Actual source code: mpibaij.c
1: #include <../src/mat/impls/baij/mpi/mpibaij.h>
3: #include <petsc/private/hashseti.h>
4: #include <petscblaslapack.h>
5: #include <petscsf.h>
7: static PetscErrorCode MatDestroy_MPIBAIJ(Mat mat)
8: {
9: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
11: PetscFunctionBegin;
12: PetscCall(PetscLogObjectState((PetscObject)mat, "Rows=%" PetscInt_FMT ",Cols=%" PetscInt_FMT, mat->rmap->N, mat->cmap->N));
13: PetscCall(MatStashDestroy_Private(&mat->stash));
14: PetscCall(MatStashDestroy_Private(&mat->bstash));
15: PetscCall(MatDestroy(&baij->A));
16: PetscCall(MatDestroy(&baij->B));
17: #if defined(PETSC_USE_CTABLE)
18: PetscCall(PetscHMapIDestroy(&baij->colmap));
19: #else
20: PetscCall(PetscFree(baij->colmap));
21: #endif
22: PetscCall(PetscFree(baij->garray));
23: PetscCall(VecDestroy(&baij->lvec));
24: PetscCall(VecScatterDestroy(&baij->Mvctx));
25: PetscCall(PetscFree2(baij->rowvalues, baij->rowindices));
26: PetscCall(PetscFree(baij->barray));
27: PetscCall(PetscFree2(baij->hd, baij->ht));
28: PetscCall(PetscFree(baij->rangebs));
29: PetscCall(PetscFree(mat->data));
31: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, NULL));
32: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatStoreValues_C", NULL));
33: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatRetrieveValues_C", NULL));
34: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIBAIJSetPreallocation_C", NULL));
35: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIBAIJSetPreallocationCSR_C", NULL));
36: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDiagonalScaleLocal_C", NULL));
37: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetHashTableFactor_C", NULL));
38: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpibaij_mpisbaij_C", NULL));
39: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpibaij_mpiadj_C", NULL));
40: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpibaij_mpiaij_C", NULL));
41: #if defined(PETSC_HAVE_HYPRE)
42: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpibaij_hypre_C", NULL));
43: #endif
44: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpibaij_is_C", NULL));
45: PetscFunctionReturn(PETSC_SUCCESS);
46: }
48: /* defines MatSetValues_MPI_Hash(), MatAssemblyBegin_MPI_Hash(), and MatAssemblyEnd_MPI_Hash() */
49: #define TYPE BAIJ
50: #include "../src/mat/impls/aij/mpi/mpihashmat.h"
51: #undef TYPE
53: #if defined(PETSC_HAVE_HYPRE)
54: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
55: #endif
57: static PetscErrorCode MatGetRowMaxAbs_MPIBAIJ(Mat A, Vec v, PetscInt idx[])
58: {
59: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
60: PetscInt i, *idxb = NULL, m = A->rmap->n, bs = A->cmap->bs;
61: PetscScalar *vv;
62: Vec vB, vA;
63: const PetscScalar *va, *vb;
65: PetscFunctionBegin;
66: PetscCall(MatCreateVecs(a->A, NULL, &vA));
67: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
69: PetscCall(VecGetArrayRead(vA, &va));
70: if (idx) {
71: for (i = 0; i < m; i++) {
72: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
73: }
74: }
76: PetscCall(MatCreateVecs(a->B, NULL, &vB));
77: PetscCall(PetscMalloc1(m, &idxb));
78: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
80: PetscCall(VecGetArrayWrite(v, &vv));
81: PetscCall(VecGetArrayRead(vB, &vb));
82: for (i = 0; i < m; i++) {
83: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
84: vv[i] = vb[i];
85: if (idx) idx[i] = bs * a->garray[idxb[i] / bs] + (idxb[i] % bs);
86: } else {
87: vv[i] = va[i];
88: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > bs * a->garray[idxb[i] / bs] + (idxb[i] % bs)) idx[i] = bs * a->garray[idxb[i] / bs] + (idxb[i] % bs);
89: }
90: }
91: PetscCall(VecRestoreArrayWrite(v, &vv));
92: PetscCall(VecRestoreArrayRead(vA, &va));
93: PetscCall(VecRestoreArrayRead(vB, &vb));
94: PetscCall(PetscFree(idxb));
95: PetscCall(VecDestroy(&vA));
96: PetscCall(VecDestroy(&vB));
97: PetscFunctionReturn(PETSC_SUCCESS);
98: }
100: static PetscErrorCode MatGetRowSumAbs_MPIBAIJ(Mat A, Vec v)
101: {
102: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
103: Vec vB, vA;
105: PetscFunctionBegin;
106: PetscCall(MatCreateVecs(a->A, NULL, &vA));
107: PetscCall(MatGetRowSumAbs(a->A, vA));
108: PetscCall(MatCreateVecs(a->B, NULL, &vB));
109: PetscCall(MatGetRowSumAbs(a->B, vB));
110: PetscCall(VecAXPY(vA, 1.0, vB));
111: PetscCall(VecDestroy(&vB));
112: PetscCall(VecCopy(vA, v));
113: PetscCall(VecDestroy(&vA));
114: PetscFunctionReturn(PETSC_SUCCESS);
115: }
117: static PetscErrorCode MatStoreValues_MPIBAIJ(Mat mat)
118: {
119: Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data;
121: PetscFunctionBegin;
122: PetscCall(MatStoreValues(aij->A));
123: PetscCall(MatStoreValues(aij->B));
124: PetscFunctionReturn(PETSC_SUCCESS);
125: }
127: static PetscErrorCode MatRetrieveValues_MPIBAIJ(Mat mat)
128: {
129: Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data;
131: PetscFunctionBegin;
132: PetscCall(MatRetrieveValues(aij->A));
133: PetscCall(MatRetrieveValues(aij->B));
134: PetscFunctionReturn(PETSC_SUCCESS);
135: }
137: /*
138: Local utility routine that creates a mapping from the global column
139: number to the local number in the off-diagonal part of the local
140: storage of the matrix. This is done in a non scalable way since the
141: length of colmap equals the global matrix length.
142: */
143: PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat mat)
144: {
145: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
146: Mat_SeqBAIJ *B = (Mat_SeqBAIJ *)baij->B->data;
147: PetscInt nbs = B->nbs, i, bs = mat->rmap->bs;
149: PetscFunctionBegin;
150: #if defined(PETSC_USE_CTABLE)
151: PetscCall(PetscHMapICreateWithSize(baij->nbs, &baij->colmap));
152: for (i = 0; i < nbs; i++) PetscCall(PetscHMapISet(baij->colmap, baij->garray[i] + 1, i * bs + 1));
153: #else
154: PetscCall(PetscCalloc1(baij->Nbs + 1, &baij->colmap));
155: for (i = 0; i < nbs; i++) baij->colmap[baij->garray[i]] = i * bs + 1;
156: #endif
157: PetscFunctionReturn(PETSC_SUCCESS);
158: }
160: #define MatSetValues_SeqBAIJ_A_Private(row, col, value, addv, orow, ocol) \
161: do { \
162: brow = row / bs; \
163: rp = PetscSafePointerPlusOffset(aj, ai[brow]); \
164: ap = PetscSafePointerPlusOffset(aa, bs2 * ai[brow]); \
165: rmax = aimax[brow]; \
166: nrow = ailen[brow]; \
167: bcol = col / bs; \
168: ridx = row % bs; \
169: cidx = col % bs; \
170: low = 0; \
171: high = nrow; \
172: while (high - low > 3) { \
173: t = (low + high) / 2; \
174: if (rp[t] > bcol) high = t; \
175: else low = t; \
176: } \
177: for (_i = low; _i < high; _i++) { \
178: if (rp[_i] > bcol) break; \
179: if (rp[_i] == bcol) { \
180: bap = ap + bs2 * _i + bs * cidx + ridx; \
181: if (addv == ADD_VALUES) *bap += value; \
182: else *bap = value; \
183: goto a_noinsert; \
184: } \
185: } \
186: if (a->nonew == 1) goto a_noinsert; \
187: PetscCheck(a->nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", orow, ocol); \
188: MatSeqXAIJReallocateAIJ(A, a->mbs, bs2, nrow, brow, bcol, rmax, aa, ai, aj, rp, ap, aimax, a->nonew, MatScalar); \
189: N = nrow++ - 1; \
190: /* shift up all the later entries in this row */ \
191: PetscCall(PetscArraymove(rp + _i + 1, rp + _i, N - _i + 1)); \
192: PetscCall(PetscArraymove(ap + bs2 * (_i + 1), ap + bs2 * _i, bs2 * (N - _i + 1))); \
193: PetscCall(PetscArrayzero(ap + bs2 * _i, bs2)); \
194: rp[_i] = bcol; \
195: ap[bs2 * _i + bs * cidx + ridx] = value; \
196: a_noinsert:; \
197: ailen[brow] = nrow; \
198: } while (0)
200: #define MatSetValues_SeqBAIJ_B_Private(row, col, value, addv, orow, ocol) \
201: do { \
202: brow = row / bs; \
203: rp = PetscSafePointerPlusOffset(bj, bi[brow]); \
204: ap = PetscSafePointerPlusOffset(ba, bs2 * bi[brow]); \
205: rmax = bimax[brow]; \
206: nrow = bilen[brow]; \
207: bcol = col / bs; \
208: ridx = row % bs; \
209: cidx = col % bs; \
210: low = 0; \
211: high = nrow; \
212: while (high - low > 3) { \
213: t = (low + high) / 2; \
214: if (rp[t] > bcol) high = t; \
215: else low = t; \
216: } \
217: for (_i = low; _i < high; _i++) { \
218: if (rp[_i] > bcol) break; \
219: if (rp[_i] == bcol) { \
220: bap = ap + bs2 * _i + bs * cidx + ridx; \
221: if (addv == ADD_VALUES) *bap += value; \
222: else *bap = value; \
223: goto b_noinsert; \
224: } \
225: } \
226: if (b->nonew == 1) goto b_noinsert; \
227: PetscCheck(b->nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", orow, ocol); \
228: MatSeqXAIJReallocateAIJ(B, b->mbs, bs2, nrow, brow, bcol, rmax, ba, bi, bj, rp, ap, bimax, b->nonew, MatScalar); \
229: N = nrow++ - 1; \
230: /* shift up all the later entries in this row */ \
231: PetscCall(PetscArraymove(rp + _i + 1, rp + _i, N - _i + 1)); \
232: PetscCall(PetscArraymove(ap + bs2 * (_i + 1), ap + bs2 * _i, bs2 * (N - _i + 1))); \
233: PetscCall(PetscArrayzero(ap + bs2 * _i, bs2)); \
234: rp[_i] = bcol; \
235: ap[bs2 * _i + bs * cidx + ridx] = value; \
236: b_noinsert:; \
237: bilen[brow] = nrow; \
238: } while (0)
240: PetscErrorCode MatSetValues_MPIBAIJ(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
241: {
242: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
243: MatScalar value;
244: PetscBool roworiented = baij->roworiented;
245: PetscInt i, j, row, col;
246: PetscInt rstart_orig = mat->rmap->rstart;
247: PetscInt rend_orig = mat->rmap->rend, cstart_orig = mat->cmap->rstart;
248: PetscInt cend_orig = mat->cmap->rend, bs = mat->rmap->bs;
250: /* Some Variables required in the macro */
251: Mat A = baij->A;
252: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
253: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
254: MatScalar *aa = a->a;
256: Mat B = baij->B;
257: Mat_SeqBAIJ *b = (Mat_SeqBAIJ *)B->data;
258: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j;
259: MatScalar *ba = b->a;
261: PetscInt *rp, ii, nrow, _i, rmax, N, brow, bcol;
262: PetscInt low, high, t, ridx, cidx, bs2 = a->bs2;
263: MatScalar *ap, *bap;
265: PetscFunctionBegin;
266: for (i = 0; i < m; i++) {
267: if (im[i] < 0) continue;
268: PetscCheck(im[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, im[i], mat->rmap->N - 1);
269: if (im[i] >= rstart_orig && im[i] < rend_orig) {
270: row = im[i] - rstart_orig;
271: for (j = 0; j < n; j++) {
272: if (in[j] >= cstart_orig && in[j] < cend_orig) {
273: col = in[j] - cstart_orig;
274: if (roworiented) value = v[i * n + j];
275: else value = v[i + j * m];
276: MatSetValues_SeqBAIJ_A_Private(row, col, value, addv, im[i], in[j]);
277: } else if (in[j] < 0) {
278: continue;
279: } else {
280: PetscCheck(in[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
281: if (mat->was_assembled) {
282: if (!baij->colmap) PetscCall(MatCreateColmap_MPIBAIJ_Private(mat));
283: #if defined(PETSC_USE_CTABLE)
284: PetscCall(PetscHMapIGetWithDefault(baij->colmap, in[j] / bs + 1, 0, &col));
285: col = col - 1;
286: #else
287: col = baij->colmap[in[j] / bs] - 1;
288: #endif
289: if (col < 0 && !((Mat_SeqBAIJ *)baij->B->data)->nonew) {
290: PetscCall(MatDisAssemble_MPIBAIJ(mat));
291: col = in[j];
292: /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */
293: B = baij->B;
294: b = (Mat_SeqBAIJ *)B->data;
295: bimax = b->imax;
296: bi = b->i;
297: bilen = b->ilen;
298: bj = b->j;
299: ba = b->a;
300: } else {
301: PetscCheck(col >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", im[i], in[j]);
302: col += in[j] % bs;
303: }
304: } else col = in[j];
305: if (roworiented) value = v[i * n + j];
306: else value = v[i + j * m];
307: MatSetValues_SeqBAIJ_B_Private(row, col, value, addv, im[i], in[j]);
308: /* PetscCall(MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv)); */
309: }
310: }
311: } else {
312: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Setting off process row %" PetscInt_FMT " even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set", im[i]);
313: if (!baij->donotstash) {
314: mat->assembled = PETSC_FALSE;
315: if (roworiented) {
316: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, PETSC_FALSE));
317: } else {
318: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, PETSC_FALSE));
319: }
320: }
321: }
322: }
323: PetscFunctionReturn(PETSC_SUCCESS);
324: }
326: static inline PetscErrorCode MatSetValuesBlocked_SeqBAIJ_Inlined(Mat A, PetscInt row, PetscInt col, const PetscScalar v[], InsertMode is, PetscInt orow, PetscInt ocol)
327: {
328: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
329: PetscInt *rp, low, high, t, ii, jj, nrow, i, rmax, N;
330: PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen;
331: PetscInt *aj = a->j, nonew = a->nonew, bs2 = a->bs2, bs = A->rmap->bs;
332: PetscBool roworiented = a->roworiented;
333: const PetscScalar *value = v;
334: MatScalar *ap, *aa = a->a, *bap;
336: PetscFunctionBegin;
337: rp = aj + ai[row];
338: ap = aa + bs2 * ai[row];
339: rmax = imax[row];
340: nrow = ailen[row];
341: value = v;
342: low = 0;
343: high = nrow;
344: while (high - low > 7) {
345: t = (low + high) / 2;
346: if (rp[t] > col) high = t;
347: else low = t;
348: }
349: for (i = low; i < high; i++) {
350: if (rp[i] > col) break;
351: if (rp[i] == col) {
352: bap = ap + bs2 * i;
353: if (roworiented) {
354: if (is == ADD_VALUES) {
355: for (ii = 0; ii < bs; ii++) {
356: for (jj = ii; jj < bs2; jj += bs) bap[jj] += *value++;
357: }
358: } else {
359: for (ii = 0; ii < bs; ii++) {
360: for (jj = ii; jj < bs2; jj += bs) bap[jj] = *value++;
361: }
362: }
363: } else {
364: if (is == ADD_VALUES) {
365: for (ii = 0; ii < bs; ii++, value += bs) {
366: for (jj = 0; jj < bs; jj++) bap[jj] += value[jj];
367: bap += bs;
368: }
369: } else {
370: for (ii = 0; ii < bs; ii++, value += bs) {
371: for (jj = 0; jj < bs; jj++) bap[jj] = value[jj];
372: bap += bs;
373: }
374: }
375: }
376: goto noinsert2;
377: }
378: }
379: if (nonew == 1) goto noinsert2;
380: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new global block indexed nonzero block (%" PetscInt_FMT ", %" PetscInt_FMT ") in the matrix", orow, ocol);
381: MatSeqXAIJReallocateAIJ(A, a->mbs, bs2, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
382: N = nrow++ - 1;
383: high++;
384: /* shift up all the later entries in this row */
385: PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1));
386: PetscCall(PetscArraymove(ap + bs2 * (i + 1), ap + bs2 * i, bs2 * (N - i + 1)));
387: rp[i] = col;
388: bap = ap + bs2 * i;
389: if (roworiented) {
390: for (ii = 0; ii < bs; ii++) {
391: for (jj = ii; jj < bs2; jj += bs) bap[jj] = *value++;
392: }
393: } else {
394: for (ii = 0; ii < bs; ii++) {
395: for (jj = 0; jj < bs; jj++) *bap++ = *value++;
396: }
397: }
398: noinsert2:;
399: ailen[row] = nrow;
400: PetscFunctionReturn(PETSC_SUCCESS);
401: }
403: /*
404: This routine should be optimized so that the block copy at ** Here a copy is required ** below is not needed
405: by passing additional stride information into the MatSetValuesBlocked_SeqBAIJ_Inlined() routine
406: */
