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: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
53: if (A->ops->copy != B->ops->copy) {
54: PetscCall(MatCopy_Basic(A, B, s));
55: PetscFunctionReturn(PETSC_SUCCESS);
56: }
57: PetscCall(MatCopy(Amat->A, Bmat->A, s));
58: PetscFunctionReturn(PETSC_SUCCESS);
59: }
61: PetscErrorCode MatShift_MPIDense(Mat A, PetscScalar alpha)
62: {
63: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
64: PetscInt j, lda, rstart = A->rmap->rstart, rend = A->rmap->rend, rend2;
65: PetscScalar *v;
67: PetscFunctionBegin;
68: PetscCall(MatDenseGetArray(mat->A, &v));
69: PetscCall(MatDenseGetLDA(mat->A, &lda));
70: rend2 = PetscMin(rend, A->cmap->N);
71: if (rend2 > rstart) {
72: for (j = rstart; j < rend2; j++) v[j - rstart + j * lda] += alpha;
73: PetscCall(PetscLogFlops(rend2 - rstart));
74: }
75: PetscCall(MatDenseRestoreArray(mat->A, &v));
76: PetscFunctionReturn(PETSC_SUCCESS);
77: }
79: static PetscErrorCode MatGetRow_MPIDense(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
80: {
81: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
82: PetscInt lrow, rstart = A->rmap->rstart, rend = A->rmap->rend;
84: PetscFunctionBegin;
85: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "only local rows");
86: lrow = row - rstart;
87: PetscCall(MatGetRow(mat->A, lrow, nz, (const PetscInt **)idx, (const PetscScalar **)v));
88: PetscFunctionReturn(PETSC_SUCCESS);
89: }
91: static PetscErrorCode MatRestoreRow_MPIDense(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
92: {
93: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
94: PetscInt lrow, rstart = A->rmap->rstart, rend = A->rmap->rend;
96: PetscFunctionBegin;
97: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "only local rows");
98: lrow = row - rstart;
99: PetscCall(MatRestoreRow(mat->A, lrow, nz, (const PetscInt **)idx, (const PetscScalar **)v));
100: PetscFunctionReturn(PETSC_SUCCESS);
101: }
103: static PetscErrorCode MatGetDiagonalBlock_MPIDense(Mat A, Mat *a)
104: {
105: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
106: PetscInt m = A->rmap->n, rstart = A->rmap->rstart;
107: PetscScalar *array;
108: MPI_Comm comm;
109: PetscBool flg;
110: Mat B;
112: PetscFunctionBegin;
113: PetscCall(MatHasCongruentLayouts(A, &flg));
114: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only square matrices supported.");
115: PetscCall(PetscObjectQuery((PetscObject)A, "DiagonalBlock", (PetscObject *)&B));
116: if (!B) { /* This should use MatDenseGetSubMatrix (not create), but we would need a call like MatRestoreDiagonalBlock */
117: #if PetscDefined(HAVE_CUDA)
118: PetscCall(PetscObjectTypeCompare((PetscObject)mdn->A, MATSEQDENSECUDA, &flg));
119: 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);
120: #elif PetscDefined(HAVE_HIP)
121: PetscCall(PetscObjectTypeCompare((PetscObject)mdn->A, MATSEQDENSEHIP, &flg));
122: 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);
123: #endif
124: PetscCall(PetscObjectGetComm((PetscObject)mdn->A, &comm));
125: PetscCall(MatCreate(comm, &B));
126: PetscCall(MatSetSizes(B, m, m, m, m));
127: PetscCall(MatSetType(B, ((PetscObject)mdn->A)->type_name));
128: PetscCall(MatDenseGetArrayRead(mdn->A, (const PetscScalar **)&array));
129: PetscCall(MatSeqDenseSetPreallocation(B, PetscSafePointerPlusOffset(array, m * rstart)));
130: PetscCall(MatDenseRestoreArrayRead(mdn->A, (const PetscScalar **)&array));
131: PetscCall(PetscObjectCompose((PetscObject)A, "DiagonalBlock", (PetscObject)B));
132: *a = B;
133: PetscCall(MatDestroy(&B));
134: } else *a = B;
135: PetscFunctionReturn(PETSC_SUCCESS);
136: }
138: static PetscErrorCode MatSetValues_MPIDense(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], const PetscScalar v[], InsertMode addv)
139: {
140: Mat_MPIDense *A = (Mat_MPIDense *)mat->data;
141: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend, row;
142: PetscBool roworiented = A->roworiented;
144: PetscFunctionBegin;
145: for (i = 0; i < m; i++) {
146: if (idxm[i] < 0) continue;
147: PetscCheck(idxm[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large");
148: if (idxm[i] >= rstart && idxm[i] < rend) {
149: row = idxm[i] - rstart;
150: if (roworiented) {
151: PetscCall(MatSetValues(A->A, 1, &row, n, idxn, PetscSafePointerPlusOffset(v, i * n), addv));
152: } else {
153: for (j = 0; j < n; j++) {
154: if (idxn[j] < 0) continue;
155: PetscCheck(idxn[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large");
156: PetscCall(MatSetValues(A->A, 1, &row, 1, &idxn[j], PetscSafePointerPlusOffset(v, i + j * m), addv));
157: }
158: }
159: } else if (!A->donotstash) {
160: mat->assembled = PETSC_FALSE;
161: if (roworiented) {
162: PetscCall(MatStashValuesRow_Private(&mat->stash, idxm[i], n, idxn, PetscSafePointerPlusOffset(v, i * n), PETSC_FALSE));
163: } else {
164: PetscCall(MatStashValuesCol_Private(&mat->stash, idxm[i], n, idxn, PetscSafePointerPlusOffset(v, i), m, PETSC_FALSE));
165: }
166: }
167: }
168: PetscFunctionReturn(PETSC_SUCCESS);
169: }
171: static PetscErrorCode MatGetValues_MPIDense(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
172: {
173: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
174: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend, row;
175: PetscBool roworiented = mdn->roworiented;
176: PetscScalar *value;
178: PetscFunctionBegin;
179: for (i = 0; i < m; i++) {
180: if (idxm[i] < 0) continue; /* negative row */
181: PetscCheck(idxm[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large");
182: PetscCheck(idxm[i] >= rstart && idxm[i] < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only local values currently supported");
183: row = idxm[i] - rstart;
184: for (j = 0; j < n; j++) {
185: if (idxn[j] < 0) continue; /* negative column */
186: PetscCheck(idxn[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large");
187: value = roworiented ? &v[j + i * n] : &v[i + j * m];
188: PetscCall(MatGetValues(mdn->A, 1, &row, 1, &idxn[j], value));
189: }
190: }
191: PetscFunctionReturn(PETSC_SUCCESS);
192: }
194: static PetscErrorCode MatDenseGetLDA_MPIDense(Mat A, PetscInt *lda)
195: {
196: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
198: PetscFunctionBegin;
199: PetscCall(MatDenseGetLDA(a->A, lda));
200: PetscFunctionReturn(PETSC_SUCCESS);
201: }
203: static PetscErrorCode MatDenseSetLDA_MPIDense(Mat A, PetscInt lda)
204: {
205: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
206: MatType mtype = MATSEQDENSE;
208: PetscFunctionBegin;
209: if (!a->A) {
210: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
211: PetscCall(PetscLayoutSetUp(A->rmap));
212: PetscCall(PetscLayoutSetUp(A->cmap));
213: PetscCall(MatCreate(PETSC_COMM_SELF, &a->A));
214: PetscCall(MatSetSizes(a->A, A->rmap->n, A->cmap->N, A->rmap->n, A->cmap->N));
215: #if PetscDefined(HAVE_CUDA)
216: PetscBool iscuda;
217: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIDENSECUDA, &iscuda));
218: if (iscuda) mtype = MATSEQDENSECUDA;
219: #elif PetscDefined(HAVE_HIP)
220: PetscBool iship;
221: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIDENSEHIP, &iship));
222: if (iship) mtype = MATSEQDENSEHIP;
223: #endif
224: PetscCall(MatSetType(a->A, mtype));
225: }
226: PetscCall(MatDenseSetLDA(a->A, lda));
227: PetscFunctionReturn(PETSC_SUCCESS);
228: }
230: static PetscErrorCode MatDenseGetArray_MPIDense(Mat A, PetscScalar **array)
231: {
232: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
234: PetscFunctionBegin;
235: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
236: PetscCall(MatDenseGetArray(a->A, array));
237: PetscFunctionReturn(PETSC_SUCCESS);
238: }
240: static PetscErrorCode MatDenseGetArrayRead_MPIDense(Mat A, PetscScalar **array)
241: {
242: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
244: PetscFunctionBegin;
245: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
246: PetscCall(MatDenseGetArrayRead(a->A, (const PetscScalar **)array));
247: PetscFunctionReturn(PETSC_SUCCESS);
248: }
250: static PetscErrorCode MatDenseGetArrayWrite_MPIDense(Mat A, PetscScalar **array)
251: {
252: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
254: PetscFunctionBegin;
255: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
256: PetscCall(MatDenseGetArrayWrite(a->A, array));
257: PetscFunctionReturn(PETSC_SUCCESS);
258: }
260: static PetscErrorCode MatDensePlaceArray_MPIDense(Mat A, const PetscScalar *array)
261: {
262: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
264: PetscFunctionBegin;
265: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
266: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
267: PetscCall(MatDensePlaceArray(a->A, array));
268: PetscFunctionReturn(PETSC_SUCCESS);
269: }
271: static PetscErrorCode MatDenseResetArray_MPIDense(Mat A)
272: {
273: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
275: PetscFunctionBegin;
276: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
277: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
278: PetscCall(MatDenseResetArray(a->A));
279: PetscFunctionReturn(PETSC_SUCCESS);
280: }
282: static PetscErrorCode MatDenseReplaceArray_MPIDense(Mat A, const PetscScalar *array)
283: {
284: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
286: PetscFunctionBegin;
287: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
288: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
289: PetscCall(MatDenseReplaceArray(a->A, array));
290: PetscFunctionReturn(PETSC_SUCCESS);
291: }
293: static PetscErrorCode MatCreateSubMatrix_MPIDense(Mat A, IS isrow, IS iscol, MatReuse scall, Mat *B)
294: {
295: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
296: PetscInt lda, i, j, rstart, rend, nrows, ncols, Ncols, nlrows, nlcols;
297: const PetscInt *irow, *icol;
298: const PetscScalar *v;
299: PetscScalar *bv;
300: Mat newmat;
301: IS iscol_local;
302: MPI_Comm comm_is, comm_mat;
304: PetscFunctionBegin;
305: PetscCall(PetscObjectGetComm((PetscObject)A, &comm_mat));
306: PetscCall(PetscObjectGetComm((PetscObject)iscol, &comm_is));
307: PetscCheck(comm_mat == comm_is, PETSC_COMM_SELF, PETSC_ERR_ARG_NOTSAMECOMM, "IS communicator must match matrix communicator");
309: PetscCall(ISAllGather(iscol, &iscol_local));
310: PetscCall(ISGetIndices(isrow, &irow));
311: PetscCall(ISGetIndices(iscol_local, &icol));
312: PetscCall(ISGetLocalSize(isrow, &nrows));
313: PetscCall(ISGetLocalSize(iscol, &ncols));
314: PetscCall(ISGetSize(iscol, &Ncols)); /* global number of columns, size of iscol_local */
316: /* No parallel redistribution currently supported! Should really check each index set
317: to confirm that it is OK. ... Currently supports only submatrix same partitioning as
318: original matrix! */
320: PetscCall(MatGetLocalSize(A, &nlrows, &nlcols));
321: PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
323: /* Check submatrix call */
324: if (scall == MAT_REUSE_MATRIX) {
325: /* SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size"); */
326: /* Really need to test rows and column sizes! */
327: newmat = *B;
328: } else {
329: /* Create and fill new matrix */
330: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &newmat));
331: PetscCall(MatSetSizes(newmat, nrows, ncols, PETSC_DECIDE, Ncols));
332: PetscCall(MatSetType(newmat, ((PetscObject)A)->type_name));
333: PetscCall(MatMPIDenseSetPreallocation(newmat, NULL));
334: }
336: /* Now extract the data pointers and do the copy, column at a time */
337: PetscCall(MatDenseGetArray(newmat, &bv));
338: PetscCall(MatDenseGetArrayRead(mat->A, &v));
339: PetscCall(MatDenseGetLDA(mat->A, &lda));
340: for (i = 0; i < Ncols; i++) {
341: const PetscScalar *av = v + lda * icol[i];
342: for (j = 0; j < nrows; j++) *bv++ = av[irow[j] - rstart];
343: }
344: PetscCall(MatDenseRestoreArrayRead(mat->A, &v));
345: PetscCall(MatDenseRestoreArray(newmat, &bv));
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 MatSetInf_MPIDense(Mat A)
440: {
441: Mat_MPIDense *l = (Mat_MPIDense *)A->data;
443: PetscFunctionBegin;
444: PetscCall(MatFlag(l->A, 1));
445: PetscFunctionReturn(PETSC_SUCCESS);
446: }
448: static PetscErrorCode MatZeroRows_MPIDense(Mat A, PetscInt n, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
449: {
450: Mat_MPIDense *l = (Mat_MPIDense *)A->data;
451: PetscInt i, len, *lrows;
453: PetscFunctionBegin;
454: /* get locally owned rows */
455: PetscCall(PetscLayoutMapLocal(A->rmap, n, rows, &len, &lrows, NULL));
456: /* fix right-hand side if needed */
457: if (x && b) {
458: const PetscScalar *xx;
459: PetscScalar *bb;
461: PetscCall(VecGetArrayRead(x, &xx));
462: PetscCall(VecGetArrayWrite(b, &bb));
463: for (i = 0; i < len; ++i) bb[lrows[i]] = diag * xx[lrows[i]];
464: PetscCall(VecRestoreArrayRead(x, &xx));
465: PetscCall(VecRestoreArrayWrite(b, &bb));
466: }
467: PetscCall(MatZeroRows(l->A, len, lrows, 0.0, NULL, NULL));
468: if (diag != 0.0) {
469: Vec d;
471: PetscCall(MatCreateVecs(A, NULL, &d));
472: PetscCall(VecSet(d, diag));
473: PetscCall(MatDiagonalSet(A, d, INSERT_VALUES));
474: PetscCall(VecDestroy(&d));
475: }
476: PetscCall(PetscFree(lrows));
477: PetscFunctionReturn(PETSC_SUCCESS);
478: }
480: PETSC_INTERN PetscErrorCode MatMult_SeqDense(Mat, Vec, Vec);
481: PETSC_INTERN PetscErrorCode MatMultAdd_SeqDense(Mat, Vec, Vec, Vec);
482: PETSC_INTERN PetscErrorCode MatMultTranspose_SeqDense(Mat, Vec, Vec);
483: PETSC_INTERN PetscErrorCode MatMultTransposeAdd_SeqDense(Mat, Vec, Vec, Vec);
485: static PetscErrorCode MatMultColumnRange_MPIDense(Mat mat, Vec xx, Vec yy, PetscInt c_start, PetscInt c_end)
486: {
487: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
488: const PetscScalar *ax;
489: PetscScalar *ay;
490: PetscMemType axmtype, aymtype;
492: PetscFunctionBegin;
493: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(mat));
494: PetscCall(VecGetArrayReadAndMemType(xx, &ax, &axmtype));
495: PetscCall(VecGetArrayWriteAndMemType(mdn->lvec, &ay, &aymtype));
496: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPI_REPLACE));
497: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, ax, ay, MPI_REPLACE));
498: PetscCall(VecRestoreArrayWriteAndMemType(mdn->lvec, &ay));
499: PetscCall(VecRestoreArrayReadAndMemType(xx, &ax));
500: PetscUseMethod(mdn->A, "MatMultColumnRange_C", (Mat, Vec, Vec, PetscInt, PetscInt), (mdn->A, mdn->lvec, yy, c_start, c_end));
501: PetscFunctionReturn(PETSC_SUCCESS);
502: }
504: static PetscErrorCode MatMult_MPIDense(Mat mat, Vec xx, Vec yy)
505: {
506: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
507: const PetscScalar *ax;
508: PetscScalar *ay;
509: PetscMemType axmtype, aymtype;
511: PetscFunctionBegin;
512: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(mat));
