1c6db04a5SJed Brown #include <../src/mat/impls/sbaij/seq/sbaij.h>
2af0996ceSBarry Smith #include <petsc/private/kernels/blockinvert.h>
381278733SSatish Balay
481278733SSatish Balay /* Version for when blocks are 6 by 6 */
MatCholeskyFactorNumeric_SeqSBAIJ_6(Mat C,Mat A,const MatFactorInfo * info)5d71ae5a4SJacob Faibussowitsch PetscErrorCode MatCholeskyFactorNumeric_SeqSBAIJ_6(Mat C, Mat A, const MatFactorInfo *info)
6d71ae5a4SJacob Faibussowitsch {
781278733SSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data, *b = (Mat_SeqSBAIJ *)C->data;
881278733SSatish Balay IS perm = b->row;
95d0c19d7SBarry Smith const PetscInt *ai, *aj, *perm_ptr, mbs = a->mbs, *bi = b->i, *bj = b->j;
105d0c19d7SBarry Smith PetscInt i, j, *a2anew, k, k1, jmin, jmax, *jl, *il, vj, nexti, ili;
1181278733SSatish Balay MatScalar *ba = b->a, *aa, *ap, *dk, *uik;
121b3064deSBarry Smith MatScalar *u, *d, *w, *wp, u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12;
131b3064deSBarry Smith MatScalar u13, u14, u15, u16, u17, u18, u19, u20, u21, u22, u23, u24, u25, u26, u27;
141b3064deSBarry Smith MatScalar u28, u29, u30, u31, u32, u33, u34, u35;
15182b8fbaSHong Zhang PetscReal shift = info->shiftamount;
16a455e926SHong Zhang PetscBool allowzeropivot, zeropivotdetected;
1781278733SSatish Balay
1881278733SSatish Balay PetscFunctionBegin;
1981278733SSatish Balay /* initialization */
200164db54SHong Zhang allowzeropivot = PetscNot(A->erroriffailure);
219566063dSJacob Faibussowitsch PetscCall(PetscCalloc1(36 * mbs, &w));
229566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(mbs, &il, mbs, &jl));
236df5ee2eSHong Zhang il[0] = 0;
246df5ee2eSHong Zhang for (i = 0; i < mbs; i++) jl[i] = mbs;
256df5ee2eSHong Zhang
269566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(36, &dk, 36, &uik));
279566063dSJacob Faibussowitsch PetscCall(ISGetIndices(perm, &perm_ptr));
2881278733SSatish Balay
2981278733SSatish Balay /* check permutation */
3081278733SSatish Balay if (!a->permute) {
319371c9d4SSatish Balay ai = a->i;
329371c9d4SSatish Balay aj = a->j;
339371c9d4SSatish Balay aa = a->a;
3481278733SSatish Balay } else {
359371c9d4SSatish Balay ai = a->inew;
369371c9d4SSatish Balay aj = a->jnew;
379566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(36 * ai[mbs], &aa));
389566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(aa, a->a, 36 * ai[mbs]));
399566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(ai[mbs], &a2anew));
409566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(a2anew, a->a2anew, ai[mbs]));
4181278733SSatish Balay
4281278733SSatish Balay for (i = 0; i < mbs; i++) {
439371c9d4SSatish Balay jmin = ai[i];
449371c9d4SSatish Balay jmax = ai[i + 1];
4581278733SSatish Balay for (j = jmin; j < jmax; j++) {
4681278733SSatish Balay while (a2anew[j] != j) {
479371c9d4SSatish Balay k = a2anew[j];
489371c9d4SSatish Balay a2anew[j] = a2anew[k];
499371c9d4SSatish Balay a2anew[k] = k;
5081278733SSatish Balay for (k1 = 0; k1 < 36; k1++) {
5181278733SSatish Balay dk[k1] = aa[k * 36 + k1];
5281278733SSatish Balay aa[k * 36 + k1] = aa[j * 36 + k1];
5381278733SSatish Balay aa[j * 36 + k1] = dk[k1];
5481278733SSatish Balay }