407: static PetscErrorCode MatSetValuesBlocked_MPIBAIJ(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
408: {
409: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
410: const PetscScalar *value;
411: MatScalar *barray = baij->barray;
412: PetscBool roworiented = baij->roworiented;
413: PetscInt i, j, ii, jj, row, col, rstart = baij->rstartbs;
414: PetscInt rend = baij->rendbs, cstart = baij->cstartbs, stepval;
415: PetscInt cend = baij->cendbs, bs = mat->rmap->bs, bs2 = baij->bs2;
417: PetscFunctionBegin;
418: if (!barray) {
419: PetscCall(PetscMalloc1(bs2, &barray));
420: baij->barray = barray;
421: }
423: if (roworiented) stepval = (n - 1) * bs;
424: else stepval = (m - 1) * bs;
426: for (i = 0; i < m; i++) {
427: if (im[i] < 0) continue;
428: PetscCheck(im[i] < baij->Mbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Block indexed row too large %" PetscInt_FMT " max %" PetscInt_FMT, im[i], baij->Mbs - 1);
429: if (im[i] >= rstart && im[i] < rend) {
430: row = im[i] - rstart;
431: for (j = 0; j < n; j++) {
432: /* If NumCol = 1 then a copy is not required */
433: if ((roworiented) && (n == 1)) {
434: barray = (MatScalar *)v + i * bs2;
435: } else if ((!roworiented) && (m == 1)) {
436: barray = (MatScalar *)v + j * bs2;
437: } else { /* Here a copy is required */
438: if (roworiented) {
439: value = v + (i * (stepval + bs) + j) * bs;
440: } else {
441: value = v + (j * (stepval + bs) + i) * bs;
442: }
443: for (ii = 0; ii < bs; ii++, value += bs + stepval) {
444: for (jj = 0; jj < bs; jj++) barray[jj] = value[jj];
445: barray += bs;
446: }
447: barray -= bs2;
448: }
450: if (in[j] >= cstart && in[j] < cend) {
451: col = in[j] - cstart;
452: PetscCall(MatSetValuesBlocked_SeqBAIJ_Inlined(baij->A, row, col, barray, addv, im[i], in[j]));
453: } else if (in[j] < 0) {
454: continue;
455: } else {
456: PetscCheck(in[j] < baij->Nbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Block indexed column too large %" PetscInt_FMT " max %" PetscInt_FMT, in[j], baij->Nbs - 1);
457: if (mat->was_assembled) {
458: if (!baij->colmap) PetscCall(MatCreateColmap_MPIBAIJ_Private(mat));
460: #if defined(PETSC_USE_CTABLE)
461: PetscCall(PetscHMapIGetWithDefault(baij->colmap, in[j] + 1, 0, &col));
462: col = col < 1 ? -1 : (col - 1) / bs;
463: #else
464: col = baij->colmap[in[j]] < 1 ? -1 : (baij->colmap[in[j]] - 1) / bs;
465: #endif
466: if (col < 0 && !((Mat_SeqBAIJ *)baij->B->data)->nonew) {
467: PetscCall(MatDisAssemble_MPIBAIJ(mat));
468: col = in[j];
469: } else PetscCheck(col >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new blocked indexed nonzero block (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", im[i], in[j]);
470: } else col = in[j];
471: PetscCall(MatSetValuesBlocked_SeqBAIJ_Inlined(baij->B, row, col, barray, addv, im[i], in[j]));
472: }
473: }
474: } else {
475: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Setting off process block indexed row %" PetscInt_FMT " even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set", im[i]);
476: if (!baij->donotstash) {
477: if (roworiented) {
478: PetscCall(MatStashValuesRowBlocked_Private(&mat->bstash, im[i], n, in, v, m, n, i));
479: } else {
480: PetscCall(MatStashValuesColBlocked_Private(&mat->bstash, im[i], n, in, v, m, n, i));
481: }
482: }
483: }
484: }
485: PetscFunctionReturn(PETSC_SUCCESS);
486: }
488: #define HASH_KEY 0.6180339887
489: #define HASH(size, key, tmp) (tmp = (key) * HASH_KEY, (PetscInt)((size) * (tmp - (PetscInt)tmp)))
490: /* #define HASH(size,key) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
491: /* #define HASH(size,key,tmp) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
492: static PetscErrorCode MatSetValues_MPIBAIJ_HT(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
493: {
494: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
495: PetscBool roworiented = baij->roworiented;
496: PetscInt i, j, row, col;
497: PetscInt rstart_orig = mat->rmap->rstart;
498: PetscInt rend_orig = mat->rmap->rend, Nbs = baij->Nbs;
499: PetscInt h1, key, size = baij->ht_size, bs = mat->rmap->bs, *HT = baij->ht, idx;
500: PetscReal tmp;
501: MatScalar **HD = baij->hd, value;
502: PetscInt total_ct = baij->ht_total_ct, insert_ct = baij->ht_insert_ct;
504: PetscFunctionBegin;
505: for (i = 0; i < m; i++) {
506: if (PetscDefined(USE_DEBUG)) {
507: PetscCheck(im[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative row");
508: PetscCheck(im[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, im[i], mat->rmap->N - 1);
509: }
510: row = im[i];
511: if (row >= rstart_orig && row < rend_orig) {
512: for (j = 0; j < n; j++) {
513: col = in[j];
514: if (roworiented) value = v[i * n + j];
515: else value = v[i + j * m];
516: /* Look up PetscInto the Hash Table */
517: key = (row / bs) * Nbs + (col / bs) + 1;
518: h1 = HASH(size, key, tmp);
520: idx = h1;
521: if (PetscDefined(USE_DEBUG)) {
522: insert_ct++;
523: total_ct++;
524: if (HT[idx] != key) {
525: for (idx = h1; (idx < size) && (HT[idx] != key); idx++, total_ct++);
526: if (idx == size) {
527: for (idx = 0; (idx < h1) && (HT[idx] != key); idx++, total_ct++);
528: PetscCheck(idx != h1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "(%" PetscInt_FMT ",%" PetscInt_FMT ") has no entry in the hash table", row, col);
529: }
530: }
531: } else if (HT[idx] != key) {
532: for (idx = h1; (idx < size) && (HT[idx] != key); idx++);
533: if (idx == size) {
534: for (idx = 0; (idx < h1) && (HT[idx] != key); idx++);
535: PetscCheck(idx != h1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "(%" PetscInt_FMT ",%" PetscInt_FMT ") has no entry in the hash table", row, col);
536: }
537: }
538: /* A HASH table entry is found, so insert the values at the correct address */
539: if (addv == ADD_VALUES) *(HD[idx] + (col % bs) * bs + (row % bs)) += value;
540: else *(HD[idx] + (col % bs) * bs + (row % bs)) = value;
541: }
542: } else if (!baij->donotstash) {
543: if (roworiented) {
544: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, PETSC_FALSE));
545: } else {
546: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, PETSC_FALSE));
547: }
548: }
549: }
550: if (PetscDefined(USE_DEBUG)) {
551: baij->ht_total_ct += total_ct;
552: baij->ht_insert_ct += insert_ct;
553: }
554: PetscFunctionReturn(PETSC_SUCCESS);
555: }
557: static PetscErrorCode MatSetValuesBlocked_MPIBAIJ_HT(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
558: {
559: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
560: PetscBool roworiented = baij->roworiented;
561: PetscInt i, j, ii, jj, row, col;
562: PetscInt rstart = baij->rstartbs;
563: PetscInt rend = mat->rmap->rend, stepval, bs = mat->rmap->bs, bs2 = baij->bs2, nbs2 = n * bs2;
564: PetscInt h1, key, size = baij->ht_size, idx, *HT = baij->ht, Nbs = baij->Nbs;
565: PetscReal tmp;
566: MatScalar **HD = baij->hd, *baij_a;
567: const PetscScalar *v_t, *value;
568: PetscInt total_ct = baij->ht_total_ct, insert_ct = baij->ht_insert_ct;
570: PetscFunctionBegin;
571: if (roworiented) stepval = (n - 1) * bs;
572: else stepval = (m - 1) * bs;
574: for (i = 0; i < m; i++) {
575: if (PetscDefined(USE_DEBUG)) {
576: PetscCheck(im[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative row: %" PetscInt_FMT, im[i]);
577: PetscCheck(im[i] < baij->Mbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, im[i], baij->Mbs - 1);
578: }
579: row = im[i];
580: v_t = v + i * nbs2;
581: if (row >= rstart && row < rend) {
582: for (j = 0; j < n; j++) {
583: col = in[j];
585: /* Look up into the Hash Table */
586: key = row * Nbs + col + 1;
587: h1 = HASH(size, key, tmp);
589: idx = h1;
590: if (PetscDefined(USE_DEBUG)) {
591: total_ct++;
592: insert_ct++;
593: if (HT[idx] != key) {
594: for (idx = h1; (idx < size) && (HT[idx] != key); idx++, total_ct++);
595: if (idx == size) {
596: for (idx = 0; (idx < h1) && (HT[idx] != key); idx++, total_ct++);
597: PetscCheck(idx != h1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "(%" PetscInt_FMT ",%" PetscInt_FMT ") has no entry in the hash table", row, col);
598: }
599: }
600: } else if (HT[idx] != key) {
601: for (idx = h1; (idx < size) && (HT[idx] != key); idx++);
602: if (idx == size) {
603: for (idx = 0; (idx < h1) && (HT[idx] != key); idx++);
604: PetscCheck(idx != h1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "(%" PetscInt_FMT ",%" PetscInt_FMT ") has no entry in the hash table", row, col);
605: }
606: }
607: baij_a = HD[idx];
608: if (roworiented) {
609: /*value = v + i*(stepval+bs)*bs + j*bs;*/
610: /* value = v + (i*(stepval+bs)+j)*bs; */
611: value = v_t;
612: v_t += bs;
613: if (addv == ADD_VALUES) {
614: for (ii = 0; ii < bs; ii++, value += stepval) {
615: for (jj = ii; jj < bs2; jj += bs) baij_a[jj] += *value++;
616: }
617: } else {
618: for (ii = 0; ii < bs; ii++, value += stepval) {
619: for (jj = ii; jj < bs2; jj += bs) baij_a[jj] = *value++;
620: }
621: }
622: } else {
623: value = v + j * (stepval + bs) * bs + i * bs;
624: if (addv == ADD_VALUES) {
625: for (ii = 0; ii < bs; ii++, value += stepval, baij_a += bs) {
626: for (jj = 0; jj < bs; jj++) baij_a[jj] += *value++;
627: }
628: } else {
629: for (ii = 0; ii < bs; ii++, value += stepval, baij_a += bs) {
630: for (jj = 0; jj < bs; jj++) baij_a[jj] = *value++;
631: }
632: }
633: }
634: }
635: } else {
636: if (!baij->donotstash) {
637: if (roworiented) {
638: PetscCall(MatStashValuesRowBlocked_Private(&mat->bstash, im[i], n, in, v, m, n, i));
639: } else {
640: PetscCall(MatStashValuesColBlocked_Private(&mat->bstash, im[i], n, in, v, m, n, i));
641: }
642: }
643: }
644: }
645: if (PetscDefined(USE_DEBUG)) {
646: baij->ht_total_ct += total_ct;
647: baij->ht_insert_ct += insert_ct;
648: }
649: PetscFunctionReturn(PETSC_SUCCESS);
650: }
652: static PetscErrorCode MatGetValues_MPIBAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
653: {
654: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
655: PetscInt bs = mat->rmap->bs, i, j, bsrstart = mat->rmap->rstart, bsrend = mat->rmap->rend;
656: PetscInt bscstart = mat->cmap->rstart, bscend = mat->cmap->rend, row, col, data;
658: PetscFunctionBegin;
659: for (i = 0; i < m; i++) {
660: if (idxm[i] < 0) continue; /* negative row */
661: PetscCheck(idxm[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, idxm[i], mat->rmap->N - 1);
662: PetscCheck(idxm[i] >= bsrstart && idxm[i] < bsrend, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only local values currently supported");
663: row = idxm[i] - bsrstart;
664: for (j = 0; j < n; j++) {
665: if (idxn[j] < 0) continue; /* negative column */
666: PetscCheck(idxn[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, idxn[j], mat->cmap->N - 1);
667: if (idxn[j] >= bscstart && idxn[j] < bscend) {
668: col = idxn[j] - bscstart;
669: PetscCall(MatGetValues_SeqBAIJ(baij->A, 1, &row, 1, &col, v + i * n + j));
670: } else {
671: if (!baij->colmap) PetscCall(MatCreateColmap_MPIBAIJ_Private(mat));
672: #if defined(PETSC_USE_CTABLE)
673: PetscCall(PetscHMapIGetWithDefault(baij->colmap, idxn[j] / bs + 1, 0, &data));
674: data--;
675: #else
676: data = baij->colmap[idxn[j] / bs] - 1;
677: #endif
678: if (data < 0 || baij->garray[data / bs] != idxn[j] / bs) *(v + i * n + j) = 0.0;
679: else {
680: col = data + idxn[j] % bs;
681: PetscCall(MatGetValues_SeqBAIJ(baij->B, 1, &row, 1, &col, v + i * n + j));
682: }
683: }
684: }
685: }
686: PetscFunctionReturn(PETSC_SUCCESS);
687: }
689: static PetscErrorCode MatNorm_MPIBAIJ(Mat mat, NormType type, PetscReal *nrm)
690: {
691: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
692: Mat_SeqBAIJ *amat = (Mat_SeqBAIJ *)baij->A->data, *bmat = (Mat_SeqBAIJ *)baij->B->data;
693: PetscInt i, j, bs2 = baij->bs2, bs = baij->A->rmap->bs, nz, row, col;
694: PetscReal sum = 0.0;
695: MatScalar *v;
697: PetscFunctionBegin;
698: if (baij->size == 1) {
699: PetscCall(MatNorm(baij->A, type, nrm));
700: } else {
701: if (type == NORM_FROBENIUS) {
702: v = amat->a;
703: nz = amat->nz * bs2;
704: for (i = 0; i < nz; i++) {
705: sum += PetscRealPart(PetscConj(*v) * (*v));
706: v++;
707: }
708: v = bmat->a;
709: nz = bmat->nz * bs2;
710: for (i = 0; i < nz; i++) {
711: sum += PetscRealPart(PetscConj(*v) * (*v));
712: v++;
713: }
714: PetscCallMPI(MPIU_Allreduce(&sum, nrm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
715: *nrm = PetscSqrtReal(*nrm);
716: } else if (type == NORM_1) { /* max column sum */
717: Vec col, bcol;
718: PetscScalar *array;
719: PetscInt *jj, *garray = baij->garray;
721: PetscCall(MatCreateVecs(mat, &col, NULL));
722: PetscCall(VecSet(col, 0.0));
723: PetscCall(VecGetArrayWrite(col, &array));
724: v = amat->a;
725: jj = amat->j;
726: for (i = 0; i < amat->nz; i++) {
727: for (j = 0; j < bs; j++) {
728: PetscInt col = bs * *jj + j; /* column index */
730: for (row = 0; row < bs; row++) array[col] += PetscAbsScalar(*v++);
731: }
732: jj++;
733: }
734: PetscCall(VecRestoreArrayWrite(col, &array));
735: PetscCall(MatCreateVecs(baij->B, &bcol, NULL));
736: PetscCall(VecSet(bcol, 0.0));
737: PetscCall(VecGetArrayWrite(bcol, &array));
738: v = bmat->a;
739: jj = bmat->j;
740: for (i = 0; i < bmat->nz; i++) {
741: for (j = 0; j < bs; j++) {
742: PetscInt col = bs * *jj + j; /* column index */
744: for (row = 0; row < bs; row++) array[col] += PetscAbsScalar(*v++);
745: }
746: jj++;
747: }
748: PetscCall(VecSetValuesBlocked(col, bmat->nbs, garray, array, ADD_VALUES));
749: PetscCall(VecRestoreArrayWrite(bcol, &array));
750: PetscCall(VecDestroy(&bcol));
751: PetscCall(VecAssemblyBegin(col));
752: PetscCall(VecAssemblyEnd(col));
753: PetscCall(VecNorm(col, NORM_INFINITY, nrm));
754: PetscCall(VecDestroy(&col));
755: } else if (type == NORM_INFINITY) { /* max row sum */
756: PetscReal *sums;
757: PetscCall(PetscMalloc1(bs, &sums));
758: sum = 0.0;
759: for (j = 0; j < amat->mbs; j++) {
760: for (row = 0; row < bs; row++) sums[row] = 0.0;
761: v = amat->a + bs2 * amat->i[j];
762: nz = amat->i[j + 1] - amat->i[j];
763: for (i = 0; i < nz; i++) {
764: for (col = 0; col < bs; col++) {
765: for (row = 0; row < bs; row++) {
766: sums[row] += PetscAbsScalar(*v);
767: v++;
768: }
769: }
770: }
771: v = bmat->a + bs2 * bmat->i[j];
772: nz = bmat->i[j + 1] - bmat->i[j];
773: for (i = 0; i < nz; i++) {
774: for (col = 0; col < bs; col++) {
775: for (row = 0; row < bs; row++) {
776: sums[row] += PetscAbsScalar(*v);
777: v++;
778: }
779: }
780: }
781: for (row = 0; row < bs; row++) {
782: if (sums[row] > sum) sum = sums[row];
783: }
784: }
785: PetscCallMPI(MPIU_Allreduce(&sum, nrm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
786: PetscCall(PetscFree(sums));
787: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for this norm yet");
788: }
789: PetscFunctionReturn(PETSC_SUCCESS);
790: }
792: /*
793: Creates the hash table, and sets the table
794: This table is created only once.
795: If new entries need to be added to the matrix
796: then the hash table has to be destroyed and
797: recreated.
