
1893: /*@
1894: MatDenseResetArray - Resets the matrix array to that it previously had before the call to `MatDensePlaceArray()`
1896: Not Collective
1898: Input Parameter:
1899: . mat - the matrix
1901: Level: developer
1903: Note:
1904: You can only call this after a call to `MatDensePlaceArray()`
1906: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArray()`, `MatDensePlaceArray()`, `VecPlaceArray()`, `VecGetArray()`, `VecRestoreArray()`, `VecReplaceArray()`, `VecResetArray()`
1907: @*/
1908: PetscErrorCode MatDenseResetArray(Mat mat)
1909: {
1910: PetscFunctionBegin;
1912: PetscUseMethod(mat, "MatDenseResetArray_C", (Mat), (mat));
1913: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
1914: PetscFunctionReturn(PETSC_SUCCESS);
1915: }
1917: /*@
1918: MatDenseReplaceArray - Allows one to replace the array in a dense matrix with an
1919: array provided by the user. This is useful to avoid copying an array
1920: into a matrix
1922: Not Collective
1924: Input Parameters:
1925: + mat - the matrix
1926: - array - the array in column major order
1928: Level: developer
1930: Note:
1931: Adding `const` to `array` was an oversight, see notes in `VecPlaceArray()`.
1933: The memory passed in MUST be obtained with `PetscMalloc()` and CANNOT be
1934: freed by the user. It will be freed when the matrix is destroyed.
1936: .seealso: [](ch_matrices), `Mat`, `MatDensePlaceArray()`, `MatDenseGetArray()`, `VecReplaceArray()`
1937: @*/
1938: PetscErrorCode MatDenseReplaceArray(Mat mat, const PetscScalar *array)
1939: {
1940: PetscFunctionBegin;
1942: PetscUseMethod(mat, "MatDenseReplaceArray_C", (Mat, const PetscScalar *), (mat, array));
1943: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
1944: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
1945: mat->offloadmask = PETSC_OFFLOAD_CPU;
1946: #endif
1947: PetscFunctionReturn(PETSC_SUCCESS);
1948: }
1950: /*@
1951: MatCreateDense - Creates a matrix in `MATDENSE` format.
1953: Collective
1955: Input Parameters:
1956: + comm - MPI communicator
1957: . m - number of local rows (or `PETSC_DECIDE` to have calculated if `M` is given)
1958: . n - number of local columns (or `PETSC_DECIDE` to have calculated if `N` is given)
1959: . M - number of global rows (or `PETSC_DECIDE` to have calculated if `m` is given)
1960: . N - number of global columns (or `PETSC_DECIDE` to have calculated if `n` is given)
1961: - data - optional location of matrix data. Set data to `NULL` (`PETSC_NULL_SCALAR_ARRAY` for Fortran users) for PETSc
1962: to control all matrix memory allocation.
1964: Output Parameter:
1965: . A - the matrix
1967: Level: intermediate
1969: Notes:
1970: The dense format is fully compatible with standard Fortran
1971: storage by columns.
1973: Although local portions of the matrix are stored in column-major
1974: order, the matrix is partitioned across MPI ranks by row.
1976: The data input variable is intended primarily for Fortran programmers
1977: who wish to allocate their own matrix memory space. Most users should
1978: set `data` to `NULL` (`PETSC_NULL_SCALAR_ARRAY` for Fortran users).
1980: The user MUST specify either the local or global matrix dimensions
1981: (possibly both).
1983: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatCreate()`, `MatCreateSeqDense()`, `MatSetValues()`
1984: @*/
1985: PetscErrorCode MatCreateDense(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscScalar data[], Mat *A)
1986: {
1987: PetscFunctionBegin;
1988: PetscCall(MatCreate(comm, A));
1989: PetscCall(MatSetSizes(*A, m, n, M, N));
1990: PetscCall(MatSetType(*A, MATDENSE));
1991: PetscCall(MatSeqDenseSetPreallocation(*A, data));
1992: PetscCall(MatMPIDenseSetPreallocation(*A, data));
1993: PetscFunctionReturn(PETSC_SUCCESS);
1994: }
1996: static PetscErrorCode MatDuplicate_MPIDense(Mat A, MatDuplicateOption cpvalues, Mat *newmat)
1997: {
1998: Mat mat;
1999: Mat_MPIDense *a, *oldmat = (Mat_MPIDense *)A->data;
2001: PetscFunctionBegin;
2002: *newmat = NULL;
2003: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &mat));
2004: PetscCall(MatSetSizes(mat, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
2005: PetscCall(MatSetType(mat, ((PetscObject)A)->type_name));
2006: a = (Mat_MPIDense *)mat->data;
2008: mat->factortype = A->factortype;
2009: mat->assembled = PETSC_TRUE;
2010: mat->preallocated = PETSC_TRUE;
2012: mat->insertmode = NOT_SET_VALUES;
2013: a->donotstash = oldmat->donotstash;