5581278733SSatish Balay }
56*5e116b59SBarry Smith /* transform column-oriented blocks that lie in the lower triangle to row-oriented blocks */
5781278733SSatish Balay if (i > aj[j]) {
5881278733SSatish Balay /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */
5981278733SSatish Balay ap = aa + j * 36; /* ptr to the beginning of j-th block of aa */
6081278733SSatish Balay for (k = 0; k < 36; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */
6181278733SSatish Balay for (k = 0; k < 6; k++) { /* j-th block of aa <- dk^T */
6281278733SSatish Balay for (k1 = 0; k1 < 6; k1++) *ap++ = dk[k + 6 * k1];
6381278733SSatish Balay }
6481278733SSatish Balay }
6581278733SSatish Balay }
6681278733SSatish Balay }
679566063dSJacob Faibussowitsch PetscCall(PetscFree(a2anew));
6881278733SSatish Balay }
6981278733SSatish Balay
7081278733SSatish Balay /* for each row k */
7181278733SSatish Balay for (k = 0; k < mbs; k++) {
7281278733SSatish Balay /*initialize k-th row with elements nonzero in row perm(k) of A */
739371c9d4SSatish Balay jmin = ai[perm_ptr[k]];
749371c9d4SSatish Balay jmax = ai[perm_ptr[k] + 1];
7581278733SSatish Balay if (jmin < jmax) {
7681278733SSatish Balay ap = aa + jmin * 36;
7781278733SSatish Balay for (j = jmin; j < jmax; j++) {
7881278733SSatish Balay vj = perm_ptr[aj[j]]; /* block col. index */
7981278733SSatish Balay wp = w + vj * 36;
8081278733SSatish Balay for (i = 0; i < 36; i++) *wp++ = *ap++;
8181278733SSatish Balay }
8281278733SSatish Balay }
8381278733SSatish Balay
8481278733SSatish Balay /* modify k-th row by adding in those rows i with U(i,k) != 0 */
859566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(dk, w + k * 36, 36));
8681278733SSatish Balay i = jl[k]; /* first row to be added to k_th row */
8781278733SSatish Balay
8881278733SSatish Balay while (i < mbs) {
8981278733SSatish Balay nexti = jl[i]; /* next row to be added to k_th row */
9081278733SSatish Balay
9181278733SSatish Balay /* compute multiplier */
9281278733SSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */
9381278733SSatish Balay
9481278733SSatish Balay /* uik = -inv(Di)*U_bar(i,k) */
9581278733SSatish Balay d = ba + i * 36;
9681278733SSatish Balay u = ba + ili * 36;
9781278733SSatish Balay
989371c9d4SSatish Balay u0 = u[0];
999371c9d4SSatish Balay u1 = u[1];
1009371c9d4SSatish Balay u2 = u[2];
1019371c9d4SSatish Balay u3 = u[3];
1029371c9d4SSatish Balay u4 = u[4];
1039371c9d4SSatish Balay u5 = u[5];
1049371c9d4SSatish Balay u6 = u[6];
1059371c9d4SSatish Balay u7 = u[7];
1069371c9d4SSatish Balay u8 = u[8];
1079371c9d4SSatish Balay u9 = u[9];
1089371c9d4SSatish Balay u10 = u[10];
1099371c9d4SSatish Balay u11 = u[11];
1109371c9d4SSatish Balay u12 = u[12];
1119371c9d4SSatish Balay u13 = u[13];
1129371c9d4SSatish Balay u14 = u[14];
1139371c9d4SSatish Balay u15 = u[15];
1149371c9d4SSatish Balay u16 = u[16];
1159371c9d4SSatish Balay u17 = u[17];
1169371c9d4SSatish Balay u18 = u[18];
1179371c9d4SSatish Balay u19 = u[19];
1189371c9d4SSatish Balay u20 = u[20];
1199371c9d4SSatish Balay u21 = u[21];
1209371c9d4SSatish Balay u22 = u[22];
1219371c9d4SSatish Balay u23 = u[23];
1229371c9d4SSatish Balay u24 = u[24];
