Actual source code: baij.c
1: /*
2: Defines the basic matrix operations for the BAIJ (compressed row)
3: matrix storage format.
4: */
5: #include <../src/mat/impls/baij/seq/baij.h>
6: #include <petscblaslapack.h>
7: #include <petsc/private/kernels/blockinvert.h>
8: #include <petsc/private/kernels/blockmatmult.h>
10: /* defines MatSetValues_Seq_Hash(), MatAssemblyEnd_Seq_Hash(), MatSetUp_Seq_Hash() */
11: #define TYPE BAIJ
12: #define TYPE_BS
13: #include "../src/mat/impls/aij/seq/seqhashmatsetvalues.h"
14: #undef TYPE_BS
15: #define TYPE_BS _BS
16: #define TYPE_BS_ON
17: #include "../src/mat/impls/aij/seq/seqhashmatsetvalues.h"
18: #undef TYPE_BS
19: #include "../src/mat/impls/aij/seq/seqhashmat.h"
20: #undef TYPE
21: #undef TYPE_BS_ON
23: #if PetscDefined(HAVE_HYPRE)
24: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
25: #endif
27: #if PetscDefined(HAVE_MKL_SPARSE_OPTIMIZE)
28: PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqBAIJMKL(Mat, MatType, MatReuse, Mat *);
29: #endif
30: #if PetscDefined(HAVE_LIBXSMM)
31: PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqBAIJLIBXSMM(Mat, MatType, MatReuse, Mat *);
32: #endif
33: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
35: MatGetDiagonalMarkers(SeqBAIJ, A->rmap->bs)
37: static PetscErrorCode MatGetColumnReductions_SeqBAIJ(Mat A, PetscInt type, PetscReal *reductions)
38: {
39: Mat_SeqBAIJ *a_aij = (Mat_SeqBAIJ *)A->data;
40: PetscInt m, n, ib, jb, bs = A->rmap->bs;
41: MatScalar *a_val = a_aij->a;
43: PetscFunctionBegin;
44: PetscCall(MatGetSize(A, &m, &n));
45: PetscCall(PetscArrayzero(reductions, n));
46: if (type == NORM_2) {
47: for (PetscInt i = a_aij->i[0]; i < a_aij->i[A->rmap->n / bs]; i++) {
48: for (jb = 0; jb < bs; jb++) {
49: for (ib = 0; ib < bs; ib++) {
50: reductions[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val * *a_val);
51: a_val++;
52: }
53: }
54: }
55: } else if (type == NORM_1) {
56: for (PetscInt i = a_aij->i[0]; i < a_aij->i[A->rmap->n / bs]; i++) {
57: for (jb = 0; jb < bs; jb++) {
58: for (ib = 0; ib < bs; ib++) {
59: reductions[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val);
60: a_val++;
61: }
62: }
63: }
64: } else if (type == NORM_INFINITY) {
65: for (PetscInt i = a_aij->i[0]; i < a_aij->i[A->rmap->n / bs]; i++) {
66: for (jb = 0; jb < bs; jb++) {
67: for (ib = 0; ib < bs; ib++) {
68: PetscInt col = A->cmap->rstart + a_aij->j[i] * bs + jb;
69: reductions[col] = PetscMax(PetscAbsScalar(*a_val), reductions[col]);
70: a_val++;
71: }
72: }
73: }
74: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
75: for (PetscInt i = a_aij->i[0]; i < a_aij->i[A->rmap->n / bs]; i++) {
76: for (jb = 0; jb < bs; jb++) {
77: for (ib = 0; ib < bs; ib++) {
78: reductions[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscRealPart(*a_val);
79: a_val++;
80: }
81: }
82: }
83: } else {
84: PetscCheck(type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
85: for (PetscInt i = a_aij->i[0]; i < a_aij->i[A->rmap->n / bs]; i++) {
86: for (jb = 0; jb < bs; jb++) {
87: for (ib = 0; ib < bs; ib++) {
88: reductions[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscImaginaryPart(*a_val);
89: a_val++;
90: }
91: }
92: }
93: }
94: if (type == NORM_2) {
95: for (PetscInt i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
96: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
97: for (PetscInt i = 0; i < n; i++) reductions[i] /= m;
98: }
99: PetscFunctionReturn(PETSC_SUCCESS);
100: }
102: static PetscErrorCode MatInvertBlockDiagonal_SeqBAIJ(Mat A, const PetscScalar **values)
103: {
104: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
105: PetscInt i, bs = A->rmap->bs, mbs = a->mbs, ipvt[5], bs2 = bs * bs, *v_pivots;
106: MatScalar *v = a->a, *odiag, *diag, work[25], *v_work;
107: PetscReal shift = 0.0;
108: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
109: const PetscInt *adiag;
111: PetscFunctionBegin;
112: allowzeropivot = PetscNot(A->erroriffailure);
114: if (a->idiag && a->idiagState == ((PetscObject)A)->state) {
115: if (values) *values = a->idiag;
116: PetscFunctionReturn(PETSC_SUCCESS);
117: }
118: PetscCall(MatGetDiagonalMarkers_SeqBAIJ(A, &adiag, NULL));
119: if (!a->idiag) PetscCall(PetscMalloc1(bs2 * mbs, &a->idiag));
120: diag = a->idiag;
121: if (values) *values = a->idiag;
122: /* factor and invert each block */
123: switch (bs) {
124: case 1:
125: for (i = 0; i < mbs; i++) {
126: odiag = v + 1 * adiag[i];
127: diag[0] = odiag[0];
129: if (PetscAbsScalar(diag[0] + shift) < PETSC_MACHINE_EPSILON) {
130: PetscCheck(allowzeropivot, PETSC_COMM_SELF, PETSC_ERR_MAT_LU_ZRPVT, "Zero pivot, row %" PetscInt_FMT " pivot value %g tolerance %g", i, (double)PetscAbsScalar(diag[0]), (double)PETSC_MACHINE_EPSILON);
131: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
132: A->factorerror_zeropivot_value = PetscAbsScalar(diag[0]);
133: A->factorerror_zeropivot_row = i;
134: PetscCall(PetscInfo(A, "Zero pivot, row %" PetscInt_FMT "\n", i));
135: }
137: diag[0] = (PetscScalar)1.0 / (diag[0] + shift);
138: diag += 1;
139: }
140: break;
141: case 2:
142: for (i = 0; i < mbs; i++) {
143: odiag = v + 4 * adiag[i];
144: diag[0] = odiag[0];
145: diag[1] = odiag[1];
146: diag[2] = odiag[2];
147: diag[3] = odiag[3];
148: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
149: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
150: diag += 4;
151: }
152: break;
153: case 3:
154: for (i = 0; i < mbs; i++) {
155: odiag = v + 9 * adiag[i];
156: diag[0] = odiag[0];
157: diag[1] = odiag[1];
158: diag[2] = odiag[2];
159: diag[3] = odiag[3];
160: diag[4] = odiag[4];
161: diag[5] = odiag[5];
162: diag[6] = odiag[6];
163: diag[7] = odiag[7];
164: diag[8] = odiag[8];
165: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
166: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
167: diag += 9;
168: }
169: break;
170: case 4:
171: for (i = 0; i < mbs; i++) {
172: odiag = v + 16 * adiag[i];
173: PetscCall(PetscArraycpy(diag, odiag, 16));
174: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
175: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
176: diag += 16;
177: }
178: break;
179: case 5:
180: for (i = 0; i < mbs; i++) {
181: odiag = v + 25 * adiag[i];
182: PetscCall(PetscArraycpy(diag, odiag, 25));
183: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
184: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
185: diag += 25;
186: }
187: break;
188: case 6:
189: for (i = 0; i < mbs; i++) {
190: odiag = v + 36 * adiag[i];
191: PetscCall(PetscArraycpy(diag, odiag, 36));
192: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
193: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
194: diag += 36;
195: }
196: break;
197: case 7:
198: for (i = 0; i < mbs; i++) {
199: odiag = v + 49 * adiag[i];
200: PetscCall(PetscArraycpy(diag, odiag, 49));
201: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
202: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
203: diag += 49;
204: }
205: break;
206: default:
207: PetscCall(PetscMalloc2(bs, &v_work, bs, &v_pivots));
208: for (i = 0; i < mbs; i++) {
209: odiag = v + bs2 * adiag[i];
210: PetscCall(PetscArraycpy(diag, odiag, bs2));
211: PetscCall(PetscKernel_A_gets_inverse_A(bs, diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
212: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
213: diag += bs2;
214: }
215: PetscCall(PetscFree2(v_work, v_pivots));
216: }
217: a->idiagState = ((PetscObject)A)->state;
218: PetscFunctionReturn(PETSC_SUCCESS);
219: }
221: static PetscErrorCode MatSOR_SeqBAIJ(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
222: {
223: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
224: PetscScalar *x, *work, *w, *workt, *t;
225: const MatScalar *v, *aa = a->a, *idiag;
226: const PetscScalar *b, *xb;
227: PetscScalar s[7], xw[7] = {0}; /* avoid some compilers thinking xw is uninitialized */
228: PetscInt m = a->mbs, i, i2, nz, bs = A->rmap->bs, bs2 = bs * bs, k, j, idx, it;
229: const PetscInt *diag, *ai = a->i, *aj = a->j, *vi;
231: PetscFunctionBegin;
232: its = its * lits;
233: PetscCheck(!(flag & SOR_EISENSTAT), PETSC_COMM_SELF, PETSC_ERR_SUP, "No support yet for Eisenstat");
234: PetscCheck(its > 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Relaxation requires global its %" PetscInt_FMT " and local its %" PetscInt_FMT " both positive", its, lits);
235: PetscCheck(!fshift, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for diagonal shift");
236: PetscCheck(omega == 1.0, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for non-trivial relaxation factor");
237: PetscCheck(!(flag & SOR_APPLY_UPPER) && !(flag & SOR_APPLY_LOWER), PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for applying upper or lower triangular parts");
239: PetscCall(MatInvertBlockDiagonal(A, NULL)); /* a no-op if the cached inverse is still current */
241: if (!m) PetscFunctionReturn(PETSC_SUCCESS);
242: diag = a->diag;
243: idiag = a->idiag;
244: k = PetscMax(A->rmap->n, A->cmap->n);
245: if (!a->mult_work) PetscCall(PetscMalloc1(k + 1, &a->mult_work));
246: if (!a->sor_workt) PetscCall(PetscMalloc1(k, &a->sor_workt));
247: if (!a->sor_work) PetscCall(PetscMalloc1(bs, &a->sor_work));
248: work = a->mult_work;
249: t = a->sor_workt;
250: w = a->sor_work;
252: PetscCall(VecGetArray(xx, &x));
253: PetscCall(VecGetArrayRead(bb, &b));
255: if (flag & SOR_ZERO_INITIAL_GUESS) {
256: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
257: switch (bs) {
258: case 1:
259: PetscKernel_v_gets_A_times_w_1(x, idiag, b);
260: t[0] = b[0];
261: i2 = 1;
262: idiag += 1;
263: for (i = 1; i < m; i++) {
264: v = aa + ai[i];
265: vi = aj + ai[i];
266: nz = diag[i] - ai[i];
267: s[0] = b[i2];
268: for (j = 0; j < nz; j++) {
269: xw[0] = x[vi[j]];
270: PetscKernel_v_gets_v_minus_A_times_w_1(s, (v + j), xw);
271: }
272: t[i2] = s[0];
273: PetscKernel_v_gets_A_times_w_1(xw, idiag, s);
274: x[i2] = xw[0];
275: idiag += 1;
276: i2 += 1;
277: }
278: break;
279: case 2:
280: PetscKernel_v_gets_A_times_w_2(x, idiag, b);
281: t[0] = b[0];
282: t[1] = b[1];
283: i2 = 2;
284: idiag += 4;
285: for (i = 1; i < m; i++) {
286: v = aa + 4 * ai[i];
287: vi = aj + ai[i];
288: nz = diag[i] - ai[i];
289: s[0] = b[i2];
290: s[1] = b[i2 + 1];
291: for (j = 0; j < nz; j++) {
292: idx = 2 * vi[j];
293: it = 4 * j;
294: xw[0] = x[idx];
295: xw[1] = x[1 + idx];
296: PetscKernel_v_gets_v_minus_A_times_w_2(s, (v + it), xw);
297: }
298: t[i2] = s[0];
299: t[i2 + 1] = s[1];
300: PetscKernel_v_gets_A_times_w_2(xw, idiag, s);
301: x[i2] = xw[0];
302: x[i2 + 1] = xw[1];
303: idiag += 4;
304: i2 += 2;
305: }
306: break;
307: case 3:
308: PetscKernel_v_gets_A_times_w_3(x, idiag, b);
309: t[0] = b[0];
310: t[1] = b[1];
311: t[2] = b[2];
312: i2 = 3;
313: idiag += 9;
314: for (i = 1; i < m; i++) {
315: v = aa + 9 * ai[i];
316: vi = aj + ai[i];
317: nz = diag[i] - ai[i];
318: s[0] = b[i2];
319: s[1] = b[i2 + 1];
320: s[2] = b[i2 + 2];
321: while (nz--) {
322: idx = 3 * (*vi++);
323: xw[0] = x[idx];
324: xw[1] = x[1 + idx];
325: xw[2] = x[2 + idx];
326: PetscKernel_v_gets_v_minus_A_times_w_3(s, v, xw);
327: v += 9;
328: }
329: t[i2] = s[0];
330: t[i2 + 1] = s[1];
331: t[i2 + 2] = s[2];
332: PetscKernel_v_gets_A_times_w_3(xw, idiag, s);
333: x[i2] = xw[0];
334: x[i2 + 1] = xw[1];
335: x[i2 + 2] = xw[2];
336: idiag += 9;
337: i2 += 3;
338: }
339: break;
340: case 4:
341: PetscKernel_v_gets_A_times_w_4(x, idiag, b);
342: t[0] = b[0];
343: t[1] = b[1];
344: t[2] = b[2];
345: t[3] = b[3];
346: i2 = 4;
347: idiag += 16;
348: for (i = 1; i < m; i++) {
349: v = aa + 16 * ai[i];
350: vi = aj + ai[i];
351: nz = diag[i] - ai[i];
352: s[0] = b[i2];
353: s[1] = b[i2 + 1];
354: s[2] = b[i2 + 2];
355: s[3] = b[i2 + 3];
356: while (nz--) {
357: idx = 4 * (*vi++);
358: xw[0] = x[idx];
359: xw[1] = x[1 + idx];
360: xw[2] = x[2 + idx];
361: xw[3] = x[3 + idx];
362: PetscKernel_v_gets_v_minus_A_times_w_4(s, v, xw);
363: v += 16;
364: }
365: t[i2] = s[0];
366: t[i2 + 1] = s[1];
367: t[i2 + 2] = s[2];
368: t[i2 + 3] = s[3];
369: PetscKernel_v_gets_A_times_w_4(xw, idiag, s);
370: x[i2] = xw[0];
371: x[i2 + 1] = xw[1];
372: x[i2 + 2] = xw[2];
373: x[i2 + 3] = xw[3];
374: idiag += 16;
375: i2 += 4;
376: }
377: break;
378: case 5:
379: PetscKernel_v_gets_A_times_w_5(x, idiag, b);
380: t[0] = b[0];
381: t[1] = b[1];
382: t[2] = b[2];
383: t[3] = b[3];
384: t[4] = b[4];
385: i2 = 5;
386: idiag += 25;
387: for (i = 1; i < m; i++) {
388: v = aa + 25 * ai[i];
389: vi = aj + ai[i];
390: nz = diag[i] - ai[i];
391: s[0] = b[i2];
392: s[1] = b[i2 + 1];
393: s[2] = b[i2 + 2];
394: s[3] = b[i2 + 3];
395: s[4] = b[i2 + 4];
396: while (nz--) {
397: idx = 5 * (*vi++);
398: xw[0] = x[idx];
399: xw[1] = x[1 + idx];
400: xw[2] = x[2 + idx];
401: xw[3] = x[3 + idx];
402: xw[4] = x[4 + idx];
403: PetscKernel_v_gets_v_minus_A_times_w_5(s, v, xw);
404: v += 25;
405: }
406: t[i2] = s[0];
407: t[i2 + 1] = s[1];
408: t[i2 + 2] = s[2];
409: t[i2 + 3] = s[3];
410: t[i2 + 4] = s[4];
411: PetscKernel_v_gets_A_times_w_5(xw, idiag, s);
412: x[i2] = xw[0];
413: x[i2 + 1] = xw[1];
414: x[i2 + 2] = xw[2];
415: x[i2 + 3] = xw[3];
416: x[i2 + 4] = xw[4];
417: idiag += 25;
418: i2 += 5;
419: }
420: break;
421: case 6:
422: PetscKernel_v_gets_A_times_w_6(x, idiag, b);
423: t[0] = b[0];
424: t[1] = b[1];
425: t[2] = b[2];
426: t[3] = b[3];
427: t[4] = b[4];
428: t[5] = b[5];
429: i2 = 6;
430: idiag += 36;
431: for (i = 1; i < m; i++) {
432: v = aa + 36 * ai[i];
433: vi = aj + ai[i];
434: nz = diag[i] - ai[i];
435: s[0] = b[i2];
436: s[1] = b[i2 + 1];
437: s[2] = b[i2 + 2];
438: s[3] = b[i2 + 3];
439: s[4] = b[i2 + 4];
440: s[5] = b[i2 + 5];
441: while (nz--) {
442: idx = 6 * (*vi++);
443: xw[0] = x[idx];
444: xw[1] = x[1 + idx];
445: xw[2] = x[2 + idx];
446: xw[3] = x[3 + idx];
447: xw[4] = x[4 + idx];
448: xw[5] = x[5 + idx];
449: PetscKernel_v_gets_v_minus_A_times_w_6(s, v, xw);
450: v += 36;
451: }
452: t[i2] = s[0];
453: t[i2 + 1] = s[1];
454: t[i2 + 2] = s[2];
455: t[i2 + 3] = s[3];
456: t[i2 + 4] = s[4];
457: t[i2 + 5] = s[5];
458: PetscKernel_v_gets_A_times_w_6(xw, idiag, s);
459: x[i2] = xw[0];
460: x[i2 + 1] = xw[1];
461: x[i2 + 2] = xw[2];
462: x[i2 + 3] = xw[3];
463: x[i2 + 4] = xw[4];
464: x[i2 + 5] = xw[5];
465: idiag += 36;
466: i2 += 6;
467: }
468: break;
469: case 7:
470: PetscKernel_v_gets_A_times_w_7(x, idiag, b);
471: t[0] = b[0];
472: t[1] = b[1];
473: t[2] = b[2];
474: t[3] = b[3];
475: t[4] = b[4];
476: t[5] = b[5];
477: t[6] = b[6];
478: i2 = 7;
479: idiag += 49;
480: for (i = 1; i < m; i++) {
481: v = aa + 49 * ai[i];
482: vi = aj + ai[i];
483: nz = diag[i] - ai[i];
484: s[0] = b[i2];
485: s[1] = b[i2 + 1];
486: s[2] = b[i2 + 2];
487: s[3] = b[i2 + 3];
488: s[4] = b[i2 + 4];
489: s[5] = b[i2 + 5];
490: s[6] = b[i2 + 6];
491: while (nz--) {
492: idx = 7 * (*vi++);
493: xw[0] = x[idx];
494: xw[1] = x[1 + idx];
495: xw[2] = x[2 + idx];
496: xw[3] = x[3 + idx];
497: xw[4] = x[4 + idx];
498: xw[5] = x[5 + idx];
499: xw[6] = x[6 + idx];
500: PetscKernel_v_gets_v_minus_A_times_w_7(s, v, xw);
501: v += 49;
502: }
503: t[i2] = s[0];
504: t[i2 + 1] = s[1];
505: t[i2 + 2] = s[2];
506: t[i2 + 3] = s[3];
507: t[i2 + 4] = s[4];
508: t[i2 + 5] = s[5];
509: t[i2 + 6] = s[6];
510: PetscKernel_v_gets_A_times_w_7(xw, idiag, s);
511: x[i2] = xw[0];
512: x[i2 + 1] = xw[1];
513: x[i2 + 2] = xw[2];
514: x[i2 + 3] = xw[3];
515: x[i2 + 4] = xw[4];
516: x[i2 + 5] = xw[5];
517: x[i2 + 6] = xw[6];
518: idiag += 49;
519: i2 += 7;
520: }
521: break;
522: default:
523: PetscKernel_w_gets_Ar_times_v(bs, bs, b, idiag, x);
524: PetscCall(PetscArraycpy(t, b, bs));
525: i2 = bs;
526: idiag += bs2;
527: for (i = 1; i < m; i++) {
528: v = aa + bs2 * ai[i];
529: vi = aj + ai[i];
530: nz = diag[i] - ai[i];
532: PetscCall(PetscArraycpy(w, b + i2, bs));
533: /* copy all rows of x that are needed into contiguous space */
534: workt = work;
535: for (j = 0; j < nz; j++) {
536: PetscCall(PetscArraycpy(workt, x + bs * (*vi++), bs));
537: workt += bs;
538: }
539: PetscKernel_w_gets_w_minus_Ar_times_v(bs, bs * nz, w, v, work);
540: PetscCall(PetscArraycpy(t + i2, w, bs));
541: PetscKernel_w_gets_Ar_times_v(bs, bs, w, idiag, x + i2);
543: idiag += bs2;
544: i2 += bs;
545: }
546: break;
547: }
548: /* for logging purposes assume number of nonzero in lower half is 1/2 of total */
549: PetscCall(PetscLogFlops(1.0 * bs2 * a->nz));
550: xb = t;
551: } else xb = b;
552: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
553: idiag = a->idiag + bs2 * (a->mbs - 1);
554: i2 = bs * (m - 1);
555: switch (bs) {
556: case 1:
557: s[0] = xb[i2];
