1*d866deddSBarry Smith /*$Id: sbaijfact.c,v 1.56 2001/01/31 19:06:52 balay Exp bsmith $*/ 281278733SSatish Balay #include "sbaij.h" 381278733SSatish Balay #include "src/inline/ilu.h" 481278733SSatish Balay 581278733SSatish Balay /* 681278733SSatish Balay Version for when blocks are 6 by 6 Using natural ordering 781278733SSatish Balay */ 881278733SSatish Balay #undef __FUNC__ 981278733SSatish Balay #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering" 1081278733SSatish Balay int MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering(Mat A,Mat *B) 1181278733SSatish Balay { 1281278733SSatish Balay Mat C = *B; 1381278733SSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1481278733SSatish Balay int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 1581278733SSatish Balay int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1681278733SSatish Balay MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 17*d866deddSBarry Smith MatScalar *u,*d,*w,*wp,u0,u1,u2,u3,u4,u5,u6,u7,u8,u9,u10,u11,u12; 18*d866deddSBarry Smith MatScalar u13,u14,u15,u16,u17,u18,u19,u20,u21,u22,u23,u24,u25,u26,u27; 19*d866deddSBarry Smith MatScalar u28,u29,u30,u31,u32,u33,u34,u35; 2081278733SSatish Balay 2181278733SSatish Balay PetscFunctionBegin; 2281278733SSatish Balay 2381278733SSatish Balay /* initialization */ 2481278733SSatish Balay ierr = PetscMalloc(36*mbs*sizeof(MatScalar),&w);CHKERRQ(ierr); 2581278733SSatish Balay ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2681278733SSatish Balay ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr); 2781278733SSatish Balay jl = il + mbs; 2881278733SSatish Balay for (i=0; i<mbs; i++) { 2981278733SSatish Balay jl[i] = mbs; il[0] = 0; 3081278733SSatish Balay } 3181278733SSatish Balay ierr = PetscMalloc(72*sizeof(MatScalar),&dk);CHKERRQ(ierr); 3281278733SSatish Balay uik = dk + 36; 3381278733SSatish Balay 3481278733SSatish Balay ai = a->i; aj = a->j; aa = a->a; 3581278733SSatish Balay 3681278733SSatish Balay /* for each row k */ 3781278733SSatish Balay for (k = 0; k<mbs; k++){ 3881278733SSatish Balay 3981278733SSatish Balay /*initialize k-th row with elements nonzero in row k of A */ 4081278733SSatish Balay jmin = ai[k]; jmax = ai[k+1]; 4181278733SSatish Balay if (jmin < jmax) { 4281278733SSatish Balay ap = aa + jmin*36; 4381278733SSatish Balay for (j = jmin; j < jmax; j++){ 4481278733SSatish Balay vj = aj[j]; /* block col. index */ 4581278733SSatish Balay wp = w + vj*36; 4681278733SSatish Balay for (i=0; i<36; i++) *wp++ = *ap++; 4781278733SSatish Balay } 4881278733SSatish Balay } 4981278733SSatish Balay 5081278733SSatish Balay /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 5181278733SSatish Balay ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr); 5281278733SSatish Balay i = jl[k]; /* first row to be added to k_th row */ 