Actual source code: mpidense.c
1: /*
2: Basic functions for basic parallel dense matrices.
3: Portions of this code are under:
4: Copyright (c) 2022 Advanced Micro Devices, Inc. All rights reserved.
5: */
7: #include <../src/mat/impls/dense/mpi/mpidense.h>
8: #include <../src/mat/impls/aij/mpi/mpiaij.h>
9: #include <petscblaslapack.h>
10: #include <petsc/private/vecimpl.h>
11: #include <petsc/private/deviceimpl.h>
12: #include <petsc/private/sfimpl.h>
14: /*@
15: MatDenseGetLocalMatrix - For a `MATMPIDENSE` or `MATSEQDENSE` matrix returns the sequential
16: matrix that represents the operator. For sequential matrices it returns itself.
18: Input Parameter:
19: . A - the sequential or MPI `MATDENSE` matrix
21: Output Parameter:
22: . B - the inner matrix
24: Level: intermediate
26: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MATMPIDENSE`, `MATSEQDENSE`
27: @*/
28: PetscErrorCode MatDenseGetLocalMatrix(Mat A, Mat *B)
29: {
30: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
31: PetscBool flg;
33: PetscFunctionBegin;
35: PetscAssertPointer(B, 2);
36: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIDENSE, &flg));
37: if (flg) *B = mat->A;
38: else {
39: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQDENSE, &flg));
40: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for matrix type %s", ((PetscObject)A)->type_name);
41: *B = A;
42: }
43: PetscFunctionReturn(PETSC_SUCCESS);
44: }
46: static PetscErrorCode MatCopy_MPIDense(Mat A, Mat B, MatStructure s)
47: {
48: Mat_MPIDense *Amat = (Mat_MPIDense *)A->data;
49: Mat_MPIDense *Bmat = (Mat_MPIDense *)B->data;
51: PetscFunctionBegin;
52: PetscCall(MatCopy(Amat->A, Bmat->A, s));
53: PetscFunctionReturn(PETSC_SUCCESS);
54: }
56: PetscErrorCode MatShift_MPIDense(Mat A, PetscScalar alpha)
57: {
58: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
59: PetscInt j, lda, rstart = A->rmap->rstart, rend = A->rmap->rend, rend2;
60: PetscScalar *v;
62: PetscFunctionBegin;
63: PetscCall(MatDenseGetArray(mat->A, &v));
64: PetscCall(MatDenseGetLDA(mat->A, &lda));
65: rend2 = PetscMin(rend, A->cmap->N);
66: if (rend2 > rstart) {
67: for (j = rstart; j < rend2; j++) v[j - rstart + j * lda] += alpha;
68: PetscCall(PetscLogFlops(rend2 - rstart));
69: }
70: PetscCall(MatDenseRestoreArray(mat->A, &v));
71: PetscFunctionReturn(PETSC_SUCCESS);
72: }
74: static PetscErrorCode MatGetRow_MPIDense(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
75: {
76: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
77: PetscInt lrow, rstart = A->rmap->rstart, rend = A->rmap->rend;
79: PetscFunctionBegin;
80: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "only local rows");
81: lrow = row - rstart;
82: PetscCall(MatGetRow(mat->A, lrow, nz, (const PetscInt **)idx, (const PetscScalar **)v));
83: PetscFunctionReturn(PETSC_SUCCESS);
84: }
86: static PetscErrorCode MatRestoreRow_MPIDense(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
87: {
88: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
89: PetscInt lrow, rstart = A->rmap->rstart, rend = A->rmap->rend;
91: PetscFunctionBegin;
92: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "only local rows");
93: lrow = row - rstart;
94: PetscCall(MatRestoreRow(mat->A, lrow, nz, (const PetscInt **)idx, (const PetscScalar **)v));
95: PetscFunctionReturn(PETSC_SUCCESS);
96: }
98: static PetscErrorCode MatGetDiagonalBlock_MPIDense(Mat A, Mat *a)
99: {
100: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
101: PetscInt m = A->rmap->n, rstart = A->rmap->rstart;
102: PetscScalar *array;
103: MPI_Comm comm;
104: PetscBool flg;
105: Mat B;
107: PetscFunctionBegin;
108: PetscCall(MatHasCongruentLayouts(A, &flg));
109: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only square matrices supported.");
110: PetscCall(PetscObjectQuery((PetscObject)A, "DiagonalBlock", (PetscObject *)&B));
111: if (!B) { /* This should use MatDenseGetSubMatrix (not create), but we would need a call like MatRestoreDiagonalBlock */
112: #if PetscDefined(HAVE_CUDA)
113: PetscCall(PetscObjectTypeCompare((PetscObject)mdn->A, MATSEQDENSECUDA, &flg));
114: PetscCheck(!flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not coded for %s. Send an email to petsc-dev@mcs.anl.gov to request this feature", MATSEQDENSECUDA);
115: #elif PetscDefined(HAVE_HIP)
116: PetscCall(PetscObjectTypeCompare((PetscObject)mdn->A, MATSEQDENSEHIP, &flg));
117: PetscCheck(!flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not coded for %s. Send an email to petsc-dev@mcs.anl.gov to request this feature", MATSEQDENSEHIP);
118: #endif
119: PetscCall(PetscObjectGetComm((PetscObject)mdn->A, &comm));
120: PetscCall(MatCreate(comm, &B));
121: PetscCall(MatSetSizes(B, m, m, m, m));
122: PetscCall(MatSetType(B, ((PetscObject)mdn->A)->type_name));
123: PetscCall(MatDenseGetArrayRead(mdn->A, (const PetscScalar **)&array));
124: PetscCall(MatSeqDenseSetPreallocation(B, array + m * rstart));
125: PetscCall(MatDenseRestoreArrayRead(mdn->A, (const PetscScalar **)&array));
126: PetscCall(PetscObjectCompose((PetscObject)A, "DiagonalBlock", (PetscObject)B));
127: *a = B;
128: PetscCall(MatDestroy(&B));
129: } else *a = B;
130: PetscFunctionReturn(PETSC_SUCCESS);
131: }
133: static PetscErrorCode MatSetValues_MPIDense(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], const PetscScalar v[], InsertMode addv)
134: {
135: Mat_MPIDense *A = (Mat_MPIDense *)mat->data;
136: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend, row;
137: PetscBool roworiented = A->roworiented;
139: PetscFunctionBegin;
140: for (i = 0; i < m; i++) {
141: if (idxm[i] < 0) continue;
142: PetscCheck(idxm[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large");
143: if (idxm[i] >= rstart && idxm[i] < rend) {
144: row = idxm[i] - rstart;
145: if (roworiented) {
146: PetscCall(MatSetValues(A->A, 1, &row, n, idxn, PetscSafePointerPlusOffset(v, i * n), addv));
147: } else {
148: for (j = 0; j < n; j++) {
149: if (idxn[j] < 0) continue;
150: PetscCheck(idxn[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large");
151: PetscCall(MatSetValues(A->A, 1, &row, 1, &idxn[j], PetscSafePointerPlusOffset(v, i + j * m), addv));
152: }
153: }
154: } else if (!A->donotstash) {
155: mat->assembled = PETSC_FALSE;
156: if (roworiented) {
157: PetscCall(MatStashValuesRow_Private(&mat->stash, idxm[i], n, idxn, PetscSafePointerPlusOffset(v, i * n), PETSC_FALSE));
158: } else {
159: PetscCall(MatStashValuesCol_Private(&mat->stash, idxm[i], n, idxn, PetscSafePointerPlusOffset(v, i), m, PETSC_FALSE));
160: }
161: }
162: }
163: PetscFunctionReturn(PETSC_SUCCESS);
164: }
166: static PetscErrorCode MatGetValues_MPIDense(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
167: {
168: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
169: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend, row;
170: PetscBool roworiented = mdn->roworiented;
171: PetscScalar *value;
173: PetscFunctionBegin;
174: for (i = 0; i < m; i++) {
175: if (idxm[i] < 0) continue; /* negative row */
176: PetscCheck(idxm[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large");
177: PetscCheck(idxm[i] >= rstart && idxm[i] < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only local values currently supported");
178: row = idxm[i] - rstart;
179: for (j = 0; j < n; j++) {
180: if (idxn[j] < 0) continue; /* negative column */
181: PetscCheck(idxn[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large");
182: value = roworiented ? &v[j + i * n] : &v[i + j * m];
183: PetscCall(MatGetValues(mdn->A, 1, &row, 1, &idxn[j], value));
184: }
185: }
186: PetscFunctionReturn(PETSC_SUCCESS);
187: }
189: static PetscErrorCode MatDenseGetLDA_MPIDense(Mat A, PetscInt *lda)
190: {
191: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
193: PetscFunctionBegin;
194: PetscCall(MatDenseGetLDA(a->A, lda));
195: PetscFunctionReturn(PETSC_SUCCESS);
196: }
198: static PetscErrorCode MatDenseSetLDA_MPIDense(Mat A, PetscInt lda)
199: {
200: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
201: MatType mtype = MATSEQDENSE;
203: PetscFunctionBegin;
204: if (!a->A) {
205: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
206: PetscCall(PetscLayoutSetUp(A->rmap));
207: PetscCall(PetscLayoutSetUp(A->cmap));
208: PetscCall(MatCreate(PETSC_COMM_SELF, &a->A));
209: PetscCall(MatSetSizes(a->A, A->rmap->n, A->cmap->N, A->rmap->n, A->cmap->N));
210: #if PetscDefined(HAVE_CUDA)
211: PetscBool iscuda;
212: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIDENSECUDA, &iscuda));
213: if (iscuda) mtype = MATSEQDENSECUDA;
214: #elif PetscDefined(HAVE_HIP)
215: PetscBool iship;
216: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIDENSEHIP, &iship));
217: if (iship) mtype = MATSEQDENSEHIP;
218: #endif
219: PetscCall(MatSetType(a->A, mtype));
220: }
221: PetscCall(MatDenseSetLDA(a->A, lda));
222: PetscFunctionReturn(PETSC_SUCCESS);
223: }
225: static PetscErrorCode MatDenseGetArray_MPIDense(Mat A, PetscScalar **array)
226: {
227: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
229: PetscFunctionBegin;
230: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
231: PetscCall(MatDenseGetArray(a->A, array));
232: PetscFunctionReturn(PETSC_SUCCESS);
233: }
235: static PetscErrorCode MatDenseGetArrayRead_MPIDense(Mat A, PetscScalar **array)
236: {
237: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
239: PetscFunctionBegin;
240: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
241: PetscCall(MatDenseGetArrayRead(a->A, (const PetscScalar **)array));
242: PetscFunctionReturn(PETSC_SUCCESS);
243: }
245: static PetscErrorCode MatDenseGetArrayWrite_MPIDense(Mat A, PetscScalar **array)
246: {
247: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
249: PetscFunctionBegin;
250: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
251: PetscCall(MatDenseGetArrayWrite(a->A, array));
252: PetscFunctionReturn(PETSC_SUCCESS);
253: }
255: static PetscErrorCode MatDensePlaceArray_MPIDense(Mat A, const PetscScalar *array)
256: {
257: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
259: PetscFunctionBegin;
260: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
261: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
262: PetscCall(MatDensePlaceArray(a->A, array));
263: PetscFunctionReturn(PETSC_SUCCESS);
264: }