558: PetscKernel_v_gets_A_times_w_1(xw, idiag, s);
559: x[i2] = xw[0];
560: i2 -= 1;
561: for (i = m - 2; i >= 0; i--) {
562: v = aa + (diag[i] + 1);
563: vi = aj + diag[i] + 1;
564: nz = ai[i + 1] - diag[i] - 1;
565: s[0] = xb[i2];
566: for (j = 0; j < nz; j++) {
567: xw[0] = x[vi[j]];
568: PetscKernel_v_gets_v_minus_A_times_w_1(s, (v + j), xw);
569: }
570: PetscKernel_v_gets_A_times_w_1(xw, idiag, s);
571: x[i2] = xw[0];
572: idiag -= 1;
573: i2 -= 1;
574: }
575: break;
576: case 2:
577: s[0] = xb[i2];
578: s[1] = xb[i2 + 1];
579: PetscKernel_v_gets_A_times_w_2(xw, idiag, s);
580: x[i2] = xw[0];
581: x[i2 + 1] = xw[1];
582: i2 -= 2;
583: idiag -= 4;
584: for (i = m - 2; i >= 0; i--) {
585: v = aa + 4 * (diag[i] + 1);
586: vi = aj + diag[i] + 1;
587: nz = ai[i + 1] - diag[i] - 1;
588: s[0] = xb[i2];
589: s[1] = xb[i2 + 1];
590: for (j = 0; j < nz; j++) {
591: idx = 2 * vi[j];
592: it = 4 * j;
593: xw[0] = x[idx];
594: xw[1] = x[1 + idx];
595: PetscKernel_v_gets_v_minus_A_times_w_2(s, (v + it), xw);
596: }
597: PetscKernel_v_gets_A_times_w_2(xw, idiag, s);
598: x[i2] = xw[0];
599: x[i2 + 1] = xw[1];
600: idiag -= 4;
601: i2 -= 2;
602: }
603: break;
604: case 3:
605: s[0] = xb[i2];
606: s[1] = xb[i2 + 1];
607: s[2] = xb[i2 + 2];
608: PetscKernel_v_gets_A_times_w_3(xw, idiag, s);
609: x[i2] = xw[0];
610: x[i2 + 1] = xw[1];
611: x[i2 + 2] = xw[2];
612: i2 -= 3;
613: idiag -= 9;
614: for (i = m - 2; i >= 0; i--) {
615: v = aa + 9 * (diag[i] + 1);
616: vi = aj + diag[i] + 1;
617: nz = ai[i + 1] - diag[i] - 1;
618: s[0] = xb[i2];
619: s[1] = xb[i2 + 1];
620: s[2] = xb[i2 + 2];
621: while (nz--) {
622: idx = 3 * (*vi++);
623: xw[0] = x[idx];
624: xw[1] = x[1 + idx];
625: xw[2] = x[2 + idx];
626: PetscKernel_v_gets_v_minus_A_times_w_3(s, v, xw);
627: v += 9;
628: }
629: PetscKernel_v_gets_A_times_w_3(xw, idiag, s);
630: x[i2] = xw[0];
631: x[i2 + 1] = xw[1];
632: x[i2 + 2] = xw[2];
633: idiag -= 9;
634: i2 -= 3;
635: }
636: break;
637: case 4:
638: s[0] = xb[i2];
639: s[1] = xb[i2 + 1];
640: s[2] = xb[i2 + 2];
641: s[3] = xb[i2 + 3];
642: PetscKernel_v_gets_A_times_w_4(xw, idiag, s);
643: x[i2] = xw[0];
644: x[i2 + 1] = xw[1];
645: x[i2 + 2] = xw[2];
646: x[i2 + 3] = xw[3];
647: i2 -= 4;
648: idiag -= 16;
649: for (i = m - 2; i >= 0; i--) {
650: v = aa + 16 * (diag[i] + 1);
651: vi = aj + diag[i] + 1;
652: nz = ai[i + 1] - diag[i] - 1;
653: s[0] = xb[i2];
654: s[1] = xb[i2 + 1];
655: s[2] = xb[i2 + 2];
656: s[3] = xb[i2 + 3];
657: while (nz--) {
658: idx = 4 * (*vi++);
659: xw[0] = x[idx];
660: xw[1] = x[1 + idx];
661: xw[2] = x[2 + idx];
662: xw[3] = x[3 + idx];
663: PetscKernel_v_gets_v_minus_A_times_w_4(s, v, xw);
664: v += 16;
665: }
666: PetscKernel_v_gets_A_times_w_4(xw, idiag, s);
667: x[i2] = xw[0];
668: x[i2 + 1] = xw[1];
669: x[i2 + 2] = xw[2];
670: x[i2 + 3] = xw[3];
671: idiag -= 16;
672: i2 -= 4;
673: }
674: break;
675: case 5:
676: s[0] = xb[i2];
677: s[1] = xb[i2 + 1];
678: s[2] = xb[i2 + 2];
679: s[3] = xb[i2 + 3];
680: s[4] = xb[i2 + 4];
681: PetscKernel_v_gets_A_times_w_5(xw, idiag, s);
682: x[i2] = xw[0];
683: x[i2 + 1] = xw[1];
684: x[i2 + 2] = xw[2];
685: x[i2 + 3] = xw[3];
686: x[i2 + 4] = xw[4];
687: i2 -= 5;
688: idiag -= 25;
689: for (i = m - 2; i >= 0; i--) {
690: v = aa + 25 * (diag[i] + 1);
691: vi = aj + diag[i] + 1;
692: nz = ai[i + 1] - diag[i] - 1;
693: s[0] = xb[i2];
694: s[1] = xb[i2 + 1];
695: s[2] = xb[i2 + 2];
696: s[3] = xb[i2 + 3];
697: s[4] = xb[i2 + 4];
698: while (nz--) {
699: idx = 5 * (*vi++);
700: xw[0] = x[idx];
701: xw[1] = x[1 + idx];
702: xw[2] = x[2 + idx];
703: xw[3] = x[3 + idx];
704: xw[4] = x[4 + idx];
705: PetscKernel_v_gets_v_minus_A_times_w_5(s, v, xw);
706: v += 25;
707: }
708: PetscKernel_v_gets_A_times_w_5(xw, idiag, s);
709: x[i2] = xw[0];
710: x[i2 + 1] = xw[1];
711: x[i2 + 2] = xw[2];
712: x[i2 + 3] = xw[3];
713: x[i2 + 4] = xw[4];
714: idiag -= 25;
715: i2 -= 5;
716: }
717: break;
718: case 6:
719: s[0] = xb[i2];
720: s[1] = xb[i2 + 1];
721: s[2] = xb[i2 + 2];
722: s[3] = xb[i2 + 3];
723: s[4] = xb[i2 + 4];
724: s[5] = xb[i2 + 5];
725: PetscKernel_v_gets_A_times_w_6(xw, idiag, s);
726: x[i2] = xw[0];
727: x[i2 + 1] = xw[1];
728: x[i2 + 2] = xw[2];
729: x[i2 + 3] = xw[3];
730: x[i2 + 4] = xw[4];
731: x[i2 + 5] = xw[5];
732: i2 -= 6;
733: idiag -= 36;
734: for (i = m - 2; i >= 0; i--) {
735: v = aa + 36 * (diag[i] + 1);
736: vi = aj + diag[i] + 1;
737: nz = ai[i + 1] - diag[i] - 1;
738: s[0] = xb[i2];
739: s[1] = xb[i2 + 1];
740: s[2] = xb[i2 + 2];
741: s[3] = xb[i2 + 3];
742: s[4] = xb[i2 + 4];
743: s[5] = xb[i2 + 5];
744: while (nz--) {
745: idx = 6 * (*vi++);
746: xw[0] = x[idx];
747: xw[1] = x[1 + idx];
748: xw[2] = x[2 + idx];
749: xw[3] = x[3 + idx];
750: xw[4] = x[4 + idx];
751: xw[5] = x[5 + idx];
752: PetscKernel_v_gets_v_minus_A_times_w_6(s, v, xw);
753: v += 36;
754: }
755: PetscKernel_v_gets_A_times_w_6(xw, idiag, s);
756: x[i2] = xw[0];
757: x[i2 + 1] = xw[1];
758: x[i2 + 2] = xw[2];
759: x[i2 + 3] = xw[3];
760: x[i2 + 4] = xw[4];
761: x[i2 + 5] = xw[5];
762: idiag -= 36;
763: i2 -= 6;
764: }
765: break;
766: case 7:
767: s[0] = xb[i2];
768: s[1] = xb[i2 + 1];
769: s[2] = xb[i2 + 2];
770: s[3] = xb[i2 + 3];
771: s[4] = xb[i2 + 4];
772: s[5] = xb[i2 + 5];
773: s[6] = xb[i2 + 6];
774: PetscKernel_v_gets_A_times_w_7(x, idiag, b);
775: x[i2] = xw[0];
776: x[i2 + 1] = xw[1];
777: x[i2 + 2] = xw[2];
778: x[i2 + 3] = xw[3];
779: x[i2 + 4] = xw[4];
780: x[i2 + 5] = xw[5];
781: x[i2 + 6] = xw[6];
782: i2 -= 7;
783: idiag -= 49;
784: for (i = m - 2; i >= 0; i--) {
785: v = aa + 49 * (diag[i] + 1);
786: vi = aj + diag[i] + 1;
787: nz = ai[i + 1] - diag[i] - 1;
788: s[0] = xb[i2];
789: s[1] = xb[i2 + 1];
790: s[2] = xb[i2 + 2];
791: s[3] = xb[i2 + 3];
792: s[4] = xb[i2 + 4];
793: s[5] = xb[i2 + 5];
794: s[6] = xb[i2 + 6];
795: while (nz--) {
796: idx = 7 * (*vi++);
797: xw[0] = x[idx];
798: xw[1] = x[1 + idx];
799: xw[2] = x[2 + idx];
800: xw[3] = x[3 + idx];
801: xw[4] = x[4 + idx];
802: xw[5] = x[5 + idx];
803: xw[6] = x[6 + idx];
804: PetscKernel_v_gets_v_minus_A_times_w_7(s, v, xw);
805: v += 49;
806: }
807: PetscKernel_v_gets_A_times_w_7(xw, idiag, s);
808: x[i2] = xw[0];
809: x[i2 + 1] = xw[1];
810: x[i2 + 2] = xw[2];
811: x[i2 + 3] = xw[3];
812: x[i2 + 4] = xw[4];
813: x[i2 + 5] = xw[5];
814: x[i2 + 6] = xw[6];
815: idiag -= 49;
816: i2 -= 7;
817: }
818: break;
819: default:
820: PetscCall(PetscArraycpy(w, xb + i2, bs));
821: PetscKernel_w_gets_Ar_times_v(bs, bs, w, idiag, x + i2);
822: i2 -= bs;
823: idiag -= bs2;
824: for (i = m - 2; i >= 0; i--) {
825: v = aa + bs2 * (diag[i] + 1);
826: vi = aj + diag[i] + 1;
827: nz = ai[i + 1] - diag[i] - 1;
829: PetscCall(PetscArraycpy(w, xb + i2, bs));
830: /* copy all rows of x that are needed into contiguous space */
831: workt = work;
832: for (j = 0; j < nz; j++) {
833: PetscCall(PetscArraycpy(workt, x + bs * (*vi++), bs));
834: workt += bs;
835: }
836: PetscKernel_w_gets_w_minus_Ar_times_v(bs, bs * nz, w, v, work);
837: PetscKernel_w_gets_Ar_times_v(bs, bs, w, idiag, x + i2);
839: idiag -= bs2;
840: i2 -= bs;
841: }
842: break;
843: }
844: PetscCall(PetscLogFlops(1.0 * bs2 * (a->nz)));
845: }
846: its--;
847: }
848: while (its--) {
849: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
850: idiag = a->idiag;
851: i2 = 0;
852: switch (bs) {
853: case 1:
854: for (i = 0; i < m; i++) {
855: v = aa + ai[i];
856: vi = aj + ai[i];
857: nz = ai[i + 1] - ai[i];
858: s[0] = b[i2];
859: for (j = 0; j < nz; j++) {
860: xw[0] = x[vi[j]];
861: PetscKernel_v_gets_v_minus_A_times_w_1(s, (v + j), xw);
862: }
863: PetscKernel_v_gets_A_times_w_1(xw, idiag, s);
864: x[i2] += xw[0];
865: idiag += 1;
866: i2 += 1;
867: }
868: break;
869: case 2:
870: for (i = 0; i < m; i++) {
871: v = aa + 4 * ai[i];
872: vi = aj + ai[i];
873: nz = ai[i + 1] - ai[i];
874: s[0] = b[i2];
875: s[1] = b[i2 + 1];
876: for (j = 0; j < nz; j++) {
877: idx = 2 * vi[j];
878: it = 4 * j;
879: xw[0] = x[idx];
880: xw[1] = x[1 + idx];
881: PetscKernel_v_gets_v_minus_A_times_w_2(s, (v + it), xw);
882: }
883: PetscKernel_v_gets_A_times_w_2(xw, idiag, s);
884: x[i2] += xw[0];
885: x[i2 + 1] += xw[1];
886: idiag += 4;
887: i2 += 2;
888: }
889: break;
890: case 3:
891: for (i = 0; i < m; i++) {
892: v = aa + 9 * ai[i];
893: vi = aj + ai[i];
894: nz = ai[i + 1] - ai[i];
895: s[0] = b[i2];
896: s[1] = b[i2 + 1];
897: s[2] = b[i2 + 2];
898: while (nz--) {
899: idx = 3 * (*vi++);
900: xw[0] = x[idx];
901: xw[1] = x[1 + idx];
902: xw[2] = x[2 + idx];
903: PetscKernel_v_gets_v_minus_A_times_w_3(s, v, xw);
904: v += 9;
905: }
906: PetscKernel_v_gets_A_times_w_3(xw, idiag, s);
907: x[i2] += xw[0];
908: x[i2 + 1] += xw[1];
909: x[i2 + 2] += xw[2];
910: idiag += 9;
911: i2 += 3;
912: }
913: break;
914: case 4:
915: for (i = 0; i < m; i++) {
916: v = aa + 16 * ai[i];
917: vi = aj + ai[i];
918: nz = ai[i + 1] - ai[i];
919: s[0] = b[i2];
920: s[1] = b[i2 + 1];
921: s[2] = b[i2 + 2];
922: s[3] = b[i2 + 3];
923: while (nz--) {
924: idx = 4 * (*vi++);
925: xw[0] = x[idx];
926: xw[1] = x[1 + idx];
927: xw[2] = x[2 + idx];
928: xw[3] = x[3 + idx];
929: PetscKernel_v_gets_v_minus_A_times_w_4(s, v, xw);
930: v += 16;
931: }
932: PetscKernel_v_gets_A_times_w_4(xw, idiag, s);
933: x[i2] += xw[0];
934: x[i2 + 1] += xw[1];
935: x[i2 + 2] += xw[2];
936: x[i2 + 3] += xw[3];
937: idiag += 16;
938: i2 += 4;
939: }
940: break;
941: case 5:
942: for (i = 0; i < m; i++) {
943: v = aa + 25 * ai[i];
944: vi = aj + ai[i];
945: nz = ai[i + 1] - ai[i];
946: s[0] = b[i2];
947: s[1] = b[i2 + 1];
948: s[2] = b[i2 + 2];
949: s[3] = b[i2 + 3];
950: s[4] = b[i2 + 4];
951: while (nz--) {
952: idx = 5 * (*vi++);
953: xw[0] = x[idx];
954: xw[1] = x[1 + idx];
955: xw[2] = x[2 + idx];
956: xw[3] = x[3 + idx];
957: xw[4] = x[4 + idx];
958: PetscKernel_v_gets_v_minus_A_times_w_5(s, v, xw);
959: v += 25;
960: }
961: PetscKernel_v_gets_A_times_w_5(xw, idiag, s);
962: x[i2] += xw[0];
963: x[i2 + 1] += xw[1];
964: x[i2 + 2] += xw[2];
965: x[i2 + 3] += xw[3];
966: x[i2 + 4] += xw[4];
967: idiag += 25;
968: i2 += 5;
969: }
970: break;
971: case 6:
972: for (i = 0; i < m; i++) {
973: v = aa + 36 * ai[i];
974: vi = aj + ai[i];
975: nz = ai[i + 1] - ai[i];
976: s[0] = b[i2];
977: s[1] = b[i2 + 1];
978: s[2] = b[i2 + 2];
979: s[3] = b[i2 + 3];
980: s[4] = b[i2 + 4];
981: s[5] = b[i2 + 5];
982: while (nz--) {
983: idx = 6 * (*vi++);
984: xw[0] = x[idx];
985: xw[1] = x[1 + idx];
986: xw[2] = x[2 + idx];
987: xw[3] = x[3 + idx];
988: xw[4] = x[4 + idx];
989: xw[5] = x[5 + idx];
990: PetscKernel_v_gets_v_minus_A_times_w_6(s, v, xw);
991: v += 36;
992: }
993: PetscKernel_v_gets_A_times_w_6(xw, idiag, s);
994: x[i2] += xw[0];
995: x[i2 + 1] += xw[1];
996: x[i2 + 2] += xw[2];
997: x[i2 + 3] += xw[3];
998: x[i2 + 4] += xw[4];
999: x[i2 + 5] += xw[5];
1000: idiag += 36;
1001: i2 += 6;
1002: }
1003: break;
1004: case 7:
1005: for (i = 0; i < m; i++) {
1006: v = aa + 49 * ai[i];
1007: vi = aj + ai[i];
1008: nz = ai[i + 1] - ai[i];
1009: s[0] = b[i2];
1010: s[1] = b[i2 + 1];
1011: s[2] = b[i2 + 2];
1012: s[3] = b[i2 + 3];
1013: s[4] = b[i2 + 4];
1014: s[5] = b[i2 + 5];
1015: s[6] = b[i2 + 6];
1016: while (nz--) {
1017: idx = 7 * (*vi++);
1018: xw[0] = x[idx];
1019: xw[1] = x[1 + idx];
1020: xw[2] = x[2 + idx];
1021: xw[3] = x[3 + idx];
1022: xw[4] = x[4 + idx];
1023: xw[5] = x[5 + idx];
1024: xw[6] = x[6 + idx];
1025: PetscKernel_v_gets_v_minus_A_times_w_7(s, v, xw);
1026: v += 49;
1027: }
1028: PetscKernel_v_gets_A_times_w_7(xw, idiag, s);
1029: x[i2] += xw[0];
1030: x[i2 + 1] += xw[1];
1031: x[i2 + 2] += xw[2];
1032: x[i2 + 3] += xw[3];
1033: x[i2 + 4] += xw[4];
1034: x[i2 + 5] += xw[5];
1035: x[i2 + 6] += xw[6];
1036: idiag += 49;
1037: i2 += 7;
1038: }
1039: break;
1040: default:
1041: for (i = 0; i < m; i++) {
1042: v = aa + bs2 * ai[i];
1043: vi = aj + ai[i];
1044: nz = ai[i + 1] - ai[i];
1046: PetscCall(PetscArraycpy(w, b + i2, bs));
1047: /* copy all rows of x that are needed into contiguous space */
1048: workt = work;
1049: for (j = 0; j < nz; j++) {
1050: PetscCall(PetscArraycpy(workt, x + bs * (*vi++), bs));
1051: workt += bs;
1052: }
1053: PetscKernel_w_gets_w_minus_Ar_times_v(bs, bs * nz, w, v, work);
1054: PetscKernel_w_gets_w_plus_Ar_times_v(bs, bs, w, idiag, x + i2);
1056: idiag += bs2;
1057: i2 += bs;
1058: }
1059: break;
1060: }
1061: PetscCall(PetscLogFlops(2.0 * bs2 * a->nz));
1062: }
1063: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1064: idiag = a->idiag + bs2 * (a->mbs - 1);
1065: i2 = bs * (m - 1);
1066: switch (bs) {
1067: case 1:
1068: for (i = m - 1; i >= 0; i--) {
1069: v = aa + ai[i];
1070: vi = aj + ai[i];
1071: nz = ai[i + 1] - ai[i];
1072: s[0] = b[i2];
1073: for (j = 0; j < nz; j++) {
1074: xw[0] = x[vi[j]];
1075: PetscKernel_v_gets_v_minus_A_times_w_1(s, (v + j), xw);
1076: }
1077: PetscKernel_v_gets_A_times_w_1(xw, idiag, s);
1078: x[i2] += xw[0];
1079: idiag -= 1;
1080: i2 -= 1;
1081: }
1082: break;
1083: case 2:
1084: for (i = m - 1; i >= 0; i--) {
1085: v = aa + 4 * ai[i];
1086: vi = aj + ai[i];
1087: nz = ai[i + 1] - ai[i];
1088: s[0] = b[i2];
1089: s[1] = b[i2 + 1];
1090: for (j = 0; j < nz; j++) {
1091: idx = 2 * vi[j];
1092: it = 4 * j;
1093: xw[0] = x[idx];
1094: xw[1] = x[1 + idx];
1095: PetscKernel_v_gets_v_minus_A_times_w_2(s, (v + it), xw);
1096: }
1097: PetscKernel_v_gets_A_times_w_2(xw, idiag, s);
1098: x[i2] += xw[0];
1099: x[i2 + 1] += xw[1];
1100: idiag -= 4;
1101: i2 -= 2;
1102: }
1103: break;
1104: case 3:
1105: for (i = m - 1; i >= 0; i--) {
1106: v = aa + 9 * ai[i];
1107: vi = aj + ai[i];
1108: nz = ai[i + 1] - ai[i];
1109: s[0] = b[i2];
1110: s[1] = b[i2 + 1];
1111: s[2] = b[i2 + 2];
1112: while (nz--) {
1113: idx = 3 * (*vi++);
1114: xw[0] = x[idx];
1115: xw[1] = x[1 + idx];
1116: xw[2] = x[2 + idx];
1117: PetscKernel_v_gets_v_minus_A_times_w_3(s, v, xw);
1118: v += 9;
1119: }
1120: PetscKernel_v_gets_A_times_w_3(xw, idiag, s);
1121: x[i2] += xw[0];
1122: x[i2 + 1] += xw[1];
1123: x[i2 + 2] += xw[2];
1124: idiag -= 9;
1125: i2 -= 3;
1126: }
1127: break;
1128: case 4:
1129: for (i = m - 1; i >= 0; i--) {
1130: v = aa + 16 * ai[i];
1131: vi = aj + ai[i];
1132: nz = ai[i + 1] - ai[i];
1133: s[0] = b[i2];
1134: s[1] = b[i2 + 1];
1135: s[2] = b[i2 + 2];
1136: s[3] = b[i2 + 3];
1137: while (nz--) {
1138: idx = 4 * (*vi++);
1139: xw[0] = x[idx];
1140: xw[1] = x[1 + idx];
1141: xw[2] = x[2 + idx];
1142: xw[3] = x[3 + idx];
1143: PetscKernel_v_gets_v_minus_A_times_w_4(s, v, xw);
1144: v += 16;
1145: }
1146: PetscKernel_v_gets_A_times_w_4(xw, idiag, s);
1147: x[i2] += xw[0];
1148: x[i2 + 1] += xw[1];
1149: x[i2 + 2] += xw[2];
1150: x[i2 + 3] += xw[3];
1151: idiag -= 16;
1152: i2 -= 4;
1153: }
1154: break;
1155: case 5:
1156: for (i = m - 1; i >= 0; i--) {
1157: v = aa + 25 * ai[i];
1158: vi = aj + ai[i];
1159: nz = ai[i + 1] - ai[i];
1160: s[0] = b[i2];
1161: s[1] = b[i2 + 1];
1162: s[2] = b[i2 + 2];
1163: s[3] = b[i2 + 3];
1164: s[4] = b[i2 + 4];
1165: while (nz--) {
1166: idx = 5 * (*vi++);
1167: xw[0] = x[idx];
1168: xw[1] = x[1 + idx];
1169: xw[2] = x[2 + idx];
1170: xw[3] = x[3 + idx];
1171: xw[4] = x[4 + idx];
1172: PetscKernel_v_gets_v_minus_A_times_w_5(s, v, xw);
1173: v += 25;
1174: }
1175: PetscKernel_v_gets_A_times_w_5(xw, idiag, s);
1176: x[i2] += xw[0];
1177: x[i2 + 1] += xw[1];
1178: x[i2 + 2] += xw[2];
1179: x[i2 + 3] += xw[3];
1180: x[i2 + 4] += xw[4];
1181: idiag -= 25;
1182: i2 -= 5;
1183: }
1184: break;
1185: case 6:
1186: for (i = m - 1; i >= 0; i--) {
1187: v = aa + 36 * ai[i];
1188: vi = aj + ai[i];
1189: nz = ai[i + 1] - ai[i];
1190: s[0] = b[i2];
1191: s[1] = b[i2 + 1];
1192: s[2] = b[i2 + 2];
1193: s[3] = b[i2 + 3];
1194: s[4] = b[i2 + 4];
1195: s[5] = b[i2 + 5];
1196: while (nz--) {
1197: idx = 6 * (*vi++);
1198: xw[0] = x[idx];
1199: xw[1] = x[1 + idx];
1200: xw[2] = x[2 + idx];
1201: xw[3] = x[3 + idx];
1202: xw[4] = x[4 + idx];
1203: xw[5] = x[5 + idx];
1204: PetscKernel_v_gets_v_minus_A_times_w_6(s, v, xw);
1205: v += 36;
1206: }
1207: PetscKernel_v_gets_A_times_w_6(xw, idiag, s);
1208: x[i2] += xw[0];
1209: x[i2 + 1] += xw[1];
1210: x[i2 + 2] += xw[2];
1211: x[i2 + 3] += xw[3];
1212: x[i2 + 4] += xw[4];
1213: x[i2 + 5] += xw[5];
1214: idiag -= 36;
1215: i2 -= 6;
1216: }
1217: break;
1218: case 7:
1219: for (i = m - 1; i >= 0; i--) {
1220: v = aa + 49 * ai[i];
1221: vi = aj + ai[i];
1222: nz = ai[i + 1] - ai[i];
1223: s[0] = b[i2];
1224: s[1] = b[i2 + 1];
1225: s[2] = b[i2 + 2];
1226: s[3] = b[i2 + 3];
1227: s[4] = b[i2 + 4];
1228: s[5] = b[i2 + 5];
1229: s[6] = b[i2 + 6];