5381278733SSatish Balay 5481278733SSatish Balay while (i < mbs){ 5581278733SSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 5681278733SSatish Balay 5781278733SSatish Balay /* compute multiplier */ 5881278733SSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 5981278733SSatish Balay 6081278733SSatish Balay /* uik = -inv(Di)*U_bar(i,k) */ 6181278733SSatish Balay d = ba + i*36; 6281278733SSatish Balay u = ba + ili*36; 6381278733SSatish Balay 64*d866deddSBarry Smith u0 = u[0]; u1 = u[1]; u2 = u[2]; u3 = u[3]; u4 = u[4]; u5 = u[5]; u6 = u[6]; 65*d866deddSBarry Smith u7 = u[7]; u8 = u[8]; u9 = u[9]; u10 = u[10]; u11 = u[11]; u12 = u[12]; u13 = u[13]; 66*d866deddSBarry Smith u14 = u[14]; u15 = u[15]; u16 = u[16]; u17 = u[17]; u18 = u[18]; u19 = u[19]; u20 = u[20]; 67*d866deddSBarry Smith u21 = u[21]; u22 = u[22]; u23 = u[23]; u24 = u[24]; u25 = u[25]; u26 = u[26]; u27 = u[27]; 68*d866deddSBarry Smith u28 = u[28]; u29 = u[29]; u30 = u[30]; u31 = u[31]; u32 = u[32]; u33 = u[33]; u34 = u[34]; 69*d866deddSBarry Smith u35 = u[35]; 70*d866deddSBarry Smith 7181278733SSatish Balay uik[0] = -(d[0]*u[0] + d[6]*u[1] + d[12]*u[2] + d[18]*u[3] + d[24]*u[4] + d[30]*u[5]); 7281278733SSatish Balay uik[1] = -(d[1]*u[0] + d[7]*u[1] + d[13]*u[2] + d[19]*u[3] + d[25]*u[4] + d[31]*u[5]); 7381278733SSatish Balay uik[2] = -(d[2]*u[0] + d[8]*u[1] + d[14]*u[2] + d[20]*u[3] + d[26]*u[4] + d[32]*u[5]); 7481278733SSatish Balay uik[3] = -(d[3]*u[0] + d[9]*u[1] + d[15]*u[2] + d[21]*u[3] + d[27]*u[4] + d[33]*u[5]); 7581278733SSatish Balay uik[4] = -(d[4]*u[0]+ d[10]*u[1] + d[16]*u[2] + d[22]*u[3] + d[28]*u[4] + d[34]*u[5]); 7681278733SSatish Balay uik[5] = -(d[5]*u[0]+ d[11]*u[1] + d[17]*u[2] + d[23]*u[3] + d[29]*u[4] + d[35]*u[5]); 7781278733SSatish Balay 7881278733SSatish Balay uik[6] = -(d[0]*u[6] + d[6]*u[7] + d[12]*u[8] + d[18]*u[9] + d[24]*u[10] + d[30]*u[11]); 7981278733SSatish Balay uik[7] = -(d[1]*u[6] + d[7]*u[7] + d[13]*u[8] + d[19]*u[9] + d[25]*u[10] + d[31]*u[11]); 8081278733SSatish Balay uik[8] = -(d[2]*u[6] + d[8]*u[7] + d[14]*u[8] + d[20]*u[9] + d[26]*u[10] + d[32]*u[11]); 8181278733SSatish Balay uik[9] = -(d[3]*u[6] + d[9]*u[7] + d[15]*u[8] + d[21]*u[9] + d[27]*u[10] + d[33]*u[11]); 8281278733SSatish Balay uik[10]= -(d[4]*u[6]+ d[10]*u[7] + d[16]*u[8] + d[22]*u[9] + d[28]*u[10] + d[34]*u[11]); 8381278733SSatish Balay uik[11]= -(d[5]*u[6]+ d[11]*u[7] + d[17]*u[8] + d[23]*u[9] + d[29]*u[10] + d[35]*u[11]); 8481278733SSatish Balay 8581278733SSatish Balay uik[12] = -(d[0]*u[12] + d[6]*u[13] + d[12]*u[14] + d[18]*u[15] + d[24]*u[16] + d[30]*u[17]); 8681278733SSatish Balay uik[13] = -(d[1]*u[12] + d[7]*u[13] + d[13]*u[14] + d[19]*u[15] + d[25]*u[16] + d[31]*u[17]); 8781278733SSatish Balay uik[14] = -(d[2]*u[12] + d[8]*u[13] + d[14]*u[14] + d[20]*u[15] + d[26]*u[16] + d[32]*u[17]); 8881278733SSatish