1239371c9d4SSatish Balay u25 = u[25];
1249371c9d4SSatish Balay u26 = u[26];
1259371c9d4SSatish Balay u27 = u[27];
1269371c9d4SSatish Balay u28 = u[28];
1279371c9d4SSatish Balay u29 = u[29];
1289371c9d4SSatish Balay u30 = u[30];
1299371c9d4SSatish Balay u31 = u[31];
1309371c9d4SSatish Balay u32 = u[32];
1319371c9d4SSatish Balay u33 = u[33];
1329371c9d4SSatish Balay u34 = u[34];
1331b3064deSBarry Smith u35 = u[35];
13481278733SSatish Balay
1351b3064deSBarry Smith uik[0] = -(d[0] * u0 + d[6] * u1 + d[12] * u2 + d[18] * u3 + d[24] * u4 + d[30] * u5);
1361b3064deSBarry Smith uik[1] = -(d[1] * u0 + d[7] * u1 + d[13] * u2 + d[19] * u3 + d[25] * u4 + d[31] * u5);
1371b3064deSBarry Smith uik[2] = -(d[2] * u0 + d[8] * u1 + d[14] * u2 + d[20] * u3 + d[26] * u4 + d[32] * u5);
1381b3064deSBarry Smith uik[3] = -(d[3] * u0 + d[9] * u1 + d[15] * u2 + d[21] * u3 + d[27] * u4 + d[33] * u5);
1391b3064deSBarry Smith uik[4] = -(d[4] * u0 + d[10] * u1 + d[16] * u2 + d[22] * u3 + d[28] * u4 + d[34] * u5);
1401b3064deSBarry Smith uik[5] = -(d[5] * u0 + d[11] * u1 + d[17] * u2 + d[23] * u3 + d[29] * u4 + d[35] * u5);
14181278733SSatish Balay
1421b3064deSBarry Smith uik[6] = -(d[0] * u6 + d[6] * u7 + d[12] * u8 + d[18] * u9 + d[24] * u10 + d[30] * u11);
1431b3064deSBarry Smith uik[7] = -(d[1] * u6 + d[7] * u7 + d[13] * u8 + d[19] * u9 + d[25] * u10 + d[31] * u11);
1441b3064deSBarry Smith uik[8] = -(d[2] * u6 + d[8] * u7 + d[14] * u8 + d[20] * u9 + d[26] * u10 + d[32] * u11);
1451b3064deSBarry Smith uik[9] = -(d[3] * u6 + d[9] * u7 + d[15] * u8 + d[21] * u9 + d[27] * u10 + d[33] * u11);
1461b3064deSBarry Smith uik[10] = -(d[4] * u6 + d[10] * u7 + d[16] * u8 + d[22] * u9 + d[28] * u10 + d[34] * u11);
1471b3064deSBarry Smith uik[11] = -(d[5] * u6 + d[11] * u7 + d[17] * u8 + d[23] * u9 + d[29] * u10 + d[35] * u11);
14881278733SSatish Balay
1491b3064deSBarry Smith uik[12] = -(d[0] * u12 + d[6] * u13 + d[12] * u14 + d[18] * u15 + d[24] * u16 + d[30] * u17);
1501b3064deSBarry Smith uik[13] = -(d[1] * u12 + d[7] * u13 + d[13] * u14 + d[19] * u15 + d[25] * u16 + d[31] * u17);
1511b3064deSBarry Smith uik[14] = -(d[2] * u12 + d[8] * u13 + d[14] * u14 + d[20] * u15 + d[26] * u16 + d[32] * u17);
1521b3064deSBarry Smith uik[15] = -(d[3] * u12 + d[9] * u13 + d[15] * u14 + d[21] * u15 + d[27] * u16 + d[33] * u17);
1531b3064deSBarry Smith uik[16] = -(d[4] * u12 + d[10] * u13 + d[16] * u14 + d[22] * u15 + d[28] * u16 + d[34] * u17);
1541b3064deSBarry Smith uik[17] = -(d[5] * u12 + d[11] * u13 + d[17] * u14 + d[23] * u15 + d[29] * u16 + d[35] * u17);
15581278733SSatish Balay
1561b3064deSBarry Smith uik[18] = -(d[0] * u18 + d[6] * u19 + d[12] * u20 + d[18] * u21 + d[24] * u22 + d[30] * u23);
1571b3064deSBarry Smith uik[19] = -(d[1] * u18 + d[7] * u19 + d[13] * u20 + d[19] * u21 + d[25] * u22 + d[31] * u23);
1581b3064deSBarry Smith uik[20] = -(d[2] * u18 + d[8] * u19 + d[14] * u20 + d[20] * u21 + d[26] * u22 + d[32] * u23);
1591b3064deSBarry Smith uik[21] = -(d[3] * u18 + d[9] * u19 + d[15] * u20 + d[21] * u21 + d[27] * u22 + d[33] * u23);
1601b3064deSBarry Smith uik[22] = -(d[4] * u18 + d[10] * u19 + d[16] * u20 + d[22] * u21 + d[28] * u22 + d[34] * u23);