798: */
799: static PetscErrorCode MatCreateHashTable_MPIBAIJ_Private(Mat mat, PetscReal factor)
800: {
801: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
802: Mat A = baij->A, B = baij->B;
803: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data, *b = (Mat_SeqBAIJ *)B->data;
804: PetscInt i, j, k, nz = a->nz + b->nz, h1, *ai = a->i, *aj = a->j, *bi = b->i, *bj = b->j;
805: PetscInt ht_size, bs2 = baij->bs2, rstart = baij->rstartbs;
806: PetscInt cstart = baij->cstartbs, *garray = baij->garray, row, col, Nbs = baij->Nbs;
807: PetscInt *HT, key;
808: MatScalar **HD;
809: PetscReal tmp;
810: #if defined(PETSC_USE_INFO)
811: PetscInt ct = 0, max = 0;
812: #endif
814: PetscFunctionBegin;
815: if (baij->ht) PetscFunctionReturn(PETSC_SUCCESS);
817: baij->ht_size = (PetscInt)(factor * nz);
818: ht_size = baij->ht_size;
820: /* Allocate Memory for Hash Table */
821: PetscCall(PetscCalloc2(ht_size, &baij->hd, ht_size, &baij->ht));
822: HD = baij->hd;
823: HT = baij->ht;
825: /* Loop Over A */
826: for (i = 0; i < a->mbs; i++) {
827: for (j = ai[i]; j < ai[i + 1]; j++) {
828: row = i + rstart;
829: col = aj[j] + cstart;
831: key = row * Nbs + col + 1;
832: h1 = HASH(ht_size, key, tmp);
833: for (k = 0; k < ht_size; k++) {
834: if (!HT[(h1 + k) % ht_size]) {
835: HT[(h1 + k) % ht_size] = key;
836: HD[(h1 + k) % ht_size] = a->a + j * bs2;
837: break;
838: #if defined(PETSC_USE_INFO)
839: } else {
840: ct++;
841: #endif
842: }
843: }
844: #if defined(PETSC_USE_INFO)
845: if (k > max) max = k;
846: #endif
847: }
848: }
849: /* Loop Over B */
850: for (i = 0; i < b->mbs; i++) {
851: for (j = bi[i]; j < bi[i + 1]; j++) {
852: row = i + rstart;
853: col = garray[bj[j]];
854: key = row * Nbs + col + 1;
855: h1 = HASH(ht_size, key, tmp);
856: for (k = 0; k < ht_size; k++) {
857: if (!HT[(h1 + k) % ht_size]) {
858: HT[(h1 + k) % ht_size] = key;
859: HD[(h1 + k) % ht_size] = b->a + j * bs2;
860: break;
861: #if defined(PETSC_USE_INFO)
862: } else {
863: ct++;
864: #endif
865: }
866: }
867: #if defined(PETSC_USE_INFO)
868: if (k > max) max = k;
869: #endif
870: }
871: }
873: /* Print Summary */
874: #if defined(PETSC_USE_INFO)
875: for (i = 0, j = 0; i < ht_size; i++) {
876: if (HT[i]) j++;
877: }
878: PetscCall(PetscInfo(mat, "Average Search = %5.2g,max search = %" PetscInt_FMT "\n", (!j) ? 0.0 : (double)(((PetscReal)(ct + j)) / j), max));
879: #endif
880: PetscFunctionReturn(PETSC_SUCCESS);
881: }
883: static PetscErrorCode MatAssemblyBegin_MPIBAIJ(Mat mat, MatAssemblyType mode)
884: {
885: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
886: PetscInt nstash, reallocs;
888: PetscFunctionBegin;
889: if (baij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
891: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
892: PetscCall(MatStashScatterBegin_Private(mat, &mat->bstash, baij->rangebs));
893: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
894: PetscCall(PetscInfo(mat, "Stash has %" PetscInt_FMT " entries,uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
895: PetscCall(MatStashGetInfo_Private(&mat->bstash, &nstash, &reallocs));
896: PetscCall(PetscInfo(mat, "Block-Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
897: PetscFunctionReturn(PETSC_SUCCESS);
898: }
900: static PetscErrorCode MatAssemblyEnd_MPIBAIJ(Mat mat, MatAssemblyType mode)
901: {
902: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
903: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)baij->A->data;
904: PetscInt i, j, rstart, ncols, flg, bs2 = baij->bs2;
905: PetscInt *row, *col;
906: PetscBool r1, r2, r3, all_assembled;
907: MatScalar *val;
908: PetscMPIInt n;
910: PetscFunctionBegin;
911: /* do not use 'b=(Mat_SeqBAIJ*)baij->B->data' as B can be reset in disassembly */
912: if (!baij->donotstash && !mat->nooffprocentries) {
913: while (1) {
914: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
915: if (!flg) break;
917: for (i = 0; i < n;) {
918: /* Now identify the consecutive vals belonging to the same row */
919: for (j = i, rstart = row[j]; j < n; j++) {
920: if (row[j] != rstart) break;
921: }
922: if (j < n) ncols = j - i;
923: else ncols = n - i;
924: /* Now assemble all these values with a single function call */
925: PetscCall(MatSetValues_MPIBAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
926: i = j;
927: }
928: }
929: PetscCall(MatStashScatterEnd_Private(&mat->stash));
930: /* Now process the block-stash. Since the values are stashed column-oriented,
931: set the row-oriented flag to column-oriented, and after MatSetValues()
932: restore the original flags */
933: r1 = baij->roworiented;
934: r2 = a->roworiented;
935: r3 = ((Mat_SeqBAIJ *)baij->B->data)->roworiented;
937: baij->roworiented = PETSC_FALSE;
938: a->roworiented = PETSC_FALSE;
939: ((Mat_SeqBAIJ *)baij->B->data)->roworiented = PETSC_FALSE;
940: while (1) {
941: PetscCall(MatStashScatterGetMesg_Private(&mat->bstash, &n, &row, &col, &val, &flg));
942: if (!flg) break;
944: for (i = 0; i < n;) {
945: /* Now identify the consecutive vals belonging to the same row */
946: for (j = i, rstart = row[j]; j < n; j++) {
947: if (row[j] != rstart) break;
948: }
949: if (j < n) ncols = j - i;
950: else ncols = n - i;
951: PetscCall(MatSetValuesBlocked_MPIBAIJ(mat, 1, row + i, ncols, col + i, val + i * bs2, mat->insertmode));
952: i = j;
953: }
954: }
955: PetscCall(MatStashScatterEnd_Private(&mat->bstash));
957: baij->roworiented = r1;
958: a->roworiented = r2;
959: ((Mat_SeqBAIJ *)baij->B->data)->roworiented = r3;
960: }
962: PetscCall(MatAssemblyBegin(baij->A, mode));
963: PetscCall(MatAssemblyEnd(baij->A, mode));
965: /* determine if any process has disassembled, if so we must
966: also disassemble ourselves, in order that we may reassemble. */
967: /*
968: if nonzero structure of submatrix B cannot change then we know that
969: no process disassembled thus we can skip this stuff
970: */
971: if (!((Mat_SeqBAIJ *)baij->B->data)->nonew) {
972: PetscCallMPI(MPIU_Allreduce(&mat->was_assembled, &all_assembled, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
973: if (mat->was_assembled && !all_assembled) PetscCall(MatDisAssemble_MPIBAIJ(mat));
974: }
976: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIBAIJ(mat));
977: PetscCall(MatAssemblyBegin(baij->B, mode));
978: PetscCall(MatAssemblyEnd(baij->B, mode));
980: #if defined(PETSC_USE_INFO)
981: if (baij->ht && mode == MAT_FINAL_ASSEMBLY) {
982: PetscCall(PetscInfo(mat, "Average Hash Table Search in MatSetValues = %5.2f\n", (double)((PetscReal)baij->ht_total_ct) / baij->ht_insert_ct));
984: baij->ht_total_ct = 0;
985: baij->ht_insert_ct = 0;
986: }
987: #endif
988: if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) {
989: PetscCall(MatCreateHashTable_MPIBAIJ_Private(mat, baij->ht_fact));
991: mat->ops->setvalues = MatSetValues_MPIBAIJ_HT;
992: mat->ops->setvaluesblocked = MatSetValuesBlocked_MPIBAIJ_HT;
993: }
995: PetscCall(PetscFree2(baij->rowvalues, baij->rowindices));
997: baij->rowvalues = NULL;
999: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
1000: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqBAIJ *)baij->A->data)->nonew) {
1001: PetscObjectState state = baij->A->nonzerostate + baij->B->nonzerostate;
1002: PetscCallMPI(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
1003: }
1004: PetscFunctionReturn(PETSC_SUCCESS);
1005: }
1007: #include <petscdraw.h>
1008: static PetscErrorCode MatView_MPIBAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1009: {
1010: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
1011: PetscMPIInt rank = baij->rank;
1012: PetscInt bs = mat->rmap->bs;
1013: PetscBool isascii, isdraw;
1014: PetscViewer sviewer;
1015: PetscViewerFormat format;
1017: PetscFunctionBegin;
1018: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1019: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1020: if (isascii) {
1021: PetscCall(PetscViewerGetFormat(viewer, &format));
1022: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1023: MatInfo info;
1024: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1025: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1026: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1027: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Local rows %" PetscInt_FMT " nz %" PetscInt_FMT " nz alloced %" PetscInt_FMT " bs %" PetscInt_FMT " mem %g\n", rank, mat->rmap->n, (PetscInt)info.nz_used, (PetscInt)info.nz_allocated,
1028: mat->rmap->bs, info.memory));
1029: PetscCall(MatGetInfo(baij->A, MAT_LOCAL, &info));
1030: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1031: PetscCall(MatGetInfo(baij->B, MAT_LOCAL, &info));
1032: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1033: PetscCall(PetscViewerFlush(viewer));
1034: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1035: PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1036: PetscCall(VecScatterView(baij->Mvctx, viewer));
1037: PetscFunctionReturn(PETSC_SUCCESS);
1038: } else if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_FACTOR_INFO) PetscFunctionReturn(PETSC_SUCCESS);
1039: }
1041: if (isdraw) {
1042: PetscDraw draw;
1043: PetscBool isnull;
1044: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1045: PetscCall(PetscDrawIsNull(draw, &isnull));
1046: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1047: }
1049: {
1050: /* assemble the entire matrix onto first processor. */
1051: Mat A;
1052: Mat_SeqBAIJ *Aloc;
1053: PetscInt M = mat->rmap->N, N = mat->cmap->N, *ai, *aj, col, i, j, k, *rvals, mbs = baij->mbs;
1054: MatScalar *a;
1055: const char *matname;
1057: /* Here we are creating a temporary matrix, so will assume MPIBAIJ is acceptable */
1058: /* Perhaps this should be the type of mat? */
1059: PetscCall(MatCreate(PetscObjectComm((PetscObject)mat), &A));
1060: if (rank == 0) {
1061: PetscCall(MatSetSizes(A, M, N, M, N));
1062: } else {
1063: PetscCall(MatSetSizes(A, 0, 0, M, N));
1064: }
1065: PetscCall(MatSetType(A, MATMPIBAIJ));
1066: PetscCall(MatMPIBAIJSetPreallocation(A, mat->rmap->bs, 0, NULL, 0, NULL));
1067: PetscCall(MatSetOption(A, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_FALSE));
1069: /* copy over the A part */
1070: Aloc = (Mat_SeqBAIJ *)baij->A->data;
1071: ai = Aloc->i;
1072: aj = Aloc->j;
1073: a = Aloc->a;
1074: PetscCall(PetscMalloc1(bs, &rvals));
1076: for (i = 0; i < mbs; i++) {
1077: rvals[0] = bs * (baij->rstartbs + i);
1078: for (j = 1; j < bs; j++) rvals[j] = rvals[j - 1] + 1;
1079: for (j = ai[i]; j < ai[i + 1]; j++) {
1080: col = (baij->cstartbs + aj[j]) * bs;
1081: for (k = 0; k < bs; k++) {
1082: PetscCall(MatSetValues_MPIBAIJ(A, bs, rvals, 1, &col, a, INSERT_VALUES));
1083: col++;
1084: a += bs;
1085: }
1086: }
1087: }
1088: /* copy over the B part */
1089: Aloc = (Mat_SeqBAIJ *)baij->B->data;
1090: ai = Aloc->i;
1091: aj = Aloc->j;
1092: a = Aloc->a;
1093: for (i = 0; i < mbs; i++) {
1094: rvals[0] = bs * (baij->rstartbs + i);
1095: for (j = 1; j < bs; j++) rvals[j] = rvals[j - 1] + 1;
1096: for (j = ai[i]; j < ai[i + 1]; j++) {
1097: col = baij->garray[aj[j]] * bs;
1098: for (k = 0; k < bs; k++) {
1099: PetscCall(MatSetValues_MPIBAIJ(A, bs, rvals, 1, &col, a, INSERT_VALUES));
1100: col++;
1101: a += bs;
1102: }
1103: }
1104: }
1105: PetscCall(PetscFree(rvals));
1106: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
1107: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
1108: /*
1109: Everyone has to call to draw the matrix since the graphics waits are
1110: synchronized across all processors that share the PetscDraw object
1111: */
1112: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1113: if (((PetscObject)mat)->name) PetscCall(PetscObjectGetName((PetscObject)mat, &matname));
1114: if (rank == 0) {
1115: if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)((Mat_MPIBAIJ *)A->data)->A, matname));
1116: PetscCall(MatView_SeqBAIJ(((Mat_MPIBAIJ *)A->data)->A, sviewer));
1117: }
1118: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1119: PetscCall(MatDestroy(&A));
1120: }
1121: PetscFunctionReturn(PETSC_SUCCESS);
1122: }
1124: /* Used for both MPIBAIJ and MPISBAIJ matrices */
1125: PetscErrorCode MatView_MPIBAIJ_Binary(Mat mat, PetscViewer viewer)
1126: {
1127: Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)mat->data;
1128: Mat_SeqBAIJ *A = (Mat_SeqBAIJ *)aij->A->data;
1129: Mat_SeqBAIJ *B = (Mat_SeqBAIJ *)aij->B->data;
1130: const PetscInt *garray = aij->garray;
1131: PetscInt header[4], M, N, m, rs, cs, bs, cnt, i, j, ja, jb, k, l;
1132: PetscCount nz, hnz;
1133: PetscInt *rowlens, *colidxs;
1134: PetscScalar *matvals;
1135: PetscMPIInt rank;
1137: PetscFunctionBegin;
1138: PetscCall(PetscViewerSetUp(viewer));
1140: M = mat->rmap->N;
1141: N = mat->cmap->N;
1142: m = mat->rmap->n;
1143: rs = mat->rmap->rstart;
1144: cs = mat->cmap->rstart;
1145: bs = mat->rmap->bs;
1146: nz = bs * bs * (A->nz + B->nz);
1148: /* write matrix header */
1149: header[0] = MAT_FILE_CLASSID;
1150: header[1] = M;
1151: header[2] = N;
1152: PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_COUNT, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1153: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1154: if (rank == 0) PetscCall(PetscIntCast(hnz, &header[3]));
1155: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1157: /* fill in and store row lengths */
1158: PetscCall(PetscMalloc1(m, &rowlens));
1159: for (cnt = 0, i = 0; i < A->mbs; i++)
1160: for (j = 0; j < bs; j++) rowlens[cnt++] = bs * (A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i]);
1161: PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1162: PetscCall(PetscFree(rowlens));
1164: /* fill in and store column indices */
1165: PetscCall(PetscMalloc1(nz, &colidxs));
1166: for (cnt = 0, i = 0; i < A->mbs; i++) {
1167: for (k = 0; k < bs; k++) {
1168: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1169: if (garray[B->j[jb]] > cs / bs) break;
1170: for (l = 0; l < bs; l++) colidxs[cnt++] = bs * garray[B->j[jb]] + l;
1171: }
1172: for (ja = A->i[i]; ja < A->i[i + 1]; ja++)
1173: for (l = 0; l < bs; l++) colidxs[cnt++] = bs * A->j[ja] + l + cs;
1174: for (; jb < B->i[i + 1]; jb++)
1175: for (l = 0; l < bs; l++) colidxs[cnt++] = bs * garray[B->j[jb]] + l;
1176: }
1177: }
1178: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscCount_FMT, cnt, nz);
1179: PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DECIDE, PETSC_DECIDE, PETSC_INT));
1180: PetscCall(PetscFree(colidxs));
1182: /* fill in and store nonzero values */
1183: PetscCall(PetscMalloc1(nz, &matvals));
1184: for (cnt = 0, i = 0; i < A->mbs; i++) {
1185: for (k = 0; k < bs; k++) {
1186: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1187: if (garray[B->j[jb]] > cs / bs) break;
1188: for (l = 0; l < bs; l++) matvals[cnt++] = B->a[bs * (bs * jb + l) + k];
1189: }
1190: for (ja = A->i[i]; ja < A->i[i + 1]; ja++)
1191: for (l = 0; l < bs; l++) matvals[cnt++] = A->a[bs * (bs * ja + l) + k];
1192: for (; jb < B->i[i + 1]; jb++)
1193: for (l = 0; l < bs; l++) matvals[cnt++] = B->a[bs * (bs * jb + l) + k];
1194: }
1195: }
1196: PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DECIDE, PETSC_DECIDE, PETSC_SCALAR));
1197: PetscCall(PetscFree(matvals));
1199: /* write block size option to the viewer's .info file */
1200: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1201: PetscFunctionReturn(PETSC_SUCCESS);
1202: }
1204: PetscErrorCode MatView_MPIBAIJ(Mat mat, PetscViewer viewer)
1205: {
1206: PetscBool isascii, isdraw, issocket, isbinary;
1208: PetscFunctionBegin;
1209: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1210: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1211: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1212: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1213: if (isascii || isdraw || issocket) PetscCall(MatView_MPIBAIJ_ASCIIorDraworSocket(mat, viewer));
1214: else if (isbinary) PetscCall(MatView_MPIBAIJ_Binary(mat, viewer));
1215: PetscFunctionReturn(PETSC_SUCCESS);
1216: }
1218: static PetscErrorCode MatMult_MPIBAIJ(Mat A, Vec xx, Vec yy)
1219: {
1220: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1221: PetscInt nt;
1223: PetscFunctionBegin;
1224: PetscCall(VecGetLocalSize(xx, &nt));
1225: PetscCheck(nt == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Incompatible partition of A and xx");
1226: PetscCall(VecGetLocalSize(yy, &nt));
1227: PetscCheck(nt == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Incompatible partition of A and yy");
1228: PetscCall(VecScatterBegin(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1229: PetscCall((*a->A->ops->mult)(a->A, xx, yy));
1230: PetscCall(VecScatterEnd(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1231: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, yy, yy));
1232: PetscFunctionReturn(PETSC_SUCCESS);
1233: }
1235: static PetscErrorCode MatMultAdd_MPIBAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1236: {
1237: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1239: PetscFunctionBegin;
1240: PetscCall(VecScatterBegin(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1241: PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
1242: PetscCall(VecScatterEnd(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1243: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
1244: PetscFunctionReturn(PETSC_SUCCESS);
1245: }
1247: static PetscErrorCode MatMultTranspose_MPIBAIJ(Mat A, Vec xx, Vec yy)
1248: {
1249: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1251: PetscFunctionBegin;
1252: /* do nondiagonal part */
1253: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1254: /* do local part */
1255: PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
1256: /* add partial results together */
1257: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1258: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1259: PetscFunctionReturn(PETSC_SUCCESS);
1260: }
1262: static PetscErrorCode MatMultTransposeAdd_MPIBAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1263: {
1264: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1266: PetscFunctionBegin;
1267: /* do nondiagonal part */
1268: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1269: /* do local part */
1270: PetscCall((*a->A->ops->multtransposeadd)(a->A, xx, yy, zz));
1271: /* add partial results together */
1272: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1273: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1274: PetscFunctionReturn(PETSC_SUCCESS);
1275: }
1277: /*
1278: This only works correctly for square matrices where the subblock A->A is the
1279: diagonal block
1280: */
1281: static PetscErrorCode MatGetDiagonal_MPIBAIJ(Mat A, Vec v)
1282: {
1283: PetscFunctionBegin;
1284: PetscCheck(A->rmap->N == A->cmap->N, PETSC_COMM_SELF, PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1285: PetscCall(MatGetDiagonal(((Mat_MPIBAIJ *)A->data)->A, v));
1286: PetscFunctionReturn(PETSC_SUCCESS);
1287: }
1289: static PetscErrorCode MatScale_MPIBAIJ(Mat A, PetscScalar aa)
1290: {
1291: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1293: PetscFunctionBegin;
1294: PetscCall(MatScale(a->A, aa));
1295: PetscCall(MatScale(a->B, aa));
1296: PetscFunctionReturn(PETSC_SUCCESS);
1297: }
1299: static PetscErrorCode MatGetRow_MPIBAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1300: {
1301: Mat_MPIBAIJ *mat = (Mat_MPIBAIJ *)matin->data;
1302: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1303: PetscInt bs = matin->rmap->bs, bs2 = mat->bs2, i, *cworkA, *cworkB, **pcA, **pcB;
1304: PetscInt nztot, nzA, nzB, lrow, brstart = matin->rmap->rstart, brend = matin->rmap->rend;
1305: PetscInt *cmap, *idx_p, cstart = mat->cstartbs;
1307: PetscFunctionBegin;
1308: PetscCheck(row >= brstart && row < brend, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only local rows");
1309: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1310: mat->getrowactive = PETSC_TRUE;
1312: if (!mat->rowvalues && (idx || v)) {
1313: /*
1314: allocate enough space to hold information from the longest row.