266: static PetscErrorCode MatDenseResetArray_MPIDense(Mat A)
267: {
268: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
270: PetscFunctionBegin;
271: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
272: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
273: PetscCall(MatDenseResetArray(a->A));
274: PetscFunctionReturn(PETSC_SUCCESS);
275: }
277: static PetscErrorCode MatDenseReplaceArray_MPIDense(Mat A, const PetscScalar *array)
278: {
279: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
281: PetscFunctionBegin;
282: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
283: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
284: PetscCall(MatDenseReplaceArray(a->A, array));
285: PetscFunctionReturn(PETSC_SUCCESS);
286: }
288: static PetscErrorCode MatCreateSubMatrix_MPIDense(Mat A, IS isrow, IS iscol, MatReuse scall, Mat *B)
289: {
290: Mat_MPIDense *mat = (Mat_MPIDense *)A->data, *newmatd;
291: PetscInt lda, i, j, rstart, rend, nrows, ncols, Ncols, nlrows, nlcols;
292: const PetscInt *irow, *icol;
293: const PetscScalar *v;
294: PetscScalar *bv;
295: Mat newmat;
296: IS iscol_local;
297: MPI_Comm comm_is, comm_mat;
299: PetscFunctionBegin;
300: PetscCall(PetscObjectGetComm((PetscObject)A, &comm_mat));
301: PetscCall(PetscObjectGetComm((PetscObject)iscol, &comm_is));
302: PetscCheck(comm_mat == comm_is, PETSC_COMM_SELF, PETSC_ERR_ARG_NOTSAMECOMM, "IS communicator must match matrix communicator");
304: PetscCall(ISAllGather(iscol, &iscol_local));
305: PetscCall(ISGetIndices(isrow, &irow));
306: PetscCall(ISGetIndices(iscol_local, &icol));
307: PetscCall(ISGetLocalSize(isrow, &nrows));
308: PetscCall(ISGetLocalSize(iscol, &ncols));
309: PetscCall(ISGetSize(iscol, &Ncols)); /* global number of columns, size of iscol_local */
311: /* No parallel redistribution currently supported! Should really check each index set
312: to confirm that it is OK. ... Currently supports only submatrix same partitioning as
313: original matrix! */
315: PetscCall(MatGetLocalSize(A, &nlrows, &nlcols));
316: PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
318: /* Check submatrix call */
319: if (scall == MAT_REUSE_MATRIX) {
320: /* SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size"); */
321: /* Really need to test rows and column sizes! */
322: newmat = *B;
323: } else {
324: /* Create and fill new matrix */
325: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &newmat));
326: PetscCall(MatSetSizes(newmat, nrows, ncols, PETSC_DECIDE, Ncols));
327: PetscCall(MatSetType(newmat, ((PetscObject)A)->type_name));
328: PetscCall(MatMPIDenseSetPreallocation(newmat, NULL));
329: }
331: /* Now extract the data pointers and do the copy, column at a time */
332: newmatd = (Mat_MPIDense *)newmat->data;
333: PetscCall(MatDenseGetArray(newmatd->A, &bv));
334: PetscCall(MatDenseGetArrayRead(mat->A, &v));
335: PetscCall(MatDenseGetLDA(mat->A, &lda));
336: for (i = 0; i < Ncols; i++) {
337: const PetscScalar *av = v + lda * icol[i];
338: for (j = 0; j < nrows; j++) *bv++ = av[irow[j] - rstart];
339: }
340: PetscCall(MatDenseRestoreArrayRead(mat->A, &v));
341: PetscCall(MatDenseRestoreArray(newmatd->A, &bv));
343: /* Assemble the matrices so that the correct flags are set */
344: PetscCall(MatAssemblyBegin(newmat, MAT_FINAL_ASSEMBLY));
345: PetscCall(MatAssemblyEnd(newmat, MAT_FINAL_ASSEMBLY));
347: /* Free work space */
348: PetscCall(ISRestoreIndices(isrow, &irow));
349: PetscCall(ISRestoreIndices(iscol_local, &icol));
350: PetscCall(ISDestroy(&iscol_local));
351: *B = newmat;
352: PetscFunctionReturn(PETSC_SUCCESS);
353: }
355: static PetscErrorCode MatDenseRestoreArray_MPIDense(Mat A, PetscScalar **array)
356: {
357: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
359: PetscFunctionBegin;
360: PetscCall(MatDenseRestoreArray(a->A, array));
361: PetscFunctionReturn(PETSC_SUCCESS);
362: }
364: static PetscErrorCode MatDenseRestoreArrayRead_MPIDense(Mat A, PetscScalar **array)
365: {
366: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
368: PetscFunctionBegin;
369: PetscCall(MatDenseRestoreArrayRead(a->A, (const PetscScalar **)array));
370: PetscFunctionReturn(PETSC_SUCCESS);
371: }
373: static PetscErrorCode MatDenseRestoreArrayWrite_MPIDense(Mat A, PetscScalar **array)
374: {
375: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
377: PetscFunctionBegin;
378: PetscCall(MatDenseRestoreArrayWrite(a->A, array));
379: PetscFunctionReturn(PETSC_SUCCESS);
380: }
382: static PetscErrorCode MatAssemblyBegin_MPIDense(Mat mat, MatAssemblyType mode)
383: {
384: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
385: PetscInt nstash, reallocs;
387: PetscFunctionBegin;
388: if (mdn->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
390: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
391: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
392: PetscCall(PetscInfo(mdn->A, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
393: PetscFunctionReturn(PETSC_SUCCESS);
394: }
396: static PetscErrorCode MatAssemblyEnd_MPIDense(Mat mat, MatAssemblyType mode)
397: {
398: Mat_MPIDense *mdn = (Mat_MPIDense *)mat->data;
399: PetscInt i, *row, *col, flg, j, rstart, ncols;
400: PetscMPIInt n;
401: PetscScalar *val;
403: PetscFunctionBegin;
404: if (!mdn->donotstash && !mat->nooffprocentries) {
405: /* wait on receives */
406: while (1) {
407: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
408: if (!flg) break;
410: for (i = 0; i < n;) {
411: /* Now identify the consecutive vals belonging to the same row */
412: for (j = i, rstart = row[j]; j < n; j++) {
413: if (row[j] != rstart) break;
414: }
415: if (j < n) ncols = j - i;
416: else ncols = n - i;
417: /* Now assemble all these values with a single function call */
418: PetscCall(MatSetValues_MPIDense(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
419: i = j;
420: }
421: }
422: PetscCall(MatStashScatterEnd_Private(&mat->stash));
423: }
425: PetscCall(MatAssemblyBegin(mdn->A, mode));
426: PetscCall(MatAssemblyEnd(mdn->A, mode));
427: PetscFunctionReturn(PETSC_SUCCESS);
428: }
430: static PetscErrorCode MatZeroEntries_MPIDense(Mat A)
431: {
432: Mat_MPIDense *l = (Mat_MPIDense *)A->data;
434: PetscFunctionBegin;
435: PetscCall(MatZeroEntries(l->A));
436: PetscFunctionReturn(PETSC_SUCCESS);
437: }
439: static PetscErrorCode MatSetInf_MPIDense(Mat A)
440: {
441: Mat_MPIDense *l = (Mat_MPIDense *)A->data;
443: PetscFunctionBegin;
444: PetscCall(MatSetInf(l->A));
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: PetscFunctionReturn(PETSC_SUCCESS);
1205: }
1207: static PetscErrorCode MatMatTransposeMultSymbolic_MPIDense_MPIDense(Mat, Mat, PetscReal, Mat);
1208: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense(Mat, Mat, Mat);
1209: static PetscErrorCode MatTransposeMatMultSymbolic_MPIDense_MPIDense(Mat, Mat, PetscReal, Mat);
1210: static PetscErrorCode MatTransposeMatMultNumeric_MPIDense_MPIDense(Mat, Mat, Mat);
1211: static PetscErrorCode MatEqual_MPIDense(Mat, Mat, PetscBool *);
1212: static PetscErrorCode MatLoad_MPIDense(Mat, PetscViewer);
1213: static PetscErrorCode MatProductSetFromOptions_MPIDense(Mat);
1215: static struct _MatOps MatOps_Values = {MatSetValues_MPIDense,
1216: MatGetRow_MPIDense,
1217: MatRestoreRow_MPIDense,
1218: MatMult_MPIDense,
1219: /* 4*/ MatMultAdd_MPIDense,
1220: MatMultTranspose_MPIDense,
1221: MatMultTransposeAdd_MPIDense,
1222: NULL,
1223: NULL,
1224: NULL,
1225: /* 10*/ NULL,
1226: NULL,
1227: NULL,
1228: NULL,
1229: MatTranspose_MPIDense,
1230: /* 15*/ MatGetInfo_MPIDense,
1231: MatEqual_MPIDense,
1232: MatGetDiagonal_MPIDense,
1233: MatDiagonalScale_MPIDense,
1234: MatNorm_MPIDense,
1235: /* 20*/ MatAssemblyBegin_MPIDense,
1236: MatAssemblyEnd_MPIDense,
1237: MatSetOption_MPIDense,
1238: MatZeroEntries_MPIDense,
1239: /* 24*/ MatZeroRows_MPIDense,
1240: NULL,
1241: NULL,
1242: NULL,
1243: NULL,
1244: /* 29*/ MatSetUp_MPIDense,
1245: NULL,
1246: NULL,
1247: MatGetDiagonalBlock_MPIDense,
1248: MatSetInf_MPIDense,
1249: /* 34*/ MatDuplicate_MPIDense,
1250: NULL,
1251: NULL,
1252: NULL,
1253: NULL,
1254: /* 39*/ MatAXPY_MPIDense,
1255: MatCreateSubMatrices_MPIDense,
1256: NULL,
1257: MatGetValues_MPIDense,
1258: MatCopy_MPIDense,
1259: /* 44*/ NULL,
1260: MatScale_MPIDense,
1261: MatShift_MPIDense,
1262: NULL,
1263: NULL,
1264: /* 49*/ MatSetRandom_MPIDense,
1265: NULL,
1266: NULL,
1267: NULL,
1268: NULL,
1269: /* 54*/ NULL,
1270: NULL,
1271: NULL,
1272: NULL,
1273: NULL,
1274: /* 59*/ MatCreateSubMatrix_MPIDense,
1275: MatDestroy_MPIDense,
1276: MatView_MPIDense,
1277: NULL,
1278: NULL,
1279: /* 64*/ NULL,
1280: NULL,
1281: NULL,
1282: NULL,
1283: NULL,
1284: /* 69*/ NULL,
1285: NULL,
1286: NULL,
1287: NULL,
1288: NULL,
1289: /* 74*/ NULL,
1290: NULL,
1291: NULL,
1292: NULL,
1293: MatLoad_MPIDense,
1294: /* 79*/ NULL,
1295: NULL,
1296: NULL,
1297: NULL,
1298: /* 83*/ NULL,
1299: NULL,
1300: NULL,
1301: NULL,
1302: MatMatTransposeMultSymbolic_MPIDense_MPIDense,
1303: MatMatTransposeMultNumeric_MPIDense_MPIDense,
1304: /* 89*/ NULL,
1305: MatProductSetFromOptions_MPIDense,
1306: NULL,
1307: NULL,
1308: MatConjugate_MPIDense,
1309: /* 94*/ NULL,
1310: NULL,
1311: MatRealPart_MPIDense,
1312: MatImaginaryPart_MPIDense,
1313: NULL,
1314: /*99*/ NULL,
1315: NULL,
1316: NULL,
1317: NULL,
1318: MatGetColumnVector_MPIDense,
1319: /*104*/ NULL,
1320: NULL,
1321: NULL,
1322: NULL,
1323: NULL,
1324: /*109*/ NULL,
1325: NULL,
1326: MatMultHermitianTranspose_MPIDense,
1327: MatMultHermitianTransposeAdd_MPIDense,
1328: NULL,
1329: /*114*/ NULL,
1330: MatGetColumnReductions_MPIDense,
1331: NULL,
1332: NULL,
1333: NULL,
1334: /*120*/ MatTransposeMatMultSymbolic_MPIDense_MPIDense,
1335: MatTransposeMatMultNumeric_MPIDense_MPIDense,
1336: NULL,
1337: NULL,
1338: /*124*/ NULL,
1339: NULL,
1340: NULL,
1341: NULL,
1342: NULL,
1343: /*129*/ NULL,
1344: MatCreateMPIMatConcatenateSeqMat_MPIDense,
1345: NULL,
1346: NULL,
1347: NULL,
1348: /*134*/ NULL,
1349: NULL,
1350: NULL,
1351: NULL,
1352: NULL,
1353: /*139*/ NULL,
1354: NULL,
1355: NULL,
1356: NULL,
1357: NULL,
1358: NULL,
1359: /*144*/ NULL,
1360: NULL,
1361: NULL,
1362: NULL};
1364: static PetscErrorCode MatMPIDenseSetPreallocation_MPIDense(Mat mat, PetscScalar *data)
1365: {
1366: Mat_MPIDense *a = (Mat_MPIDense *)mat->data;
1367: MatType mtype = MATSEQDENSE;
1369: PetscFunctionBegin;