513: PetscCall(VecGetArrayReadAndMemType(xx, &ax, &axmtype));
514: PetscCall(VecGetArrayWriteAndMemType(mdn->lvec, &ay, &aymtype));
515: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPI_REPLACE));
516: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, ax, ay, MPI_REPLACE));
517: PetscCall(VecRestoreArrayWriteAndMemType(mdn->lvec, &ay));
518: PetscCall(VecRestoreArrayReadAndMemType(xx, &ax));
519: PetscUseTypeMethod(mdn->A, mult, mdn->lvec, yy);
520: PetscFunctionReturn(PETSC_SUCCESS);
521: }
523: static PetscErrorCode MatGetMultPetscSF_MPIDense(Mat A, PetscSF *sf)
524: {
525: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
527: PetscFunctionBegin;
528: *sf = mdn->Mvctx;
529: PetscFunctionReturn(PETSC_SUCCESS);
530: }
532: static PetscErrorCode MatMultAddColumnRange_MPIDense(Mat mat, Vec xx, Vec yy, Vec zz, PetscInt c_start, PetscInt c_end)
533: {
534: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
535: const PetscScalar *ax;
536: PetscScalar *ay;
537: PetscMemType axmtype, aymtype;
539: PetscFunctionBegin;
540: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(mat));
541: PetscCall(VecGetArrayReadAndMemType(xx, &ax, &axmtype));
542: PetscCall(VecGetArrayAndMemType(mdn->lvec, &ay, &aymtype));
543: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPI_REPLACE));
544: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, ax, ay, MPI_REPLACE));
545: PetscCall(VecRestoreArrayAndMemType(mdn->lvec, &ay));
546: PetscCall(VecRestoreArrayReadAndMemType(xx, &ax));
547: PetscUseMethod(mdn->A, "MatMultAddColumnRange_C", (Mat, Vec, Vec, Vec, PetscInt, PetscInt), (mdn->A, mdn->lvec, yy, zz, c_start, c_end));
548: PetscFunctionReturn(PETSC_SUCCESS);
549: }
551: static PetscErrorCode MatMultAdd_MPIDense(Mat mat, Vec xx, Vec yy, Vec zz)
552: {
553: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
554: const PetscScalar *ax;
555: PetscScalar *ay;
556: PetscMemType axmtype, aymtype;
558: PetscFunctionBegin;
559: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(mat));
560: PetscCall(VecGetArrayReadAndMemType(xx, &ax, &axmtype));
561: PetscCall(VecGetArrayAndMemType(mdn->lvec, &ay, &aymtype));
562: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPI_REPLACE));
563: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, ax, ay, MPI_REPLACE));
564: PetscCall(VecRestoreArrayAndMemType(mdn->lvec, &ay));
565: PetscCall(VecRestoreArrayReadAndMemType(xx, &ax));
566: PetscUseTypeMethod(mdn->A, multadd, mdn->lvec, yy, zz);
567: PetscFunctionReturn(PETSC_SUCCESS);
568: }
570: static PetscErrorCode MatMultHermitianTransposeColumnRange_MPIDense(Mat A, Vec xx, Vec yy, PetscInt c_start, PetscInt c_end)
571: {
572: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
573: const PetscScalar *ax;
574: PetscScalar *ay;
575: PetscMemType axmtype, aymtype;
576: PetscInt r_start, r_end;
577: PetscInt c_start_local, c_end_local;
579: PetscFunctionBegin;
580: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
581: PetscCall(VecZeroEntries(a->lvec));
582: PetscCall(VecGetOwnershipRange(yy, &r_start, &r_end));
583: c_start_local = PetscMax(c_start, r_start);
584: c_end_local = PetscMin(c_end, r_end);
585: PetscCall(VecGetArrayAndMemType(yy, &ay, &aymtype));
586: if (c_end_local > c_start_local) {
587: if (PetscMemTypeHost(aymtype)) {
588: PetscCall(PetscArrayzero(&ay[c_start_local], (size_t)(c_end_local - c_start_local)));
589: } else {
590: PetscCall(PetscDeviceRegisterMemory(ay, aymtype, sizeof(*ay) * ((size_t)(r_end - r_start))));
591: PetscCall(PetscDeviceArrayZero(NULL, &ay[c_start_local], (size_t)(c_end_local - c_start_local)));
592: }
593: }
594: PetscUseMethod(a->A, "MatMultHermitianTransposeColumnRange_C", (Mat, Vec, Vec, PetscInt, PetscInt), (a->A, xx, a->lvec, c_start, c_end));
595: PetscCall(VecGetArrayReadAndMemType(a->lvec, &ax, &axmtype));
596: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPIU_SUM));
597: PetscCall(PetscSFReduceEnd(a->Mvctx, MPIU_SCALAR, ax, ay, MPIU_SUM));
598: PetscCall(VecRestoreArrayReadAndMemType(a->lvec, &ax));
599: PetscCall(VecRestoreArrayAndMemType(yy, &ay));
600: PetscFunctionReturn(PETSC_SUCCESS);
601: }
603: static PetscErrorCode MatMultTransposeKernel_MPIDense(Mat A, Vec xx, Vec yy, PetscBool herm)
604: {
605: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
606: const PetscScalar *ax;
607: PetscScalar *ay;
608: PetscMemType axmtype, aymtype;
610: PetscFunctionBegin;
611: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
612: PetscCall(VecSet(yy, 0.0));
613: if (herm) PetscUseTypeMethod(a->A, multhermitiantranspose, xx, a->lvec);
614: else PetscUseTypeMethod(a->A, multtranspose, xx, a->lvec);
615: PetscCall(VecGetArrayReadAndMemType(a->lvec, &ax, &axmtype));
616: PetscCall(VecGetArrayAndMemType(yy, &ay, &aymtype));
617: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPIU_SUM));
618: PetscCall(PetscSFReduceEnd(a->Mvctx, MPIU_SCALAR, ax, ay, MPIU_SUM));
619: PetscCall(VecRestoreArrayReadAndMemType(a->lvec, &ax));
620: PetscCall(VecRestoreArrayAndMemType(yy, &ay));
621: PetscFunctionReturn(PETSC_SUCCESS);
622: }
624: static PetscErrorCode MatMultHermitianTransposeAddColumnRange_MPIDense(Mat A, Vec xx, Vec yy, Vec zz, PetscInt c_start, PetscInt c_end)
625: {
626: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
627: const PetscScalar *ax;
628: PetscScalar *ay;
629: PetscMemType axmtype, aymtype;
631: PetscFunctionBegin;
632: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
633: PetscCall(VecCopy(yy, zz));
634: PetscCall(VecZeroEntries(a->lvec));
635: PetscUseMethod(a->A, "MatMultHermitianTransposeColumnRange_C", (Mat, Vec, Vec, PetscInt, PetscInt), (a->A, xx, a->lvec, c_start, c_end));
636: PetscCall(VecGetArrayReadAndMemType(a->lvec, &ax, &axmtype));
637: PetscCall(VecGetArrayAndMemType(zz, &ay, &aymtype));
638: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPIU_SUM));
639: PetscCall(PetscSFReduceEnd(a->Mvctx, MPIU_SCALAR, ax, ay, MPIU_SUM));
640: PetscCall(VecRestoreArrayReadAndMemType(a->lvec, &ax));
641: PetscCall(VecRestoreArrayAndMemType(zz, &ay));
642: PetscFunctionReturn(PETSC_SUCCESS);
643: }
645: static PetscErrorCode MatMultTransposeAddKernel_MPIDense(Mat A, Vec xx, Vec yy, Vec zz, PetscBool herm)
646: {
647: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
648: const PetscScalar *ax;
649: PetscScalar *ay;
650: PetscMemType axmtype, aymtype;
652: PetscFunctionBegin;
653: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
654: PetscCall(VecCopy(yy, zz));
655: if (herm) PetscUseTypeMethod(a->A, multhermitiantranspose, xx, a->lvec);
656: else PetscUseTypeMethod(a->A, multtranspose, xx, a->lvec);
657: PetscCall(VecGetArrayReadAndMemType(a->lvec, &ax, &axmtype));
658: PetscCall(VecGetArrayAndMemType(zz, &ay, &aymtype));
659: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, MPIU_SCALAR, axmtype, ax, aymtype, ay, MPIU_SUM));
660: PetscCall(PetscSFReduceEnd(a->Mvctx, MPIU_SCALAR, ax, ay, MPIU_SUM));
661: PetscCall(VecRestoreArrayReadAndMemType(a->lvec, &ax));
662: PetscCall(VecRestoreArrayAndMemType(zz, &ay));
663: PetscFunctionReturn(PETSC_SUCCESS);
664: }
666: static PetscErrorCode MatMultTranspose_MPIDense(Mat A, Vec xx, Vec yy)
667: {
668: PetscFunctionBegin;
669: PetscCall(MatMultTransposeKernel_MPIDense(A, xx, yy, PETSC_FALSE));
670: PetscFunctionReturn(PETSC_SUCCESS);
671: }
673: static PetscErrorCode MatMultTransposeAdd_MPIDense(Mat A, Vec xx, Vec yy, Vec zz)
674: {
675: PetscFunctionBegin;
676: PetscCall(MatMultTransposeAddKernel_MPIDense(A, xx, yy, zz, PETSC_FALSE));
677: PetscFunctionReturn(PETSC_SUCCESS);
678: }
680: static PetscErrorCode MatMultHermitianTranspose_MPIDense(Mat A, Vec xx, Vec yy)
681: {
682: PetscFunctionBegin;
683: PetscCall(MatMultTransposeKernel_MPIDense(A, xx, yy, PETSC_TRUE));
684: PetscFunctionReturn(PETSC_SUCCESS);
685: }
687: static PetscErrorCode MatMultHermitianTransposeAdd_MPIDense(Mat A, Vec xx, Vec yy, Vec zz)
688: {
689: PetscFunctionBegin;
690: PetscCall(MatMultTransposeAddKernel_MPIDense(A, xx, yy, zz, PETSC_TRUE));
691: PetscFunctionReturn(PETSC_SUCCESS);
692: }
694: PetscErrorCode MatGetDiagonal_MPIDense(Mat A, Vec v)
695: {
696: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
697: PetscInt lda, len, i, nl, ng, m = A->rmap->n, radd;
698: PetscScalar *x;
699: const PetscScalar *av;
701: PetscFunctionBegin;
702: PetscCall(VecGetArray(v, &x));
703: PetscCall(VecGetSize(v, &ng));
704: PetscCheck(ng == A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming mat and vec");
705: PetscCall(VecGetLocalSize(v, &nl));
706: len = PetscMin(a->A->rmap->n, a->A->cmap->n);
707: radd = A->rmap->rstart * m;
708: PetscCall(MatDenseGetArrayRead(a->A, &av));
709: PetscCall(MatDenseGetLDA(a->A, &lda));
710: for (i = 0; i < len; i++) x[i] = av[radd + i * lda + i];
711: PetscCall(MatDenseRestoreArrayRead(a->A, &av));
712: if (nl - i > 0) PetscCall(PetscArrayzero(x + i, nl - i));
713: PetscCall(VecRestoreArray(v, &x));
714: PetscFunctionReturn(PETSC_SUCCESS);
715: }
717: static PetscErrorCode MatDestroy_MPIDense(Mat mat)
718: {
719: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
721: PetscFunctionBegin;
722: PetscCall(PetscLogObjectState((PetscObject)mat, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT, mat->rmap->N, mat->cmap->N));
723: PetscCall(MatStashDestroy_Private(&mat->stash));
724: PetscCheck(!mdn->vecinuse, PetscObjectComm((PetscObject)mat), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
725: PetscCheck(!mdn->matinuse, PetscObjectComm((PetscObject)mat), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
726: PetscCall(MatDestroy(&mdn->A));
727: PetscCall(VecDestroy(&mdn->lvec));
728: PetscCall(PetscSFDestroy(&mdn->Mvctx));
729: PetscCall(VecDestroy(&mdn->cvec));
730: PetscCall(MatDestroy(&mdn->cmat));
732: PetscCall(PetscFree(mat->data));
733: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, NULL));
735: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetLDA_C", NULL));
736: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseSetLDA_C", NULL));
737: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArray_C", NULL));
738: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArray_C", NULL));
739: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayRead_C", NULL));
740: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayRead_C", NULL));
741: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayWrite_C", NULL));
742: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayWrite_C", NULL));
743: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDensePlaceArray_C", NULL));
744: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseResetArray_C", NULL));
745: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseReplaceArray_C", NULL));
746: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", NULL));
747: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpiaij_C", NULL));
748: #if PetscDefined(HAVE_ELEMENTAL)
749: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_elemental_C", NULL));
750: #endif
751: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
752: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_scalapack_C", NULL));
753: #endif
754: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIDenseSetPreallocation_C", NULL));
755: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpidense_C", NULL));
756: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaij_C", NULL));
757: #if PetscDefined(HAVE_CUDA)
758: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijcusparse_mpidense_C", NULL));
759: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijcusparse_C", NULL));
760: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensecuda_C", NULL));
761: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidensecuda_mpidense_C", NULL));
762: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpidensecuda_C", NULL));
763: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijcusparse_mpidensecuda_C", NULL));
764: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidensecuda_mpiaij_C", NULL));
765: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidensecuda_mpiaijcusparse_C", NULL));
766: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAGetArray_C", NULL));
767: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAGetArrayRead_C", NULL));
768: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAGetArrayWrite_C", NULL));
769: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDARestoreArray_C", NULL));
770: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDARestoreArrayRead_C", NULL));
771: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDARestoreArrayWrite_C", NULL));
772: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAPlaceArray_C", NULL));
773: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAResetArray_C", NULL));
774: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDAReplaceArray_C", NULL));
775: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseCUDASetPreallocation_C", NULL));
776: #endif
777: #if PetscDefined(HAVE_HIP)
778: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijhipsparse_mpidense_C", NULL));
779: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijhipsparse_C", NULL));
780: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensehip_C", NULL));
781: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidensehip_mpidense_C", NULL));
782: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpidensehip_C", NULL));
783: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijhipsparse_mpidensehip_C", NULL));
784: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidensehip_mpiaij_C", NULL));
785: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidensehip_mpiaijhipsparse_C", NULL));
786: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPGetArray_C", NULL));
787: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPGetArrayRead_C", NULL));
788: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPGetArrayWrite_C", NULL));
789: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPRestoreArray_C", NULL));
790: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPRestoreArrayRead_C", NULL));
791: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPRestoreArrayWrite_C", NULL));
792: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPPlaceArray_C", NULL));
793: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPResetArray_C", NULL));
794: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPReplaceArray_C", NULL));
795: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseHIPSetPreallocation_C", NULL));
796: #endif
797: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumn_C", NULL));
798: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumn_C", NULL));