1230: while (nz--) {
1231: idx = 7 * (*vi++);
1232: xw[0] = x[idx];
1233: xw[1] = x[1 + idx];
1234: xw[2] = x[2 + idx];
1235: xw[3] = x[3 + idx];
1236: xw[4] = x[4 + idx];
1237: xw[5] = x[5 + idx];
1238: xw[6] = x[6 + idx];
1239: PetscKernel_v_gets_v_minus_A_times_w_7(s, v, xw);
1240: v += 49;
1241: }
1242: PetscKernel_v_gets_A_times_w_7(xw, idiag, s);
1243: x[i2] += xw[0];
1244: x[i2 + 1] += xw[1];
1245: x[i2 + 2] += xw[2];
1246: x[i2 + 3] += xw[3];
1247: x[i2 + 4] += xw[4];
1248: x[i2 + 5] += xw[5];
1249: x[i2 + 6] += xw[6];
1250: idiag -= 49;
1251: i2 -= 7;
1252: }
1253: break;
1254: default:
1255: for (i = m - 1; i >= 0; i--) {
1256: v = aa + bs2 * ai[i];
1257: vi = aj + ai[i];
1258: nz = ai[i + 1] - ai[i];
1260: PetscCall(PetscArraycpy(w, b + i2, bs));
1261: /* copy all rows of x that are needed into contiguous space */
1262: workt = work;
1263: for (j = 0; j < nz; j++) {
1264: PetscCall(PetscArraycpy(workt, x + bs * (*vi++), bs));
1265: workt += bs;
1266: }
1267: PetscKernel_w_gets_w_minus_Ar_times_v(bs, bs * nz, w, v, work);
1268: PetscKernel_w_gets_w_plus_Ar_times_v(bs, bs, w, idiag, x + i2);
1270: idiag -= bs2;
1271: i2 -= bs;
1272: }
1273: break;
1274: }
1275: PetscCall(PetscLogFlops(2.0 * bs2 * (a->nz)));
1276: }
1277: }
1278: PetscCall(VecRestoreArray(xx, &x));
1279: PetscCall(VecRestoreArrayRead(bb, &b));
1280: PetscFunctionReturn(PETSC_SUCCESS);
1281: }
1283: /*
1284: Special version for direct calls from Fortran (Used in PETSc-fun3d)
1285: */
1286: #if PetscDefined(HAVE_FORTRAN_CAPS)
1287: #define matsetvaluesblocked4_ MATSETVALUESBLOCKED4
1288: #elif !PetscDefined(HAVE_FORTRAN_UNDERSCORE)
1289: #define matsetvaluesblocked4_ matsetvaluesblocked4
1290: #endif
1292: PETSC_EXTERN void matsetvaluesblocked4_(Mat *AA, PetscInt *mm, const PetscInt im[], PetscInt *nn, const PetscInt in[], const PetscScalar v[])
1293: {
1294: Mat A = *AA;
1295: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
1296: PetscInt *rp, k, low, high, t, ii, jj, row, nrow, i, col, l, N, m = *mm, n = *nn;
1297: PetscInt *ai = a->i, *ailen = a->ilen;
1298: PetscInt *aj = a->j, stepval, lastcol = -1;
1299: const PetscScalar *value = v;
1300: MatScalar *ap, *aa = a->a, *bap;
1302: PetscFunctionBegin;
1303: if (A->rmap->bs != 4) SETERRABORT(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Can only be called with a block size of 4");
1304: stepval = (n - 1) * 4;
1305: for (k = 0; k < m; k++) { /* loop over added rows */
1306: row = im[k];
1307: rp = aj + ai[row];
1308: ap = aa + 16 * ai[row];
1309: nrow = ailen[row];
1310: low = 0;
1311: high = nrow;
1312: for (l = 0; l < n; l++) { /* loop over added columns */
1313: col = in[l];
1314: if (col <= lastcol) low = 0;
1315: else high = nrow;
1316: lastcol = col;
1317: value = v + k * (stepval + 4 + l) * 4;
1318: while (high - low > 7) {
1319: t = (low + high) / 2;
1320: if (rp[t] > col) high = t;
1321: else low = t;
1322: }
1323: for (i = low; i < high; i++) {
1324: if (rp[i] > col) break;
1325: if (rp[i] == col) {
1326: bap = ap + 16 * i;
1327: for (ii = 0; ii < 4; ii++, value += stepval) {
1328: for (jj = ii; jj < 16; jj += 4) bap[jj] += *value++;
1329: }
1330: goto noinsert2;
1331: }
1332: }
1333: N = nrow++ - 1;
1334: high++; /* added new column index thus must search to one higher than before */
1335: /* shift up all the later entries in this row */
1336: for (ii = N; ii >= i; ii--) {
1337: rp[ii + 1] = rp[ii];
1338: PetscCallVoid(PetscArraycpy(ap + 16 * (ii + 1), ap + 16 * (ii), 16));
1339: }
1340: if (N >= i) PetscCallVoid(PetscArrayzero(ap + 16 * i, 16));
1341: rp[i] = col;
1342: bap = ap + 16 * i;
1343: for (ii = 0; ii < 4; ii++, value += stepval) {
1344: for (jj = ii; jj < 16; jj += 4) bap[jj] = *value++;
1345: }
1346: noinsert2:;
1347: low = i;
1348: }
1349: ailen[row] = nrow;
1350: }
1351: PetscFunctionReturnVoid();
1352: }
1354: #if PetscDefined(HAVE_FORTRAN_CAPS)
1355: #define matsetvalues4_ MATSETVALUES4
1356: #elif !PetscDefined(HAVE_FORTRAN_UNDERSCORE)
1357: #define matsetvalues4_ matsetvalues4
1358: #endif
1360: PETSC_EXTERN void matsetvalues4_(Mat *AA, PetscInt *mm, PetscInt *im, PetscInt *nn, PetscInt *in, PetscScalar *v)
1361: {
1362: Mat A = *AA;
1363: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
1364: PetscInt *rp, k, low, high, t, row, nrow, i, col, l, N, n = *nn, m = *mm;
1365: PetscInt *ai = a->i, *ailen = a->ilen;
1366: PetscInt *aj = a->j, brow, bcol;
1367: PetscInt ridx, cidx, lastcol = -1;
1368: MatScalar *ap, value, *aa = a->a, *bap;
1370: PetscFunctionBegin;
1371: for (k = 0; k < m; k++) { /* loop over added rows */
1372: row = im[k];
1373: brow = row / 4;
1374: rp = aj + ai[brow];
1375: ap = aa + 16 * ai[brow];
1376: nrow = ailen[brow];
1377: low = 0;
1378: high = nrow;
1379: for (l = 0; l < n; l++) { /* loop over added columns */
1380: col = in[l];
1381: bcol = col / 4;
1382: ridx = row % 4;
1383: cidx = col % 4;
1384: value = v[l + k * n];
1385: if (col <= lastcol) low = 0;
1386: else high = nrow;
1387: lastcol = col;
1388: while (high - low > 7) {
1389: t = (low + high) / 2;
1390: if (rp[t] > bcol) high = t;
1391: else low = t;
1392: }
1393: for (i = low; i < high; i++) {
1394: if (rp[i] > bcol) break;
1395: if (rp[i] == bcol) {
1396: bap = ap + 16 * i + 4 * cidx + ridx;
1397: *bap += value;
1398: goto noinsert1;
1399: }
1400: }
1401: N = nrow++ - 1;
1402: high++; /* added new column thus must search to one higher than before */
1403: /* shift up all the later entries in this row */
1404: PetscCallVoid(PetscArraymove(rp + i + 1, rp + i, N - i + 1));
1405: PetscCallVoid(PetscArraymove(ap + 16 * i + 16, ap + 16 * i, 16 * (N - i + 1)));
1406: PetscCallVoid(PetscArrayzero(ap + 16 * i, 16));
1407: rp[i] = bcol;
1408: ap[16 * i + 4 * cidx + ridx] = value;
1409: noinsert1:;
1410: low = i;
1411: }
1412: ailen[brow] = nrow;
1413: }
1414: PetscFunctionReturnVoid();
1415: }
1417: static PetscErrorCode MatGetRowIJ_SeqBAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool blockcompressed, PetscInt *nn, const PetscInt *inia[], const PetscInt *inja[], PetscBool *done)
1418: {
1419: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
1420: PetscInt i, j, n = a->mbs, nz = a->i[n], *tia, *tja, bs = A->rmap->bs, k, l, cnt;
1421: PetscInt **ia = (PetscInt **)inia, **ja = (PetscInt **)inja;
1423: PetscFunctionBegin;
1424: *nn = n;
1425: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
1426: if (symmetric) {
1427: PetscCall(MatToSymmetricIJ_SeqAIJ(n, a->i, a->j, PETSC_TRUE, 0, 0, &tia, &tja));
1428: nz = tia[n];
1429: } else {
1430: tia = a->i;
1431: tja = a->j;
1432: }
1434: if (!blockcompressed && bs > 1) {
1435: (*nn) *= bs;
1436: /* malloc & create the natural set of indices */
1437: PetscCall(PetscMalloc1((n + 1) * bs, ia));
1438: if (n) {
1439: (*ia)[0] = oshift;
1440: for (j = 1; j < bs; j++) (*ia)[j] = (tia[1] - tia[0]) * bs + (*ia)[j - 1];
1441: }
1443: for (i = 1; i < n; i++) {
1444: (*ia)[i * bs] = (tia[i] - tia[i - 1]) * bs + (*ia)[i * bs - 1];
1445: for (j = 1; j < bs; j++) (*ia)[i * bs + j] = (tia[i + 1] - tia[i]) * bs + (*ia)[i * bs + j - 1];
1446: }
1447: if (n) (*ia)[n * bs] = (tia[n] - tia[n - 1]) * bs + (*ia)[n * bs - 1];
1449: if (inja) {
1450: PetscCall(PetscMalloc1(nz * bs * bs, ja));
1451: cnt = 0;
1452: for (i = 0; i < n; i++) {
1453: for (j = 0; j < bs; j++) {
1454: for (k = tia[i]; k < tia[i + 1]; k++) {
1455: for (l = 0; l < bs; l++) (*ja)[cnt++] = bs * tja[k] + l;
1456: }
1457: }
1458: }
1459: }
1461: if (symmetric) { /* deallocate memory allocated in MatToSymmetricIJ_SeqAIJ() */
1462: PetscCall(PetscFree(tia));
1463: PetscCall(PetscFree(tja));
1464: }
1465: } else if (oshift == 1) {
1466: if (symmetric) {
1467: nz = tia[A->rmap->n / bs];
1468: /* add 1 to i and j indices */
1469: for (i = 0; i < A->rmap->n / bs + 1; i++) tia[i] = tia[i] + 1;
1470: *ia = tia;
1471: if (ja) {
1472: for (i = 0; i < nz; i++) tja[i] = tja[i] + 1;
1473: *ja = tja;
1474: }
1475: } else {
1476: nz = a->i[A->rmap->n / bs];
1477: /* malloc space and add 1 to i and j indices */
1478: PetscCall(PetscMalloc1(A->rmap->n / bs + 1, ia));
1479: for (i = 0; i < A->rmap->n / bs + 1; i++) (*ia)[i] = a->i[i] + 1;
1480: if (ja) {
1481: PetscCall(PetscMalloc1(nz, ja));
1482: for (i = 0; i < nz; i++) (*ja)[i] = a->j[i] + 1;
1483: }
1484: }
1485: } else {
1486: *ia = tia;
1487: if (ja) *ja = tja;
1488: }
1489: PetscFunctionReturn(PETSC_SUCCESS);
1490: }
1492: static PetscErrorCode MatRestoreRowIJ_SeqBAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool blockcompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
1493: {
1494: PetscFunctionBegin;
1495: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
1496: if ((!blockcompressed && A->rmap->bs > 1) || (symmetric || oshift == 1)) {
1497: PetscCall(PetscFree(*ia));
1498: if (ja) PetscCall(PetscFree(*ja));
1499: }
1500: PetscFunctionReturn(PETSC_SUCCESS);
1501: }
1503: PetscErrorCode MatDestroy_SeqBAIJ(Mat A)
1504: {
1505: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
1507: PetscFunctionBegin;
1508: if (A->hash_active) {
1509: PetscInt bs;
1510: A->ops[0] = a->cops;
1511: PetscCall(PetscHMapIJVDestroy(&a->ht));
1512: PetscCall(MatGetBlockSize(A, &bs));
1513: if (bs > 1) PetscCall(PetscHSetIJDestroy(&a->bht));
1514: PetscCall(PetscFree(a->dnz));
1515: PetscCall(PetscFree(a->bdnz));
1516: A->hash_active = PETSC_FALSE;
1517: }
1518: PetscCall(PetscLogObjectState((PetscObject)A, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT ", NZ=%" PetscInt_FMT, A->rmap->N, A->cmap->n, a->nz));
1519: PetscCall(MatSeqXAIJFreeAIJ(A, &a->a, &a->j, &a->i));
1520: PetscCall(ISDestroy(&a->row));
1521: PetscCall(ISDestroy(&a->col));
1522: PetscCall(PetscFree(a->diag));
1523: PetscCall(PetscFree(a->idiag));
1524: if (a->free_imax_ilen) PetscCall(PetscFree2(a->imax, a->ilen));
1525: PetscCall(PetscFree(a->solve_work));
1526: PetscCall(PetscFree(a->mult_work));
1527: PetscCall(PetscFree(a->sor_workt));
1528: PetscCall(PetscFree(a->sor_work));
1529: PetscCall(ISDestroy(&a->icol));
1530: PetscCall(PetscFree(a->saved_values));
1531: PetscCall(PetscFree2(a->compressedrow.i, a->compressedrow.rindex));
1533: PetscCall(MatDestroy(&a->sbaijMat));
1534: PetscCall(MatDestroy(&a->parent));
1535: PetscCall(PetscFree(A->data));
1537: PetscCall(PetscObjectChangeTypeName((PetscObject)A, NULL));
1538: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqBAIJGetArray_C", NULL));
1539: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqBAIJRestoreArray_C", NULL));
1540: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatStoreValues_C", NULL));
1541: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatRetrieveValues_C", NULL));
1542: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqBAIJSetColumnIndices_C", NULL));
1543: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqbaij_seqaij_C", NULL));
1544: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqbaij_seqsbaij_C", NULL));
1545: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqBAIJSetPreallocation_C", NULL));
1546: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqBAIJSetPreallocationCSR_C", NULL));
1547: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqbaij_seqbstrm_C", NULL));
1548: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsTranspose_C", NULL));
1549: #if PetscDefined(HAVE_HYPRE)
1550: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqbaij_hypre_C", NULL));
1551: #endif
1552: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqbaij_is_C", NULL));
1553: #if PetscDefined(HAVE_LIBXSMM)
1554: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqbaij_seqbaijlibxsmm_C", NULL));
1555: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqbaijlibxsmm_seqdense_C", NULL));
1556: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqbaijlibxsmm_seqbaij_C", NULL));
1557: #endif
1558: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
1559: PetscFunctionReturn(PETSC_SUCCESS);
1560: }
1562: static PetscErrorCode MatSetOption_SeqBAIJ(Mat A, MatOption op, PetscBool flg)
1563: {
1564: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
1566: PetscFunctionBegin;
1567: switch (op) {
1568: case MAT_ROW_ORIENTED:
1569: a->roworiented = flg;
1570: break;
1571: case MAT_KEEP_NONZERO_PATTERN:
1572: a->keepnonzeropattern = flg;
1573: break;
1574: case MAT_NEW_NONZERO_LOCATIONS:
1575: a->nonew = (flg ? 0 : 1);
1576: break;
1577: case MAT_NEW_NONZERO_LOCATION_ERR:
1578: a->nonew = (flg ? -1 : 0);
1579: break;
1580: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1581: a->nonew = (flg ? -2 : 0);
1582: break;
1583: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1584: a->nounused = (flg ? -1 : 0);
1585: break;
1586: default:
1587: break;
1588: }
1589: PetscFunctionReturn(PETSC_SUCCESS);
1590: }
1592: /* used for both SeqBAIJ and SeqSBAIJ matrices */
1593: PetscErrorCode MatGetRow_SeqBAIJ_private(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v, PetscInt *ai, PetscInt *aj, PetscScalar *aa)
1594: {
1595: PetscInt itmp, i, j, k, M, bn, bp, *idx_i, bs, bs2;
1596: MatScalar *aa_i;
1597: PetscScalar *v_i;
1599: PetscFunctionBegin;
1600: bs = A->rmap->bs;
1601: bs2 = bs * bs;
1602: PetscCheck(row >= 0 && row < A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %" PetscInt_FMT " out of range", row);
1604: bn = row / bs; /* Block number */
1605: bp = row % bs; /* Block Position */
1606: M = ai[bn + 1] - ai[bn];
1607: *nz = bs * M;
1609: if (v) {
1610: *v = NULL;
1611: if (*nz) {
1612: PetscCall(PetscMalloc1(*nz, v));
1613: for (i = 0; i < M; i++) { /* for each block in the block row */
1614: v_i = *v + i * bs;
1615: aa_i = aa + bs2 * (ai[bn] + i);
1616: for (j = bp, k = 0; j < bs2; j += bs, k++) v_i[k] = aa_i[j];
1617: }
1618: }
1619: }
1621: if (idx) {
1622: *idx = NULL;
1623: if (*nz) {
1624: PetscCall(PetscMalloc1(*nz, idx));
1625: for (i = 0; i < M; i++) { /* for each block in the block row */
1626: idx_i = *idx + i * bs;
1627: itmp = bs * aj[ai[bn] + i];
1628: for (j = 0; j < bs; j++) idx_i[j] = itmp++;
1629: }
1630: }
1631: }
1632: PetscFunctionReturn(PETSC_SUCCESS);
1633: }
1635: PetscErrorCode MatGetRow_SeqBAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1636: {
1637: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
1639: PetscFunctionBegin;
1640: PetscCall(MatGetRow_SeqBAIJ_private(A, row, nz, idx, v, a->i, a->j, a->a));
1641: PetscFunctionReturn(PETSC_SUCCESS);
1642: }
1644: PetscErrorCode MatRestoreRow_SeqBAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1645: {
1646: PetscFunctionBegin;
1647: if (idx) PetscCall(PetscFree(*idx));
1648: if (v) PetscCall(PetscFree(*v));
1649: PetscFunctionReturn(PETSC_SUCCESS);
1650: }
1652: static PetscErrorCode MatTranspose_SeqBAIJ(Mat A, MatReuse reuse, Mat *B)
1653: {
1654: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data, *at;
1655: Mat C;
1656: PetscInt i, j, k, *aj = a->j, *ai = a->i, bs = A->rmap->bs, mbs = a->mbs, nbs = a->nbs, *atfill;
1657: PetscInt bs2 = a->bs2, *ati, *atj, anzj, kr;
1658: MatScalar *ata, *aa = a->a;
1660: PetscFunctionBegin;
1661: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *B));
1662: PetscCall(PetscCalloc1(1 + nbs, &atfill));
1663: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_INPLACE_MATRIX) {
1664: for (i = 0; i < ai[mbs]; i++) atfill[aj[i]] += 1; /* count num of non-zeros in row aj[i] */
1666: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
1667: PetscCall(MatSetSizes(C, A->cmap->n, A->rmap->N, A->cmap->n, A->rmap->N));
1668: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
1669: PetscCall(MatSeqBAIJSetPreallocation(C, bs, 0, atfill));
1671: at = (Mat_SeqBAIJ *)C->data;
1672: ati = at->i;
1673: for (i = 0; i < nbs; i++) at->ilen[i] = at->imax[i] = ati[i + 1] - ati[i];
1674: } else {
1675: C = *B;
1676: at = (Mat_SeqBAIJ *)C->data;
1677: ati = at->i;
1678: }
1680: atj = at->j;
1681: ata = at->a;
1683: /* Copy ati into atfill so we have locations of the next free space in atj */
1684: PetscCall(PetscArraycpy(atfill, ati, nbs));
1686: /* Walk through A row-wise and mark nonzero entries of A^T. */
1687: for (i = 0; i < mbs; i++) {
1688: anzj = ai[i + 1] - ai[i];
1689: for (j = 0; j < anzj; j++) {
1690: atj[atfill[*aj]] = i;
1691: for (kr = 0; kr < bs; kr++) {
1692: for (k = 0; k < bs; k++) ata[bs2 * atfill[*aj] + k * bs + kr] = *aa++;
1693: }
1694: atfill[*aj++] += 1;
1695: }
1696: }
1697: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
1698: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
1700: /* Clean up temporary space and complete requests. */
1701: PetscCall(PetscFree(atfill));