1370: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)mat), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1371: PetscCall(PetscLayoutSetUp(mat->rmap));
1372: PetscCall(PetscLayoutSetUp(mat->cmap));
1373: if (!a->A) {
1374: PetscCall(MatCreate(PETSC_COMM_SELF, &a->A));
1375: PetscCall(MatSetSizes(a->A, mat->rmap->n, mat->cmap->N, mat->rmap->n, mat->cmap->N));
1376: }
1377: #if PetscDefined(HAVE_CUDA)
1378: PetscBool iscuda;
1379: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATMPIDENSECUDA, &iscuda));
1380: if (iscuda) mtype = MATSEQDENSECUDA;
1381: #endif
1382: #if PetscDefined(HAVE_HIP)
1383: PetscBool iship;
1384: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATMPIDENSEHIP, &iship));
1385: if (iship) mtype = MATSEQDENSEHIP;
1386: #endif
1387: PetscCall(MatSetType(a->A, mtype));
1388: PetscCall(MatSeqDenseSetPreallocation(a->A, data));
1389: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
1390: mat->offloadmask = a->A->offloadmask;
1391: #endif
1392: mat->preallocated = PETSC_TRUE;
1393: mat->assembled = PETSC_TRUE;
1394: PetscFunctionReturn(PETSC_SUCCESS);
1395: }
1397: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIDense(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
1398: {
1399: Mat B, C;
1401: PetscFunctionBegin;
1402: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &C));
1403: PetscCall(MatConvert_SeqAIJ_SeqDense(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &B));
1404: PetscCall(MatDestroy(&C));
1405: if (reuse == MAT_REUSE_MATRIX) {
1406: C = *newmat;
1407: } else C = NULL;
1408: PetscCall(MatCreateMPIMatConcatenateSeqMat(PetscObjectComm((PetscObject)A), B, A->cmap->n, !C ? MAT_INITIAL_MATRIX : MAT_REUSE_MATRIX, &C));
1409: PetscCall(MatDestroy(&B));
1410: if (reuse == MAT_INPLACE_MATRIX) PetscCall(MatHeaderReplace(A, &C));
1411: else if (reuse == MAT_INITIAL_MATRIX) *newmat = C;
1412: PetscFunctionReturn(PETSC_SUCCESS);
1413: }
1415: static PetscErrorCode MatConvert_MPIDense_MPIAIJ(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
1416: {
1417: Mat B, C;
1419: PetscFunctionBegin;
1420: PetscCall(MatDenseGetLocalMatrix(A, &C));
1421: PetscCall(MatConvert_SeqDense_SeqAIJ(C, MATSEQAIJ, MAT_INITIAL_MATRIX, &B));
1422: if (reuse == MAT_REUSE_MATRIX) {
1423: C = *newmat;
1424: } else C = NULL;
1425: PetscCall(MatCreateMPIMatConcatenateSeqMat(PetscObjectComm((PetscObject)A), B, A->cmap->n, !C ? MAT_INITIAL_MATRIX : MAT_REUSE_MATRIX, &C));
1426: PetscCall(MatDestroy(&B));
1427: if (reuse == MAT_INPLACE_MATRIX) PetscCall(MatHeaderReplace(A, &C));
1428: else if (reuse == MAT_INITIAL_MATRIX) *newmat = C;
1429: PetscFunctionReturn(PETSC_SUCCESS);
1430: }
1432: #if PetscDefined(HAVE_ELEMENTAL)
1433: PETSC_INTERN PetscErrorCode MatConvert_MPIDense_Elemental(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
1434: {
1435: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1436: Mat mat_elemental;
1437: PetscScalar *v;
1438: PetscInt m = A->rmap->n, N = A->cmap->N, rstart = A->rmap->rstart, i, *rows, *cols, lda;
1440: PetscFunctionBegin;
1441: if (reuse == MAT_REUSE_MATRIX) {
1442: mat_elemental = *newmat;
1443: PetscCall(MatZeroEntries(*newmat));
1444: } else {
1445: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &mat_elemental));
1446: PetscCall(MatSetSizes(mat_elemental, PETSC_DECIDE, PETSC_DECIDE, A->rmap->N, A->cmap->N));
1447: PetscCall(MatSetType(mat_elemental, MATELEMENTAL));
1448: PetscCall(MatSetUp(mat_elemental));
1449: PetscCall(MatSetOption(mat_elemental, MAT_ROW_ORIENTED, PETSC_FALSE));
1450: }
1452: PetscCall(PetscMalloc2(m, &rows, N, &cols));
1453: for (i = 0; i < N; i++) cols[i] = i;
1454: for (i = 0; i < m; i++) rows[i] = rstart + i;
1456: /* PETSc-Elemental interface uses axpy for setting off-processor entries, only ADD_VALUES is allowed */
1457: PetscCall(MatDenseGetArray(A, &v));
1458: PetscCall(MatDenseGetLDA(a->A, &lda));
1459: if (lda == m) PetscCall(MatSetValues(mat_elemental, m, rows, N, cols, v, ADD_VALUES));
1460: else {
1461: for (i = 0; i < N; i++) PetscCall(MatSetValues(mat_elemental, m, rows, 1, &i, v + lda * i, ADD_VALUES));
1462: }
1463: PetscCall(MatAssemblyBegin(mat_elemental, MAT_FINAL_ASSEMBLY));
1464: PetscCall(MatAssemblyEnd(mat_elemental, MAT_FINAL_ASSEMBLY));
1465: PetscCall(MatDenseRestoreArray(A, &v));
1466: PetscCall(PetscFree2(rows, cols));
1468: if (reuse == MAT_INPLACE_MATRIX) {
1469: PetscCall(MatHeaderReplace(A, &mat_elemental));
1470: } else {
1471: *newmat = mat_elemental;
1472: }
1473: PetscFunctionReturn(PETSC_SUCCESS);
1474: }
1475: #endif
1477: static PetscErrorCode MatDenseGetColumn_MPIDense(Mat A, PetscInt col, PetscScalar **vals)
1478: {
1479: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
1481: PetscFunctionBegin;
1482: PetscCall(MatDenseGetColumn(mat->A, col, vals));
1483: PetscFunctionReturn(PETSC_SUCCESS);
1484: }
1486: static PetscErrorCode MatDenseRestoreColumn_MPIDense(Mat A, PetscScalar **vals)
1487: {
1488: Mat_MPIDense *mat = (Mat_MPIDense *)A->data;
1490: PetscFunctionBegin;
1491: PetscCall(MatDenseRestoreColumn(mat->A, vals));
1492: PetscFunctionReturn(PETSC_SUCCESS);
1493: }
1495: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIDense(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
1496: {
1497: Mat_MPIDense *mat;
1498: PetscInt m, nloc, N;
1500: PetscFunctionBegin;
1501: PetscCall(MatGetSize(inmat, &m, &N));
1502: PetscCall(MatGetLocalSize(inmat, NULL, &nloc));
1503: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
1504: PetscInt sum;
1506: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
1507: /* Check sum(n) = N */
1508: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
1509: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
1511: PetscCall(MatCreateDense(comm, m, n, PETSC_DETERMINE, N, NULL, outmat));
1512: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
1513: }
1515: /* numeric phase */
1516: mat = (Mat_MPIDense *)(*outmat)->data;
1517: PetscCall(MatCopy(inmat, mat->A, SAME_NONZERO_PATTERN));
1518: PetscFunctionReturn(PETSC_SUCCESS);
1519: }
1521: PetscErrorCode MatDenseGetColumnVec_MPIDense(Mat A, PetscInt col, Vec *v)
1522: {
1523: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1524: PetscInt lda;
1526: PetscFunctionBegin;
1527: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1528: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1529: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
1530: a->vecinuse = col + 1;
1531: PetscCall(MatDenseGetLDA(a->A, &lda));
1532: PetscCall(MatDenseGetArray(a->A, (PetscScalar **)&a->ptrinuse));
1533: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)lda)));
1534: *v = a->cvec;
1535: PetscFunctionReturn(PETSC_SUCCESS);
1536: }
1538: PetscErrorCode MatDenseRestoreColumnVec_MPIDense(Mat A, PetscInt col, Vec *v)
1539: {
1540: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1542: PetscFunctionBegin;
1543: PetscCheck(a->vecinuse, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
1544: PetscCheck(a->cvec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing internal column vector");
1545: VecCheckAssembled(a->cvec);
1546: a->vecinuse = 0;
1547: PetscCall(MatDenseRestoreArray(a->A, (PetscScalar **)&a->ptrinuse));
1548: PetscCall(VecResetArray(a->cvec));
1549: if (v) *v = NULL;
1550: PetscFunctionReturn(PETSC_SUCCESS);
1551: }
1553: PetscErrorCode MatDenseGetColumnVecRead_MPIDense(Mat A, PetscInt col, Vec *v)
1554: {
1555: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1556: PetscInt lda;
1558: PetscFunctionBegin;
1559: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1560: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1561: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
1562: a->vecinuse = col + 1;
1563: PetscCall(MatDenseGetLDA(a->A, &lda));
1564: PetscCall(MatDenseGetArrayRead(a->A, &a->ptrinuse));
1565: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)lda)));
1566: PetscCall(VecLockReadPush(a->cvec));
1567: *v = a->cvec;
1568: PetscFunctionReturn(PETSC_SUCCESS);
1569: }
1571: PetscErrorCode MatDenseRestoreColumnVecRead_MPIDense(Mat A, PetscInt col, Vec *v)
1572: {
1573: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1575: PetscFunctionBegin;
1576: PetscCheck(a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
1577: PetscCheck(a->cvec, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing internal column vector");
1578: VecCheckAssembled(a->cvec);
1579: a->vecinuse = 0;
1580: PetscCall(MatDenseRestoreArrayRead(a->A, &a->ptrinuse));
1581: PetscCall(VecLockReadPop(a->cvec));
1582: PetscCall(VecResetArray(a->cvec));
1583: if (v) *v = NULL;
1584: PetscFunctionReturn(PETSC_SUCCESS);
1585: }
1587: PetscErrorCode MatDenseGetColumnVecWrite_MPIDense(Mat A, PetscInt col, Vec *v)
1588: {
1589: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1590: PetscInt lda;
1592: PetscFunctionBegin;
1593: PetscCheck(!a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1594: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1595: if (!a->cvec) PetscCall(MatDenseCreateColumnVec_Private(A, &a->cvec));
1596: a->vecinuse = col + 1;
1597: PetscCall(MatDenseGetLDA(a->A, &lda));
1598: PetscCall(MatDenseGetArrayWrite(a->A, (PetscScalar **)&a->ptrinuse));
1599: PetscCall(VecPlaceArray(a->cvec, PetscSafePointerPlusOffset(a->ptrinuse, (size_t)col * (size_t)lda)));
1600: *v = a->cvec;
1601: PetscFunctionReturn(PETSC_SUCCESS);
1602: }
1604: PetscErrorCode MatDenseRestoreColumnVecWrite_MPIDense(Mat A, PetscInt col, Vec *v)
1605: {
1606: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1608: PetscFunctionBegin;
1609: PetscCheck(a->vecinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseGetColumnVec() first");
1610: PetscCheck(a->cvec, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing internal column vector");
1611: VecCheckAssembled(a->cvec);
1612: a->vecinuse = 0;
1613: PetscCall(MatDenseRestoreArrayWrite(a->A, (PetscScalar **)&a->ptrinuse));
1614: PetscCall(VecResetArray(a->cvec));
1615: if (v) *v = NULL;
1616: PetscFunctionReturn(PETSC_SUCCESS);
1617: }
1619: static PetscErrorCode MatDenseGetSubMatrix_MPIDense(Mat A, PetscInt rbegin, PetscInt rend, PetscInt cbegin, PetscInt cend, Mat *v)
1620: {
1621: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1622: Mat_MPIDense *c;
1623: MPI_Comm comm;
1624: PetscInt prbegin, prend, pcbegin, pcend;
