sbaij/seq/sbaijfact12.c

9   Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data, *b = (Mat_SeqSBAIJ *)C->data;
10   PetscInt      i, j, mbs = a->mbs, *bi = b->i, *bj = b->j;
12   MatScalar    *ba = b->a, *aa, *ap, *dk, *uik;
13   MatScalar    *u, *d, *w, *wp, u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12;
17   PetscReal     shift = info->shiftamount;
22   allowzeropivot = PetscNot(A->erroriffailure);
29   ai = a->i;
30   aj = a->j;
31   aa = a->a;
35     /*initialize k-th row with elements nonzero in row k of A */
48     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
56       ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */
58       /* uik = -inv(Di)*U_bar(i,k) */
59       d = ba + i * 49;
112       uik[0] = -(d[0] * u0 + d[7] * u1 + d[14] * u2 + d[21] * u3 + d[28] * u4 + d[35] * u5 + d[42] * u6);
113       uik[1] = -(d[1] * u0 + d[8] * u1 + d[15] * u2 + d[22] * u3 + d[29] * u4 + d[36] * u5 + d[43] * u6);
114       uik[2] = -(d[2] * u0 + d[9] * u1 + d[16] * u2 + d[23] * u3 + d[30] * u4 + d[37] * u5 + d[44] * u6);
115       uik[3] = -(d[3] * u0 + d[10] * u1 + d[17] * u2 + d[24] * u3 + d[31] * u4 + d[38] * u5 + d[45] * u6);
116       uik[4] = -(d[4] * u0 + d[11] * u1 + d[18] * u2 + d[25] * u3 + d[32] * u4 + d[39] * u5 + d[46] * u6);
117       uik[5] = -(d[5] * u0 + d[12] * u1 + d[19] * u2 + d[26] * u3 + d[33] * u4 + d[40] * u5 + d[47] * u6);
118       uik[6] = -(d[6] * u0 + d[13] * u1 + d[20] * u2 + d[27] * u3 + d[34] * u4 + d[41] * u5 + d[48] * u6);
120       uik[7]  = -(d[0] * u7 + d[7] * u8 + d[14] * u9 + d[21] * u10 + d[28] * u11 + d[35] * u12 + d[42] * u13);
121       uik[8]  = -(d[1] * u7 + d[8] * u8 + d[15] * u9 + d[22] * u10 + d[29] * u11 + d[36] * u12 + d[43] * u13);
122       uik[9]  = -(d[2] * u7 + d[9] * u8 + d[16] * u9 + d[23] * u10 + d[30] * u11 + d[37] * u12 + d[44] * u13);
123       uik[10] = -(d[3] * u7 + d[10] * u8 + d[17] * u9 + d[24] * u10 + d[31] * u11 + d[38] * u12 + d[45] * u13);
124       uik[11] = -(d[4] * u7 + d[11] * u8 + d[18] * u9 + d[25] * u10 + d[32] * u11 + d[39] * u12 + d[46] * u13);
125       uik[12] = -(d[5] * u7 + d[12] * u8 + d[19] * u9 + d[26] * u10 + d[33] * u11 + d[40] * u12 + d[47] * u13);
126       uik[13] = -(d[6] * u7 + d[13] * u8 + d[20] * u9 + d[27] * u10 + d[34] * u11 + d[41] * u12 + d[48] * u13);
128       uik[14] = -(d[0] * u14 + d[7] * u15 + d[14] * u16 + d[21] * u17 + d[28] * u18 + d[35] * u19 + d[42] * u20);
129       uik[15] = -(d[1] * u14 + d[8] * u15 + d[15] * u16 + d[22] * u17 + d[29] * u18 + d[36] * u19 + d[43] * u20);
130       uik[16] = -(d[2] * u14 + d[9] * u15 + d[16] * u16 + d[23] * u17 + d[30] * u18 + d[37] * u19 + d[44] * u20);
131       uik[17] = -(d[3] * u14 + d[10] * u15 + d[17] * u16 + d[24] * u17 + d[31] * u18 + d[38] * u19 + d[45] * u20);
132       uik[18] = -(d[4] * u14 + d[11] * u15 + d[18] * u16 + d[25] * u17 + d[32] * u18 + d[39] * u19 + d[46] * u20);
133       uik[19] = -(d[5] * u14 + d[12] * u15 + d[19] * u16 + d[26] * u17 + d[33] * u18 + d[40] * u19 + d[47] * u20);
134       uik[20] = -(d[6] * u14 + d[13] * u15 + d[20] * u16 + d[27] * u17 + d[34] * u18 + d[41] * u19 + d[48] * u20);
136       uik[21] = -(d[0] * u21 + d[7] * u22 + d[14] * u23 + d[21] * u24 + d[28] * u25 + d[35] * u26 + d[42] * u27);
137       uik[22] = -(d[1] * u21 + d[8] * u22 + d[15] * u23 + d[22] * u24 + d[29] * u25 + d[36] * u26 + d[43] * u27);
138       uik[23] = -(d[2] * u21 + d[9] * u22 + d[16] * u23 + d[23] * u24 + d[30] * u25 + d[37] * u26 + d[44] * u27);
139       uik[24] = -(d[3] * u21 + d[10] * u22 + d[17] * u23 + d[24] * u24 + d[31] * u25 + d[38] * u26 + d[45] * u27);
140       uik[25] = -(d[4] * u21 + d[11] * u22 + d[18] * u23 + d[25] * u24 + d[32] * u25 + d[39] * u26 + d[46] * u27);
141       uik[26] = -(d[5] * u21 + d[12] * u22 + d[19] * u23 + d[26] * u24 + d[33] * u25 + d[40] * u26 + d[47] * u27);