1703: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1704: PetscCall(MatSetBlockSizes(C, A->cmap->bs, A->rmap->bs));
1705: *B = C;
1706: } else {
1707: PetscCall(MatHeaderMerge(A, &C));
1708: }
1709: PetscFunctionReturn(PETSC_SUCCESS);
1710: }
1712: static PetscErrorCode MatCompare_SeqBAIJ_Private(Mat A, Mat B, PetscReal tol, PetscBool *flg)
1713: {
1714: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data, *b = (Mat_SeqBAIJ *)B->data;
1716: PetscFunctionBegin;
1717: /* If the matrix/block dimensions are not equal, or no of nonzeros or shift */
1718: if (A->rmap->N != B->rmap->N || A->cmap->n != B->cmap->n || A->rmap->bs != B->rmap->bs || a->nz != b->nz) {
1719: *flg = PETSC_FALSE;
1720: PetscFunctionReturn(PETSC_SUCCESS);
1721: }
1723: /* if the a->i are the same */
1724: PetscCall(PetscArraycmp(a->i, b->i, a->mbs + 1, flg));
1725: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
1727: /* if a->j are the same */
1728: PetscCall(PetscArraycmp(a->j, b->j, a->nz, flg));
1729: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
1731: if (tol == 0.0) PetscCall(PetscArraycmp(a->a, b->a, a->nz * A->rmap->bs * A->rmap->bs, flg)); /* if a->a are the same */
1732: else {
1733: *flg = PETSC_TRUE;
1734: for (PetscInt i = 0; (i < a->nz * A->rmap->bs * A->rmap->bs) && *flg; ++i)
1735: if (PetscAbsScalar(a->a[i] - b->a[i]) > tol) *flg = PETSC_FALSE;
1736: }
1737: PetscFunctionReturn(PETSC_SUCCESS);
1738: }
1740: static PetscErrorCode MatIsTranspose_SeqBAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
1741: {
1742: Mat Btrans;
1744: PetscFunctionBegin;
1745: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &Btrans));
1746: PetscCall(MatCompare_SeqBAIJ_Private(A, Btrans, tol, f));
1747: PetscCall(MatDestroy(&Btrans));
1748: PetscFunctionReturn(PETSC_SUCCESS);
1749: }
1751: static PetscErrorCode MatEqual_SeqBAIJ(Mat A, Mat B, PetscBool *flg)
1752: {
1753: PetscFunctionBegin;
1754: PetscCall(MatCompare_SeqBAIJ_Private(A, B, 0.0, flg));
1755: PetscFunctionReturn(PETSC_SUCCESS);
1756: }
1758: /* Used for both SeqBAIJ and SeqSBAIJ matrices */
1759: PetscErrorCode MatView_SeqBAIJ_Binary(Mat mat, PetscViewer viewer)
1760: {
1761: Mat_SeqBAIJ *A = (Mat_SeqBAIJ *)mat->data;
1762: PetscInt header[4], M, N, m, bs, nz, cnt, i, j, k, l;
1763: PetscInt *rowlens, *colidxs;
1764: PetscScalar *matvals;
1766: PetscFunctionBegin;
1767: PetscCall(PetscViewerSetUp(viewer));
1769: M = mat->rmap->N;
1770: N = mat->cmap->N;
1771: m = mat->rmap->n;
1772: bs = mat->rmap->bs;
1773: nz = bs * bs * A->nz;
1775: /* write matrix header */
1776: header[0] = MAT_FILE_CLASSID;
1777: header[1] = M;
1778: header[2] = N;
1779: header[3] = nz;
1780: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1782: /* store row lengths */
1783: PetscCall(PetscMalloc1(m, &rowlens));
1784: for (cnt = 0, i = 0; i < A->mbs; i++)
1785: for (j = 0; j < bs; j++) rowlens[cnt++] = bs * (A->i[i + 1] - A->i[i]);
1786: PetscCall(PetscViewerBinaryWrite(viewer, rowlens, m, PETSC_INT));
1787: PetscCall(PetscFree(rowlens));
1789: /* store column indices */
1790: PetscCall(PetscMalloc1(nz, &colidxs));
1791: for (cnt = 0, i = 0; i < A->mbs; i++)
1792: for (k = 0; k < bs; k++)
1793: for (j = A->i[i]; j < A->i[i + 1]; j++)
1794: for (l = 0; l < bs; l++) colidxs[cnt++] = bs * A->j[j] + l;
1795: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt_FMT, cnt, nz);
1796: PetscCall(PetscViewerBinaryWrite(viewer, colidxs, nz, PETSC_INT));
1797: PetscCall(PetscFree(colidxs));
1799: /* store nonzero values */
1800: PetscCall(PetscMalloc1(nz, &matvals));
1801: for (cnt = 0, i = 0; i < A->mbs; i++)
1802: for (k = 0; k < bs; k++)
1803: for (j = A->i[i]; j < A->i[i + 1]; j++)
1804: for (l = 0; l < bs; l++) matvals[cnt++] = A->a[bs * (bs * j + l) + k];
1805: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt_FMT, cnt, nz);
1806: PetscCall(PetscViewerBinaryWrite(viewer, matvals, nz, PETSC_SCALAR));
1807: PetscCall(PetscFree(matvals));
1809: /* write block size option to the viewer's .info file */
1810: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1811: PetscFunctionReturn(PETSC_SUCCESS);
1812: }
1814: static PetscErrorCode MatView_SeqBAIJ_ASCII_structonly(Mat A, PetscViewer viewer)
1815: {
1816: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
1817: PetscInt i, bs = A->rmap->bs, k;
1819: PetscFunctionBegin;
1820: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
1821: for (i = 0; i < a->mbs; i++) {
1822: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT "-%" PetscInt_FMT ":", i * bs, i * bs + bs - 1));
1823: for (k = a->i[i]; k < a->i[i + 1]; k++) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT "-%" PetscInt_FMT ") ", bs * a->j[k], bs * a->j[k] + bs - 1));
1824: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
1825: }
1826: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
1827: PetscFunctionReturn(PETSC_SUCCESS);
1828: }
1830: static PetscErrorCode MatView_SeqBAIJ_ASCII(Mat A, PetscViewer viewer)
1831: {
1832: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
1833: PetscInt i, j, bs = A->rmap->bs, k, l, bs2 = a->bs2;
1834: PetscViewerFormat format;
1836: PetscFunctionBegin;
1837: if (A->structure_only) {
1838: PetscCall(MatView_SeqBAIJ_ASCII_structonly(A, viewer));
1839: PetscFunctionReturn(PETSC_SUCCESS);
1840: }
1842: PetscCall(PetscViewerGetFormat(viewer, &format));
1843: if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1844: } else if (format == PETSC_VIEWER_ASCII_MATLAB) {
1845: const char *matname;
1846: Mat aij;
1847: PetscCall(MatConvert(A, MATSEQAIJ, MAT_INITIAL_MATRIX, &aij));
1848: PetscCall(PetscObjectGetName((PetscObject)A, &matname));
1849: PetscCall(PetscObjectSetName((PetscObject)aij, matname));
1850: PetscCall(MatView(aij, viewer));
1851: PetscCall(MatDestroy(&aij));
1852: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1853: PetscFunctionReturn(PETSC_SUCCESS);
1854: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
1855: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
1856: for (i = 0; i < a->mbs; i++) {
1857: for (j = 0; j < bs; j++) {
1858: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i * bs + j));
1859: for (k = a->i[i]; k < a->i[i + 1]; k++) {
1860: for (l = 0; l < bs; l++) {
1861: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[bs2 * k + l * bs + j]) > 0.0 && PetscRealPart(a->a[bs2 * k + l * bs + j]) != 0.0) {
1862: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %gi) ", bs * a->j[k] + l, (double)PetscRealPart(a->a[bs2 * k + l * bs + j]), (double)PetscImaginaryPart(a->a[bs2 * k + l * bs + j])));
1863: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[bs2 * k + l * bs + j]) < 0.0 && PetscRealPart(a->a[bs2 * k + l * bs + j]) != 0.0) {
1864: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %gi) ", bs * a->j[k] + l, (double)PetscRealPart(a->a[bs2 * k + l * bs + j]), -(double)PetscImaginaryPart(a->a[bs2 * k + l * bs + j])));
1865: } else if (PetscRealPart(a->a[bs2 * k + l * bs + j]) != 0.0) {
1866: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", bs * a->j[k] + l, (double)PetscRealPart(a->a[bs2 * k + l * bs + j])));
1867: }
1868: }
1869: }
1870: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
1871: }
1872: }
1873: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
1874: } else {
1875: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
1876: for (i = 0; i < a->mbs; i++) {
1877: for (j = 0; j < bs; j++) {
1878: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i * bs + j));
1879: for (k = a->i[i]; k < a->i[i + 1]; k++) {
1880: for (l = 0; l < bs; l++) {
1881: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[bs2 * k + l * bs + j]) > 0.0) {
1882: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i) ", bs * a->j[k] + l, (double)PetscRealPart(a->a[bs2 * k + l * bs + j]), (double)PetscImaginaryPart(a->a[bs2 * k + l * bs + j])));
1883: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[bs2 * k + l * bs + j]) < 0.0) {
1884: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i) ", bs * a->j[k] + l, (double)PetscRealPart(a->a[bs2 * k + l * bs + j]), -(double)PetscImaginaryPart(a->a[bs2 * k + l * bs + j])));
1885: } else {
1886: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", bs * a->j[k] + l, (double)PetscRealPart(a->a[bs2 * k + l * bs + j])));
1887: }
1888: }
1889: }
1890: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
1891: }
1892: }
1893: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
1894: }
1895: PetscCall(PetscViewerFlush(viewer));
1896: PetscFunctionReturn(PETSC_SUCCESS);
1897: }
1899: #include <petscdraw.h>
1900: static PetscErrorCode MatView_SeqBAIJ_Draw_Zoom(PetscDraw draw, void *Aa)
1901: {
1902: Mat A = (Mat)Aa;
1903: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
1904: PetscInt row, i, j, k, l, mbs = a->mbs, bs = A->rmap->bs, bs2 = a->bs2;
1905: PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r;
1906: MatScalar *aa;
1907: PetscViewer viewer;
1908: PetscViewerFormat format;
1909: int color;
1911: PetscFunctionBegin;
1912: PetscCall(PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer));
1913: PetscCall(PetscViewerGetFormat(viewer, &format));
1914: PetscCall(PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr));
1916: /* loop over matrix elements drawing boxes */
1918: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
1919: PetscDrawCollectiveBegin(draw);
1920: /* Blue for negative, Cyan for zero and Red for positive */
1921: color = PETSC_DRAW_BLUE;
1922: for (i = 0, row = 0; i < mbs; i++, row += bs) {
1923: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1924: y_l = A->rmap->N - row - 1.0;
1925: y_r = y_l + 1.0;
1926: x_l = a->j[j] * bs;
1927: x_r = x_l + 1.0;
1928: aa = a->a + j * bs2;
1929: for (k = 0; k < bs; k++) {
1930: for (l = 0; l < bs; l++) {
1931: if (PetscRealPart(*aa++) >= 0.) continue;
1932: PetscCall(PetscDrawRectangle(draw, x_l + k, y_l - l, x_r + k, y_r - l, color, color, color, color));
1933: }
1934: }
1935: }
1936: }
1937: color = PETSC_DRAW_CYAN;
1938: for (i = 0, row = 0; i < mbs; i++, row += bs) {
1939: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1940: y_l = A->rmap->N - row - 1.0;
1941: y_r = y_l + 1.0;
1942: x_l = a->j[j] * bs;
1943: x_r = x_l + 1.0;
1944: aa = a->a + j * bs2;
1945: for (k = 0; k < bs; k++) {
1946: for (l = 0; l < bs; l++) {
1947: if (PetscRealPart(*aa++) != 0.) continue;
1948: PetscCall(PetscDrawRectangle(draw, x_l + k, y_l - l, x_r + k, y_r - l, color, color, color, color));
1949: }
1950: }
1951: }
1952: }
1953: color = PETSC_DRAW_RED;
1954: for (i = 0, row = 0; i < mbs; i++, row += bs) {
1955: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1956: y_l = A->rmap->N - row - 1.0;
1957: y_r = y_l + 1.0;
1958: x_l = a->j[j] * bs;
1959: x_r = x_l + 1.0;
1960: aa = a->a + j * bs2;
1961: for (k = 0; k < bs; k++) {
1962: for (l = 0; l < bs; l++) {
1963: if (PetscRealPart(*aa++) <= 0.) continue;
1964: PetscCall(PetscDrawRectangle(draw, x_l + k, y_l - l, x_r + k, y_r - l, color, color, color, color));
1965: }
1966: }
1967: }
1968: }
1969: PetscDrawCollectiveEnd(draw);
1970: } else {
1971: /* use contour shading to indicate magnitude of values */
1972: /* first determine max of all nonzero values */
1973: PetscReal minv = 0.0, maxv = 0.0;
1974: PetscDraw popup;
1976: for (i = 0; i < a->nz * a->bs2; i++) {
1977: if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]);
1978: }
1979: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1980: PetscCall(PetscDrawGetPopup(draw, &popup));
1981: PetscCall(PetscDrawScalePopup(popup, 0.0, maxv));
1983: PetscDrawCollectiveBegin(draw);
1984: for (i = 0, row = 0; i < mbs; i++, row += bs) {
1985: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1986: y_l = A->rmap->N - row - 1.0;
1987: y_r = y_l + 1.0;
1988: x_l = a->j[j] * bs;
1989: x_r = x_l + 1.0;
1990: aa = a->a + j * bs2;
1991: for (k = 0; k < bs; k++) {
1992: for (l = 0; l < bs; l++) {
1993: MatScalar v = *aa++;
1994: color = PetscDrawRealToColor(PetscAbsScalar(v), minv, maxv);
1995: PetscCall(PetscDrawRectangle(draw, x_l + k, y_l - l, x_r + k, y_r - l, color, color, color, color));
1996: }
1997: }
1998: }
1999: }
2000: PetscDrawCollectiveEnd(draw);
2001: }
2002: PetscFunctionReturn(PETSC_SUCCESS);
2003: }
2005: static PetscErrorCode MatView_SeqBAIJ_Draw(Mat A, PetscViewer viewer)
2006: {
2007: PetscReal xl, yl, xr, yr, w, h;
2008: PetscDraw draw;
2009: PetscBool isnull;
2011: PetscFunctionBegin;
2012: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
2013: PetscCall(PetscDrawIsNull(draw, &isnull));
2014: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
2016: xr = A->cmap->n;
2017: yr = A->rmap->N;
2018: h = yr / 10.0;
2019: w = xr / 10.0;
2020: xr += w;
2021: yr += h;
2022: xl = -w;
2023: yl = -h;
2024: PetscCall(PetscDrawSetCoordinates(draw, xl, yl, xr, yr));
2025: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer));
2026: PetscCall(PetscDrawZoom(draw, MatView_SeqBAIJ_Draw_Zoom, A));
2027: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL));
2028: PetscCall(PetscDrawSave(draw));
2029: PetscFunctionReturn(PETSC_SUCCESS);
2030: }
2032: PetscErrorCode MatView_SeqBAIJ(Mat A, PetscViewer viewer)
2033: {
2034: PetscBool isascii, isbinary, isdraw;
2036: PetscFunctionBegin;
2037: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
2038: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
2039: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
2040: if (isascii) {
2041: PetscCall(MatView_SeqBAIJ_ASCII(A, viewer));
2042: } else if (isbinary) {
2043: PetscCall(MatView_SeqBAIJ_Binary(A, viewer));
2044: } else if (isdraw) {
2045: PetscCall(MatView_SeqBAIJ_Draw(A, viewer));
2046: } else {
2047: Mat B;
2048: PetscCall(MatConvert(A, MATSEQAIJ, MAT_INITIAL_MATRIX, &B));
2049: PetscCall(MatView(B, viewer));
2050: PetscCall(MatDestroy(&B));
2051: }
2052: PetscFunctionReturn(PETSC_SUCCESS);
2053: }
2055: PetscErrorCode MatGetValues_SeqBAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], PetscScalar v[])
2056: {
2057: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2058: PetscInt *rp, k, low, high, t, row, nrow, i, col, l, *aj = a->j;
2059: PetscInt *ai = a->i, *ailen = a->ilen;
2060: PetscInt brow, bcol, ridx, cidx, bs = A->rmap->bs, bs2 = a->bs2;
2061: MatScalar *ap, *aa = a->a;
2062: PetscBool roworiented = a->roworiented;
2063: PetscScalar *value;
2065: PetscFunctionBegin;
2066: for (k = 0; k < m; k++) { /* loop over rows */
2067: row = im[k];
2068: if (row < 0) continue; /* negative row */
2069: brow = row / bs;
2070: PetscCheck(row < A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %" PetscInt_FMT " too large", row);
2071: rp = PetscSafePointerPlusOffset(aj, ai[brow]);
2072: ap = PetscSafePointerPlusOffset(aa, bs2 * ai[brow]);
2073: nrow = ailen[brow];
2074: for (l = 0; l < n; l++) { /* loop over columns */
2075: if (in[l] < 0) continue; /* negative column */
2076: PetscCheck(in[l] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column %" PetscInt_FMT " too large", in[l]);
2077: value = roworiented ? &v[l + k * n] : &v[k + l * m];
2078: col = in[l];
2079: bcol = col / bs;
2080: cidx = col % bs;
2081: ridx = row % bs;
2082: high = nrow;
2083: low = 0; /* assume unsorted */
2084: while (high - low > 5) {
2085: t = (low + high) / 2;
2086: if (rp[t] > bcol) high = t;
2087: else low = t;
2088: }
2089: for (i = low; i < high; i++) {
2090: if (rp[i] > bcol) break;
2091: if (rp[i] == bcol) {
2092: *value = ap[bs2 * i + bs * cidx + ridx];
2093: goto finished;
2094: }
2095: }
2096: *value = 0.0;
2097: finished:;
2098: }
2099: }
2100: PetscFunctionReturn(PETSC_SUCCESS);
2101: }
2103: PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
2104: {
2105: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2106: PetscInt *rp, k, low, high, t, ii, jj, row, nrow, i, col, l, rmax, N, lastcol = -1;
2107: PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen;
2108: PetscInt *aj = a->j, nonew = a->nonew, bs2 = a->bs2, bs = A->rmap->bs, stepval;
2109: PetscBool roworiented = a->roworiented;
2110: const PetscScalar *value = v;
2111: MatScalar *ap = NULL, *aa = a->a, *bap;
2113: PetscFunctionBegin;
2114: if (roworiented) {
2115: stepval = (n - 1) * bs;
2116: } else {
2117: stepval = (m - 1) * bs;
2118: }
2119: for (k = 0; k < m; k++) { /* loop over added rows */
2120: row = im[k];
2121: if (row < 0) continue;
2122: PetscCheck(row < a->mbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Block row index too large %" PetscInt_FMT " max %" PetscInt_FMT, row, a->mbs - 1);
2123: rp = aj + ai[row];
2124: if (!A->structure_only) ap = aa + bs2 * ai[row];