2015: PetscCall(PetscLayoutReference(A->rmap, &mat->rmap));
2016: PetscCall(PetscLayoutReference(A->cmap, &mat->cmap));
2018: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
2020: *newmat = mat;
2021: PetscFunctionReturn(PETSC_SUCCESS);
2022: }
2024: static PetscErrorCode MatLoad_MPIDense(Mat newMat, PetscViewer viewer)
2025: {
2026: PetscBool isbinary;
2027: #if PetscDefined(HAVE_HDF5)
2028: PetscBool ishdf5;
2029: #endif
2031: PetscFunctionBegin;
2034: /* force binary viewer to load .info file if it has not yet done so */
2035: PetscCall(PetscViewerSetUp(viewer));
2036: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
2037: #if PetscDefined(HAVE_HDF5)
2038: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
2039: #endif
2040: if (isbinary) {
2041: PetscCall(MatLoad_Dense_Binary(newMat, viewer));
2042: #if PetscDefined(HAVE_HDF5)
2043: } else if (ishdf5) {
2044: PetscCall(MatLoad_Dense_HDF5(newMat, viewer));
2045: #endif
2046: } else 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);
2047: PetscFunctionReturn(PETSC_SUCCESS);
2048: }
2050: static PetscErrorCode MatEqual_MPIDense(Mat A, Mat B, PetscBool *flag)
2051: {
2052: Mat_MPIDense *matB = (Mat_MPIDense *)B->data, *matA = (Mat_MPIDense *)A->data;
2053: Mat a, b;
2055: PetscFunctionBegin;
2056: a = matA->A;
2057: b = matB->A;
2058: PetscCall(MatEqual(a, b, flag));
2059: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, flag, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2060: PetscFunctionReturn(PETSC_SUCCESS);
2061: }
2063: static PetscErrorCode MatProductCtxDestroy_MatTransMatMult_MPIDense_MPIDense(PetscCtxRt data)
2064: {
2065: MatProductCtx_TransMatMultDense *atb = *(MatProductCtx_TransMatMultDense **)data;
2067: PetscFunctionBegin;
2068: PetscCall(PetscFree2(atb->sendbuf, atb->recvcounts));
2069: PetscCall(MatDestroy(&atb->atb));
2070: PetscCall(PetscFree(atb));
2071: PetscFunctionReturn(PETSC_SUCCESS);
2072: }
2074: static PetscErrorCode MatProductCtxDestroy_MatMatTransMult_MPIDense_MPIDense(PetscCtxRt data)
2075: {
2076: MatProductCtx_MatTransMultDense *abt = *(MatProductCtx_MatTransMultDense **)data;
2078: PetscFunctionBegin;
2079: PetscCall(PetscFree2(abt->buf[0], abt->buf[1]));
2080: PetscCall(PetscFree2(abt->recvcounts, abt->recvdispls));
2081: PetscCall(PetscFree(abt));
2082: PetscFunctionReturn(PETSC_SUCCESS);
2083: }
2085: static PetscErrorCode MatTransposeMatMultNumeric_MPIDense_MPIDense(Mat A, Mat B, Mat C)
2086: {
2087: Mat_MPIDense *a = (Mat_MPIDense *)A->data, *b = (Mat_MPIDense *)B->data, *c = (Mat_MPIDense *)C->data;
2088: MatProductCtx_TransMatMultDense *atb;
2089: MPI_Comm comm;
2090: PetscMPIInt size, *recvcounts;
2091: PetscScalar *carray, *sendbuf;
2092: const PetscScalar *atbarray;
2093: PetscInt i, cN = C->cmap->N, proc, k, j, lda;
2094: const PetscInt *ranges;
2096: PetscFunctionBegin;
2097: MatCheckProduct(C, 3);
2098: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2099: atb = (MatProductCtx_TransMatMultDense *)C->product->data;
2100: recvcounts = atb->recvcounts;
2101: sendbuf = atb->sendbuf;
2103: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2104: PetscCallMPI(MPI_Comm_size(comm, &size));
2106: /* compute atbarray = aseq^T * bseq */
2107: PetscCall(MatTransposeMatMult(a->A, b->A, atb->atb ? MAT_REUSE_MATRIX : MAT_INITIAL_MATRIX, PETSC_DETERMINE, &atb->atb));
2109: PetscCall(MatGetOwnershipRanges(C, &ranges));
2111: if (ranges[1] == C->rmap->N) {
2112: /* all of the values are being reduced to rank 0: optimize this case to use MPI_Reduce and GPU aware MPI if available */
2113: PetscInt atb_lda, c_lda;
2114: Mat atb_local = atb->atb;
2115: Mat atb_alloc = NULL;
2116: Mat c_local = c->A;
2117: Mat c_alloc = NULL;
2118: PetscMemType atb_memtype, c_memtype;
2119: const PetscScalar *atb_array = NULL;
2120: MPI_Datatype vector_type;
2121: PetscScalar *c_array = NULL;
2122: PetscMPIInt rank;
2124: PetscCallMPI(MPI_Comm_rank(comm, &rank));
2126: PetscCall(MatDenseGetLDA(atb_local, &atb_lda));
2127: if (atb_lda != C->rmap->N) {
2128: // copy atb to a matrix that will have lda == the number of rows
2129: PetscCall(MatDuplicate(atb_local, MAT_DO_NOT_COPY_VALUES, &atb_alloc));
2130: PetscCall(MatCopy(atb_local, atb_alloc, DIFFERENT_NONZERO_PATTERN));
2131: atb_local = atb_alloc;
2132: }
2134: if (rank == 0) {
2135: PetscCall(MatDenseGetLDA(c_local, &c_lda));
2136: if (c_lda != C->rmap->N) {
2137: // copy c to a matrix that will have lda == the number of rows