1315: */
1316: Mat_SeqBAIJ *Aa = (Mat_SeqBAIJ *)mat->A->data, *Ba = (Mat_SeqBAIJ *)mat->B->data;
1317: PetscInt max = 1, mbs = mat->mbs, tmp;
1318: for (i = 0; i < mbs; i++) {
1319: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1320: if (max < tmp) max = tmp;
1321: }
1322: PetscCall(PetscMalloc2(max * bs2, &mat->rowvalues, max * bs2, &mat->rowindices));
1323: }
1324: lrow = row - brstart;
1326: pvA = &vworkA;
1327: pcA = &cworkA;
1328: pvB = &vworkB;
1329: pcB = &cworkB;
1330: if (!v) {
1331: pvA = NULL;
1332: pvB = NULL;
1333: }
1334: if (!idx) {
1335: pcA = NULL;
1336: if (!v) pcB = NULL;
1337: }
1338: PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1339: PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1340: nztot = nzA + nzB;
1342: cmap = mat->garray;
1343: if (v || idx) {
1344: if (nztot) {
1345: /* Sort by increasing column numbers, assuming A and B already sorted */
1346: PetscInt imark = -1;
1347: if (v) {
1348: *v = v_p = mat->rowvalues;
1349: for (i = 0; i < nzB; i++) {
1350: if (cmap[cworkB[i] / bs] < cstart) v_p[i] = vworkB[i];
1351: else break;
1352: }
1353: imark = i;
1354: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1355: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1356: }
1357: if (idx) {
1358: *idx = idx_p = mat->rowindices;
1359: if (imark > -1) {
1360: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i] / bs] * bs + cworkB[i] % bs;
1361: } else {
1362: for (i = 0; i < nzB; i++) {
1363: if (cmap[cworkB[i] / bs] < cstart) idx_p[i] = cmap[cworkB[i] / bs] * bs + cworkB[i] % bs;
1364: else break;
1365: }
1366: imark = i;
1367: }
1368: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart * bs + cworkA[i];
1369: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i] / bs] * bs + cworkB[i] % bs;
1370: }
1371: } else {
1372: if (idx) *idx = NULL;
1373: if (v) *v = NULL;
1374: }
1375: }
1376: *nz = nztot;
1377: PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1378: PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1379: PetscFunctionReturn(PETSC_SUCCESS);
1380: }
1382: static PetscErrorCode MatRestoreRow_MPIBAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1383: {
1384: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
1386: PetscFunctionBegin;
1387: PetscCheck(baij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow not called");
1388: baij->getrowactive = PETSC_FALSE;
1389: PetscFunctionReturn(PETSC_SUCCESS);
1390: }
1392: static PetscErrorCode MatZeroEntries_MPIBAIJ(Mat A)
1393: {
1394: Mat_MPIBAIJ *l = (Mat_MPIBAIJ *)A->data;
1396: PetscFunctionBegin;
1397: PetscCall(MatZeroEntries(l->A));
1398: PetscCall(MatZeroEntries(l->B));
1399: PetscFunctionReturn(PETSC_SUCCESS);
1400: }
1402: static PetscErrorCode MatGetInfo_MPIBAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1403: {
1404: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)matin->data;
1405: Mat A = a->A, B = a->B;
1406: PetscLogDouble isend[5], irecv[5];
1408: PetscFunctionBegin;
1409: info->block_size = (PetscReal)matin->rmap->bs;
1411: PetscCall(MatGetInfo(A, MAT_LOCAL, info));
1413: isend[0] = info->nz_used;
1414: isend[1] = info->nz_allocated;
1415: isend[2] = info->nz_unneeded;
1416: isend[3] = info->memory;
1417: isend[4] = info->mallocs;
1419: PetscCall(MatGetInfo(B, MAT_LOCAL, info));
1421: isend[0] += info->nz_used;
1422: isend[1] += info->nz_allocated;
1423: isend[2] += info->nz_unneeded;
1424: isend[3] += info->memory;
1425: isend[4] += info->mallocs;
1427: if (flag == MAT_LOCAL) {
1428: info->nz_used = isend[0];
1429: info->nz_allocated = isend[1];
1430: info->nz_unneeded = isend[2];
1431: info->memory = isend[3];
1432: info->mallocs = isend[4];
1433: } else if (flag == MAT_GLOBAL_MAX) {
1434: PetscCallMPI(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));
1436: info->nz_used = irecv[0];
1437: info->nz_allocated = irecv[1];
1438: info->nz_unneeded = irecv[2];
1439: info->memory = irecv[3];
1440: info->mallocs = irecv[4];
1441: } else if (flag == MAT_GLOBAL_SUM) {
1442: PetscCallMPI(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));
1444: info->nz_used = irecv[0];
1445: info->nz_allocated = irecv[1];
1446: info->nz_unneeded = irecv[2];
1447: info->memory = irecv[3];
1448: info->mallocs = irecv[4];
1449: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_ARG_WRONG, "Unknown MatInfoType argument %d", (int)flag);
1450: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1451: info->fill_ratio_needed = 0;
1452: info->factor_mallocs = 0;
1453: PetscFunctionReturn(PETSC_SUCCESS);
1454: }
1456: static PetscErrorCode MatSetOption_MPIBAIJ(Mat A, MatOption op, PetscBool flg)
1457: {
1458: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1460: PetscFunctionBegin;
1461: switch (op) {
1462: case MAT_NEW_NONZERO_LOCATIONS:
1463: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1464: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1465: case MAT_KEEP_NONZERO_PATTERN:
1466: case MAT_NEW_NONZERO_LOCATION_ERR:
1467: MatCheckPreallocated(A, 1);
1468: PetscCall(MatSetOption(a->A, op, flg));
1469: PetscCall(MatSetOption(a->B, op, flg));
1470: break;
1471: case MAT_ROW_ORIENTED:
1472: MatCheckPreallocated(A, 1);
1473: a->roworiented = flg;
1475: PetscCall(MatSetOption(a->A, op, flg));
1476: PetscCall(MatSetOption(a->B, op, flg));
1477: break;
1478: case MAT_IGNORE_OFF_PROC_ENTRIES:
1479: a->donotstash = flg;
1480: break;
1481: case MAT_USE_HASH_TABLE:
1482: a->ht_flag = flg;
1483: a->ht_fact = 1.39;
1484: break;
1485: case MAT_SPD:
1486: case MAT_SYMMETRIC:
1487: case MAT_STRUCTURALLY_SYMMETRIC:
1488: case MAT_HERMITIAN:
1489: case MAT_SYMMETRY_ETERNAL:
1490: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1491: case MAT_SPD_ETERNAL:
1492: /* if the diagonal matrix is square it inherits some of the properties above */
1493: if (a->A && A->rmap->n == A->cmap->n) PetscCall(MatSetOption(a->A, op, flg));
1494: break;
1495: default:
1496: break;
1497: }
1498: PetscFunctionReturn(PETSC_SUCCESS);
1499: }
1501: static PetscErrorCode MatTranspose_MPIBAIJ(Mat A, MatReuse reuse, Mat *matout)
1502: {
1503: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)A->data;
1504: Mat_SeqBAIJ *Aloc;
1505: Mat B;
1506: PetscInt M = A->rmap->N, N = A->cmap->N, *ai, *aj, i, *rvals, j, k, col;
1507: PetscInt bs = A->rmap->bs, mbs = baij->mbs;
1508: MatScalar *a;
1510: PetscFunctionBegin;
1511: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1512: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_INPLACE_MATRIX) {
1513: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1514: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1515: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1516: /* Do not know preallocation information, but must set block size */
1517: PetscCall(MatMPIBAIJSetPreallocation(B, A->rmap->bs, PETSC_DECIDE, NULL, PETSC_DECIDE, NULL));
1518: } else {
1519: B = *matout;
1520: }
1522: /* copy over the A part */
1523: Aloc = (Mat_SeqBAIJ *)baij->A->data;
1524: ai = Aloc->i;
1525: aj = Aloc->j;
1526: a = Aloc->a;
1527: PetscCall(PetscMalloc1(bs, &rvals));
1529: for (i = 0; i < mbs; i++) {
1530: rvals[0] = bs * (baij->rstartbs + i);
1531: for (j = 1; j < bs; j++) rvals[j] = rvals[j - 1] + 1;
1532: for (j = ai[i]; j < ai[i + 1]; j++) {
1533: col = (baij->cstartbs + aj[j]) * bs;
1534: for (k = 0; k < bs; k++) {
1535: PetscCall(MatSetValues_MPIBAIJ(B, 1, &col, bs, rvals, a, INSERT_VALUES));
1537: col++;
1538: a += bs;
1539: }
1540: }
1541: }
1542: /* copy over the B part */
1543: Aloc = (Mat_SeqBAIJ *)baij->B->data;
1544: ai = Aloc->i;
1545: aj = Aloc->j;
1546: a = Aloc->a;
1547: for (i = 0; i < mbs; i++) {
1548: rvals[0] = bs * (baij->rstartbs + i);
1549: for (j = 1; j < bs; j++) rvals[j] = rvals[j - 1] + 1;
1550: for (j = ai[i]; j < ai[i + 1]; j++) {
1551: col = baij->garray[aj[j]] * bs;
1552: for (k = 0; k < bs; k++) {
1553: PetscCall(MatSetValues_MPIBAIJ(B, 1, &col, bs, rvals, a, INSERT_VALUES));
1554: col++;
1555: a += bs;
1556: }
1557: }
1558: }
1559: PetscCall(PetscFree(rvals));
1560: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1561: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1563: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) *matout = B;
1564: else PetscCall(MatHeaderMerge(A, &B));
1565: PetscFunctionReturn(PETSC_SUCCESS);
1566: }
1568: static PetscErrorCode MatDiagonalScale_MPIBAIJ(Mat mat, Vec ll, Vec rr)
1569: {
1570: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
1571: Mat a = baij->A, b = baij->B;
1572: PetscInt s1, s2, s3;
1574: PetscFunctionBegin;
1575: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1576: if (rr) {
1577: PetscCall(VecGetLocalSize(rr, &s1));
1578: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1579: /* Overlap communication with computation. */
1580: PetscCall(VecScatterBegin(baij->Mvctx, rr, baij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1581: }
1582: if (ll) {
1583: PetscCall(VecGetLocalSize(ll, &s1));
1584: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1585: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1586: }
1587: /* scale the diagonal block */
1588: PetscUseTypeMethod(a, diagonalscale, ll, rr);
1590: if (rr) {
1591: /* Do a scatter end and then right scale the off-diagonal block */
1592: PetscCall(VecScatterEnd(baij->Mvctx, rr, baij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1593: PetscUseTypeMethod(b, diagonalscale, NULL, baij->lvec);
1594: }
1595: PetscFunctionReturn(PETSC_SUCCESS);
1596: }
1598: static PetscErrorCode MatZeroRows_MPIBAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
1599: {
1600: Mat_MPIBAIJ *l = (Mat_MPIBAIJ *)A->data;
1601: PetscInt *lrows;
1602: PetscInt r, len;
1603: PetscBool cong;
1605: PetscFunctionBegin;
1606: /* get locally owned rows */
1607: PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
1608: /* fix right-hand side if needed */
1609: if (x && b) {
1610: const PetscScalar *xx;
1611: PetscScalar *bb;
1613: PetscCall(VecGetArrayRead(x, &xx));
1614: PetscCall(VecGetArray(b, &bb));
1615: for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
1616: PetscCall(VecRestoreArrayRead(x, &xx));
1617: PetscCall(VecRestoreArray(b, &bb));
1618: }
1620: /* actually zap the local rows */
1621: /*
1622: Zero the required rows. If the "diagonal block" of the matrix
1623: is square and the user wishes to set the diagonal we use separate
1624: code so that MatSetValues() is not called for each diagonal allocating
1625: new memory, thus calling lots of mallocs and slowing things down.
1627: */
1628: /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
1629: PetscCall(MatZeroRows_SeqBAIJ(l->B, len, lrows, 0.0, NULL, NULL));
1630: PetscCall(MatHasCongruentLayouts(A, &cong));
1631: if ((diag != 0.0) && cong) {
1632: PetscCall(MatZeroRows_SeqBAIJ(l->A, len, lrows, diag, NULL, NULL));
1633: } else if (diag != 0.0) {
1634: PetscCall(MatZeroRows_SeqBAIJ(l->A, len, lrows, 0.0, NULL, NULL));
1635: PetscCheck(!((Mat_SeqBAIJ *)l->A->data)->nonew, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatZeroRows() on rectangular matrices cannot be used with the Mat options MAT_NEW_NONZERO_LOCATIONS, MAT_NEW_NONZERO_LOCATION_ERR, and MAT_NEW_NONZERO_ALLOCATION_ERR");
1636: for (r = 0; r < len; ++r) {
1637: const PetscInt row = lrows[r] + A->rmap->rstart;
1638: PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
1639: }
1640: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
1641: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
1642: } else {
1643: PetscCall(MatZeroRows_SeqBAIJ(l->A, len, lrows, 0.0, NULL, NULL));
1644: }
1645: PetscCall(PetscFree(lrows));
1647: /* only change matrix nonzero state if pattern was allowed to be changed */
1648: if (!((Mat_SeqBAIJ *)l->A->data)->keepnonzeropattern || !((Mat_SeqBAIJ *)l->A->data)->nonew) {
1649: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1650: PetscCallMPI(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1651: }
1652: PetscFunctionReturn(PETSC_SUCCESS);
1653: }
1655: static PetscErrorCode MatZeroRowsColumns_MPIBAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
1656: {
1657: Mat_MPIBAIJ *l = (Mat_MPIBAIJ *)A->data;
1658: PetscMPIInt n, p = 0;
1659: PetscInt i, j, k, r, len = 0, row, col, count;
1660: PetscInt *lrows, *owners = A->rmap->range;
1661: PetscSFNode *rrows;
1662: PetscSF sf;
1663: const PetscScalar *xx;
1664: PetscScalar *bb, *mask;
1665: Vec xmask, lmask;
1666: Mat_SeqBAIJ *baij = (Mat_SeqBAIJ *)l->B->data;
1667: PetscInt bs = A->rmap->bs, bs2 = baij->bs2;
1668: PetscScalar *aa;
1670: PetscFunctionBegin;
1671: PetscCall(PetscMPIIntCast(A->rmap->n, &n));
1672: /* create PetscSF where leaves are input rows and roots are owned rows */
1673: PetscCall(PetscMalloc1(n, &lrows));
1674: for (r = 0; r < n; ++r) lrows[r] = -1;
1675: PetscCall(PetscMalloc1(N, &rrows));
1676: for (r = 0; r < N; ++r) {
1677: const PetscInt idx = rows[r];
1678: PetscCheck(idx >= 0 && A->rmap->N > idx, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %" PetscInt_FMT " out of range [0,%" PetscInt_FMT ")", idx, A->rmap->N);
1679: if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
1680: PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
1681: }
1682: rrows[r].rank = p;
1683: rrows[r].index = rows[r] - owners[p];
1684: }
1685: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1686: PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
1687: /* collect flags for rows to be zeroed */
1688: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
1689: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
1690: PetscCall(PetscSFDestroy(&sf));
1691: /* compress and put in row numbers */
1692: for (r = 0; r < n; ++r)
1693: if (lrows[r] >= 0) lrows[len++] = r;
1694: /* zero diagonal part of matrix */
1695: PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
1696: /* handle off-diagonal part of matrix */
1697: PetscCall(MatCreateVecs(A, &xmask, NULL));
1698: PetscCall(VecDuplicate(l->lvec, &lmask));
1699: PetscCall(VecGetArray(xmask, &bb));
1700: for (i = 0; i < len; i++) bb[lrows[i]] = 1;
1701: PetscCall(VecRestoreArray(xmask, &bb));
1702: PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
1703: PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
1704: PetscCall(VecDestroy(&xmask));
1705: if (x) {
1706: PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
1707: PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
1708: PetscCall(VecGetArrayRead(l->lvec, &xx));
1709: PetscCall(VecGetArray(b, &bb));
1710: }
1711: PetscCall(VecGetArray(lmask, &mask));
1712: /* remove zeroed rows of off-diagonal matrix */
1713: for (i = 0; i < len; ++i) {
1714: row = lrows[i];
1715: count = (baij->i[row / bs + 1] - baij->i[row / bs]) * bs;
1716: aa = baij->a + baij->i[row / bs] * bs2 + (row % bs);
1717: for (k = 0; k < count; ++k) {
1718: aa[0] = 0.0;
1719: aa += bs;
1720: }
1721: }
1722: /* loop over all elements of off process part of matrix zeroing removed columns */
1723: for (i = 0; i < l->B->rmap->N; ++i) {
1724: row = i / bs;
1725: for (j = baij->i[row]; j < baij->i[row + 1]; ++j) {
1726: for (k = 0; k < bs; ++k) {
1727: col = bs * baij->j[j] + k;
1728: if (PetscAbsScalar(mask[col])) {
1729: aa = baij->a + j * bs2 + (i % bs) + bs * k;
1730: if (x) bb[i] -= aa[0] * xx[col];
1731: aa[0] = 0.0;
1732: }
1733: }
1734: }
1735: }
1736: if (x) {
1737: PetscCall(VecRestoreArray(b, &bb));
1738: PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1739: }
1740: PetscCall(VecRestoreArray(lmask, &mask));
1741: PetscCall(VecDestroy(&lmask));
1742: PetscCall(PetscFree(lrows));
1744: /* only change matrix nonzero state if pattern was allowed to be changed */
1745: if (!((Mat_SeqBAIJ *)l->A->data)->nonew) {
1746: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1747: PetscCallMPI(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1748: }
1749: PetscFunctionReturn(PETSC_SUCCESS);
1750: }
1752: static PetscErrorCode MatSetUnfactored_MPIBAIJ(Mat A)
1753: {
1754: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1756: PetscFunctionBegin;
1757: PetscCall(MatSetUnfactored(a->A));
1758: PetscFunctionReturn(PETSC_SUCCESS);
1759: }
1761: static PetscErrorCode MatDuplicate_MPIBAIJ(Mat, MatDuplicateOption, Mat *);
1763: static PetscErrorCode MatEqual_MPIBAIJ(Mat A, Mat B, PetscBool *flag)
1764: {
1765: Mat_MPIBAIJ *matB = (Mat_MPIBAIJ *)B->data, *matA = (Mat_MPIBAIJ *)A->data;
1766: Mat a, b, c, d;
1767: PetscBool flg;
1769: PetscFunctionBegin;
1770: a = matA->A;
1771: b = matA->B;
1772: c = matB->A;
1773: d = matB->B;
1775: PetscCall(MatEqual(a, c, &flg));
1776: if (flg) PetscCall(MatEqual(b, d, &flg));
1777: PetscCallMPI(MPIU_Allreduce(&flg, flag, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
1778: PetscFunctionReturn(PETSC_SUCCESS);
1779: }
1781: static PetscErrorCode MatCopy_MPIBAIJ(Mat A, Mat B, MatStructure str)
1782: {
1783: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1784: Mat_MPIBAIJ *b = (Mat_MPIBAIJ *)B->data;
1786: PetscFunctionBegin;
1787: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
1788: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
1789: PetscCall(MatCopy_Basic(A, B, str));
1790: } else {
1791: PetscCall(MatCopy(a->A, b->A, str));
1792: PetscCall(MatCopy(a->B, b->B, str));
1793: }
1794: PetscCall(PetscObjectStateIncrease((PetscObject)B));
1795: PetscFunctionReturn(PETSC_SUCCESS);
1796: }
1798: PetscErrorCode MatAXPYGetPreallocation_MPIBAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
1799: {
1800: PetscInt bs = Y->rmap->bs, m = Y->rmap->N / bs;
1801: Mat_SeqBAIJ *x = (Mat_SeqBAIJ *)X->data;
1802: Mat_SeqBAIJ *y = (Mat_SeqBAIJ *)Y->data;
1804: PetscFunctionBegin;
1805: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
1806: PetscFunctionReturn(PETSC_SUCCESS);
1807: }
1809: static PetscErrorCode MatAXPY_MPIBAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
1810: {
1811: Mat_MPIBAIJ *xx = (Mat_MPIBAIJ *)X->data, *yy = (Mat_MPIBAIJ *)Y->data;
1812: PetscBLASInt bnz, one = 1;
1813: Mat_SeqBAIJ *x, *y;
1814: PetscInt bs2 = Y->rmap->bs * Y->rmap->bs;
1816: PetscFunctionBegin;
1817: if (str == SAME_NONZERO_PATTERN) {
1818: PetscScalar alpha = a;
1819: x = (Mat_SeqBAIJ *)xx->A->data;
1820: y = (Mat_SeqBAIJ *)yy->A->data;
1821: PetscCall(PetscBLASIntCast(x->nz * bs2, &bnz));
1822: PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, x->a, &one, y->a, &one));
1823: x = (Mat_SeqBAIJ *)xx->B->data;
1824: y = (Mat_SeqBAIJ *)yy->B->data;
1825: PetscCall(PetscBLASIntCast(x->nz * bs2, &bnz));
1826: PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, x->a, &one, y->a, &one));
1827: PetscCall(PetscObjectStateIncrease((PetscObject)Y));
1828: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
1829: PetscCall(MatAXPY_Basic(Y, a, X, str));
1830: } else {
1831: Mat B;
1832: PetscInt *nnz_d, *nnz_o, bs = Y->rmap->bs;
1833: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
1834: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
1835: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
1836: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
1837: PetscCall(MatSetSizes(B, Y->rmap->n, Y->cmap->n, Y->rmap->N, Y->cmap->N));
1838: PetscCall(MatSetBlockSizesFromMats(B, Y, Y));
1839: PetscCall(MatSetType(B, MATMPIBAIJ));
1840: PetscCall(MatAXPYGetPreallocation_SeqBAIJ(yy->A, xx->A, nnz_d));
1841: PetscCall(MatAXPYGetPreallocation_MPIBAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
1842: PetscCall(MatMPIBAIJSetPreallocation(B, bs, 0, nnz_d, 0, nnz_o));
1843: /* MatAXPY_BasicWithPreallocation() for BAIJ matrix is much slower than AIJ, even for bs=1 ! */
1844: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
1845: PetscCall(MatHeaderMerge(Y, &B));
1846: PetscCall(PetscFree(nnz_d));
1847: PetscCall(PetscFree(nnz_o));
1848: }
1849: PetscFunctionReturn(PETSC_SUCCESS);
1850: }
1852: static PetscErrorCode MatConjugate_MPIBAIJ(Mat mat)
1853: {
1854: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)mat->data;
1856: PetscFunctionBegin;
1857: PetscCall(MatConjugate_SeqBAIJ(a->A));
1858: PetscCall(MatConjugate_SeqBAIJ(a->B));
1859: PetscFunctionReturn(PETSC_SUCCESS);
1860: }
1862: static PetscErrorCode MatRealPart_MPIBAIJ(Mat A)
1863: {
1864: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1866: PetscFunctionBegin;
1867: PetscCall(MatRealPart(a->A));
1868: PetscCall(MatRealPart(a->B));
1869: PetscFunctionReturn(PETSC_SUCCESS);
1870: }
1872: static PetscErrorCode MatImaginaryPart_MPIBAIJ(Mat A)
1873: {
1874: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
1876: PetscFunctionBegin;
1877: PetscCall(MatImaginaryPart(a->A));
1878: PetscCall(MatImaginaryPart(a->B));
1879: PetscFunctionReturn(PETSC_SUCCESS);
1880: }
1882: static PetscErrorCode MatCreateSubMatrix_MPIBAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
1883: {
1884: IS iscol_local;
1885: PetscInt csize;
1887: PetscFunctionBegin;
1888: PetscCall(ISGetLocalSize(iscol, &csize));
1889: if (call == MAT_REUSE_MATRIX) {
1890: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
1891: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
1892: } else {
1893: PetscCall(ISAllGather(iscol, &iscol_local));
1894: }
1895: PetscCall(MatCreateSubMatrix_MPIBAIJ_Private(mat, isrow, iscol_local, csize, call, newmat, PETSC_FALSE));
1896: if (call == MAT_INITIAL_MATRIX) {
1897: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
1898: PetscCall(ISDestroy(&iscol_local));
1899: }
1900: PetscFunctionReturn(PETSC_SUCCESS);
1901: }
1903: /*
1904: Not great since it makes two copies of the submatrix, first an SeqBAIJ
1905: in local and then by concatenating the local matrices the end result.