2125: rmax = imax[row];
2126: nrow = ailen[row];
2127: low = 0;
2128: high = nrow;
2129: for (l = 0; l < n; l++) { /* loop over added columns */
2130: if (in[l] < 0) continue;
2131: PetscCheck(in[l] < a->nbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Block column index too large %" PetscInt_FMT " max %" PetscInt_FMT, in[l], a->nbs - 1);
2132: col = in[l];
2133: if (!A->structure_only) {
2134: if (roworiented) {
2135: value = v + (k * (stepval + bs) + l) * bs;
2136: } else {
2137: value = v + (l * (stepval + bs) + k) * bs;
2138: }
2139: }
2140: if (col <= lastcol) low = 0;
2141: else high = nrow;
2142: lastcol = col;
2143: while (high - low > 7) {
2144: t = (low + high) / 2;
2145: if (rp[t] > col) high = t;
2146: else low = t;
2147: }
2148: for (i = low; i < high; i++) {
2149: if (rp[i] > col) break;
2150: if (rp[i] == col) {
2151: if (A->structure_only) goto noinsert2;
2152: bap = ap + bs2 * i;
2153: if (roworiented) {
2154: if (is == ADD_VALUES) {
2155: for (ii = 0; ii < bs; ii++, value += stepval) {
2156: for (jj = ii; jj < bs2; jj += bs) bap[jj] += *value++;
2157: }
2158: } else {
2159: for (ii = 0; ii < bs; ii++, value += stepval) {
2160: for (jj = ii; jj < bs2; jj += bs) bap[jj] = *value++;
2161: }
2162: }
2163: } else {
2164: if (is == ADD_VALUES) {
2165: for (ii = 0; ii < bs; ii++, value += bs + stepval) {
2166: for (jj = 0; jj < bs; jj++) bap[jj] += value[jj];
2167: bap += bs;
2168: }
2169: } else {
2170: for (ii = 0; ii < bs; ii++, value += bs + stepval) {
2171: for (jj = 0; jj < bs; jj++) bap[jj] = value[jj];
2172: bap += bs;
2173: }
2174: }
2175: }
2176: goto noinsert2;
2177: }
2178: }
2179: if (nonew == 1) goto noinsert2;
2180: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new blocked index new nonzero block (%" PetscInt_FMT ", %" PetscInt_FMT ") in the matrix", row, col);
2181: if (A->structure_only) {
2182: MatSeqXAIJReallocateAIJ_structure_only(A, a->mbs, bs2, nrow, row, col, rmax, ai, aj, rp, imax, nonew, MatScalar);
2183: } else {
2184: MatSeqXAIJReallocateAIJ(A, a->mbs, bs2, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
2185: }
2186: N = nrow++ - 1;
2187: high++;
2188: /* shift up all the later entries in this row */
2189: PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1));
2190: rp[i] = col;
2191: if (!A->structure_only) {
2192: PetscCall(PetscArraymove(ap + bs2 * (i + 1), ap + bs2 * i, bs2 * (N - i + 1)));
2193: bap = ap + bs2 * i;
2194: if (roworiented) {
2195: for (ii = 0; ii < bs; ii++, value += stepval) {
2196: for (jj = ii; jj < bs2; jj += bs) bap[jj] = *value++;
2197: }
2198: } else {
2199: for (ii = 0; ii < bs; ii++, value += stepval) {
2200: for (jj = 0; jj < bs; jj++) *bap++ = *value++;
2201: }
2202: }
2203: }
2204: noinsert2:;
2205: low = i;
2206: }
2207: ailen[row] = nrow;
2208: }
2209: PetscFunctionReturn(PETSC_SUCCESS);
2210: }
2212: PetscErrorCode MatAssemblyEnd_SeqBAIJ(Mat A, MatAssemblyType mode)
2213: {
2214: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2215: PetscInt fshift = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax;
2216: PetscInt m = A->rmap->N, *ip, N, *ailen = a->ilen;
2217: PetscInt mbs = a->mbs, bs2 = a->bs2, rmax = 0;
2218: MatScalar *aa = a->a, *ap;
2219: PetscReal ratio = 0.6;
2221: PetscFunctionBegin;
2222: if (mode == MAT_FLUSH_ASSEMBLY || (A->was_assembled && A->ass_nonzerostate == A->nonzerostate)) PetscFunctionReturn(PETSC_SUCCESS);
2224: if (m) rmax = ailen[0];
2225: for (i = 1; i < mbs; i++) {
2226: /* move each row back by the amount of empty slots (fshift) before it*/
2227: fshift += imax[i - 1] - ailen[i - 1];
2228: rmax = PetscMax(rmax, ailen[i]);
2229: if (fshift) {
2230: ip = aj + ai[i];
2231: ap = aa + bs2 * ai[i];
2232: N = ailen[i];
2233: PetscCall(PetscArraymove(ip - fshift, ip, N));
2234: if (!A->structure_only) PetscCall(PetscArraymove(ap - bs2 * fshift, ap, bs2 * N));
2235: }
2236: ai[i] = ai[i - 1] + ailen[i - 1];
2237: }
2238: if (mbs) {
2239: fshift += imax[mbs - 1] - ailen[mbs - 1];
2240: ai[mbs] = ai[mbs - 1] + ailen[mbs - 1];
2241: }
2243: /* reset ilen and imax for each row */
2244: a->nonzerorowcnt = 0;
2245: if (A->structure_only) {
2246: PetscCall(PetscFree2(a->imax, a->ilen));
2247: } else { /* !A->structure_only */
2248: for (i = 0; i < mbs; i++) {
2249: ailen[i] = imax[i] = ai[i + 1] - ai[i];
2250: a->nonzerorowcnt += ((ai[i + 1] - ai[i]) > 0);
2251: }
2252: }
2253: a->nz = ai[mbs];
2255: if (fshift) PetscCheck(a->nounused != -1, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unused space detected in matrix: %" PetscInt_FMT " X %" PetscInt_FMT " block size %" PetscInt_FMT ", %" PetscInt_FMT " unneeded", m, A->cmap->n, A->rmap->bs, fshift * bs2);
2256: PetscCall(PetscInfo(A, "Matrix size: %" PetscInt_FMT " X %" PetscInt_FMT ", block size %" PetscInt_FMT "; storage space: %" PetscInt_FMT " unneeded, %" PetscInt_FMT " used\n", m, A->cmap->n, A->rmap->bs, fshift * bs2, a->nz * bs2));
2257: PetscCall(PetscInfo(A, "Number of mallocs during MatSetValues is %" PetscInt_FMT "\n", a->reallocs));
2258: PetscCall(PetscInfo(A, "Most nonzeros blocks in any row is %" PetscInt_FMT "\n", rmax));
2260: A->info.mallocs += a->reallocs;
2261: a->reallocs = 0;
2262: A->info.nz_unneeded = (PetscReal)fshift * bs2;
2263: a->rmax = rmax;
2265: if (!A->structure_only) PetscCall(MatCheckCompressedRow(A, a->nonzerorowcnt, &a->compressedrow, a->i, mbs, ratio));
2266: PetscFunctionReturn(PETSC_SUCCESS);
2267: }
2269: /*
2270: This function returns an array of flags which indicate the locations of contiguous
2271: blocks that should be zeroed. for eg: if bs = 3 and is = [0,1,2,3,5,6,7,8,9]
2272: then the resulting sizes = [3,1,1,3,1] corresponding to sets [(0,1,2),(3),(5),(6,7,8),(9)]
2273: Assume: sizes should be long enough to hold all the values.
2274: */
2275: static PetscErrorCode MatZeroRows_SeqBAIJ_Check_Blocks(PetscInt idx[], PetscInt n, PetscInt bs, PetscInt sizes[], PetscInt *bs_max)
2276: {
2277: PetscInt j = 0;
2279: PetscFunctionBegin;
2280: for (PetscInt i = 0; i < n; j++) {
2281: PetscInt row = idx[i];
2282: if (row % bs != 0) { /* Not the beginning of a block */
2283: sizes[j] = 1;
2284: i++;
2285: } else if (i + bs > n) { /* complete block doesn't exist (at idx end) */
2286: sizes[j] = 1; /* Also makes sure at least 'bs' values exist for next else */
2287: i++;
2288: } else { /* Beginning of the block, so check if the complete block exists */
2289: PetscBool flg = PETSC_TRUE;
2290: for (PetscInt k = 1; k < bs; k++) {
2291: if (row + k != idx[i + k]) { /* break in the block */
2292: flg = PETSC_FALSE;
2293: break;
2294: }
2295: }
2296: if (flg) { /* No break in the bs */
2297: sizes[j] = bs;
2298: i += bs;
2299: } else {
2300: sizes[j] = 1;
2301: i++;
2302: }
2303: }
2304: }
2305: *bs_max = j;
2306: PetscFunctionReturn(PETSC_SUCCESS);
2307: }
2309: PetscErrorCode MatZeroRows_SeqBAIJ(Mat A, PetscInt is_n, const PetscInt is_idx[], PetscScalar diag, Vec x, Vec b)
2310: {
2311: Mat_SeqBAIJ *baij = (Mat_SeqBAIJ *)A->data;
2312: PetscInt i, j, k, count, *rows;
2313: PetscInt bs = A->rmap->bs, bs2 = baij->bs2, *sizes, row, bs_max;
2314: PetscScalar zero = 0.0;
2315: MatScalar *aa;
2316: const PetscScalar *xx;
2317: PetscScalar *bb;
2319: PetscFunctionBegin;
2320: /* fix right-hand side if needed */
2321: if (x && b) {
2322: PetscCall(VecGetArrayRead(x, &xx));
2323: PetscCall(VecGetArray(b, &bb));
2324: for (i = 0; i < is_n; i++) bb[is_idx[i]] = diag * xx[is_idx[i]];
2325: PetscCall(VecRestoreArrayRead(x, &xx));
2326: PetscCall(VecRestoreArray(b, &bb));
2327: }
2329: /* Make a copy of the IS and sort it */
2330: /* allocate memory for rows,sizes */
2331: PetscCall(PetscMalloc2(is_n, &rows, 2 * is_n, &sizes));
2333: /* copy IS values to rows, and sort them */
2334: for (i = 0; i < is_n; i++) rows[i] = is_idx[i];
2335: PetscCall(PetscSortInt(is_n, rows));
2337: if (baij->keepnonzeropattern) {
2338: for (i = 0; i < is_n; i++) sizes[i] = 1;
2339: bs_max = is_n;
2340: } else {
2341: PetscCall(MatZeroRows_SeqBAIJ_Check_Blocks(rows, is_n, bs, sizes, &bs_max));
2342: A->nonzerostate++;
2343: }
2345: for (i = 0, j = 0; i < bs_max; j += sizes[i], i++) {
2346: row = rows[j];
2347: PetscCheck(row >= 0 && row <= A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", row);
2348: count = (baij->i[row / bs + 1] - baij->i[row / bs]) * bs;
2349: aa = PetscSafePointerPlusOffset(baij->a, baij->i[row / bs] * bs2 + (row % bs));
2350: if (sizes[i] == bs && !baij->keepnonzeropattern) {
2351: if (diag != (PetscScalar)0.0) {
2352: if (baij->ilen[row / bs] > 0) {
2353: baij->ilen[row / bs] = 1;
2354: baij->j[baij->i[row / bs]] = row / bs;
2356: PetscCall(PetscArrayzero(aa, count * bs));
2357: }
2358: /* Now insert all the diagonal values for this bs */
2359: for (k = 0; k < bs; k++) PetscUseTypeMethod(A, setvalues, 1, rows + j + k, 1, rows + j + k, &diag, INSERT_VALUES);
2360: } else { /* (diag == 0.0) */
2361: baij->ilen[row / bs] = 0;
2362: } /* end (diag == 0.0) */
2363: } else { /* (sizes[i] != bs) */
2364: PetscAssert(sizes[i] == 1, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal Error. Value should be 1");
2365: for (k = 0; k < count; k++) {
2366: aa[0] = zero;
2367: aa += bs;
2368: }
2369: if (diag != (PetscScalar)0.0) PetscUseTypeMethod(A, setvalues, 1, rows + j, 1, rows + j, &diag, INSERT_VALUES);
2370: }
2371: }
2373: PetscCall(PetscFree2(rows, sizes));
2374: PetscCall(MatAssemblyEnd_SeqBAIJ(A, MAT_FINAL_ASSEMBLY));
2375: PetscFunctionReturn(PETSC_SUCCESS);
2376: }
2378: static PetscErrorCode MatZeroRowsColumns_SeqBAIJ(Mat A, PetscInt is_n, const PetscInt is_idx[], PetscScalar diag, Vec x, Vec b)
2379: {
2380: Mat_SeqBAIJ *baij = (Mat_SeqBAIJ *)A->data;
2381: PetscInt i, j, k, count;
2382: PetscInt bs = A->rmap->bs, bs2 = baij->bs2, row, col;
2383: PetscScalar zero = 0.0;
2384: MatScalar *aa;
2385: const PetscScalar *xx;
2386: PetscScalar *bb;
2387: PetscBool *zeroed, vecs = PETSC_FALSE;
2389: PetscFunctionBegin;
2390: /* fix right-hand side if needed */
2391: if (x && b) {
2392: PetscCall(VecGetArrayRead(x, &xx));
2393: PetscCall(VecGetArray(b, &bb));
2394: vecs = PETSC_TRUE;
2395: }
2397: /* zero the columns */
2398: PetscCall(PetscCalloc1(A->rmap->n, &zeroed));
2399: for (i = 0; i < is_n; i++) {
2400: PetscCheck(is_idx[i] >= 0 && is_idx[i] < A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", is_idx[i]);
2401: zeroed[is_idx[i]] = PETSC_TRUE;
2402: }
2403: for (i = 0; i < A->rmap->N; i++) {
2404: if (!zeroed[i]) {
2405: row = i / bs;
2406: for (j = baij->i[row]; j < baij->i[row + 1]; j++) {
2407: for (k = 0; k < bs; k++) {
2408: col = bs * baij->j[j] + k;
2409: if (zeroed[col]) {
2410: aa = baij->a + j * bs2 + (i % bs) + bs * k;
2411: if (vecs) bb[i] -= aa[0] * xx[col];
2412: aa[0] = 0.0;
2413: }
2414: }
2415: }
2416: } else if (vecs) bb[i] = diag * xx[i];
2417: }
2418: PetscCall(PetscFree(zeroed));
2419: if (vecs) {
2420: PetscCall(VecRestoreArrayRead(x, &xx));
2421: PetscCall(VecRestoreArray(b, &bb));
2422: }
2424: /* zero the rows */
2425: for (i = 0; i < is_n; i++) {
2426: row = is_idx[i];
2427: count = (baij->i[row / bs + 1] - baij->i[row / bs]) * bs;
2428: aa = PetscSafePointerPlusOffset(baij->a, baij->i[row / bs] * bs2 + (row % bs));
2429: for (k = 0; k < count; k++) {
2430: aa[0] = zero;
2431: aa += bs;
2432: }
2433: if (diag != (PetscScalar)0.0) PetscUseTypeMethod(A, setvalues, 1, &row, 1, &row, &diag, INSERT_VALUES);
2434: }
2435: PetscCall(MatAssemblyEnd_SeqBAIJ(A, MAT_FINAL_ASSEMBLY));
2436: PetscFunctionReturn(PETSC_SUCCESS);
2437: }
2439: PetscErrorCode MatSetValues_SeqBAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
2440: {
2441: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2442: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N, lastcol = -1;
2443: PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen;
2444: PetscInt *aj = a->j, nonew = a->nonew, bs = A->rmap->bs, brow, bcol;
2445: PetscInt ridx, cidx, bs2 = a->bs2;
2446: PetscBool roworiented = a->roworiented;
2447: MatScalar *ap = NULL, value = 0.0, *aa = a->a, *bap;
2449: PetscFunctionBegin;
2450: for (k = 0; k < m; k++) { /* loop over added rows */
2451: row = im[k];
2452: brow = row / bs;
2453: if (row < 0) continue;
2454: PetscCheck(row < A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, row, A->rmap->N - 1);
2455: rp = PetscSafePointerPlusOffset(aj, ai[brow]);
2456: if (!A->structure_only) ap = PetscSafePointerPlusOffset(aa, bs2 * ai[brow]);
2457: rmax = imax[brow];
2458: nrow = ailen[brow];
2459: low = 0;
2460: high = nrow;
2461: for (l = 0; l < n; l++) { /* loop over added columns */
2462: if (in[l] < 0) continue;
2463: PetscCheck(in[l] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[l], A->cmap->n - 1);
2464: col = in[l];
2465: bcol = col / bs;
2466: ridx = row % bs;
2467: cidx = col % bs;
2468: if (!A->structure_only) {
2469: if (roworiented) {
2470: value = v[l + k * n];
2471: } else {
2472: value = v[k + l * m];
2473: }
2474: }
2475: if (col <= lastcol) low = 0;
2476: else high = nrow;
2477: lastcol = col;
2478: while (high - low > 7) {
2479: t = (low + high) / 2;
2480: if (rp[t] > bcol) high = t;
2481: else low = t;
2482: }
2483: for (i = low; i < high; i++) {
2484: if (rp[i] > bcol) break;
2485: if (rp[i] == bcol) {
2486: bap = PetscSafePointerPlusOffset(ap, bs2 * i + bs * cidx + ridx);
2487: if (!A->structure_only) {
2488: if (is == ADD_VALUES) *bap += value;
2489: else *bap = value;
2490: }
2491: goto noinsert1;
2492: }
2493: }
2494: if (nonew == 1) goto noinsert1;
2495: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero (%" PetscInt_FMT ", %" PetscInt_FMT ") in the matrix", row, col);
2496: if (A->structure_only) {
2497: MatSeqXAIJReallocateAIJ_structure_only(A, a->mbs, bs2, nrow, brow, bcol, rmax, ai, aj, rp, imax, nonew, MatScalar);
2498: } else {
2499: MatSeqXAIJReallocateAIJ(A, a->mbs, bs2, nrow, brow, bcol, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
2500: }
2501: N = nrow++ - 1;
2502: high++;
2503: /* shift up all the later entries in this row */
2504: PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1));
2505: rp[i] = bcol;
2506: if (!A->structure_only) {
2507: PetscCall(PetscArraymove(ap + bs2 * (i + 1), ap + bs2 * i, bs2 * (N - i + 1)));
2508: PetscCall(PetscArrayzero(ap + bs2 * i, bs2));
2509: ap[bs2 * i + bs * cidx + ridx] = value;
2510: }
2511: a->nz++;
2512: noinsert1:;
2513: low = i;
2514: }
2515: ailen[brow] = nrow;
2516: }
2517: PetscFunctionReturn(PETSC_SUCCESS);
2518: }
2520: static PetscErrorCode MatILUFactor_SeqBAIJ(Mat inA, IS row, IS col, const MatFactorInfo *info)
2521: {
2522: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)inA->data;
2523: Mat outA;
2524: PetscBool row_identity, col_identity;
2526: PetscFunctionBegin;
2527: PetscCheck(info->levels == 0, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only levels = 0 supported for in-place ILU");
2528: PetscCall(ISIdentity(row, &row_identity));
2529: PetscCall(ISIdentity(col, &col_identity));
2530: PetscCheck(row_identity && col_identity, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Row and column permutations must be identity for in-place ILU");
2532: outA = inA;
2533: inA->factortype = MAT_FACTOR_LU;
2534: PetscCall(PetscFree(inA->solvertype));
2535: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &inA->solvertype));
2537: PetscCall(PetscObjectReference((PetscObject)row));
2538: PetscCall(ISDestroy(&a->row));
2539: a->row = row;
2540: PetscCall(PetscObjectReference((PetscObject)col));
2541: PetscCall(ISDestroy(&a->col));
2542: a->col = col;
2544: /* Create the invert permutation so that it can be used in MatLUFactorNumeric() */
2545: PetscCall(ISDestroy(&a->icol));
2546: PetscCall(ISInvertPermutation(col, PETSC_DECIDE, &a->icol));
2548: PetscCall(MatSeqBAIJSetNumericFactorization_inplace(inA, (PetscBool)(row_identity && col_identity)));
2549: if (!a->solve_work) PetscCall(PetscMalloc1(inA->rmap->N + inA->rmap->bs, &a->solve_work));
2550: PetscCall(MatLUFactorNumeric(outA, inA, info));
2551: PetscFunctionReturn(PETSC_SUCCESS);
2552: }
2554: static PetscErrorCode MatSeqBAIJSetColumnIndices_SeqBAIJ(Mat mat, const PetscInt *indices)
2555: {
2556: Mat_SeqBAIJ *baij = (Mat_SeqBAIJ *)mat->data;
2558: PetscFunctionBegin;
2559: baij->nz = baij->maxnz;
2560: PetscCall(PetscArraycpy(baij->j, indices, baij->nz));
2561: PetscCall(PetscArraycpy(baij->ilen, baij->imax, baij->mbs));
2562: PetscFunctionReturn(PETSC_SUCCESS);
2563: }
2565: /*@
2566: MatSeqBAIJSetColumnIndices - Set the column indices for all the block rows in the matrix.