2138: PetscCall(MatDuplicate(c_local, MAT_DO_NOT_COPY_VALUES, &c_alloc));
2139: c_local = c_alloc;
2140: }
2141: PetscCall(MatZeroEntries(c_local));
2142: }
2143: /* atb_local and c_local have nrows = lda = A->cmap->N and ncols =
2144: * B->cmap->N: use the a->Mvctx to use the best reduction method */
2145: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
2146: vector_type = MPIU_SCALAR;
2147: if (B->cmap->N > 1) {
2148: PetscMPIInt mpi_N;
2150: PetscCall(PetscMPIIntCast(B->cmap->N, &mpi_N));
2151: PetscCallMPI(MPI_Type_contiguous(mpi_N, MPIU_SCALAR, &vector_type));
2152: PetscCallMPI(MPI_Type_commit(&vector_type));
2153: }
2154: PetscCall(MatDenseGetArrayReadAndMemType(atb_local, &atb_array, &atb_memtype));
2155: PetscCall(MatDenseGetArrayWriteAndMemType(c_local, &c_array, &c_memtype));
2156: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, vector_type, atb_memtype, atb_array, c_memtype, c_array, MPIU_SUM));
2157: PetscCall(PetscSFReduceEnd(a->Mvctx, vector_type, atb_array, c_array, MPIU_SUM));
2158: PetscCall(MatDenseRestoreArrayWriteAndMemType(c_local, &c_array));
2159: PetscCall(MatDenseRestoreArrayReadAndMemType(atb_local, &atb_array));
2160: if (rank == 0 && c_local != c->A) PetscCall(MatCopy(c_local, c->A, DIFFERENT_NONZERO_PATTERN));
2161: if (B->cmap->N > 1) PetscCallMPI(MPI_Type_free(&vector_type));
2162: PetscCall(MatDestroy(&atb_alloc));
2163: PetscCall(MatDestroy(&c_alloc));
2164: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2165: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2166: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2167: PetscFunctionReturn(PETSC_SUCCESS);
2168: }
2170: /* arrange atbarray into sendbuf */
2171: PetscCall(MatDenseGetArrayRead(atb->atb, &atbarray));
2172: PetscCall(MatDenseGetLDA(atb->atb, &lda));
2173: for (proc = 0, k = 0; proc < size; proc++) {
2174: for (j = 0; j < cN; j++) {
2175: for (i = ranges[proc]; i < ranges[proc + 1]; i++) sendbuf[k++] = atbarray[i + j * lda];
2176: }
2177: }
2178: PetscCall(MatDenseRestoreArrayRead(atb->atb, &atbarray));
2180: /* sum all atbarray to local values of C */
2181: PetscCall(MatDenseGetArrayWrite(c->A, &carray));
2182: PetscCallMPI(MPI_Reduce_scatter(sendbuf, carray, recvcounts, MPIU_SCALAR, MPIU_SUM, comm));
2183: PetscCall(MatDenseRestoreArrayWrite(c->A, &carray));
2184: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2185: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2186: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2187: PetscFunctionReturn(PETSC_SUCCESS);
2188: }
2190: static PetscErrorCode MatTransposeMatMultSymbolic_MPIDense_MPIDense(Mat A, Mat B, PetscReal fill, Mat C)
2191: {
2192: MPI_Comm comm;
2193: PetscMPIInt size;
2194: PetscInt cm = A->cmap->n, cM, cN = B->cmap->N;
2195: MatProductCtx_TransMatMultDense *atb;
2196: PetscBool cisdense = PETSC_FALSE;
2197: const PetscInt *ranges;
2199: PetscFunctionBegin;
2200: MatCheckProduct(C, 4);
2201: PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
2202: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2203: PetscCheck(A->rmap->rstart == B->rmap->rstart && A->rmap->rend == B->rmap->rend, comm, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, A (%" PetscInt_FMT ", %" PetscInt_FMT ") != B (%" PetscInt_FMT ",%" PetscInt_FMT ")", A->rmap->rstart,
2204: A->rmap->rend, B->rmap->rstart, B->rmap->rend);
2206: /* create matrix product C */
2207: PetscCall(MatSetSizes(C, cm, B->cmap->n, A->cmap->N, B->cmap->N));
2208: #if PetscDefined(HAVE_CUDA)
2209: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, ""));
2210: #endif
2211: #if PetscDefined(HAVE_HIP)
2212: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSEHIP, ""));
2213: #endif
2214: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2215: PetscCall(MatSetUp(C));
2217: /* create data structure for reuse C */
2218: PetscCallMPI(MPI_Comm_size(comm, &size));
2219: PetscCall(PetscNew(&atb));
2220: cM = C->rmap->N;
2221: PetscCall(PetscMalloc2(cM * cN, &atb->sendbuf, size, &atb->recvcounts));
2222: PetscCall(MatGetOwnershipRanges(C, &ranges));
2223: for (PetscMPIInt i = 0; i < size; i++) PetscCall(PetscMPIIntCast((ranges[i + 1] - ranges[i]) * cN, &atb->recvcounts[i]));
2224: C->product->data = atb;
2225: C->product->destroy = MatProductCtxDestroy_MatTransMatMult_MPIDense_MPIDense;
2226: PetscFunctionReturn(PETSC_SUCCESS);
2227: }
2229: static PetscErrorCode MatMatTransposeMultSymbolic_MPIDense_MPIDense(Mat A, Mat B, PetscReal fill, Mat C)