799: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVec_C", NULL));
800: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVec_C", NULL));
801: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecRead_C", NULL));
802: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecRead_C", NULL));
803: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecWrite_C", NULL));
804: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecWrite_C", NULL));
805: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetSubMatrix_C", NULL));
806: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreSubMatrix_C", NULL));
807: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultColumnRange_C", NULL));
808: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultAddColumnRange_C", NULL));
809: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeColumnRange_C", NULL));
810: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeAddColumnRange_C", NULL));
811: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatGetMultPetscSF_C", NULL));
812: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseUpdateColumnLayout_C", NULL));
814: PetscCall(PetscObjectCompose((PetscObject)mat, "DiagonalBlock", NULL));
815: PetscFunctionReturn(PETSC_SUCCESS);
816: }
818: #include <petscdraw.h>
819: static PetscErrorCode MatView_MPIDense_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
820: {
821: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
822: PetscMPIInt rank;
823: PetscViewerType vtype;
824: PetscBool isascii, isdraw;
825: PetscViewer sviewer;
826: PetscViewerFormat format;
828: PetscFunctionBegin;
829: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
830: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
831: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
832: if (isascii) {
833: PetscCall(PetscViewerGetType(viewer, &vtype));
834: PetscCall(PetscViewerGetFormat(viewer, &format));
835: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
836: MatInfo info;
837: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
838: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
839: 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,
840: (PetscInt)info.memory));
841: PetscCall(PetscViewerFlush(viewer));
842: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
843: if (mdn->Mvctx) PetscCall(PetscSFView(mdn->Mvctx, viewer));
844: PetscFunctionReturn(PETSC_SUCCESS);
845: } else if (format == PETSC_VIEWER_ASCII_INFO) {
846: PetscFunctionReturn(PETSC_SUCCESS);
847: }
848: } else if (isdraw) {
849: PetscDraw draw;
850: PetscBool isnull;
852: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
853: PetscCall(PetscDrawIsNull(draw, &isnull));
854: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
855: }
857: {
858: /* assemble the entire matrix onto first processor. */
859: Mat A;
860: PetscInt M = mat->rmap->N, N = mat->cmap->N, m, row, i, nz;
861: PetscInt *cols;
862: PetscScalar *vals;
864: PetscCall(MatCreate(PetscObjectComm((PetscObject)mat), &A));
865: if (rank == 0) {
866: PetscCall(MatSetSizes(A, M, N, M, N));
867: } else {
868: PetscCall(MatSetSizes(A, 0, 0, M, N));
869: }
870: /* Since this is a temporary matrix, MATMPIDENSE instead of ((PetscObject)A)->type_name here is probably acceptable. */
871: PetscCall(MatSetType(A, MATMPIDENSE));
872: PetscCall(MatMPIDenseSetPreallocation(A, NULL));
874: /* Copy the matrix ... This isn't the most efficient means,
875: but it's quick for now */
876: A->insertmode = INSERT_VALUES;
878: row = mat->rmap->rstart;
879: m = mdn->A->rmap->n;
880: for (i = 0; i < m; i++) {
881: PetscCall(MatGetRow_MPIDense(mat, row, &nz, &cols, &vals));
882: PetscCall(MatSetValues_MPIDense(A, 1, &row, nz, cols, vals, INSERT_VALUES));
883: PetscCall(MatRestoreRow_MPIDense(mat, row, &nz, &cols, &vals));
884: row++;
885: }
887: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
888: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
889: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
890: if (rank == 0) {
891: PetscCall(PetscObjectSetName((PetscObject)((Mat_MPIDense *)A->data)->A, ((PetscObject)mat)->name));
892: PetscCall(MatView_SeqDense(((Mat_MPIDense *)A->data)->A, sviewer));
893: }
894: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
895: PetscCall(MatDestroy(&A));
896: }
897: PetscFunctionReturn(PETSC_SUCCESS);
898: }
900: static PetscErrorCode MatView_MPIDense(Mat mat, PetscViewer viewer)
901: {
902: PetscBool isascii, isbinary, isdraw, issocket;
904: PetscFunctionBegin;
905: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
906: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
907: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
908: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
910: if (isascii || issocket || isdraw) PetscCall(MatView_MPIDense_ASCIIorDraworSocket(mat, viewer));
911: else if (isbinary) PetscCall(MatView_Dense_Binary(mat, viewer));
912: PetscFunctionReturn(PETSC_SUCCESS);
913: }
915: static PetscErrorCode MatGetInfo_MPIDense(Mat A, MatInfoType flag, MatInfo *info)
916: {
917: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
918: Mat mdn = mat->A;
919: PetscLogDouble irecv[5];
921: PetscFunctionBegin;
922: info->block_size = 1.0;
924: PetscCall(MatGetInfo(mdn, MAT_LOCAL, info));
926: irecv[0] = info->nz_used;
927: irecv[1] = info->nz_allocated;
928: irecv[2] = info->nz_unneeded;
929: irecv[3] = info->memory;
930: irecv[4] = info->mallocs;
931: if (flag == MAT_LOCAL) {
932: info->nz_used = irecv[0];
933: info->nz_allocated = irecv[1];
934: info->nz_unneeded = irecv[2];
935: info->memory = irecv[3];
936: info->mallocs = irecv[4];
937: } else if (flag == MAT_GLOBAL_MAX) {
938: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)A)));
940: info->nz_used = irecv[0];
941: info->nz_allocated = irecv[1];
942: info->nz_unneeded = irecv[2];
943: info->memory = irecv[3];
944: info->mallocs = irecv[4];
945: } else if (flag == MAT_GLOBAL_SUM) {
946: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)A)));
948: info->nz_used = irecv[0];
949: info->nz_allocated = irecv[1];
950: info->nz_unneeded = irecv[2];
951: info->memory = irecv[3];
952: info->mallocs = irecv[4];
953: }
954: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
955: info->fill_ratio_needed = 0;
956: info->factor_mallocs = 0;
957: PetscFunctionReturn(PETSC_SUCCESS);
958: }
960: static PetscErrorCode MatSetOption_MPIDense(Mat A, MatOption op, PetscBool flg)
961: {
962: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
964: PetscFunctionBegin;
965: switch (op) {
966: case MAT_NEW_NONZERO_LOCATIONS:
967: case MAT_NEW_NONZERO_LOCATION_ERR:
968: case MAT_NEW_NONZERO_ALLOCATION_ERR:
969: MatCheckPreallocated(A, 1);
970: PetscCall(MatSetOption(a->A, op, flg));
971: break;
972: case MAT_ROW_ORIENTED:
973: MatCheckPreallocated(A, 1);
974: a->roworiented = flg;
975: PetscCall(MatSetOption(a->A, op, flg));
976: break;
977: case MAT_IGNORE_OFF_PROC_ENTRIES:
978: a->donotstash = flg;
979: break;
980: case MAT_SYMMETRIC:
981: case MAT_STRUCTURALLY_SYMMETRIC:
982: case MAT_HERMITIAN:
983: case MAT_SYMMETRY_ETERNAL:
984: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
985: case MAT_SPD:
986: case MAT_SPD_ETERNAL:
987: /* if the diagonal matrix is square it inherits some of the properties above */
988: if (a->A && A->rmap->n == A->cmap->n) PetscCall(MatSetOption(a->A, op, flg));
989: break;
990: default:
991: break;
992: }
993: PetscFunctionReturn(PETSC_SUCCESS);
994: }
996: static PetscErrorCode MatDiagonalScale_MPIDense(Mat A, Vec ll, Vec rr)
997: {
998: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
999: PetscInt s1, s2, s3;
1001: PetscFunctionBegin;
1002: PetscCall(MatGetLocalSize(A, &s2, &s3));
1003: if (ll) {
1004: PetscCall(VecGetLocalSize(ll, &s1));
1005: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vector non-conforming local size, %" PetscInt_FMT " != %" PetscInt_FMT, s1, s2);
1006: }
1007: if (rr) {
1008: PetscCall(VecGetLocalSize(rr, &s1));
1009: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vector non-conforming local size, %" PetscInt_FMT " != %" PetscInt_FMT, s1, s3);
1010: /* gather the right scaling into the local column layout, staying on the device when the Vecs live there */
1011: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
1012: PetscCall(VecScatterBegin(mdn->Mvctx, rr, mdn->lvec, INSERT_VALUES, SCATTER_FORWARD));
1013: PetscCall(VecScatterEnd(mdn->Mvctx, rr, mdn->lvec, INSERT_VALUES, SCATTER_FORWARD));
1014: }
1015: /* the local matrix holds exactly the local rows and all the columns, so the local part of ll applies to it directly;
1016: the operation is called rather than MatDiagonalScale() because ll is the parallel Vec and the interface checks that
1017: it shares the communicator of the sequential matrix, as in MatDiagonalScale_MPIAIJ() */
1018: PetscUseTypeMethod(mdn->A, diagonalscale, ll, rr ? mdn->lvec : NULL);
1019: PetscCall(PetscObjectStateIncrease((PetscObject)mdn->A));
1020: PetscFunctionReturn(PETSC_SUCCESS);
1021: }
1023: static PetscErrorCode MatNorm_MPIDense(Mat A, NormType type, PetscReal *nrm)
1024: {
1025: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
1026: PetscInt i, j, lda;
1027: PetscMPIInt size;
1028: const PetscScalar *av;
1030: PetscFunctionBegin;
1031: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
1032: if (size == 1) {
1033: PetscCall(MatNorm(mdn->A, type, nrm));
1034: } else {
1035: if (type == NORM_FROBENIUS) {
1036: PetscCall(MatNorm(mdn->A, NORM_FROBENIUS, nrm));
1037: *nrm *= *nrm;
1038: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, nrm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
1039: *nrm = PetscSqrtReal(*nrm);
1040: } else if (type == NORM_1) {
1041: PetscReal *tmp;
1043: PetscCall(PetscCalloc1(A->cmap->N, &tmp));
1044: *nrm = 0.0;
1045: PetscCall(MatDenseGetArrayRead(mdn->A, &av));
1046: PetscCall(MatDenseGetLDA(mdn->A, &lda));
1047: for (j = 0; j < mdn->A->cmap->n; j++) {
1048: for (i = 0; i < mdn->A->rmap->n; i++) tmp[j] += PetscAbsScalar(av[i + j * lda]);
1049: }
1050: PetscCall(MatDenseRestoreArrayRead(mdn->A, &av));
1051: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, tmp, A->cmap->N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
1052: for (j = 0; j < A->cmap->N; j++) {
1053: if (tmp[j] > *nrm) *nrm = tmp[j];
1054: }
1055: PetscCall(PetscFree(tmp));
1056: PetscCall(PetscLogFlops(A->cmap->n * A->rmap->n));
1057: } else if (type == NORM_INFINITY) { /* max row norm */
1058: PetscCall(MatNorm(mdn->A, type, nrm));
1059: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, nrm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)A)));
1060: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Unsupported norm type %s", NormTypes[type]);
1061: }
1062: PetscFunctionReturn(PETSC_SUCCESS);
1063: }
1065: static PetscErrorCode MatTranspose_MPIDense(Mat A, MatReuse reuse, Mat *matout)
1066: {
1067: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1068: Mat B;
1069: PetscInt M = A->rmap->N, N = A->cmap->N, m, n, *rwork, rstart = A->rmap->rstart;
1070: PetscInt j, i, lda;
1071: PetscScalar *v;
1073: PetscFunctionBegin;
1074: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1075: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_INPLACE_MATRIX) {
1076: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1077: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1078: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1079: PetscCall(MatMPIDenseSetPreallocation(B, NULL));
1080: } else B = *matout;
1082: m = a->A->rmap->n;
1083: n = a->A->cmap->n;
1084: PetscCall(MatDenseGetArrayRead(a->A, (const PetscScalar **)&v));
1085: PetscCall(MatDenseGetLDA(a->A, &lda));
1086: PetscCall(PetscMalloc1(m, &rwork));
1087: for (i = 0; i < m; i++) rwork[i] = rstart + i;
1088: for (j = 0; j < n; j++) {
1089: PetscCall(MatSetValues(B, 1, &j, m, rwork, v, INSERT_VALUES));
1090: v = PetscSafePointerPlusOffset(v, lda);
1091: }
1092: PetscCall(MatDenseRestoreArrayRead(a->A, (const PetscScalar **)&v));
1093: PetscCall(PetscFree(rwork));
1094: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1095: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1096: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1097: *matout = B;
1098: } else {
1099: PetscCall(MatHeaderMerge(A, &B));
1100: }
1101: PetscFunctionReturn(PETSC_SUCCESS);
1102: }
1104: static PetscErrorCode MatDuplicate_MPIDense(Mat, MatDuplicateOption, Mat *);
1105: PETSC_INTERN PetscErrorCode MatScale_MPIDense(Mat, PetscScalar);
1107: static PetscErrorCode MatSetUp_MPIDense(Mat A)
1108: {
1109: PetscFunctionBegin;
1110: PetscCall(PetscLayoutSetUp(A->rmap));
1111: PetscCall(PetscLayoutSetUp(A->cmap));
1112: if (!A->preallocated) PetscCall(MatMPIDenseSetPreallocation(A, NULL));
1113: PetscFunctionReturn(PETSC_SUCCESS);
1114: }
1116: static PetscErrorCode MatAXPY_MPIDense(Mat Y, PetscScalar alpha, Mat X, MatStructure str)
1117: {
1118: Mat_MPIDense *A = (Mat_MPIDense *)Y->data, *B = (Mat_MPIDense *)X->data;
1120: PetscFunctionBegin;
1121: PetscCall(MatAXPY(A->A, alpha, B->A, str));
1122: PetscFunctionReturn(PETSC_SUCCESS);
1123: }
1125: static PetscErrorCode MatConjugate_MPIDense(Mat mat)
1126: {
1127: Mat_MPIDense *a = (Mat_MPIDense *)mat->data;
1129: PetscFunctionBegin;
1130: PetscCall(MatConjugate(a->A));
1131: PetscFunctionReturn(PETSC_SUCCESS);
1132: }
1134: static PetscErrorCode MatRealPart_MPIDense(Mat A)
1135: {
1136: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1138: PetscFunctionBegin;
1139: PetscCall(MatRealPart(a->A));
1140: PetscFunctionReturn(PETSC_SUCCESS);
1141: }
1143: static PetscErrorCode MatImaginaryPart_MPIDense(Mat A)
1144: {
1145: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1147: PetscFunctionBegin;
1148: PetscCall(MatImaginaryPart(a->A));
1149: PetscFunctionReturn(PETSC_SUCCESS);
1150: }
1152: static PetscErrorCode MatGetColumnVector_MPIDense(Mat A, Vec v, PetscInt col)
1153: {
1154: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1156: PetscFunctionBegin;
1157: PetscCheck(a->A, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Missing local matrix");
1158: PetscCheck(a->A->ops->getcolumnvector, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Missing get column operation");
1159: PetscUseTypeMethod(a->A, getcolumnvector, v, col);
1160: PetscFunctionReturn(PETSC_SUCCESS);
1161: }
1163: PETSC_INTERN PetscErrorCode MatGetColumnReductions_SeqDense(Mat, PetscInt, PetscReal *);
1165: static PetscErrorCode MatGetColumnReductions_MPIDense(Mat A, PetscInt type, PetscReal *reductions)
1166: {
1167: PetscInt i, m, n;
1168: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1170: PetscFunctionBegin;
1171: PetscCall(MatGetSize(A, &m, &n));
1172: if (type == REDUCTION_MEAN_REALPART) {