1906: Writing it directly would be much like MatCreateSubMatrices_MPIBAIJ().
1907: This routine is used for BAIJ and SBAIJ matrices (unfortunate dependency).
1908: */
1909: PetscErrorCode MatCreateSubMatrix_MPIBAIJ_Private(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat, PetscBool sym)
1910: {
1911: PetscMPIInt rank, size;
1912: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs;
1913: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
1914: Mat M, Mreuse;
1915: MatScalar *vwork, *aa;
1916: MPI_Comm comm;
1917: IS isrow_new, iscol_new;
1918: Mat_SeqBAIJ *aij;
1920: PetscFunctionBegin;
1921: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
1922: PetscCallMPI(MPI_Comm_rank(comm, &rank));
1923: PetscCallMPI(MPI_Comm_size(comm, &size));
1924: /* The compression and expansion should be avoided. Doesn't point
1925: out errors, might change the indices, hence buggey */
1926: PetscCall(ISCompressIndicesGeneral(mat->rmap->N, mat->rmap->n, mat->rmap->bs, 1, &isrow, &isrow_new));
1927: if (isrow == iscol) {
1928: iscol_new = isrow_new;
1929: PetscCall(PetscObjectReference((PetscObject)iscol_new));
1930: } else PetscCall(ISCompressIndicesGeneral(mat->cmap->N, mat->cmap->n, mat->cmap->bs, 1, &iscol, &iscol_new));
1932: if (call == MAT_REUSE_MATRIX) {
1933: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
1934: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
1935: PetscCall(MatCreateSubMatrices_MPIBAIJ_local(mat, 1, &isrow_new, &iscol_new, MAT_REUSE_MATRIX, &Mreuse, sym));
1936: } else {
1937: PetscCall(MatCreateSubMatrices_MPIBAIJ_local(mat, 1, &isrow_new, &iscol_new, MAT_INITIAL_MATRIX, &Mreuse, sym));
1938: }
1939: PetscCall(ISDestroy(&isrow_new));
1940: PetscCall(ISDestroy(&iscol_new));
1941: /*
1942: m - number of local rows
1943: n - number of columns (same on all processors)
1944: rstart - first row in new global matrix generated
1945: */
1946: PetscCall(MatGetBlockSize(mat, &bs));
1947: PetscCall(MatGetSize(Mreuse, &m, &n));
1948: m = m / bs;
1949: n = n / bs;
1951: if (call == MAT_INITIAL_MATRIX) {
1952: aij = (Mat_SeqBAIJ *)Mreuse->data;
1953: ii = aij->i;
1954: jj = aij->j;
1956: /*
1957: Determine the number of non-zeros in the diagonal and off-diagonal
1958: portions of the matrix in order to do correct preallocation
1959: */
1961: /* first get start and end of "diagonal" columns */
1962: if (csize == PETSC_DECIDE) {
1963: PetscCall(ISGetSize(isrow, &mglobal));
1964: if (mglobal == n * bs) { /* square matrix */
1965: nlocal = m;
1966: } else {
1967: nlocal = n / size + ((n % size) > rank);
1968: }
1969: } else {
1970: nlocal = csize / bs;
1971: }
1972: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
1973: rstart = rend - nlocal;
1974: PetscCheck(rank != size - 1 || rend == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Local column sizes %" PetscInt_FMT " do not add up to total number of columns %" PetscInt_FMT, rend, n);
1976: /* next, compute all the lengths */
1977: PetscCall(PetscMalloc2(m + 1, &dlens, m + 1, &olens));
1978: for (i = 0; i < m; i++) {
1979: jend = ii[i + 1] - ii[i];
1980: olen = 0;
1981: dlen = 0;
1982: for (j = 0; j < jend; j++) {
1983: if (*jj < rstart || *jj >= rend) olen++;
1984: else dlen++;
1985: jj++;
1986: }
1987: olens[i] = olen;
1988: dlens[i] = dlen;
1989: }
1990: PetscCall(MatCreate(comm, &M));
1991: PetscCall(MatSetSizes(M, bs * m, bs * nlocal, PETSC_DECIDE, bs * n));
1992: PetscCall(MatSetType(M, sym ? ((PetscObject)mat)->type_name : MATMPIBAIJ));
1993: PetscCall(MatMPIBAIJSetPreallocation(M, bs, 0, dlens, 0, olens));
1994: PetscCall(MatMPISBAIJSetPreallocation(M, bs, 0, dlens, 0, olens));
1995: PetscCall(PetscFree2(dlens, olens));
1996: } else {
1997: PetscInt ml, nl;
1999: M = *newmat;
2000: PetscCall(MatGetLocalSize(M, &ml, &nl));
2001: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
2002: PetscCall(MatZeroEntries(M));
2003: /*
2004: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
2005: rather than the slower MatSetValues().
2006: */
2007: M->was_assembled = PETSC_TRUE;
2008: M->assembled = PETSC_FALSE;
2009: }
2010: PetscCall(MatSetOption(M, MAT_ROW_ORIENTED, PETSC_FALSE));
2011: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
2012: aij = (Mat_SeqBAIJ *)Mreuse->data;
2013: ii = aij->i;
2014: jj = aij->j;
2015: aa = aij->a;
2016: for (i = 0; i < m; i++) {
2017: row = rstart / bs + i;
2018: nz = ii[i + 1] - ii[i];
2019: cwork = jj;
2020: jj = PetscSafePointerPlusOffset(jj, nz);
2021: vwork = aa;
2022: aa = PetscSafePointerPlusOffset(aa, nz * bs * bs);
2023: PetscCall(MatSetValuesBlocked_MPIBAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
2024: }
2026: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
2027: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
2028: *newmat = M;
2030: /* save submatrix used in processor for next request */
2031: if (call == MAT_INITIAL_MATRIX) {
2032: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
2033: PetscCall(PetscObjectDereference((PetscObject)Mreuse));
2034: }
2035: PetscFunctionReturn(PETSC_SUCCESS);
2036: }
2038: static PetscErrorCode MatPermute_MPIBAIJ(Mat A, IS rowp, IS colp, Mat *B)
2039: {
2040: MPI_Comm comm, pcomm;
2041: PetscInt clocal_size, nrows;
2042: const PetscInt *rows;
2043: PetscMPIInt size;
2044: IS crowp, lcolp;
2046: PetscFunctionBegin;
2047: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
2048: /* make a collective version of 'rowp' */
2049: PetscCall(PetscObjectGetComm((PetscObject)rowp, &pcomm));
2050: if (pcomm == comm) {
2051: crowp = rowp;
2052: } else {
2053: PetscCall(ISGetSize(rowp, &nrows));
2054: PetscCall(ISGetIndices(rowp, &rows));
2055: PetscCall(ISCreateGeneral(comm, nrows, rows, PETSC_COPY_VALUES, &crowp));
2056: PetscCall(ISRestoreIndices(rowp, &rows));
2057: }
2058: PetscCall(ISSetPermutation(crowp));
2059: /* make a local version of 'colp' */
2060: PetscCall(PetscObjectGetComm((PetscObject)colp, &pcomm));
2061: PetscCallMPI(MPI_Comm_size(pcomm, &size));
2062: if (size == 1) {
2063: lcolp = colp;
2064: } else {
2065: PetscCall(ISAllGather(colp, &lcolp));
2066: }
2067: PetscCall(ISSetPermutation(lcolp));
2068: /* now we just get the submatrix */
2069: PetscCall(MatGetLocalSize(A, NULL, &clocal_size));
2070: PetscCall(MatCreateSubMatrix_MPIBAIJ_Private(A, crowp, lcolp, clocal_size, MAT_INITIAL_MATRIX, B, PETSC_FALSE));
2071: /* clean up */
2072: if (pcomm != comm) PetscCall(ISDestroy(&crowp));
2073: if (size > 1) PetscCall(ISDestroy(&lcolp));
2074: PetscFunctionReturn(PETSC_SUCCESS);
2075: }
2077: static PetscErrorCode MatGetGhosts_MPIBAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
2078: {
2079: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
2080: Mat_SeqBAIJ *B = (Mat_SeqBAIJ *)baij->B->data;
2082: PetscFunctionBegin;
2083: if (nghosts) *nghosts = B->nbs;
2084: if (ghosts) *ghosts = baij->garray;
2085: PetscFunctionReturn(PETSC_SUCCESS);
2086: }
2088: static PetscErrorCode MatGetSeqNonzeroStructure_MPIBAIJ(Mat A, Mat *newmat)
2089: {
2090: Mat B;
2091: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
2092: Mat_SeqBAIJ *ad = (Mat_SeqBAIJ *)a->A->data, *bd = (Mat_SeqBAIJ *)a->B->data;
2093: Mat_SeqAIJ *b;
2094: PetscMPIInt size, rank, *recvcounts = NULL, *displs = NULL;
2095: PetscInt sendcount, i, *rstarts = A->rmap->range, n, cnt, j, bs = A->rmap->bs;
2096: PetscInt m, *garray = a->garray, *lens, *jsendbuf, *a_jsendbuf, *b_jsendbuf;
2098: PetscFunctionBegin;
2099: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
2100: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)A), &rank));
2102: /* Tell every processor the number of nonzeros per row */
2103: PetscCall(PetscMalloc1(A->rmap->N / bs, &lens));
2104: for (i = A->rmap->rstart / bs; i < A->rmap->rend / bs; i++) lens[i] = ad->i[i - A->rmap->rstart / bs + 1] - ad->i[i - A->rmap->rstart / bs] + bd->i[i - A->rmap->rstart / bs + 1] - bd->i[i - A->rmap->rstart / bs];
2105: PetscCall(PetscMalloc1(2 * size, &recvcounts));
2106: displs = recvcounts + size;
2107: for (i = 0; i < size; i++) {
2108: PetscCall(PetscMPIIntCast(A->rmap->range[i + 1] / bs - A->rmap->range[i] / bs, &recvcounts[i]));
2109: PetscCall(PetscMPIIntCast(A->rmap->range[i] / bs, &displs[i]));
2110: }
2111: PetscCallMPI(MPI_Allgatherv(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL, lens, recvcounts, displs, MPIU_INT, PetscObjectComm((PetscObject)A)));
2112: /* Create the sequential matrix of the same type as the local block diagonal */
2113: PetscCall(MatCreate(PETSC_COMM_SELF, &B));
2114: PetscCall(MatSetSizes(B, A->rmap->N / bs, A->cmap->N / bs, PETSC_DETERMINE, PETSC_DETERMINE));
2115: PetscCall(MatSetType(B, MATSEQAIJ));
2116: PetscCall(MatSeqAIJSetPreallocation(B, 0, lens));
2117: b = (Mat_SeqAIJ *)B->data;
2119: /* Copy my part of matrix column indices over */
2120: sendcount = ad->nz + bd->nz;
2121: jsendbuf = b->j + b->i[rstarts[rank] / bs];
2122: a_jsendbuf = ad->j;
2123: b_jsendbuf = bd->j;
2124: n = A->rmap->rend / bs - A->rmap->rstart / bs;
2125: cnt = 0;
2126: for (i = 0; i < n; i++) {
2127: /* put in lower diagonal portion */
2128: m = bd->i[i + 1] - bd->i[i];
2129: while (m > 0) {
2130: /* is it above diagonal (in bd (compressed) numbering) */
2131: if (garray[*b_jsendbuf] > A->rmap->rstart / bs + i) break;
2132: jsendbuf[cnt++] = garray[*b_jsendbuf++];
2133: m--;
2134: }
2136: /* put in diagonal portion */
2137: for (j = ad->i[i]; j < ad->i[i + 1]; j++) jsendbuf[cnt++] = A->rmap->rstart / bs + *a_jsendbuf++;
2139: /* put in upper diagonal portion */
2140: while (m-- > 0) jsendbuf[cnt++] = garray[*b_jsendbuf++];
2141: }
2142: PetscCheck(cnt == sendcount, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Corrupted PETSc matrix: nz given %" PetscInt_FMT " actual nz %" PetscInt_FMT, sendcount, cnt);
2144: /* Gather all column indices to all processors */
2145: for (i = 0; i < size; i++) {
2146: recvcounts[i] = 0;
2147: for (j = A->rmap->range[i] / bs; j < A->rmap->range[i + 1] / bs; j++) recvcounts[i] += lens[j];
2148: }
2149: displs[0] = 0;
2150: for (i = 1; i < size; i++) displs[i] = displs[i - 1] + recvcounts[i - 1];
2151: PetscCallMPI(MPI_Allgatherv(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL, b->j, recvcounts, displs, MPIU_INT, PetscObjectComm((PetscObject)A)));
2152: /* Assemble the matrix into usable form (note numerical values not yet set) */
2153: /* set the b->ilen (length of each row) values */
2154: PetscCall(PetscArraycpy(b->ilen, lens, A->rmap->N / bs));
2155: /* set the b->i indices */
2156: b->i[0] = 0;
2157: for (i = 1; i <= A->rmap->N / bs; i++) b->i[i] = b->i[i - 1] + lens[i - 1];
2158: PetscCall(PetscFree(lens));
2159: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
2160: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
2161: PetscCall(PetscFree(recvcounts));
2163: PetscCall(MatPropagateSymmetryOptions(A, B));
2164: *newmat = B;
2165: PetscFunctionReturn(PETSC_SUCCESS);
2166: }
2168: static PetscErrorCode MatSOR_MPIBAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
2169: {
2170: Mat_MPIBAIJ *mat = (Mat_MPIBAIJ *)matin->data;
2171: Vec bb1 = NULL;
2173: PetscFunctionBegin;
2174: if (flag == SOR_APPLY_UPPER) {
2175: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
2176: PetscFunctionReturn(PETSC_SUCCESS);
2177: }
2179: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS) PetscCall(VecDuplicate(bb, &bb1));
2181: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
2182: if (flag & SOR_ZERO_INITIAL_GUESS) {
2183: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
2184: its--;
2185: }
2187: while (its--) {
2188: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
2189: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
2191: /* update rhs: bb1 = bb - B*x */
2192: PetscCall(VecScale(mat->lvec, -1.0));
2193: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
2195: /* local sweep */
2196: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx));
2197: }
2198: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
2199: if (flag & SOR_ZERO_INITIAL_GUESS) {
2200: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
2201: its--;
2202: }
2203: while (its--) {
2204: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
2205: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
2207: /* update rhs: bb1 = bb - B*x */
2208: PetscCall(VecScale(mat->lvec, -1.0));
2209: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
2211: /* local sweep */
2212: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx));
2213: }
2214: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
2215: if (flag & SOR_ZERO_INITIAL_GUESS) {
2216: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
2217: its--;
2218: }
2219: while (its--) {
2220: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
2221: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
2223: /* update rhs: bb1 = bb - B*x */
2224: PetscCall(VecScale(mat->lvec, -1.0));
2225: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
2227: /* local sweep */
2228: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx));
2229: }
2230: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel version of SOR requested not supported");
2232: PetscCall(VecDestroy(&bb1));
2233: PetscFunctionReturn(PETSC_SUCCESS);
2234: }
2236: static PetscErrorCode MatGetColumnReductions_MPIBAIJ(Mat A, PetscInt type, PetscReal *reductions)
2237: {
2238: Mat_MPIBAIJ *aij = (Mat_MPIBAIJ *)A->data;
2239: PetscInt m, N, i, *garray = aij->garray;
2240: PetscInt ib, jb, bs = A->rmap->bs;
2241: Mat_SeqBAIJ *a_aij = (Mat_SeqBAIJ *)aij->A->data;
2242: MatScalar *a_val = a_aij->a;
2243: Mat_SeqBAIJ *b_aij = (Mat_SeqBAIJ *)aij->B->data;
2244: MatScalar *b_val = b_aij->a;
2245: PetscReal *work;
2247: PetscFunctionBegin;
2248: PetscCall(MatGetSize(A, &m, &N));
2249: PetscCall(PetscCalloc1(N, &work));
2250: if (type == NORM_2) {
2251: for (i = a_aij->i[0]; i < a_aij->i[aij->A->rmap->n / bs]; i++) {
2252: for (jb = 0; jb < bs; jb++) {
2253: for (ib = 0; ib < bs; ib++) {
2254: work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val * *a_val);
2255: a_val++;
2256: }
2257: }
2258: }
2259: for (i = b_aij->i[0]; i < b_aij->i[aij->B->rmap->n / bs]; i++) {
2260: for (jb = 0; jb < bs; jb++) {
2261: for (ib = 0; ib < bs; ib++) {
2262: work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val * *b_val);
2263: b_val++;
2264: }
2265: }
2266: }
2267: } else if (type == NORM_1) {
2268: for (i = a_aij->i[0]; i < a_aij->i[aij->A->rmap->n / bs]; i++) {
2269: for (jb = 0; jb < bs; jb++) {
2270: for (ib = 0; ib < bs; ib++) {
2271: work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val);
2272: a_val++;
2273: }
2274: }
2275: }
2276: for (i = b_aij->i[0]; i < b_aij->i[aij->B->rmap->n / bs]; i++) {
2277: for (jb = 0; jb < bs; jb++) {
2278: for (ib = 0; ib < bs; ib++) {
2279: work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val);
2280: b_val++;
2281: }
2282: }
2283: }
2284: } else if (type == NORM_INFINITY) {
2285: for (i = a_aij->i[0]; i < a_aij->i[aij->A->rmap->n / bs]; i++) {
2286: for (jb = 0; jb < bs; jb++) {
2287: for (ib = 0; ib < bs; ib++) {
2288: PetscInt col = A->cmap->rstart + a_aij->j[i] * bs + jb;
2289: work[col] = PetscMax(PetscAbsScalar(*a_val), work[col]);
2290: a_val++;
2291: }
2292: }
2293: }
2294: for (i = b_aij->i[0]; i < b_aij->i[aij->B->rmap->n / bs]; i++) {
2295: for (jb = 0; jb < bs; jb++) {
2296: for (ib = 0; ib < bs; ib++) {
2297: PetscInt col = garray[b_aij->j[i]] * bs + jb;
2298: work[col] = PetscMax(PetscAbsScalar(*b_val), work[col]);
2299: b_val++;
2300: }
2301: }
2302: }
2303: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
2304: for (i = a_aij->i[0]; i < a_aij->i[aij->A->rmap->n / bs]; i++) {
2305: for (jb = 0; jb < bs; jb++) {
2306: for (ib = 0; ib < bs; ib++) {
2307: work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscRealPart(*a_val);
2308: a_val++;
2309: }
2310: }
2311: }
2312: for (i = b_aij->i[0]; i < b_aij->i[aij->B->rmap->n / bs]; i++) {
2313: for (jb = 0; jb < bs; jb++) {
2314: for (ib = 0; ib < bs; ib++) {
2315: work[garray[b_aij->j[i]] * bs + jb] += PetscRealPart(*b_val);
2316: b_val++;
2317: }
2318: }
2319: }
2320: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
2321: for (i = a_aij->i[0]; i < a_aij->i[aij->A->rmap->n / bs]; i++) {
2322: for (jb = 0; jb < bs; jb++) {
2323: for (ib = 0; ib < bs; ib++) {
2324: work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscImaginaryPart(*a_val);
2325: a_val++;
2326: }
2327: }
2328: }
2329: for (i = b_aij->i[0]; i < b_aij->i[aij->B->rmap->n / bs]; i++) {
2330: for (jb = 0; jb < bs; jb++) {
2331: for (ib = 0; ib < bs; ib++) {
2332: work[garray[b_aij->j[i]] * bs + jb] += PetscImaginaryPart(*b_val);
2333: b_val++;
2334: }
2335: }
2336: }
2337: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
2338: if (type == NORM_INFINITY) {
2339: PetscCallMPI(MPIU_Allreduce(work, reductions, N, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)A)));
2340: } else {
2341: PetscCallMPI(MPIU_Allreduce(work, reductions, N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
2342: }
2343: PetscCall(PetscFree(work));
2344: if (type == NORM_2) {
2345: for (i = 0; i < N; i++) reductions[i] = PetscSqrtReal(reductions[i]);
