1*81278733SSatish Balay #include "sbaij.h" 2*81278733SSatish Balay #include "src/inline/ilu.h" 3*81278733SSatish Balay 4*81278733SSatish Balay /* Version for when blocks are 6 by 6 */ 5*81278733SSatish Balay #undef __FUNC__ 6*81278733SSatish Balay #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6" 7*81278733SSatish Balay int MatCholeskyFactorNumeric_SeqSBAIJ_6(Mat A,Mat *B) 8*81278733SSatish Balay { 9*81278733SSatish Balay Mat C = *B; 10*81278733SSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 11*81278733SSatish Balay IS perm = b->row; 12*81278733SSatish Balay int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 13*81278733SSatish Balay int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 14*81278733SSatish Balay MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 15*81278733SSatish Balay MatScalar *u,*d,*w,*wp; 16*81278733SSatish Balay 17*81278733SSatish Balay PetscFunctionBegin; 18*81278733SSatish Balay /* initialization */ 19*81278733SSatish Balay ierr = PetscMalloc(36*mbs*sizeof(MatScalar),&w);CHKERRQ(ierr); 20*81278733SSatish Balay ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr); 21*81278733SSatish Balay ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr); 22*81278733SSatish Balay jl = il + mbs; 23*81278733SSatish Balay for (i=0; i<mbs; i++) { 24*81278733SSatish Balay jl[i] = mbs; il[0] = 0; 25*81278733SSatish Balay } 26*81278733SSatish Balay ierr = PetscMalloc(72*sizeof(MatScalar),&dk);CHKERRQ(ierr); 27*81278733SSatish Balay uik = dk + 36; 28*81278733SSatish Balay ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 29*81278733SSatish Balay 30*81278733SSatish Balay /* check permutation */ 31*81278733SSatish Balay if (!a->permute){ 32*81278733SSatish Balay ai = a->i; aj = a->j; aa = a->a; 33*81278733SSatish Balay } else { 34*81278733SSatish Balay ai = a->inew; aj = a->jnew; 35*81278733SSatish Balay ierr = PetscMalloc(36*ai[mbs]*sizeof(MatScalar),&aa);CHKERRQ(ierr); 36*81278733SSatish Balay ierr = PetscMemcpy(aa,a->a,36*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 37*81278733SSatish Balay ierr = PetscMalloc(ai[mbs]*sizeof(int),&a2anew);CHKERRQ(ierr); 38*81278733SSatish Balay ierr = PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 39*81278733SSatish Balay 40*81278733SSatish Balay for (i=0; i<mbs; i++){ 41*81278733SSatish Balay jmin = ai[i]; jmax = ai[i+1]; 42*81278733SSatish Balay for (j=jmin; j<jmax; j++){ 43*81278733SSatish Balay while (a2anew[j] != j){ 44*81278733SSatish Balay k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 45*81278733SSatish Balay for (k1=0; k1<36; k1++){ 46*81278733SSatish Balay dk[k1] = aa[k*36+k1]; 47*81278733SSatish Balay aa[k*36+k1] = aa[j*36+k1]; 48*81278733SSatish Balay aa[j*36+k1] = dk[k1]; 49*81278733SSatish Balay } 50*81278733SSatish Balay } 51*81278733SSatish Balay /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 52*81278733SSatish Balay if (i > aj[j]){ 53*81278733SSatish Balay /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 54*81278733SSatish Balay ap = aa + j*36; /* ptr to the beginning of j-th block of aa */ 55*81278733SSatish Balay for (k=0; k<36; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 56*81278733SSatish Balay for (k=0; k<6; k++){ /* j-th block of aa <- dk^T */ 57*81278733SSatish Balay for (k1=0; k1<6; k1++) *ap++ = dk[k + 6*k1]; 58*81278733SSatish Balay } 59*81278733SSatish Balay } 60*81278733SSatish Balay } 61*81278733SSatish Balay } 62*81278733SSatish Balay ierr = PetscFree(a2anew);CHKERRQ(ierr); 63*81278733SSatish Balay } 64*81278733SSatish Balay 65*81278733SSatish Balay /* for each row k */ 66*81278733SSatish Balay for (k = 0; k<mbs; k++){ 67*81278733SSatish Balay 68*81278733SSatish Balay /*initialize k-th row with elements nonzero in row perm(k) of A */ 69*81278733SSatish Balay jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 70*81278733SSatish Balay if (jmin < jmax) { 71*81278733SSatish Balay ap = aa + jmin*36; 72*81278733SSatish Balay for (j = jmin; j < jmax; j++){ 73*81278733SSatish Balay vj = perm_ptr[aj[j]]; /* block col. index */ 74*81278733SSatish Balay wp = w + vj*36; 75*81278733SSatish Balay for (i=0; i<36; i++) *wp++ = *ap++; 76*81278733SSatish Balay } 77*81278733SSatish Balay } 78*81278733SSatish Balay 79*81278733SSatish Balay /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 80*81278733SSatish Balay ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr); 81*81278733SSatish Balay i = jl[k]; /* first row to be added to k_th row */ 82*81278733SSatish Balay 83*81278733SSatish Balay while (i < mbs){ 84*81278733SSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 85*81278733SSatish Balay 86*81278733SSatish Balay /* compute multiplier */ 87*81278733SSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 88*81278733SSatish Balay 89*81278733SSatish Balay /* uik = -inv(Di)*U_bar(i,k) */ 90*81278733SSatish Balay d = ba + i*36; 91*81278733SSatish Balay u = ba + ili*36; 92*81278733SSatish Balay 93*81278733SSatish 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]); 94*81278733SSatish 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]); 95*81278733SSatish 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]); 96*81278733SSatish 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]); 97*81278733SSatish 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]); 98*81278733SSatish 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]); 99*81278733SSatish Balay 100*81278733SSatish 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]); 101*81278733SSatish 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]); 102*81278733SSatish 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]); 103*81278733SSatish 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]); 104*81278733SSatish 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]); 105*81278733SSatish 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]); 106*81278733SSatish Balay 107*81278733SSatish 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]); 108*81278733SSatish 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]); 109*81278733SSatish 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]); 110*81278733SSatish 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]); 111*81278733SSatish 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]); 112*81278733SSatish 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]); 113*81278733SSatish Balay 114*81278733SSatish 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]); 115*81278733SSatish 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]); 116*81278733SSatish 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]); 117*81278733SSatish 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]); 118*81278733SSatish 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]); 119*81278733SSatish 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]); 120*81278733SSatish Balay 121*81278733SSatish 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]); 122*81278733SSatish 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]); 123*81278733SSatish 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]); 124*81278733SSatish 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]); 125*81278733SSatish 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]); 126*81278733SSatish 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]); 127*81278733SSatish Balay 128*81278733SSatish 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]); 129*81278733SSatish 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]); 130*81278733SSatish 