1173: PetscCall(MatGetColumnReductions_SeqDense(a->A, (PetscInt)REDUCTION_SUM_REALPART, reductions));
1174: } else if (type == REDUCTION_MEAN_IMAGINARYPART) {
1175: PetscCall(MatGetColumnReductions_SeqDense(a->A, (PetscInt)REDUCTION_SUM_IMAGINARYPART, reductions));
1176: } else {
1177: PetscCall(MatGetColumnReductions_SeqDense(a->A, type, reductions));
1178: }
1179: if (type == NORM_2) {
1180: for (i = 0; i < n; i++) reductions[i] *= reductions[i];
1181: }
1182: if (type == NORM_INFINITY) {
1183: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, reductions, n, MPIU_REAL, MPIU_MAX, A->hdr.comm));
1184: } else {
1185: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, reductions, n, MPIU_REAL, MPIU_SUM, A->hdr.comm));
1186: }
1187: if (type == NORM_2) {
1188: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
1189: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
1190: for (i = 0; i < n; i++) reductions[i] /= m;
1191: }
1192: PetscFunctionReturn(PETSC_SUCCESS);
1193: }
1195: static PetscErrorCode MatSetRandom_MPIDense(Mat x, PetscRandom rctx)
1196: {
1197: Mat_MPIDense *d = (Mat_MPIDense *)x->data;
1199: PetscFunctionBegin;
1200: PetscCall(MatSetRandom(d->A, rctx));
1201: #if PetscDefined(HAVE_DEVICE)
1202: x->offloadmask = d->A->offloadmask;
1203: #endif
1204: PetscCall(PetscObjectStateIncrease((PetscObject)x));
1205: PetscFunctionReturn(PETSC_SUCCESS);
1206: }
1208: static PetscErrorCode MatMatTransposeMultSymbolic_MPIDense_MPIDense(Mat, Mat, PetscReal, Mat);
1209: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense(Mat, Mat, Mat);
1210: static PetscErrorCode MatTransposeMatMultSymbolic_MPIDense_MPIDense(Mat, Mat, PetscReal, Mat);
1211: static PetscErrorCode MatTransposeMatMultNumeric_MPIDense_MPIDense(Mat, Mat, Mat);
1212: static PetscErrorCode MatEqual_MPIDense(Mat, Mat, PetscBool *);
1213: static PetscErrorCode MatLoad_MPIDense(Mat, PetscViewer);
1214: static PetscErrorCode MatProductSetFromOptions_MPIDense(Mat);
1216: static struct _MatOps MatOps_Values = {MatSetValues_MPIDense,
1217: MatGetRow_MPIDense,
1218: MatRestoreRow_MPIDense,
1219: MatMult_MPIDense,
1220: /* 4*/ MatMultAdd_MPIDense,
1221: MatMultTranspose_MPIDense,
1222: MatMultTransposeAdd_MPIDense,
1223: NULL,
1224: NULL,
1225: NULL,
1226: /* 10*/ NULL,
1227: NULL,
1228: NULL,
1229: NULL,
1230: MatTranspose_MPIDense,
1231: /* 15*/ MatGetInfo_MPIDense,
1232: MatEqual_MPIDense,
1233: MatGetDiagonal_MPIDense,
1234: MatDiagonalScale_MPIDense,
1235: MatNorm_MPIDense,
1236: /* 20*/ MatAssemblyBegin_MPIDense,
1237: MatAssemblyEnd_MPIDense,
1238: MatSetOption_MPIDense,
1239: MatZeroEntries_MPIDense,
1240: /* 24*/ MatZeroRows_MPIDense,
1241: NULL,
1242: NULL,
1243: NULL,
1244: NULL,
1245: /* 29*/ MatSetUp_MPIDense,
1246: NULL,
1247: NULL,
1248: MatGetDiagonalBlock_MPIDense,
1249: MatSetInf_MPIDense,
1250: /* 34*/ MatDuplicate_MPIDense,
1251: NULL,
1252: NULL,
1253: NULL,
1254: NULL,
1255: /* 39*/ MatAXPY_MPIDense,
1256: MatCreateSubMatrices_MPIDense,
1257: NULL,
1258: MatGetValues_MPIDense,
1259: MatCopy_MPIDense,
1260: /* 44*/ NULL,
1261: MatScale_MPIDense,
1262: MatShift_MPIDense,
1263: NULL,
1264: NULL,
1265: /* 49*/ MatSetRandom_MPIDense,
1266: NULL,
1267: NULL,
1268: NULL,
1269: NULL,
1270: /* 54*/ NULL,
1271: NULL,
1272: NULL,
1273: NULL,
1274: NULL,
1275: /* 59*/ MatCreateSubMatrix_MPIDense,
1276: MatDestroy_MPIDense,
1277: MatView_MPIDense,
1278: NULL,
1279: NULL,
1280: /* 64*/ NULL,
1281: NULL,
1282: NULL,
1283: NULL,
1284: NULL,
1285: /* 69*/ NULL,
1286: NULL,
1287: NULL,
1288: NULL,
1289: NULL,
1290: /* 74*/ NULL,
1291: NULL,
1292: NULL,
1293: NULL,
1294: MatLoad_MPIDense,
1295: /* 79*/ NULL,
1296: NULL,
1297: NULL,
1298: NULL,
1299: /* 83*/ NULL,
1300: NULL,
1301: NULL,
1302: NULL,
1303: MatMatTransposeMultSymbolic_MPIDense_MPIDense,
1304: MatMatTransposeMultNumeric_MPIDense_MPIDense,
1305: /* 89*/ NULL,
1306: MatProductSetFromOptions_MPIDense,
1307: NULL,
1308: NULL,
1309: MatConjugate_MPIDense,
1310: /* 94*/ NULL,
1311: NULL,
1312: MatRealPart_MPIDense,
1313: MatImaginaryPart_MPIDense,
1314: NULL,
1315: /*99*/ NULL,
1316: NULL,
1317: NULL,
1318: NULL,
1319: MatGetColumnVector_MPIDense,
1320: /*104*/ NULL,
1321: NULL,
1322: NULL,
1323: NULL,
1324: NULL,
1325: /*109*/ NULL,
1326: NULL,
1327: MatMultHermitianTranspose_MPIDense,
1328: MatMultHermitianTransposeAdd_MPIDense,
1329: NULL,
1330: /*114*/ NULL,
1331: MatGetColumnReductions_MPIDense,
1332: NULL,
1333: NULL,
1334: NULL,
1335: /*120*/ MatTransposeMatMultSymbolic_MPIDense_MPIDense,
1336: MatTransposeMatMultNumeric_MPIDense_MPIDense,
1337: NULL,
1338: NULL,
1339: /*124*/ NULL,
1340: NULL,
1341: NULL,
1342: NULL,
1343: NULL,
1344: /*129*/ NULL,
1345: MatCreateMPIMatConcatenateSeqMat_MPIDense,
1346: NULL,
1347: NULL,
1348: NULL,
1349: /*134*/ NULL,
1350: NULL,
1351: NULL,
1352: NULL,
1353: NULL,
1354: /*139*/ NULL,
1355: NULL,
1356: NULL,
1357: NULL,
1358: NULL,
1359: NULL,
1360: /*144*/ NULL,
1361: NULL,
1362: NULL,
1363: NULL};
1365: static PetscErrorCode MatMPIDenseSetPreallocation_MPIDense(Mat mat, PetscScalar *data)
1366: {
1367: Mat_MPIDense *a = (Mat_MPIDense *)mat->data;
1368: MatType mtype = MATSEQDENSE;
1370: PetscFunctionBegin;
1371: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)mat), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1372: PetscCall(PetscLayoutSetUp(mat->rmap));
1373: PetscCall(PetscLayoutSetUp(mat->cmap));
1374: if (!a->A) {
1375: PetscCall(MatCreate(PETSC_COMM_SELF, &a->A));
1376: PetscCall(MatSetSizes(a->A, mat->rmap->n, mat->cmap->N, mat->rmap->n, mat->cmap->N));
1377: }
1378: #if PetscDefined(HAVE_CUDA)
1379: PetscBool iscuda;
1380: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATMPIDENSECUDA, &iscuda));
1381: if (iscuda) mtype = MATSEQDENSECUDA;
1382: #endif
1383: #if PetscDefined(HAVE_HIP)
1384: PetscBool iship;
1385: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATMPIDENSEHIP, &iship));
1386: if (iship) mtype = MATSEQDENSEHIP;
1387: #endif
1388: PetscCall(MatSetType(a->A, mtype));
1389: PetscCall(MatSeqDenseSetPreallocation(a->A, data));
1390: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
1391: mat->offloadmask = a->A->offloadmask;
1392: #endif
1393: mat->preallocated = PETSC_TRUE;
1394: mat->assembled = PETSC_TRUE;
1395: PetscFunctionReturn(PETSC_SUCCESS);
1396: }
1398: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIDense(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
1399: {
1400: Mat B, C;
1402: PetscFunctionBegin;
1403: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &C));
1404: PetscCall(MatConvert_SeqAIJ_SeqDense(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &B));
1405: PetscCall(MatDestroy(&C));
1406: if (reuse == MAT_REUSE_MATRIX) {
1407: C = *newmat;
1408: } else C = NULL;
1409: PetscCall(MatCreateMPIMatConcatenateSeqMat(PetscObjectComm((PetscObject)A), B, A->cmap->n, !C ? MAT_INITIAL_MATRIX : MAT_REUSE_MATRIX, &C));
1410: PetscCall(MatDestroy(&B));
1411: if (reuse == MAT_INPLACE_MATRIX) PetscCall(MatHeaderReplace(A, &C));
1412: else if (reuse == MAT_INITIAL_MATRIX) *newmat = C;
1413: PetscFunctionReturn(PETSC_SUCCESS);
1414: }
1416: static PetscErrorCode MatConvert_MPIDense_MPIAIJ(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
1417: {
1418: Mat B, C;
1420: PetscFunctionBegin;
1421: PetscCall(MatDenseGetLocalMatrix(A, &C));
1422: PetscCall(MatConvert_SeqDense_SeqAIJ(C, MATSEQAIJ, MAT_INITIAL_MATRIX, &B));
1423: if (reuse == MAT_REUSE_MATRIX) {
1424: C = *newmat;
1425: } else C = NULL;
1426: PetscCall(MatCreateMPIMatConcatenateSeqMat(PetscObjectComm((PetscObject)A), B, A->cmap->n, !C ? MAT_INITIAL_MATRIX : MAT_REUSE_MATRIX, &C));
1427: PetscCall(MatDestroy(&B));
1428: if (reuse == MAT_INPLACE_MATRIX) PetscCall(MatHeaderReplace(A, &C));
1429: else if (reuse == MAT_INITIAL_MATRIX) *newmat = C;
1430: PetscFunctionReturn(PETSC_SUCCESS);
1431: }
1433: #if PetscDefined(HAVE_ELEMENTAL)
1434: PETSC_INTERN PetscErrorCode MatConvert_MPIDense_Elemental(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
1435: {
1436: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1437: Mat mat_elemental;
1438: PetscScalar *v;
1439: PetscInt m = A->rmap->n, N = A->cmap->N, rstart = A->rmap->rstart, i, *rows, *cols, lda;
1441: PetscFunctionBegin;
1442: if (reuse == MAT_REUSE_MATRIX) {
1443: mat_elemental = *newmat;
1444: PetscCall(MatZeroEntries(*newmat));
1445: } else {
1446: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental));
1447: PetscCall(MatSetSizes(mat_elemental, PETSC_DECIDE, PETSC_DECIDE, A->rmap->N, A->cmap->N));
1448: PetscCall(MatSetType(mat_elemental, MATELEMENTAL));
1449: PetscCall(MatSetUp(mat_elemental));
1450: PetscCall(MatSetOption(mat_elemental, MAT_ROW_ORIENTED, PETSC_FALSE));
1451: }
1453: PetscCall(PetscMalloc2(m, &rows, N, &cols));
1454: for (i = 0; i < N; i++) cols[i] = i;
1455: for (i = 0; i < m; i++) rows[i] = rstart + i;
1457: /* PETSc-Elemental interface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
1458: PetscCall(MatDenseGetArray(A, &v));
1459: PetscCall(MatDenseGetLDA(a->A, &lda));
1460: if (lda == m) PetscCall(MatSetValues(mat_elemental, m, rows, N, cols, v, ADD_VALUES));
1461: else {
1462: for (i = 0; i < N; i++) PetscCall(MatSetValues(mat_elemental, m, rows, 1, &i, v + lda * i, ADD_VALUES));
1463: }
1464: PetscCall(MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY));
1465: PetscCall(MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY));
1466: PetscCall(MatDenseRestoreArray(A, &v));
1467: PetscCall(PetscFree2(rows, cols));
1469: if (reuse == MAT_INPLACE_MATRIX) {
1470: PetscCall(MatHeaderReplace(A, &mat_elemental));
1471: } else {
1472: *newmat = mat_elemental;
1473: }
1474: PetscFunctionReturn(PETSC_SUCCESS);
1475: }
1476: #endif
1478: static PetscErrorCode MatDenseGetColumn_MPIDense(Mat A, PetscInt col, PetscScalar **vals)
1479: {
1480: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
1482: PetscFunctionBegin;
1483: PetscCall(MatDenseGetColumn(mat->A, col, vals));
1484: PetscFunctionReturn(PETSC_SUCCESS);
1485: }
1487: static PetscErrorCode MatDenseRestoreColumn_MPIDense(Mat A, PetscScalar **vals)
1488: {
1489: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
1491: PetscFunctionBegin;
1492: PetscCall(MatDenseRestoreColumn(mat->A, vals));
1493: PetscFunctionReturn(PETSC_SUCCESS);
1494: }
1496: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIDense(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
1497: {
1498: Mat_MPIDense *mat;
1499: PetscInt m, nloc, N;
1501: PetscFunctionBegin;
1502: PetscCall(MatGetSize(inmat, &m, &N));
1503: PetscCall(MatGetLocalSize(inmat, NULL, &nloc));
1504: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
1505: PetscInt sum;
1507: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
1508: /* Check sum(n) = N */
1509: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
1510: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
1512: PetscCall(MatCreateDense(comm, m, n, PETSC_DETERMINE, N, NULL, outmat));
1513: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
1514: }
1516: /* numeric phase */
1517: mat = (Mat_MPIDense *)(*outmat)->data;
1518: PetscCall(MatCopy(inmat, mat->A, SAME_NONZERO_PATTERN));
1519: PetscFunctionReturn(PETSC_SUCCESS);
1520: }
1522: PetscErrorCode MatDenseGetColumnVec_MPIDense(Mat A, PetscInt col, Vec *v)
1523: {
1524: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1525: PetscInt lda;
1527: PetscFunctionBegin;
1528: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1529: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1530: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
1531: a->vecinuse = col + 1;
1532: PetscCall(MatDenseGetLDA(a->A, &lda));
1533: PetscCall(MatDenseGetArray(a->A, (PetscScalar **)&a->ptrinuse));
1534: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)lda)));
1535: *v = a->cvec;
1536: PetscFunctionReturn(PETSC_SUCCESS);
1537: }
1539: PetscErrorCode MatDenseRestoreColumnVec_MPIDense(Mat A, PetscInt col, Vec *v)
1540: {
1541: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1543: PetscFunctionBegin;
1544: PetscCheck(a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
1545: PetscCheck(a->cvec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column vector");
1546: VecCheckAssembled(a->cvec);
1547: a->vecinuse = 0;
1548: PetscCall(MatDenseRestoreArray(a->A, (PetscScalar **)&a->ptrinuse));
1549: PetscCall(VecResetArray(a->cvec));
1550: if (v) *v = NULL;
1551: PetscFunctionReturn(PETSC_SUCCESS);
1552: }
1554: PetscErrorCode MatDenseGetColumnVecRead_MPIDense(Mat A, PetscInt col, Vec *v)
1555: {
1556: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1557: PetscInt lda;
1559: PetscFunctionBegin;
1560: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1561: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1562: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
1563: a->vecinuse = col + 1;
1564: PetscCall(MatDenseGetLDA(a->A, &lda));
1565: PetscCall(MatDenseGetArrayRead(a->A, &a->ptrinuse));
1566: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)lda)));
1567: PetscCall(VecLockReadPush(a->cvec));
1568: *v = a->cvec;
1569: PetscFunctionReturn(PETSC_SUCCESS);
1570: }
1572: PetscErrorCode MatDenseRestoreColumnVecRead_MPIDense(Mat A, PetscInt col, Vec *v)
1573: {
1574: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1576: PetscFunctionBegin;
1577: PetscCheck(a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
1578: PetscCheck(a->cvec, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing internal column vector");
1579: VecCheckAssembled(a->cvec);
1580: a->vecinuse = 0;
1581: PetscCall(MatDenseRestoreArrayRead(a->A, &a->ptrinuse));
1582: PetscCall(VecLockReadPop(a->cvec));
1583: PetscCall(VecResetArray(a->cvec));
1584: if (v) *v = NULL;
1585: PetscFunctionReturn(PETSC_SUCCESS);
1586: }
1588: PetscErrorCode MatDenseGetColumnVecWrite_MPIDense(Mat A, PetscInt col, Vec *v)
1589: {
1590: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1591: PetscInt lda;
1593: PetscFunctionBegin;
1594: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1595: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1596: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
1597: a->vecinuse = col + 1;
1598: PetscCall(MatDenseGetLDA(a->A, &lda));
1599: PetscCall(MatDenseGetArrayWrite(a->A, (PetscScalar **)&a->ptrinuse));
1600: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)lda)));
1601: *v = a->cvec;
1602: PetscFunctionReturn(PETSC_SUCCESS);
1603: }
1605: PetscErrorCode MatDenseRestoreColumnVecWrite_MPIDense(Mat A, PetscInt col, Vec *v)
1606: {
1607: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1609: PetscFunctionBegin;
1610: PetscCheck(a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
