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 7 by 7 */ 5d71ae5a4SJacob Faibussowitsch PetscErrorCode MatCholeskyFactorNumeric_SeqSBAIJ_7(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, *bi = b->i, *bj = b->j; 105d0c19d7SBarry Smith PetscInt *a2anew, k, k1, jmin, jmax, *jl, *il, vj, nexti, ili, i, j, mbs = a->mbs; 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, u36, u37, u38, u39, u40, u41; 151b3064deSBarry Smith MatScalar u42, u43, u44, u45, u46, u47, u48; 16182b8fbaSHong Zhang PetscReal shift = info->shiftamount; 17a455e926SHong Zhang PetscBool allowzeropivot, zeropivotdetected; 1881278733SSatish Balay 1981278733SSatish Balay PetscFunctionBegin; 2081278733SSatish Balay /* initialization */ 210164db54SHong Zhang allowzeropivot = PetscNot(A->erroriffailure); 229566063dSJacob Faibussowitsch PetscCall(PetscCalloc1(49 * mbs, &w)); 239566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(mbs, &il, mbs, &jl)); 246df5ee2eSHong Zhang il[0] = 0; 256df5ee2eSHong Zhang for (i = 0; i < mbs; i++) jl[i] = mbs; 266df5ee2eSHong Zhang 279566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(49, &dk, 49, &uik)); 289566063dSJacob Faibussowitsch PetscCall(ISGetIndices(perm, &perm_ptr)); 2981278733SSatish Balay 3081278733SSatish Balay /* check permutation */ 3181278733SSatish Balay if (!a->permute) { 329371c9d4SSatish Balay ai = a->i; 339371c9d4SSatish Balay aj = a->j; 349371c9d4SSatish Balay aa = a->a; 3581278733SSatish Balay } else { 369371c9d4SSatish Balay ai = a->inew; 379371c9d4SSatish Balay aj = a->jnew; 389566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(49 * ai[mbs], &aa)); 399566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(aa, a->a, 49 * ai[mbs])); 409566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(ai[mbs], &a2anew)); 419566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(a2anew, a->a2anew, ai[mbs])); 4281278733SSatish Balay 4381278733SSatish Balay for (i = 0; i < mbs; i++) { 449371c9d4SSatish Balay jmin = ai[i]; 459371c9d4SSatish Balay jmax = ai[i + 1]; 4681278733SSatish Balay for (j = jmin; j < jmax; j++) { 4781278733SSatish Balay while (a2anew[j] != j) { 489371c9d4SSatish Balay k = a2anew[j]; 499371c9d4SSatish Balay a2anew[j] = a2anew[k]; 509371c9d4SSatish Balay a2anew[k] = k; 5181278733SSatish Balay for (k1 = 0; k1 < 49; k1++) { 5281278733SSatish Balay dk[k1] = aa[k * 49 + k1]; 5381278733SSatish Balay aa[k * 49 + k1] = aa[j * 49 + k1]; 5481278733SSatish Balay aa[j * 49 + k1] = dk[k1]; 5581278733SSatish Balay } 5681278733SSatish Balay } 57*5e116b59SBarry Smith /* transform column-oriented blocks that lie in the lower triangle to row-oriented blocks */ 5881278733SSatish Balay if (i > aj[j]) { 5981278733SSatish Balay /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 6081278733SSatish Balay ap = aa + j * 49; /* ptr to the beginning of j-th block of aa */ 6181278733SSatish Balay for (k = 0; k < 49; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 6281278733SSatish Balay for (k = 0; k < 7; k++) { /* j-th block of aa <- dk^T */ 6381278733SSatish Balay for (k1 = 0; k1 < 7; k1++) *ap++ = dk[k + 7 * k1]; 6481278733SSatish Balay } 6581278733SSatish Balay } 6681278733SSatish Balay } 6781278733SSatish Balay } 689566063dSJacob Faibussowitsch PetscCall(PetscFree(a2anew)); 6981278733SSatish Balay } 7081278733SSatish Balay 7181278733SSatish Balay /* for each row k */ 7281278733SSatish Balay for (k = 0; k < mbs; k++) { 7381278733SSatish Balay /*initialize k-th row with elements nonzero in row perm(k) of A */ 749371c9d4SSatish Balay jmin = ai[perm_ptr[k]]; 759371c9d4SSatish Balay jmax = ai[perm_ptr[k] + 1]; 7681278733SSatish Balay if (jmin < jmax) { 7781278733SSatish Balay ap = aa + jmin * 49; 7881278733SSatish Balay for (j = jmin; j < jmax; j++) { 7981278733SSatish Balay vj = perm_ptr[aj[j]]; /* block col. index */ 8081278733SSatish Balay wp = w + vj * 49; 8181278733SSatish Balay for (i = 0; i < 49; i++) *wp++ = *ap++; 8281278733SSatish Balay } 8381278733SSatish Balay } 8481278733SSatish Balay 8581278733SSatish Balay /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 