1611b3064deSBarry Smith uik[23] = -(d[5] * u18 + d[11] * u19 + d[17] * u20 + d[23] * u21 + d[29] * u22 + d[35] * u23);
16281278733SSatish Balay
1631b3064deSBarry Smith uik[24] = -(d[0] * u24 + d[6] * u25 + d[12] * u26 + d[18] * u27 + d[24] * u28 + d[30] * u29);
1641b3064deSBarry Smith uik[25] = -(d[1] * u24 + d[7] * u25 + d[13] * u26 + d[19] * u27 + d[25] * u28 + d[31] * u29);
1651b3064deSBarry Smith uik[26] = -(d[2] * u24 + d[8] * u25 + d[14] * u26 + d[20] * u27 + d[26] * u28 + d[32] * u29);
1661b3064deSBarry Smith uik[27] = -(d[3] * u24 + d[9] * u25 + d[15] * u26 + d[21] * u27 + d[27] * u28 + d[33] * u29);
1671b3064deSBarry Smith uik[28] = -(d[4] * u24 + d[10] * u25 + d[16] * u26 + d[22] * u27 + d[28] * u28 + d[34] * u29);
1681b3064deSBarry Smith uik[29] = -(d[5] * u24 + d[11] * u25 + d[17] * u26 + d[23] * u27 + d[29] * u28 + d[35] * u29);
1691b3064deSBarry Smith
1701b3064deSBarry Smith uik[30] = -(d[0] * u30 + d[6] * u31 + d[12] * u32 + d[18] * u33 + d[24] * u34 + d[30] * u35);
1711b3064deSBarry Smith uik[31] = -(d[1] * u30 + d[7] * u31 + d[13] * u32 + d[19] * u33 + d[25] * u34 + d[31] * u35);
1721b3064deSBarry Smith uik[32] = -(d[2] * u30 + d[8] * u31 + d[14] * u32 + d[20] * u33 + d[26] * u34 + d[32] * u35);
1731b3064deSBarry Smith uik[33] = -(d[3] * u30 + d[9] * u31 + d[15] * u32 + d[21] * u33 + d[27] * u34 + d[33] * u35);
1741b3064deSBarry Smith uik[34] = -(d[4] * u30 + d[10] * u31 + d[16] * u32 + d[22] * u33 + d[28] * u34 + d[34] * u35);
1751b3064deSBarry Smith uik[35] = -(d[5] * u30 + d[11] * u31 + d[17] * u32 + d[23] * u33 + d[29] * u34 + d[35] * u35);
17681278733SSatish Balay
17781278733SSatish Balay /* update D(k) += -U(i,k)^T * U_bar(i,k) */
1781b3064deSBarry Smith dk[0] += uik[0] * u0 + uik[1] * u1 + uik[2] * u2 + uik[3] * u3 + uik[4] * u4 + uik[5] * u5;
1791b3064deSBarry Smith dk[1] += uik[6] * u0 + uik[7] * u1 + uik[8] * u2 + uik[9] * u3 + uik[10] * u4 + uik[11] * u5;
1801b3064deSBarry Smith dk[2] += uik[12] * u0 + uik[13] * u1 + uik[14] * u2 + uik[15] * u3 + uik[16] * u4 + uik[17] * u5;
1811b3064deSBarry Smith dk[3] += uik[18] * u0 + uik[19] * u1 + uik[20] * u2 + uik[21] * u3 + uik[22] * u4 + uik[23] * u5;
1821b3064deSBarry Smith dk[4] += uik[24] * u0 + uik[25] * u1 + uik[26] * u2 + uik[27] * u3 + uik[28] * u4 + uik[29] * u5;
1831b3064deSBarry Smith dk[5] += uik[30] * u0 + uik[31] * u1 + uik[32] * u2 + uik[33] * u3 + uik[34] * u4 + uik[35] * u5;
18481278733SSatish Balay
1851b3064deSBarry Smith dk[6] += uik[0] * u6 + uik[1] * u7 + uik[2] * u8 + uik[3] * u9 + uik[4] * u10 + uik[5] * u11;
1861b3064deSBarry Smith dk[7] += uik[6] * u6 + uik[7] * u7 + uik[8] * u8 + uik[9] * u9 + uik[10] * u10 + uik[11] * u11;
1871b3064deSBarry Smith dk[8] += uik[12] * u6 + uik[13] * u7 + uik[14] * u8 + uik[15] * u9 + uik[16] * u10 + uik[17] * u11;
1881b3064deSBarry Smith dk[9] += uik[18] * u6 + uik[19] * u7 + uik[20] * u8 + uik[21] * u9 + uik[22] * u10 + uik[23] * u11;
1891b3064deSBarry Smith dk[10] += uik[24] * u6 + uik[25] * u7 + uik[26] * u8 + uik[27] * u9 + uik[28] * u10 + uik[29] * u11;
1901b3064deSBarry Smith dk[11] += uik[30] * u6 + uik[31] * u7 + uik[32] * u8 + uik[33] * u9 + uik[34] * u10 + uik[35] * u11;