1611: PetscCheck(a->cvec, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing internal column vector");
1612: VecCheckAssembled(a->cvec);
1613: a->vecinuse = 0;
1614: PetscCall(MatDenseRestoreArrayWrite(a->A, (PetscScalar **)&a->ptrinuse));
1615: PetscCall(VecResetArray(a->cvec));
1616: if (v) *v = NULL;
1617: PetscFunctionReturn(PETSC_SUCCESS);
1618: }
1620: static PetscErrorCode MatDenseGetSubMatrix_MPIDense(Mat A, PetscInt rbegin, PetscInt rend, PetscInt cbegin, PetscInt cend, Mat *v)
1621: {
1622: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1623: Mat_MPIDense *c;
1624: MPI_Comm comm;
1625: PetscInt prbegin, prend, pcbegin, pcend;
1627: PetscFunctionBegin;
1628: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
1629: PetscCheck(!a->vecinuse, comm, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1630: PetscCheck(!a->matinuse, comm, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1631: prbegin = PetscMax(0, PetscMin(A->rmap->rend, rbegin) - A->rmap->rstart);
1632: prend = PetscMin(A->rmap->n, PetscMax(0, rend - A->rmap->rstart));
1633: pcbegin = PetscMax(0, PetscMin(A->cmap->rend, cbegin) - A->cmap->rstart);
1634: pcend = PetscMin(A->cmap->n, PetscMax(0, cend - A->cmap->rstart));
1635: if (!a->cmat) {
1636: PetscCall(MatCreate(comm, &a->cmat));
1637: PetscCall(MatSetType(a->cmat, ((PetscObject)A)->type_name));
1638: PetscCall(MatSetVecType(a->cmat, A->defaultvectype));
1639: if (rend - rbegin == A->rmap->N) PetscCall(PetscLayoutReference(A->rmap, &a->cmat->rmap));
1640: else {
1641: PetscCall(PetscLayoutSetLocalSize(a->cmat->rmap, prend - prbegin));
1642: PetscCall(PetscLayoutSetSize(a->cmat->rmap, rend - rbegin));
1643: PetscCall(PetscLayoutSetUp(a->cmat->rmap));
1644: }
1645: if (cend - cbegin == A->cmap->N) PetscCall(PetscLayoutReference(A->cmap, &a->cmat->cmap));
1646: else {
1647: PetscCall(PetscLayoutSetLocalSize(a->cmat->cmap, pcend - pcbegin));
1648: PetscCall(PetscLayoutSetSize(a->cmat->cmap, cend - cbegin));
1649: PetscCall(PetscLayoutSetUp(a->cmat->cmap));
1650: }
1651: c = (Mat_MPIDense *)a->cmat->data;
1652: c->sub_rbegin = rbegin;
1653: c->sub_rend = rend;
1654: c->sub_cbegin = cbegin;
1655: c->sub_cend = cend;
1656: }
1657: c = (Mat_MPIDense *)a->cmat->data;
1658: if (c->sub_rbegin != rbegin || c->sub_rend != rend) {
1659: PetscCall(PetscLayoutDestroy(&a->cmat->rmap));
1660: PetscCall(PetscLayoutCreate(comm, &a->cmat->rmap));
1661: PetscCall(PetscLayoutSetLocalSize(a->cmat->rmap, prend - prbegin));
1662: PetscCall(PetscLayoutSetSize(a->cmat->rmap, rend - rbegin));
1663: PetscCall(PetscLayoutSetUp(a->cmat->rmap));
1664: c->sub_rbegin = rbegin;
1665: c->sub_rend = rend;
1666: }
1667: if (c->sub_cbegin != cbegin || c->sub_cend != cend) {
1668: // special optimization: check if all columns are owned by rank 0, in which case no communication is necessary
1669: if (cend - cbegin != a->cmat->cmap->N || A->cmap->range[1] != A->cmap->N) {
1670: PetscCall(PetscLayoutDestroy(&a->cmat->cmap));
1671: PetscCall(PetscLayoutCreate(comm, &a->cmat->cmap));
1672: PetscCall(PetscLayoutSetLocalSize(a->cmat->cmap, pcend - pcbegin));
1673: PetscCall(PetscLayoutSetSize(a->cmat->cmap, cend - cbegin));
1674: PetscCall(PetscLayoutSetUp(a->cmat->cmap));
1675: PetscCall(VecDestroy(&c->lvec));
1676: PetscCall(PetscSFDestroy(&c->Mvctx));
1677: }
1678: c->sub_cbegin = cbegin;
1679: c->sub_cend = cend;
1680: }
1681: PetscCheck(!c->A, comm, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1682: PetscCall(MatDenseGetSubMatrix(a->A, prbegin, prend, cbegin, cend, &c->A));
1684: a->cmat->preallocated = PETSC_TRUE;
1685: a->cmat->assembled = PETSC_TRUE;
1686: #if PetscDefined(HAVE_DEVICE)
1687: a->cmat->offloadmask = c->A->offloadmask;
1688: #endif
1689: a->matinuse = cbegin + 1;
1690: *v = a->cmat;
1691: PetscFunctionReturn(PETSC_SUCCESS);
1692: }
1694: static PetscErrorCode MatDenseRestoreSubMatrix_MPIDense(Mat A, Mat *v)
1695: {
1696: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1697: Mat_MPIDense *c;
1699: PetscFunctionBegin;
1700: PetscCheck(a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseGetSubMatrix() first");
1701: PetscCheck(a->cmat, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing internal matrix");
1702: PetscCheck(*v == a->cmat, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Not the matrix obtained from MatDenseGetSubMatrix()");
1703: a->matinuse = 0;
1704: c = (Mat_MPIDense *)a->cmat->data;
1705: PetscCall(MatDenseRestoreSubMatrix(a->A, &c->A));
1706: *v = NULL;
1707: #if PetscDefined(HAVE_DEVICE)
1708: A->offloadmask = a->A->offloadmask;
1709: #endif
1710: PetscFunctionReturn(PETSC_SUCCESS);
1711: }
1713: static PetscErrorCode MatDenseUpdateColumnLayout_MPIDense(Mat A, PetscLayout clayout)
1714: {
1715: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1717: PetscFunctionBegin;
1718: if (A->cmap == clayout) PetscFunctionReturn(PETSC_SUCCESS);
1719: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1720: PetscCall(PetscLayoutReference(clayout, &A->cmap));
1721: PetscCall(MatDestroy(&a->cmat));
1722: PetscCall(PetscSFDestroy(&a->Mvctx));
1723: PetscFunctionReturn(PETSC_SUCCESS);
1724: }
1726: /*MC
1727: MATMPIDENSE - MATMPIDENSE = "mpidense" - A matrix type to be used for distributed dense matrices.
1729: Options Database Key:
1730: . -mat_type mpidense - sets the matrix type to `MATMPIDENSE` during a call to `MatSetFromOptions()`
1732: Level: beginner
1734: .seealso: [](ch_matrices), `Mat`, `MatCreateDense()`, `MATSEQDENSE`, `MATDENSE`
1735: M*/
1736: PetscErrorCode MatCreate_MPIDense(Mat mat)
1737: {
1738: Mat_MPIDense *a;
1740: PetscFunctionBegin;
1741: PetscCall(PetscNew(&a));
1742: mat->data = (void *)a;
1743: mat->ops[0] = MatOps_Values;
1745: mat->insertmode = NOT_SET_VALUES;
1747: /* build cache for off array entries formed */
1748: a->donotstash = PETSC_FALSE;
1750: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)mat), 1, &mat->stash));
1752: /* stuff used for matrix vector multiply */
1753: a->lvec = NULL;
1754: a->Mvctx = NULL;
1755: a->roworiented = PETSC_TRUE;
1757: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetLDA_C", MatDenseGetLDA_MPIDense));
1758: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseSetLDA_C", MatDenseSetLDA_MPIDense));
1759: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArray_C", MatDenseGetArray_MPIDense));
1760: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArray_C", MatDenseRestoreArray_MPIDense));
1761: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayRead_C", MatDenseGetArrayRead_MPIDense));
1762: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayRead_C", MatDenseRestoreArrayRead_MPIDense));
1763: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayWrite_C", MatDenseGetArrayWrite_MPIDense));
1764: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayWrite_C", MatDenseRestoreArrayWrite_MPIDense));
1765: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDensePlaceArray_C", MatDensePlaceArray_MPIDense));
1766: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseResetArray_C", MatDenseResetArray_MPIDense));
1767: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseReplaceArray_C", MatDenseReplaceArray_MPIDense));
1768: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVec_C", MatDenseGetColumnVec_MPIDense));
1769: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVec_C", MatDenseRestoreColumnVec_MPIDense));
1770: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecRead_C", MatDenseGetColumnVecRead_MPIDense));
1771: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecRead_C", MatDenseRestoreColumnVecRead_MPIDense));
1772: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecWrite_C", MatDenseGetColumnVecWrite_MPIDense));
1773: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecWrite_C", MatDenseRestoreColumnVecWrite_MPIDense));
1774: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetSubMatrix_C", MatDenseGetSubMatrix_MPIDense));
1775: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreSubMatrix_C", MatDenseRestoreSubMatrix_MPIDense));
1776: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
1777: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpiaij_C", MatConvert_MPIDense_MPIAIJ));
1778: #if PetscDefined(HAVE_ELEMENTAL)
1779: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_elemental_C", MatConvert_MPIDense_Elemental));
1780: #endif
1781: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
1782: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_scalapack_C", MatConvert_Dense_ScaLAPACK));
1783: #endif
1784: #if PetscDefined(HAVE_CUDA)
1785: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensecuda_C", MatConvert_MPIDense_MPIDenseCUDA));
1786: #endif
1787: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIDenseSetPreallocation_C", MatMPIDenseSetPreallocation_MPIDense));
1788: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpidense_C", MatProductSetFromOptions_MPIAIJ_MPIDense));
1789: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaij_C", MatProductSetFromOptions_MPIDense_MPIAIJ));
1790: #if PetscDefined(HAVE_CUDA)
1791: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijcusparse_mpidense_C", MatProductSetFromOptions_MPIAIJ_MPIDense));
1792: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijcusparse_C", MatProductSetFromOptions_MPIDense_MPIAIJ));
1793: #endif
1794: #if PetscDefined(HAVE_HIP)
1795: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensehip_C", MatConvert_MPIDense_MPIDenseHIP));
1796: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijhipsparse_mpidense_C", MatProductSetFromOptions_MPIAIJ_MPIDense));
1797: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijhipsparse_C", MatProductSetFromOptions_MPIDense_MPIAIJ));
1798: #endif
1799: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumn_C", MatDenseGetColumn_MPIDense));
1800: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumn_C", MatDenseRestoreColumn_MPIDense));
1801: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultColumnRange_C", MatMultColumnRange_MPIDense));
1802: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultAddColumnRange_C", MatMultAddColumnRange_MPIDense));
1803: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeColumnRange_C", MatMultHermitianTransposeColumnRange_MPIDense));
1804: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeAddColumnRange_C", MatMultHermitianTransposeAddColumnRange_MPIDense));
1805: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatGetMultPetscSF_C", MatGetMultPetscSF_MPIDense));
1806: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseUpdateColumnLayout_C", MatDenseUpdateColumnLayout_MPIDense));
1807: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, MATMPIDENSE));
1808: PetscFunctionReturn(PETSC_SUCCESS);
1809: }
1811: /*MC
1812: MATDENSE - MATDENSE = "dense" - A matrix type to be used for dense matrices.
1814: This matrix type is identical to `MATSEQDENSE` when constructed with a single process communicator,
1815: and `MATMPIDENSE` otherwise.
1817: Options Database Key:
1818: . -mat_type dense - sets the matrix type to `MATDENSE` during a call to `MatSetFromOptions()`
1820: Level: beginner
1822: .seealso: [](ch_matrices), `Mat`, `MATSEQDENSE`, `MATMPIDENSE`, `MATDENSECUDA`, `MATDENSEHIP`
1823: M*/
1825: /*@
1826: MatMPIDenseSetPreallocation - Sets the array used to store the matrix entries
1828: Collective
1830: Input Parameters:
1831: + B - the matrix
1832: - data - optional location of matrix data. Set to `NULL` for PETSc
1833: to control all matrix memory allocation.
1835: Level: intermediate
1837: Notes:
1838: The dense format is fully compatible with standard Fortran
1839: storage by columns.
1841: The data input variable is intended primarily for Fortran programmers
1842: who wish to allocate their own matrix memory space. Most users should
1843: set `data` to `NULL`.
1845: .seealso: [](ch_matrices), `Mat`, `MATMPIDENSE`, `MatCreate()`, `MatCreateSeqDense()`, `MatSetValues()`
1846: @*/
1847: PetscErrorCode MatMPIDenseSetPreallocation(Mat B, PetscScalar *data)
1848: {
1849: PetscFunctionBegin;
1851: PetscTryMethod(B, "MatMPIDenseSetPreallocation_C", (Mat, PetscScalar *), (B, data));
1852: PetscFunctionReturn(PETSC_SUCCESS);
1853: }
1855: /*@
1856: MatDensePlaceArray - Allows one to replace the array in a `MATDENSE` matrix with an
1857: array provided by the user. This is useful to avoid copying an array
1858: into a matrix
1860: Not Collective
1862: Input Parameters:
1863: + mat - the matrix
1864: - array - the array in column major order
1866: Level: developer
1868: Note:
1869: Adding `const` to `array` was an oversight, see notes in `VecPlaceArray()`.
1871: You can return to the original array with a call to `MatDenseResetArray()`. The user is responsible for freeing this array; it will not be
1872: freed when the matrix is destroyed.
1874: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArray()`, `MatDenseResetArray()`, `VecPlaceArray()`, `VecGetArray()`, `VecRestoreArray()`, `VecReplaceArray()`, `VecResetArray()`,
1875: `MatDenseReplaceArray()`
1876: @*/
1877: PetscErrorCode MatDensePlaceArray(Mat mat, const PetscScalar *array)
1878: {
1879: PetscFunctionBegin;
1881: PetscUseMethod(mat, "MatDensePlaceArray_C", (Mat, const PetscScalar *), (mat, array));
1882: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
1883: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
1884: mat->offloadmask = PETSC_OFFLOAD_CPU;
1885: #endif
1886: PetscFunctionReturn(PETSC_SUCCESS);
1887: }
1889: /*@
1890: MatDenseResetArray - Resets the matrix array to that it previously had before the call to `MatDensePlaceArray()`
1892: Not Collective
1894: Input Parameter:
1895: . mat - the matrix
1897: Level: developer
1899: Note:
1900: You can only call this after a call to `MatDensePlaceArray()`
1902: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArray()`, `MatDensePlaceArray()`, `VecPlaceArray()`, `VecGetArray()`, `VecRestoreArray()`, `VecReplaceArray()`, `VecResetArray()`
1903: @*/
1904: PetscErrorCode MatDenseResetArray(Mat mat)
1905: {
1906: PetscFunctionBegin;
1908: PetscUseMethod(mat, "MatDenseResetArray_C", (Mat), (mat));
1909: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
1910: PetscFunctionReturn(PETSC_SUCCESS);
1911: }
1913: /*@
1914: MatDenseReplaceArray - Allows one to replace the array in a dense matrix with an
1915: array provided by the user. This is useful to avoid copying an array
1916: into a matrix
1918: Not Collective
1920: Input Parameters:
1921: + mat - the matrix
1922: - array - the array in column major order
1924: Level: developer
1926: Note:
1927: Adding `const` to `array` was an oversight, see notes in `VecPlaceArray()`.