869566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(dk, w + k * 49, 49)); 8781278733SSatish Balay i = jl[k]; /* first row to be added to k_th row */ 8881278733SSatish Balay 8981278733SSatish Balay while (i < mbs) { 9081278733SSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 9181278733SSatish Balay 9281278733SSatish Balay /* compute multiplier */ 9381278733SSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 9481278733SSatish Balay 9581278733SSatish Balay /* uik = -inv(Di)*U_bar(i,k) */ 9681278733SSatish Balay d = ba + i * 49; 9781278733SSatish Balay u = ba + ili * 49; 9881278733SSatish Balay 999371c9d4SSatish Balay u0 = u[0]; 1009371c9d4SSatish Balay u1 = u[1]; 1019371c9d4SSatish Balay u2 = u[2]; 1029371c9d4SSatish Balay u3 = u[3]; 1039371c9d4SSatish Balay u4 = u[4]; 1049371c9d4SSatish Balay u5 = u[5]; 1059371c9d4SSatish Balay u6 = u[6]; 1069371c9d4SSatish Balay u7 = u[7]; 1079371c9d4SSatish Balay u8 = u[8]; 1089371c9d4SSatish Balay u9 = u[9]; 1099371c9d4SSatish Balay u10 = u[10]; 1109371c9d4SSatish Balay u11 = u[11]; 1119371c9d4SSatish Balay u12 = u[12]; 1129371c9d4SSatish Balay u13 = u[13]; 1139371c9d4SSatish Balay u14 = u[14]; 1149371c9d4SSatish Balay u15 = u[15]; 1159371c9d4SSatish Balay u16 = u[16]; 1169371c9d4SSatish Balay u17 = u[17]; 1179371c9d4SSatish Balay u18 = u[18]; 1189371c9d4SSatish Balay u19 = u[19]; 1199371c9d4SSatish Balay u20 = u[20]; 1209371c9d4SSatish Balay u21 = u[21]; 1219371c9d4SSatish Balay u22 = u[22]; 1229371c9d4SSatish Balay u23 = u[23]; 1239371c9d4SSatish Balay u24 = u[24]; 1249371c9d4SSatish Balay u25 = u[25]; 1259371c9d4SSatish Balay u26 = u[26]; 1269371c9d4SSatish Balay u27 = u[27]; 1279371c9d4SSatish Balay u28 = u[28]; 1289371c9d4SSatish Balay u29 = u[29]; 1299371c9d4SSatish Balay u30 = u[30]; 1309371c9d4SSatish Balay u31 = u[31]; 1319371c9d4SSatish Balay u32 = u[32]; 1329371c9d4SSatish Balay u33 = u[33]; 1339371c9d4SSatish Balay u34 = u[34]; 1349371c9d4SSatish Balay u35 = u[35]; 1359371c9d4SSatish Balay u36 = u[36]; 1369371c9d4SSatish Balay u37 = u[37]; 1379371c9d4SSatish Balay u38 = u[38]; 1389371c9d4SSatish Balay u39 = u[39]; 1399371c9d4SSatish Balay u40 = u[40]; 1409371c9d4SSatish Balay u41 = u[41]; 1419371c9d4SSatish Balay u42 = u[42]; 1429371c9d4SSatish Balay u43 = u[43]; 1439371c9d4SSatish Balay u44 = u[44]; 1449371c9d4SSatish Balay u45 = u[45]; 1459371c9d4SSatish Balay u46 = u[46]; 1469371c9d4SSatish Balay u47 = u[47]; 1479371c9d4SSatish Balay u48 = u[48]; 14881278733SSatish Balay 1491b3064deSBarry Smith uik[0] = -(d[0] * u0 + d[7] * u1 + d[14] * u2 + d[21] * u3 + d[28] * u4 + d[35] * u5 + d[42] * u6); 1501b3064deSBarry Smith uik[1] = -(d[1] * u0 + d[8] * u1 + d[15] * u2 + d[22] * u3 + d[29] * u4 + d[36] * u5 + d[43] * u6); 1511b3064deSBarry Smith uik[2] = -(d[2] * u0 + d[9] * u1 + d[16] * u2 + d[23] * u3 + d[30] * u4 + d[37] * u5 + d[44] * u6); 1521b3064deSBarry Smith uik[3] = -(d[3] * u0 + d[10] * u1 + d[17] * u2 + d[24] * u3 + d[31] * u4 + d[38] * u5 + d[45] * u6); 1531b3064deSBarry Smith uik[4] = -(d[4] * u0 + d[11] * u1 + d[18] * u2 + d[25] * u3 + d[32] * u4 + d[39] * u5 + d[46] * u6); 1541b3064deSBarry Smith uik[5] = -(d[5] * u0 + d[12] * u1 + d[19] * u2 + d[26] * u3 + d[33] * u4 + d[40] * u5 + d[47] * u6); 1551b3064deSBarry Smith uik[6] = -(d[6] * u0 + d[13] * u1 + d[20] * u2 + d[27] * u3 + d[34] * u4 + d[41] * u5 + d[48] * u6); 15681278733SSatish Balay 1571b3064deSBarry Smith uik[7] = -(d[0] * u7 + d[7] * u8 + d[14] * u9 + d[21] * u10 + d[28] * u11 + d[35] * u12 + d[42] * u13); 1581b3064deSBarry Smith uik[8] = -(d[1] * u7 + d[8] * u8 + d[15] * u9 + d[22] * u10 + d[29] * u11 + d[36] * u12 + d[43] * u13); 1591b3064deSBarry Smith uik[9] = -(d[2] * u7 + d[9] * u8 + d[16] * u9 + d[23] * u10 + d[30] * u11 + d[37] * u12 + d[44] * u13); 1601b3064deSBarry Smith uik[10] = -(d[3] * u7 + d[10] * u8 + d[17] * u9 + d[24] * u10 + d[31] * u11 + d[38] * u12 + d[45] * u13); 1611b3064deSBarry Smith uik[11] = -(d[4] * u7 + d[11] * u8 + d[18] * u9 + d[25] * u10 + d[32] * u11 + d[39] * u12 + d[46] * u13); 1621b3064deSBarry Smith uik[12] = -(d[5] * u7 + d[12] * u8 + d[19] * u9 + d[26] * u10 + d[33] * u11 + d[40] * u12 + d[47] * u13); 1631b3064deSBarry Smith uik[13] = -(d[6] * u7 + d[13] * u8 + d[20] * u9 + d[27] * u10 + d[34] * u11 + d[41] * u12 + d[48] * u13); 16481278733SSatish Balay 1651b3064deSBarry Smith uik[14] = -(d[0] * u14 + d[7] * u15 + d[14] * u16 + d[21] * u17 + d[28] * u18 + d[35] * u19 + d[42] * u20); 1661b3064deSBarry Smith uik[15] = -(d[1] * u14 + d[8] * u15 + d[15] * u16 + d[22] * u17 + d[29] * u18 + d[36] * u19 + d[43] * u20); 1671b3064deSBarry Smith uik[16] = -(d[2] * u14 + d[9] * u15 + d[16] * u16 + d[23] * u17 + d[30] * u18 + d[37] * u19 + d[44] * u20); 1681b3064deSBarry Smith uik[17] = -(d[3] * u14 + d[10] * u15 + d[17] * u16 + d[24] * u17 + d[31] * u18 + d[38] * u19 + d[45] * u20); 1691b3064deSBarry Smith uik[18] = -(d[4] * u14 + d[11] * u15 + d[18] * u16 + d[25] * u17 + d[32] * u18 + d[39] * u19 + d[46] * u20); 1701b3064deSBarry Smith uik[19] = -(d[5] * u14 + d[12] * u15 + d[19] * u16 + d[26] * u17 + d[33] * u18 + d[40] * u19 + d[47] * u20); 1711b3064deSBarry Smith uik[20] = -(d[6] * u14 + d[13] * u15 + d[20] * u16 + d[27] * u17 + d[34] * u18 + d[41] * u19 + d[48] * u20); 17281278733SSatish Balay 1731b3064deSBarry Smith uik[21] = -(d[0] * u21 + d[7] * u22 + d[14] * u23 + d[21] * u24 + d[28] * u25 + d[35] * u26 + d[42] * u27); 1741b3064deSBarry Smith uik[22] = -(d[1] * u21 + d[8] * u22 + d[15] * u23 + d[22] * u24 + d[29] * u25 + d[36] * u26 + d[43] * u27); 1751b3064deSBarry Smith uik[23] = -(d[2] * u21 + d[9] * u22 + d[16] * u23 + d[23] * u24 + d[30] * u25 + d[37] * u26 + d[44] * u27); 1761b3064deSBarry Smith uik[24] = -(d[3] * u21 + d[10] * u22 + d[17] * u23 + d[24] * u24 + d[31] * u25 + d[38] * u26 + d[45] * u27); 1771b3064deSBarry Smith uik[25] = -(d[4] * u21 + d[11] * u22 + d[18] * u23 + d[25] * u24 + d[32] * u25 + d[39] * u26 + d[46] * u27); 1781b3064deSBarry Smith uik[26] = -(d[5] * u21 + d[12] * u22 + d[19] * u23 + d[26] * u24 + d[33] * u25 + d[40] * u26 + d[47] * u27); 1791b3064deSBarry Smith uik[27] = -(d[6] * u21 + d[13] * u22 + d[20] * u23 + d[27] * u24 + d[34] * u25 + d[41] * u26 + d[48] * u27); 18081278733SSatish Balay 1811b3064deSBarry Smith uik[28] = -(d[0] * u28 + d[7] * u29 + d[14] * u30 + d[21] * u31 + d[28] * u32 + d[35] * u33 + d[42] * u34); 1821b3064deSBarry Smith uik[29] = -(d[1] * u28 + d[8] * u29 + d[15] * u30 + d[22] * u31 + d[29] * u32 + d[36] * u33 + d[43] * u34); 1831b3064deSBarry Smith uik[30] = -(d[2] * u28 + d[9] * u29 + d[16] * u30 + d[23] * u31 + d[30] * u32 + d[37] * u33 + d[44] * u34); 1841b3064deSBarry Smith uik[31] = -(d[3] * u28 + d[10] * u29 + d[17] * u30 + d[24] * u31 + d[31] * u32 + d[38] * u33 + d[45] * u34); 1851b3064deSBarry Smith uik[32] = -(d[4] * u28 + d[11] * u29 + d[18] * u30 + d[25] * u31 + d[32] * u32 + d[39] * u33 + d[46] * u34); 1861b3064deSBarry Smith uik[33] = -(d[5] * u28 + d[12] * u29 + d[19] * u30 + d[26] * u31 + d[33] * u32 + d[40] * u33 + d[47] * u34); 1871b3064deSBarry Smith uik[34] = -(d[6] * u28 + d[13] * u29 + d[20] * u30 + d[27] * u31 + d[34] * u32 + d[41] * u33 + d[48] * u34); 18881278733SSatish Balay 1891b3064deSBarry Smith uik[35] = -(d[0] * u35 + d[7] * u36 + d[14] * u37 + d[21] * u38 + d[28] * u39 + d[35] * u40 + d[42] * u41); 1901b3064deSBarry Smith uik[36] = -(d[1] * u35 + d[8] * u36 + d[15] * u37 + d[22] * u38 + d[29] * u39 + d[36] * u40 + d[43] * u41); 1911b3064deSBarry Smith uik[37] = -(d[2] * u35 + d[9] * u36 + d[16] * u37 + d[23] * u38 + d[30] * u39 + d[37] * u40 + d[44] * u41); 1921b3064deSBarry Smith uik[38] = -(d[3] * u35 + d[10] * u36 + d[17] * u37 + d[24] * u38 + d[31] * u39 + d[38] * u40 + d[45] * u41); 1931b3064deSBarry Smith uik[39] = -(d[4] * u35 + d[11] * u36 + d[18] * u37 + d[25] * u38 + d[32] * u39 + d[39] * u40 + d[46] * u41); 1941b3064deSBarry Smith uik[40] = -(d[5] * u35 + d[12] * u36 + d[19] * u37 + d[26] * u38 + d[33] * u39 + d[40] * u40 + d[47] * u41); 1951b3064deSBarry Smith uik[41] = -(d[6] * u35 + d[13] * u36 + d[20] * u37 + d[27] * u38 + d[34] * u39 + d[41] * u40 + d[48] * u41); 1961b3064deSBarry Smith 1971b3064deSBarry Smith uik[42] = -(d[0] * u42 + d[7] * u43 + d[14] * u44 + d[21] * u45 + d[28] * u46 + d[35] * u47 + d[42] * u48); 1981b3064deSBarry Smith uik[43] = -(d[1] * u42 + d[8] * u43 + d[15] * u44 + d[22] * u45 + d[29] * u46 + d[36] * u47 + d[43] * u48); 1991b3064deSBarry Smith uik[44] = -(d[2] * u42 + d[9] * u43 + d[16] * u44 + d[23] * u45 + d[30] * u46 + d[37] * u47 + d[44] * u48); 2001b3064deSBarry Smith uik[45] = -(d[3] * u42 + d[10] * u43 + d[17] * u44 + d[24] * u45 + d[31] * u46 + d[38] * u47 + d[45] * u48); 2011b3064deSBarry Smith uik[46] = -(d[4] * u42 + d[11] * u43 + d[18] * u44 + d[25] * u45 + d[32] * u46 + d[39] * u47 + d[46] * u48); 2021b3064deSBarry Smith uik[47] = -(d[5] * u42 + d[12] * u43 + d[19] * u44 + d[26] * u45 + d[33] * u46 + d[40] * u47 + d[47] * u48); 2031b3064deSBarry Smith uik[48] = -(d[6] * u42 + d[13] * u43 + d[20] * u44 + d[27] * u45 + d[34] * u46 + d[41] * u47 + d[48] * u48); 20481278733SSatish Balay 20581278733SSatish Balay /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 2061b3064deSBarry Smith dk[0] += uik[0] * u0 + uik[1] * u1 + uik[2] * u2 + uik[3] * u3 + uik[4] * u4 + uik[5] * u5 + uik[6] * u6; 2071b3064deSBarry Smith dk[1] += uik[7] * u0 + uik[8] * u1 + uik[9] * u2 + uik[10] * u3 + uik[11] * u4 + uik[12] * u5 + uik[13] * u6; 2081b3064deSBarry Smith dk[2] += uik[14] * u0 + uik[15] * u1 + uik[16] * u2 + uik[17] * u3 + uik[18] * u4 + uik[19] * u5 + uik[20] * u6; 2091b3064deSBarry Smith dk[3] += uik[21] * u0 + uik[22] * u1 + uik[23] * u2 + uik[24] * u3 + uik[25] * u4 + uik[26] * u5 + uik[27] * u6; 2101b3064deSBarry Smith dk[4] += uik[28] * u0 + uik[29] * u1 + uik[30] * u2 + uik[31] * u3 + uik[32] * u4 + uik[33] * u5 + uik[34] * u6; 2111b3064deSBarry Smith dk[5] += uik[35] * u0 + uik[36] * u1 + uik[37] * u2 + uik[38] * u3 + uik[39] * u4 + uik[40] * u5 + uik[41] * u6; 2121b3064deSBarry Smith dk[6] += uik[42] * u0 + uik[43] * u1 + uik[44] * u2 + uik[45] * u3 + uik[46] * u4 + uik[47] * u5 + uik[48] * u6; 21381278733SSatish Balay 2141b3064deSBarry Smith dk[7] += uik[0] * u7 + uik[1] * u8 + uik[2] * u9 + uik[3] * u10 + uik[4] * u11 + uik[5] * u12 + uik[6] * u13; 2151b3064deSBarry Smith dk[8] += uik[7] * u7 + uik[8] * u8 + uik[9] * u9 + uik[10] * u10 + uik[11] * u11 + uik[12] * u12 + uik[13] * u13; 2161b3064deSBarry Smith dk[9] += uik[14] * u7 + uik[15] * u8 + uik[16] * u9 + uik[17] * u10 + uik[18] * u11 + uik[19] * u12 + uik[20] * u13; 2171b3064deSBarry Smith dk[10] += uik[21] * u7 + uik[22] * u8 + uik[23] * u9 + uik[24] * u10 + uik[25] * u11 + uik[26] * u12 + uik[27] * u13; 2181b3064deSBarry Smith dk[11] += uik[28] * u7 + uik[29] * u8 + uik[30] * u9 + uik[31] * u10 + uik[32] * u11 + uik[33] * u12 + uik[34] * u13; 2191b3064deSBarry Smith dk[12] += uik[35] * u7 + uik[36] * u8 + uik[37] * u9 + uik[38] * u10 + uik[39] * u11 + uik[40] * u12 + uik[41] * u13; 2201b3064deSBarry Smith dk[13] += uik[42] * u7 + uik[43] * u8 + uik[44] * u9 + uik[45] * u10 + uik[46] * u11 + uik[47] * u12 + uik[48] * u13; 22181278733SSatish Balay 2221b3064deSBarry Smith dk[14] += uik[0] * u14 + uik[1] * u15 + uik[2] * u16 + uik[3] * u17 + uik[4] * u18 + uik[5] * u19 + uik[6] * u20; 2231b3064deSBarry Smith dk[15] += uik[7] * u14 + uik[8] * u15 + uik[9] * u16 + uik[10] * u17 + uik[11] * u18 + uik[12] * u19 + uik[13] * u20; 2241b3064deSBarry Smith dk[16] += uik[14] * u14 + uik[15] * u15 + uik[16] * u16 + uik[17] * u17 + uik[18] * u18 + uik[19] * u19 + uik[20] * u20; 2251b3064deSBarry Smith dk[17] += uik[21] * u14 + uik[22] * u15 + uik[23] * u16 + uik[24] * u17 + uik[25] * u18 + uik[26] * u19 + uik[27] * u20; 2261b3064deSBarry Smith dk[18] += uik[28] * u14 + uik[29] * u15 + uik[30] * u16 + uik[31] * u17 + uik[32] * u18 + uik[33] * u19 + uik[34] * u20; 2271b3064deSBarry Smith dk[19] += uik[35] * u14 + uik[36] * u15 + uik[37] * u16 + uik[38] * u17 + uik[39] * u18 + uik[40] * u19 + uik[41] * u20; 2281b3064deSBarry Smith dk[20] += uik[42] * u14 + uik[43] * u15 + uik[44] * u16 + uik[45] * u17 + uik[46] * u18 + uik[47] * u19 + uik[48] * u20; 22981278733SSatish Balay 2301b3064deSBarry Smith dk[21] += uik[0] * u21 + uik[1] * u22 + uik[2] * u23 + uik[3] * u24 + uik[4] * u25 + uik[5] * u26 + uik[6] * u27; 2311b3064deSBarry