Balay uik[15] = -(d[3]*u[12] + d[9]*u[13] + d[15]*u[14] + d[21]*u[15] + d[27]*u[16] + d[33]*u[17]); 8981278733SSatish Balay uik[16] = -(d[4]*u[12]+ d[10]*u[13] + d[16]*u[14] + d[22]*u[15] + d[28]*u[16] + d[34]*u[17]); 9081278733SSatish Balay uik[17] = -(d[5]*u[12]+ d[11]*u[13] + d[17]*u[14] + d[23]*u[15] + d[29]*u[16] + d[35]*u[17]); 9181278733SSatish Balay 9281278733SSatish Balay uik[18] = -(d[0]*u[18] + d[6]*u[19] + d[12]*u[20] + d[18]*u[21] + d[24]*u[22] + d[30]*u[23]); 9381278733SSatish Balay uik[19] = -(d[1]*u[18] + d[7]*u[19] + d[13]*u[20] + d[19]*u[21] + d[25]*u[22] + d[31]*u[23]); 9481278733SSatish Balay uik[20] = -(d[2]*u[18] + d[8]*u[19] + d[14]*u[20] + d[20]*u[21] + d[26]*u[22] + d[32]*u[23]); 9581278733SSatish Balay uik[21] = -(d[3]*u[18] + d[9]*u[19] + d[15]*u[20] + d[21]*u[21] + d[27]*u[22] + d[33]*u[23]); 9681278733SSatish Balay uik[22] = -(d[4]*u[18]+ d[10]*u[19] + d[16]*u[20] + d[22]*u[21] + d[28]*u[22] + d[34]*u[23]); 9781278733SSatish Balay uik[23] = -(d[5]*u[18]+ d[11]*u[19] + d[17]*u[20] + d[23]*u[21] + d[29]*u[22] + d[35]*u[23]); 9881278733SSatish Balay 9981278733SSatish Balay uik[24] = -(d[0]*u[24] + d[6]*u[25] + d[12]*u[26] + d[18]*u[27] + d[24]*u[28] + d[30]*u[29]); 10081278733SSatish Balay uik[25] = -(d[1]*u[24] + d[7]*u[25] + d[13]*u[26] + d[19]*u[27] + d[25]*u[28] + d[31]*u[29]); 10181278733SSatish Balay uik[26] = -(d[2]*u[24] + d[8]*u[25] + d[14]*u[26] + d[20]*u[27] + d[26]*u[28] + d[32]*u[29]); 10281278733SSatish Balay uik[27] = -(d[3]*u[24] + d[9]*u[25] + d[15]*u[26] + d[21]*u[27] + d[27]*u[28] + d[33]*u[29]); 10381278733SSatish Balay uik[28] = -(d[4]*u[24]+ d[10]*u[25] + d[16]*u[26] + d[22]*u[27] + d[28]*u[28] + d[34]*u[29]); 10481278733SSatish Balay uik[29] = -(d[5]*u[24]+ d[11]*u[25] + d[17]*u[26] + d[23]*u[27] + d[29]*u[28] + d[35]*u[29]); 10581278733SSatish Balay 10681278733SSatish Balay uik[30] = -(d[0]*u[30] + d[6]*u[31] + d[12]*u[32] + d[18]*u[33] + d[24]*u[34] + d[30]*u[35]); 10781278733SSatish Balay uik[31] = -(d[1]*u[30] + d[7]*u[31] + d[13]*u[32] + d[19]*u[33] + d[25]*u[34] + d[31]*u[35]); 10881278733SSatish Balay uik[32] = -(d[2]*u[30] + d[8]*u[31] + d[14]*u[32] + d[20]*u[33] + d[26]*u[34] + d[32]*u[35]); 10981278733SSatish Balay uik[33] = -(d[3]*u[30] + d[9]*u[31] + d[15]*u[32] + d[21]*u[33] + d[27]*u[34] + d[33]*u[35]); 11081278733SSatish Balay uik[34] = -(d[4]*u[30]+ d[10]*u[31] + d[16]*u[32] + d[22]*u[33] + d[28]*u[34] + d[34]*u[35]); 11181278733SSatish Balay uik[35] = -(d[5]*u[30]+ d[11]*u[31] + d[17]*u[32] + d[23]*u[33] + d[29]*u[34] + d[35]*u[35]); 11281278733SSatish Balay 11381278733SSatish Balay /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 11481278733SSatish Balay dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 11581278733SSatish Balay dk[1] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2] + uik[9]*u[3]+ uik[10]*u[4]+ uik[11]*u[5]; 11681278733SSatish Balay dk[2] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]+ uik[16]*u[4]+ uik[17]*u[5]; 11781278733SSatish Balay dk[3] += uik[18]*u[0]+ uik[19]*u[1]+ uik[20]*u[2]+ uik[21]*u[3]+ uik[22]*u[4]+ uik[23]*u[5]; 11881278733SSatish Balay dk[4] += uik[24]*u[0]+ uik[25]*u[1]+ uik[26]*u[2]+ uik[27]*u[3]+ uik[28]*u[4]+ uik[29]*u[5]; 11981278733SSatish Balay dk[5] += uik[30]*u[0]+ uik[31]*u[1]+ uik[32]*u[2]+ uik[33]*u[3]+ uik[34]*u[4]+ uik[35]*u[5]; 12081278733SSatish Balay 12181278733SSatish Balay dk[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8] + uik[3]*u[9] + uik[4]*u[10] + uik[5]*u[11]; 12281278733SSatish Balay dk[7] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]; 12381278733SSatish Balay dk[8] += uik[12]*u[6]+ uik[13]*u[7]+ uik[14]*u[8]+ uik[15]*u[9]+ uik[16]*u[10]+ uik[17]*u[11]; 12481278733SSatish Balay dk[9] += uik[18]*u[6]+ uik[19]*u[7]+ uik[20]*u[8]+ uik[21]*u[9]+ uik[22]*u[10]+ uik[23]*u[11]; 12581278733SSatish Balay dk[10]+= uik[24]*u[6]+ uik[25]*u[7]+ uik[26]*u[8]+ uik[27]*u[9]+ uik[28]*u[10]+ uik[29]*u[11]; 12681278733SSatish Balay dk[11]+= uik[30]*u[6]+ uik[31]*u[7]+ uik[32]*u[8]+ uik[33]*u[9]+ uik[34]*u[10]+ uik[35]*u[11]; 12781278733SSatish Balay 12881278733SSatish Balay dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15] + uik[4]*u[16] + uik[5]*u[17]; 12981278733SSatish Balay dk[13]+= uik[6]*u[12] + uik[7]*u[13] + uik[8]*u[14] + uik[9]*u[15]+ uik[10]*u[16]+ uik[11]*u[17]; 13081278733SSatish Balay dk[14]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]; 13181278733SSatish Balay dk[15]+= uik[18]*u[12]+ uik[19]*u[13]+ uik[20]*u[14]+ uik[21]*u[15]+ uik[22]*u[16]+ uik[23]*u[17]; 13281278733SSatish Balay dk[16]+= uik[24]*u[12]+ uik[25]*u[13]+ uik[26]*u[14]+ uik[27]*u[15]+ uik[28]*u[16]+ uik[29]*u[17]; 13381278733SSatish Balay dk[17]+= uik[30]*u[12]+ uik[31]*u[13]+ uik[32]*u[14]+ uik[33]*u[15]+ uik[34]*u[16]+ uik[35]*u[17]; 13481278733SSatish Balay 13581278733SSatish Balay dk[18]+= uik[0]*u[18] + uik[1]*u[19] + uik[2]*u[20] + uik[3]*u[21] + uik[4]*u[22] + uik[5]*u[23]; 13681278733SSatish Balay dk[19]+= uik[6]*u[18] + uik[7]*u[19] + uik[8]*u[20] + uik[9]*u[21]+ uik[10]*u[22]+ uik[11]*u[23]; 13781278733SSatish Balay dk[20]+= uik[12]*u[18]+ uik[13]*u[19]+ uik[14]*u[20]+ uik[15]*u[21]+ uik[16]*u[22]+ uik[17]*u[23]; 13881278733SSatish Balay dk[21]+= uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]; 13981278733SSatish Balay dk[22]+= uik[24]*u[18]+ uik[25]*u[19]+ uik[26]*u[20]+ uik[27]*u[21]+ uik[28]*u[22]+ uik[29]*u[23]; 14081278733SSatish Balay dk[23]+= uik[30]*u[18]+ uik[31]*u[19]+ uik[32]*u[20]+ uik[33]*u[21]+ uik[34]*u[22]+ uik[35]*u[23]; 14181278733SSatish Balay 14281278733SSatish Balay dk[24]+= uik[0]*u[24] + uik[1]*u[25] + uik[2]*u[26] + uik[3]*u[27] + uik[4]*u[28] + uik[5]*u[29]; 