1626: PetscFunctionBegin;
1627: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
1628: PetscCheck(!a->vecinuse, comm, PETSC_ERR_ORDER, "Need to call MatDenseRestoreColumnVec() first");
1629: PetscCheck(!a->matinuse, comm, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1630: prbegin = PetscMax(0, PetscMin(A->rmap->rend, rbegin) - A->rmap->rstart);
1631: prend = PetscMin(A->rmap->n, PetscMax(0, rend - A->rmap->rstart));
1632: pcbegin = PetscMax(0, PetscMin(A->cmap->rend, cbegin) - A->cmap->rstart);
1633: pcend = PetscMin(A->cmap->n, PetscMax(0, cend - A->cmap->rstart));
1634: if (!a->cmat) {
1635: PetscCall(MatCreate(comm, &a->cmat));
1636: PetscCall(MatSetType(a->cmat, ((PetscObject)A)->type_name));
1637: if (rend - rbegin == A->rmap->N) PetscCall(PetscLayoutReference(A->rmap, &a->cmat->rmap));
1638: else {
1639: PetscCall(PetscLayoutSetLocalSize(a->cmat->rmap, prend - prbegin));
1640: PetscCall(PetscLayoutSetSize(a->cmat->rmap, rend - rbegin));
1641: PetscCall(PetscLayoutSetUp(a->cmat->rmap));
1642: }
1643: if (cend - cbegin == A->cmap->N) PetscCall(PetscLayoutReference(A->cmap, &a->cmat->cmap));
1644: else {
1645: PetscCall(PetscLayoutSetLocalSize(a->cmat->cmap, pcend - pcbegin));
1646: PetscCall(PetscLayoutSetSize(a->cmat->cmap, cend - cbegin));
1647: PetscCall(PetscLayoutSetUp(a->cmat->cmap));
1648: }
1649: c = (Mat_MPIDense *)a->cmat->data;
1650: c->sub_rbegin = rbegin;
1651: c->sub_rend = rend;
1652: c->sub_cbegin = cbegin;
1653: c->sub_cend = cend;
1654: }
1655: c = (Mat_MPIDense *)a->cmat->data;
1656: if (c->sub_rbegin != rbegin || c->sub_rend != rend) {
1657: PetscCall(PetscLayoutDestroy(&a->cmat->rmap));
1658: PetscCall(PetscLayoutCreate(comm, &a->cmat->rmap));
1659: PetscCall(PetscLayoutSetLocalSize(a->cmat->rmap, prend - prbegin));
1660: PetscCall(PetscLayoutSetSize(a->cmat->rmap, rend - rbegin));
1661: PetscCall(PetscLayoutSetUp(a->cmat->rmap));
1662: c->sub_rbegin = rbegin;
1663: c->sub_rend = rend;
1664: }
1665: if (c->sub_cbegin != cbegin || c->sub_cend != cend) {
1666: // special optimization: check if all columns are owned by rank 0, in which case no communication is necessary
1667: if ((cend - cbegin != a->cmat->cmap->N) || (A->cmap->range[1] != A->cmap->N)) {
1668: PetscCall(PetscLayoutDestroy(&a->cmat->cmap));
1669: PetscCall(PetscLayoutCreate(comm, &a->cmat->cmap));
1670: PetscCall(PetscLayoutSetLocalSize(a->cmat->cmap, pcend - pcbegin));
1671: PetscCall(PetscLayoutSetSize(a->cmat->cmap, cend - cbegin));
1672: PetscCall(PetscLayoutSetUp(a->cmat->cmap));
1673: PetscCall(VecDestroy(&c->lvec));
1674: PetscCall(PetscSFDestroy(&c->Mvctx));
1675: }
1676: c->sub_cbegin = cbegin;
1677: c->sub_cend = cend;
1678: }
1679: PetscCheck(!c->A, comm, PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1680: PetscCall(MatDenseGetSubMatrix(a->A, prbegin, prend, cbegin, cend, &c->A));
1682: a->cmat->preallocated = PETSC_TRUE;
1683: a->cmat->assembled = PETSC_TRUE;
1684: #if PetscDefined(HAVE_DEVICE)
1685: a->cmat->offloadmask = c->A->offloadmask;
1686: #endif
1687: a->matinuse = cbegin + 1;
1688: *v = a->cmat;
1689: PetscFunctionReturn(PETSC_SUCCESS);
1690: }
1692: static PetscErrorCode MatDenseRestoreSubMatrix_MPIDense(Mat A, Mat *v)
1693: {
1694: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1695: Mat_MPIDense *c;
1697: PetscFunctionBegin;
1698: PetscCheck(a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseGetSubMatrix() first");
1699: PetscCheck(a->cmat, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "Missing internal matrix");
1700: PetscCheck(*v == a->cmat, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Not the matrix obtained from MatDenseGetSubMatrix()");
1701: a->matinuse = 0;
1702: c = (Mat_MPIDense *)a->cmat->data;
1703: PetscCall(MatDenseRestoreSubMatrix(a->A, &c->A));
1704: *v = NULL;
1705: #if PetscDefined(HAVE_DEVICE)
1706: A->offloadmask = a->A->offloadmask;
1707: #endif
1708: PetscFunctionReturn(PETSC_SUCCESS);
1709: }
1711: static PetscErrorCode MatDenseUpdateColumnLayout_MPIDense(Mat A, PetscLayout clayout)
1712: {
1713: Mat_MPIDense *a = (Mat_MPIDense *)A->data;
1715: PetscFunctionBegin;
1716: if (A->cmap == clayout) PetscFunctionReturn(PETSC_SUCCESS);
1717: PetscCheck(!a->matinuse, PetscObjectComm((PetscObject)A), PETSC_ERR_ORDER, "Need to call MatDenseRestoreSubMatrix() first");
1718: PetscCall(PetscLayoutReference(clayout, &A->cmap));
1719: PetscCall(MatDestroy(&a->cmat));
1720: PetscCall(PetscSFDestroy(&a->Mvctx));
1721: PetscFunctionReturn(PETSC_SUCCESS);
1722: }
1724: /*MC
1725: MATMPIDENSE - MATMPIDENSE = "mpidense" - A matrix type to be used for distributed dense matrices.
1727: Options Database Key:
1728: . -mat_type mpidense - sets the matrix type to `MATMPIDENSE` during a call to `MatSetFromOptions()`
1730: Level: beginner
1732: .seealso: [](ch_matrices), `Mat`, `MatCreateDense()`, `MATSEQDENSE`, `MATDENSE`
1733: M*/
1734: PetscErrorCode MatCreate_MPIDense(Mat mat)
1735: {
1736: Mat_MPIDense *a;
1738: PetscFunctionBegin;
1739: PetscCall(PetscNew(&a));
1740: mat->data = (void *)a;
1741: mat->ops[0] = MatOps_Values;
1743: mat->insertmode = NOT_SET_VALUES;
1745: /* build cache for off array entries formed */
1746: a->donotstash = PETSC_FALSE;
1748: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)mat), 1, &mat->stash));
1750: /* stuff used for matrix vector multiply */
1751: a->lvec = NULL;
1752: a->Mvctx = NULL;
1753: a->roworiented = PETSC_TRUE;
1755: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetLDA_C", MatDenseGetLDA_MPIDense));
1756: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseSetLDA_C", MatDenseSetLDA_MPIDense));
1757: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArray_C", MatDenseGetArray_MPIDense));
1758: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArray_C", MatDenseRestoreArray_MPIDense));
1759: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayRead_C", MatDenseGetArrayRead_MPIDense));
1760: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayRead_C", MatDenseRestoreArrayRead_MPIDense));
1761: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetArrayWrite_C", MatDenseGetArrayWrite_MPIDense));
1762: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreArrayWrite_C", MatDenseRestoreArrayWrite_MPIDense));
1763: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDensePlaceArray_C", MatDensePlaceArray_MPIDense));
1764: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseResetArray_C", MatDenseResetArray_MPIDense));
1765: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseReplaceArray_C", MatDenseReplaceArray_MPIDense));
1766: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVec_C", MatDenseGetColumnVec_MPIDense));
1767: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVec_C", MatDenseRestoreColumnVec_MPIDense));
1768: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecRead_C", MatDenseGetColumnVecRead_MPIDense));
1769: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecRead_C", MatDenseRestoreColumnVecRead_MPIDense));
1770: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumnVecWrite_C", MatDenseGetColumnVecWrite_MPIDense));
1771: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumnVecWrite_C", MatDenseRestoreColumnVecWrite_MPIDense));
1772: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetSubMatrix_C", MatDenseGetSubMatrix_MPIDense));
1773: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreSubMatrix_C", MatDenseRestoreSubMatrix_MPIDense));
1774: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
1775: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpiaij_C", MatConvert_MPIDense_MPIAIJ));
1776: #if PetscDefined(HAVE_ELEMENTAL)
1777: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_elemental_C", MatConvert_MPIDense_Elemental));
1778: #endif
1779: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
1780: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_scalapack_C", MatConvert_Dense_ScaLAPACK));
1781: #endif
1782: #if PetscDefined(HAVE_CUDA)
1783: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensecuda_C", MatConvert_MPIDense_MPIDenseCUDA));
1784: #endif
1785: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIDenseSetPreallocation_C", MatMPIDenseSetPreallocation_MPIDense));
1786: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpidense_C", MatProductSetFromOptions_MPIAIJ_MPIDense));
1787: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaij_C", MatProductSetFromOptions_MPIDense_MPIAIJ));
1788: #if PetscDefined(HAVE_CUDA)
1789: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijcusparse_mpidense_C", MatProductSetFromOptions_MPIAIJ_MPIDense));
1790: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijcusparse_C", MatProductSetFromOptions_MPIDense_MPIAIJ));
1791: #endif
1792: #if PetscDefined(HAVE_HIP)
1793: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpidense_mpidensehip_C", MatConvert_MPIDense_MPIDenseHIP));
1794: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaijhipsparse_mpidense_C", MatProductSetFromOptions_MPIAIJ_MPIDense));
1795: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpidense_mpiaijhipsparse_C", MatProductSetFromOptions_MPIDense_MPIAIJ));
1796: #endif
1797: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseGetColumn_C", MatDenseGetColumn_MPIDense));
1798: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseRestoreColumn_C", MatDenseRestoreColumn_MPIDense));
1799: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultColumnRange_C", MatMultColumnRange_MPIDense));
1800: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultAddColumnRange_C", MatMultAddColumnRange_MPIDense));
1801: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeColumnRange_C", MatMultHermitianTransposeColumnRange_MPIDense));
1802: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMultHermitianTransposeAddColumnRange_C", MatMultHermitianTransposeAddColumnRange_MPIDense));
1803: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatGetMultPetscSF_C", MatGetMultPetscSF_MPIDense));
1804: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDenseUpdateColumnLayout_C", MatDenseUpdateColumnLayout_MPIDense));
1805: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, MATMPIDENSE));
1806: PetscFunctionReturn(PETSC_SUCCESS);
1807: }
1809: /*MC
1810: MATDENSE - MATDENSE = "dense" - A matrix type to be used for dense matrices.
