sbaij/seq/sbaijfact7.c

7   Mat_SeqSBAIJ   *a = (Mat_SeqSBAIJ *)A->data, *b = (Mat_SeqSBAIJ *)C->data;
8   IS              perm = b->row;
9   const PetscInt *ai, *aj, *perm_ptr, mbs = a->mbs, *bi = b->i, *bj = b->j;
11   MatScalar      *ba = b->a, *aa, *ap, *dk, *uik;
12   MatScalar      *u, *d, *rtmp, *rtmp_ptr, work[25];
13   PetscReal       shift = info->shiftamount;
18   allowzeropivot = PetscNot(A->erroriffailure);
28   if (!a->permute) {
29     ai = a->i;
30     aj = a->j;
31     aa = a->a;
33     ai = a->inew;
34     aj = a->jnew;
36     PetscCall(PetscArraycpy(aa, a->a, 25 * ai[mbs]));
38     PetscCall(PetscArraycpy(a2anew, a->a2anew, ai[mbs]));
54         /* transform column-oriented blocks that lie in the lower triangle to row-oriented blocks */
56           /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */
57           ap = aa + j * 25;                       /* ptr to the beginning of j-th block of aa */
58           for (k = 0; k < 25; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */
59           for (k = 0; k < 5; k++) {               /* j-th block of aa <- dk^T */
70     /*initialize k-th row with elements nonzero in row perm(k) of A */
82     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
90       ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */
92       /* uik = -inv(Di)*U_bar(i,k) */
93       d = ba + i * 25;
96       uik[0] = -(d[0] * u[0] + d[5] * u[1] + d[10] * u[2] + d[15] * u[3] + d[20] * u[4]);
97       uik[1] = -(d[1] * u[0] + d[6] * u[1] + d[11] * u[2] + d[16] * u[3] + d[21] * u[4]);
98       uik[2] = -(d[2] * u[0] + d[7] * u[1] + d[12] * u[2] + d[17] * u[3] + d[22] * u[4]);
99       uik[3] = -(d[3] * u[0] + d[8] * u[1] + d[13] * u[2] + d[18] * u[3] + d[23] * u[4]);
100       uik[4] = -(d[4] * u[0] + d[9] * u[1] + d[14] * u[2] + d[19] * u[3] + d[24] * u[4]);
102       uik[5] = -(d[0] * u[5] + d[5] * u[6] + d[10] * u[7] + d[15] * u[8] + d[20] * u[9]);
103       uik[6] = -(d[1] * u[5] + d[6] * u[6] + d[11] * u[7] + d[16] * u[8] + d[21] * u[9]);
104       uik[7] = -(d[2] * u[5] + d[7] * u[6] + d[12] * u[7] + d[17] * u[8] + d[22] * u[9]);
105       uik[8] = -(d[3] * u[5] + d[8] * u[6] + d[13] * u[7] + d[18] * u[8] + d[23] * u[9]);
106       uik[9] = -(d[4] * u[5] + d[9] * u[6] + d[14] * u[7] + d[19] * u[8] + d[24] * u[9]);
108       uik[10] = -(d[0] * u[10] + d[5] * u[11] + d[10] * u[12] + d[15] * u[13] + d[20] * u[14]);
109       uik[11] = -(d[1] * u[10] + d[6] * u[11] + d[11] * u[12] + d[16] * u[13] + d[21] * u[14]);
110       uik[12] = -(d[2] * u[10] + d[7] * u[11] + d[12] * u[12] + d[17] * u[13] + d[22] * u[14]);
111       uik[13] = -(d[3] * u[10] + d[8] * u[11] + d[13] * u[12] + d[18] * u[13] + d[23] * u[14]);
112       uik[14] = -(d[4] * u[10] + d[9] * u[11] + d[14] * u[12] + d[19] * u[13] + d[24] * u[14]);
114       uik[15] = -(d[0] * u[15] + d[5] * u[16] + d[10] * u[17] + d[15] * u[18] + d[20] * u[19]);
115       uik[16] = -(d[1] * u[15] + d[6] * u[16] + d[11] * u[17] + d[16] * u[18] + d[21] * u[19]);
116       uik[17] = -(d[2] * u[15] + d[7] * u[16] + d[12] * u[17] + d[17] * u[18] + d[22] * u[19]);
117       uik[18] = -(d[3] * u[15] + d[8] * u[16] + d[13] * u[17] + d[18] * u[18] + d[23] * u[19]);
118       uik[19] = -(d[4] * u[15] + d[9] * u[16] + d[14] * u[17] + d[19] * u[18] + d[24] * u[19]);
120       uik[20] = -(d[0] * u[20] + d[5] * u[21] + d[10] * u[22] + d[15] * u[23] + d[20] * u[24]);
121       uik[21] = -(d[1] * u[20] + d[6] * u[21] + d[11] * u[22] + d[16] * u[23] + d[21] * u[24]);
122       uik[22] = -(d[2] * u[20] + d[7] * u[21] + d[12] * u[22] + d[17] * u[23] + d[22] * u[24]);
123       uik[23] = -(d[3] * u[20] + d[8] * u[21] + d[13] * u[22] + d[18] * u[23] + d[23] * u[24]);
124       uik[24] = -(d[4] * u[20] + d[9] * u[21] + d[14] * u[22] + d[19] * u[23] + d[24] * u[24]);
126       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
159       /* update -U(i,k) */
162       /* add multiple of row i to k-th row ... */
167           /* rtmp += -U(i,k)^T * U_bar(i,j) */
200         PetscCall(PetscLogFlops(2.0 * 125.0 * (jmax - jmin)));
203         il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */
211     /* save nonzero entries in k-th row of U ... */
214     d = ba + k * 25;
215     PetscCall(PetscArraycpy(d, dk, 25));
216     PetscCall(PetscKernel_A_gets_inverse_A_5(d, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
217     if (zeropivotdetected) C->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
232       /* ... add k to row list for first nonzero entry in k-th row */
243   if (a->permute) PetscCall(PetscFree(aa));
247   C->ops->solve          = MatSolve_SeqSBAIJ_5_inplace;
248   C->ops->solvetranspose = MatSolve_SeqSBAIJ_5_inplace;
249   C->assembled           = PETSC_TRUE;
250   C->preallocated        = PETSC_TRUE;
252   PetscCall(PetscLogFlops(1.3333 * 125 * b->mbs)); /* from inverting diagonal blocks */