1*1b3064deSBarry 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 /* Version for when blocks are 6 by 6 */ 681278733SSatish Balay #undef __FUNC__ 781278733SSatish Balay #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6" 881278733SSatish Balay int MatCholeskyFactorNumeric_SeqSBAIJ_6(Mat A,Mat *B) 981278733SSatish Balay { 1081278733SSatish Balay Mat C = *B; 1181278733SSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 1281278733SSatish Balay IS perm = b->row; 1381278733SSatish Balay int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 1481278733SSatish Balay int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 1581278733SSatish Balay MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 16*1b3064deSBarry Smith MatScalar *u,*d,*w,*wp,u0,u1,u2,u3,u4,u5,u6,u7,u8,u9,u10,u11,u12; 17*1b3064deSBarry Smith MatScalar u13,u14,u15,u16,u17,u18,u19,u20,u21,u22,u23,u24,u25,u26,u27; 18*1b3064deSBarry Smith MatScalar u28,u29,u30,u31,u32,u33,u34,u35; 1981278733SSatish Balay 2081278733SSatish Balay PetscFunctionBegin; 2181278733SSatish Balay /* initialization */ 2281278733SSatish Balay ierr = PetscMalloc(36*mbs*sizeof(MatScalar),&w);CHKERRQ(ierr); 2381278733SSatish Balay ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2481278733SSatish Balay ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr); 2581278733SSatish Balay jl = il + mbs; 2681278733SSatish Balay for (i=0; i<mbs; i++) { 2781278733SSatish Balay jl[i] = mbs; il[0] = 0; 2881278733SSatish Balay } 2981278733SSatish Balay ierr = PetscMalloc(72*sizeof(MatScalar),&dk);CHKERRQ(ierr); 3081278733SSatish Balay uik = dk + 36; 3181278733SSatish Balay ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 3281278733SSatish Balay 3381278733SSatish Balay /* check permutation */ 3481278733SSatish Balay if (!a->permute){ 3581278733SSatish Balay ai = a->i; aj = a->j; aa = a->a; 3681278733SSatish Balay } else { 3781278733SSatish Balay ai = a->inew; aj = a->jnew; 3881278733SSatish Balay ierr = PetscMalloc(36*ai[mbs]*sizeof(MatScalar),&aa);CHKERRQ(ierr); 3981278733SSatish Balay ierr = PetscMemcpy(aa,a->a,36*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 4081278733SSatish Balay ierr = PetscMalloc(ai[mbs]*sizeof(int),&a2anew);CHKERRQ(ierr); 4181278733SSatish Balay ierr = PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 4281278733SSatish Balay 4381278733SSatish Balay for (i=0; i<mbs; i++){ 4481278733SSatish Balay jmin = ai[i]; jmax = ai[i+1]; 4581278733SSatish Balay for (j=jmin; j<jmax; j++){ 4681278733SSatish Balay while (a2anew[j] != j){ 4781278733SSatish Balay k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 4881278733SSatish Balay for (k1=0; k1<36; k1++){ 