19181278733SSatish Balay
1921b3064deSBarry Smith dk[12] += uik[0] * u12 + uik[1] * u13 + uik[2] * u14 + uik[3] * u15 + uik[4] * u16 + uik[5] * u17;
1931b3064deSBarry Smith dk[13] += uik[6] * u12 + uik[7] * u13 + uik[8] * u14 + uik[9] * u15 + uik[10] * u16 + uik[11] * u17;
1941b3064deSBarry Smith dk[14] += uik[12] * u12 + uik[13] * u13 + uik[14] * u14 + uik[15] * u15 + uik[16] * u16 + uik[17] * u17;
1951b3064deSBarry Smith dk[15] += uik[18] * u12 + uik[19] * u13 + uik[20] * u14 + uik[21] * u15 + uik[22] * u16 + uik[23] * u17;
1961b3064deSBarry Smith dk[16] += uik[24] * u12 + uik[25] * u13 + uik[26] * u14 + uik[27] * u15 + uik[28] * u16 + uik[29] * u17;
1971b3064deSBarry Smith dk[17] += uik[30] * u12 + uik[31] * u13 + uik[32] * u14 + uik[33] * u15 + uik[34] * u16 + uik[35] * u17;
19881278733SSatish Balay
1991b3064deSBarry Smith dk[18] += uik[0] * u18 + uik[1] * u19 + uik[2] * u20 + uik[3] * u21 + uik[4] * u22 + uik[5] * u23;
2001b3064deSBarry Smith dk[19] += uik[6] * u18 + uik[7] * u19 + uik[8] * u20 + uik[9] * u21 + uik[10] * u22 + uik[11] * u23;
2011b3064deSBarry Smith dk[20] += uik[12] * u18 + uik[13] * u19 + uik[14] * u20 + uik[15] * u21 + uik[16] * u22 + uik[17] * u23;
2021b3064deSBarry Smith dk[21] += uik[18] * u18 + uik[19] * u19 + uik[20] * u20 + uik[21] * u21 + uik[22] * u22 + uik[23] * u23;
2031b3064deSBarry Smith dk[22] += uik[24] * u18 + uik[25] * u19 + uik[26] * u20 + uik[27] * u21 + uik[28] * u22 + uik[29] * u23;
2041b3064deSBarry Smith dk[23] += uik[30] * u18 + uik[31] * u19 + uik[32] * u20 + uik[33] * u21 + uik[34] * u22 + uik[35] * u23;
20581278733SSatish Balay
2061b3064deSBarry Smith dk[24] += uik[0] * u24 + uik[1] * u25 + uik[2] * u26 + uik[3] * u27 + uik[4] * u28 + uik[5] * u29;
2071b3064deSBarry Smith dk[25] += uik[6] * u24 + uik[7] * u25 + uik[8] * u26 + uik[9] * u27 + uik[10] * u28 + uik[11] * u29;
2081b3064deSBarry Smith dk[26] += uik[12] * u24 + uik[13] * u25 + uik[14] * u26 + uik[15] * u27 + uik[16] * u28 + uik[17] * u29;
2091b3064deSBarry Smith dk[27] += uik[18] * u24 + uik[19] * u25 + uik[20] * u26 + uik[21] * u27 + uik[22] * u28 + uik[23] * u29;
2101b3064deSBarry Smith dk[28] += uik[24] * u24 + uik[25] * u25 + uik[26] * u26 + uik[27] * u27 + uik[28] * u28 + uik[29] * u29;
2111b3064deSBarry Smith dk[29] += uik[30] * u24 + uik[31] * u25 + uik[32] * u26 + uik[33] * u27 + uik[34] * u28 + uik[35] * u29;
21281278733SSatish Balay
2131b3064deSBarry Smith dk[30] += uik[0] * u30 + uik[1] * u31 + uik[2] * u32 + uik[3] * u33 + uik[4] * u34 + uik[5] * u35;
2141b3064deSBarry Smith dk[31] += uik[6] * u30 + uik[7] * u31 + uik[8] * u32 + uik[9] * u33 + uik[10] * u34 + uik[11] * u35;
2151b3064deSBarry Smith dk[32] += uik[12] * u30 + uik[13] * u31 + uik[14] * u32 + uik[15] * u33 + uik[16] * u34 + uik[17] * u35;
2161b3064deSBarry Smith dk[33] += uik[18] * u30 + uik[19] * u31 + uik[20] * u32 + uik[21] * u33 + uik[22] * u34 + uik[23] * u35;
2171b3064deSBarry Smith dk[34] += uik[24] * u30 + uik[25] * u31 + uik[26] * u32 + uik[27] * u33 + uik[28] * u34 + uik[29] * u35;
2181b3064deSBarry Smith dk[35] += uik[30] * u30 + uik[31] * u31 + uik[32] * u32 + uik[33] * u33 + uik[34] * u34 + uik[35] * u35;