1929: The memory passed in MUST be obtained with `PetscMalloc()` and CANNOT be
1930: freed by the user. It will be freed when the matrix is destroyed.
1932: .seealso: [](ch_matrices), `Mat`, `MatDensePlaceArray()`, `MatDenseGetArray()`, `VecReplaceArray()`
1933: @*/
1934: PetscErrorCode MatDenseReplaceArray(Mat mat, const PetscScalar *array)
1935: {
1936: PetscFunctionBegin;
1938: PetscUseMethod(mat, "MatDenseReplaceArray_C", (Mat, const PetscScalar *), (mat, array));
1939: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
1940: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
1941: mat->offloadmask = PETSC_OFFLOAD_CPU;
1942: #endif
1943: PetscFunctionReturn(PETSC_SUCCESS);
1944: }
1946: /*@
1947: MatCreateDense - Creates a matrix in `MATDENSE` format.
1949: Collective
1951: Input Parameters:
1952: + comm - MPI communicator
1953: . m - number of local rows (or `PETSC_DECIDE` to have calculated if `M` is given)
1954: . n - number of local columns (or `PETSC_DECIDE` to have calculated if `N` is given)
1955: . M - number of global rows (or `PETSC_DECIDE` to have calculated if `m` is given)
1956: . N - number of global columns (or `PETSC_DECIDE` to have calculated if `n` is given)
1957: - data - optional location of matrix data. Set data to `NULL` (`PETSC_NULL_SCALAR_ARRAY` for Fortran users) for PETSc
1958: to control all matrix memory allocation.
1960: Output Parameter:
1961: . A - the matrix
1963: Level: intermediate
1965: Notes:
1966: The dense format is fully compatible with standard Fortran
1967: storage by columns.
1969: Although local portions of the matrix are stored in column-major
1970: order, the matrix is partitioned across MPI ranks by row.
1972: The data input variable is intended primarily for Fortran programmers
1973: who wish to allocate their own matrix memory space. Most users should
1974: set `data` to `NULL` (`PETSC_NULL_SCALAR_ARRAY` for Fortran users).
1976: The user MUST specify either the local or global matrix dimensions
1977: (possibly both).
1979: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatCreate()`, `MatCreateSeqDense()`, `MatSetValues()`
1980: @*/
1981: PetscErrorCode MatCreateDense(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscScalar data[], Mat *A)
1982: {
1983: PetscFunctionBegin;
1984: PetscCall(MatCreate(comm, A));
1985: PetscCall(MatSetSizes(*A, m, n, M, N));
1986: PetscCall(MatSetType(*A, MATDENSE));
1987: PetscCall(MatSeqDenseSetPreallocation(*A, data));
1988: PetscCall(MatMPIDenseSetPreallocation(*A, data));
1989: PetscFunctionReturn(PETSC_SUCCESS);
1990: }
1992: static PetscErrorCode MatDuplicate_MPIDense(Mat A, MatDuplicateOption cpvalues, Mat *newmat)
1993: {
1994: Mat mat;
1995: Mat_MPIDense *a, *oldmat = (Mat_MPIDense *)A->data;
1997: PetscFunctionBegin;
1998: *newmat = NULL;
1999: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &mat));
2000: PetscCall(MatSetSizes(mat, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
2001: PetscCall(MatSetType(mat, ((PetscObject)A)->type_name));
2002: a = (Mat_MPIDense *)mat->data;
2004: mat->factortype = A->factortype;
2005: mat->assembled = PETSC_TRUE;
2006: mat->preallocated = PETSC_TRUE;
2008: mat->insertmode = NOT_SET_VALUES;
2009: a->donotstash = oldmat->donotstash;
2011: PetscCall(PetscLayoutReference(A->rmap, &mat->rmap));
2012: PetscCall(PetscLayoutReference(A->cmap, &mat->cmap));
2014: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
2016: *newmat = mat;
2017: PetscFunctionReturn(PETSC_SUCCESS);
2018: }
2020: static PetscErrorCode MatLoad_MPIDense(Mat newMat, PetscViewer viewer)
2021: {
2022: PetscBool isbinary;
2023: #if PetscDefined(HAVE_HDF5)
2024: PetscBool ishdf5;
2025: #endif
2027: PetscFunctionBegin;
2030: /* force binary viewer to load .info file if it has not yet done so */
2031: PetscCall(PetscViewerSetUp(viewer));
2032: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
2033: #if PetscDefined(HAVE_HDF5)
2034: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
2035: #endif
2036: if (isbinary) {
2037: PetscCall(MatLoad_Dense_Binary(newMat, viewer));
2038: #if PetscDefined(HAVE_HDF5)
2039: } else if (ishdf5) {
2040: PetscCall(MatLoad_Dense_HDF5(newMat, viewer));
2041: #endif
2042: } 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);
2043: PetscFunctionReturn(PETSC_SUCCESS);
2044: }
2046: static PetscErrorCode MatEqual_MPIDense(Mat A, Mat B, PetscBool *flag)
2047: {
2048: Mat_MPIDense *matB = (Mat_MPIDense *)B->data, *matA = (Mat_MPIDense *)A->data;
2049: Mat a, b;
2051: PetscFunctionBegin;
2052: a = matA->A;
2053: b = matB->A;
2054: PetscCall(MatEqual(a, b, flag));
2055: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, flag, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2056: PetscFunctionReturn(PETSC_SUCCESS);
2057: }
2059: static PetscErrorCode MatProductCtxDestroy_MatTransMatMult_MPIDense_MPIDense(PetscCtxRt data)
2060: {
2061: MatProductCtx_TransMatMultDense *atb = *(MatProductCtx_TransMatMultDense **)data;
2063: PetscFunctionBegin;
2064: PetscCall(PetscFree2(atb->sendbuf, atb->recvcounts));
2065: PetscCall(MatDestroy(&atb->atb));
2066: PetscCall(PetscFree(atb));
2067: PetscFunctionReturn(PETSC_SUCCESS);
2068: }
2070: static PetscErrorCode MatProductCtxDestroy_MatMatTransMult_MPIDense_MPIDense(PetscCtxRt data)
2071: {
2072: MatProductCtx_MatTransMultDense *abt = *(MatProductCtx_MatTransMultDense **)data;
2074: PetscFunctionBegin;
2075: PetscCall(PetscFree2(abt->buf[0], abt->buf[1]));
2076: PetscCall(PetscFree2(abt->recvcounts, abt->recvdispls));
2077: PetscCall(PetscFree(abt));
2078: PetscFunctionReturn(PETSC_SUCCESS);
2079: }
2081: static PetscErrorCode MatTransposeMatMultNumeric_MPIDense_MPIDense(Mat A, Mat B, Mat C)
2082: {
2083: Mat_MPIDense *a = (Mat_MPIDense *)A->data, *b = (Mat_MPIDense *)B->data, *c = (Mat_MPIDense *)C->data;
2084: MatProductCtx_TransMatMultDense *atb;
2085: MPI_Comm comm;
2086: PetscMPIInt size, *recvcounts;
2087: PetscScalar *carray, *sendbuf;
2088: const PetscScalar *atbarray;
2089: PetscInt i, cN = C->cmap->N, proc, k, j, lda;
2090: const PetscInt *ranges;
2092: PetscFunctionBegin;
2093: MatCheckProduct(C, 3);
2094: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2095: atb = (MatProductCtx_TransMatMultDense *)C->product->data;
2096: recvcounts = atb->recvcounts;
2097: sendbuf = atb->sendbuf;
2099: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2100: PetscCallMPI(MPI_Comm_size(comm, &size));
2102: /* compute atbarray = aseq^T * bseq */
2103: PetscCall(MatTransposeMatMult(a->A, b->A, atb->atb ? MAT_REUSE_MATRIX : MAT_INITIAL_MATRIX, PETSC_DETERMINE, &atb->atb));
2105: PetscCall(MatGetOwnershipRanges(C, &ranges));
2107: if (ranges[1] == C->rmap->N) {
2108: /* all of the values are being reduced to rank 0: optimize this case to use MPI_Reduce and GPU aware MPI if available */
2109: PetscInt atb_lda, c_lda;
2110: Mat atb_local = atb->atb;
2111: Mat atb_alloc = NULL;
2112: Mat c_local = c->A;
2113: Mat c_alloc = NULL;
2114: PetscMemType atb_memtype, c_memtype;
2115: const PetscScalar *atb_array = NULL;
2116: MPI_Datatype vector_type;
2117: PetscScalar *c_array = NULL;
2118: PetscMPIInt rank;
2120: PetscCallMPI(MPI_Comm_rank(comm, &rank));
2122: PetscCall(MatDenseGetLDA(atb_local, &atb_lda));
2123: if (atb_lda != C->rmap->N) {
2124: // copy atb to a matrix that will have lda == the number of rows
2125: PetscCall(MatDuplicate(atb_local, MAT_DO_NOT_COPY_VALUES, &atb_alloc));
2126: PetscCall(MatCopy(atb_local, atb_alloc, DIFFERENT_NONZERO_PATTERN));
2127: atb_local = atb_alloc;
2128: }
2130: if (rank == 0) {
2131: PetscCall(MatDenseGetLDA(c_local, &c_lda));
2132: if (c_lda != C->rmap->N) {
2133: // copy c to a matrix that will have lda == the number of rows
2134: PetscCall(MatDuplicate(c_local, MAT_DO_NOT_COPY_VALUES, &c_alloc));
2135: c_local = c_alloc;
2136: }
2137: PetscCall(MatZeroEntries(c_local));
2138: }
2139: /* atb_local and c_local have nrows = lda = A->cmap->N and ncols =
2140: * B->cmap->N: use the a->Mvctx to use the best reduction method */
2141: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
2142: vector_type = MPIU_SCALAR;
2143: if (B->cmap->N > 1) {
2144: PetscMPIInt mpi_N;
2146: PetscCall(PetscMPIIntCast(B->cmap->N, &mpi_N));
2147: PetscCallMPI(MPI_Type_contiguous(mpi_N, MPIU_SCALAR, &vector_type));
2148: PetscCallMPI(MPI_Type_commit(&vector_type));
2149: }
2150: PetscCall(MatDenseGetArrayReadAndMemType(atb_local, &atb_array, &atb_memtype));
2151: PetscCall(MatDenseGetArrayWriteAndMemType(c_local, &c_array, &c_memtype));
2152: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, vector_type, atb_memtype, atb_array, c_memtype, c_array, MPIU_SUM));
2153: PetscCall(PetscSFReduceEnd(a->Mvctx, vector_type, atb_array, c_array, MPIU_SUM));
2154: PetscCall(MatDenseRestoreArrayWriteAndMemType(c_local, &c_array));
2155: PetscCall(MatDenseRestoreArrayReadAndMemType(atb_local, &atb_array));
2156: if (rank == 0 && c_local != c->A) PetscCall(MatCopy(c_local, c->A, DIFFERENT_NONZERO_PATTERN));
2157: if (B->cmap->N > 1) PetscCallMPI(MPI_Type_free(&vector_type));
2158: PetscCall(MatDestroy(&atb_alloc));
2159: PetscCall(MatDestroy(&c_alloc));
2160: PetscFunctionReturn(PETSC_SUCCESS);
2161: }
2163: /* arrange atbarray into sendbuf */
2164: PetscCall(MatDenseGetArrayRead(atb->atb, &atbarray));
2165: PetscCall(MatDenseGetLDA(atb->atb, &lda));
2166: for (proc = 0, k = 0; proc < size; proc++) {
2167: for (j = 0; j < cN; j++) {
2168: for (i = ranges[proc]; i < ranges[proc + 1]; i++) sendbuf[k++] = atbarray[i + j * lda];
2169: }
2170: }
2171: PetscCall(MatDenseRestoreArrayRead(atb->atb, &atbarray));
2173: /* sum all atbarray to local values of C */
2174: PetscCall(MatDenseGetLDA(c->A, &lda));
2175: PetscCall(MatDenseGetArrayWrite(c->A, &carray));
2176: if (lda > C->rmap->n && cN > 1 && C->rmap->n) {
2177: PetscScalar *recvbuf;
2179: /* the local product is no longer needed after packing sendbuf; reuse its storage for the contiguous reduction */
2180: PetscCall(MatDenseGetArrayWrite(atb->atb, &recvbuf));
2181: PetscCallMPI(MPI_Reduce_scatter(sendbuf, recvbuf, recvcounts, MPIU_SCALAR, MPIU_SUM, comm));
2182: for (j = 0; j < cN; j++) PetscCall(PetscArraycpy(carray + (size_t)j * lda, recvbuf + (size_t)j * C->rmap->n, C->rmap->n));
2183: PetscCall(MatDenseRestoreArrayWrite(atb->atb, &recvbuf));
2184: } else PetscCallMPI(MPI_Reduce_scatter(sendbuf, carray, recvcounts, MPIU_SCALAR, MPIU_SUM, comm));
2185: PetscCall(MatDenseRestoreArrayWrite(c->A, &carray));
2186: PetscFunctionReturn(PETSC_SUCCESS);
2187: }
2189: static PetscErrorCode MatTransposeMatMultSymbolic_MPIDense_MPIDense(Mat A, Mat B, PetscReal fill, Mat C)
2190: {
2191: MPI_Comm comm;
2192: PetscMPIInt size;
2193: PetscInt cm = A->cmap->n, cM, cN = B->cmap->N;
2194: MatProductCtx_TransMatMultDense *atb;
2195: PetscBool cisdense = PETSC_FALSE;
2196: const PetscInt *ranges;
2198: PetscFunctionBegin;
2199: MatCheckProduct(C, 4);
2200: PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
2201: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2202: 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,
2203: A->rmap->rend, B->rmap->rstart, B->rmap->rend);
2205: /* create matrix product C */
2206: PetscCall(MatSetSizes(C, cm, B->cmap->n, A->cmap->N, B->cmap->N));
2207: #if PetscDefined(HAVE_CUDA)
2208: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, ""));
2209: #endif
2210: #if PetscDefined(HAVE_HIP)
2211: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSEHIP, ""));
2212: #endif
2213: if (!cisdense) {
2214: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2215: PetscCall(MatSetVecType(C, A->defaultvectype));
2216: }
2217: PetscCall(MatSetUp(C));
2219: /* create data structure for reuse C */
2220: PetscCallMPI(MPI_Comm_size(comm, &size));
2221: PetscCall(PetscNew(&atb));
2222: cM = C->rmap->N;
2223: PetscCall(PetscMalloc2(cM * cN, &atb->sendbuf, size, &atb->recvcounts));
2224: PetscCall(MatGetOwnershipRanges(C, &ranges));
2225: for (PetscMPIInt i = 0; i < size; i++) PetscCall(PetscMPIIntCast((ranges[i + 1] - ranges[i]) * cN, &atb->recvcounts[i]));
2226: C->product->data = atb;
2227: C->product->destroy = MatProductCtxDestroy_MatTransMatMult_MPIDense_MPIDense;
2228: PetscFunctionReturn(PETSC_SUCCESS);
2229: }
2231: static PetscErrorCode MatMatTransposeMultSymbolic_MPIDense_MPIDense(Mat A, Mat B, PetscReal fill, Mat C)
2232: {
2233: MPI_Comm comm;
2234: PetscMPIInt i, size;
2235: PetscInt maxRows, bufsiz;