4981278733SSatish Balay dk[k1] = aa[k*36+k1]; 5081278733SSatish Balay aa[k*36+k1] = aa[j*36+k1]; 5181278733SSatish Balay aa[j*36+k1] = dk[k1]; 5281278733SSatish Balay } 5381278733SSatish Balay } 5481278733SSatish Balay /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 5581278733SSatish Balay if (i > aj[j]){ 5681278733SSatish Balay /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 5781278733SSatish Balay ap = aa + j*36; /* ptr to the beginning of j-th block of aa */ 5881278733SSatish Balay for (k=0; k<36; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 5981278733SSatish Balay for (k=0; k<6; k++){ /* j-th block of aa <- dk^T */ 6081278733SSatish Balay for (k1=0; k1<6; k1++) *ap++ = dk[k + 6*k1]; 6181278733SSatish Balay } 6281278733SSatish Balay } 6381278733SSatish Balay } 6481278733SSatish Balay } 6581278733SSatish Balay ierr = PetscFree(a2anew);CHKERRQ(ierr); 6681278733SSatish Balay } 6781278733SSatish Balay 6881278733SSatish Balay /* for each row k */ 6981278733SSatish Balay for (k = 0; k<mbs; k++){ 7081278733SSatish Balay 7181278733SSatish Balay /*initialize k-th row with elements nonzero in row perm(k) of A */ 7281278733SSatish Balay jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 7381278733SSatish Balay if (jmin < jmax) { 7481278733SSatish Balay ap = aa + jmin*36; 7581278733SSatish Balay for (j = jmin; j < jmax; j++){ 7681278733SSatish Balay vj = perm_ptr[aj[j]]; /* block col. index */ 7781278733SSatish Balay wp = w + vj*36; 7881278733SSatish Balay for (i=0; i<36; i++) *wp++ = *ap++; 7981278733SSatish Balay } 8081278733SSatish Balay } 8181278733SSatish Balay 8281278733SSatish Balay /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 8381278733SSatish Balay ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr); 8481278733SSatish Balay i = jl[k]; /* first row to be added to k_th row */ 8581278733SSatish Balay 8681278733SSatish Balay while (i < mbs){ 8781278733SSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 8881278733SSatish Balay 8981278733SSatish Balay /* compute multiplier */ 9081278733SSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 9181278733SSatish Balay 9281278733SSatish Balay /* uik = -inv(Di)*U_bar(i,k) */ 9381278733SSatish Balay d = ba + i*36; 9481278733SSatish Balay u = ba + ili*36; 9581278733SSatish Balay 96*1b3064deSBarry Smith u0 = u[0]; u1 = u[1]; u2 = u[2]; u3 = u[3]; u4 = u[4]; u5 = u[5]; u6 = u[6]; 97*1b3064deSBarry Smith u7 = u[7]; u8 = u[8]; u9 = u[9]; u10 = u[10]; u11 = u[11]; u12 = u[12]; u13 = u[13]; 98*1b3064deSBarry Smith u14 = u[14]; u15 = u[15]; u16 = u[16]; u17 = u[17]; u18 = u[18]; u19 = u[19]; u20 = u[20]; 99*1b3064deSBarry Smith u21 = u[21]; u22 = u[22]; u23 = u[23]; u24 = u[24]; u25 = u[25]; u26 = u[26]; u27 = u[27]; 100*1b3064deSBarry Smith u28 = u[28]; u29 = u[29]; u30 = u[30]; u31 = u[31]; u32 = u[32]; u33 = u[33]; u34 = u[34]; 101*1b3064deSBarry Smith u35 = u[35]; 10281278733SSatish Balay 103*1b3064deSBarry Smith uik[0] = -(d[0]*u0 + d[6]*u1 + d[12]*u2 + d[18]*u3 + d[24]*u4 + d[30]*u5); 104*1b3064deSBarry Smith uik[1] = -(d[1]*u0 + d[7]*u1 + d[13]*u2 + d[19]*u3 + d[25]*u4 + d[31]*u5); 105*1b3064deSBarry Smith uik[2] = -(d[2]*u0 + d[8]*u1 + d[14]*u2 + d[20]*u3 + d[26]*u4 + d[32]*u5); 106*1b3064deSBarry Smith uik[3] = -(d[3]*u0 + d[9]*u1 + d[15]*u2 + d[21]*u3 + d[27]*u4 + d[33]*u5); 107*1b3064deSBarry Smith uik[4] = -(d[4]*u0+ d[10]*u1 + d[16]*u2 + d[22]*u3 + d[28]*u4 + d[34]*u5); 108*1b3064deSBarry Smith uik[5] = -(d[5]*u0+ d[11]*u1 + d[17]*u2 + d[23]*u3 + d[29]*u4 + d[35]*u5); 10981278733SSatish Balay 110*1b3064deSBarry Smith uik[6] = -(d[0]*u6 + d[6]*u7 + d[12]*u8 + d[18]*u9 + d[24]*u10 + d[30]*u11); 111*1b3064deSBarry Smith uik[7] = -(d[1]*u6 + d[7]*u7 + d[13]*u8 + d[19]*u9 + d[25]*u10 + d[31]*u11); 112*1b3064deSBarry Smith uik[8] = -(d[2]*u6 + d[8]*u7 + d[14]*u8 + d[20]*u9 + d[26]*u10 + d[32]*u11); 113*1b3064deSBarry Smith uik[9] = -(d[3]*u6 + d[9]*u7 + d[15]*u8 + d[21]*u9 + d[27]*u10 + d[33]*u11); 114*1b3064deSBarry Smith uik[10]= -(d[4]*u6+ d[10]*u7 + d[16]*u8 + d[22]*u9 + d[28]*u10 + d[34]*u11); 115*1b3064deSBarry Smith uik[11]= -(d[5]*u6+ d[11]*u7 + d[17]*u8 + d[23]*u9 + d[29]*u10 + d[35]*u11); 11681278733SSatish Balay 117*1b3064deSBarry Smith uik[12] = -(d[0]*u12 + d[6]*u13 + d[12]*u14 + d[18]*u15 + d[24]*u16 + d[30]*u17); 118*1b3064deSBarry Smith uik[13] = -(d[1]*u12 + d[7]*u13 + d[13]*u14 + d[19]*u15 + d[25]*u16 + d[31]*u17); 119*1b3064deSBarry Smith uik[14] = -(d[2]*u12 + d[8]*u13 + d[14]*u14 + d[20]*u15 + d[26]*u16 + d[32]*u17); 120*1b3064deSBarry Smith uik[15] = -(d[3]*u12 + d[9]*u13 + d[15]*u14 + d[21]*u15 + d[27]*u16 + d[33]*u17); 121*1b3064deSBarry Smith uik[16] = -(d[4]*u12+ d[10]*u13 + d[16]*u14 + d[22]*u15 + d[28]*u16 + d[34]*u17); 122*1b3064deSBarry Smith uik[17] = -(d[5]*u12+ d[11]*u13 + d[17]*u14 + d[23]*u15 + d[29]*u16 + d[35]*u17); 12381278733SSatish Balay 124*1b3064deSBarry Smith uik[18] = -(d[0]*u18 + d[6]*u19 + d[12]*u20 + d[18]*u21 + d[24]*u22 + d[30]*u23); 125*1b3064deSBarry Smith uik[19] = -(d[1]*u18 + d[7]*u19 + d[13]*u20 + d[19]*u21 + d[25]*u22 + d[31]*u23); 126*1b3064deSBarry Smith uik[20] = -(d[2]*u18 + d[8]*u19 + d[14]*u20 + d[20]*u21 + d[26]*u22 + d[32]*u23); 127*1b3064deSBarry Smith uik[21] = -(d[3]*u18 + d[9]*u19 + d[15]*u20 + d[21]*u21 + d[27]*u22 + d[33]*u23); 128*1b3064deSBarry Smith uik[22] = -(d[4]*u18+ d[10]*u19 + d[16]*u20 + d[22]*u21 + d[28]*u22 + d[34]*u23); 129*1b3064deSBarry Smith uik[23] = -(d[5]*u18+ d[11]*u19 + d[17]*u20 + d[23]*u21 + d[29]*u22 + d[35]*u23); 13081278733SSatish Balay 131*1b3064deSBarry Smith uik[24] = -(d[0]*u24 + d[6]*u25 + d[12]*u26 + d[18]*u27 + d[24]*u28 + d[30]*u29); 132*1b3064deSBarry Smith uik[25] = -(d[1]*u24 + d[7]*u25 + d[13]*u26 + d[19]*u27 + d[25]*u28 + d[31]*u29); 133*1b3064deSBarry Smith uik[26] = -(d[2]*u24 + d[8]*u25 + d[14]*u26 + d[20]*u27 + d[26]*u28 + d[32]*u29); 134*1b3064deSBarry Smith uik[27] = -(d[3]*u24 + d[9]*u25 + d[15]*u26 + d[21]*u27 + d[27]*u28 + d[33]*u29); 135*1b3064deSBarry Smith uik[28] = -(d[4]*u24+ d[10]*u25 + d[16]*u26 + d[22]*u27 + d[28]*u28 + d[34]*u29); 136*1b3064deSBarry Smith uik[29] = -(d[5]*u24+ d[11]*u25 + d[17]*u26 + d[23]*u27 + d[29]*u28 + d[35]*u29); 137*1b3064deSBarry Smith 138*1b3064deSBarry Smith uik[30] = -(d[0]*u30 + d[6]*u31 + d[12]*u32 + d[18]*u33 + d[24]*u34 + d[30]*u35); 139*1b3064deSBarry Smith uik[31] = -(d[1]*u30 + d[7]*u31 + d[13]*u32 + d[19]*u33 + d[25]*u34 + d[31]*u35); 140*1b3064deSBarry Smith uik[32] = -(d[2]*u30 + d[8]*u31 + d[14]*u32 + d[20]*u33 + d[26]*u34 + d[32]*u35); 141*1b3064deSBarry Smith uik[33] = -(d[3]*u30 + d[9]*u31 + d[15]*u32 + d[21]*u33 + d[27]*u34 + d[33]*u35); 142*1b3064deSBarry Smith uik[34] = -(d[4]*u30+ d[10]*u31 + d[16]*u32 + d[22]*u33 + d[28]*u34 + d[34]*u35); 143*1b3064deSBarry Smith uik[35] = -(d[5]*u30+ d[11]*u31 + d[17]*u32 + d[23]*u33 + d[29]*u34 + d[35]*u35); 14481278733SSatish Balay 14581278733SSatish Balay /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 146*1b3064deSBarry Smith dk[0] += uik[0]*u0 + uik[1]*u1 + uik[2]*u2 + uik[3]*u3 + uik[4]*u4 + uik[5]*u5; 147*1b3064deSBarry Smith dk[1] += uik[6]*u0 + uik[7]*u1 + uik[8]*u2 + uik[9]*u3+ uik[10]*u4+ uik[11]*u5; 148*1b3064deSBarry Smith dk[2] += uik[12]*u0+ uik[13]*u1+ uik[14]*u2+ uik[15]*u3+ uik[16]*u4+ uik[17]*u5; 149*1b3064deSBarry Smith dk[3] += uik[18]*u0+ uik[19]*u1+ uik[20]*u2+ uik[21]*u3+ uik[22]*u4+ uik[23]*u5; 150*1b3064deSBarry Smith dk[4] += uik[24]*u0+ uik[25]*u1+ uik[26]*u2+ uik[27]*u3+ uik[28]*u4+ uik[29]*u5; 151*1b3064deSBarry Smith dk[5] += uik[30]*u0+ uik[31]*u1+ uik[32]*u2+ uik[33]*u3+ uik[34]*u4+ uik[35]*u5; 15281278733SSatish Balay 153*1b3064deSBarry Smith dk[6] += uik[0]*u6 + uik[1]*u7 + uik[2]*u8 + uik[3]*u9 + uik[4]*u10 + uik[5]*u11; 154*1b3064deSBarry Smith dk[7] += uik[6]*u6 + uik[7]*u7 + uik[8]*u8 + uik[9]*u9+ uik[10]*u10+ uik[11]*u11; 155*1b3064deSBarry Smith dk[8] += uik[12]*u6+ uik[13]*u7+ uik[14]*u8+ uik[15]*u9+ uik[16]*u10+ uik[17]*u11; 156*1b3064deSBarry Smith dk[9] += uik[18]*u6+ uik[19]*u7+ uik[20]*u8+ uik[21]*u9+ uik[22]*u10+ uik[23]*u11; 157*1b3064deSBarry Smith dk[10]+= uik[24]*u6+ uik[25]*u7+ uik[26]*u8+ uik[27]*u9+ uik[28]*u10+ uik[29]*u11; 158*1b3064deSBarry Smith dk[11]+= uik[30]*u6+ uik[31]*u7+ uik[32]*u8+ uik[33]*u9+ uik[34]*u10+ uik[35]*u11; 15981278733SSatish Balay 160*1b3064deSBarry Smith dk[12]+= uik[0]*u12 + uik[1]*u13 + uik[2]*u14 + uik[3]*u15 + uik[4]*u16 + uik[5]*u17; 161*1b3064deSBarry Smith dk[13]+= uik[6]*u12 + uik[7]*u13 + uik[8]*u14 + uik[9]*u15+ uik[10]*u16+ uik[11]*u17; 162*1b3064deSBarry Smith dk[14]+= uik[12]*u12+ uik[13]*u13+ uik[14]*u14+ uik[15]*u15+ uik[16]*u16+ uik[17]*u17; 163*1b3064deSBarry Smith dk[15]+= uik[18]*u12+ uik[19]*u13+ uik[20]*u14+ uik[21]*u15+ uik[22]*u16+ uik[23]*u17; 164*1b3064deSBarry Smith dk[16]+= uik[24]*u12+ uik[25]*u13+ uik[26]*u14+ uik[27]*u15+ uik[28]*u16+ uik[29]*u17; 165*1b3064deSBarry Smith dk[17]+= uik[30]*u12+ uik[31]*u13+ uik[32]*u14+ uik[33]*u15+ uik[34]*u16+ uik[35]*u17; 16681278733SSatish Balay 167*1b3064deSBarry Smith dk[18]+= uik[0]*u18 + uik[1]*u19 + uik[2]*u20 + uik[3]*u21 + uik[4]*u22 + uik[5]*u23; 168*1b3064deSBarry Smith dk[19]+= uik[6]*u18 + uik[7]*u19 + uik[8]*u20 + uik[9]*u21+ uik[10]*u22+ uik[11]*u23; 169*1b3064deSBarry Smith dk[20]+= uik[12]*u18+ uik[13]*u19+ uik[14]*u20+ uik[15]*u21+ uik[16]*u22+ uik[17]*u23; 170*1b3064deSBarry Smith dk[21]+= uik[18]*u18+ uik[19]*u19+ uik[20]*u20+ uik[21]*u21+ uik[22]*u22+ uik[23]*u23; 171*1b3064deSBarry Smith dk[22]+= uik[24]*u18+ uik[25]*u19+ uik[26]*u20+ uik[27]*u21+ uik[28]*u22+ uik[29]*u23; 172*1b3064deSBarry Smith dk[23]+= uik[30]*u18+ uik[31]*u19+ uik[32]*u20+ uik[33]*u21+ uik[34]*u22+ uik[35]*u23; 17381278733SSatish Balay 174*1b3064deSBarry Smith dk[24]+= uik[0]*u24 + uik[1]*u25 + uik[2]*u26 + uik[3]*u27 + uik[4]*u28 + uik[5]*u29; 175*1b3064deSBarry Smith dk[25]+= uik[6]*u24 + uik[7]*u25 + uik[8]*u26 + uik[9]*u27+ uik[10]*u28+ uik[11]*u29; 176*1b3064deSBarry Smith dk[26]+= uik[12]*u24+ uik[13]*u25+ uik[14]*u26+ uik[15]*u27+ uik[16]*u28+ uik[17]*u29; 177*1b3064deSBarry Smith dk[27]+= uik[18]*u24+ uik[19]*u25+ uik[20]*u26+ uik[21]*u27+ uik[22]*u28+ uik[23]*u29; 178*1b3064deSBarry Smith dk[28]+= uik[24]*u24+ uik[25]*u25+ uik[26]*u26+ uik[27]*u27+ uik[28]*u28+ uik[29]*u29; 179*1b3064deSBarry Smith dk[29]+= uik[30]*u24+ uik[31]*u25+ uik[32]*u26+ uik[33]*u27+ uik[34]*u28+ uik[35]*u29; 18081278733SSatish Balay 181*1b3064deSBarry Smith dk[30]+= uik[0]*u30 + uik[1]*u31 + uik[2]*u32 + uik[3]*u33 + uik[4]*u34 + uik[5]*u35; 182*1b3064deSBarry Smith dk[31]+= uik[6]*u30 + uik[7]*u31 + uik[8]*u32 + uik[9]*u33+ uik[10]*u34+ uik[11]*u35; 183*1b3064deSBarry Smith dk[32]+= uik[12]*u30+ uik[13]*u31+ uik[14]*u32+ uik[15]*u33+ uik[16]*u34+ uik[17]*u35; 184*1b3064deSBarry Smith dk[33]+= uik[18]*u30+ uik[19]*u31+ uik[20]*u32+ uik[21]*u33+ uik[22]*u34+ uik[23]*u35; 185*1b3064deSBarry Smith dk[34]+= uik[24]*u30+ uik[25]*u31+ uik[26]*u32+ uik[27]*u33+ uik[28]*u34+ uik[29]*u35; 186*1b3064deSBarry Smith dk[35]+= uik[30]*u30+ uik[31]*u31+ uik[32]*u32+ uik[33]*u33+ uik[34]*u34+ uik[35]*u35; 18781278733SSatish Balay 18881278733SSatish Balay /* update -U(i,k) */ 18981278733SSatish Balay ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr); 19081278733SSatish Balay 19181278733SSatish Balay /* add multiple of row i to k-th row ... */ 19281278733SSatish Balay jmin = ili + 1; jmax = bi[i+1]; 19381278733SSatish Balay if (jmin < jmax){ 19481278733SSatish Balay for (j=jmin; j<jmax; j++) { 19581278733SSatish Balay /* w += -U(i,k)^T * U_bar(i,j) */ 19681278733SSatish Balay wp = w + bj[j]*36; 19781278733SSatish Balay u = ba + j*36; 19881278733SSatish Balay 199*1b3064deSBarry Smith u0 = u[0]; u1 = u[1]; u2 = u[2]; u3 = u[3]; u4 = u[4]; u5 = u[5]; u6 = u[6]; 200*1b3064deSBarry Smith u7 = u[7]; u8 = u[8]; u9 = u[9]; u10 = u[10]; u11 = u[11]; u12 = u[12]; u13 = u[13]; 201*1b3064deSBarry Smith u14 = u[14]; u15 = u[15]; u16 = u[16]; u17 = u[17]; u18 = u[18]; u19 = u[19]; u20 = u[20]; 202*1b3064deSBarry Smith u21 = u[21]; u22 = u[22]; u23 = u[23]; u24 = u[24]; u25 = u[25]; u26 = u[26]; u27 = u[27]; 203*1b3064deSBarry Smith u28 = u[28]; u29 = u[29]; u30 = u[30]; u31 = u[31]; u32 = u[32]; u33 = u[33]; u34 = u[34]; 204*1b3064deSBarry Smith u35 = u[35]; 20581278733SSatish Balay 206*1b3064deSBarry Smith wp[0] += uik[0]*u0 + uik[1]*u1 + uik[2]*u2 + uik[3]*u3 + uik[4]*u4 + uik[5]*u5; 207*1b3064deSBarry Smith wp[1] += uik[6]*u0 + uik[7]*u1 + uik[8]*u2 + uik[9]*u3+ uik[10]*u4+ uik[11]*u5; 208*1b3064deSBarry Smith wp[2] += uik[12]*u0+ uik[13]*u1+ uik[14]*u2+ uik[15]*u3+ uik[16]*u4+ uik[17]*u5; 209*1b3064deSBarry Smith wp[3] += uik[18]*u0+ uik[19]*u1+ uik[20]*u2+ uik[21]*u3+ uik[22]*u4+ uik[23]*u5; 210*1b3064deSBarry Smith wp[4] += uik[24]*u0+ uik[25]*u1+ uik[26]*u2+ uik[27]*u3+ uik[28]*u4+ uik[29]*u5; 211*1b3064deSBarry Smith wp[5] += uik[30]*u0+ uik[31]*u1+ uik[32]*u2+ uik[33]*u3+ uik[34]*u4+ uik[35]*u5; 21281278733SSatish Balay 213*1b3064deSBarry Smith wp[6] += uik[0]*u6 + uik[1]*u7 + uik[2]*u8 + uik[3]*u9 + uik[4]*u10 + uik[5]*u11; 214*1b3064deSBarry Smith wp[7] += uik[6]*u6 + uik[7]*u7 + uik[8]*u8 + uik[9]*u9+ uik[10]*u10+ uik[11]*u11; 215*1b3064deSBarry Smith wp[8] += uik[12]*u6+ uik[13]*u7+ uik[14]*u8+ uik[15]*u9+ uik[16]*u10+ uik[17]*u11; 216*1b3064deSBarry Smith wp[9] += uik[18]*u6+ uik[19]*u7+ uik[20]*u8+ uik[21]*u9+ uik[22]*u10+ uik[23]*u11; 