Smith dk[22] += uik[7] * u21 + uik[8] * u22 + uik[9] * u23 + uik[10] * u24 + uik[11] * u25 + uik[12] * u26 + uik[13] * u27; 2321b3064deSBarry Smith dk[23] += uik[14] * u21 + uik[15] * u22 + uik[16] * u23 + uik[17] * u24 + uik[18] * u25 + uik[19] * u26 + uik[20] * u27; 2331b3064deSBarry Smith dk[24] += uik[21] * u21 + uik[22] * u22 + uik[23] * u23 + uik[24] * u24 + uik[25] * u25 + uik[26] * u26 + uik[27] * u27; 2341b3064deSBarry Smith dk[25] += uik[28] * u21 + uik[29] * u22 + uik[30] * u23 + uik[31] * u24 + uik[32] * u25 + uik[33] * u26 + uik[34] * u27; 2351b3064deSBarry Smith dk[26] += uik[35] * u21 + uik[36] * u22 + uik[37] * u23 + uik[38] * u24 + uik[39] * u25 + uik[40] * u26 + uik[41] * u27; 2361b3064deSBarry Smith dk[27] += uik[42] * u21 + uik[43] * u22 + uik[44] * u23 + uik[45] * u24 + uik[46] * u25 + uik[47] * u26 + uik[48] * u27; 23781278733SSatish Balay 2381b3064deSBarry Smith dk[28] += uik[0] * u28 + uik[1] * u29 + uik[2] * u30 + uik[3] * u31 + uik[4] * u32 + uik[5] * u33 + uik[6] * u34; 2391b3064deSBarry Smith dk[29] += uik[7] * u28 + uik[8] * u29 + uik[9] * u30 + uik[10] * u31 + uik[11] * u32 + uik[12] * u33 + uik[13] * u34; 2401b3064deSBarry Smith dk[30] += uik[14] * u28 + uik[15] * u29 + uik[16] * u30 + uik[17] * u31 + uik[18] * u32 + uik[19] * u33 + uik[20] * u34; 2411b3064deSBarry Smith dk[31] += uik[21] * u28 + uik[22] * u29 + uik[23] * u30 + uik[24] * u31 + uik[25] * u32 + uik[26] * u33 + uik[27] * u34; 2421b3064deSBarry Smith dk[32] += uik[28] * u28 + uik[29] * u29 + uik[30] * u30 + uik[31] * u31 + uik[32] * u32 + uik[33] * u33 + uik[34] * u34; 2431b3064deSBarry Smith dk[33] += uik[35] * u28 + uik[36] * u29 + uik[37] * u30 + uik[38] * u31 + uik[39] * u32 + uik[40] * u33 + uik[41] * u34; 2441b3064deSBarry Smith dk[34] += uik[42] * u28 + uik[43] * u29 + uik[44] * u30 + uik[45] * u31 + uik[46] * u32 + uik[47] * u33 + uik[48] * u34; 24581278733SSatish Balay 2461b3064deSBarry Smith dk[35] += uik[0] * u35 + uik[1] * u36 + uik[2] * u37 + uik[3] * u38 + uik[4] * u39 + uik[5] * u40 + uik[6] * u41; 2471b3064deSBarry Smith dk[36] += uik[7] * u35 + uik[8] * u36 + uik[9] * u37 + uik[10] * u38 + uik[11] * u39 + uik[12] * u40 + uik[13] * u41; 2481b3064deSBarry Smith dk[37] += uik[14] * u35 + uik[15] * u36 + uik[16] * u37 + uik[17] * u38 + uik[18] * u39 + uik[19] * u40 + uik[20] * u41; 2491b3064deSBarry Smith dk[38] += uik[21] * u35 + uik[22] * u36 + uik[23] * u37 + uik[24] * u38 + uik[25] * u39 + uik[26] * u40 + uik[27] * u41; 2501b3064deSBarry Smith dk[39] += uik[28] * u35 + uik[29] * u36 + uik[30] * u37 + uik[31] * u38 + uik[32] * u39 + uik[33] * u40 + uik[34] * u41; 2511b3064deSBarry Smith dk[40] += uik[35] * u35 + uik[36] * u36 + uik[37] * u37 + uik[38] * u38 + uik[39] * u39 + uik[40] * u40 + uik[41] * u41; 2521b3064deSBarry Smith dk[41] += uik[42] * u35 + uik[43] * u36 + uik[44] * u37 + uik[45] * u38 + uik[46] * u39 + uik[47] * u40 + uik[48] * u41; 25381278733SSatish Balay 2541b3064deSBarry Smith dk[42] += uik[0] * u42 + uik[1] * u43 + uik[2] * u44 + uik[3] * u45 + uik[4] * u46 + uik[5] * u47 + uik[6] * u48; 2551b3064deSBarry Smith dk[43] += uik[7] * u42 + uik[8] * u43 + uik[9] * u44 + uik[10] * u45 + uik[11] * u46 + uik[12] * u47 + uik[13] * u48; 2561b3064deSBarry Smith dk[44] += uik[14] * u42 + uik[15] * u43 + uik[16] * u44 + uik[17] * u45 + uik[18] * u46 + uik[19] * u47 + uik[20] * u48; 2571b3064deSBarry Smith dk[45] += uik[21] * u42 + uik[22] * u43 + uik[23] * u44 + uik[24] * u45 + uik[25] * u46 + uik[26] * u47 + uik[27] * u48; 2581b3064deSBarry Smith dk[46] += uik[28] * u42 + uik[29] * u43 + uik[30] * u44 + uik[31] * u45 + uik[32] * u46 + uik[33] * u47 + uik[34] * u48; 2591b3064deSBarry Smith dk[47] += uik[35] * u42 + uik[36] * u43 + uik[37] * u44 + uik[38] * u45 + uik[39] * u46 + uik[40] * u47 + uik[41] * u48; 2601b3064deSBarry Smith dk[48] += uik[42] * u42 + uik[43] * u43 + uik[44] * u44 + uik[45] * u45 + uik[46] * u46 + uik[47] * u47 + uik[48] * u48; 26181278733SSatish Balay 2629566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(343.0 * 4.0)); 263187a9f4bSHong Zhang 26481278733SSatish Balay /* update -U(i,k) */ 2659566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(ba + ili * 49, uik, 49)); 26681278733SSatish Balay 26781278733SSatish Balay /* add multiple of row i to k-th row ... */ 2689371c9d4SSatish Balay jmin = ili + 1; 2699371c9d4SSatish Balay jmax = bi[i + 1]; 27081278733SSatish Balay if (jmin < jmax) { 27181278733SSatish Balay for (j = jmin; j < jmax; j++) { 27281278733SSatish Balay /* w += -U(i,k)^T * U_bar(i,j) */ 27381278733SSatish Balay wp = w + bj[j] * 49; 27481278733SSatish Balay u = ba + j * 49; 27581278733SSatish Balay 2769371c9d4SSatish Balay u0 = u[0]; 2779371c9d4SSatish Balay u1 = u[1]; 2789371c9d4SSatish Balay u2 = u[2]; 2799371c9d4SSatish Balay u3 = u[3]; 2809371c9d4SSatish Balay u4 = u[4]; 2819371c9d4SSatish Balay u5 = u[5]; 2829371c9d4SSatish Balay u6 = u[6]; 2839371c9d4SSatish Balay u7 = u[7]; 2849371c9d4SSatish Balay u8 = u[8]; 2859371c9d4SSatish Balay u9 = u[9]; 2869371c9d4SSatish Balay u10 = u[10]; 2879371c9d4SSatish Balay u11 = u[11]; 2889371c9d4SSatish Balay u12 = u[12]; 2899371c9d4SSatish Balay u13 = u[13]; 2909371c9d4SSatish Balay u14 = u[14]; 2919371c9d4SSatish Balay u15 = u[15]; 2929371c9d4SSatish Balay u16 = u[16]; 2939371c9d4SSatish Balay u17 = u[17]; 2949371c9d4SSatish Balay u18 = u[18]; 2959371c9d4SSatish Balay u19 = u[19]; 2969371c9d4SSatish Balay u20 = u[20]; 2979371c9d4SSatish Balay u21 = u[21]; 2989371c9d4SSatish Balay u22 = u[22]; 2999371c9d4SSatish Balay u23 = u[23]; 3009371c9d4SSatish Balay u24 = u[24]; 3019371c9d4SSatish Balay u25 = u[25]; 3029371c9d4SSatish Balay u26 = u[26]; 3039371c9d4SSatish Balay u27 = u[27]; 3049371c9d4SSatish Balay u28 = u[28]; 3059371c9d4SSatish Balay u29 = u[29]; 3069371c9d4SSatish Balay u30 = u[30]; 3079371c9d4SSatish Balay u31 = u[31]; 3089371c9d4SSatish Balay u32 = u[32]; 3099371c9d4SSatish Balay u33 = u[33]; 3109371c9d4SSatish Balay u34 = u[34]; 3119371c9d4SSatish Balay u35 = u[35]; 3129371c9d4SSatish Balay u36 = u[36]; 3139371c9d4SSatish Balay u37 = u[37]; 3149371c9d4SSatish Balay u38 = u[38]; 3159371c9d4SSatish Balay u39 = u[39]; 3169371c9d4SSatish Balay u40 = u[40]; 3179371c9d4SSatish Balay u41 = u[41]; 3189371c9d4SSatish Balay u42 = u[42]; 3199371c9d4SSatish Balay u43 = u[43]; 3209371c9d4SSatish Balay u44 = u[44]; 3219371c9d4SSatish Balay u45 = u[45]; 3229371c9d4SSatish Balay u46 = u[46]; 3239371c9d4SSatish Balay u47 = u[47]; 3249371c9d4SSatish Balay u48 = u[48]; 32581278733SSatish Balay 3261b3064deSBarry Smith wp[0] += uik[0] * u0 + uik[1] * u1 + uik[2] * u2 + uik[3] * u3 + uik[4] * u4 + uik[5] * u5 + uik[6] * u6; 3271b3064deSBarry Smith wp[1] += uik[7] * u0 + uik[8] * u1 + uik[9] * u2 + uik[10] * u3 + uik[11] * u4 + uik[12] * u5 + uik[13] * u6; 3281b3064deSBarry Smith wp[2] += uik[14] * u0 + uik[15] * u1 + uik[16] * u2 + uik[17] * u3 + uik[18] * u4 + uik[19] * u5 + uik[20] * u6; 3291b3064deSBarry Smith wp[3] += uik[21] * u0 + uik[22] * u1 + uik[23] * u2 + uik[24] * u3 + uik[25] * u4 + uik[26] * u5 + uik[27] * u6; 3301b3064deSBarry Smith wp[4] += uik[28] * u0 + uik[29] * u1 + uik[30] * u2 + uik[31] * u3 + uik[32] * u4 + uik[33] * u5 + uik[34] * u6; 3311b3064deSBarry Smith wp[5] += uik[35] * u0 + uik[36] * u1 + uik[37] * u2 + uik[38] * u3 + uik[39] * u4 + uik[40] * u5 + uik[41] * u6; 3321b3064deSBarry Smith wp[6] += uik[42] * u0 + uik[43] * u1 + uik[44] * u2 + uik[45] * u3 + uik[46] * u4 + uik[47] * u5 + uik[48] * u6; 33381278733SSatish Balay 3341b3064deSBarry Smith wp[7] += uik[0] * u7 + uik[1] * u8 + uik[2] * u9 + uik[3] * u10 + uik[4] * u11 + uik[5] * u12 + uik[6] * u13; 3351b3064deSBarry Smith wp[8] += uik[7] * u7 + uik[8] * u8 + uik[9] * u9 + uik[10] * u10 + uik[11] * u11 + uik[12] * u12 + uik[13] * u13; 3361b3064deSBarry Smith wp[9] += uik[14] * u7 + uik[15] * u8 + uik[16] * u9 + uik[17] * u10 + uik[18] * u11 + uik[19] * u12 + uik[20] * u13; 3371b3064deSBarry Smith wp[10] += uik[21] * u7 + uik[22] * u8 + uik[23] * u9 + uik[24] * u10 + uik[25] * u11 + uik[26] * u12 + uik[27] * u13; 3381b3064deSBarry Smith wp[11] += uik[28] * u7 + uik[29] * u8 + uik[30] * u9 + uik[31] * u10 + uik[32] * u11 + uik[33] * u12 + uik[34] * u13; 3391b3064deSBarry