21981278733SSatish Balay
2209566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(216.0 * 4.0));
221187a9f4bSHong Zhang
22281278733SSatish Balay /* update -U(i,k) */
2239566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(ba + ili * 36, uik, 36));
22481278733SSatish Balay
22581278733SSatish Balay /* add multiple of row i to k-th row ... */
2269371c9d4SSatish Balay jmin = ili + 1;
2279371c9d4SSatish Balay jmax = bi[i + 1];
22881278733SSatish Balay if (jmin < jmax) {
22981278733SSatish Balay for (j = jmin; j < jmax; j++) {
23081278733SSatish Balay /* w += -U(i,k)^T * U_bar(i,j) */
23181278733SSatish Balay wp = w + bj[j] * 36;
23281278733SSatish Balay u = ba + j * 36;
23381278733SSatish Balay
2349371c9d4SSatish Balay u0 = u[0];
2359371c9d4SSatish Balay u1 = u[1];
2369371c9d4SSatish Balay u2 = u[2];
2379371c9d4SSatish Balay u3 = u[3];
2389371c9d4SSatish Balay u4 = u[4];
2399371c9d4SSatish Balay u5 = u[5];
2409371c9d4SSatish Balay u6 = u[6];
2419371c9d4SSatish Balay u7 = u[7];
2429371c9d4SSatish Balay u8 = u[8];
2439371c9d4SSatish Balay u9 = u[9];
2449371c9d4SSatish Balay u10 = u[10];
2459371c9d4SSatish Balay u11 = u[11];
2469371c9d4SSatish Balay u12 = u[12];
2479371c9d4SSatish Balay u13 = u[13];
2489371c9d4SSatish Balay u14 = u[14];
2499371c9d4SSatish Balay u15 = u[15];
2509371c9d4SSatish Balay u16 = u[16];
2519371c9d4SSatish Balay u17 = u[17];
2529371c9d4SSatish Balay u18 = u[18];
2539371c9d4SSatish Balay u19 = u[19];
2549371c9d4SSatish Balay u20 = u[20];
2559371c9d4SSatish Balay u21 = u[21];
2569371c9d4SSatish Balay u22 = u[22];
2579371c9d4SSatish Balay u23 = u[23];
2589371c9d4SSatish Balay u24 = u[24];
2599371c9d4SSatish Balay u25 = u[25];
2609371c9d4SSatish Balay u26 = u[26];
2619371c9d4SSatish Balay u27 = u[27];
2629371c9d4SSatish Balay u28 = u[28];
2639371c9d4SSatish Balay u29 = u[29];
2649371c9d4SSatish Balay u30 = u[30];
2659371c9d4SSatish Balay u31 = u[31];
2669371c9d4SSatish Balay u32 = u[32];
2679371c9d4SSatish Balay u33 = u[33];
2689371c9d4SSatish Balay u34 = u[34];
2691b3064deSBarry Smith u35 = u[35];
27081278733SSatish Balay
2711b3064deSBarry Smith wp[0] += uik[0] * u0 + uik[1] * u1 + uik[2] * u2 + uik[3] * u3 + uik[4] * u4 + uik[5] * u5;
2721b3064deSBarry Smith wp[1] += uik[6] * u0 + uik[7] * u1 + uik[8] * u2 + uik[9] * u3 + uik[10] * u4 + uik[11] * u5;
2731b3064deSBarry Smith wp[2] += uik[12] * u0 + uik[13] * u1 + uik[14] * u2 + uik[15] * u3 + uik[16] * u4 + uik[17] * u5;
2741b3064deSBarry Smith wp[3] += uik[18] * u0 + uik[19] * u1 + uik[20] * u2 + uik[21] * u3 + uik[22] * u4 + uik[23] * u5;
2751b3064deSBarry Smith wp[4] += uik[24] * u0 + uik[25] * u1 + uik[26] * u2 + uik[27] * u3 + uik[28] * u4 + uik[29] * u5;
2761b3064deSBarry Smith wp[5] += uik[30] * u0 + uik[31] * u1 + uik[32] * u2 + uik[33] * u3 + uik[34] * u4 + uik[35] * u5;
27781278733SSatish Balay
2781b3064deSBarry Smith wp[6] += uik[0] * u6 + uik[1] * u7 + uik[2] * u8 + uik[3] * u9 + uik[4] * u10 + uik[5] * u11;
2791b3064deSBarry Smith wp[7] += uik[6] * u6 + uik[7] * u7 + uik[8] * u8 + uik[9] * u9 + uik[10] * u10 + uik[11] * u11;
2801b3064deSBarry Smith wp[8] += uik[12] * u6 + uik[13] * u7 + uik[14] * u8 + uik[15] * u9 + uik[16] * u10 + uik[17] * u11;