217*1b3064deSBarry Smith wp[10]+= uik[24]*u6+ uik[25]*u7+ uik[26]*u8+ uik[27]*u9+ uik[28]*u10+ uik[29]*u11; 218*1b3064deSBarry Smith wp[11]+= uik[30]*u6+ uik[31]*u7+ uik[32]*u8+ uik[33]*u9+ uik[34]*u10+ uik[35]*u11; 21981278733SSatish Balay 220*1b3064deSBarry Smith wp[12]+= uik[0]*u12 + uik[1]*u13 + uik[2]*u14 + uik[3]*u15 + uik[4]*u16 + uik[5]*u17; 221*1b3064deSBarry Smith wp[13]+= uik[6]*u12 + uik[7]*u13 + uik[8]*u14 + uik[9]*u15+ uik[10]*u16+ uik[11]*u17; 222*1b3064deSBarry Smith wp[14]+= uik[12]*u12+ uik[13]*u13+ uik[14]*u14+ uik[15]*u15+ uik[16]*u16+ uik[17]*u17; 223*1b3064deSBarry Smith wp[15]+= uik[18]*u12+ uik[19]*u13+ uik[20]*u14+ uik[21]*u15+ uik[22]*u16+ uik[23]*u17; 224*1b3064deSBarry Smith wp[16]+= uik[24]*u12+ uik[25]*u13+ uik[26]*u14+ uik[27]*u15+ uik[28]*u16+ uik[29]*u17; 225*1b3064deSBarry Smith wp[17]+= uik[30]*u12+ uik[31]*u13+ uik[32]*u14+ uik[33]*u15+ uik[34]*u16+ uik[35]*u17; 22681278733SSatish Balay 227*1b3064deSBarry Smith wp[18]+= uik[0]*u18 + uik[1]*u19 + uik[2]*u20 + uik[3]*u21 + uik[4]*u22 + uik[5]*u23; 228*1b3064deSBarry Smith wp[19]+= uik[6]*u18 + uik[7]*u19 + uik[8]*u20 + uik[9]*u21+ uik[10]*u22+ uik[11]*u23; 229*1b3064deSBarry Smith wp[20]+= uik[12]*u18+ uik[13]*u19+ uik[14]*u20+ uik[15]*u21+ uik[16]*u22+ uik[17]*u23; 230*1b3064deSBarry Smith wp[21]+= uik[18]*u18+ uik[19]*u19+ uik[20]*u20+ uik[21]*u21+ uik[22]*u22+ uik[23]*u23; 231*1b3064deSBarry Smith wp[22]+= uik[24]*u18+ uik[25]*u19+ uik[26]*u20+ uik[27]*u21+ uik[28]*u22+ uik[29]*u23; 232*1b3064deSBarry Smith wp[23]+= uik[30]*u18+ uik[31]*u19+ uik[32]*u20+ uik[33]*u21+ uik[34]*u22+ uik[35]*u23; 233*1b3064deSBarry Smith 234*1b3064deSBarry Smith wp[24]+= uik[0]*u24 + uik[1]*u25 + uik[2]*u26 + uik[3]*u27 + uik[4]*u28 + uik[5]*u29; 235*1b3064deSBarry Smith wp[25]+= uik[6]*u24 + uik[7]*u25 + uik[8]*u26 + uik[9]*u27+ uik[10]*u28+ uik[11]*u29; 236*1b3064deSBarry Smith wp[26]+= uik[12]*u24+ uik[13]*u25+ uik[14]*u26+ uik[15]*u27+ uik[16]*u28+ uik[17]*u29; 237*1b3064deSBarry Smith wp[27]+= uik[18]*u24+ uik[19]*u25+ uik[20]*u26+ uik[21]*u27+ uik[22]*u28+ uik[23]*u29; 238*1b3064deSBarry Smith wp[28]+= uik[24]*u24+ uik[25]*u25+ uik[26]*u26+ uik[27]*u27+ uik[28]*u28+ uik[29]*u29; 239*1b3064deSBarry Smith wp[29]+= uik[30]*u24+ uik[31]*u25+ uik[32]*u26+ uik[33]*u27+ uik[34]*u28+ uik[35]*u29; 240*1b3064deSBarry Smith 241*1b3064deSBarry Smith wp[30]+= uik[0]*u30 + uik[1]*u31 + uik[2]*u32 + uik[3]*u33 + uik[4]*u34 + uik[5]*u35; 242*1b3064deSBarry Smith wp[31]+= uik[6]*u30 + uik[7]*u31 + uik[8]*u32 + uik[9]*u33+ uik[10]*u34+ uik[11]*u35; 243*1b3064deSBarry Smith wp[32]+= uik[12]*u30+ uik[13]*u31+ uik[14]*u32+ uik[15]*u33+ uik[16]*u34+ uik[17]*u35; 