2568: Input Parameters:
2569: + mat - the `MATSEQBAIJ` matrix
2570: - indices - the block column indices
2572: Level: advanced
2574: Notes:
2575: This can be called if you have precomputed the nonzero structure of the
2576: matrix and want to provide it to the matrix object to improve the performance
2577: of the `MatSetValues()` operation.
2579: You MUST have set the correct numbers of nonzeros per row in the call to
2580: `MatCreateSeqBAIJ()`, and the columns indices MUST be sorted.
2582: MUST be called before any calls to `MatSetValues()`
2584: .seealso: [](ch_matrices), `Mat`, `MATSEQBAIJ`, `MatSetValues()`
2585: @*/
2586: PetscErrorCode MatSeqBAIJSetColumnIndices(Mat mat, PetscInt *indices)
2587: {
2588: PetscFunctionBegin;
2590: PetscAssertPointer(indices, 2);
2591: PetscUseMethod(mat, "MatSeqBAIJSetColumnIndices_C", (Mat, const PetscInt *), (mat, (const PetscInt *)indices));
2592: PetscFunctionReturn(PETSC_SUCCESS);
2593: }
2595: static PetscErrorCode MatGetRowMaxAbs_SeqBAIJ(Mat A, Vec v, PetscInt idx[])
2596: {
2597: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2598: PetscInt i, j, n, row, bs, *ai, *aj, mbs;
2599: PetscReal atmp;
2600: PetscScalar *x, zero = 0.0;
2601: MatScalar *aa;
2602: PetscInt ncols, brow, krow, kcol;
2604: PetscFunctionBegin;
2605: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2606: bs = A->rmap->bs;
2607: aa = a->a;
2608: ai = a->i;
2609: aj = a->j;
2610: mbs = a->mbs;
2612: PetscCall(VecSet(v, zero));
2613: PetscCall(VecGetArray(v, &x));
2614: PetscCall(VecGetLocalSize(v, &n));
2615: PetscCheck(n == A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
2616: for (i = 0; i < mbs; i++) {
2617: ncols = ai[1] - ai[0];
2618: ai++;
2619: brow = bs * i;
2620: for (j = 0; j < ncols; j++) {
2621: for (kcol = 0; kcol < bs; kcol++) {
2622: for (krow = 0; krow < bs; krow++) {
2623: atmp = PetscAbsScalar(*aa);
2624: aa++;
2625: row = brow + krow; /* row index */
2626: if (PetscAbsScalar(x[row]) < atmp) {
2627: x[row] = atmp;
2628: if (idx) idx[row] = bs * (*aj) + kcol;
2629: }
2630: }
2631: }
2632: aj++;
2633: }
2634: }
2635: PetscCall(VecRestoreArray(v, &x));
2636: PetscFunctionReturn(PETSC_SUCCESS);
2637: }
2639: static PetscErrorCode MatGetRowSumAbs_SeqBAIJ(Mat A, Vec v)
2640: {
2641: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2642: PetscInt i, j, n, row, bs, *ai, mbs;
2643: PetscReal atmp;
2644: PetscScalar *x, zero = 0.0;
2645: MatScalar *aa;
2646: PetscInt ncols, brow, krow, kcol;
2648: PetscFunctionBegin;
2649: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2650: bs = A->rmap->bs;
2651: aa = a->a;
2652: ai = a->i;
2653: mbs = a->mbs;
2655: PetscCall(VecSet(v, zero));
2656: PetscCall(VecGetArrayWrite(v, &x));
2657: PetscCall(VecGetLocalSize(v, &n));
2658: PetscCheck(n == A->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
2659: for (i = 0; i < mbs; i++) {
2660: ncols = ai[1] - ai[0];
2661: ai++;
2662: brow = bs * i;
2663: for (j = 0; j < ncols; j++) {
2664: for (kcol = 0; kcol < bs; kcol++) {
2665: for (krow = 0; krow < bs; krow++) {
2666: atmp = PetscAbsScalar(*aa);
2667: aa++;
2668: row = brow + krow; /* row index */
2669: x[row] += atmp;
2670: }
2671: }
2672: }
2673: }
2674: PetscCall(VecRestoreArrayWrite(v, &x));
2675: PetscFunctionReturn(PETSC_SUCCESS);
2676: }
2678: static PetscErrorCode MatCopy_SeqBAIJ(Mat A, Mat B, MatStructure str)
2679: {
2680: PetscFunctionBegin;
2681: /* If the two matrices have the same copy implementation, use fast copy. */
2682: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2683: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2684: Mat_SeqBAIJ *b = (Mat_SeqBAIJ *)B->data;
2685: PetscInt ambs = a->mbs, bmbs = b->mbs, abs = A->rmap->bs, bbs = B->rmap->bs, bs2 = abs * abs;
2687: PetscCheck(a->i[ambs] == b->i[bmbs], PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Number of nonzero blocks in matrices A %" PetscInt_FMT " and B %" PetscInt_FMT " are different", a->i[ambs], b->i[bmbs]);
2688: PetscCheck(abs == bbs, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Block size A %" PetscInt_FMT " and B %" PetscInt_FMT " are different", abs, bbs);
2689: PetscCall(PetscArraycpy(b->a, a->a, bs2 * a->i[ambs]));
2690: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2691: } else {
2692: PetscCall(MatCopy_Basic(A, B, str));
2693: }
2694: PetscFunctionReturn(PETSC_SUCCESS);
2695: }
2697: static PetscErrorCode MatSeqBAIJGetArray_SeqBAIJ(Mat A, PetscScalar *array[])
2698: {
2699: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2701: PetscFunctionBegin;
2702: *array = a->a;
2703: PetscFunctionReturn(PETSC_SUCCESS);
2704: }
2706: static PetscErrorCode MatSeqBAIJRestoreArray_SeqBAIJ(Mat A, PetscScalar *array[])
2707: {
2708: PetscFunctionBegin;
2709: *array = NULL;
2710: PetscFunctionReturn(PETSC_SUCCESS);
2711: }
2713: PetscErrorCode MatAXPYGetPreallocation_SeqBAIJ(Mat Y, Mat X, PetscInt *nnz)
2714: {
2715: PetscInt bs = Y->rmap->bs, mbs = Y->rmap->N / bs;
2716: Mat_SeqBAIJ *x = (Mat_SeqBAIJ *)X->data;
2717: Mat_SeqBAIJ *y = (Mat_SeqBAIJ *)Y->data;
2719: PetscFunctionBegin;
2720: /* Set the number of nonzeros in the new matrix */
2721: PetscCall(MatAXPYGetPreallocation_SeqX_private(mbs, x->i, x->j, y->i, y->j, nnz));
2722: PetscFunctionReturn(PETSC_SUCCESS);
2723: }
2725: PetscErrorCode MatAXPY_SeqBAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2726: {
2727: Mat_SeqBAIJ *x = (Mat_SeqBAIJ *)X->data, *y = (Mat_SeqBAIJ *)Y->data;
2728: PetscInt bs = Y->rmap->bs, bs2 = bs * bs;
2729: PetscBLASInt one = 1;
2731: PetscFunctionBegin;
2732: if (str == UNKNOWN_NONZERO_PATTERN || (PetscDefined(USE_DEBUG) && str == SAME_NONZERO_PATTERN)) {
2733: PetscBool e = x->nz == y->nz && x->mbs == y->mbs && bs == X->rmap->bs ? PETSC_TRUE : PETSC_FALSE;
2734: if (e) {
2735: PetscCall(PetscArraycmp(x->i, y->i, x->mbs + 1, &e));
2736: if (e) {
2737: PetscCall(PetscArraycmp(x->j, y->j, x->i[x->mbs], &e));
2738: if (e) str = SAME_NONZERO_PATTERN;
2739: }
2740: }
2741: if (!e) PetscCheck(str != SAME_NONZERO_PATTERN, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "MatStructure is not SAME_NONZERO_PATTERN");
2742: }
2743: if (str == SAME_NONZERO_PATTERN) {
2744: PetscScalar alpha = a;
2745: PetscBLASInt bnz;
2746: PetscCall(PetscBLASIntCast(x->nz * bs2, &bnz));
2747: PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, x->a, &one, y->a, &one));
2748: PetscCall(PetscObjectStateIncrease((PetscObject)Y));
2749: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2750: PetscCall(MatAXPY_Basic(Y, a, X, str));
2751: } else {
2752: Mat B;
2753: PetscInt *nnz;
2754: PetscCheck(bs == X->rmap->bs, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrices must have same block size");
2755: PetscCall(PetscMalloc1(Y->rmap->N, &nnz));
2756: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2757: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2758: PetscCall(MatSetSizes(B, Y->rmap->n, Y->cmap->n, Y->rmap->N, Y->cmap->N));
2759: PetscCall(MatSetBlockSizesFromMats(B, Y, Y));
2760: PetscCall(MatSetType(B, (MatType)((PetscObject)Y)->type_name));
2761: PetscCall(MatAXPYGetPreallocation_SeqBAIJ(Y, X, nnz));
2762: PetscCall(MatSeqBAIJSetPreallocation(B, bs, 0, nnz));
2763: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2764: PetscCall(MatHeaderMerge(Y, &B));
2765: PetscCall(PetscFree(nnz));
2766: }
2767: PetscFunctionReturn(PETSC_SUCCESS);
2768: }
2770: PETSC_INTERN PetscErrorCode MatConjugate_SeqBAIJ(Mat A)
2771: {
2772: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2773: PetscInt i, nz = a->bs2 * a->i[a->mbs];
2774: MatScalar *aa = a->a;
2776: PetscFunctionBegin;
2777: for (i = 0; i < nz; i++) aa[i] = PetscConj(aa[i]);
2778: PetscFunctionReturn(PETSC_SUCCESS);
2779: }
2781: static PetscErrorCode MatRealPart_SeqBAIJ(Mat A)
2782: {
2783: #if PetscDefined(USE_COMPLEX)
2784: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2785: PetscInt i, nz = a->bs2 * a->i[a->mbs];
2786: MatScalar *aa = a->a;
2788: PetscFunctionBegin;
2789: for (i = 0; i < nz; i++) aa[i] = PetscRealPart(aa[i]);
2790: PetscFunctionReturn(PETSC_SUCCESS);
2791: #else
2792: (void)A;
2793: return PETSC_SUCCESS;
2794: #endif
2795: }
2797: static PetscErrorCode MatImaginaryPart_SeqBAIJ(Mat A)
2798: {
2799: #if PetscDefined(USE_COMPLEX)
2800: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2801: PetscInt i, nz = a->bs2 * a->i[a->mbs];
2802: MatScalar *aa = a->a;
2804: PetscFunctionBegin;
2805: for (i = 0; i < nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
2806: PetscFunctionReturn(PETSC_SUCCESS);
2807: #else
2808: (void)A;
2809: return PETSC_SUCCESS;
2810: #endif
2811: }
2813: /*
2814: Code almost identical to MatGetColumnIJ_SeqAIJ() should share common code
2815: */
2816: static PetscErrorCode MatGetColumnIJ_SeqBAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
2817: {
2818: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2819: PetscInt bs = A->rmap->bs, i, *collengths, *cia, *cja, n = A->cmap->n / bs, m = A->rmap->n / bs;
2820: PetscInt nz = a->i[m], row, *jj, mr, col;
2822: PetscFunctionBegin;
2823: *nn = n;
2824: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
2825: PetscCheck(!symmetric, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not for BAIJ matrices");
2826: PetscCall(PetscCalloc1(n, &collengths));
2827: PetscCall(PetscMalloc1(n + 1, &cia));
2828: PetscCall(PetscMalloc1(nz, &cja));
2829: jj = a->j;
2830: for (i = 0; i < nz; i++) collengths[jj[i]]++;
2831: cia[0] = oshift;
2832: for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i];
2833: PetscCall(PetscArrayzero(collengths, n));
2834: jj = a->j;
2835: for (row = 0; row < m; row++) {
2836: mr = a->i[row + 1] - a->i[row];
2837: for (i = 0; i < mr; i++) {
2838: col = *jj++;
2840: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
2841: }
2842: }
2843: PetscCall(PetscFree(collengths));
2844: *ia = cia;
2845: *ja = cja;
2846: PetscFunctionReturn(PETSC_SUCCESS);
2847: }
2849: static PetscErrorCode MatRestoreColumnIJ_SeqBAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
2850: {
2851: PetscFunctionBegin;
2852: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
2853: PetscCall(PetscFree(*ia));
2854: PetscCall(PetscFree(*ja));
2855: PetscFunctionReturn(PETSC_SUCCESS);
2856: }
2858: /*
2859: MatGetColumnIJ_SeqBAIJ_Color() and MatRestoreColumnIJ_SeqBAIJ_Color() are customized from
2860: MatGetColumnIJ_SeqBAIJ() and MatRestoreColumnIJ_SeqBAIJ() by adding an output
2861: spidx[], index of a->a, to be used in MatTransposeColoringCreate() and MatFDColoringCreate()
2862: */
2863: PetscErrorCode MatGetColumnIJ_SeqBAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done)
2864: {
2865: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2866: PetscInt i, *collengths, *cia, *cja, n = a->nbs, m = a->mbs;
2867: PetscInt nz = a->i[m], row, *jj, mr, col;
2868: PetscInt *cspidx;
2870: PetscFunctionBegin;
2871: *nn = n;
2872: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
2874: PetscCall(PetscCalloc1(n, &collengths));
2875: PetscCall(PetscMalloc1(n + 1, &cia));
2876: PetscCall(PetscMalloc1(nz, &cja));
2877: PetscCall(PetscMalloc1(nz, &cspidx));
2878: jj = a->j;
2879: for (i = 0; i < nz; i++) collengths[jj[i]]++;
2880: cia[0] = oshift;
2881: for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i];
2882: PetscCall(PetscArrayzero(collengths, n));
2883: jj = a->j;
2884: for (row = 0; row < m; row++) {
2885: mr = a->i[row + 1] - a->i[row];
2886: for (i = 0; i < mr; i++) {
2887: col = *jj++;
2888: cspidx[cia[col] + collengths[col] - oshift] = a->i[row] + i; /* index of a->j */
2889: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
2890: }
2891: }
2892: PetscCall(PetscFree(collengths));
2893: *ia = cia;
2894: *ja = cja;
2895: *spidx = cspidx;
2896: PetscFunctionReturn(PETSC_SUCCESS);
2897: }
2899: PetscErrorCode MatRestoreColumnIJ_SeqBAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done)
2900: {
2901: PetscFunctionBegin;
2902: PetscCall(MatRestoreColumnIJ_SeqBAIJ(A, oshift, symmetric, inodecompressed, n, ia, ja, done));
2903: PetscCall(PetscFree(*spidx));
2904: PetscFunctionReturn(PETSC_SUCCESS);
2905: }
2907: static PetscErrorCode MatShift_SeqBAIJ(Mat Y, PetscScalar a)
2908: {
2909: Mat_SeqBAIJ *aij = (Mat_SeqBAIJ *)Y->data;
2911: PetscFunctionBegin;
2912: if (!Y->preallocated || !aij->nz) PetscCall(MatSeqBAIJSetPreallocation(Y, Y->rmap->bs, 1, NULL));
2913: PetscCall(MatShift_Basic(Y, a));
2914: PetscFunctionReturn(PETSC_SUCCESS);
2915: }
2917: PetscErrorCode MatEliminateZeros_SeqBAIJ(Mat A, PetscBool keep)
2918: {
2919: Mat_SeqBAIJ *a = (Mat_SeqBAIJ *)A->data;
2920: PetscInt fshift = 0, fshift_prev = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax, j, k;
2921: PetscInt m = A->rmap->N, *ailen = a->ilen;
2922: PetscInt mbs = a->mbs, bs2 = a->bs2, rmax = 0;
2923: MatScalar *aa = a->a, *ap;
2924: PetscBool zero;
2926: PetscFunctionBegin;
2927: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot eliminate zeros for unassembled matrix");
2928: if (m) rmax = ailen[0];
2929: for (i = 1, a->nonzerorowcnt = 0; i <= mbs; i++) {
2930: for (k = ai[i - 1]; k < ai[i]; k++) {
2931: zero = PETSC_TRUE;
2932: ap = aa + bs2 * k;
2933: for (j = 0; j < bs2 && zero; j++) {
2934: if (ap[j] != 0.0) zero = PETSC_FALSE;
2935: }
2936: if (zero && (aj[k] != i - 1 || !keep)) fshift++;
2937: else {
2938: if (zero && aj[k] == i - 1) PetscCall(PetscInfo(A, "Keep the diagonal block at row %" PetscInt_FMT "\n", i - 1));
2939: aj[k - fshift] = aj[k];
2940: PetscCall(PetscArraymove(ap - bs2 * fshift, ap, bs2));
2941: }
2942: }
2943: ai[i - 1] -= fshift_prev;
2944: fshift_prev = fshift;
2945: ailen[i - 1] = imax[i - 1] = ai[i] - fshift - ai[i - 1];
2946: a->nonzerorowcnt += ((ai[i] - fshift - ai[i - 1]) > 0);
2947: rmax = PetscMax(rmax, ailen[i - 1]);
2948: }
2949: if (fshift) {
2950: if (mbs) {
2951: ai[mbs] -= fshift;
2952: a->nz = ai[mbs];
2953: }
2954: PetscCall(PetscInfo(A, "Matrix size: %" PetscInt_FMT " X %" PetscInt_FMT "; zeros eliminated: %" PetscInt_FMT "; nonzeros left: %" PetscInt_FMT "\n", m, A->cmap->n, fshift, a->nz));
2955: A->nonzerostate++;
2956: A->info.nz_unneeded += (PetscReal)fshift;
2957: a->rmax = rmax;
2958: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
2959: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
2960: }
2961: PetscFunctionReturn(PETSC_SUCCESS);
2962: }
2964: static struct _MatOps MatOps_Values = {MatSetValues_SeqBAIJ,
2965: MatGetRow_SeqBAIJ,
2966: MatRestoreRow_SeqBAIJ,
2967: MatMult_SeqBAIJ_N,
2968: /* 4*/ MatMultAdd_SeqBAIJ_N,
2969: MatMultTranspose_SeqBAIJ,
2970: MatMultTransposeAdd_SeqBAIJ,
2971: NULL,
2972: NULL,
2973: NULL,
2974: /* 10*/ NULL,
2975: MatLUFactor_SeqBAIJ,
2976: NULL,
2977: NULL,
2978: MatTranspose_SeqBAIJ,
2979: /* 15*/ MatGetInfo_SeqBAIJ,
2980: MatEqual_SeqBAIJ,
2981: MatGetDiagonal_SeqBAIJ,
2982: MatDiagonalScale_SeqBAIJ,
2983: MatNorm_SeqBAIJ,
2984: /* 20*/ NULL,
2985: MatAssemblyEnd_SeqBAIJ,
2986: MatSetOption_SeqBAIJ,
2987: MatZeroEntries_SeqBAIJ,
2988: /* 24*/ MatZeroRows_SeqBAIJ,
2989: NULL,
2990: NULL,
2991: NULL,
2992: NULL,
2993: /* 29*/ MatSetUp_Seq_Hash,
2994: NULL,
2995: NULL,
2996: NULL,
2997: NULL,
2998: /* 34*/ MatDuplicate_SeqBAIJ,
2999: NULL,
3000: NULL,
3001: MatILUFactor_SeqBAIJ,
3002: NULL,
3003: /* 39*/ MatAXPY_SeqBAIJ,
3004: MatCreateSubMatrices_SeqBAIJ,
3005: MatIncreaseOverlap_SeqBAIJ,
3006: MatGetValues_SeqBAIJ,
3007: MatCopy_SeqBAIJ,
3008: /* 44*/ NULL,
3009: MatScale_SeqBAIJ,
3010: MatShift_SeqBAIJ,
3011: NULL,
3012: MatZeroRowsColumns_SeqBAIJ,
3013: /* 49*/ NULL,
3014: MatGetRowIJ_SeqBAIJ,
3015: MatRestoreRowIJ_SeqBAIJ,
3016: MatGetColumnIJ_SeqBAIJ,
3017: MatRestoreColumnIJ_SeqBAIJ,
3018: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3019: NULL,