2236: PetscMPIInt tag;
2237: PetscInt alg;
2238: MatProductCtx_MatTransMultDense *abt;
2239: Mat_Product *product = C->product;
2240: PetscBool flg;
2242: PetscFunctionBegin;
2243: MatCheckProduct(C, 4);
2244: PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
2245: /* check local size of A and B */
2246: 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);
2248: PetscCall(PetscStrcmp(product->alg, "allgatherv", &flg));
2249: alg = flg ? 0 : 1;
2251: /* setup matrix product C */
2252: PetscCall(MatSetSizes(C, A->rmap->n, B->rmap->n, A->rmap->N, B->rmap->N));
2253: PetscCall(MatSetType(C, MATMPIDENSE));
2254: /* keep the VecType of A, e.g. VECKOKKOS from MatCreateDenseFromVecType(), but only if A is a MATMPIDENSE like C, since the VecType of a MATMPIDENSECUDA or MATMPIDENSEHIP A, e.g. VECCUDA, is not valid for C */
2255: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIDENSE, &flg));
2256: if (flg) PetscCall(MatSetVecType(C, A->defaultvectype));
2257: PetscCall(MatSetUp(C));
2258: PetscCall(PetscObjectGetNewTag((PetscObject)C, &tag));
2260: /* create data structure for reuse C */
2261: PetscCall(PetscObjectGetComm((PetscObject)C, &comm));
2262: PetscCallMPI(MPI_Comm_size(comm, &size));
2263: PetscCall(PetscNew(&abt));
2264: abt->tag = tag;
2265: abt->alg = alg;
2266: switch (alg) {
2267: case 1: /* alg: "cyclic" */
2268: for (maxRows = 0, i = 0; i < size; i++) maxRows = PetscMax(maxRows, B->rmap->range[i + 1] - B->rmap->range[i]);
2269: bufsiz = A->cmap->N * maxRows;
2270: PetscCall(PetscMalloc2(bufsiz, &abt->buf[0], bufsiz, &abt->buf[1]));
2271: break;
2272: default: /* alg: "allgatherv" */
2273: PetscCall(PetscMalloc2(B->rmap->n * B->cmap->N, &abt->buf[0], B->rmap->N * B->cmap->N, &abt->buf[1]));
2274: PetscCall(PetscMalloc2(size, &abt->recvcounts, size + 1, &abt->recvdispls));
2275: for (i = 0; i <= size; i++) PetscCall(PetscMPIIntCast(B->rmap->range[i] * A->cmap->N, &abt->recvdispls[i]));
2276: for (i = 0; i < size; i++) PetscCall(PetscMPIIntCast(abt->recvdispls[i + 1] - abt->recvdispls[i], &abt->recvcounts[i]));
2277: break;
2278: }
2280: C->product->data = abt;
2281: C->product->destroy = MatProductCtxDestroy_MatMatTransMult_MPIDense_MPIDense;
2282: C->ops->mattransposemultnumeric = MatMatTransposeMultNumeric_MPIDense_MPIDense;
2283: PetscFunctionReturn(PETSC_SUCCESS);
2284: }
2286: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense_Cyclic(Mat A, Mat B, Mat C)
2287: {
2288: Mat_MPIDense *a = (Mat_MPIDense *)A->data, *b = (Mat_MPIDense *)B->data, *c = (Mat_MPIDense *)C->data;
2289: MatProductCtx_MatTransMultDense *abt;
2290: MPI_Comm comm;
2291: PetscMPIInt rank, size, sendto, recvfrom, recvisfrom;
2292: PetscScalar *sendbuf, *recvbuf = NULL, *cv;
2293: PetscInt i, cK = A->cmap->N, sendsiz, recvsiz, k, j, bn;
2294: PetscScalar _DOne = 1.0, _DZero = 0.0;
2295: const PetscScalar *av, *bv;
2296: PetscBLASInt cm, cn, ck, alda, blda = 0, clda;
2297: MPI_Request reqs[2];
2298: const PetscInt *ranges;
2300: PetscFunctionBegin;
2301: MatCheckProduct(C, 3);
2302: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2303: abt = (MatProductCtx_MatTransMultDense *)C->product->data;
2304: PetscCall(PetscObjectGetComm((PetscObject)C, &comm));
2305: PetscCallMPI(MPI_Comm_rank(comm, &rank));
2306: PetscCallMPI(MPI_Comm_size(comm, &size));
2307: PetscCall(MatDenseGetArrayRead(a->A, &av));
2308: PetscCall(MatDenseGetArrayRead(b->A, &bv));
2309: PetscCall(MatDenseGetArrayWrite(c->A, &cv));
2310: PetscCall(MatDenseGetLDA(a->A, &i));
2311: PetscCall(PetscBLASIntCast(i, &alda));
2312: PetscCall(MatDenseGetLDA(b->A, &i));
2313: PetscCall(PetscBLASIntCast(i, &blda));
2314: PetscCall(MatDenseGetLDA(c->A, &i));
2315: PetscCall(PetscBLASIntCast(i, &clda));
2316: PetscCall(MatGetOwnershipRanges(B, &ranges));
2317: bn = B->rmap->n;
2318: if (blda == bn) {
2319: sendbuf = (PetscScalar *)bv;
2320: } else {
2321: sendbuf = abt->buf[0];
2322: for (k = 0, i = 0; i < cK; i++) {
2323: for (j = 0; j < bn; j++, k++) sendbuf[k] = bv[i * blda + j];
2324: }
2325: }
2326: if (size > 1) {
2327: sendto = (rank + size - 1) % size;
2328: recvfrom = (rank + size + 1) % size;
2329: } else {
2330: sendto = recvfrom = 0;
2331: }
2332: PetscCall(PetscBLASIntCast(cK, &ck));
2333: PetscCall(PetscBLASIntCast(c->A->rmap->n, &cm));
2334: recvisfrom = rank;
2335: for (i = 0; i < size; i++) {
2336: /* we have finished receiving in sending, bufs can be read/modified */
2337: PetscMPIInt nextrecvisfrom = (recvisfrom + 1) % size; /* which process the next recvbuf will originate on */
2338: PetscInt nextbn = ranges[nextrecvisfrom + 1] - ranges[nextrecvisfrom];
2340: if (nextrecvisfrom != rank) {
2341: /* start the cyclic sends from sendbuf, to recvbuf (which will switch to sendbuf) */
2342: sendsiz = cK * bn;
2343: recvsiz = cK * nextbn;
2344: recvbuf = (i & 1) ? abt->buf[0] : abt->buf[1];
2345: PetscCallMPI(MPIU_Isend(sendbuf, sendsiz, MPIU_SCALAR, sendto, abt->tag, comm, &reqs[0]));
2346: PetscCallMPI(MPIU_Irecv(recvbuf, recvsiz, MPIU_SCALAR, recvfrom, abt->tag, comm, &reqs[1]));
2347: }
2349: /* local aseq * sendbuf^T */
2350: PetscCall(PetscBLASIntCast(ranges[recvisfrom + 1] - ranges[recvisfrom], &cn));
2351: if (cm && cn && ck) PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &cm, &cn, &ck, &_DOne, av, &alda, sendbuf, &cn, &_DZero, cv + clda * ranges[recvisfrom], &clda));
2353: if (nextrecvisfrom != rank) {
2354: /* wait for the sends and receives to complete, swap sendbuf and recvbuf */
2355: PetscCallMPI(MPI_Waitall(2, reqs, MPI_STATUSES_IGNORE));
2356: }
2357: bn = nextbn;
2358: recvisfrom = nextrecvisfrom;
2359: sendbuf = recvbuf;
2360: }
2361: PetscCall(MatDenseRestoreArrayRead(a->A, &av));
2362: PetscCall(MatDenseRestoreArrayRead(b->A, &bv));
2363: PetscCall(MatDenseRestoreArrayWrite(c->A, &cv));
2364: PetscFunctionReturn(PETSC_SUCCESS);
2365: }
2367: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense_Allgatherv(Mat A, Mat B, Mat C)
2368: {
2369: Mat_MPIDense *a = (Mat_MPIDense *)A->data, *b = (Mat_MPIDense *)B->data, *c = (Mat_MPIDense *)C->data;
2370: MatProductCtx_MatTransMultDense *abt;
2371: MPI_Comm comm;
2372: PetscMPIInt size, ibn;
2373: PetscScalar *cv, *sendbuf, *recvbuf;
2374: const PetscScalar *av, *bv;
2375: PetscInt blda, i, cK = A->cmap->N, k, j, bn;
2376: PetscScalar _DOne = 1.0, _DZero = 0.0;
2377: PetscBLASInt cm, cn, ck, alda, clda;
2379: PetscFunctionBegin;
2380: MatCheckProduct(C, 3);
2381: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2382: abt = (MatProductCtx_MatTransMultDense *)C->product->data;
2383: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2384: PetscCallMPI(MPI_Comm_size(comm, &size));
2385: PetscCall(MatDenseGetArrayRead(a->A, &av));
2386: PetscCall(MatDenseGetArrayRead(b->A, &bv));
2387: PetscCall(MatDenseGetArrayWrite(c->A, &cv));
2388: PetscCall(MatDenseGetLDA(a->A, &i));
2389: PetscCall(PetscBLASIntCast(i, &alda));
2390: PetscCall(MatDenseGetLDA(b->A, &blda));
2391: PetscCall(MatDenseGetLDA(c->A, &i));
2392: PetscCall(PetscBLASIntCast(i, &clda));
2393: /* copy transpose of B into buf[0] */
2394: bn = B->rmap->n;
2395: sendbuf = abt->buf[0];
2396: recvbuf = abt->buf[1];
2397: for (k = 0, j = 0; j < bn; j++) {
2398: for (i = 0; i < cK; i++, k++) sendbuf[k] = bv[i * blda + j];
2399: }
2400: PetscCall(MatDenseRestoreArrayRead(b->A, &bv));
2401: PetscCall(PetscMPIIntCast(bn * cK, &ibn));
2402: PetscCallMPI(MPI_Allgatherv(sendbuf, ibn, MPIU_SCALAR, recvbuf, abt->recvcounts, abt->recvdispls, MPIU_SCALAR, comm));
2403: PetscCall(PetscBLASIntCast(cK, &ck));
2404: PetscCall(PetscBLASIntCast(c->A->rmap->n, &cm));
2405: PetscCall(PetscBLASIntCast(c->A->cmap->n, &cn));
2406: if (cm && cn && ck) PetscCallBLAS("BLASgemm", BLASgemm_("N", "N", &cm, &cn, &ck, &_DOne, av, &alda, recvbuf, &ck, &_DZero, cv, &clda));
2407: PetscCall(MatDenseRestoreArrayRead(a->A, &av));
2408: PetscCall(MatDenseRestoreArrayRead(b->A, &bv));
2409: PetscCall(MatDenseRestoreArrayWrite(c->A, &cv));
2410: PetscFunctionReturn(PETSC_SUCCESS);
2411: }
2413: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense(Mat A, Mat B, Mat C)
2414: {
2415: MatProductCtx_MatTransMultDense *abt;
2417: PetscFunctionBegin;
2418: MatCheckProduct(C, 3);
2419: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2420: abt = (MatProductCtx_MatTransMultDense *)C->product->data;
2421: switch (abt->alg) {
2422: case 1:
2423: PetscCall(MatMatTransposeMultNumeric_MPIDense_MPIDense_Cyclic(A, B, C));
2424: break;
2425: default:
2426: PetscCall(MatMatTransposeMultNumeric_MPIDense_MPIDense_Allgatherv(A, B, C));
2427: break;
2428: }
2429: PetscFunctionReturn(PETSC_SUCCESS);
2430: }
2432: static PetscErrorCode MatProductCtxDestroy_MatMatMult_MPIDense_MPIDense(PetscCtxRt data)
2433: {
2434: MatProductCtx_MatMultDense *ab = *(MatProductCtx_MatMultDense **)data;
2436: PetscFunctionBegin;
2437: PetscCall(MatDestroy(&ab->Ce));
2438: PetscCall(MatDestroy(&ab->Ae));
2439: PetscCall(MatDestroy(&ab->Be));
2440: PetscCall(PetscFree(ab));
2441: PetscFunctionReturn(PETSC_SUCCESS);
2442: }
2444: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIDense(Mat A, Mat B, Mat C)
2445: {
2446: MatProductCtx_MatMultDense *ab;
2447: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
2448: Mat_MPIDense *b = (Mat_MPIDense *)B->data;
2450: PetscFunctionBegin;
2451: MatCheckProduct(C, 3);
2452: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Missing product data");
2453: ab = (MatProductCtx_MatMultDense *)C->product->data;
2454: if (ab->Ae && ab->Ce) {
2455: #if PetscDefined(HAVE_ELEMENTAL)
2456: PetscCall(MatConvert_MPIDense_Elemental(A, MATELEMENTAL, MAT_REUSE_MATRIX, &ab->Ae));
2457: PetscCall(MatConvert_MPIDense_Elemental(B, MATELEMENTAL, MAT_REUSE_MATRIX, &ab->Be));
2458: PetscCall(MatMatMultNumeric_Elemental(ab->Ae, ab->Be, ab->Ce));
2459: PetscCall(MatConvert(ab->Ce, MATMPIDENSE, MAT_REUSE_MATRIX, &C));
2460: #else
2461: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "PETSC_HAVE_ELEMENTAL not defined");
2462: #endif
2463: } else {
2464: MPI_Comm comm;
2465: Mat clocal;
2466: const PetscInt *ranges;
2467: PetscMPIInt size;
2469: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A)); /* cannot be done during the symbolic phase because of possible calls to MatProductReplaceMats() */
2470: comm = PetscObjectComm((PetscObject)B);
2471: PetscCallMPI(MPI_Comm_size(comm, &size));
2472: PetscCall(PetscLayoutGetRanges(B->rmap, &ranges));
2473: if (ranges[1] == ranges[size]) {
2474: // optimize for the case where the B matrix is broadcast from rank 0
2475: PetscInt b_lda, be_lda;
2476: Mat b_local = b->A;
2477: Mat b_alloc = NULL;
2478: Mat be_local = ab->Be;
2479: Mat be_alloc = NULL;
2480: PetscMemType b_memtype, be_memtype;
2481: const PetscScalar *b_array = NULL;
2482: MPI_Datatype vector_type;
2483: PetscScalar *be_array = NULL;
2484: PetscMPIInt rank;
2486: PetscCallMPI(MPI_Comm_rank(comm, &rank));
2487: PetscCall(MatDenseGetLDA(be_local, &be_lda));
2488: if (be_lda != B->rmap->N) {
2489: PetscCall(MatDuplicate(be_local, MAT_DO_NOT_COPY_VALUES, &be_alloc));
2490: be_local = be_alloc;
2491: }
2493: if (rank == 0) {
2494: PetscCall(MatDenseGetLDA(b_local, &b_lda));
2495: if (b_lda != B->rmap->N) {
2496: PetscCall(MatDuplicate(b_local, MAT_DO_NOT_COPY_VALUES, &b_alloc));
2497: PetscCall(MatCopy(b_local, b_alloc, DIFFERENT_NONZERO_PATTERN));
2498: b_local = b_alloc;
2499: }
2500: }
2501: vector_type = MPIU_SCALAR;
2502: if (B->cmap->N > 1) {
2503: PetscMPIInt mpi_N;
2505: PetscCall(PetscMPIIntCast(B->cmap->N, &mpi_N));
2506: PetscCallMPI(MPI_Type_contiguous(mpi_N, MPIU_SCALAR, &vector_type));
2507: PetscCallMPI(MPI_Type_commit(&vector_type));
2508: }
2509: PetscCall(MatDenseGetArrayReadAndMemType(b_local, &b_array, &b_memtype));
2510: PetscCall(MatDenseGetArrayWriteAndMemType(be_local, &be_array, &be_memtype));
2511: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, vector_type, b_memtype, b_array, be_memtype, be_array, MPI_REPLACE));
2512: PetscCall(PetscSFBcastEnd(mdn->Mvctx, vector_type, b_array, be_array, MPI_REPLACE));
2513: PetscCall(MatDenseRestoreArrayWriteAndMemType(be_local, &be_array));
2514: PetscCall(MatDenseRestoreArrayReadAndMemType(b_local, &b_array));
2515: if (be_local != ab->Be) PetscCall(MatCopy(be_local, ab->Be, DIFFERENT_NONZERO_PATTERN));
2516: if (B->cmap->N > 1) PetscCallMPI(MPI_Type_free(&vector_type));
2517: PetscCall(MatDestroy(&be_alloc));