2230: {
2231: MPI_Comm comm;
2232: PetscMPIInt i, size;
2233: PetscInt maxRows, bufsiz;
2234: PetscMPIInt tag;
2235: PetscInt alg;
2236: MatProductCtx_MatTransMultDense *abt;
2237: Mat_Product *product = C->product;
2238: PetscBool flg;
2240: PetscFunctionBegin;
2241: MatCheckProduct(C, 4);
2242: PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
2243: /* check local size of A and B */
2244: PetscCheck(A->cmap->n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local column dimensions are incompatible, A (%" PetscInt_FMT ") != B (%" PetscInt_FMT ")", A->cmap->n, B->cmap->n);
2246: PetscCall(PetscStrcmp(product->alg, "allgatherv", &flg));
2247: alg = flg ? 0 : 1;
2249: /* setup matrix product C */
2250: PetscCall(MatSetSizes(C, A->rmap->n, B->rmap->n, A->rmap->N, B->rmap->N));
2251: PetscCall(MatSetType(C, MATMPIDENSE));
2252: PetscCall(MatSetUp(C));
2253: PetscCall(PetscObjectGetNewTag((PetscObject)C, &tag));
2255: /* create data structure for reuse C */
2256: PetscCall(PetscObjectGetComm((PetscObject)C, &comm));
2257: PetscCallMPI(MPI_Comm_size(comm, &size));
2258: PetscCall(PetscNew(&abt));
2259: abt->tag = tag;
2260: abt->alg = alg;
2261: switch (alg) {
2262: case 1: /* alg: "cyclic" */
2263: for (maxRows = 0, i = 0; i < size; i++) maxRows = PetscMax(maxRows, B->rmap->range[i + 1] - B->rmap->range[i]);
2264: bufsiz = A->cmap->N * maxRows;
2265: PetscCall(PetscMalloc2(bufsiz, &abt->buf[0], bufsiz, &abt->buf[1]));
2266: break;
2267: default: /* alg: "allgatherv" */
2268: PetscCall(PetscMalloc2(B->rmap->n * B->cmap->N, &abt->buf[0], B->rmap->N * B->cmap->N, &abt->buf[1]));
2269: PetscCall(PetscMalloc2(size, &abt->recvcounts, size + 1, &abt->recvdispls));
2270: for (i = 0; i <= size; i++) PetscCall(PetscMPIIntCast(B->rmap->range[i] * A->cmap->N, &abt->recvdispls[i]));
2271: for (i = 0; i < size; i++) PetscCall(PetscMPIIntCast(abt->recvdispls[i + 1] - abt->recvdispls[i], &abt->recvcounts[i]));
2272: break;
2273: }
2275: C->product->data = abt;
2276: C->product->destroy = MatProductCtxDestroy_MatMatTransMult_MPIDense_MPIDense;
2277: C->ops->mattransposemultnumeric = MatMatTransposeMultNumeric_MPIDense_MPIDense;
2278: PetscFunctionReturn(PETSC_SUCCESS);
2279: }
2281: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense_Cyclic(Mat A, Mat B, Mat C)
2282: {
2283: Mat_MPIDense *a = (Mat_MPIDense *)A->data, *b = (Mat_MPIDense *)B->data, *c = (Mat_MPIDense *)C->data;
2284: MatProductCtx_MatTransMultDense *abt;
2285: MPI_Comm comm;
2286: PetscMPIInt rank, size, sendto, recvfrom, recvisfrom;
2287: PetscScalar *sendbuf, *recvbuf = NULL, *cv;
2288: PetscInt i, cK = A->cmap->N, sendsiz, recvsiz, k, j, bn;
2289: PetscScalar _DOne = 1.0, _DZero = 0.0;
2290: const PetscScalar *av, *bv;
2291: PetscBLASInt cm, cn, ck, alda, blda = 0, clda;
2292: MPI_Request reqs[2];
2293: const PetscInt *ranges;
2295: PetscFunctionBegin;
2296: MatCheckProduct(C, 3);
2297: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2298: abt = (MatProductCtx_MatTransMultDense *)C->product->data;
2299: PetscCall(PetscObjectGetComm((PetscObject)C, &comm));
2300: PetscCallMPI(MPI_Comm_rank(comm, &rank));
2301: PetscCallMPI(MPI_Comm_size(comm, &size));
2302: PetscCall(MatDenseGetArrayRead(a->A, &av));
2303: PetscCall(MatDenseGetArrayRead(b->A, &bv));
2304: PetscCall(MatDenseGetArrayWrite(c->A, &cv));
2305: PetscCall(MatDenseGetLDA(a->A, &i));
2306: PetscCall(PetscBLASIntCast(i, &alda));
2307: PetscCall(MatDenseGetLDA(b->A, &i));
2308: PetscCall(PetscBLASIntCast(i, &blda));
2309: PetscCall(MatDenseGetLDA(c->A, &i));
2310: PetscCall(PetscBLASIntCast(i, &clda));
2311: PetscCall(MatGetOwnershipRanges(B, &ranges));
2312: bn = B->rmap->n;
2313: if (blda == bn) {
2314: sendbuf = (PetscScalar *)bv;
2315: } else {
2316: sendbuf = abt->buf[0];
2317: for (k = 0, i = 0; i < cK; i++) {
2318: for (j = 0; j < bn; j++, k++) sendbuf[k] = bv[i * blda + j];
2319: }
2320: }
2321: if (size > 1) {
2322: sendto = (rank + size - 1) % size;
2323: recvfrom = (rank + size + 1) % size;
2324: } else {
2325: sendto = recvfrom = 0;
2326: }
2327: PetscCall(PetscBLASIntCast(cK, &ck));
2328: PetscCall(PetscBLASIntCast(c->A->rmap->n, &cm));
2329: recvisfrom = rank;
2330: for (i = 0; i < size; i++) {
2331: /* we have finished receiving in sending, bufs can be read/modified */
2332: PetscMPIInt nextrecvisfrom = (recvisfrom + 1) % size; /* which process the next recvbuf will originate on */