2811b3064deSBarry Smith wp[9] += uik[18] * u6 + uik[19] * u7 + uik[20] * u8 + uik[21] * u9 + uik[22] * u10 + uik[23] * u11;
2821b3064deSBarry Smith wp[10] += uik[24] * u6 + uik[25] * u7 + uik[26] * u8 + uik[27] * u9 + uik[28] * u10 + uik[29] * u11;
2831b3064deSBarry Smith wp[11] += uik[30] * u6 + uik[31] * u7 + uik[32] * u8 + uik[33] * u9 + uik[34] * u10 + uik[35] * u11;
28481278733SSatish Balay
2851b3064deSBarry Smith wp[12] += uik[0] * u12 + uik[1] * u13 + uik[2] * u14 + uik[3] * u15 + uik[4] * u16 + uik[5] * u17;
2861b3064deSBarry Smith wp[13] += uik[6] * u12 + uik[7] * u13 + uik[8] * u14 + uik[9] * u15 + uik[10] * u16 + uik[11] * u17;
2871b3064deSBarry Smith wp[14] += uik[12] * u12 + uik[13] * u13 + uik[14] * u14 + uik[15] * u15 + uik[16] * u16 + uik[17] * u17;
2881b3064deSBarry Smith wp[15] += uik[18] * u12 + uik[19] * u13 + uik[20] * u14 + uik[21] * u15 + uik[22] * u16 + uik[23] * u17;
2891b3064deSBarry Smith wp[16] += uik[24] * u12 + uik[25] * u13 + uik[26] * u14 + uik[27] * u15 + uik[28] * u16 + uik[29] * u17;
2901b3064deSBarry Smith wp[17] += uik[30] * u12 + uik[31] * u13 + uik[32] * u14 + uik[33] * u15 + uik[34] * u16 + uik[35] * u17;
29181278733SSatish Balay
2921b3064deSBarry Smith wp[18] += uik[0] * u18 + uik[1] * u19 + uik[2] * u20 + uik[3] * u21 + uik[4] * u22 + uik[5] * u23;
2931b3064deSBarry Smith wp[19] += uik[6] * u18 + uik[7] * u19 + uik[8] * u20 + uik[9] * u21 + uik[10] * u22 + uik[11] * u23;
2941b3064deSBarry Smith wp[20] += uik[12] * u18 + uik[13] * u19 + uik[14] * u20 + uik[15] * u21 + uik[16] * u22 + uik[17] * u23;
2951b3064deSBarry Smith wp[21] += uik[18] * u18 + uik[19] * u19 + uik[20] * u20 + uik[21] * u21 + uik[22] * u22 + uik[23] * u23;
2961b3064deSBarry Smith wp[22] += uik[24] * u18 + uik[25] * u19 + uik[26] * u20 + uik[27] * u21 + uik[28] * u22 + uik[29] * u23;
2971b3064deSBarry Smith wp[23] += uik[30] * u18 + uik[31] * u19 + uik[32] * u20 + uik[33] * u21 + uik[34] * u22 + uik[35] * u23;
2981b3064deSBarry Smith
2991b3064deSBarry Smith wp[24] += uik[0] * u24 + uik[1] * u25 + uik[2] * u26 + uik[3] * u27 + uik[4] * u28 + uik[5] * u29;
3001b3064deSBarry Smith wp[25] += uik[6] * u24 + uik[7] * u25 + uik[8] * u26 + uik[9] * u27 + uik[10] * u28 + uik[11] * u29;
3011b3064deSBarry Smith wp[26] += uik[12] * u24 + uik[13] * u25 + uik[14] * u26 + uik[15] * u27 + uik[16] * u28 + uik[17] * u29;
3021b3064deSBarry Smith wp[27] += uik[18] * u24 + uik[19] * u25 + uik[20] * u26 + uik[21] * u27 + uik[22] * u28 + uik[23] * u29;
3031b3064deSBarry Smith wp[28] += uik[24] * u24 + uik[25] * u25 + uik[26] * u26 + uik[27] * u27 + uik[28] * u28 + uik[29] * u29;
3041b3064deSBarry Smith wp[29] += uik[30] * u24 + uik[31] * u25 + uik[32] * u26 + uik[33] * u27 + uik[34] * u28 + uik[35] * u29;
3051b3064deSBarry Smith
3061b3064deSBarry Smith wp[30] += uik[0] * u30 + uik[1] * u31 + uik[2] * u32 + uik[3] * u33 + uik[4] * u34 + uik[5] * u35;
3071b3064deSBarry Smith wp[31] += uik[6] * u30 + uik[7] * u31 + uik[8] * u32 + uik[9] * u33 + uik[10] * u34 + uik[11] * u35;
3081b3064deSBarry Smith wp[32] += uik[12] * u30 + uik[13] * u31 + uik[14] * u32 + uik[15] * u33 + uik[16] * u34 + uik[17] * u35;