14381278733SSatish Balay dk[25]+= uik[6]*u[24] + uik[7]*u[25] + uik[8]*u[26] + uik[9]*u[27]+ uik[10]*u[28]+ uik[11]*u[29]; 14481278733SSatish Balay dk[26]+= uik[12]*u[24]+ uik[13]*u[25]+ uik[14]*u[26]+ uik[15]*u[27]+ uik[16]*u[28]+ uik[17]*u[29]; 14581278733SSatish Balay dk[27]+= uik[18]*u[24]+ uik[19]*u[25]+ uik[20]*u[26]+ uik[21]*u[27]+ uik[22]*u[28]+ uik[23]*u[29]; 14681278733SSatish Balay dk[28]+= uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]+ uik[28]*u[28]+ uik[29]*u[29]; 14781278733SSatish Balay dk[29]+= uik[30]*u[24]+ uik[31]*u[25]+ uik[32]*u[26]+ uik[33]*u[27]+ uik[34]*u[28]+ uik[35]*u[29]; 14881278733SSatish Balay 14981278733SSatish Balay dk[30]+= uik[0]*u[30] + uik[1]*u[31] + uik[2]*u[32] + uik[3]*u[33] + uik[4]*u[34] + uik[5]*u[35]; 15081278733SSatish Balay dk[31]+= uik[6]*u[30] + uik[7]*u[31] + uik[8]*u[32] + uik[9]*u[33]+ uik[10]*u[34]+ uik[11]*u[35]; 15181278733SSatish Balay dk[32]+= uik[12]*u[30]+ uik[13]*u[31]+ uik[14]*u[32]+ uik[15]*u[33]+ uik[16]*u[34]+ uik[17]*u[35]; 15281278733SSatish Balay dk[33]+= uik[18]*u[30]+ uik[19]*u[31]+ uik[20]*u[32]+ uik[21]*u[33]+ uik[22]*u[34]+ uik[23]*u[35]; 15381278733SSatish Balay dk[34]+= uik[24]*u[30]+ uik[25]*u[31]+ uik[26]*u[32]+ uik[27]*u[33]+ uik[28]*u[34]+ uik[29]*u[35]; 15481278733SSatish Balay dk[35]+= uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]+ uik[35]*u[35]; 15581278733SSatish Balay 15681278733SSatish Balay /* update -U(i,k) */ 15781278733SSatish Balay ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr); 15881278733SSatish Balay 15981278733SSatish Balay /* add multiple of row i to k-th row ... */ 16081278733SSatish Balay jmin = ili + 1; jmax = bi[i+1]; 16181278733SSatish Balay if (jmin < jmax){ 16281278733SSatish Balay for (j=jmin; j<jmax; j++) { 16381278733SSatish Balay /* w += -U(i,k)^T * U_bar(i,j) */ 16481278733SSatish Balay wp = w + bj[j]*36; 16581278733SSatish Balay u = ba + j*36; 16681278733SSatish Balay wp[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4] + uik[5]*u[5]; 16781278733SSatish Balay wp[1] += uik[6]*u[0] + uik[7]*u[1] + uik[8]*u[2] + uik[9]*u[3]+ uik[10]*u[4]+ uik[11]*u[5]; 16881278733SSatish Balay wp[2] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]+ uik[16]*u[4]+ uik[17]*u[5]; 16981278733SSatish Balay wp[3] += uik[18]*u[0]+ uik[19]*u[1]+ uik[20]*u[2]+ uik[21]*u[3]+ uik[22]*u[4]+ uik[23]*u[5]; 17081278733SSatish Balay wp[4] += uik[24]*u[0]+ uik[25]*u[1]+ uik[26]*u[2]+ uik[27]*u[3]+ uik[28]*u[4]+ uik[29]*u[5]; 17181278733SSatish Balay wp[5] += uik[30]*u[0]+ uik[31]*u[1]+ uik[32]*u[2]+ uik[33]*u[3]+ uik[34]*u[4]+ uik[35]*u[5]; 17281278733SSatish Balay 17381278733SSatish Balay wp[6] += uik[0]*u[6] + uik[1]*u[7] + uik[2]*u[8] + uik[3]*u[9] + uik[4]*u[10] + uik[5]*u[11]; 17481278733SSatish Balay wp[7] += uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9]+ uik[10]*u[10]+ uik[11]*u[11]; 17581278733SSatish Balay wp[8] += uik[12]*u[6]+ uik[13]*u[7]+ uik[14]*u[8]+ uik[15]*u[9]+ uik[16]*u[10]+ uik[17]*u[11]; 17681278733SSatish Balay wp[9] += uik[18]*u[6]+ uik[19]*u[7]+ uik[20]*u[8]+ uik[21]*u[9]+ uik[22]*u[10]+ uik[23]*u[11]; 17781278733SSatish Balay wp[10]+= uik[24]*u[6]+ uik[25]*u[7]+ uik[26]*u[8]+ uik[27]*u[9]+ uik[28]*u[10]+ uik[29]*u[11]; 17881278733SSatish Balay wp[11]+= uik[30]*u[6]+ uik[31]*u[7]+ uik[32]*u[8]+ uik[33]*u[9]+ uik[34]*u[10]+ uik[35]*u[11]; 17981278733SSatish Balay 18081278733SSatish Balay wp[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15] + uik[4]*u[16] + uik[5]*u[17]; 18181278733SSatish Balay wp[13]+= uik[6]*u[12] + uik[7]*u[13] + uik[8]*u[14] + uik[9]*u[15]+ uik[10]*u[16]+ uik[11]*u[17]; 18281278733SSatish Balay wp[14]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]; 18381278733SSatish Balay wp[15]+= uik[18]*u[12]+ uik[19]*u[13]+ uik[20]*u[14]+ uik[21]*u[15]+ uik[22]*u[16]+ uik[23]*u[17]; 18481278733SSatish Balay wp[16]+= uik[24]*u[12]+ uik[25]*u[13]+ uik[26]*u[14]+ uik[27]*u[15]+ uik[28]*u[16]+ uik[29]*u[17]; 18581278733SSatish Balay wp[17]+= uik[30]*u[12]+ uik[31]*u[13]+ uik[32]*u[14]+ uik[33]*u[15]+ uik[34]*u[16]+ uik[35]*u[17]; 18681278733SSatish Balay 18781278733SSatish Balay wp[18]+= uik[0]*u[18] + uik[1]*u[19] + uik[2]*u[20] + uik[3]*u[21] + uik[4]*u[22] + uik[5]*u[23]; 18881278733SSatish Balay wp[19]+= uik[6]*u[18] + uik[7]*u[19] + uik[8]*u[20] + uik[9]*u[21]+ uik[10]*u[22]+ uik[11]*u[23]; 18981278733SSatish Balay wp[20]+= uik[12]*u[18]+ uik[13]*u[19]+ uik[14]*u[20]+ uik[15]*u[21]+ uik[16]*u[22]+ uik[17]*u[23]; 19081278733SSatish Balay wp[21]+= uik[18]*u[18]+ uik[19]*u[19]+ uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]; 19181278733SSatish Balay wp[22]+= uik[24]*u[18]+ uik[25]*u[19]+ uik[26]*u[20]+ uik[27]*u[21]+ uik[28]*u[22]+ uik[29]*u[23]; 19281278733SSatish Balay wp[23]+= uik[30]*u[18]+ uik[31]*u[19]+ uik[32]*u[20]+ uik[33]*u[21]+ uik[34]*u[22]+ uik[35]*u[23]; 19381278733SSatish Balay 19481278733SSatish Balay wp[24]+= uik[0]*u[24] + uik[1]*u[25] + uik[2]*u[26] + uik[3]*u[27] + uik[4]*u[28] + uik[5]*u[29]; 19581278733SSatish Balay wp[25]+= uik[6]*u[24] + uik[7]*u[25] + uik[8]*u[26] + uik[9]*u[27]+ uik[10]*u[28]+ uik[11]*u[29]; 19681278733SSatish Balay wp[26]+= uik[12]*u[24]+ uik[13]*u[25]+ uik[14]*u[26]+ uik[15]*u[27]+ uik[16]*u[28]+ uik[17]*u[29]; 19781278733SSatish Balay wp[27]+= uik[18]*u[24]+ uik[19]*u[25]+ uik[20]*u[26]+ uik[21]*u[27]+ uik[22]*u[28]+ uik[23]*u[29]; 19881278733SSatish Balay wp[28]+= uik[24]*u[24]+ uik[25]*u[25]+ uik[26]*u[26]+ uik[27]*u[27]+ uik[28]*u[28]+ uik[29]*u[29]; 19981278733SSatish Balay wp[29]+= uik[30]*u[24]+ uik[31]*u[25]+ uik[32]*u[26]+ uik[33]*u[27]+ uik[34]*u[28]+ uik[35]*u[29]; 20081278733SSatish Balay 20181278733SSatish Balay