2333: PetscInt nextbn = ranges[nextrecvisfrom + 1] - ranges[nextrecvisfrom];
2335: if (nextrecvisfrom != rank) {
2336: /* start the cyclic sends from sendbuf, to recvbuf (which will switch to sendbuf) */
2337: sendsiz = cK * bn;
2338: recvsiz = cK * nextbn;
2339: recvbuf = (i & 1) ? abt->buf[0] : abt->buf[1];
2340: PetscCallMPI(MPIU_Isend(sendbuf, sendsiz, MPIU_SCALAR, sendto, abt->tag, comm, &reqs[0]));
2341: PetscCallMPI(MPIU_Irecv(recvbuf, recvsiz, MPIU_SCALAR, recvfrom, abt->tag, comm, &reqs[1]));
2342: }
2344: /* local aseq * sendbuf^T */
2345: PetscCall(PetscBLASIntCast(ranges[recvisfrom + 1] - ranges[recvisfrom], &cn));
2346: if (cm && cn && ck) PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &cm, &cn, &ck, &_DOne, av, &alda, sendbuf, &cn, &_DZero, cv + clda * ranges[recvisfrom], &clda));
2348: if (nextrecvisfrom != rank) {
2349: /* wait for the sends and receives to complete, swap sendbuf and recvbuf */
2350: PetscCallMPI(MPI_Waitall(2, reqs, MPI_STATUSES_IGNORE));
2351: }
2352: bn = nextbn;
2353: recvisfrom = nextrecvisfrom;
2354: sendbuf = recvbuf;
2355: }
2356: PetscCall(MatDenseRestoreArrayRead(a->A, &av));
2357: PetscCall(MatDenseRestoreArrayRead(b->A, &bv));
2358: PetscCall(MatDenseRestoreArrayWrite(c->A, &cv));
2359: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2360: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2361: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2362: PetscFunctionReturn(PETSC_SUCCESS);
2363: }
2365: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense_Allgatherv(Mat A, Mat B, Mat C)
2366: {
2367: Mat_MPIDense *a = (Mat_MPIDense *)A->data, *b = (Mat_MPIDense *)B->data, *c = (Mat_MPIDense *)C->data;
2368: MatProductCtx_MatTransMultDense *abt;
2369: MPI_Comm comm;
2370: PetscMPIInt size, ibn;
2371: PetscScalar *cv, *sendbuf, *recvbuf;
2372: const PetscScalar *av, *bv;
2373: PetscInt blda, i, cK = A->cmap->N, k, j, bn;
2374: PetscScalar _DOne = 1.0, _DZero = 0.0;
2375: PetscBLASInt cm, cn, ck, alda, clda;
2377: PetscFunctionBegin;
2378: MatCheckProduct(C, 3);
2379: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2380: abt = (MatProductCtx_MatTransMultDense *)C->product->data;
2381: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2382: PetscCallMPI(MPI_Comm_size(comm, &size));
2383: PetscCall(MatDenseGetArrayRead(a->A, &av));
2384: PetscCall(MatDenseGetArrayRead(b->A, &bv));
2385: PetscCall(MatDenseGetArrayWrite(c->A, &cv));
2386: PetscCall(MatDenseGetLDA(a->A, &i));
2387: PetscCall(PetscBLASIntCast(i, &alda));
2388: PetscCall(MatDenseGetLDA(b->A, &blda));
2389: PetscCall(MatDenseGetLDA(c->A, &i));
2390: PetscCall(PetscBLASIntCast(i, &clda));
2391: /* copy transpose of B into buf[0] */
2392: bn = B->rmap->n;
2393: sendbuf = abt->buf[0];
2394: recvbuf = abt->buf[1];
2395: for (k = 0, j = 0; j < bn; j++) {
2396: for (i = 0; i < cK; i++, k++) sendbuf[k] = bv[i * blda + j];
2397: }
2398: PetscCall(MatDenseRestoreArrayRead(b->A, &bv));
2399: PetscCall(PetscMPIIntCast(bn * cK, &ibn));
2400: PetscCallMPI(MPI_Allgatherv(sendbuf, ibn, MPIU_SCALAR, recvbuf, abt->recvcounts, abt->recvdispls, MPIU_SCALAR, comm));
2401: PetscCall(PetscBLASIntCast(cK, &ck));
2402: PetscCall(PetscBLASIntCast(c->A->rmap->n, &cm));
2403: PetscCall(PetscBLASIntCast(c->A->cmap->n, &cn));
2404: if (cm && cn && ck) PetscCallBLAS("BLASgemm", BLASgemm_("N", "N", &cm, &cn, &ck, &_DOne, av, &alda, recvbuf, &ck, &_DZero, cv, &clda));
2405: PetscCall(MatDenseRestoreArrayRead(a->A, &av));
2406: PetscCall(MatDenseRestoreArrayRead(b->A, &bv));
2407: PetscCall(MatDenseRestoreArrayWrite(c->A, &cv));
2408: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2409: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2410: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2411: PetscFunctionReturn(PETSC_SUCCESS);
2412: }
2414: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense(Mat A, Mat B, Mat C)
2415: {
2416: MatProductCtx_MatTransMultDense *abt;
2418: PetscFunctionBegin;
2419: MatCheckProduct(C, 3);
2420: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2421: abt = (MatProductCtx_MatTransMultDense *)C->product->data;
2422: switch (abt->alg) {
2423: case 1:
2424: PetscCall(MatMatTransposeMultNumeric_MPIDense_MPIDense_Cyclic(A, B, C));
2425: break;
2426: default:
2427: PetscCall(MatMatTransposeMultNumeric_MPIDense_MPIDense_Allgatherv(A, B, C));
2428: break;
2429: }
2430: PetscFunctionReturn(PETSC_SUCCESS);