2518: PetscCall(MatDestroy(&b_alloc));
2519: } else PetscCall(MatDenseScatter_Private(mdn->Mvctx, B, ab->Be, INSERT_VALUES, SCATTER_FORWARD));
2520: PetscCall(MatDenseGetLocalMatrix(C, &clocal));
2521: PetscCall((*clocal->ops->matmultnumeric)(mdn->A, ab->Be, clocal));
2522: }
2523: PetscFunctionReturn(PETSC_SUCCESS);
2524: }
2526: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIDense(Mat A, Mat B, PetscReal fill, Mat C)
2527: {
2528: Mat_Product *product = C->product;
2529: PetscInt alg;
2530: MatProductCtx_MatMultDense *ab;
2531: PetscBool flg;
2532: VecType vtype;
2534: PetscFunctionBegin;
2535: MatCheckProduct(C, 4);
2536: PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
2537: /* check local size of A and B */
2538: 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 ")",
2539: A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend);
2541: PetscCall(PetscStrcmp(product->alg, "petsc", &flg));
2542: alg = flg ? 0 : 1;
2544: /* setup C */
2545: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
2546: PetscCall(PetscObjectBaseTypeCompare((PetscObject)C, MATMPIDENSE, &flg));
2547: if (!flg) {
2548: PetscCall(PetscObjectTypeCompare((PetscObject)B, ((PetscObject)A)->type_name, &flg));
2549: PetscCall(MatSetType(C, flg ? ((PetscObject)A)->type_name : MATMPIDENSE));
2550: }
2551: /* preserve a supplied dense result and its storage; copy the VecType only from a matching MatType */
2552: PetscCall(PetscObjectTypeCompare((PetscObject)A, ((PetscObject)C)->type_name, &flg));
2553: if (flg) PetscCall(MatSetVecType(C, A->defaultvectype));
2554: PetscCall(MatSetUp(C));
2556: /* create data structure for reuse Cdense */
2557: PetscCall(PetscNew(&ab));
2559: switch (alg) {
2560: case 1: /* alg: "elemental" */
2561: #if PetscDefined(HAVE_ELEMENTAL)
2562: /* create elemental matrices Ae and Be */
2563: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &ab->Ae));
2564: PetscCall(MatSetSizes(ab->Ae, PETSC_DECIDE, PETSC_DECIDE, A->rmap->N, A->cmap->N));
2565: PetscCall(MatSetType(ab->Ae, MATELEMENTAL));
2566: PetscCall(MatSetUp(ab->Ae));
2567: PetscCall(MatSetOption(ab->Ae, MAT_ROW_ORIENTED, PETSC_FALSE));
2569: PetscCall(MatCreate(PetscObjectComm((PetscObject)B), &ab->Be));
2570: PetscCall(MatSetSizes(ab->Be, PETSC_DECIDE, PETSC_DECIDE, B->rmap->N, B->cmap->N));
2571: PetscCall(MatSetType(ab->Be, MATELEMENTAL));
2572: PetscCall(MatSetUp(ab->Be));
2573: PetscCall(MatSetOption(ab->Be, MAT_ROW_ORIENTED, PETSC_FALSE));
2575: /* compute symbolic Ce = Ae*Be */
2576: PetscCall(MatCreate(PetscObjectComm((PetscObject)C), &ab->Ce));
2577: PetscCall(MatMatMultSymbolic_Elemental(ab->Ae, ab->Be, fill, ab->Ce));
2578: #else
2579: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "PETSC_HAVE_ELEMENTAL not defined");
2580: #endif
2581: break;
2582: default: /* alg: "petsc" */
2583: ab->Ae = NULL;
2584: PetscCall(MatGetVecType(C, &vtype));
2585: PetscCall(MatCreateDenseFromVecType(PETSC_COMM_SELF, vtype, A->cmap->N, B->cmap->N, A->cmap->N, B->cmap->N, PETSC_DECIDE, NULL, &ab->Be));
2586: ab->Ce = NULL;
2587: break;
2588: }
2590: C->product->data = ab;
2591: C->product->destroy = MatProductCtxDestroy_MatMatMult_MPIDense_MPIDense;
2592: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIDense;
2593: PetscFunctionReturn(PETSC_SUCCESS);
2594: }
2596: static PetscErrorCode MatProductSetFromOptions_MPIDense_AB(Mat C)
2597: {
2598: Mat_Product *product = C->product;
2599: const char *algTypes[2] = {"petsc", "elemental"};
2600: PetscInt alg, nalg = PetscDefined(HAVE_ELEMENTAL) ? 2 : 1;
2601: PetscBool flg = PETSC_FALSE;
2603: PetscFunctionBegin;
2604: /* Set default algorithm */
2605: alg = 0; /* default is PETSc */
2606: PetscCall(PetscStrcmp(product->alg, "default", &flg));
2607: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2609: /* Get runtime option */
2610: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
2611: PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatProduct_AB", algTypes, nalg, algTypes[alg], &alg, &flg));
2612: PetscOptionsEnd();
2613: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2615: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIDense;
2616: C->ops->productsymbolic = MatProductSymbolic_AB;
2617: PetscFunctionReturn(PETSC_SUCCESS);
2618: }
2620: static PetscErrorCode MatProductSetFromOptions_MPIDense_AtB(Mat C)
2621: {
2622: Mat_Product *product = C->product;
2623: Mat A = product->A, B = product->B;
2625: PetscFunctionBegin;
2626: 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 ")",
2627: A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend);
2628: C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_MPIDense_MPIDense;
2629: C->ops->productsymbolic = MatProductSymbolic_AtB;
2630: PetscFunctionReturn(PETSC_SUCCESS);
2631: }
2633: static PetscErrorCode MatProductSetFromOptions_MPIDense_ABt(Mat C)
2634: {
2635: Mat_Product *product = C->product;
2636: const char *algTypes[2] = {"allgatherv", "cyclic"};
2637: PetscInt alg, nalg = 2;
2638: PetscBool flg = PETSC_FALSE;
2640: PetscFunctionBegin;
2641: /* Set default algorithm */
2642: alg = 0; /* default is allgatherv */
2643: PetscCall(PetscStrcmp(product->alg, "default", &flg));
2644: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2646: /* Get runtime option */
2647: if (product->api_user) {
2648: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatTransposeMult", "Mat");
2649: PetscCall(PetscOptionsEList("-matmattransmult_mpidense_mpidense_via", "Algorithmic approach", "MatMatTransposeMult", algTypes, nalg, algTypes[alg], &alg, &flg));
2650: PetscOptionsEnd();
2651: } else {
2652: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_ABt", "Mat");
2653: PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatProduct_ABt", algTypes, nalg, algTypes[alg], &alg, &flg));
2654: PetscOptionsEnd();
2655: }
2656: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2658: C->ops->mattransposemultsymbolic = MatMatTransposeMultSymbolic_MPIDense_MPIDense;
2659: C->ops->productsymbolic = MatProductSymbolic_ABt;
2660: PetscFunctionReturn(PETSC_SUCCESS);
2661: }
2663: static PetscErrorCode MatProductSetFromOptions_MPIDense(Mat C)
2664: {
2665: Mat_Product *product = C->product;
2667: PetscFunctionBegin;
2668: switch (product->type) {
2669: case MATPRODUCT_AB:
2670: PetscCall(MatProductSetFromOptions_MPIDense_AB(C));
2671: break;
2672: case MATPRODUCT_AtB:
2673: PetscCall(MatProductSetFromOptions_MPIDense_AtB(C));
2674: break;
2675: case MATPRODUCT_ABt:
2676: PetscCall(MatProductSetFromOptions_MPIDense_ABt(C));
2677: break;
2678: default:
2679: break;
2680: }
2681: PetscFunctionReturn(PETSC_SUCCESS);
2682: }
2684: typedef struct {
2685: PetscSF sf;
2686: PetscInt N, rld, lld, nr, nl;
2687: } MatDenseScatterCtx;
2689: static PetscErrorCode MatDenseScatterCtxDestroy_Private(PetscCtxRt ptr)
2690: {
2691: MatDenseScatterCtx **ctx = (MatDenseScatterCtx **)ptr;
2693: PetscFunctionBegin;
2694: PetscCall(PetscSFDestroy(&(*ctx)->sf));
2695: PetscCall(PetscFree(*ctx));
2696: PetscFunctionReturn(PETSC_SUCCESS);
2697: }
2699: // The communication of MatDenseScatter_Private(), for a caller that already has a PetscSF for the layouts of X and Y
2700: PetscErrorCode MatDenseScatterWithSF_Private(PetscSF vsf, Mat X, Mat Y, MPI_Op op, ScatterMode smode, PetscBool writeonly)
2701: {
2702: const PetscScalar *in;
2703: PetscScalar *out;
2704: PetscMemType mtype[2];
2706: PetscFunctionBegin;
2707: PetscCall(MatDenseGetArrayReadAndMemType(X, &in, &mtype[0]));
2708: if (writeonly == PETSC_TRUE) PetscCall(MatDenseGetArrayWriteAndMemType(Y, &out, &mtype[1]));
2709: else PetscCall(MatDenseGetArrayAndMemType(Y, &out, &mtype[1]));
2710: if (smode == SCATTER_FORWARD) {
2711: PetscCall(PetscSFBcastWithMemTypeBegin(vsf, vsf->vscat.unit, mtype[0], in, mtype[1], out, op));
2712: PetscCall(PetscSFBcastEnd(vsf, vsf->vscat.unit, in, out, op));
2713: } else {
2714: PetscCall(PetscSFReduceWithMemTypeBegin(vsf, vsf->vscat.unit, mtype[0], in, mtype[1], out, op));
2715: PetscCall(PetscSFReduceEnd(vsf, vsf->vscat.unit, in, out, op));
2716: }
2717: if (writeonly == PETSC_TRUE) PetscCall(MatDenseRestoreArrayWriteAndMemType(Y, &out));
2718: else PetscCall(MatDenseRestoreArrayAndMemType(Y, &out));
2719: PetscCall(MatDenseRestoreArrayReadAndMemType(X, &in));
2720: PetscFunctionReturn(PETSC_SUCCESS);
2721: }
2723: /* MatDenseScatter_Private - Scatters the columns of the dense matrix X into the dense matrix Y
2725: Collective
2727: Input Parameters:
2728: + sf - the PetscSF for one column
2729: . X - the matrix to read
2730: . Y - the matrix to update
2731: . mode - INSERT_VALUES, ADD_VALUES, MAX_VALUES, or MIN_VALUES
2732: - smode - SCATTER_FORWARD for a broadcast from X to Y, or SCATTER_REVERSE for a reduction from X to Y
2734: Level: developer
2736: Note:
2737: With more than one column, a strided PetscSF is cached on sf. It is rebuilt if the number of columns changes.
2738: Any other change of layout, such as another leading dimension, is an error unless the cache is first cleared
2739: on all processes with PetscObjectCompose((PetscObject)sf, "_MatDenseScatter_StridedSF", NULL).
2741: .seealso: `MatDenseScatterWithSF_Private()`, `PetscSFCreateStridedSF()`
2742: */
2743: PetscErrorCode MatDenseScatter_Private(PetscSF sf, Mat X, Mat Y, InsertMode mode, ScatterMode smode)
2744: {
2745: PetscSF vsf;
2746: PetscInt N, ny, rld, lld;
2747: MPI_Op op = MPI_OP_NULL;
2748: PetscBool writeonly = PETSC_FALSE;
2750: PetscFunctionBegin;
2754: if (mode == INSERT_VALUES) op = MPI_REPLACE;
2755: else if (mode == ADD_VALUES) op = MPIU_SUM;
2756: else if (mode == MAX_VALUES) op = MPIU_MAX;
2757: else if (mode == MIN_VALUES) op = MPIU_MIN;
2758: PetscCheck(op != MPI_OP_NULL, PetscObjectComm((PetscObject)sf), PETSC_ERR_SUP, "Unsupported InsertMode %d in MatDenseScatter_Private()", mode);
2759: PetscCheck(smode == SCATTER_FORWARD || smode == SCATTER_REVERSE, PetscObjectComm((PetscObject)sf), PETSC_ERR_SUP, "Unsupported ScatterMode %d in MatDenseScatter_Private()", smode);
2760: PetscCall(MatGetSize(X, NULL, &N));
2761: PetscCall(MatGetSize(Y, NULL, &ny));
2762: PetscCheck(N == ny, PetscObjectComm((PetscObject)sf), PETSC_ERR_ARG_SIZ, "Matrix column sizes must match: %" PetscInt_FMT " != %" PetscInt_FMT, N, ny);
2763: if (!N) PetscFunctionReturn(PETSC_SUCCESS);
2764: PetscCall(MatDenseGetLDA(smode == SCATTER_FORWARD ? X : Y, &rld));
2765: PetscCall(MatDenseGetLDA(smode == SCATTER_FORWARD ? Y : X, &lld));
2766: /* get cached or create new strided PetscSF when the number of columns is greater than one */
2767: if (N > 1) {
2768: MatDenseScatterCtx *ctx;
2769: PetscInt nr, nl;
2771: PetscCall(PetscSFGetGraph(sf, &nr, &nl, NULL, NULL));
2772: PetscCall(PetscObjectContainerQuery((PetscObject)sf, "_MatDenseScatter_StridedSF", &ctx));
2773: if (ctx && ctx->N != N) {
2774: PetscCall(PetscObjectCompose((PetscObject)sf, "_MatDenseScatter_StridedSF", NULL));
2775: ctx = NULL;
2776: }
2777: /* callers must clear the cache on all processes before changing the layout (which happens, e.g., with matrices with different leading dimensions) on any rank */
2778: PetscCheck(ctx == NULL || (ctx->rld == rld && ctx->lld == lld && ctx->nr == nr && ctx->nl == nl), PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "The cached strided SF has a different layout; the calling function must destroy it first on all processes with PetscObjectCompose((PetscObject)sf, \"_MatDenseScatter_StridedSF\", NULL)");
2779: if (ctx == NULL) {
2780: PetscCall(PetscNew(&ctx));
2781: PetscCall(PetscSFCreateStridedSF(sf, N, rld, lld, &ctx->sf));
2782: ctx->N = N;
2783: ctx->rld = rld;
2784: ctx->lld = lld;
2785: ctx->nr = nr;
2786: ctx->nl = nl;
2787: PetscCall(PetscObjectContainerCompose((PetscObject)sf, "_MatDenseScatter_StridedSF", ctx, MatDenseScatterCtxDestroy_Private));
2788: }
2789: vsf = ctx->sf;
2790: } else vsf = sf;
2791: if (mode == INSERT_VALUES && smode == SCATTER_FORWARD) {
2792: const PetscInt *ilocal;
2793: PetscInt nl;
2795: PetscCall(PetscSFGetGraph(sf, NULL, &nl, &ilocal, NULL));
2796: PetscCall(MatGetLocalSize(Y, &ny, NULL));
2797: // preserve untouched rows and padding; contiguous leaves cover every row only in this case
2798: writeonly = (PetscBool)(ilocal == NULL && (PetscCount)nl * PetscMax(sf->vscat.bs, 1) == ny && lld == ny);
2799: }
2800: PetscCall(MatDenseScatterWithSF_Private(vsf, X, Y, op, smode, writeonly));
2801: PetscFunctionReturn(PETSC_SUCCESS);
2802: }