3091b3064deSBarry Smith wp[33] += uik[18] * u30 + uik[19] * u31 + uik[20] * u32 + uik[21] * u33 + uik[22] * u34 + uik[23] * u35;
3101b3064deSBarry Smith wp[34] += uik[24] * u30 + uik[25] * u31 + uik[26] * u32 + uik[27] * u33 + uik[28] * u34 + uik[29] * u35;
3111b3064deSBarry Smith wp[35] += uik[30] * u30 + uik[31] * u31 + uik[32] * u32 + uik[33] * u33 + uik[34] * u34 + uik[35] * u35;
31281278733SSatish Balay }
3139566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(2.0 * 216.0 * (jmax - jmin)));
31481278733SSatish Balay
31581278733SSatish Balay /* ... add i to row list for next nonzero entry */
31681278733SSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */
31781278733SSatish Balay j = bj[jmin];
3189371c9d4SSatish Balay jl[i] = jl[j];
3199371c9d4SSatish Balay jl[j] = i; /* update jl */
32081278733SSatish Balay }
32181278733SSatish Balay i = nexti;
32281278733SSatish Balay }
32381278733SSatish Balay
32481278733SSatish Balay /* save nonzero entries in k-th row of U ... */
32581278733SSatish Balay
32681278733SSatish Balay /* invert diagonal block */
32781278733SSatish Balay d = ba + k * 36;
3289566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(d, dk, 36));
3299566063dSJacob Faibussowitsch PetscCall(PetscKernel_A_gets_inverse_A_6(d, shift, allowzeropivot, &zeropivotdetected));
3307b6c816cSBarry Smith if (zeropivotdetected) C->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
33181278733SSatish Balay
3329371c9d4SSatish Balay jmin = bi[k];
3339371c9d4SSatish Balay jmax = bi[k + 1];
33481278733SSatish Balay if (jmin < jmax) {
33581278733SSatish Balay for (j = jmin; j < jmax; j++) {
33681278733SSatish Balay vj = bj[j]; /* block col. index of U */
33781278733SSatish Balay u = ba + j * 36;
33881278733SSatish Balay wp = w + vj * 36;
33981278733SSatish Balay for (k1 = 0; k1 < 36; k1++) {
34081278733SSatish Balay *u++ = *wp;
34181278733SSatish Balay *wp++ = 0.0;
34281278733SSatish Balay }
34381278733SSatish Balay }
34481278733SSatish Balay
34581278733SSatish Balay /* ... add k to row list for first nonzero entry in k-th row */
34681278733SSatish Balay il[k] = jmin;
34781278733SSatish Balay i = bj[jmin];
3489371c9d4SSatish Balay jl[k] = jl[i];
3499371c9d4SSatish Balay jl[i] = k;
35081278733SSatish Balay }
35181278733SSatish Balay }
35281278733SSatish Balay
3539566063dSJacob Faibussowitsch PetscCall(PetscFree(w));
3549566063dSJacob Faibussowitsch PetscCall(PetscFree2(il, jl));
3559566063dSJacob Faibussowitsch PetscCall(PetscFree2(dk, uik));
3561baa6e33SBarry Smith if (a->permute) PetscCall(PetscFree(aa));
35781278733SSatish Balay
3589566063dSJacob Faibussowitsch PetscCall(ISRestoreIndices(perm, &perm_ptr));
35926fbe8dcSKarl Rupp
3604f79d315SHong Zhang C->ops->solve = MatSolve_SeqSBAIJ_6_inplace;
3614f79d315SHong Zhang C->ops->solvetranspose = MatSolve_SeqSBAIJ_6_inplace;
36281278733SSatish Balay C->assembled = PETSC_TRUE;
36381278733SSatish Balay C->preallocated = PETSC_TRUE;
36426fbe8dcSKarl Rupp
3659566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(1.3333 * 216 * b->mbs)); /* from inverting diagonal blocks */
3663ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS);
36781278733SSatish Balay }
368