2346: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
2347: for (i = 0; i < N; i++) reductions[i] /= m;
2348: }
2349: PetscFunctionReturn(PETSC_SUCCESS);
2350: }
2352: static PetscErrorCode MatInvertBlockDiagonal_MPIBAIJ(Mat A, const PetscScalar **values)
2353: {
2354: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
2356: PetscFunctionBegin;
2357: PetscCall(MatInvertBlockDiagonal(a->A, values));
2358: A->factorerrortype = a->A->factorerrortype;
2359: A->factorerror_zeropivot_value = a->A->factorerror_zeropivot_value;
2360: A->factorerror_zeropivot_row = a->A->factorerror_zeropivot_row;
2361: PetscFunctionReturn(PETSC_SUCCESS);
2362: }
2364: static PetscErrorCode MatShift_MPIBAIJ(Mat Y, PetscScalar a)
2365: {
2366: Mat_MPIBAIJ *maij = (Mat_MPIBAIJ *)Y->data;
2367: Mat_SeqBAIJ *aij = (Mat_SeqBAIJ *)maij->A->data;
2369: PetscFunctionBegin;
2370: if (!Y->preallocated) {
2371: PetscCall(MatMPIBAIJSetPreallocation(Y, Y->rmap->bs, 1, NULL, 0, NULL));
2372: } else if (!aij->nz) {
2373: PetscInt nonew = aij->nonew;
2374: PetscCall(MatSeqBAIJSetPreallocation(maij->A, Y->rmap->bs, 1, NULL));
2375: aij->nonew = nonew;
2376: }
2377: PetscCall(MatShift_Basic(Y, a));
2378: PetscFunctionReturn(PETSC_SUCCESS);
2379: }
2381: static PetscErrorCode MatGetDiagonalBlock_MPIBAIJ(Mat A, Mat *a)
2382: {
2383: PetscFunctionBegin;
2384: *a = ((Mat_MPIBAIJ *)A->data)->A;
2385: PetscFunctionReturn(PETSC_SUCCESS);
2386: }
2388: static PetscErrorCode MatEliminateZeros_MPIBAIJ(Mat A, PetscBool keep)
2389: {
2390: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
2392: PetscFunctionBegin;
2393: PetscCall(MatEliminateZeros_SeqBAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2394: PetscCall(MatEliminateZeros_SeqBAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2395: PetscFunctionReturn(PETSC_SUCCESS);
2396: }
2398: static struct _MatOps MatOps_Values = {MatSetValues_MPIBAIJ,
2399: MatGetRow_MPIBAIJ,
2400: MatRestoreRow_MPIBAIJ,
2401: MatMult_MPIBAIJ,
2402: /* 4*/ MatMultAdd_MPIBAIJ,
2403: MatMultTranspose_MPIBAIJ,
2404: MatMultTransposeAdd_MPIBAIJ,
2405: NULL,
2406: NULL,
2407: NULL,
2408: /*10*/ NULL,
2409: NULL,
2410: NULL,
2411: MatSOR_MPIBAIJ,
2412: MatTranspose_MPIBAIJ,
2413: /*15*/ MatGetInfo_MPIBAIJ,
2414: MatEqual_MPIBAIJ,
2415: MatGetDiagonal_MPIBAIJ,
2416: MatDiagonalScale_MPIBAIJ,
2417: MatNorm_MPIBAIJ,
2418: /*20*/ MatAssemblyBegin_MPIBAIJ,
2419: MatAssemblyEnd_MPIBAIJ,
2420: MatSetOption_MPIBAIJ,
2421: MatZeroEntries_MPIBAIJ,
2422: /*24*/ MatZeroRows_MPIBAIJ,
2423: NULL,
2424: NULL,
2425: NULL,
2426: NULL,
2427: /*29*/ MatSetUp_MPI_Hash,
2428: NULL,
2429: NULL,
2430: MatGetDiagonalBlock_MPIBAIJ,
2431: NULL,
2432: /*34*/ MatDuplicate_MPIBAIJ,
2433: NULL,
2434: NULL,
2435: NULL,
2436: NULL,
2437: /*39*/ MatAXPY_MPIBAIJ,
2438: MatCreateSubMatrices_MPIBAIJ,
2439: MatIncreaseOverlap_MPIBAIJ,
2440: MatGetValues_MPIBAIJ,
2441: MatCopy_MPIBAIJ,
2442: /*44*/ NULL,
2443: MatScale_MPIBAIJ,
2444: MatShift_MPIBAIJ,
2445: NULL,
2446: MatZeroRowsColumns_MPIBAIJ,
2447: /*49*/ NULL,
2448: NULL,
2449: NULL,
2450: NULL,
2451: NULL,
2452: /*54*/ MatFDColoringCreate_MPIXAIJ,
2453: NULL,
2454: MatSetUnfactored_MPIBAIJ,
2455: MatPermute_MPIBAIJ,
2456: MatSetValuesBlocked_MPIBAIJ,
2457: /*59*/ MatCreateSubMatrix_MPIBAIJ,
2458: MatDestroy_MPIBAIJ,
2459: MatView_MPIBAIJ,
2460: NULL,
2461: NULL,
2462: /*64*/ NULL,
2463: NULL,
2464: NULL,
2465: NULL,
2466: MatGetRowMaxAbs_MPIBAIJ,
2467: /*69*/ NULL,
2468: NULL,
2469: NULL,
2470: MatFDColoringApply_BAIJ,
2471: NULL,
2472: /*74*/ NULL,
2473: NULL,
2474: NULL,
2475: NULL,
2476: MatLoad_MPIBAIJ,
2477: /*79*/ NULL,
2478: NULL,
2479: NULL,
2480: NULL,
2481: NULL,
2482: /*84*/ NULL,
2483: NULL,
2484: NULL,
2485: NULL,
2486: NULL,
2487: /*89*/ NULL,
2488: NULL,
2489: NULL,
2490: NULL,
2491: MatConjugate_MPIBAIJ,
2492: /*94*/ NULL,
2493: NULL,
2494: MatRealPart_MPIBAIJ,
2495: MatImaginaryPart_MPIBAIJ,
2496: NULL,
2497: /*99*/ NULL,
2498: NULL,
2499: NULL,
2500: NULL,
2501: NULL,
2502: /*104*/ MatGetSeqNonzeroStructure_MPIBAIJ,
2503: NULL,
2504: MatGetGhosts_MPIBAIJ,
2505: NULL,
2506: NULL,
2507: /*109*/ NULL,
2508: NULL,
2509: NULL,
2510: NULL,
2511: MatGetMultiProcBlock_MPIBAIJ,
2512: /*114*/ NULL,
2513: MatGetColumnReductions_MPIBAIJ,
2514: MatInvertBlockDiagonal_MPIBAIJ,
2515: NULL,
2516: NULL,
2517: /*119*/ NULL,
2518: NULL,
2519: NULL,
2520: NULL,
2521: NULL,
2522: /*124*/ NULL,
2523: NULL,
2524: MatSetBlockSizes_Default,
2525: NULL,
2526: MatFDColoringSetUp_MPIXAIJ,
2527: /*129*/ NULL,
2528: MatCreateMPIMatConcatenateSeqMat_MPIBAIJ,
2529: NULL,
2530: NULL,
2531: NULL,
2532: /*134*/ NULL,
2533: NULL,
2534: MatEliminateZeros_MPIBAIJ,
2535: MatGetRowSumAbs_MPIBAIJ,
2536: NULL,
2537: /*139*/ NULL,
2538: NULL,
2539: MatCopyHashToXAIJ_MPI_Hash,
2540: NULL,
2541: NULL,
2542: /*144*/ MatADot_Default,
2543: MatANorm_Default,
2544: NULL,
2545: NULL};
2547: PETSC_INTERN PetscErrorCode MatConvert_MPIBAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
2548: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
2550: static PetscErrorCode MatMPIBAIJSetPreallocationCSR_MPIBAIJ(Mat B, PetscInt bs, const PetscInt ii[], const PetscInt jj[], const PetscScalar V[])
2551: {
2552: PetscInt m, rstart, cstart, cend;
2553: PetscInt i, j, dlen, olen, nz, nz_max = 0, *d_nnz = NULL, *o_nnz = NULL;
2554: const PetscInt *JJ = NULL;
2555: PetscScalar *values = NULL;
2556: PetscBool roworiented = ((Mat_MPIBAIJ *)B->data)->roworiented;
2557: PetscBool nooffprocentries;
2559: PetscFunctionBegin;
2560: PetscCall(PetscLayoutSetBlockSize(B->rmap, bs));
2561: PetscCall(PetscLayoutSetBlockSize(B->cmap, bs));
2562: PetscCall(PetscLayoutSetUp(B->rmap));
2563: PetscCall(PetscLayoutSetUp(B->cmap));
2564: PetscCall(PetscLayoutGetBlockSize(B->rmap, &bs));
2565: m = B->rmap->n / bs;
2566: rstart = B->rmap->rstart / bs;
2567: cstart = B->cmap->rstart / bs;
2568: cend = B->cmap->rend / bs;
2570: PetscCheck(!ii[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "ii[0] must be 0 but it is %" PetscInt_FMT, ii[0]);
2571: PetscCall(PetscMalloc2(m, &d_nnz, m, &o_nnz));
2572: for (i = 0; i < m; i++) {
2573: nz = ii[i + 1] - ii[i];
2574: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative number of columns %" PetscInt_FMT, i, nz);
2575: nz_max = PetscMax(nz_max, nz);
2576: dlen = 0;
2577: olen = 0;
2578: JJ = jj + ii[i];
2579: for (j = 0; j < nz; j++) {
2580: if (*JJ < cstart || *JJ >= cend) olen++;
2581: else dlen++;
2582: JJ++;
2583: }
2584: d_nnz[i] = dlen;
2585: o_nnz[i] = olen;
2586: }
2587: PetscCall(MatMPIBAIJSetPreallocation(B, bs, 0, d_nnz, 0, o_nnz));
2588: PetscCall(PetscFree2(d_nnz, o_nnz));
2590: values = (PetscScalar *)V;
2591: if (!values) PetscCall(PetscCalloc1(bs * bs * nz_max, &values));
2592: for (i = 0; i < m; i++) {
2593: PetscInt row = i + rstart;
2594: PetscInt ncols = ii[i + 1] - ii[i];
2595: const PetscInt *icols = jj + ii[i];
2596: if (bs == 1 || !roworiented) { /* block ordering matches the non-nested layout of MatSetValues so we can insert entire rows */
2597: const PetscScalar *svals = values + (V ? (bs * bs * ii[i]) : 0);
2598: PetscCall(MatSetValuesBlocked_MPIBAIJ(B, 1, &row, ncols, icols, svals, INSERT_VALUES));
2599: } else { /* block ordering does not match so we can only insert one block at a time. */
2600: PetscInt j;
2601: for (j = 0; j < ncols; j++) {
2602: const PetscScalar *svals = values + (V ? (bs * bs * (ii[i] + j)) : 0);
2603: PetscCall(MatSetValuesBlocked_MPIBAIJ(B, 1, &row, 1, &icols[j], svals, INSERT_VALUES));
2604: }
2605: }
2606: }
2608: if (!V) PetscCall(PetscFree(values));
2609: nooffprocentries = B->nooffprocentries;
2610: B->nooffprocentries = PETSC_TRUE;
2611: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
2612: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
2613: B->nooffprocentries = nooffprocentries;
2615: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
2616: PetscFunctionReturn(PETSC_SUCCESS);
2617: }
2619: /*@C
2620: MatMPIBAIJSetPreallocationCSR - Creates a sparse parallel matrix in `MATBAIJ` format using the given nonzero structure and (optional) numerical values
2622: Collective
2624: Input Parameters:
2625: + B - the matrix
2626: . bs - the block size
2627: . i - the indices into `j` for the start of each local row (starts with zero)
2628: . j - the column indices for each local row (starts with zero) these must be sorted for each row
2629: - v - optional values in the matrix, use `NULL` if not provided
2631: Level: advanced
2633: Notes:
2634: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
2635: thus you CANNOT change the matrix entries by changing the values of `v` after you have
2636: called this routine.
2638: The order of the entries in values is specified by the `MatOption` `MAT_ROW_ORIENTED`. For example, C programs
2639: may want to use the default `MAT_ROW_ORIENTED` with value `PETSC_TRUE` and use an array v[nnz][bs][bs] where the second index is
2640: over rows within a block and the last index is over columns within a block row. Fortran programs will likely set
2641: `MAT_ROW_ORIENTED` with value `PETSC_FALSE` and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a
2642: block column and the second index is over columns within a block.
2644: Though this routine has Preallocation() in the name it also sets the exact nonzero locations of the matrix entries and usually the numerical values as well
2646: .seealso: `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIBAIJSetPreallocation()`, `MatCreateAIJ()`, `MATMPIAIJ`, `MatCreateMPIBAIJWithArrays()`, `MATMPIBAIJ`
2647: @*/
2648: PetscErrorCode MatMPIBAIJSetPreallocationCSR(Mat B, PetscInt bs, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
2649: {
2650: PetscFunctionBegin;
2654: PetscTryMethod(B, "MatMPIBAIJSetPreallocationCSR_C", (Mat, PetscInt, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, bs, i, j, v));
2655: PetscFunctionReturn(PETSC_SUCCESS);
2656: }
2658: PetscErrorCode MatMPIBAIJSetPreallocation_MPIBAIJ(Mat B, PetscInt bs, PetscInt d_nz, const PetscInt *d_nnz, PetscInt o_nz, const PetscInt *o_nnz)
2659: {
2660: Mat_MPIBAIJ *b = (Mat_MPIBAIJ *)B->data;
2661: PetscInt i;
2662: PetscMPIInt size;
2664: PetscFunctionBegin;
2665: if (B->hash_active) {
2666: B->ops[0] = b->cops;
2667: B->hash_active = PETSC_FALSE;
2668: }
2669: if (!B->preallocated) PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), bs, &B->bstash));
2670: PetscCall(MatSetBlockSize(B, bs));
2671: PetscCall(PetscLayoutSetUp(B->rmap));
2672: PetscCall(PetscLayoutSetUp(B->cmap));
2673: PetscCall(PetscLayoutGetBlockSize(B->rmap, &bs));
2675: if (d_nnz) {
2676: for (i = 0; i < B->rmap->n / bs; i++) PetscCheck(d_nnz[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "d_nnz cannot be less than -1: local row %" PetscInt_FMT " value %" PetscInt_FMT, i, d_nnz[i]);
2677: }
2678: if (o_nnz) {
2679: for (i = 0; i < B->rmap->n / bs; i++) PetscCheck(o_nnz[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "o_nnz cannot be less than -1: local row %" PetscInt_FMT " value %" PetscInt_FMT, i, o_nnz[i]);
2680: }
2682: b->bs2 = bs * bs;
2683: b->mbs = B->rmap->n / bs;
2684: b->nbs = B->cmap->n / bs;
2685: b->Mbs = B->rmap->N / bs;
2686: b->Nbs = B->cmap->N / bs;
2688: for (i = 0; i <= b->size; i++) b->rangebs[i] = B->rmap->range[i] / bs;
2689: b->rstartbs = B->rmap->rstart / bs;
2690: b->rendbs = B->rmap->rend / bs;
2691: b->cstartbs = B->cmap->rstart / bs;
2692: b->cendbs = B->cmap->rend / bs;
2694: #if defined(PETSC_USE_CTABLE)
2695: PetscCall(PetscHMapIDestroy(&b->colmap));
2696: #else
2697: PetscCall(PetscFree(b->colmap));
2698: #endif
2699: PetscCall(PetscFree(b->garray));
2700: PetscCall(VecDestroy(&b->lvec));
2701: PetscCall(VecScatterDestroy(&b->Mvctx));
2703: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2705: MatSeqXAIJGetOptions_Private(b->B);
2706: PetscCall(MatDestroy(&b->B));
2707: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2708: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2709: PetscCall(MatSetType(b->B, MATSEQBAIJ));
2710: MatSeqXAIJRestoreOptions_Private(b->B);
2712: MatSeqXAIJGetOptions_Private(b->A);
2713: PetscCall(MatDestroy(&b->A));
2714: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2715: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2716: PetscCall(MatSetType(b->A, MATSEQBAIJ));
2717: MatSeqXAIJRestoreOptions_Private(b->A);
2719: PetscCall(MatSeqBAIJSetPreallocation(b->A, bs, d_nz, d_nnz));
2720: PetscCall(MatSeqBAIJSetPreallocation(b->B, bs, o_nz, o_nnz));
2721: B->preallocated = PETSC_TRUE;
2722: B->was_assembled = PETSC_FALSE;
2723: B->assembled = PETSC_FALSE;
2724: PetscFunctionReturn(PETSC_SUCCESS);
2725: }
2727: extern PetscErrorCode MatDiagonalScaleLocal_MPIBAIJ(Mat, Vec);
2728: extern PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat, PetscReal);
2730: PETSC_INTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAdj(Mat B, MatType newtype, MatReuse reuse, Mat *adj)
2731: {
2732: Mat_MPIBAIJ *b = (Mat_MPIBAIJ *)B->data;
2733: Mat_SeqBAIJ *d = (Mat_SeqBAIJ *)b->A->data, *o = (Mat_SeqBAIJ *)b->B->data;
2734: PetscInt M = B->rmap->n / B->rmap->bs, i, *ii, *jj, cnt, j, k, rstart = B->rmap->rstart / B->rmap->bs;
2735: const PetscInt *id = d->i, *jd = d->j, *io = o->i, *jo = o->j, *garray = b->garray;
2737: PetscFunctionBegin;
2738: PetscCall(PetscMalloc1(M + 1, &ii));
2739: ii[0] = 0;
2740: for (i = 0; i < M; i++) {
2741: PetscCheck((id[i + 1] - id[i]) >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Indices wrong %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT, i, id[i], id[i + 1]);
2742: PetscCheck((io[i + 1] - io[i]) >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Indices wrong %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT, i, io[i], io[i + 1]);
2743: ii[i + 1] = ii[i] + id[i + 1] - id[i] + io[i + 1] - io[i];
2744: /* remove one from count of matrix has diagonal */
2745: for (j = id[i]; j < id[i + 1]; j++) {
2746: if (jd[j] == i) {
2747: ii[i + 1]--;
2748: break;
2749: }
2750: }
2751: }
2752: PetscCall(PetscMalloc1(ii[M], &jj));
2753: cnt = 0;
2754: for (i = 0; i < M; i++) {
2755: for (j = io[i]; j < io[i + 1]; j++) {
2756: if (garray[jo[j]] > rstart) break;
2757: jj[cnt++] = garray[jo[j]];
2758: }
2759: for (k = id[i]; k < id[i + 1]; k++) {
2760: if (jd[k] != i) jj[cnt++] = rstart + jd[k];
2761: }
2762: for (; j < io[i + 1]; j++) jj[cnt++] = garray[jo[j]];
2763: }
2764: PetscCall(MatCreateMPIAdj(PetscObjectComm((PetscObject)B), M, B->cmap->N / B->rmap->bs, ii, jj, NULL, adj));
2765: PetscFunctionReturn(PETSC_SUCCESS);
2766: }
2768: #include <../src/mat/impls/aij/mpi/mpiaij.h>
2770: PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat, MatType, MatReuse, Mat *);
2772: PETSC_INTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAIJ(Mat A, MatType newtype, MatReuse reuse, Mat *newmat)
2773: {
2774: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
2775: Mat_MPIAIJ *b;
2776: Mat B;
2778: PetscFunctionBegin;
2779: PetscCheck(A->assembled, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Matrix must be assembled");
2781: if (reuse == MAT_REUSE_MATRIX) {
2782: B = *newmat;
2783: } else {
2784: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
2785: PetscCall(MatSetType(B, MATMPIAIJ));
2786: PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
2787: PetscCall(MatSetBlockSizes(B, A->rmap->bs, A->cmap->bs));
2788: PetscCall(MatSeqAIJSetPreallocation(B, 0, NULL));
2789: PetscCall(MatMPIAIJSetPreallocation(B, 0, NULL, 0, NULL));
2790: }
2791: b = (Mat_MPIAIJ *)B->data;
2793: if (reuse == MAT_REUSE_MATRIX) {
2794: PetscCall(MatConvert_SeqBAIJ_SeqAIJ(a->A, MATSEQAIJ, MAT_REUSE_MATRIX, &b->A));
2795: PetscCall(MatConvert_SeqBAIJ_SeqAIJ(a->B, MATSEQAIJ, MAT_REUSE_MATRIX, &b->B));
2796: } else {
2797: PetscInt *garray = a->garray;
2798: Mat_SeqAIJ *bB;
2799: PetscInt bs, nnz;
2800: PetscCall(MatDestroy(&b->A));
2801: PetscCall(MatDestroy(&b->B));
2802: /* just clear out the data structure */
2803: PetscCall(MatDisAssemble_MPIAIJ(B, PETSC_FALSE));
2804: PetscCall(MatConvert_SeqBAIJ_SeqAIJ(a->A, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->A));
2805: PetscCall(MatConvert_SeqBAIJ_SeqAIJ(a->B, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->B));
2807: /* Global numbering for b->B columns */
2808: bB = (Mat_SeqAIJ *)b->B->data;
2809: bs = A->rmap->bs;
2810: nnz = bB->i[A->rmap->n];
2811: for (PetscInt k = 0; k < nnz; k++) {
2812: PetscInt bj = bB->j[k] / bs;
2813: PetscInt br = bB->j[k] % bs;
2814: bB->j[k] = garray[bj] * bs + br;
2815: }
2816: }
2817: PetscCall(MatSetOption(B, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2818: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
2819: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
2820: PetscCall(MatSetOption(B, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
2822: if (reuse == MAT_INPLACE_MATRIX) {
2823: PetscCall(MatHeaderReplace(A, &B));
2824: } else {
2825: *newmat = B;
2826: }
2827: PetscFunctionReturn(PETSC_SUCCESS);
2828: }
2830: /*MC
2831: MATMPIBAIJ - MATMPIBAIJ = "mpibaij" - A matrix type to be used for distributed block sparse matrices.