1812: This matrix type is identical to `MATSEQDENSE` when constructed with a single process communicator,
1813: and `MATMPIDENSE` otherwise.
1815: Options Database Key:
1816: . -mat_type dense - sets the matrix type to `MATDENSE` during a call to `MatSetFromOptions()`
1818: Level: beginner
1820: .seealso: [](ch_matrices), `Mat`, `MATSEQDENSE`, `MATMPIDENSE`, `MATDENSECUDA`, `MATDENSEHIP`
1821: M*/
1823: /*@
1824: MatMPIDenseSetPreallocation - Sets the array used to store the matrix entries
1826: Collective
1828: Input Parameters:
1829: + B - the matrix
1830: - data - optional location of matrix data. Set to `NULL` for PETSc
1831: to control all matrix memory allocation.
1833: Level: intermediate
1835: Notes:
1836: The dense format is fully compatible with standard Fortran
1837: storage by columns.
1839: The data input variable is intended primarily for Fortran programmers
1840: who wish to allocate their own matrix memory space. Most users should
1841: set `data` to `NULL`.
1843: .seealso: [](ch_matrices), `Mat`, `MATMPIDENSE`, `MatCreate()`, `MatCreateSeqDense()`, `MatSetValues()`
1844: @*/
1845: PetscErrorCode MatMPIDenseSetPreallocation(Mat B, PetscScalar *data)
1846: {
1847: PetscFunctionBegin;
1849: PetscTryMethod(B, "MatMPIDenseSetPreallocation_C", (Mat, PetscScalar *), (B, data));
1850: PetscFunctionReturn(PETSC_SUCCESS);
1851: }
1853: /*@
1854: MatDensePlaceArray - Allows one to replace the array in a `MATDENSE` matrix with an
1855: array provided by the user. This is useful to avoid copying an array
1856: into a matrix
1858: Not Collective
1860: Input Parameters:
1861: + mat - the matrix
1862: - array - the array in column major order
1864: Level: developer
1866: Note:
1867: Adding `const` to `array` was an oversight, see notes in `VecPlaceArray()`.
1869: You can return to the original array with a call to `MatDenseResetArray()`. The user is responsible for freeing this array; it will not be
1870: freed when the matrix is destroyed.
1872: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArray()`, `MatDenseResetArray()`, `VecPlaceArray()`, `VecGetArray()`, `VecRestoreArray()`, `VecReplaceArray()`, `VecResetArray()`,
1873: `MatDenseReplaceArray()`
1874: @*/
1875: PetscErrorCode MatDensePlaceArray(Mat mat, const PetscScalar *array)
1876: {
1877: PetscFunctionBegin;
1879: PetscUseMethod(mat, "MatDensePlaceArray_C", (Mat, const PetscScalar *), (mat, array));
1880: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
1881: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
1882: mat->offloadmask = PETSC_OFFLOAD_CPU;
1883: #endif
1884: PetscFunctionReturn(PETSC_SUCCESS);
1885: }
1887: /*@
1888: MatDenseResetArray - Resets the matrix array to that it previously had before the call to `MatDensePlaceArray()`
1890: Not Collective
1892: Input Parameter:
1893: . mat - the matrix
1895: Level: developer
1897: Note:
1898: You can only call this after a call to `MatDensePlaceArray()`
1900: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatDenseGetArray()`, `MatDensePlaceArray()`, `VecPlaceArray()`, `VecGetArray()`, `VecRestoreArray()`, `VecReplaceArray()`, `VecResetArray()`
1901: @*/
1902: PetscErrorCode MatDenseResetArray(Mat mat)
1903: {
1904: PetscFunctionBegin;
1906: PetscUseMethod(mat, "MatDenseResetArray_C", (Mat), (mat));
1907: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
1908: PetscFunctionReturn(PETSC_SUCCESS);
1909: }
1911: /*@
1912: MatDenseReplaceArray - Allows one to replace the array in a dense matrix with an
1913: array provided by the user. This is useful to avoid copying an array
1914: into a matrix
1916: Not Collective
1918: Input Parameters:
1919: + mat - the matrix
1920: - array - the array in column major order
1922: Level: developer
1924: Note:
1925: Adding `const` to `array` was an oversight, see notes in `VecPlaceArray()`.
1927: The memory passed in MUST be obtained with `PetscMalloc()` and CANNOT be
1928: freed by the user. It will be freed when the matrix is destroyed.
1930: .seealso: [](ch_matrices), `Mat`, `MatDensePlaceArray()`, `MatDenseGetArray()`, `VecReplaceArray()`
1931: @*/
1932: PetscErrorCode MatDenseReplaceArray(Mat mat, const PetscScalar *array)
1933: {
1934: PetscFunctionBegin;
1936: PetscUseMethod(mat, "MatDenseReplaceArray_C", (Mat, const PetscScalar *), (mat, array));
1937: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
1938: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP)
1939: mat->offloadmask = PETSC_OFFLOAD_CPU;
1940: #endif
1941: PetscFunctionReturn(PETSC_SUCCESS);
1942: }
1944: /*@
1945: MatCreateDense - Creates a matrix in `MATDENSE` format.
1947: Collective
1949: Input Parameters:
1950: + comm - MPI communicator
1951: . m - number of local rows (or `PETSC_DECIDE` to have calculated if `M` is given)
1952: . n - number of local columns (or `PETSC_DECIDE` to have calculated if `N` is given)
1953: . M - number of global rows (or `PETSC_DECIDE` to have calculated if `m` is given)
1954: . N - number of global columns (or `PETSC_DECIDE` to have calculated if `n` is given)
1955: - data - optional location of matrix data. Set data to `NULL` (`PETSC_NULL_SCALAR_ARRAY` for Fortran users) for PETSc
1956: to control all matrix memory allocation.
1958: Output Parameter:
1959: . A - the matrix
1961: Level: intermediate
1963: Notes:
1964: The dense format is fully compatible with standard Fortran
1965: storage by columns.
1967: Although local portions of the matrix are stored in column-major
1968: order, the matrix is partitioned across MPI ranks by row.
1970: The data input variable is intended primarily for Fortran programmers
1971: who wish to allocate their own matrix memory space. Most users should
1972: set `data` to `NULL` (`PETSC_NULL_SCALAR_ARRAY` for Fortran users).
1974: The user MUST specify either the local or global matrix dimensions
1975: (possibly both).
1977: .seealso: [](ch_matrices), `Mat`, `MATDENSE`, `MatCreate()`, `MatCreateSeqDense()`, `MatSetValues()`
1978: @*/
1979: PetscErrorCode MatCreateDense(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscScalar data[], Mat *A)
1980: {
1981: PetscFunctionBegin;
1982: PetscCall(MatCreate(comm, A));
1983: PetscCall(MatSetSizes(*A, m, n, M, N));
1984: PetscCall(MatSetType(*A, MATDENSE));
1985: PetscCall(MatSeqDenseSetPreallocation(*A, data));
1986: PetscCall(MatMPIDenseSetPreallocation(*A, data));
1987: PetscFunctionReturn(PETSC_SUCCESS);
1988: }
1990: static PetscErrorCode MatDuplicate_MPIDense(Mat A, MatDuplicateOption cpvalues, Mat *newmat)
1991: {
1992: Mat mat;
1993: Mat_MPIDense *a, *oldmat = (Mat_MPIDense *)A->data;
1995: PetscFunctionBegin;
1996: *newmat = NULL;
1997: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &mat));
1998: PetscCall(MatSetSizes(mat, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
1999: PetscCall(MatSetType(mat, ((PetscObject)A)->type_name));
2000: a = (Mat_MPIDense *)mat->data;
2002: mat->factortype = A->factortype;
2003: mat->assembled = PETSC_TRUE;
2004: mat->preallocated = PETSC_TRUE;
2006: mat->insertmode = NOT_SET_VALUES;
2007: a->donotstash = oldmat->donotstash;
2009: PetscCall(PetscLayoutReference(A->rmap, &mat->rmap));
2010: PetscCall(PetscLayoutReference(A->cmap, &mat->cmap));
2012: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
2014: *newmat = mat;
2015: PetscFunctionReturn(PETSC_SUCCESS);
2016: }
2018: static PetscErrorCode MatLoad_MPIDense(Mat newMat, PetscViewer viewer)
2019: {
2020: PetscBool isbinary;
2021: #if PetscDefined(HAVE_HDF5)
2022: PetscBool ishdf5;
2023: #endif
2025: PetscFunctionBegin;
2028: /* force binary viewer to load .info file if it has not yet done so */
2029: PetscCall(PetscViewerSetUp(viewer));
2030: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
2031: #if PetscDefined(HAVE_HDF5)
2032: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
2033: #endif
2034: if (isbinary) {
2035: PetscCall(MatLoad_Dense_Binary(newMat, viewer));
2036: #if PetscDefined(HAVE_HDF5)
2037: } else if (ishdf5) {
2038: PetscCall(MatLoad_Dense_HDF5(newMat, viewer));
2039: #endif
2040: } 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);
2041: PetscFunctionReturn(PETSC_SUCCESS);
2042: }
2044: static PetscErrorCode MatEqual_MPIDense(Mat A, Mat B, PetscBool *flag)
2045: {
2046: Mat_MPIDense *matB = (Mat_MPIDense *)B->data, *matA = (Mat_MPIDense *)A->data;
2047: Mat a, b;
2049: PetscFunctionBegin;
2050: a = matA->A;
2051: b = matB->A;
2052: PetscCall(MatEqual(a, b, flag));
2053: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, flag, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2054: PetscFunctionReturn(PETSC_SUCCESS);
2055: }
2057: static PetscErrorCode MatProductCtxDestroy_MatTransMatMult_MPIDense_MPIDense(PetscCtxRt data)
2058: {
2059: MatProductCtx_TransMatMultDense *atb = *(MatProductCtx_TransMatMultDense **)data;
2061: PetscFunctionBegin;
2062: PetscCall(PetscFree2(atb->sendbuf, atb->recvcounts));
2063: PetscCall(MatDestroy(&atb->atb));
2064: PetscCall(PetscFree(atb));
2065: PetscFunctionReturn(PETSC_SUCCESS);
2066: }
2068: static PetscErrorCode MatProductCtxDestroy_MatMatTransMult_MPIDense_MPIDense(PetscCtxRt data)
2069: {
2070: MatProductCtx_MatTransMultDense *abt = *(MatProductCtx_MatTransMultDense **)data;
2072: PetscFunctionBegin;
2073: PetscCall(PetscFree2(abt->buf[0], abt->buf[1]));
2074: PetscCall(PetscFree2(abt->recvcounts, abt->recvdispls));
2075: PetscCall(PetscFree(abt));
2076: PetscFunctionReturn(PETSC_SUCCESS);