3020: NULL,
3021: NULL,
3022: MatSetValuesBlocked_SeqBAIJ,
3023: /* 59*/ MatCreateSubMatrix_SeqBAIJ,
3024: MatDestroy_SeqBAIJ,
3025: MatView_SeqBAIJ,
3026: NULL,
3027: NULL,
3028: /* 64*/ NULL,
3029: NULL,
3030: NULL,
3031: NULL,
3032: MatGetRowMaxAbs_SeqBAIJ,
3033: /* 69*/ NULL,
3034: MatConvert_Basic,
3035: NULL,
3036: MatFDColoringApply_BAIJ,
3037: NULL,
3038: /* 74*/ NULL,
3039: NULL,
3040: NULL,
3041: NULL,
3042: MatLoad_SeqBAIJ,
3043: /* 79*/ NULL,
3044: NULL,
3045: NULL,
3046: NULL,
3047: NULL,
3048: /* 84*/ NULL,
3049: NULL,
3050: NULL,
3051: NULL,
3052: NULL,
3053: /* 89*/ NULL,
3054: NULL,
3055: NULL,
3056: NULL,
3057: MatConjugate_SeqBAIJ,
3058: /* 94*/ NULL,
3059: NULL,
3060: MatRealPart_SeqBAIJ,
3061: MatImaginaryPart_SeqBAIJ,
3062: NULL,
3063: /* 99*/ NULL,
3064: NULL,
3065: NULL,
3066: NULL,
3067: NULL,
3068: /*104*/ NULL,
3069: NULL,
3070: NULL,
3071: NULL,
3072: NULL,
3073: /*109*/ NULL,
3074: NULL,
3075: MatMultHermitianTranspose_SeqBAIJ,
3076: MatMultHermitianTransposeAdd_SeqBAIJ,
3077: NULL,
3078: /*114*/ NULL,
3079: MatGetColumnReductions_SeqBAIJ,
3080: MatInvertBlockDiagonal_SeqBAIJ,
3081: NULL,
3082: NULL,
3083: /*119*/ NULL,
3084: NULL,
3085: NULL,
3086: NULL,
3087: NULL,
3088: /*124*/ NULL,
3089: MatSetBlockSizes_Default,
3090: NULL,
3091: MatFDColoringSetUp_SeqXAIJ,
3092: NULL,
3093: /*129*/ MatCreateMPIMatConcatenateSeqMat_SeqBAIJ,
3094: MatDestroySubMatrices_SeqBAIJ,
3095: NULL,
3096: NULL,
3097: NULL,
3098: /*134*/ NULL,
3099: MatEliminateZeros_SeqBAIJ,
3100: MatGetRowSumAbs_SeqBAIJ,
3101: NULL,
3102: NULL,
3103: /*139*/ NULL,
3104: MatCopyHashToXAIJ_Seq_Hash,
3105: NULL,
3106: NULL,
3107: NULL,
3108: /*144*/ NULL,
3109: NULL,
3110: NULL,
3111: NULL};
3113: static PetscErrorCode MatStoreValues_SeqBAIJ(Mat mat)
3114: {
3115: Mat_SeqBAIJ *aij = (Mat_SeqBAIJ *)mat->data;
3116: PetscInt nz = aij->i[aij->mbs] * aij->bs2;
3118: PetscFunctionBegin;
3119: PetscCheck(aij->nonew == 1, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3121: /* allocate space for values if not already there */
3122: if (!aij->saved_values) PetscCall(PetscMalloc1(nz + 1, &aij->saved_values));
3124: /* copy values over */
3125: PetscCall(PetscArraycpy(aij->saved_values, aij->a, nz));
3126: PetscFunctionReturn(PETSC_SUCCESS);
3127: }
3129: static PetscErrorCode MatRetrieveValues_SeqBAIJ(Mat mat)
3130: {
3131: Mat_SeqBAIJ *aij = (Mat_SeqBAIJ *)mat->data;
3132: PetscInt nz = aij->i[aij->mbs] * aij->bs2;
3134: PetscFunctionBegin;
3135: PetscCheck(aij->nonew == 1, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3136: PetscCheck(aij->saved_values, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatStoreValues(A);first");
3138: /* copy values over */
3139: PetscCall(PetscArraycpy(aij->a, aij->saved_values, nz));
3140: PetscFunctionReturn(PETSC_SUCCESS);
3141: }
3143: PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat, MatType, MatReuse, Mat *);
3144: PETSC_INTERN PetscErrorCode MatConvert_SeqBAIJ_SeqSBAIJ(Mat, MatType, MatReuse, Mat *);
3146: PetscErrorCode MatSeqBAIJSetPreallocation_SeqBAIJ(Mat B, PetscInt bs, PetscInt nz, const PetscInt nnz[])
3147: {
3148: Mat_SeqBAIJ *b = (Mat_SeqBAIJ *)B->data;
3149: PetscInt i, mbs, nbs, bs2;
3150: PetscBool flg = PETSC_FALSE, skipallocation = PETSC_FALSE, realalloc = PETSC_FALSE;
3152: PetscFunctionBegin;
3153: if (B->hash_active) {
3154: PetscInt bs;
3155: B->ops[0] = b->cops;
3156: PetscCall(PetscHMapIJVDestroy(&b->ht));
3157: PetscCall(MatGetBlockSize(B, &bs));
3158: if (bs > 1) PetscCall(PetscHSetIJDestroy(&b->bht));
3159: PetscCall(PetscFree(b->dnz));
3160: PetscCall(PetscFree(b->bdnz));
3161: B->hash_active = PETSC_FALSE;
3162: }
3163: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3164: if (nz == MAT_SKIP_ALLOCATION) {
3165: skipallocation = PETSC_TRUE;
3166: nz = 0;
3167: }
3169: PetscCall(MatSetBlockSize(B, bs));
3170: PetscCall(PetscLayoutSetUp(B->rmap));
3171: PetscCall(PetscLayoutSetUp(B->cmap));
3172: PetscCall(PetscLayoutGetBlockSize(B->rmap, &bs));
3174: B->preallocated = PETSC_TRUE;
3176: mbs = B->rmap->n / bs;
3177: nbs = B->cmap->n / bs;
3178: bs2 = bs * bs;
3180: PetscCheck(mbs * bs == B->rmap->n && nbs * bs == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Number rows %" PetscInt_FMT ", cols %" PetscInt_FMT " must be divisible by blocksize %" PetscInt_FMT, B->rmap->N, B->cmap->n, bs);
3182: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3183: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nz cannot be less than 0: value %" PetscInt_FMT, nz);
3184: if (nnz) {
3185: for (i = 0; i < mbs; i++) {
3186: PetscCheck(nnz[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nnz cannot be less than 0: local row %" PetscInt_FMT " value %" PetscInt_FMT, i, nnz[i]);
3187: PetscCheck(nnz[i] <= nbs, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nnz cannot be greater than block row length: local row %" PetscInt_FMT " value %" PetscInt_FMT " rowlength %" PetscInt_FMT, i, nnz[i], nbs);
3188: }
3189: }
3191: PetscOptionsBegin(PetscObjectComm((PetscObject)B), NULL, "Optimize options for SEQBAIJ matrix 2 ", "Mat");
3192: PetscCall(PetscOptionsBool("-mat_no_unroll", "Do not optimize for block size (slow)", NULL, flg, &flg, NULL));
3193: PetscOptionsEnd();
3195: if (!flg) {
3196: switch (bs) {
3197: case 1:
3198: B->ops->mult = MatMult_SeqBAIJ_1;
3199: B->ops->multadd = MatMultAdd_SeqBAIJ_1;
3200: break;
3201: case 2:
3202: B->ops->mult = MatMult_SeqBAIJ_2;
3203: B->ops->multadd = MatMultAdd_SeqBAIJ_2;
3204: break;
3205: case 3:
3206: B->ops->mult = MatMult_SeqBAIJ_3;
3207: B->ops->multadd = MatMultAdd_SeqBAIJ_3;
3208: break;
3209: case 4:
3210: B->ops->mult = MatMult_SeqBAIJ_4;
3211: B->ops->multadd = MatMultAdd_SeqBAIJ_4;
3212: break;
3213: case 5:
3214: B->ops->mult = MatMult_SeqBAIJ_5;
3215: B->ops->multadd = MatMultAdd_SeqBAIJ_5;
3216: break;
3217: case 6:
3218: B->ops->mult = MatMult_SeqBAIJ_6;
3219: B->ops->multadd = MatMultAdd_SeqBAIJ_6;
3220: break;
3221: case 7:
3222: B->ops->mult = MatMult_SeqBAIJ_7;
3223: B->ops->multadd = MatMultAdd_SeqBAIJ_7;
3224: break;
3225: case 9: {
3226: PetscInt version = 1;
3227: PetscCall(PetscOptionsGetInt(NULL, ((PetscObject)B)->prefix, "-mat_baij_mult_version", &version, NULL));
3228: switch (version) {
3229: #if PetscDefined(HAVE_IMMINTRIN_H) && defined(__AVX2__) && defined(__FMA__) && PetscDefined(USE_REAL_DOUBLE) && !PetscDefined(USE_COMPLEX) && !PetscDefined(USE_64BIT_INDICES)
3230: case 1:
3231: B->ops->mult = MatMult_SeqBAIJ_9_AVX2;
3232: B->ops->multadd = MatMultAdd_SeqBAIJ_9_AVX2;
3233: PetscCall(PetscInfo(B, "Using AVX2 for MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", bs));
3234: break;
3235: #endif
3236: default:
3237: B->ops->mult = MatMult_SeqBAIJ_N;
3238: B->ops->multadd = MatMultAdd_SeqBAIJ_N;
3239: PetscCall(PetscInfo(B, "Using BLAS for MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", bs));
3240: break;
3241: }
3242: break;
3243: }
3244: case 11:
3245: B->ops->mult = MatMult_SeqBAIJ_11;
3246: B->ops->multadd = MatMultAdd_SeqBAIJ_11;
3247: break;
3248: case 12: {
3249: PetscInt version = 1;
3250: PetscCall(PetscOptionsGetInt(NULL, ((PetscObject)B)->prefix, "-mat_baij_mult_version", &version, NULL));
3251: switch (version) {
3252: case 1:
3253: B->ops->mult = MatMult_SeqBAIJ_12_ver1;
3254: B->ops->multadd = MatMultAdd_SeqBAIJ_12_ver1;
3255: PetscCall(PetscInfo(B, "Using version %" PetscInt_FMT " of MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", version, bs));
3256: break;
3257: case 2:
3258: B->ops->mult = MatMult_SeqBAIJ_12_ver2;
3259: B->ops->multadd = MatMultAdd_SeqBAIJ_12_ver2;
3260: PetscCall(PetscInfo(B, "Using version %" PetscInt_FMT " of MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", version, bs));
3261: break;
3262: #if PetscDefined(HAVE_IMMINTRIN_H) && defined(__AVX2__) && defined(__FMA__) && PetscDefined(USE_REAL_DOUBLE) && !PetscDefined(USE_COMPLEX) && !PetscDefined(USE_64BIT_INDICES)
3263: case 3:
3264: B->ops->mult = MatMult_SeqBAIJ_12_AVX2;
3265: B->ops->multadd = MatMultAdd_SeqBAIJ_12_ver1;
3266: PetscCall(PetscInfo(B, "Using AVX2 for MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", bs));
3267: break;
3268: #endif
3269: default:
3270: B->ops->mult = MatMult_SeqBAIJ_N;
3271: B->ops->multadd = MatMultAdd_SeqBAIJ_N;
3272: PetscCall(PetscInfo(B, "Using BLAS for MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", bs));
3273: break;
3274: }
3275: break;
3276: }
3277: case 15: {
3278: PetscInt version = 1;
3279: PetscCall(PetscOptionsGetInt(NULL, ((PetscObject)B)->prefix, "-mat_baij_mult_version", &version, NULL));
3280: switch (version) {
3281: case 1:
3282: B->ops->mult = MatMult_SeqBAIJ_15_ver1;
3283: PetscCall(PetscInfo(B, "Using version %" PetscInt_FMT " of MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", version, bs));
3284: break;
3285: case 2:
3286: B->ops->mult = MatMult_SeqBAIJ_15_ver2;
3287: PetscCall(PetscInfo(B, "Using version %" PetscInt_FMT " of MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", version, bs));
3288: break;
3289: case 3:
3290: B->ops->mult = MatMult_SeqBAIJ_15_ver3;
3291: PetscCall(PetscInfo(B, "Using version %" PetscInt_FMT " of MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", version, bs));
3292: break;
3293: case 4:
3294: B->ops->mult = MatMult_SeqBAIJ_15_ver4;
3295: PetscCall(PetscInfo(B, "Using version %" PetscInt_FMT " of MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", version, bs));
3296: break;
3297: default:
3298: B->ops->mult = MatMult_SeqBAIJ_N;
3299: PetscCall(PetscInfo(B, "Using BLAS for MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", bs));
3300: break;
3301: }
3302: B->ops->multadd = MatMultAdd_SeqBAIJ_N;
3303: break;
3304: }
3305: default:
3306: B->ops->mult = MatMult_SeqBAIJ_N;
3307: B->ops->multadd = MatMultAdd_SeqBAIJ_N;
3308: PetscCall(PetscInfo(B, "Using BLAS for MatMult for BAIJ for blocksize %" PetscInt_FMT "\n", bs));
3309: break;
3310: }
3311: }
3312: B->ops->sor = MatSOR_SeqBAIJ;
3313: b->mbs = mbs;
3314: b->nbs = nbs;
3315: if (!skipallocation) {
3316: if (!b->imax) {
3317: PetscCall(PetscMalloc2(mbs, &b->imax, mbs, &b->ilen));
3319: b->free_imax_ilen = PETSC_TRUE;
3320: }
3321: /* b->ilen will count nonzeros in each block row so far. */
3322: for (i = 0; i < mbs; i++) b->ilen[i] = 0;
3323: if (!nnz) {
3324: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3325: else if (nz < 0) nz = 1;
3326: nz = PetscMin(nz, nbs);
3327: for (i = 0; i < mbs; i++) b->imax[i] = nz;
3328: PetscCall(PetscIntMultError(nz, mbs, &nz));
3329: } else {
3330: PetscInt64 nz64 = 0;
3331: for (i = 0; i < mbs; i++) {
3332: b->imax[i] = nnz[i];
3333: nz64 += nnz[i];
3334: }
3335: PetscCall(PetscIntCast(nz64, &nz));
3336: }
3338: /* allocate the matrix space */
3339: PetscCall(MatSeqXAIJFreeAIJ(B, &b->a, &b->j, &b->i));
3340: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscInt), (void **)&b->j));
3341: PetscCall(PetscShmgetAllocateArray(B->rmap->N + 1, sizeof(PetscInt), (void **)&b->i));
3342: if (B->structure_only) {
3343: b->free_a = PETSC_FALSE;
3344: } else {
3345: PetscInt nzbs2 = 0;
3346: PetscCall(PetscIntMultError(nz, bs2, &nzbs2));
3347: PetscCall(PetscShmgetAllocateArray(nzbs2, sizeof(PetscScalar), (void **)&b->a));
3348: b->free_a = PETSC_TRUE;
3349: PetscCall(PetscArrayzero(b->a, nzbs2));
3350: }
3351: b->free_ij = PETSC_TRUE;
3352: PetscCall(PetscArrayzero(b->j, nz));
3354: b->i[0] = 0;
3355: for (i = 1; i < mbs + 1; i++) b->i[i] = b->i[i - 1] + b->imax[i - 1];
3356: } else {
3357: b->free_a = PETSC_FALSE;
3358: b->free_ij = PETSC_FALSE;
3359: }
3361: b->bs2 = bs2;
3362: b->mbs = mbs;
3363: b->nz = 0;
3364: b->maxnz = nz;
3365: B->info.nz_unneeded = (PetscReal)b->maxnz * bs2;
3366: B->was_assembled = PETSC_FALSE;
3367: B->assembled = PETSC_FALSE;
3368: if (realalloc) PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
3369: PetscFunctionReturn(PETSC_SUCCESS);
3370: }
3372: static PetscErrorCode MatSeqBAIJSetPreallocationCSR_SeqBAIJ(Mat B, PetscInt bs, const PetscInt ii[], const PetscInt jj[], const PetscScalar V[])
3373: {
3374: PetscInt i, m, nz, nz_max = 0, *nnz;
3375: PetscScalar *values = NULL;
3376: PetscBool roworiented = ((Mat_SeqBAIJ *)B->data)->roworiented;
3378: PetscFunctionBegin;
3379: PetscCheck(bs >= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Invalid block size specified, must be positive but it is %" PetscInt_FMT, bs);
3380: PetscCall(PetscLayoutSetBlockSize(B->rmap, bs));
3381: PetscCall(PetscLayoutSetBlockSize(B->cmap, bs));
3382: PetscCall(PetscLayoutSetUp(B->rmap));
3383: PetscCall(PetscLayoutSetUp(B->cmap));
3384: PetscCall(PetscLayoutGetBlockSize(B->rmap, &bs));
3385: m = B->rmap->n / bs;
3387: PetscCheck(ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "ii[0] must be 0 but it is %" PetscInt_FMT, ii[0]);
3388: PetscCall(PetscMalloc1(m + 1, &nnz));
3389: for (i = 0; i < m; i++) {
3390: nz = ii[i + 1] - ii[i];
3391: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative number of columns %" PetscInt_FMT, i, nz);
3392: nz_max = PetscMax(nz_max, nz);
3393: nnz[i] = nz;
3394: }
3395: PetscCall(MatSeqBAIJSetPreallocation(B, bs, 0, nnz));
3396: PetscCall(PetscFree(nnz));
3398: values = (PetscScalar *)V;
3399: if (!values) PetscCall(PetscCalloc1(bs * bs * (nz_max + 1), &values));
3400: for (i = 0; i < m; i++) {
3401: PetscInt ncols = ii[i + 1] - ii[i];
3402: const PetscInt *icols = jj + ii[i];
3403: if (bs == 1 || !roworiented) {
3404: const PetscScalar *svals = values + (V ? (bs * bs * ii[i]) : 0);
3405: PetscCall(MatSetValuesBlocked_SeqBAIJ(B, 1, &i, ncols, icols, svals, INSERT_VALUES));
3406: } else {
3407: for (PetscInt j = 0; j < ncols; j++) {
3408: const PetscScalar *svals = values + (V ? (bs * bs * (ii[i] + j)) : 0);
3409: PetscCall(MatSetValuesBlocked_SeqBAIJ(B, 1, &i, 1, &icols[j], svals, INSERT_VALUES));
3410: }
3411: }
3412: }
3413: if (!V) PetscCall(PetscFree(values));
3414: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3415: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3416: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3417: PetscFunctionReturn(PETSC_SUCCESS);
3418: }
3420: /*@
3421: MatSeqBAIJGetArray - gives read/write access to the array where the data for a `MATSEQBAIJ` matrix is stored
3423: Not Collective
3425: Input Parameter:
3426: . A - a `MATSEQBAIJ` matrix
3428: Output Parameter:
3429: . array - pointer to the data
3431: Level: intermediate
3433: .seealso: [](ch_matrices), `Mat`, `MATSEQBAIJ`, `MatSeqBAIJRestoreArray()`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArray()`
3434: @*/
3435: PetscErrorCode MatSeqBAIJGetArray(Mat A, PetscScalar *array[])
3436: {
3437: PetscFunctionBegin;
3438: PetscUseMethod(A, "MatSeqBAIJGetArray_C", (Mat, PetscScalar **), (A, array));
3439: PetscFunctionReturn(PETSC_SUCCESS);
3440: }
3442: /*@
3443: MatSeqBAIJRestoreArray - returns access to the array where the data for a `MATSEQBAIJ` matrix is stored obtained by `MatSeqBAIJGetArray()`
3445: Not Collective
3447: Input Parameters:
3448: + A - a `MATSEQBAIJ` matrix
3449: - array - pointer to the data
3451: Level: intermediate
3453: .seealso: [](ch_matrices), `Mat`, `MatSeqBAIJGetArray()`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArray()`
3454: @*/
3455: PetscErrorCode MatSeqBAIJRestoreArray(Mat A, PetscScalar *array[])
3456: {
3457: PetscFunctionBegin;
3458: PetscUseMethod(A, "MatSeqBAIJRestoreArray_C", (Mat, PetscScalar **), (A, array));
3459: PetscCall(PetscObjectStateIncrease((PetscObject)A));
3460: PetscFunctionReturn(PETSC_SUCCESS);
3461: }
3463: /*MC
3464: MATSEQBAIJ - MATSEQBAIJ = "seqbaij" - A matrix type to be used for sequential block sparse matrices, based on
3465: block sparse compressed row format.