Smith wp[12] += uik[35] * u7 + uik[36] * u8 + uik[37] * u9 + uik[38] * u10 + uik[39] * u11 + uik[40] * u12 + uik[41] * u13; 3401b3064deSBarry Smith wp[13] += uik[42] * u7 + uik[43] * u8 + uik[44] * u9 + uik[45] * u10 + uik[46] * u11 + uik[47] * u12 + uik[48] * u13; 34181278733SSatish Balay 3421b3064deSBarry Smith wp[14] += uik[0] * u14 + uik[1] * u15 + uik[2] * u16 + uik[3] * u17 + uik[4] * u18 + uik[5] * u19 + uik[6] * u20; 3431b3064deSBarry Smith wp[15] += uik[7] * u14 + uik[8] * u15 + uik[9] * u16 + uik[10] * u17 + uik[11] * u18 + uik[12] * u19 + uik[13] * u20; 3441b3064deSBarry Smith wp[16] += uik[14] * u14 + uik[15] * u15 + uik[16] * u16 + uik[17] * u17 + uik[18] * u18 + uik[19] * u19 + uik[20] * u20; 3451b3064deSBarry Smith wp[17] += uik[21] * u14 + uik[22] * u15 + uik[23] * u16 + uik[24] * u17 + uik[25] * u18 + uik[26] * u19 + uik[27] * u20; 3461b3064deSBarry Smith wp[18] += uik[28] * u14 + uik[29] * u15 + uik[30] * u16 + uik[31] * u17 + uik[32] * u18 + uik[33] * u19 + uik[34] * u20; 3471b3064deSBarry Smith wp[19] += uik[35] * u14 + uik[36] * u15 + uik[37] * u16 + uik[38] * u17 + uik[39] * u18 + uik[40] * u19 + uik[41] * u20; 3481b3064deSBarry Smith wp[20] += uik[42] * u14 + uik[43] * u15 + uik[44] * u16 + uik[45] * u17 + uik[46] * u18 + uik[47] * u19 + uik[48] * u20; 34981278733SSatish Balay 3501b3064deSBarry Smith wp[21] += uik[0] * u21 + uik[1] * u22 + uik[2] * u23 + uik[3] * u24 + uik[4] * u25 + uik[5] * u26 + uik[6] * u27; 3511b3064deSBarry Smith wp[22] += uik[7] * u21 + uik[8] * u22 + uik[9] * u23 + uik[10] * u24 + uik[11] * u25 + uik[12] * u26 + uik[13] * u27; 3521b3064deSBarry Smith wp[23] += uik[14] * u21 + uik[15] * u22 + uik[16] * u23 + uik[17] * u24 + uik[18] * u25 + uik[19] * u26 + uik[20] * u27; 3531b3064deSBarry Smith wp[24] += uik[21] * u21 + uik[22] * u22 + uik[23] * u23 + uik[24] * u24 + uik[25] * u25 + uik[26] * u26 + uik[27] * u27; 3541b3064deSBarry Smith wp[25] += uik[28] * u21 + uik[29] * u22 + uik[30] * u23 + uik[31] * u24 + uik[32] * u25 + uik[33] * u26 + uik[34] * u27; 3551b3064deSBarry Smith wp[26] += uik[35] * u21 + uik[36] * u22 + uik[37] * u23 + uik[38] * u24 + uik[39] * u25 + uik[40] * u26 + uik[41] * u27; 3561b3064deSBarry Smith wp[27] += uik[42] * u21 + uik[43] * u22 + uik[44] * u23 + uik[45] * u24 + uik[46] * u25 + uik[47] * u26 + uik[48] * u27; 35781278733SSatish Balay 3581b3064deSBarry Smith wp[28] += uik[0] * u28 + uik[1] * u29 + uik[2] * u30 + uik[3] * u31 + uik[4] * u32 + uik[5] * u33 + uik[6] * u34; 3591b3064deSBarry Smith wp[29] += uik[7] * u28 + uik[8] * u29 + uik[9] * u30 + uik[10] * u31 + uik[11] * u32 + uik[12] * u33 + uik[13] * u34; 3601b3064deSBarry Smith wp[30] += uik[14] * u28 + uik[15] * u29 + uik[16] * u30 + uik[17] * u31 + uik[18] * u32 + uik[19] * u33 + uik[20] * u34; 3611b3064deSBarry Smith wp[31] += uik[21] * u28 + uik[22] * u29 + uik[23] * u30 + uik[24] * u31 + uik[25] * u32 + uik[26] * u33 + uik[27] * u34; 3621b3064deSBarry Smith wp[32] += uik[28] * u28 + uik[29] * u29 + uik[30] * u30 + uik[31] * u31 + uik[32] * u32 + uik[33] * u33 + uik[34] * u34; 3631b3064deSBarry Smith wp[33] += uik[35] * u28 + uik[36] * u29 + uik[37] * u30 + uik[38] * u31 + uik[39] * u32 + uik[40] * u33 + uik[41] * u34; 3641b3064deSBarry Smith wp[34] += uik[42] * u28 + uik[43] * u29 + uik[44] * u30 + uik[45] * u31 + uik[46] * u32 + uik[47] * u33 + uik[48] * u34; 36581278733SSatish Balay 3661b3064deSBarry Smith wp[35] += uik[0] * u35 + uik[1] * u36 + uik[2] * u37 + uik[3] * u38 + uik[4] * u39 + uik[5] * u40 + uik[6] * u41; 3671b3064deSBarry Smith wp[36] += uik[7] * u35 + uik[8] * u36 + uik[9] * u37 + uik[10] * u38 + uik[11] * u39 + uik[12] * u40 + uik[13] * u41; 3681b3064deSBarry Smith wp[37] += uik[14] * u35 + uik[15] * u36 + uik[16] * u37 + uik[17] * u38 + uik[18] * u39 + uik[19] * u40 + uik[20] * u41; 3691b3064deSBarry Smith wp[38] += uik[21] * u35 + uik[22] * u36 + uik[23] * u37 + uik[24] * u38 + uik[25] * u39 + uik[26] * u40 + uik[27] * u41; 3701b3064deSBarry Smith wp[39] += uik[28] * u35 + uik[29] * u36 + uik[30] * u37 + uik[31] * u38 + uik[32] * u39 + uik[33] * u40 + uik[34] * u41; 