2431: }
2433: static PetscErrorCode MatProductCtxDestroy_MatMatMult_MPIDense_MPIDense(PetscCtxRt data)
2434: {
2435: MatProductCtx_MatMultDense *ab = *(MatProductCtx_MatMultDense **)data;
2437: PetscFunctionBegin;
2438: PetscCall(MatDestroy(&ab->Ce));
2439: PetscCall(MatDestroy(&ab->Ae));
2440: PetscCall(MatDestroy(&ab->Be));
2441: PetscCall(PetscFree(ab));
2442: PetscFunctionReturn(PETSC_SUCCESS);
2443: }
2445: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIDense(Mat A, Mat B, Mat C)
2446: {
2447: MatProductCtx_MatMultDense *ab;
2448: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
2449: Mat_MPIDense *b = (Mat_MPIDense *)B->data;
2451: PetscFunctionBegin;
2452: MatCheckProduct(C, 3);
2453: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Missing product data");
2454: ab = (MatProductCtx_MatMultDense *)C->product->data;
2455: if (ab->Ae && ab->Ce) {
2456: #if PetscDefined(HAVE_ELEMENTAL)
2457: PetscCall(MatConvert_MPIDense_Elemental(A, MATELEMENTAL, MAT_REUSE_MATRIX, &ab->Ae));
2458: PetscCall(MatConvert_MPIDense_Elemental(B, MATELEMENTAL, MAT_REUSE_MATRIX, &ab->Be));
2459: PetscCall(MatMatMultNumeric_Elemental(ab->Ae, ab->Be, ab->Ce));
2460: PetscCall(MatConvert(ab->Ce, MATMPIDENSE, MAT_REUSE_MATRIX, &C));
2461: #else
2462: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "PETSC_HAVE_ELEMENTAL not defined");
2463: #endif
2464: } else {
2465: MPI_Comm comm;
2466: const PetscScalar *read;
2467: PetscScalar *write;
2468: PetscInt lda;
2469: const PetscInt *ranges;
2470: PetscMPIInt size;
2472: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A)); /* cannot be done during the symbolic phase because of possible calls to MatProductReplaceMats() */
2473: comm = PetscObjectComm((PetscObject)B);
2474: PetscCallMPI(MPI_Comm_size(comm, &size));
2475: PetscCall(PetscLayoutGetRanges(B->rmap, &ranges));
2476: if (ranges[1] == ranges[size]) {
2477: // optimize for the case where the B matrix is broadcast from rank 0
2478: PetscInt b_lda, be_lda;
2479: Mat b_local = b->A;
2480: Mat b_alloc = NULL;
2481: Mat be_local = ab->Be;
2482: Mat be_alloc = NULL;
2483: PetscMemType b_memtype, be_memtype;
2484: const PetscScalar *b_array = NULL;
2485: MPI_Datatype vector_type;
2486: PetscScalar *be_array = NULL;
2487: PetscMPIInt rank;
2489: PetscCallMPI(MPI_Comm_rank(comm, &rank));
2490: PetscCall(MatDenseGetLDA(be_local, &be_lda));
2491: if (be_lda != B->rmap->N) {
2492: PetscCall(MatDuplicate(be_local, MAT_DO_NOT_COPY_VALUES, &be_alloc));
2493: be_local = be_alloc;
2494: }
2496: if (rank == 0) {
2497: PetscCall(MatDenseGetLDA(b_local, &b_lda));
2498: if (b_lda != B->rmap->N) {
2499: PetscCall(MatDuplicate(b_local, MAT_DO_NOT_COPY_VALUES, &b_alloc));
2500: PetscCall(MatCopy(b_local, b_alloc, DIFFERENT_NONZERO_PATTERN));
2501: b_local = b_alloc;
2502: }
2503: }
2504: vector_type = MPIU_SCALAR;
2505: if (B->cmap->N > 1) {
2506: PetscMPIInt mpi_N;
2508: PetscCall(PetscMPIIntCast(B->cmap->N, &mpi_N));
2509: PetscCallMPI(MPI_Type_contiguous(mpi_N, MPIU_SCALAR, &vector_type));
2510: PetscCallMPI(MPI_Type_commit(&vector_type));
2511: }
2512: PetscCall(MatDenseGetArrayReadAndMemType(b_local, &b_array, &b_memtype));
2513: PetscCall(MatDenseGetArrayWriteAndMemType(be_local, &be_array, &be_memtype));
2514: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, vector_type, b_memtype, b_array, be_memtype, be_array, MPI_REPLACE));
2515: PetscCall(PetscSFBcastEnd(mdn->Mvctx, vector_type, b_array, be_array, MPI_REPLACE));
2516: PetscCall(MatDenseRestoreArrayWriteAndMemType(be_local, &be_array));
2517: PetscCall(MatDenseRestoreArrayReadAndMemType(b_local, &b_array));
2518: if (be_local != ab->Be) PetscCall(MatCopy(be_local, ab->Be, DIFFERENT_NONZERO_PATTERN));
2519: if (B->cmap->N > 1) PetscCallMPI(MPI_Type_free(&vector_type));
2520: PetscCall(MatDestroy(&be_alloc));
2521: PetscCall(MatDestroy(&b_alloc));
2522: } else {
2523: PetscCall(MatDenseGetLDA(B, &lda));
2524: PetscCall(MatDenseGetArrayRead(B, &read));
2525: PetscCall(MatDenseGetArrayWrite(ab->Be, &write));
2526: for (PetscInt i = 0; i < C->cmap->N; ++i) {
2527: PetscCall(PetscSFBcastBegin(mdn->Mvctx, MPIU_SCALAR, read + i * lda, write + i * ab->Be->rmap->n, MPI_REPLACE));
2528: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, read + i * lda, write + i * ab->Be->rmap->n, MPI_REPLACE));
2529: }
2530: PetscCall(MatDenseRestoreArrayWrite(ab->Be, &write));
2531: PetscCall(MatDenseRestoreArrayRead(B, &read));