3467: Options Database Keys:
3468: + -mat_type seqbaij - sets the matrix type to `MATSEQBAIJ` during a call to `MatSetFromOptions()`
3469: - -mat_baij_mult_version version - indicate the version of the matrix-vector product to use (0 often indicates using BLAS)
3471: Level: beginner
3473: Notes:
3474: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
3475: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
3477: Run with `-info` to see what version of the matrix-vector product is being used
3479: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqBAIJ()`
3480: M*/
3482: PETSC_EXTERN PetscErrorCode MatCreate_SeqBAIJ(Mat B)
3483: {
3484: PetscMPIInt size;
3485: Mat_SeqBAIJ *b;
3487: PetscFunctionBegin;
3488: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
3489: PetscCheck(size == 1, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Comm must be of size 1");
3491: PetscCall(PetscNew(&b));
3492: B->data = (void *)b;
3493: B->ops[0] = MatOps_Values;
3495: b->row = NULL;
3496: b->col = NULL;
3497: b->icol = NULL;
3498: b->reallocs = 0;
3499: b->saved_values = NULL;
3501: b->roworiented = PETSC_TRUE;
3502: b->nonew = 0;
3503: b->diag = NULL;
3504: B->spptr = NULL;
3505: B->info.nz_unneeded = (PetscReal)b->maxnz * b->bs2;
3506: b->keepnonzeropattern = PETSC_FALSE;
3508: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqBAIJGetArray_C", MatSeqBAIJGetArray_SeqBAIJ));
3509: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqBAIJRestoreArray_C", MatSeqBAIJRestoreArray_SeqBAIJ));
3510: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_SeqBAIJ));
3511: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_SeqBAIJ));
3512: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqBAIJSetColumnIndices_C", MatSeqBAIJSetColumnIndices_SeqBAIJ));
3513: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqbaij_seqaij_C", MatConvert_SeqBAIJ_SeqAIJ));
3514: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqbaij_seqsbaij_C", MatConvert_SeqBAIJ_SeqSBAIJ));
3515: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqBAIJSetPreallocation_C", MatSeqBAIJSetPreallocation_SeqBAIJ));
3516: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqBAIJSetPreallocationCSR_C", MatSeqBAIJSetPreallocationCSR_SeqBAIJ));
3517: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_SeqBAIJ));
3518: #if PetscDefined(HAVE_HYPRE)
3519: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqbaij_hypre_C", MatConvert_AIJ_HYPRE));
3520: #endif
3521: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqbaij_is_C", MatConvert_XAIJ_IS));
3522: #if PetscDefined(HAVE_LIBXSMM)
3523: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqbaij_seqbaijlibxsmm_C", MatConvert_SeqBAIJ_SeqBAIJLIBXSMM));
3524: #endif
3525: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQBAIJ));
3526: PetscFunctionReturn(PETSC_SUCCESS);
3527: }
3529: PETSC_INTERN PetscErrorCode MatDuplicateNoCreate_SeqBAIJ(Mat C, Mat A, MatDuplicateOption cpvalues, PetscBool mallocmatspace)
3530: {
3531: Mat_SeqBAIJ *c = (Mat_SeqBAIJ *)C->data, *a = (Mat_SeqBAIJ *)A->data;
3532: PetscInt i, mbs = a->mbs, nz = a->nz, bs2 = a->bs2;
3534: PetscFunctionBegin;
3535: PetscCheck(A->assembled, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Cannot duplicate unassembled matrix");
3536: PetscCheck(a->i[mbs] == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Corrupt matrix");
3538: if (cpvalues == MAT_SHARE_NONZERO_PATTERN) {
3539: c->imax = a->imax;
3540: c->ilen = a->ilen;
3541: c->free_imax_ilen = PETSC_FALSE;
3542: } else {
3543: PetscCall(PetscMalloc2(mbs, &c->imax, mbs, &c->ilen));
3544: for (i = 0; i < mbs; i++) {
3545: c->imax[i] = a->imax[i];
3546: c->ilen[i] = a->ilen[i];
3547: }
3548: c->free_imax_ilen = PETSC_TRUE;
3549: }
3551: /* allocate the matrix space */
3552: if (mallocmatspace) {
3553: if (cpvalues == MAT_SHARE_NONZERO_PATTERN) {
3554: PetscCall(PetscShmgetAllocateArray(bs2 * nz, sizeof(PetscScalar), (void **)&c->a));
3555: PetscCall(PetscArrayzero(c->a, bs2 * nz));
3556: c->free_a = PETSC_TRUE;
3557: c->i = a->i;
3558: c->j = a->j;
3559: c->free_ij = PETSC_FALSE;
3560: c->parent = A;
3561: C->preallocated = PETSC_TRUE;
3562: C->assembled = PETSC_TRUE;
3564: PetscCall(PetscObjectReference((PetscObject)A));
3565: PetscCall(MatSetOption(A, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3566: PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3567: } else {
3568: PetscCall(PetscShmgetAllocateArray(bs2 * nz, sizeof(PetscScalar), (void **)&c->a));
3569: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscInt), (void **)&c->j));
3570: PetscCall(PetscShmgetAllocateArray(mbs + 1, sizeof(PetscInt), (void **)&c->i));
3571: c->free_a = PETSC_TRUE;
3572: c->free_ij = PETSC_TRUE;
3574: PetscCall(PetscArraycpy(c->i, a->i, mbs + 1));
3575: if (mbs > 0) {
3576: PetscCall(PetscArraycpy(c->j, a->j, nz));
3577: if (cpvalues == MAT_COPY_VALUES) {
3578: PetscCall(PetscArraycpy(c->a, a->a, bs2 * nz));
3579: } else {
3580: PetscCall(PetscArrayzero(c->a, bs2 * nz));
3581: }
3582: }
3583: C->preallocated = PETSC_TRUE;
3584: C->assembled = PETSC_TRUE;
3585: }
3586: }
3588: c->roworiented = a->roworiented;
3589: c->nonew = a->nonew;
3591: PetscCall(PetscLayoutReference(A->rmap, &C->rmap));
3592: PetscCall(PetscLayoutReference(A->cmap, &C->cmap));
3594: c->bs2 = a->bs2;
3595: c->mbs = a->mbs;
3596: c->nbs = a->nbs;
3597: c->nz = a->nz;
3598: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
3599: c->solve_work = NULL;
3600: c->mult_work = NULL;
3601: c->sor_workt = NULL;
3602: c->sor_work = NULL;
3604: c->compressedrow.use = a->compressedrow.use;
3605: c->compressedrow.nrows = a->compressedrow.nrows;
3606: if (a->compressedrow.use) {
3607: i = a->compressedrow.nrows;
3608: PetscCall(PetscMalloc2(i + 1, &c->compressedrow.i, i + 1, &c->compressedrow.rindex));
3609: PetscCall(PetscArraycpy(c->compressedrow.i, a->compressedrow.i, i + 1));
3610: PetscCall(PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, i));
3611: } else {
3612: c->compressedrow.use = PETSC_FALSE;
3613: c->compressedrow.i = NULL;
3614: c->compressedrow.rindex = NULL;
3615: }
3616: c->nonzerorowcnt = a->nonzerorowcnt;
3617: C->nonzerostate = A->nonzerostate;
3619: PetscCall(PetscFunctionListDuplicate(((PetscObject)A)->qlist, &((PetscObject)C)->qlist));
3620: PetscFunctionReturn(PETSC_SUCCESS);
3621: }
3623: PetscErrorCode MatDuplicate_SeqBAIJ(Mat A, MatDuplicateOption cpvalues, Mat *B)
3624: {
3625: PetscFunctionBegin;
3626: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
3627: PetscCall(MatSetSizes(*B, A->rmap->N, A->cmap->n, A->rmap->N, A->cmap->n));
3628: PetscCall(MatSetType(*B, MATSEQBAIJ));
3629: PetscCall(MatDuplicateNoCreate_SeqBAIJ(*B, A, cpvalues, PETSC_TRUE));
3630: PetscFunctionReturn(PETSC_SUCCESS);
3631: }
3633: /* Used for both SeqBAIJ and SeqSBAIJ matrices */
3634: PetscErrorCode MatLoad_SeqBAIJ_Binary(Mat mat, PetscViewer viewer)
3635: {
3636: PetscInt header[4], M, N, nz, bs, m, n, mbs, nbs, rows, cols, sum, i, j, k;
3637: PetscInt *rowidxs, *colidxs;
3638: PetscScalar *matvals;
3640: PetscFunctionBegin;
3641: PetscCall(PetscViewerSetUp(viewer));
3643: /* read matrix header */
3644: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3645: PetscCheck(header[0] == MAT_FILE_CLASSID, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3646: M = header[1];
3647: N = header[2];
3648: nz = header[3];
3649: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3650: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3651: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as SeqBAIJ");
3653: /* set block sizes from the viewer's .info file */
3654: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3655: /* set local and global sizes if not set already */
3656: if (mat->rmap->n < 0) mat->rmap->n = M;
3657: if (mat->cmap->n < 0) mat->cmap->n = N;
3658: if (mat->rmap->N < 0) mat->rmap->N = M;
3659: if (mat->cmap->N < 0) mat->cmap->N = N;
3660: PetscCall(PetscLayoutSetUp(mat->rmap));
3661: PetscCall(PetscLayoutSetUp(mat->cmap));
3663: /* check if the matrix sizes are correct */
3664: PetscCall(MatGetSize(mat, &rows, &cols));
3665: 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);
3666: PetscCall(MatGetBlockSize(mat, &bs));
3667: PetscCall(MatGetLocalSize(mat, &m, &n));
3668: mbs = m / bs;
3669: nbs = n / bs;
3671: /* read in row lengths, column indices and nonzero values */
3672: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3673: PetscCall(PetscViewerBinaryRead(viewer, rowidxs + 1, m, NULL, PETSC_INT));
3674: rowidxs[0] = 0;
3675: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3676: sum = rowidxs[m];
3677: PetscCheck(sum == nz, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Inconsistent matrix data in file: nonzeros = %" PetscInt_FMT ", sum-row-lengths = %" PetscInt_FMT, nz, sum);
3679: /* read in column indices and nonzero values */
3680: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, nz, &matvals));
3681: PetscCall(PetscViewerBinaryRead(viewer, colidxs, rowidxs[m], NULL, PETSC_INT));
3682: PetscCall(PetscViewerBinaryRead(viewer, matvals, rowidxs[m], NULL, PETSC_SCALAR));
3684: { /* preallocate matrix storage */
3685: PetscBT bt; /* helper bit set to count nonzeros */
3686: PetscInt *nnz;
3687: PetscBool sbaij;
3689: PetscCall(PetscBTCreate(nbs, &bt));
3690: PetscCall(PetscCalloc1(mbs, &nnz));
3691: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATSEQSBAIJ, &sbaij));
3692: for (i = 0; i < mbs; i++) {
3693: PetscCall(PetscBTMemzero(nbs, bt));
3694: for (k = 0; k < bs; k++) {
3695: PetscInt row = bs * i + k;
3696: for (j = rowidxs[row]; j < rowidxs[row + 1]; j++) {
3697: PetscInt col = colidxs[j];
3698: if (!sbaij || col >= row)
3699: if (!PetscBTLookupSet(bt, col / bs)) nnz[i]++;
3700: }
3701: }
3702: }
3703: PetscCall(PetscBTDestroy(&bt));
3704: PetscCall(MatSeqBAIJSetPreallocation(mat, bs, 0, nnz));
3705: PetscCall(MatSeqSBAIJSetPreallocation(mat, bs, 0, nnz));
3706: PetscCall(PetscFree(nnz));
3707: }
3709: /* store matrix values */
3710: for (i = 0; i < m; i++) {
3711: PetscInt row = i, s = rowidxs[i], e = rowidxs[i + 1];
3712: PetscUseTypeMethod(mat, setvalues, 1, &row, e - s, colidxs + s, matvals + s, INSERT_VALUES);
3713: }
3715: PetscCall(PetscFree(rowidxs));
3716: PetscCall(PetscFree2(colidxs, matvals));
3717: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
3718: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
3719: PetscFunctionReturn(PETSC_SUCCESS);
3720: }
3722: PetscErrorCode MatLoad_SeqBAIJ(Mat mat, PetscViewer viewer)
3723: {
3724: PetscBool isbinary;
3726: PetscFunctionBegin;
3727: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3728: 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);
3729: PetscCall(MatLoad_SeqBAIJ_Binary(mat, viewer));
3730: PetscFunctionReturn(PETSC_SUCCESS);
3731: }
3733: /*@
3734: MatCreateSeqBAIJ - Creates a sparse matrix in `MATSEQAIJ` (block
3735: compressed row) format. For good matrix assembly performance the
3736: user should preallocate the matrix storage by setting the parameter `nz`
3737: (or the array `nnz`).
3739: Collective
3741: Input Parameters:
3742: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3743: . bs - size of block, the blocks are ALWAYS square. One can use `MatSetBlockSizes()` to set a different row and column blocksize but the row
3744: blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with `MatCreateVecs()`
3745: . m - number of rows
3746: . n - number of columns
3747: . nz - number of nonzero blocks per block row (same for all rows)
3748: - nnz - array containing the number of nonzero blocks in the various block rows
3749: (possibly different for each block row) or `NULL`
3751: Output Parameter:
3752: . A - the matrix
3754: Options Database Keys:
3755: + -mat_no_unroll - uses code that does not unroll the loops in the block calculations (much slower)
3756: - -mat_block_size - size of the blocks to use
3758: Level: intermediate
3760: Notes:
3761: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
3762: MatXXXXSetPreallocation() paradigm instead of this routine directly.
3763: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
3765: The number of rows and columns must be divisible by blocksize.
3767: If the `nnz` parameter is given then the `nz` parameter is ignored
3769: A nonzero block is any block that as 1 or more nonzeros in it
3771: The `MATSEQBAIJ` format is fully compatible with standard Fortran
3772: storage. That is, the stored row and column indices can begin at
3773: either one (as in Fortran) or zero.
3775: Specify the preallocated storage with either `nz` or `nnz` (not both).
3776: Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3777: allocation. See [Sparse Matrices](sec_matsparse) for details.
3778: matrices.
3780: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateBAIJ()`
3781: @*/
3782: PetscErrorCode MatCreateSeqBAIJ(MPI_Comm comm, PetscInt bs, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
3783: {
3784: PetscFunctionBegin;
3785: PetscCall(MatCreate(comm, A));
3786: PetscCall(MatSetSizes(*A, m, n, m, n));
3787: PetscCall(MatSetType(*A, MATSEQBAIJ));
3788: PetscCall(MatSeqBAIJSetPreallocation(*A, bs, nz, (PetscInt *)nnz));
3789: PetscFunctionReturn(PETSC_SUCCESS);
3790: }
3792: /*@
3793: MatSeqBAIJSetPreallocation - Sets the block size and expected nonzeros
3794: per row in the matrix. For good matrix assembly performance the
3795: user should preallocate the matrix storage by setting the parameter `nz`
3796: (or the array `nnz`).
3798: Collective
3800: Input Parameters:
3801: + B - the matrix
3802: . bs - size of block, the blocks are ALWAYS square. One can use `MatSetBlockSizes()` to set a different row and column blocksize but the row
3803: blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with `MatCreateVecs()`
3804: . nz - number of block nonzeros per block row (same for all rows)
3805: - nnz - array containing the number of block nonzeros in the various block rows
3806: (possibly different for each block row) or `NULL`
3808: Options Database Keys:
3809: + -mat_no_unroll - uses code that does not unroll the loops in the block calculations (much slower)
3810: - -mat_block_size - size of the blocks to use
3812: Level: intermediate
3814: Notes:
3815: If the `nnz` parameter is given then the `nz` parameter is ignored
3817: You can call `MatGetInfo()` to get information on how effective the preallocation was;
3818: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3819: You can also run with the option `-info` and look for messages with the string
3820: malloc in them to see if additional memory allocation was needed.
3822: The `MATSEQBAIJ` format is fully compatible with standard Fortran
3823: storage. That is, the stored row and column indices can begin at
3824: either one (as in Fortran) or zero.
3826: Specify the preallocated storage with either `nz` or `nnz` (not both).
3827: Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3828: allocation. See [Sparse Matrices](sec_matsparse) for details.
3830: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateBAIJ()`, `MatGetInfo()`
3831: @*/
3832: PetscErrorCode MatSeqBAIJSetPreallocation(Mat B, PetscInt bs, PetscInt nz, const PetscInt nnz[])
3833: {
3834: PetscFunctionBegin;
3838: PetscTryMethod(B, "MatSeqBAIJSetPreallocation_C", (Mat, PetscInt, PetscInt, const PetscInt[]), (B, bs, nz, nnz));
3839: PetscFunctionReturn(PETSC_SUCCESS);
3840: }
3842: /*@
3843: MatSeqBAIJSetPreallocationCSR - Creates a sparse sequential matrix in `MATSEQBAIJ` format using the given nonzero structure and (optional) numerical values
3845: Collective
3847: Input Parameters:
3848: + B - the matrix
3849: . bs - the blocksize
3850: . i - the indices into `j` for the start of each local row (indices start with zero)
3851: . j - the column indices for each local row (indices start with zero) these must be sorted for each row
3852: - v - optional values in the matrix, use `NULL` if not provided
3854: Level: advanced
3856: Notes:
3857: The `i`,`j`,`v` values are COPIED with this routine; to avoid the copy use `MatCreateSeqBAIJWithArrays()`
3859: The order of the entries in values is specified by the `MatOption` `MAT_ROW_ORIENTED`. For example, C programs
3860: may want to use the default `MAT_ROW_ORIENTED` of `PETSC_TRUE` and use an array v[nnz][bs][bs] where the second index is
3861: over rows within a block and the last index is over columns within a block row. Fortran programs will likely set
3862: `MAT_ROW_ORIENTED` of `PETSC_FALSE` and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a
3863: block column and the second index is over columns within a block.
3865: 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
3867: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqBAIJ()`, `MatSetValues()`, `MatSeqBAIJSetPreallocation()`, `MATSEQBAIJ`
3868: @*/
3869: PetscErrorCode MatSeqBAIJSetPreallocationCSR(Mat B, PetscInt bs, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
3870: {
3871: PetscFunctionBegin;
3875: PetscTryMethod(B, "MatSeqBAIJSetPreallocationCSR_C", (Mat, PetscInt, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, bs, i, j, v));
3876: PetscFunctionReturn(PETSC_SUCCESS);
3877: }
3879: /*@
3880: MatCreateSeqBAIJWithArrays - Creates a `MATSEQBAIJ` matrix using matrix elements provided by the user.
3882: Collective
3884: Input Parameters:
3885: + comm - must be an MPI communicator of size 1
3886: . bs - size of block
3887: . m - number of rows
3888: . n - number of columns
3889: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row block row of the matrix
3890: . j - column indices
3891: - a - matrix values
3893: Output Parameter:
3894: . mat - the matrix
3896: Level: advanced
3898: Notes:
3899: The `i`, `j`, and `a` arrays are not copied by this routine, the user must free these arrays
3900: once the matrix is destroyed
3902: You cannot set new nonzero locations into this matrix, that will generate an error.
3904: The `i` and `j` indices are 0 based
3906: When block size is greater than 1 the matrix values must be stored using the `MATSEQBAIJ` storage format
3908: The order of the entries in values is the same as the block compressed sparse row storage format; that is, it is
3909: the same as a three dimensional array in Fortran values(bs,bs,nnz) that contains the first column of the first
3910: block, followed by the second column of the first block etc etc. That is, the blocks are contiguous in memory
3911: with column-major ordering within blocks.
3913: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateBAIJ()`, `MatCreateSeqBAIJ()`
3914: @*/
3915: PetscErrorCode MatCreateSeqBAIJWithArrays(MPI_Comm comm, PetscInt bs, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat)
3916: {
3917: Mat_SeqBAIJ *baij;
3919: PetscFunctionBegin;
3920: PetscCheck(bs == 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "block size %" PetscInt_FMT " > 1 is not supported yet", bs);
3921: if (m > 0) PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
3923: PetscCall(MatCreate(comm, mat));
3924: PetscCall(MatSetSizes(*mat, m, n, m, n));
3925: PetscCall(MatSetType(*mat, MATSEQBAIJ));
3926: PetscCall(MatSeqBAIJSetPreallocation(*mat, bs, MAT_SKIP_ALLOCATION, NULL));
3927: baij = (Mat_SeqBAIJ *)(*mat)->data;
3928: PetscCall(PetscMalloc2(m, &baij->imax, m, &baij->ilen));
3930: baij->i = i;
3931: baij->j = j;
3932: baij->a = a;
3934: baij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
3935: baij->free_a = PETSC_FALSE;
3936: baij->free_ij = PETSC_FALSE;
3937: baij->free_imax_ilen = PETSC_TRUE;
3939: for (PetscInt ii = 0; ii < m; ii++) {
3940: const PetscInt row_len = i[ii + 1] - i[ii];
3942: baij->ilen[ii] = baij->imax[ii] = row_len;
3943: PetscCheck(row_len >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative row length in i (row indices) row = %" PetscInt_FMT " length = %" PetscInt_FMT, ii, row_len);
3944: }
3945: if (PetscDefined(USE_DEBUG)) {
3946: for (PetscInt ii = 0; ii < baij->i[m]; ii++) {
3947: PetscCheck(j[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative column index at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]);
3948: PetscCheck(j[ii] <= n - 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column index to large at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]);
3949: }
3950: }
3952: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3953: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3954: PetscFunctionReturn(PETSC_SUCCESS);
3955: }
3957: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqBAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
3958: {
3959: PetscFunctionBegin;
3960: PetscCall(MatCreateMPIMatConcatenateSeqMat_MPIBAIJ(comm, inmat, n, scall, outmat));
3961: PetscFunctionReturn(PETSC_SUCCESS);
3962: }