2077: }
2079: static PetscErrorCode MatTransposeMatMultNumeric_MPIDense_MPIDense(Mat A, Mat B, Mat C)
2080: {
2081: Mat_MPIDense *a = (Mat_MPIDense *)A->data, *b = (Mat_MPIDense *)B->data, *c = (Mat_MPIDense *)C->data;
2082: MatProductCtx_TransMatMultDense *atb;
2083: MPI_Comm comm;
2084: PetscMPIInt size, *recvcounts;
2085: PetscScalar *carray, *sendbuf;
2086: const PetscScalar *atbarray;
2087: PetscInt i, cN = C->cmap->N, proc, k, j, lda;
2088: const PetscInt *ranges;
2090: PetscFunctionBegin;
2091: MatCheckProduct(C, 3);
2092: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2093: atb = (MatProductCtx_TransMatMultDense *)C->product->data;
2094: recvcounts = atb->recvcounts;
2095: sendbuf = atb->sendbuf;
2097: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2098: PetscCallMPI(MPI_Comm_size(comm, &size));
2100: /* compute atbarray = aseq^T * bseq */
2101: PetscCall(MatTransposeMatMult(a->A, b->A, atb->atb ? MAT_REUSE_MATRIX : MAT_INITIAL_MATRIX, PETSC_DETERMINE, &atb->atb));
2103: PetscCall(MatGetOwnershipRanges(C, &ranges));
2105: if (ranges[1] == C->rmap->N) {
2106: /* all of the values are being reduced to rank 0: optimize this case to use MPI_Reduce and GPU aware MPI if available */
2107: PetscInt atb_lda, c_lda;
2108: Mat atb_local = atb->atb;
2109: Mat atb_alloc = NULL;
2110: Mat c_local = c->A;
2111: Mat c_alloc = NULL;
2112: PetscMemType atb_memtype, c_memtype;
2113: const PetscScalar *atb_array = NULL;
2114: MPI_Datatype vector_type;
2115: PetscScalar *c_array = NULL;
2116: PetscMPIInt rank;
2118: PetscCallMPI(MPI_Comm_rank(comm, &rank));
2120: PetscCall(MatDenseGetLDA(atb_local, &atb_lda));
2121: if (atb_lda != C->rmap->N) {
2122: // copy atb to a matrix that will have lda == the number of rows
2123: PetscCall(MatDuplicate(atb_local, MAT_DO_NOT_COPY_VALUES, &atb_alloc));
2124: PetscCall(MatCopy(atb_local, atb_alloc, DIFFERENT_NONZERO_PATTERN));
2125: atb_local = atb_alloc;
2126: }
2128: if (rank == 0) {
2129: PetscCall(MatDenseGetLDA(c_local, &c_lda));
2130: if (c_lda != C->rmap->N) {
2131: // copy c to a matrix that will have lda == the number of rows
2132: PetscCall(MatDuplicate(c_local, MAT_DO_NOT_COPY_VALUES, &c_alloc));
2133: c_local = c_alloc;
2134: }
2135: PetscCall(MatZeroEntries(c_local));
2136: }
2137: /* atb_local and c_local have nrows = lda = A->cmap->N and ncols =
2138: * B->cmap->N: use the a->Mvctx to use the best reduction method */
2139: if (!a->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A));
2140: vector_type = MPIU_SCALAR;
2141: if (B->cmap->N > 1) {
2142: PetscMPIInt mpi_N;
2144: PetscCall(PetscMPIIntCast(B->cmap->N, &mpi_N));
2145: PetscCallMPI(MPI_Type_contiguous(mpi_N, MPIU_SCALAR, &vector_type));
2146: PetscCallMPI(MPI_Type_commit(&vector_type));
2147: }
2148: PetscCall(MatDenseGetArrayReadAndMemType(atb_local, &atb_array, &atb_memtype));
2149: PetscCall(MatDenseGetArrayWriteAndMemType(c_local, &c_array, &c_memtype));
2150: PetscCall(PetscSFReduceWithMemTypeBegin(a->Mvctx, vector_type, atb_memtype, atb_array, c_memtype, c_array, MPIU_SUM));
2151: PetscCall(PetscSFReduceEnd(a->Mvctx, vector_type, atb_array, c_array, MPIU_SUM));
2152: PetscCall(MatDenseRestoreArrayWriteAndMemType(c_local, &c_array));
2153: PetscCall(MatDenseRestoreArrayReadAndMemType(atb_local, &atb_array));
2154: if (rank == 0 && c_local != c->A) PetscCall(MatCopy(c_local, c->A, DIFFERENT_NONZERO_PATTERN));
2155: if (B->cmap->N > 1) PetscCallMPI(MPI_Type_free(&vector_type));
2156: PetscCall(MatDestroy(&atb_alloc));
2157: PetscCall(MatDestroy(&c_alloc));
2158: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2159: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2160: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2161: PetscFunctionReturn(PETSC_SUCCESS);
2162: }
2164: /* arrange atbarray into sendbuf */
2165: PetscCall(MatDenseGetArrayRead(atb->atb, &atbarray));
2166: PetscCall(MatDenseGetLDA(atb->atb, &lda));
2167: for (proc = 0, k = 0; proc < size; proc++) {
2168: for (j = 0; j < cN; j++) {
2169: for (i = ranges[proc]; i < ranges[proc + 1]; i++) sendbuf[k++] = atbarray[i + j * lda];
2170: }
2171: }
2172: PetscCall(MatDenseRestoreArrayRead(atb->atb, &atbarray));
2174: /* sum all atbarray to local values of C */
2175: PetscCall(MatDenseGetArrayWrite(c->A, &carray));
2176: PetscCallMPI(MPI_Reduce_scatter(sendbuf, carray, recvcounts, MPIU_SCALAR, MPIU_SUM, comm));
2177: PetscCall(MatDenseRestoreArrayWrite(c->A, &carray));
2178: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2179: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2180: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2181: PetscFunctionReturn(PETSC_SUCCESS);
2182: }
2184: static PetscErrorCode MatTransposeMatMultSymbolic_MPIDense_MPIDense(Mat A, Mat B, PetscReal fill, Mat C)
2185: {
2186: MPI_Comm comm;
2187: PetscMPIInt size;
2188: PetscInt cm = A->cmap->n, cM, cN = B->cmap->N;
2189: MatProductCtx_TransMatMultDense *atb;
2190: PetscBool cisdense = PETSC_FALSE;
2191: const PetscInt *ranges;
2193: PetscFunctionBegin;
2194: MatCheckProduct(C, 4);
2195: PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
2196: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2197: 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,
2198: A->rmap->rend, B->rmap->rstart, B->rmap->rend);
2200: /* create matrix product C */
2201: PetscCall(MatSetSizes(C, cm, B->cmap->n, A->cmap->N, B->cmap->N));
2202: #if PetscDefined(HAVE_CUDA)
2203: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, ""));
2204: #endif
2205: #if PetscDefined(HAVE_HIP)
2206: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSEHIP, ""));
2207: #endif
2208: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2209: PetscCall(MatSetUp(C));
2211: /* create data structure for reuse C */
2212: PetscCallMPI(MPI_Comm_size(comm, &size));
2213: PetscCall(PetscNew(&atb));
2214: cM = C->rmap->N;
2215: PetscCall(PetscMalloc2(cM * cN, &atb->sendbuf, size, &atb->recvcounts));
2216: PetscCall(MatGetOwnershipRanges(C, &ranges));
2217: for (PetscMPIInt i = 0; i < size; i++) PetscCall(PetscMPIIntCast((ranges[i + 1] - ranges[i]) * cN, &atb->recvcounts[i]));
2218: C->product->data = atb;
2219: C->product->destroy = MatProductCtxDestroy_MatTransMatMult_MPIDense_MPIDense;
2220: PetscFunctionReturn(PETSC_SUCCESS);
2221: }
2223: static PetscErrorCode MatMatTransposeMultSymbolic_MPIDense_MPIDense(Mat A, Mat B, PetscReal fill, Mat C)
2224: {
2225: MPI_Comm comm;
2226: PetscMPIInt i, size;
2227: PetscInt maxRows, bufsiz;
2228: PetscMPIInt tag;
2229: PetscInt alg;
2230: MatProductCtx_MatTransMultDense *abt;
2231: Mat_Product *product = C->product;
2232: PetscBool flg;
2234: PetscFunctionBegin;
2235: MatCheckProduct(C, 4);
2236: PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
2237: /* check local size of A and B */
2238: 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);
2240: PetscCall(PetscStrcmp(product->alg, "allgatherv", &flg));
2241: alg = flg ? 0 : 1;
2243: /* setup matrix product C */
2244: PetscCall(MatSetSizes(C, A->rmap->n, B->rmap->n, A->rmap->N, B->rmap->N));
2245: PetscCall(MatSetType(C, MATMPIDENSE));
2246: PetscCall(MatSetUp(C));
2247: PetscCall(PetscObjectGetNewTag((PetscObject)C, &tag));
2249: /* create data structure for reuse C */
2250: PetscCall(PetscObjectGetComm((PetscObject)C, &comm));
2251: PetscCallMPI(MPI_Comm_size(comm, &size));
2252: PetscCall(PetscNew(&abt));
2253: abt->tag = tag;
2254: abt->alg = alg;
2255: switch (alg) {
2256: case 1: /* alg: "cyclic" */
2257: for (maxRows = 0, i = 0; i < size; i++) maxRows = PetscMax(maxRows, B->rmap->range[i + 1] - B->rmap->range[i]);
2258: bufsiz = A->cmap->N * maxRows;
2259: PetscCall(PetscMalloc2(bufsiz, &abt->buf[0], bufsiz, &abt->buf[1]));
2260: break;
2261: default: /* alg: "allgatherv" */
2262: PetscCall(PetscMalloc2(B->rmap->n * B->cmap->N, &abt->buf[0], B->rmap->N * B->cmap->N, &abt->buf[1]));
2263: PetscCall(PetscMalloc2(size, &abt->recvcounts, size + 1, &abt->recvdispls));
2264: for (i = 0; i <= size; i++) PetscCall(PetscMPIIntCast(B->rmap->range[i] * A->cmap->N, &abt->recvdispls[i]));
2265: for (i = 0; i < size; i++) PetscCall(PetscMPIIntCast(abt->recvdispls[i + 1] - abt->recvdispls[i], &abt->recvcounts[i]));
2266: break;
2267: }
2269: C->product->data = abt;
2270: C->product->destroy = MatProductCtxDestroy_MatMatTransMult_MPIDense_MPIDense;
2271: C->ops->mattransposemultnumeric = MatMatTransposeMultNumeric_MPIDense_MPIDense;
2272: PetscFunctionReturn(PETSC_SUCCESS);
2273: }
2275: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense_Cyclic(Mat A, Mat B, Mat C)
2276: {
2277: Mat_MPIDense *a = (Mat_MPIDense *)A->data, *b = (Mat_MPIDense *)B->data, *c = (Mat_MPIDense *)C->data;
2278: MatProductCtx_MatTransMultDense *abt;
2279: MPI_Comm comm;
2280: PetscMPIInt rank, size, sendto, recvfrom, recvisfrom;
2281: PetscScalar *sendbuf, *recvbuf = NULL, *cv;
2282: PetscInt i, cK = A->cmap->N, sendsiz, recvsiz, k, j, bn;
2283: PetscScalar _DOne = 1.0, _DZero = 0.0;
2284: const PetscScalar *av, *bv;
2285: PetscBLASInt cm, cn, ck, alda, blda = 0, clda;
2286: MPI_Request reqs[2];
2287: const PetscInt *ranges;
2289: PetscFunctionBegin;
2290: MatCheckProduct(C, 3);
2291: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2292: abt = (MatProductCtx_MatTransMultDense *)C->product->data;
2293: PetscCall(PetscObjectGetComm((PetscObject)C, &comm));
2294: PetscCallMPI(MPI_Comm_rank(comm, &rank));