2532: }
2533: PetscCall(MatMatMultNumeric_SeqDense_SeqDense(((Mat_MPIDense *)A->data)->A, ab->Be, ((Mat_MPIDense *)C->data)->A));
2534: }
2535: PetscFunctionReturn(PETSC_SUCCESS);
2536: }
2538: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIDense(Mat A, Mat B, PetscReal fill, Mat C)
2539: {
2540: Mat_Product *product = C->product;
2541: PetscInt alg;
2542: MatProductCtx_MatMultDense *ab;
2543: PetscBool flg;
2545: PetscFunctionBegin;
2546: MatCheckProduct(C, 4);
2547: PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
2548: /* check local size of A and B */
2549: PetscCheck(A->cmap->rstart == B->rmap->rstart && A->cmap->rend == B->rmap->rend, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, A (%" PetscInt_FMT ", %" PetscInt_FMT ") != B (%" PetscInt_FMT ", %" PetscInt_FMT ")",
2550: A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend);
2552: PetscCall(PetscStrcmp(product->alg, "petsc", &flg));
2553: alg = flg ? 0 : 1;
2555: /* setup C */
2556: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
2557: PetscCall(MatSetType(C, MATMPIDENSE));
2558: PetscCall(MatSetUp(C));
2560: /* create data structure for reuse Cdense */
2561: PetscCall(PetscNew(&ab));
2563: switch (alg) {
2564: case 1: /* alg: "elemental" */
2565: #if PetscDefined(HAVE_ELEMENTAL)
2566: /* create elemental matrices Ae and Be */
2567: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &ab->Ae));
2568: PetscCall(MatSetSizes(ab->Ae, PETSC_DECIDE, PETSC_DECIDE, A->rmap->N, A->cmap->N));
2569: PetscCall(MatSetType(ab->Ae, MATELEMENTAL));
2570: PetscCall(MatSetUp(ab->Ae));
2571: PetscCall(MatSetOption(ab->Ae, MAT_ROW_ORIENTED, PETSC_FALSE));
2573: PetscCall(MatCreate(PetscObjectComm((PetscObject)B), &ab->Be));
2574: PetscCall(MatSetSizes(ab->Be, PETSC_DECIDE, PETSC_DECIDE, B->rmap->N, B->cmap->N));
2575: PetscCall(MatSetType(ab->Be, MATELEMENTAL));
2576: PetscCall(MatSetUp(ab->Be));
2577: PetscCall(MatSetOption(ab->Be, MAT_ROW_ORIENTED, PETSC_FALSE));
2579: /* compute symbolic Ce = Ae*Be */
2580: PetscCall(MatCreate(PetscObjectComm((PetscObject)C), &ab->Ce));
2581: PetscCall(MatMatMultSymbolic_Elemental(ab->Ae, ab->Be, fill, ab->Ce));
2582: #else
2583: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "PETSC_HAVE_ELEMENTAL not defined");
2584: #endif
2585: break;
2586: default: /* alg: "petsc" */
2587: ab->Ae = NULL;
2588: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, A->cmap->N, B->cmap->N, NULL, &ab->Be));
2589: ab->Ce = NULL;
2590: break;
2591: }
2593: C->product->data = ab;
2594: C->product->destroy = MatProductCtxDestroy_MatMatMult_MPIDense_MPIDense;
2595: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIDense;
2596: PetscFunctionReturn(PETSC_SUCCESS);
2597: }
2599: static PetscErrorCode MatProductSetFromOptions_MPIDense_AB(Mat C)
2600: {
2601: Mat_Product *product = C->product;
2602: const char *algTypes[2] = {"petsc", "elemental"};
2603: PetscInt alg, nalg = PetscDefined(HAVE_ELEMENTAL) ? 2 : 1;
2604: PetscBool flg = PETSC_FALSE;
2606: PetscFunctionBegin;
2607: /* Set default algorithm */
2608: alg = 0; /* default is PETSc */
2609: PetscCall(PetscStrcmp(product->alg, "default", &flg));
2610: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2612: /* Get runtime option */
2613: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
2614: PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatProduct_AB", algTypes, nalg, algTypes[alg], &alg, &flg));
2615: PetscOptionsEnd();
2616: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2618: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIDense;
2619: C->ops->productsymbolic = MatProductSymbolic_AB;
2620: PetscFunctionReturn(PETSC_SUCCESS);
2621: }
2623: static PetscErrorCode MatProductSetFromOptions_MPIDense_AtB(Mat C)
2624: {
2625: Mat_Product *product = C->product;
2626: Mat A = product->A, B = product->B;
2628: PetscFunctionBegin;
2629: PetscCheck(A->rmap->rstart == B->rmap->rstart && A->rmap->rend == B->rmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")",
2630: A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend);
2631: C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_MPIDense_MPIDense;
2632: C->ops->productsymbolic = MatProductSymbolic_AtB;
2633: PetscFunctionReturn(PETSC_SUCCESS);
2634: }
2636: static PetscErrorCode MatProductSetFromOptions_MPIDense_ABt(Mat C)
2637: {
2638: Mat_Product *product = C->product;
2639: const char *algTypes[2] = {"allgatherv", "cyclic"};