3711b3064deSBarry Smith wp[40] += uik[35] * u35 + uik[36] * u36 + uik[37] * u37 + uik[38] * u38 + uik[39] * u39 + uik[40] * u40 + uik[41] * u41; 3721b3064deSBarry Smith wp[41] += uik[42] * u35 + uik[43] * u36 + uik[44] * u37 + uik[45] * u38 + uik[46] * u39 + uik[47] * u40 + uik[48] * u41; 3731b3064deSBarry Smith 3741b3064deSBarry Smith wp[42] += uik[0] * u42 + uik[1] * u43 + uik[2] * u44 + uik[3] * u45 + uik[4] * u46 + uik[5] * u47 + uik[6] * u48; 3751b3064deSBarry Smith wp[43] += uik[7] * u42 + uik[8] * u43 + uik[9] * u44 + uik[10] * u45 + uik[11] * u46 + uik[12] * u47 + uik[13] * u48; 3761b3064deSBarry Smith wp[44] += uik[14] * u42 + uik[15] * u43 + uik[16] * u44 + uik[17] * u45 + uik[18] * u46 + uik[19] * u47 + uik[20] * u48; 3771b3064deSBarry Smith wp[45] += uik[21] * u42 + uik[22] * u43 + uik[23] * u44 + uik[24] * u45 + uik[25] * u46 + uik[26] * u47 + uik[27] * u48; 3781b3064deSBarry Smith wp[46] += uik[28] * u42 + uik[29] * u43 + uik[30] * u44 + uik[31] * u45 + uik[32] * u46 + uik[33] * u47 + uik[34] * u48; 3791b3064deSBarry Smith wp[47] += uik[35] * u42 + uik[36] * u43 + uik[37] * u44 + uik[38] * u45 + uik[39] * u46 + uik[40] * u47 + uik[41] * u48; 3801b3064deSBarry Smith wp[48] += uik[42] * u42 + uik[43] * u43 + uik[44] * u44 + uik[45] * u45 + uik[46] * u46 + uik[47] * u47 + uik[48] * u48; 38181278733SSatish Balay } 3829566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(2.0 * 343.0 * (jmax - jmin))); 38381278733SSatish Balay 38481278733SSatish Balay /* ... add i to row list for next nonzero entry */ 38581278733SSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 38681278733SSatish Balay j = bj[jmin]; 3879371c9d4SSatish Balay jl[i] = jl[j]; 3889371c9d4SSatish Balay jl[j] = i; /* update jl */ 38981278733SSatish Balay } 39081278733SSatish Balay i = nexti; 39181278733SSatish Balay } 39281278733SSatish Balay 39381278733SSatish Balay /* save nonzero entries in k-th row of U ... */ 39481278733SSatish Balay 39581278733SSatish Balay /* invert diagonal block */ 39681278733SSatish Balay d = ba + k * 49; 3979566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(d, dk, 49)); 3989566063dSJacob Faibussowitsch PetscCall(PetscKernel_A_gets_inverse_A_7(d, shift, allowzeropivot, &zeropivotdetected)); 3997b6c816cSBarry Smith if (zeropivotdetected) C->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 40081278733SSatish Balay 4019371c9d4SSatish Balay jmin = bi[k]; 4029371c9d4SSatish Balay jmax = bi[k + 1]; 40381278733SSatish Balay if (jmin < jmax) { 40481278733SSatish Balay for (j = jmin; j < jmax; j++) { 40581278733SSatish Balay vj = bj[j]; /* block col. index of U */ 40681278733SSatish Balay u = ba + j * 49; 40781278733SSatish Balay wp = w + vj * 49; 40881278733SSatish Balay for (k1 = 0; k1 < 49; k1++) { 40981278733SSatish Balay *u++ = *wp; 41081278733SSatish Balay *wp++ = 0.0; 41181278733SSatish Balay } 41281278733SSatish Balay } 41381278733SSatish Balay 41481278733SSatish Balay /* ... add k to row list for first nonzero entry in k-th row */ 41581278733SSatish Balay il[k] = jmin; 41681278733SSatish Balay i = bj[jmin]; 4179371c9d4SSatish Balay jl[k] = jl[i]; 4189371c9d4SSatish Balay jl[i] = k; 41981278733SSatish Balay } 42081278733SSatish Balay } 42181278733SSatish Balay 4229566063dSJacob Faibussowitsch PetscCall(PetscFree(w)); 4239566063dSJacob Faibussowitsch PetscCall(PetscFree2(il, jl)); 4249566063dSJacob Faibussowitsch PetscCall(PetscFree2(dk, uik)); 4251baa6e33SBarry Smith if (a->permute) PetscCall(PetscFree(aa)); 42681278733SSatish Balay 4279566063dSJacob Faibussowitsch PetscCall(ISRestoreIndices(perm, &perm_ptr)); 42826fbe8dcSKarl Rupp 4294f79d315SHong Zhang C->ops->solve = MatSolve_SeqSBAIJ_7_inplace; 4304f79d315SHong Zhang C->ops->solvetranspose = MatSolve_SeqSBAIJ_7_inplace; 43181278733SSatish Balay C->assembled = PETSC_TRUE; 43281278733SSatish Balay C->preallocated = PETSC_TRUE; 43326fbe8dcSKarl Rupp 4349566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(1.3333 * 343 * b->mbs)); /* from inverting diagonal blocks */ 4353ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 43681278733SSatish Balay } 437