wp[30]+= uik[0]*u[30] + uik[1]*u[31] + uik[2]*u[32] + uik[3]*u[33] + uik[4]*u[34] + uik[5]*u[35]; 20281278733SSatish Balay wp[31]+= uik[6]*u[30] + uik[7]*u[31] + uik[8]*u[32] + uik[9]*u[33]+ uik[10]*u[34]+ uik[11]*u[35]; 20381278733SSatish Balay wp[32]+= uik[12]*u[30]+ uik[13]*u[31]+ uik[14]*u[32]+ uik[15]*u[33]+ uik[16]*u[34]+ uik[17]*u[35]; 20481278733SSatish Balay wp[33]+= uik[18]*u[30]+ uik[19]*u[31]+ uik[20]*u[32]+ uik[21]*u[33]+ uik[22]*u[34]+ uik[23]*u[35]; 20581278733SSatish Balay wp[34]+= uik[24]*u[30]+ uik[25]*u[31]+ uik[26]*u[32]+ uik[27]*u[33]+ uik[28]*u[34]+ uik[29]*u[35]; 20681278733SSatish Balay wp[35]+= uik[30]*u[30]+ uik[31]*u[31]+ uik[32]*u[32]+ uik[33]*u[33]+ uik[34]*u[34]+ uik[35]*u[35]; 20781278733SSatish Balay } 20881278733SSatish Balay 20981278733SSatish Balay /* ... add i to row list for next nonzero entry */ 21081278733SSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 21181278733SSatish Balay j = bj[jmin]; 21281278733SSatish Balay jl[i] = jl[j]; jl[j] = i; /* update jl */ 21381278733SSatish Balay } 21481278733SSatish Balay i = nexti; 21581278733SSatish Balay } 21681278733SSatish Balay 21781278733SSatish Balay /* save nonzero entries in k-th row of U ... */ 21881278733SSatish Balay 21981278733SSatish Balay /* invert diagonal block */ 22081278733SSatish Balay d = ba+k*36; 22181278733SSatish Balay ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr); 22281278733SSatish Balay ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr); 22381278733SSatish Balay 22481278733SSatish Balay jmin = bi[k]; jmax = bi[k+1]; 22581278733SSatish Balay if (jmin < jmax) { 22681278733SSatish Balay for (j=jmin; j<jmax; j++){ 22781278733SSatish Balay vj = bj[j]; /* block col. index of U */ 22881278733SSatish Balay u = ba + j*36; 22981278733SSatish Balay wp = w + vj*36; 23081278733SSatish Balay for (k1=0; k1<36; k1++){ 23181278733SSatish Balay *u++ = *wp; 23281278733SSatish Balay *wp++ = 0.0; 23381278733SSatish Balay } 23481278733SSatish Balay } 23581278733SSatish Balay 23681278733SSatish Balay /* ... add k to row list for first nonzero entry in k-th row */ 23781278733SSatish Balay il[k] = jmin; 23881278733SSatish Balay i = bj[jmin]; 23981278733SSatish Balay jl[k] = jl[i]; jl[i] = k; 24081278733SSatish Balay } 24181278733SSatish Balay } 24281278733SSatish Balay 24381278733SSatish Balay ierr = PetscFree(w);CHKERRQ(ierr); 24481278733SSatish Balay ierr = PetscFree(il);CHKERRQ(ierr); 24581278733SSatish Balay ierr = PetscFree(dk);CHKERRQ(ierr); 24681278733SSatish Balay 24781278733SSatish Balay C->factor = FACTOR_CHOLESKY; 24881278733SSatish Balay C->assembled = PETSC_TRUE; 24981278733SSatish Balay C->preallocated = PETSC_TRUE; 25081278733SSatish Balay PetscLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */ 25181278733SSatish Balay PetscFunctionReturn(0); 25281278733SSatish Balay } 253