2640: PetscInt alg, nalg = 2;
2641: PetscBool flg = PETSC_FALSE;
2643: PetscFunctionBegin;
2644: /* Set default algorithm */
2645: alg = 0; /* default is allgatherv */
2646: PetscCall(PetscStrcmp(product->alg, "default", &flg));
2647: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2649: /* Get runtime option */
2650: if (product->api_user) {
2651: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatTransposeMult", "Mat");
2652: PetscCall(PetscOptionsEList("-matmattransmult_mpidense_mpidense_via", "Algorithmic approach", "MatMatTransposeMult", algTypes, nalg, algTypes[alg], &alg, &flg));
2653: PetscOptionsEnd();
2654: } else {
2655: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_ABt", "Mat");
2656: PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatProduct_ABt", algTypes, nalg, algTypes[alg], &alg, &flg));
2657: PetscOptionsEnd();
2658: }
2659: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2661: C->ops->mattransposemultsymbolic = MatMatTransposeMultSymbolic_MPIDense_MPIDense;
2662: C->ops->productsymbolic = MatProductSymbolic_ABt;
2663: PetscFunctionReturn(PETSC_SUCCESS);
2664: }
2666: static PetscErrorCode MatProductSetFromOptions_MPIDense(Mat C)
2667: {
2668: Mat_Product *product = C->product;
2670: PetscFunctionBegin;
2671: switch (product->type) {
2672: case MATPRODUCT_AB:
2673: PetscCall(MatProductSetFromOptions_MPIDense_AB(C));
2674: break;
2675: case MATPRODUCT_AtB:
2676: PetscCall(MatProductSetFromOptions_MPIDense_AtB(C));
2677: break;
2678: case MATPRODUCT_ABt:
2679: PetscCall(MatProductSetFromOptions_MPIDense_ABt(C));
2680: break;
2681: default:
2682: break;
2683: }
2684: PetscFunctionReturn(PETSC_SUCCESS);
2685: }
2687: PetscErrorCode MatDenseScatter_Private(PetscSF sf, Mat X, Mat Y, InsertMode mode, ScatterMode smode)
2688: {
2689: const PetscScalar *in;
2690: PetscScalar *out;
2691: PetscSF vsf;
2692: PetscInt N, ny, rld, lld;
2693: PetscMemType mtype[2];
2694: MPI_Op op = MPI_OP_NULL;
2696: PetscFunctionBegin;
2700: if (mode == INSERT_VALUES) op = MPI_REPLACE;
2701: else if (mode == ADD_VALUES) op = MPIU_SUM;
2702: else if (mode == MAX_VALUES) op = MPIU_MAX;
2703: else if (mode == MIN_VALUES) op = MPIU_MIN;
2704: PetscCheck(op != MPI_OP_NULL, PetscObjectComm((PetscObject)sf), PETSC_ERR_SUP, "Unsupported InsertMode %d in MatDenseScatter_Private()", mode);
2705: PetscCheck(smode == SCATTER_FORWARD || smode == SCATTER_REVERSE, PetscObjectComm((PetscObject)sf), PETSC_ERR_SUP, "Unsupported ScatterMode %d in MatDenseScatter_Private()", smode);
2706: PetscCall(MatGetSize(X, NULL, &N));
2707: PetscCall(MatGetSize(Y, NULL, &ny));
2708: PetscCheck(N == ny, PetscObjectComm((PetscObject)sf), PETSC_ERR_ARG_SIZ, "Matrix column sizes must match: %" PetscInt_FMT " != %" PetscInt_FMT, N, ny);
2709: PetscCall(MatDenseGetLDA(X, &rld));
2710: PetscCall(MatDenseGetLDA(Y, &lld));
2711: /* get cached or create new strided PetscSF when the number of columns is greater than one */
2712: if (N > 1) {
2713: PetscCall(PetscObjectQuery((PetscObject)sf, "_MatDenseScatter_StridedSF", (PetscObject *)&vsf));
2714: if (vsf) {
2715: PetscInt nr[2], nl[2];
2717: PetscCall(PetscSFGetGraph(sf, nr, nl, NULL, NULL));
2718: PetscCall(PetscSFGetGraph(vsf, nr + 1, nl + 1, NULL, NULL));
2719: if (N * nr[0] != nr[1] || N * nl[0] != nl[1]) vsf = NULL;
2720: }
2721: if (!vsf) {
2722: PetscCall(PetscSFCreateStridedSF(sf, N, rld, lld, &vsf));
2723: PetscCall(PetscObjectCompose((PetscObject)sf, "_MatDenseScatter_StridedSF", (PetscObject)vsf));
2724: PetscCall(PetscObjectDereference((PetscObject)vsf));
2725: }
2726: } else vsf = sf;
2727: /* the output array is accessed in read and write mode,
2728: but write-only in the INSERT_VALUES case could be worth exploring */
2729: PetscCall(MatDenseGetArrayReadAndMemType(X, &in, &mtype[0]));
2730: PetscCall(MatDenseGetArrayAndMemType(Y, &out, &mtype[1]));
2731: if (smode == SCATTER_FORWARD) {
2732: PetscCall(PetscSFBcastWithMemTypeBegin(vsf, vsf->vscat.unit, mtype[0], in, mtype[1], out, op));
2733: PetscCall(PetscSFBcastEnd(vsf, vsf->vscat.unit, in, out, op));
2734: } else {
2735: PetscCall(PetscSFReduceWithMemTypeBegin(vsf, vsf->vscat.unit, mtype[0], in, mtype[1], out, op));
2736: PetscCall(PetscSFReduceEnd(vsf, vsf->vscat.unit, in, out, op));
2737: }
2738: PetscCall(MatDenseRestoreArrayAndMemType(Y, &out));
2739: PetscCall(MatDenseRestoreArrayReadAndMemType(X, &in));
2740: PetscFunctionReturn(PETSC_SUCCESS);
2741: }