2833: Options Database Keys:
2834: + -mat_type mpibaij - sets the matrix type to `MATMPIBAIJ` during a call to `MatSetFromOptions()`
2835: . -mat_block_size bs - set the blocksize used to store the matrix
2836: . -mat_baij_mult_version version - indicate the version of the matrix-vector product to use (0 often indicates using BLAS)
2837: - -mat_use_hash_table fact - set hash table factor
2839: Level: beginner
2841: Note:
2842: `MatSetOption(A, MAT_STRUCTURE_ONLY, PETSC_TRUE)` may be called for this matrix type. In this no
2843: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
2845: .seealso: `Mat`, `MATBAIJ`, `MATSEQBAIJ`, `MatCreateBAIJ`
2846: M*/
2848: PETSC_INTERN PetscErrorCode MatConvert_MPIBAIJ_MPIBSTRM(Mat, MatType, MatReuse, Mat *);
2850: PETSC_EXTERN PetscErrorCode MatCreate_MPIBAIJ(Mat B)
2851: {
2852: Mat_MPIBAIJ *b;
2853: PetscBool flg = PETSC_FALSE;
2855: PetscFunctionBegin;
2856: PetscCall(PetscNew(&b));
2857: B->data = (void *)b;
2858: B->ops[0] = MatOps_Values;
2859: B->assembled = PETSC_FALSE;
2861: B->insertmode = NOT_SET_VALUES;
2862: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
2863: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &b->size));
2865: /* build local table of row and column ownerships */
2866: PetscCall(PetscMalloc1(b->size + 1, &b->rangebs));
2868: /* build cache for off array entries formed */
2869: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
2871: b->donotstash = PETSC_FALSE;
2872: b->colmap = NULL;
2873: b->garray = NULL;
2874: b->roworiented = PETSC_TRUE;
2876: /* stuff used in block assembly */
2877: b->barray = NULL;
2879: /* stuff used for matrix vector multiply */
2880: b->lvec = NULL;
2881: b->Mvctx = NULL;
2883: /* stuff for MatGetRow() */
2884: b->rowindices = NULL;
2885: b->rowvalues = NULL;
2886: b->getrowactive = PETSC_FALSE;
2888: /* hash table stuff */
2889: b->ht = NULL;
2890: b->hd = NULL;
2891: b->ht_size = 0;
2892: b->ht_flag = PETSC_FALSE;
2893: b->ht_fact = 0;
2894: b->ht_total_ct = 0;
2895: b->ht_insert_ct = 0;
2897: /* stuff for MatCreateSubMatrices_MPIBAIJ_local() */
2898: b->ijonly = PETSC_FALSE;
2900: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpibaij_mpiadj_C", MatConvert_MPIBAIJ_MPIAdj));
2901: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpibaij_mpiaij_C", MatConvert_MPIBAIJ_MPIAIJ));
2902: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpibaij_mpisbaij_C", MatConvert_MPIBAIJ_MPISBAIJ));
2903: #if defined(PETSC_HAVE_HYPRE)
2904: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpibaij_hypre_C", MatConvert_AIJ_HYPRE));
2905: #endif
2906: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIBAIJ));
2907: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIBAIJ));
2908: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIBAIJSetPreallocation_C", MatMPIBAIJSetPreallocation_MPIBAIJ));
2909: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIBAIJSetPreallocationCSR_C", MatMPIBAIJSetPreallocationCSR_MPIBAIJ));
2910: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIBAIJ));
2911: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetHashTableFactor_C", MatSetHashTableFactor_MPIBAIJ));
2912: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpibaij_is_C", MatConvert_XAIJ_IS));
2913: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIBAIJ));
2915: PetscOptionsBegin(PetscObjectComm((PetscObject)B), NULL, "Options for loading MPIBAIJ matrix 1", "Mat");
2916: PetscCall(PetscOptionsName("-mat_use_hash_table", "Use hash table to save time in constructing matrix", "MatSetOption", &flg));
2917: if (flg) {
2918: PetscReal fact = 1.39;
2919: PetscCall(MatSetOption(B, MAT_USE_HASH_TABLE, PETSC_TRUE));
2920: PetscCall(PetscOptionsReal("-mat_use_hash_table", "Use hash table factor", "MatMPIBAIJSetHashTableFactor", fact, &fact, NULL));
2921: if (fact <= 1.0) fact = 1.39;
2922: PetscCall(MatMPIBAIJSetHashTableFactor(B, fact));
2923: PetscCall(PetscInfo(B, "Hash table Factor used %5.2g\n", (double)fact));
2924: }
2925: PetscOptionsEnd();
2926: PetscFunctionReturn(PETSC_SUCCESS);
2927: }
2929: // PetscClangLinter pragma disable: -fdoc-section-header-unknown
2930: /*MC
2931: MATBAIJ - MATBAIJ = "baij" - A matrix type to be used for block sparse matrices.
2933: This matrix type is identical to `MATSEQBAIJ` when constructed with a single process communicator,
2934: and `MATMPIBAIJ` otherwise.
2936: Options Database Keys:
2937: . -mat_type baij - sets the matrix type to `MATBAIJ` during a call to `MatSetFromOptions()`
2939: Level: beginner
2941: .seealso: `Mat`, `MatCreateBAIJ()`, `MATSEQBAIJ`, `MATMPIBAIJ`, `MatMPIBAIJSetPreallocation()`, `MatMPIBAIJSetPreallocationCSR()`
2942: M*/
2944: /*@
2945: MatMPIBAIJSetPreallocation - Allocates memory for a sparse parallel matrix in `MATMPIBAIJ` format
2946: (block compressed row).
2948: Collective
2950: Input Parameters:
2951: + B - the matrix
2952: . bs - size of block, the blocks are ALWAYS square. One can use `MatSetBlockSizes()` to set a different row and column blocksize but the row
2953: blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with `MatCreateVecs()`
2954: . d_nz - number of block nonzeros per block row in diagonal portion of local
2955: submatrix (same for all local rows)
2956: . d_nnz - array containing the number of block nonzeros in the various block rows
2957: of the in diagonal portion of the local (possibly different for each block
2958: row) or `NULL`. If you plan to factor the matrix you must leave room for the diagonal entry and
2959: set it even if it is zero.
2960: . o_nz - number of block nonzeros per block row in the off-diagonal portion of local
2961: submatrix (same for all local rows).
2962: - o_nnz - array containing the number of nonzeros in the various block rows of the
2963: off-diagonal portion of the local submatrix (possibly different for
2964: each block row) or `NULL`.
2966: If the *_nnz parameter is given then the *_nz parameter is ignored
2968: Options Database Keys:
2969: + -mat_block_size - size of the blocks to use
2970: - -mat_use_hash_table fact - set hash table factor
2972: Level: intermediate
2974: Notes:
2975: For good matrix assembly performance
2976: the user should preallocate the matrix storage by setting the parameters
2977: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`). By setting these parameters accurately,
2978: performance can be increased by more than a factor of 50.
2980: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one processor
2981: than it must be used on all processors that share the object for that argument.
2983: Storage Information:
2984: For a square global matrix we define each processor's diagonal portion
2985: to be its local rows and the corresponding columns (a square submatrix);
2986: each processor's off-diagonal portion encompasses the remainder of the
2987: local matrix (a rectangular submatrix).
2989: The user can specify preallocated storage for the diagonal part of
2990: the local submatrix with either `d_nz` or `d_nnz` (not both). Set
2991: `d_nz` = `PETSC_DEFAULT` and `d_nnz` = `NULL` for PETSc to control dynamic
2992: memory allocation. Likewise, specify preallocated storage for the
2993: off-diagonal part of the local submatrix with `o_nz` or `o_nnz` (not both).
2995: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
2996: the figure below we depict these three local rows and all columns (0-11).
2998: .vb
2999: 0 1 2 3 4 5 6 7 8 9 10 11
3000: --------------------------
3001: row 3 |o o o d d d o o o o o o
3002: row 4 |o o o d d d o o o o o o
3003: row 5 |o o o d d d o o o o o o
3004: --------------------------
3005: .ve
3007: Thus, any entries in the d locations are stored in the d (diagonal)
3008: submatrix, and any entries in the o locations are stored in the
3009: o (off-diagonal) submatrix. Note that the d and the o submatrices are
3010: stored simply in the `MATSEQBAIJ` format for compressed row storage.
3012: Now `d_nz` should indicate the number of block nonzeros per row in the d matrix,
3013: and `o_nz` should indicate the number of block nonzeros per row in the o matrix.
3014: In general, for PDE problems in which most nonzeros are near the diagonal,
3015: one expects `d_nz` >> `o_nz`.
3017: You can call `MatGetInfo()` to get information on how effective the preallocation was;
3018: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3019: You can also run with the option `-info` and look for messages with the string
3020: malloc in them to see if additional memory allocation was needed.
3022: .seealso: `Mat`, `MATMPIBAIJ`, `MatCreate()`, `MatCreateSeqBAIJ()`, `MatSetValues()`, `MatCreateBAIJ()`, `MatMPIBAIJSetPreallocationCSR()`, `PetscSplitOwnership()`
3023: @*/
3024: PetscErrorCode MatMPIBAIJSetPreallocation(Mat B, PetscInt bs, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
3025: {
3026: PetscFunctionBegin;
3030: PetscTryMethod(B, "MatMPIBAIJSetPreallocation_C", (Mat, PetscInt, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, bs, d_nz, d_nnz, o_nz, o_nnz));
3031: PetscFunctionReturn(PETSC_SUCCESS);
3032: }
3034: // PetscClangLinter pragma disable: -fdoc-section-header-unknown
3035: /*@
3036: MatCreateBAIJ - Creates a sparse parallel matrix in `MATBAIJ` format
3037: (block compressed row).
3039: Collective
3041: Input Parameters:
3042: + comm - MPI communicator
3043: . bs - size of block, the blocks are ALWAYS square. One can use `MatSetBlockSizes()` to set a different row and column blocksize but the row
3044: blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with `MatCreateVecs()`
3045: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
3046: This value should be the same as the local size used in creating the
3047: y vector for the matrix-vector product y = Ax.
3048: . n - number of local columns (or `PETSC_DECIDE` to have calculated if N is given)
3049: This value should be the same as the local size used in creating the
3050: x vector for the matrix-vector product y = Ax.
3051: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
3052: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
3053: . d_nz - number of nonzero blocks per block row in diagonal portion of local
3054: submatrix (same for all local rows)
3055: . d_nnz - array containing the number of nonzero blocks in the various block rows
3056: of the in diagonal portion of the local (possibly different for each block
3057: row) or NULL. If you plan to factor the matrix you must leave room for the diagonal entry
3058: and set it even if it is zero.
3059: . o_nz - number of nonzero blocks per block row in the off-diagonal portion of local
3060: submatrix (same for all local rows).
3061: - o_nnz - array containing the number of nonzero blocks in the various block rows of the
3062: off-diagonal portion of the local submatrix (possibly different for
3063: each block row) or NULL.
3065: Output Parameter:
3066: . A - the matrix
3068: Options Database Keys:
3069: + -mat_block_size - size of the blocks to use
3070: - -mat_use_hash_table fact - set hash table factor
3072: Level: intermediate
3074: Notes:
3075: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
3076: MatXXXXSetPreallocation() paradigm instead of this routine directly.
3077: [MatXXXXSetPreallocation() is, for example, `MatSeqBAIJSetPreallocation()`]
3079: For good matrix assembly performance
3080: the user should preallocate the matrix storage by setting the parameters
3081: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`). By setting these parameters accurately,
3082: performance can be increased by more than a factor of 50.
3084: If the *_nnz parameter is given then the *_nz parameter is ignored
3086: A nonzero block is any block that as 1 or more nonzeros in it
3088: The user MUST specify either the local or global matrix dimensions
3089: (possibly both).
3091: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one processor
3092: than it must be used on all processors that share the object for that argument.
3094: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
3095: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
3097: Storage Information:
3098: For a square global matrix we define each processor's diagonal portion
3099: to be its local rows and the corresponding columns (a square submatrix);
3100: each processor's off-diagonal portion encompasses the remainder of the
3101: local matrix (a rectangular submatrix).
3103: The user can specify preallocated storage for the diagonal part of
3104: the local submatrix with either d_nz or d_nnz (not both). Set
3105: `d_nz` = `PETSC_DEFAULT` and `d_nnz` = `NULL` for PETSc to control dynamic
3106: memory allocation. Likewise, specify preallocated storage for the
3107: off-diagonal part of the local submatrix with `o_nz` or `o_nnz` (not both).
3109: Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
3110: the figure below we depict these three local rows and all columns (0-11).
3112: .vb
3113: 0 1 2 3 4 5 6 7 8 9 10 11
3114: --------------------------
3115: row 3 |o o o d d d o o o o o o
3116: row 4 |o o o d d d o o o o o o
3117: row 5 |o o o d d d o o o o o o
3118: --------------------------
3119: .ve
3121: Thus, any entries in the d locations are stored in the d (diagonal)
3122: submatrix, and any entries in the o locations are stored in the
3123: o (off-diagonal) submatrix. Note that the d and the o submatrices are
3124: stored simply in the `MATSEQBAIJ` format for compressed row storage.
3126: Now `d_nz` should indicate the number of block nonzeros per row in the d matrix,
3127: and `o_nz` should indicate the number of block nonzeros per row in the o matrix.
3128: In general, for PDE problems in which most nonzeros are near the diagonal,
3129: one expects `d_nz` >> `o_nz`.