244*1b3064deSBarry Smith wp[33]+= uik[18]*u30+ uik[19]*u31+ uik[20]*u32+ uik[21]*u33+ uik[22]*u34+ uik[23]*u35; 245*1b3064deSBarry Smith wp[34]+= uik[24]*u30+ uik[25]*u31+ uik[26]*u32+ uik[27]*u33+ uik[28]*u34+ uik[29]*u35; 246*1b3064deSBarry Smith wp[35]+= uik[30]*u30+ uik[31]*u31+ uik[32]*u32+ uik[33]*u33+ uik[34]*u34+ uik[35]*u35; 24781278733SSatish Balay } 24881278733SSatish Balay 24981278733SSatish Balay /* ... add i to row list for next nonzero entry */ 25081278733SSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 25181278733SSatish Balay j = bj[jmin]; 25281278733SSatish Balay jl[i] = jl[j]; jl[j] = i; /* update jl */ 25381278733SSatish Balay } 25481278733SSatish Balay i = nexti; 25581278733SSatish Balay } 25681278733SSatish Balay 25781278733SSatish Balay /* save nonzero entries in k-th row of U ... */ 25881278733SSatish Balay 25981278733SSatish Balay /* invert diagonal block */ 26081278733SSatish Balay d = ba+k*36; 26181278733SSatish Balay ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr); 26281278733SSatish Balay ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr); 26381278733SSatish Balay 26481278733SSatish Balay jmin = bi[k]; jmax = bi[k+1]; 26581278733SSatish Balay if (jmin < jmax) { 26681278733SSatish Balay for (j=jmin; j<jmax; j++){ 26781278733SSatish Balay vj = bj[j]; /* block col. index of U */ 26881278733SSatish Balay u = ba + j*36; 26981278733SSatish Balay wp = w + vj*36; 27081278733SSatish Balay for (k1=0; k1<36; k1++){ 27181278733SSatish Balay *u++ = *wp; 27281278733SSatish Balay *wp++ = 0.0; 27381278733SSatish Balay } 27481278733SSatish Balay } 27581278733SSatish Balay 27681278733SSatish Balay /* ... add k to row list for first nonzero entry in k-th row */ 27781278733SSatish Balay il[k] = jmin; 27881278733SSatish Balay i = bj[jmin]; 27981278733SSatish Balay jl[k] = jl[i]; jl[i] = k; 28081278733SSatish Balay } 28181278733SSatish Balay } 28281278733SSatish Balay 28381278733SSatish Balay ierr = PetscFree(w);CHKERRQ(ierr); 28481278733SSatish Balay ierr = PetscFree(il);CHKERRQ(ierr); 28581278733SSatish Balay ierr = PetscFree(dk);CHKERRQ(ierr); 28681278733SSatish Balay if (a->permute){ 28781278733SSatish Balay ierr = PetscFree(aa);CHKERRQ(ierr); 28881278733SSatish Balay } 28981278733SSatish Balay 29081278733SSatish Balay ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 29181278733SSatish Balay C->factor = FACTOR_CHOLESKY; 29281278733SSatish Balay C->assembled = PETSC_TRUE; 29381278733SSatish Balay C->preallocated = PETSC_TRUE; 29481278733SSatish Balay PetscLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */ 29581278733SSatish Balay PetscFunctionReturn(0); 29681278733SSatish Balay } 297