142       uik[27] = -(d[6] * u21 + d[13] * u22 + d[20] * u23 + d[27] * u24 + d[34] * u25 + d[41] * u26 + d[48] * u27);
144       uik[28] = -(d[0] * u28 + d[7] * u29 + d[14] * u30 + d[21] * u31 + d[28] * u32 + d[35] * u33 + d[42] * u34);
145       uik[29] = -(d[1] * u28 + d[8] * u29 + d[15] * u30 + d[22] * u31 + d[29] * u32 + d[36] * u33 + d[43] * u34);
146       uik[30] = -(d[2] * u28 + d[9] * u29 + d[16] * u30 + d[23] * u31 + d[30] * u32 + d[37] * u33 + d[44] * u34);
147       uik[31] = -(d[3] * u28 + d[10] * u29 + d[17] * u30 + d[24] * u31 + d[31] * u32 + d[38] * u33 + d[45] * u34);
148       uik[32] = -(d[4] * u28 + d[11] * u29 + d[18] * u30 + d[25] * u31 + d[32] * u32 + d[39] * u33 + d[46] * u34);
149       uik[33] = -(d[5] * u28 + d[12] * u29 + d[19] * u30 + d[26] * u31 + d[33] * u32 + d[40] * u33 + d[47] * u34);
150       uik[34] = -(d[6] * u28 + d[13] * u29 + d[20] * u30 + d[27] * u31 + d[34] * u32 + d[41] * u33 + d[48] * u34);
152       uik[35] = -(d[0] * u35 + d[7] * u36 + d[14] * u37 + d[21] * u38 + d[28] * u39 + d[35] * u40 + d[42] * u41);
153       uik[36] = -(d[1] * u35 + d[8] * u36 + d[15] * u37 + d[22] * u38 + d[29] * u39 + d[36] * u40 + d[43] * u41);
154       uik[37] = -(d[2] * u35 + d[9] * u36 + d[16] * u37 + d[23] * u38 + d[30] * u39 + d[37] * u40 + d[44] * u41);
155       uik[38] = -(d[3] * u35 + d[10] * u36 + d[17] * u37 + d[24] * u38 + d[31] * u39 + d[38] * u40 + d[45] * u41);
156       uik[39] = -(d[4] * u35 + d[11] * u36 + d[18] * u37 + d[25] * u38 + d[32] * u39 + d[39] * u40 + d[46] * u41);
157       uik[40] = -(d[5] * u35 + d[12] * u36 + d[19] * u37 + d[26] * u38 + d[33] * u39 + d[40] * u40 + d[47] * u41);
158       uik[41] = -(d[6] * u35 + d[13] * u36 + d[20] * u37 + d[27] * u38 + d[34] * u39 + d[41] * u40 + d[48] * u41);
160       uik[42] = -(d[0] * u42 + d[7] * u43 + d[14] * u44 + d[21] * u45 + d[28] * u46 + d[35] * u47 + d[42] * u48);
161       uik[43] = -(d[1] * u42 + d[8] * u43 + d[15] * u44 + d[22] * u45 + d[29] * u46 + d[36] * u47 + d[43] * u48);
162       uik[44] = -(d[2] * u42 + d[9] * u43 + d[16] * u44 + d[23] * u45 + d[30] * u46 + d[37] * u47 + d[44] * u48);
163       uik[45] = -(d[3] * u42 + d[10] * u43 + d[17] * u44 + d[24] * u45 + d[31] * u46 + d[38] * u47 + d[45] * u48);
164       uik[46] = -(d[4] * u42 + d[11] * u43 + d[18] * u44 + d[25] * u45 + d[32] * u46 + d[39] * u47 + d[46] * u48);
165       uik[47] = -(d[5] * u42 + d[12] * u43 + d[19] * u44 + d[26] * u45 + d[33] * u46 + d[40] * u47 + d[47] * u48);
166       uik[48] = -(d[6] * u42 + d[13] * u43 + d[20] * u44 + d[27] * u45 + d[34] * u46 + d[41] * u47 + d[48] * u48);
168       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
227       /* update -U(i,k) */
230       /* add multiple of row i to k-th row ... */
235           /* w += -U(i,k)^T * U_bar(i,j) */
345         PetscCall(PetscLogFlops(2.0 * 343.0 * (jmax - jmin)));
348         il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */
356     /* save nonzero entries in k-th row of U ... */
359     d = ba + k * 49;
360     PetscCall(PetscArraycpy(d, dk, 49));
361     PetscCall(PetscKernel_A_gets_inverse_A_7(d, shift, allowzeropivot, &zeropivotdetected));
362     if (zeropivotdetected) C->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
377       /* ... add k to row list for first nonzero entry in k-th row */
389   C->ops->solve          = MatSolve_SeqSBAIJ_7_NaturalOrdering_inplace;
390   C->ops->solvetranspose = MatSolve_SeqSBAIJ_7_NaturalOrdering_inplace;
391   C->ops->forwardsolve   = MatForwardSolve_SeqSBAIJ_7_NaturalOrdering_inplace;
392   C->ops->backwardsolve  = MatBackwardSolve_SeqSBAIJ_7_NaturalOrdering_inplace;
393   C->assembled           = PETSC_TRUE;
394   C->preallocated        = PETSC_TRUE;
396   PetscCall(PetscLogFlops(1.3333 * 343 * b->mbs)); /* from inverting diagonal blocks */