2295: PetscCallMPI(MPI_Comm_size(comm, &size));
2296: PetscCall(MatDenseGetArrayRead(a->A, &av));
2297: PetscCall(MatDenseGetArrayRead(b->A, &bv));
2298: PetscCall(MatDenseGetArrayWrite(c->A, &cv));
2299: PetscCall(MatDenseGetLDA(a->A, &i));
2300: PetscCall(PetscBLASIntCast(i, &alda));
2301: PetscCall(MatDenseGetLDA(b->A, &i));
2302: PetscCall(PetscBLASIntCast(i, &blda));
2303: PetscCall(MatDenseGetLDA(c->A, &i));
2304: PetscCall(PetscBLASIntCast(i, &clda));
2305: PetscCall(MatGetOwnershipRanges(B, &ranges));
2306: bn = B->rmap->n;
2307: if (blda == bn) {
2308: sendbuf = (PetscScalar *)bv;
2309: } else {
2310: sendbuf = abt->buf[0];
2311: for (k = 0, i = 0; i < cK; i++) {
2312: for (j = 0; j < bn; j++, k++) sendbuf[k] = bv[i * blda + j];
2313: }
2314: }
2315: if (size > 1) {
2316: sendto = (rank + size - 1) % size;
2317: recvfrom = (rank + size + 1) % size;
2318: } else {
2319: sendto = recvfrom = 0;
2320: }
2321: PetscCall(PetscBLASIntCast(cK, &ck));
2322: PetscCall(PetscBLASIntCast(c->A->rmap->n, &cm));
2323: recvisfrom = rank;
2324: for (i = 0; i < size; i++) {
2325: /* we have finished receiving in sending, bufs can be read/modified */
2326: PetscMPIInt nextrecvisfrom = (recvisfrom + 1) % size; /* which process the next recvbuf will originate on */
2327: PetscInt nextbn = ranges[nextrecvisfrom + 1] - ranges[nextrecvisfrom];
2329: if (nextrecvisfrom != rank) {
2330: /* start the cyclic sends from sendbuf, to recvbuf (which will switch to sendbuf) */
2331: sendsiz = cK * bn;
2332: recvsiz = cK * nextbn;
2333: recvbuf = (i & 1) ? abt->buf[0] : abt->buf[1];
2334: PetscCallMPI(MPIU_Isend(sendbuf, sendsiz, MPIU_SCALAR, sendto, abt->tag, comm, &reqs[0]));
2335: PetscCallMPI(MPIU_Irecv(recvbuf, recvsiz, MPIU_SCALAR, recvfrom, abt->tag, comm, &reqs[1]));
2336: }
2338: /* local aseq * sendbuf^T */
2339: PetscCall(PetscBLASIntCast(ranges[recvisfrom + 1] - ranges[recvisfrom], &cn));
2340: if (cm && cn && ck) PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &cm, &cn, &ck, &_DOne, av, &alda, sendbuf, &cn, &_DZero, cv + clda * ranges[recvisfrom], &clda));
2342: if (nextrecvisfrom != rank) {
2343: /* wait for the sends and receives to complete, swap sendbuf and recvbuf */
2344: PetscCallMPI(MPI_Waitall(2, reqs, MPI_STATUSES_IGNORE));
2345: }
2346: bn = nextbn;
2347: recvisfrom = nextrecvisfrom;
2348: sendbuf = recvbuf;
2349: }
2350: PetscCall(MatDenseRestoreArrayRead(a->A, &av));
2351: PetscCall(MatDenseRestoreArrayRead(b->A, &bv));
2352: PetscCall(MatDenseRestoreArrayWrite(c->A, &cv));
2353: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2354: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2355: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2356: PetscFunctionReturn(PETSC_SUCCESS);
2357: }
2359: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense_Allgatherv(Mat A, Mat B, Mat C)
2360: {
2361: Mat_MPIDense *a = (Mat_MPIDense *)A->data, *b = (Mat_MPIDense *)B->data, *c = (Mat_MPIDense *)C->data;
2362: MatProductCtx_MatTransMultDense *abt;
2363: MPI_Comm comm;
2364: PetscMPIInt size, ibn;
2365: PetscScalar *cv, *sendbuf, *recvbuf;
2366: const PetscScalar *av, *bv;
2367: PetscInt blda, i, cK = A->cmap->N, k, j, bn;
2368: PetscScalar _DOne = 1.0, _DZero = 0.0;
2369: PetscBLASInt cm, cn, ck, alda, clda;
2371: PetscFunctionBegin;
2372: MatCheckProduct(C, 3);
2373: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2374: abt = (MatProductCtx_MatTransMultDense *)C->product->data;
2375: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2376: PetscCallMPI(MPI_Comm_size(comm, &size));
2377: PetscCall(MatDenseGetArrayRead(a->A, &av));
2378: PetscCall(MatDenseGetArrayRead(b->A, &bv));
2379: PetscCall(MatDenseGetArrayWrite(c->A, &cv));
2380: PetscCall(MatDenseGetLDA(a->A, &i));
2381: PetscCall(PetscBLASIntCast(i, &alda));
2382: PetscCall(MatDenseGetLDA(b->A, &blda));
2383: PetscCall(MatDenseGetLDA(c->A, &i));
2384: PetscCall(PetscBLASIntCast(i, &clda));
2385: /* copy transpose of B into buf[0] */
2386: bn = B->rmap->n;
2387: sendbuf = abt->buf[0];
2388: recvbuf = abt->buf[1];
2389: for (k = 0, j = 0; j < bn; j++) {
2390: for (i = 0; i < cK; i++, k++) sendbuf[k] = bv[i * blda + j];
2391: }
2392: PetscCall(MatDenseRestoreArrayRead(b->A, &bv));
2393: PetscCall(PetscMPIIntCast(bn * cK, &ibn));
2394: PetscCallMPI(MPI_Allgatherv(sendbuf, ibn, MPIU_SCALAR, recvbuf, abt->recvcounts, abt->recvdispls, MPIU_SCALAR, comm));
2395: PetscCall(PetscBLASIntCast(cK, &ck));
2396: PetscCall(PetscBLASIntCast(c->A->rmap->n, &cm));
2397: PetscCall(PetscBLASIntCast(c->A->cmap->n, &cn));
2398: if (cm && cn && ck) PetscCallBLAS("BLASgemm", BLASgemm_("N", "N", &cm, &cn, &ck, &_DOne, av, &alda, recvbuf, &ck, &_DZero, cv, &clda));
2399: PetscCall(MatDenseRestoreArrayRead(a->A, &av));
2400: PetscCall(MatDenseRestoreArrayRead(b->A, &bv));
2401: PetscCall(MatDenseRestoreArrayWrite(c->A, &cv));
2402: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2403: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2404: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2405: PetscFunctionReturn(PETSC_SUCCESS);
2406: }
2408: static PetscErrorCode MatMatTransposeMultNumeric_MPIDense_MPIDense(Mat A, Mat B, Mat C)
2409: {
2410: MatProductCtx_MatTransMultDense *abt;
2412: PetscFunctionBegin;
2413: MatCheckProduct(C, 3);
2414: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
2415: abt = (MatProductCtx_MatTransMultDense *)C->product->data;
2416: switch (abt->alg) {
2417: case 1:
2418: PetscCall(MatMatTransposeMultNumeric_MPIDense_MPIDense_Cyclic(A, B, C));
2419: break;
2420: default:
2421: PetscCall(MatMatTransposeMultNumeric_MPIDense_MPIDense_Allgatherv(A, B, C));
2422: break;
2423: }
2424: PetscFunctionReturn(PETSC_SUCCESS);
2425: }
2427: static PetscErrorCode MatProductCtxDestroy_MatMatMult_MPIDense_MPIDense(PetscCtxRt data)
2428: {
2429: MatProductCtx_MatMultDense *ab = *(MatProductCtx_MatMultDense **)data;
2431: PetscFunctionBegin;
2432: PetscCall(MatDestroy(&ab->Ce));
2433: PetscCall(MatDestroy(&ab->Ae));
2434: PetscCall(MatDestroy(&ab->Be));
2435: PetscCall(PetscFree(ab));
2436: PetscFunctionReturn(PETSC_SUCCESS);
2437: }
2439: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIDense(Mat A, Mat B, Mat C)
2440: {
2441: MatProductCtx_MatMultDense *ab;
2442: Mat_MPIDense *mdn = (Mat_MPIDense *)A->data;
2443: Mat_MPIDense *b = (Mat_MPIDense *)B->data;
2445: PetscFunctionBegin;
2446: MatCheckProduct(C, 3);
2447: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Missing product data");
2448: ab = (MatProductCtx_MatMultDense *)C->product->data;
2449: if (ab->Ae && ab->Ce) {
2450: #if PetscDefined(HAVE_ELEMENTAL)
2451: PetscCall(MatConvert_MPIDense_Elemental(A, MATELEMENTAL, MAT_REUSE_MATRIX, &ab->Ae));
2452: PetscCall(MatConvert_MPIDense_Elemental(B, MATELEMENTAL, MAT_REUSE_MATRIX, &ab->Be));
2453: PetscCall(MatMatMultNumeric_Elemental(ab->Ae, ab->Be, ab->Ce));
2454: PetscCall(MatConvert(ab->Ce, MATMPIDENSE, MAT_REUSE_MATRIX, &C));
2455: #else
2456: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "PETSC_HAVE_ELEMENTAL not defined");
2457: #endif
2458: } else {
2459: MPI_Comm comm;
2460: const PetscScalar *read;
2461: PetscScalar *write;
2462: PetscInt lda;
2463: const PetscInt *ranges;
2464: PetscMPIInt size;
2466: if (!mdn->Mvctx) PetscCall(MatSetUpMultiply_MPIDense(A)); /* cannot be done during the symbolic phase because of possible calls to MatProductReplaceMats() */
2467: comm = PetscObjectComm((PetscObject)B);
2468: PetscCallMPI(MPI_Comm_size(comm, &size));
2469: PetscCall(PetscLayoutGetRanges(B->rmap, &ranges));
2470: if (ranges[1] == ranges[size]) {
2471: // optimize for the case where the B matrix is broadcast from rank 0
2472: PetscInt b_lda, be_lda;
2473: Mat b_local = b->A;
2474: Mat b_alloc = NULL;
2475: Mat be_local = ab->Be;
2476: Mat be_alloc = NULL;
2477: PetscMemType b_memtype, be_memtype;
2478: const PetscScalar *b_array = NULL;
2479: MPI_Datatype vector_type;
2480: PetscScalar *be_array = NULL;
2481: PetscMPIInt rank;
2483: PetscCallMPI(MPI_Comm_rank(comm, &rank));
2484: PetscCall(MatDenseGetLDA(be_local, &be_lda));
2485: if (be_lda != B->rmap->N) {
2486: PetscCall(MatDuplicate(be_local, MAT_DO_NOT_COPY_VALUES, &be_alloc));
2487: be_local = be_alloc;
2488: }
2490: if (rank == 0) {
2491: PetscCall(MatDenseGetLDA(b_local, &b_lda));
2492: if (b_lda != B->rmap->N) {
2493: PetscCall(MatDuplicate(b_local, MAT_DO_NOT_COPY_VALUES, &b_alloc));
2494: PetscCall(MatCopy(b_local, b_alloc, DIFFERENT_NONZERO_PATTERN));
2495: b_local = b_alloc;
2496: }
2497: }
2498: vector_type = MPIU_SCALAR;
2499: if (B->cmap->N > 1) {
2500: PetscMPIInt mpi_N;
2502: PetscCall(PetscMPIIntCast(B->cmap->N, &mpi_N));
2503: PetscCallMPI(MPI_Type_contiguous(mpi_N, MPIU_SCALAR, &vector_type));
2504: PetscCallMPI(MPI_Type_commit(&vector_type));
2505: }
2506: PetscCall(MatDenseGetArrayReadAndMemType(b_local, &b_array, &b_memtype));
2507: PetscCall(MatDenseGetArrayWriteAndMemType(be_local, &be_array, &be_memtype));
2508: PetscCall(PetscSFBcastWithMemTypeBegin(mdn->Mvctx, vector_type, b_memtype, b_array, be_memtype, be_array, MPI_REPLACE));
2509: PetscCall(PetscSFBcastEnd(mdn->Mvctx, vector_type, b_array, be_array, MPI_REPLACE));
2510: PetscCall(MatDenseRestoreArrayWriteAndMemType(be_local, &be_array));
2511: PetscCall(MatDenseRestoreArrayReadAndMemType(b_local, &b_array));
2512: if (be_local != ab->Be) PetscCall(MatCopy(be_local, ab->Be, DIFFERENT_NONZERO_PATTERN));