3131: .seealso: `Mat`, `MatCreate()`, `MatCreateSeqBAIJ()`, `MatSetValues()`, `MatMPIBAIJSetPreallocation()`, `MatMPIBAIJSetPreallocationCSR()`,
3132: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, `MatGetOwnershipRangesColumn()`, `PetscLayout`
3133: @*/
3134: PetscErrorCode MatCreateBAIJ(MPI_Comm comm, PetscInt bs, PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[], Mat *A)
3135: {
3136: PetscMPIInt size;
3138: PetscFunctionBegin;
3139: PetscCall(MatCreate(comm, A));
3140: PetscCall(MatSetSizes(*A, m, n, M, N));
3141: PetscCallMPI(MPI_Comm_size(comm, &size));
3142: if (size > 1) {
3143: PetscCall(MatSetType(*A, MATMPIBAIJ));
3144: PetscCall(MatMPIBAIJSetPreallocation(*A, bs, d_nz, d_nnz, o_nz, o_nnz));
3145: } else {
3146: PetscCall(MatSetType(*A, MATSEQBAIJ));
3147: PetscCall(MatSeqBAIJSetPreallocation(*A, bs, d_nz, d_nnz));
3148: }
3149: PetscFunctionReturn(PETSC_SUCCESS);
3150: }
3152: static PetscErrorCode MatDuplicate_MPIBAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
3153: {
3154: Mat mat;
3155: Mat_MPIBAIJ *a, *oldmat = (Mat_MPIBAIJ *)matin->data;
3156: PetscInt len = 0;
3158: PetscFunctionBegin;
3159: *newmat = NULL;
3160: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
3161: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
3162: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
3164: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
3165: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
3166: if (matin->hash_active) PetscCall(MatSetUp(mat));
3167: else {
3168: mat->factortype = matin->factortype;
3169: mat->preallocated = PETSC_TRUE;
3170: mat->assembled = PETSC_TRUE;
3171: mat->insertmode = NOT_SET_VALUES;
3173: a = (Mat_MPIBAIJ *)mat->data;
3174: mat->rmap->bs = matin->rmap->bs;
3175: a->bs2 = oldmat->bs2;
3176: a->mbs = oldmat->mbs;
3177: a->nbs = oldmat->nbs;
3178: a->Mbs = oldmat->Mbs;
3179: a->Nbs = oldmat->Nbs;
3181: a->size = oldmat->size;
3182: a->rank = oldmat->rank;
3183: a->donotstash = oldmat->donotstash;
3184: a->roworiented = oldmat->roworiented;
3185: a->rowindices = NULL;
3186: a->rowvalues = NULL;
3187: a->getrowactive = PETSC_FALSE;
3188: a->barray = NULL;
3189: a->rstartbs = oldmat->rstartbs;
3190: a->rendbs = oldmat->rendbs;
3191: a->cstartbs = oldmat->cstartbs;
3192: a->cendbs = oldmat->cendbs;
3194: /* hash table stuff */
3195: a->ht = NULL;
3196: a->hd = NULL;
3197: a->ht_size = 0;
3198: a->ht_flag = oldmat->ht_flag;
3199: a->ht_fact = oldmat->ht_fact;
3200: a->ht_total_ct = 0;
3201: a->ht_insert_ct = 0;
3203: PetscCall(PetscArraycpy(a->rangebs, oldmat->rangebs, a->size + 1));
3204: if (oldmat->colmap) {
3205: #if defined(PETSC_USE_CTABLE)
3206: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
3207: #else
3208: PetscCall(PetscMalloc1(a->Nbs, &a->colmap));
3209: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, a->Nbs));
3210: #endif
3211: } else a->colmap = NULL;
3213: if (oldmat->garray && (len = ((Mat_SeqBAIJ *)oldmat->B->data)->nbs)) {
3214: PetscCall(PetscMalloc1(len, &a->garray));
3215: PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3216: } else a->garray = NULL;
3218: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)matin), matin->rmap->bs, &mat->bstash));
3219: PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3220: PetscCall(VecScatterCopy(oldmat->Mvctx, &a->Mvctx));
3222: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3223: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3224: }
3225: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3226: *newmat = mat;
3227: PetscFunctionReturn(PETSC_SUCCESS);
3228: }
3230: /* Used for both MPIBAIJ and MPISBAIJ matrices */
3231: PetscErrorCode MatLoad_MPIBAIJ_Binary(Mat mat, PetscViewer viewer)
3232: {
3233: PetscInt header[4], M, N, nz, bs, m, n, mbs, nbs, rows, cols, sum, i, j, k;
3234: PetscInt *rowidxs, *colidxs, rs, cs, ce;
3235: PetscScalar *matvals;
3237: PetscFunctionBegin;
3238: PetscCall(PetscViewerSetUp(viewer));
3240: /* read in matrix header */
3241: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3242: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3243: M = header[1];
3244: N = header[2];
3245: nz = header[3];
3246: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3247: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3248: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIBAIJ");
3250: /* set block sizes from the viewer's .info file */
3251: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3252: /* set local sizes if not set already */
3253: if (mat->rmap->n < 0 && M == N) mat->rmap->n = mat->cmap->n;
3254: if (mat->cmap->n < 0 && M == N) mat->cmap->n = mat->rmap->n;
3255: /* set global sizes if not set already */
3256: if (mat->rmap->N < 0) mat->rmap->N = M;
3257: if (mat->cmap->N < 0) mat->cmap->N = N;
3258: PetscCall(PetscLayoutSetUp(mat->rmap));
3259: PetscCall(PetscLayoutSetUp(mat->cmap));
3261: /* check if the matrix sizes are correct */
3262: PetscCall(MatGetSize(mat, &rows, &cols));
3263: PetscCheck(M == rows && N == cols, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%" PetscInt_FMT ", %" PetscInt_FMT ") than the input matrix (%" PetscInt_FMT ", %" PetscInt_FMT ")", M, N, rows, cols);
3264: PetscCall(MatGetBlockSize(mat, &bs));
3265: PetscCall(MatGetLocalSize(mat, &m, &n));
3266: PetscCall(PetscLayoutGetRange(mat->rmap, &rs, NULL));
3267: PetscCall(PetscLayoutGetRange(mat->cmap, &cs, &ce));
3268: mbs = m / bs;
3269: nbs = n / bs;
3271: /* read in row lengths and build row indices */
3272: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3273: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3274: rowidxs[0] = 0;
3275: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3276: PetscCallMPI(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3277: PetscCheck(sum == nz, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Inconsistent matrix data in file: nonzeros = %" PetscInt_FMT ", sum-row-lengths = %" PetscInt_FMT, nz, sum);
3279: /* read in column indices and matrix values */
3280: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3281: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3282: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3284: { /* preallocate matrix storage */
3285: PetscBT bt; /* helper bit set to count diagonal nonzeros */
3286: PetscHSetI ht; /* helper hash set to count off-diagonal nonzeros */
3287: PetscBool sbaij, done;
3288: PetscInt *d_nnz, *o_nnz;
3290: PetscCall(PetscBTCreate(nbs, &bt));
3291: PetscCall(PetscHSetICreate(&ht));
3292: PetscCall(PetscCalloc2(mbs, &d_nnz, mbs, &o_nnz));
3293: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATMPISBAIJ, &sbaij));
3294: for (i = 0; i < mbs; i++) {
3295: PetscCall(PetscBTMemzero(nbs, bt));
3296: PetscCall(PetscHSetIClear(ht));
3297: for (k = 0; k < bs; k++) {
3298: PetscInt row = bs * i + k;
3299: for (j = rowidxs[row]; j < rowidxs[row + 1]; j++) {
3300: PetscInt col = colidxs[j];
3301: if (!sbaij || col >= row) {
3302: if (col >= cs && col < ce) {
3303: if (!PetscBTLookupSet(bt, (col - cs) / bs)) d_nnz[i]++;
3304: } else {
3305: PetscCall(PetscHSetIQueryAdd(ht, col / bs, &done));
3306: if (done) o_nnz[i]++;
3307: }
3308: }
3309: }
3310: }
3311: }
3312: PetscCall(PetscBTDestroy(&bt));
3313: PetscCall(PetscHSetIDestroy(&ht));
3314: PetscCall(MatMPIBAIJSetPreallocation(mat, bs, 0, d_nnz, 0, o_nnz));
3315: PetscCall(MatMPISBAIJSetPreallocation(mat, bs, 0, d_nnz, 0, o_nnz));
3316: PetscCall(PetscFree2(d_nnz, o_nnz));
3317: }
3319: /* store matrix values */
3320: for (i = 0; i < m; i++) {
3321: PetscInt row = rs + i, s = rowidxs[i], e = rowidxs[i + 1];
3322: PetscUseTypeMethod(mat, setvalues, 1, &row, e - s, colidxs + s, matvals + s, INSERT_VALUES);
3323: }
3325: PetscCall(PetscFree(rowidxs));
3326: PetscCall(PetscFree2(colidxs, matvals));
3327: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
3328: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
3329: PetscFunctionReturn(PETSC_SUCCESS);
3330: }
3332: PetscErrorCode MatLoad_MPIBAIJ(Mat mat, PetscViewer viewer)
3333: {
3334: PetscBool isbinary;
3336: PetscFunctionBegin;
3337: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3338: PetscCheck(isbinary, PetscObjectComm((PetscObject)viewer), PETSC_ERR_SUP, "Viewer type %s not yet supported for reading %s matrices", ((PetscObject)viewer)->type_name, ((PetscObject)mat)->type_name);
3339: PetscCall(MatLoad_MPIBAIJ_Binary(mat, viewer));
3340: PetscFunctionReturn(PETSC_SUCCESS);
3341: }
3343: /*@
3344: MatMPIBAIJSetHashTableFactor - Sets the factor required to compute the size of the matrices hash table
3346: Input Parameters:
3347: + mat - the matrix
3348: - fact - factor
3350: Options Database Key:
3351: . -mat_use_hash_table fact - provide the factor
3353: Level: advanced
3355: .seealso: `Mat`, `MATMPIBAIJ`, `MatSetOption()`
3356: @*/
3357: PetscErrorCode MatMPIBAIJSetHashTableFactor(Mat mat, PetscReal fact)
3358: {
3359: PetscFunctionBegin;
3360: PetscTryMethod(mat, "MatSetHashTableFactor_C", (Mat, PetscReal), (mat, fact));
3361: PetscFunctionReturn(PETSC_SUCCESS);
3362: }
3364: PetscErrorCode MatSetHashTableFactor_MPIBAIJ(Mat mat, PetscReal fact)
3365: {
3366: Mat_MPIBAIJ *baij;
3368: PetscFunctionBegin;
3369: baij = (Mat_MPIBAIJ *)mat->data;
3370: baij->ht_fact = fact;
3371: PetscFunctionReturn(PETSC_SUCCESS);
3372: }
3374: PetscErrorCode MatMPIBAIJGetSeqBAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
3375: {
3376: Mat_MPIBAIJ *a = (Mat_MPIBAIJ *)A->data;
3377: PetscBool flg;
3379: PetscFunctionBegin;
3380: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIBAIJ, &flg));
3381: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIBAIJ matrix as input");
3382: if (Ad) *Ad = a->A;
3383: if (Ao) *Ao = a->B;
3384: if (colmap) *colmap = a->garray;
3385: PetscFunctionReturn(PETSC_SUCCESS);
3386: }
3388: /*
3389: Special version for direct calls from Fortran (to eliminate two function call overheads
3390: */
3391: #if defined(PETSC_HAVE_FORTRAN_CAPS)
3392: #define matmpibaijsetvaluesblocked_ MATMPIBAIJSETVALUESBLOCKED
3393: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
3394: #define matmpibaijsetvaluesblocked_ matmpibaijsetvaluesblocked
3395: #endif
3397: // PetscClangLinter pragma disable: -fdoc-synopsis-matching-symbol-name
3398: /*@C
3399: MatMPIBAIJSetValuesBlocked - Direct Fortran call to replace call to `MatSetValuesBlocked()`
3401: Collective
3403: Input Parameters:
3404: + matin - the matrix
3405: . min - number of input rows
3406: . im - input rows
3407: . nin - number of input columns
3408: . in - input columns
3409: . v - numerical values input
3410: - addvin - `INSERT_VALUES` or `ADD_VALUES`
3412: Level: advanced
3414: Developer Notes:
3415: This has a complete copy of `MatSetValuesBlocked_MPIBAIJ()` which is terrible code un-reuse.
3417: .seealso: `Mat`, `MatSetValuesBlocked()`
3418: @*/
3419: PETSC_EXTERN PetscErrorCode matmpibaijsetvaluesblocked_(Mat *matin, PetscInt *min, const PetscInt im[], PetscInt *nin, const PetscInt in[], const MatScalar v[], InsertMode *addvin)
3420: {
3421: /* convert input arguments to C version */
3422: Mat mat = *matin;
3423: PetscInt m = *min, n = *nin;
3424: InsertMode addv = *addvin;
3426: Mat_MPIBAIJ *baij = (Mat_MPIBAIJ *)mat->data;
3427: const MatScalar *value;
3428: MatScalar *barray = baij->barray;
3429: PetscBool roworiented = baij->roworiented;
3430: PetscInt i, j, ii, jj, row, col, rstart = baij->rstartbs;
3431: PetscInt rend = baij->rendbs, cstart = baij->cstartbs, stepval;
3432: PetscInt cend = baij->cendbs, bs = mat->rmap->bs, bs2 = baij->bs2;
3434: PetscFunctionBegin;
3435: /* tasks normally handled by MatSetValuesBlocked() */
3436: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
3437: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
3438: PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3439: if (mat->assembled) {
3440: mat->was_assembled = PETSC_TRUE;
3441: mat->assembled = PETSC_FALSE;
3442: }
3443: PetscCall(PetscLogEventBegin(MAT_SetValues, mat, 0, 0, 0));
3445: if (!barray) {
3446: PetscCall(PetscMalloc1(bs2, &barray));
3447: baij->barray = barray;
3448: }
3450: if (roworiented) stepval = (n - 1) * bs;
3451: else stepval = (m - 1) * bs;
3453: for (i = 0; i < m; i++) {
3454: if (im[i] < 0) continue;
3455: PetscCheck(im[i] < baij->Mbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large, row %" PetscInt_FMT " max %" PetscInt_FMT, im[i], baij->Mbs - 1);
3456: if (im[i] >= rstart && im[i] < rend) {
3457: row = im[i] - rstart;
3458: for (j = 0; j < n; j++) {
3459: /* If NumCol = 1 then a copy is not required */
3460: if ((roworiented) && (n == 1)) {
3461: barray = (MatScalar *)v + i * bs2;
3462: } else if ((!roworiented) && (m == 1)) {
3463: barray = (MatScalar *)v + j * bs2;
3464: } else { /* Here a copy is required */
3465: if (roworiented) {
3466: value = v + i * (stepval + bs) * bs + j * bs;
3467: } else {
3468: value = v + j * (stepval + bs) * bs + i * bs;
3469: }
3470: for (ii = 0; ii < bs; ii++, value += stepval) {
3471: for (jj = 0; jj < bs; jj++) *barray++ = *value++;
3472: }
3473: barray -= bs2;
3474: }
3476: if (in[j] >= cstart && in[j] < cend) {
3477: col = in[j] - cstart;
3478: PetscCall(MatSetValuesBlocked_SeqBAIJ_Inlined(baij->A, row, col, barray, addv, im[i], in[j]));
3479: } else if (in[j] < 0) {
3480: continue;
3481: } else {
3482: PetscCheck(in[j] < baij->Nbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large, col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], baij->Nbs - 1);
3483: if (mat->was_assembled) {
3484: if (!baij->colmap) PetscCall(MatCreateColmap_MPIBAIJ_Private(mat));
3486: #if defined(PETSC_USE_DEBUG)
3487: #if defined(PETSC_USE_CTABLE)
3488: {
3489: PetscInt data;
3490: PetscCall(PetscHMapIGetWithDefault(baij->colmap, in[j] + 1, 0, &data));
3491: PetscCheck((data - 1) % bs == 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Incorrect colmap");
3492: }
3493: #else
3494: PetscCheck((baij->colmap[in[j]] - 1) % bs == 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Incorrect colmap");
3495: #endif
3496: #endif
3497: #if defined(PETSC_USE_CTABLE)
3498: PetscCall(PetscHMapIGetWithDefault(baij->colmap, in[j] + 1, 0, &col));
3499: col = (col - 1) / bs;
3500: #else
3501: col = (baij->colmap[in[j]] - 1) / bs;
3502: #endif
3503: if (col < 0 && !((Mat_SeqBAIJ *)baij->A->data)->nonew) {
3504: PetscCall(MatDisAssemble_MPIBAIJ(mat));
3505: col = in[j];
3506: }
3507: } else col = in[j];
3508: PetscCall(MatSetValuesBlocked_SeqBAIJ_Inlined(baij->B, row, col, barray, addv, im[i], in[j]));
3509: }
3510: }
3511: } else {
3512: if (!baij->donotstash) {
3513: if (roworiented) {
3514: PetscCall(MatStashValuesRowBlocked_Private(&mat->bstash, im[i], n, in, v, m, n, i));
3515: } else {
3516: PetscCall(MatStashValuesColBlocked_Private(&mat->bstash, im[i], n, in, v, m, n, i));
3517: }
3518: }
3519: }
3520: }
3522: /* task normally handled by MatSetValuesBlocked() */
3523: PetscCall(PetscLogEventEnd(MAT_SetValues, mat, 0, 0, 0));
3524: PetscFunctionReturn(PETSC_SUCCESS);
3525: }
3527: /*@
3528: MatCreateMPIBAIJWithArrays - creates a `MATMPIBAIJ` matrix using arrays that contain in standard block CSR format for the local rows.
3530: Collective
3532: Input Parameters:
3533: + comm - MPI communicator
3534: . bs - the block size, only a block size of 1 is supported
3535: . m - number of local rows (Cannot be `PETSC_DECIDE`)
3536: . n - This value should be the same as the local size used in creating the
3537: x vector for the matrix-vector product $ y = Ax $. (or `PETSC_DECIDE` to have
3538: calculated if `N` is given) For square matrices `n` is almost always `m`.
3539: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
3540: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
3541: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of block elements in that rowth block row of the matrix
3542: . j - column indices
3543: - a - matrix values
3545: Output Parameter:
3546: . mat - the matrix
3548: Level: intermediate
3550: Notes:
3551: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
3552: thus you CANNOT change the matrix entries by changing the values of a[] after you have
3553: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3555: The order of the entries in values is the same as the block compressed sparse row storage format; that is, it is
3556: the same as a three dimensional array in Fortran values(bs,bs,nnz) that contains the first column of the first
3557: block, followed by the second column of the first block etc etc. That is, the blocks are contiguous in memory
3558: with column-major ordering within blocks.
3560: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3562: .seealso: `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
3563: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`
3564: @*/
3565: PetscErrorCode MatCreateMPIBAIJWithArrays(MPI_Comm comm, PetscInt bs, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
3566: {
3567: PetscFunctionBegin;
3568: PetscCheck(!i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
3569: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
3570: PetscCall(MatCreate(comm, mat));
3571: PetscCall(MatSetSizes(*mat, m, n, M, N));
3572: PetscCall(MatSetType(*mat, MATMPIBAIJ));
3573: PetscCall(MatSetBlockSize(*mat, bs));
3574: PetscCall(MatSetUp(*mat));
3575: PetscCall(MatSetOption(*mat, MAT_ROW_ORIENTED, PETSC_FALSE));
3576: PetscCall(MatMPIBAIJSetPreallocationCSR(*mat, bs, i, j, a));
3577: PetscCall(MatSetOption(*mat, MAT_ROW_ORIENTED, PETSC_TRUE));
3578: PetscFunctionReturn(PETSC_SUCCESS);
3579: }
3581: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIBAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
3582: {
3583: PetscInt m, N, i, rstart, nnz, Ii, bs, cbs;
3584: PetscInt *indx;
3585: PetscScalar *values;
3587: PetscFunctionBegin;
3588: PetscCall(MatGetSize(inmat, &m, &N));
3589: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
3590: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)inmat->data;
3591: PetscInt *dnz, *onz, mbs, Nbs, nbs;
3592: PetscInt *bindx, rmax = a->rmax, j;
3593: PetscMPIInt rank, size;
3595: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
3596: mbs = m / bs;
3597: Nbs = N / cbs;
3598: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnershipBlock(comm, cbs, &n, &N));
3599: nbs = n / cbs;
3601: PetscCall(PetscMalloc1(rmax, &bindx));
3602: MatPreallocateBegin(comm, mbs, nbs, dnz, onz); /* inline function, output __end and __rstart are used below */
3604: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3605: PetscCallMPI(MPI_Comm_size(comm, &size));
3606: if (rank == size - 1) {
3607: /* Check sum(nbs) = Nbs */
3608: PetscCheck(__end == Nbs, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local block columns %" PetscInt_FMT " != global block columns %" PetscInt_FMT, __end, Nbs);
3609: }
3611: rstart = __rstart; /* block rstart of *outmat; see inline function MatPreallocateBegin */
3612: for (i = 0; i < mbs; i++) {
3613: PetscCall(MatGetRow_SeqBAIJ(inmat, i * bs, &nnz, &indx, NULL)); /* non-blocked nnz and indx */
3614: nnz = nnz / bs;
3615: for (j = 0; j < nnz; j++) bindx[j] = indx[j * bs] / bs;
3616: PetscCall(MatPreallocateSet(i + rstart, nnz, bindx, dnz, onz));
3617: PetscCall(MatRestoreRow_SeqBAIJ(inmat, i * bs, &nnz, &indx, NULL));
3618: }
3619: PetscCall(PetscFree(bindx));
3621: PetscCall(MatCreate(comm, outmat));
3622: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
3623: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
3624: PetscCall(MatSetType(*outmat, MATBAIJ));
3625: PetscCall(MatSeqBAIJSetPreallocation(*outmat, bs, 0, dnz));
3626: PetscCall(MatMPIBAIJSetPreallocation(*outmat, bs, 0, dnz, 0, onz));
3627: MatPreallocateEnd(dnz, onz);
3628: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3629: }
3631: /* numeric phase */
3632: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
3633: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
3635: for (i = 0; i < m; i++) {
3636: PetscCall(MatGetRow_SeqBAIJ(inmat, i, &nnz, &indx, &values));
3637: Ii = i + rstart;
3638: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
3639: PetscCall(MatRestoreRow_SeqBAIJ(inmat, i, &nnz, &indx, &values));
3640: }
3641: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
3642: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
3643: PetscFunctionReturn(PETSC_SUCCESS);
3644: }