2513: if (B->cmap->N > 1) PetscCallMPI(MPI_Type_free(&vector_type));
2514: PetscCall(MatDestroy(&be_alloc));
2515: PetscCall(MatDestroy(&b_alloc));
2516: } else {
2517: PetscCall(MatDenseGetLDA(B, &lda));
2518: PetscCall(MatDenseGetArrayRead(B, &read));
2519: PetscCall(MatDenseGetArrayWrite(ab->Be, &write));
2520: for (PetscInt i = 0; i < C->cmap->N; ++i) {
2521: PetscCall(PetscSFBcastBegin(mdn->Mvctx, MPIU_SCALAR, read + i * lda, write + i * ab->Be->rmap->n, MPI_REPLACE));
2522: PetscCall(PetscSFBcastEnd(mdn->Mvctx, MPIU_SCALAR, read + i * lda, write + i * ab->Be->rmap->n, MPI_REPLACE));
2523: }
2524: PetscCall(MatDenseRestoreArrayWrite(ab->Be, &write));
2525: PetscCall(MatDenseRestoreArrayRead(B, &read));
2526: }
2527: PetscCall(MatMatMultNumeric_SeqDense_SeqDense(((Mat_MPIDense *)A->data)->A, ab->Be, ((Mat_MPIDense *)C->data)->A));
2528: }
2529: PetscFunctionReturn(PETSC_SUCCESS);
2530: }
2532: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIDense(Mat A, Mat B, PetscReal fill, Mat C)
2533: {
2534: Mat_Product *product = C->product;
2535: PetscInt alg;
2536: MatProductCtx_MatMultDense *ab;
2537: PetscBool flg;
2539: PetscFunctionBegin;
2540: MatCheckProduct(C, 4);
2541: PetscCheck(!C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
2542: /* check local size of A and B */
2543: 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 ")",
2544: A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend);
2546: PetscCall(PetscStrcmp(product->alg, "petsc", &flg));
2547: alg = flg ? 0 : 1;
2549: /* setup C */
2550: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
2551: PetscCall(MatSetType(C, MATMPIDENSE));
2552: PetscCall(MatSetUp(C));
2554: /* create data structure for reuse Cdense */
2555: PetscCall(PetscNew(&ab));
2557: switch (alg) {
2558: case 1: /* alg: "elemental" */
2559: #if PetscDefined(HAVE_ELEMENTAL)
2560: /* create elemental matrices Ae and Be */
2561: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &ab->Ae));
2562: PetscCall(MatSetSizes(ab->Ae, PETSC_DECIDE, PETSC_DECIDE, A->rmap->N, A->cmap->N));
2563: PetscCall(MatSetType(ab->Ae, MATELEMENTAL));
2564: PetscCall(MatSetUp(ab->Ae));
2565: PetscCall(MatSetOption(ab->Ae, MAT_ROW_ORIENTED, PETSC_FALSE));
2567: PetscCall(MatCreate(PetscObjectComm((PetscObject)B), &ab->Be));
2568: PetscCall(MatSetSizes(ab->Be, PETSC_DECIDE, PETSC_DECIDE, B->rmap->N, B->cmap->N));
2569: PetscCall(MatSetType(ab->Be, MATELEMENTAL));
2570: PetscCall(MatSetUp(ab->Be));
2571: PetscCall(MatSetOption(ab->Be, MAT_ROW_ORIENTED, PETSC_FALSE));
2573: /* compute symbolic Ce = Ae*Be */
2574: PetscCall(MatCreate(PetscObjectComm((PetscObject)C), &ab->Ce));
2575: PetscCall(MatMatMultSymbolic_Elemental(ab->Ae, ab->Be, fill, ab->Ce));
2576: #else
2577: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "PETSC_HAVE_ELEMENTAL not defined");
2578: #endif
2579: break;
2580: default: /* alg: "petsc" */
2581: ab->Ae = NULL;
2582: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, A->cmap->N, B->cmap->N, NULL, &ab->Be));
2583: ab->Ce = NULL;
2584: break;
2585: }
2587: C->product->data = ab;
2588: C->product->destroy = MatProductCtxDestroy_MatMatMult_MPIDense_MPIDense;
2589: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIDense;
2590: PetscFunctionReturn(PETSC_SUCCESS);
2591: }
2593: static PetscErrorCode MatProductSetFromOptions_MPIDense_AB(Mat C)
2594: {
2595: Mat_Product *product = C->product;
2596: const char *algTypes[2] = {"petsc", "elemental"};
2597: PetscInt alg, nalg = PetscDefined(HAVE_ELEMENTAL) ? 2 : 1;
2598: PetscBool flg = PETSC_FALSE;
2600: PetscFunctionBegin;
2601: /* Set default algorithm */
2602: alg = 0; /* default is PETSc */
2603: PetscCall(PetscStrcmp(product->alg, "default", &flg));
2604: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2606: /* Get runtime option */
2607: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
2608: PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatProduct_AB", algTypes, nalg, algTypes[alg], &alg, &flg));
2609: PetscOptionsEnd();
2610: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2612: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIDense;
2613: C->ops->productsymbolic = MatProductSymbolic_AB;
2614: PetscFunctionReturn(PETSC_SUCCESS);
2615: }
2617: static PetscErrorCode MatProductSetFromOptions_MPIDense_AtB(Mat C)
2618: {
2619: Mat_Product *product = C->product;
2620: Mat A = product->A, B = product->B;
2622: PetscFunctionBegin;
2623: 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 ")",
2624: A->rmap->rstart, A->rmap->rend, B->rmap->rstart, B->rmap->rend);
2625: C->ops->transposematmultsymbolic = MatTransposeMatMultSymbolic_MPIDense_MPIDense;
2626: C->ops->productsymbolic = MatProductSymbolic_AtB;
2627: PetscFunctionReturn(PETSC_SUCCESS);
2628: }
2630: static PetscErrorCode MatProductSetFromOptions_MPIDense_ABt(Mat C)
2631: {
2632: Mat_Product *product = C->product;
2633: const char *algTypes[2] = {"allgatherv", "cyclic"};
2634: PetscInt alg, nalg = 2;
2635: PetscBool flg = PETSC_FALSE;
2637: PetscFunctionBegin;
2638: /* Set default algorithm */
2639: alg = 0; /* default is allgatherv */
2640: PetscCall(PetscStrcmp(product->alg, "default", &flg));
2641: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2643: /* Get runtime option */
2644: if (product->api_user) {
2645: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatTransposeMult", "Mat");
2646: PetscCall(PetscOptionsEList("-matmattransmult_mpidense_mpidense_via", "Algorithmic approach", "MatMatTransposeMult", algTypes, nalg, algTypes[alg], &alg, &flg));
2647: PetscOptionsEnd();
2648: } else {
2649: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_ABt", "Mat");
2650: PetscCall(PetscOptionsEList("-mat_product_algorithm", "Algorithmic approach", "MatProduct_ABt", algTypes, nalg, algTypes[alg], &alg, &flg));
2651: PetscOptionsEnd();
2652: }
2653: if (flg) PetscCall(MatProductSetAlgorithm(C, algTypes[alg]));
2655: C->ops->mattransposemultsymbolic = MatMatTransposeMultSymbolic_MPIDense_MPIDense;
2656: C->ops->productsymbolic = MatProductSymbolic_ABt;
2657: PetscFunctionReturn(PETSC_SUCCESS);
2658: }
2660: static PetscErrorCode MatProductSetFromOptions_MPIDense(Mat C)
2661: {
2662: Mat_Product *product = C->product;
2664: PetscFunctionBegin;
2665: switch (product->type) {
2666: case MATPRODUCT_AB:
2667: PetscCall(MatProductSetFromOptions_MPIDense_AB(C));
2668: break;
2669: case MATPRODUCT_AtB:
2670: PetscCall(MatProductSetFromOptions_MPIDense_AtB(C));
2671: break;
2672: case MATPRODUCT_ABt:
2673: PetscCall(MatProductSetFromOptions_MPIDense_ABt(C));
2674: break;
2675: default:
2676: break;
2677: }
2678: PetscFunctionReturn(PETSC_SUCCESS);
2679: }
2681: PetscErrorCode MatDenseScatter_Private(PetscSF sf, Mat X, Mat Y, InsertMode mode, ScatterMode smode)
2682: {
2683: const PetscScalar *in;
2684: PetscScalar *out;
2685: PetscSF vsf;
2686: PetscInt N, ny, rld, lld;
2687: PetscMemType mtype[2];
2688: MPI_Op op = MPI_OP_NULL;
2690: PetscFunctionBegin;
2694: if (mode == INSERT_VALUES) op = MPI_REPLACE;
2695: else if (mode == ADD_VALUES) op = MPIU_SUM;
2696: else if (mode == MAX_VALUES) op = MPIU_MAX;
2697: else if (mode == MIN_VALUES) op = MPIU_MIN;
2698: PetscCheck(op != MPI_OP_NULL, PetscObjectComm((PetscObject)sf), PETSC_ERR_SUP, "Unsupported InsertMode %d in MatDenseScatter_Private()", mode);
2699: PetscCheck(smode == SCATTER_FORWARD || smode == SCATTER_REVERSE, PetscObjectComm((PetscObject)sf), PETSC_ERR_SUP, "Unsupported ScatterMode %d in MatDenseScatter_Private()", smode);
2700: PetscCall(MatGetSize(X, NULL, &N));
2701: PetscCall(MatGetSize(Y, NULL, &ny));
2702: PetscCheck(N == ny, PetscObjectComm((PetscObject)sf), PETSC_ERR_ARG_SIZ, "Matrix column sizes must match: %" PetscInt_FMT " != %" PetscInt_FMT, N, ny);
2703: PetscCall(MatDenseGetLDA(X, &rld));
2704: PetscCall(MatDenseGetLDA(Y, &lld));
2705: /* get cached or create new strided PetscSF when the number of columns is greater than one */
2706: if (N > 1) {
2707: PetscCall(PetscObjectQuery((PetscObject)sf, "_MatDenseScatter_StridedSF", (PetscObject *)&vsf));
2708: if (vsf) {
2709: PetscInt nr[2], nl[2];
2711: PetscCall(PetscSFGetGraph(sf, nr, nl, NULL, NULL));
2712: PetscCall(PetscSFGetGraph(vsf, nr + 1, nl + 1, NULL, NULL));
2713: if (N * nr[0] != nr[1] || N * nl[0] != nl[1]) vsf = NULL;
2714: }
2715: if (!vsf) {
2716: PetscCall(PetscSFCreateStridedSF(sf, N, rld, lld, &vsf));
2717: PetscCall(PetscObjectCompose((PetscObject)sf, "_MatDenseScatter_StridedSF", (PetscObject)vsf));
2718: PetscCall(PetscObjectDereference((PetscObject)vsf));
2719: }
2720: } else vsf = sf;
2721: /* the output array is accessed in read and write mode,
2722: but write-only in the INSERT_VALUES case could be worth exploring */
2723: PetscCall(MatDenseGetArrayReadAndMemType(X, &in, &mtype[0]));
2724: PetscCall(MatDenseGetArrayAndMemType(Y, &out, &mtype[1]));
2725: if (smode == SCATTER_FORWARD) {
2726: PetscCall(PetscSFBcastWithMemTypeBegin(vsf, vsf->vscat.unit, mtype[0], in, mtype[1], out, op));
2727: PetscCall(PetscSFBcastEnd(vsf, vsf->vscat.unit, in, out, op));
2728: } else {
2729: PetscCall(PetscSFReduceWithMemTypeBegin(vsf, vsf->vscat.unit, mtype[0], in, mtype[1], out, op));
2730: PetscCall(PetscSFReduceEnd(vsf, vsf->vscat.unit, in, out, op));
2731: }
2732: PetscCall(MatDenseRestoreArrayAndMemType(Y, &out));
2733: PetscCall(MatDenseRestoreArrayReadAndMemType(X, &in));
2734: PetscFunctionReturn(PETSC_SUCCESS);
2735: }