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]); 131*81278733SSatish 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]); 132*81278733SSatish 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]); 133*81278733SSatish 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]); 134*81278733SSatish Balay 135*81278733SSatish Balay /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 136*81278733SSatish 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]; 137*81278733SSatish 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]; 138*81278733SSatish 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]; 139*81278733SSatish 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]; 140*81278733SSatish 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]; 141*81278733SSatish 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]; 142*81278733SSatish Balay 143*81278733SSatish 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]; 144*81278733SSatish 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]; 145*81278733SSatish 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]; 146*81278733SSatish 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]; 147*81278733SSatish 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]; 148*81278733SSatish 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]; 149*81278733SSatish Balay 150*81278733SSatish 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]; 151*81278733SSatish 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]; 152*81278733SSatish 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]; 153*81278733SSatish 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]; 154*81278733SSatish 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]; 155*81278733SSatish 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]; 156*81278733SSatish Balay 157*81278733SSatish 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]; 158*81278733SSatish 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]; 159*81278733SSatish 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]; 160*81278733SSatish 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]; 161*81278733SSatish 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]; 162*81278733SSatish 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]; 163*81278733SSatish Balay 164*81278733SSatish 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]; 165*81278733SSatish 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]; 166*81278733SSatish 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]; 167*81278733SSatish 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]; 168*81278733SSatish 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]; 169*81278733SSatish 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]; 170*81278733SSatish Balay 171*81278733SSatish 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]; 172*81278733SSatish 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]; 173*81278733SSatish 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]; 174*81278733SSatish 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]; 175*81278733SSatish 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]; 176*81278733SSatish 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]; 177*81278733SSatish Balay 178*81278733SSatish Balay /* update -U(i,k) */ 179*81278733SSatish Balay ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr); 180*81278733SSatish Balay 181*81278733SSatish Balay /* add multiple of row i to k-th row ... */ 182*81278733SSatish Balay jmin = ili + 1; jmax = bi[i+1]; 183*81278733SSatish Balay if (jmin < jmax){ 184*81278733SSatish Balay for (j=jmin; j<jmax; j++) { 185*81278733SSatish Balay /* w += -U(i,k)^T * U_bar(i,j) */ 186*81278733SSatish Balay wp = w + bj[j]*36; 187*81278733SSatish Balay u = ba + j*36; 188*81278733SSatish 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]; 189*81278733SSatish 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]; 190*81278733SSatish 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]; 191*81278733SSatish 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]; 192*81278733SSatish 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]; 193*81278733SSatish 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]; 194*81278733SSatish Balay 195*81278733SSatish 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]; 196*81278733SSatish 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]; 197*81278733SSatish 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]; 198*81278733SSatish 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]; 199*81278733SSatish 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]; 200*81278733SSatish 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]; 201*81278733SSatish Balay 202*81278733SSatish 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]; 203*81278733SSatish 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]; 204*81278733SSatish 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]; 205*81278733SSatish 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]; 206*81278733SSatish 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]; 207*81278733SSatish 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]; 208*81278733SSatish Balay 209*81278733SSatish 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]; 210*81278733SSatish 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]; 211*81278733SSatish 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]; 212*81278733SSatish 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]; 213*81278733SSatish 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]; 214*81278733SSatish 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]; 215*81278733SSatish Balay 216*81278733SSatish 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]; 217*81278733SSatish 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]; 218*81278733SSatish 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]; 219*81278733SSatish 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]; 220*81278733SSatish 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]; 221*81278733SSatish 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]; 222*81278733SSatish Balay 223*81278733SSatish 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]; 224*81278733SSatish 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]; 225*81278733SSatish 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]; 226*81278733SSatish 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]; 227*81278733SSatish 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]; 228*81278733SSatish 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]; 229*81278733SSatish Balay } 230*81278733SSatish Balay 231*81278733SSatish Balay /* ... add i to row list for next nonzero entry */ 232*81278733SSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 233*81278733SSatish Balay j = bj[jmin]; 234*81278733SSatish Balay jl[i] = jl[j]; jl[j] = i; /* update jl */ 235*81278733SSatish Balay } 236*81278733SSatish Balay i = nexti; 237*81278733SSatish Balay } 238*81278733SSatish Balay 239*81278733SSatish Balay /* save nonzero entries in k-th row of U ... */ 240*81278733SSatish Balay 241*81278733SSatish Balay /* invert diagonal block */ 242*81278733SSatish Balay d = ba+k*36; 243*81278733SSatish Balay ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr); 244*81278733SSatish Balay ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr); 245*81278733SSatish Balay 246*81278733SSatish Balay jmin = bi[k]; jmax = bi[k+1]; 247*81278733SSatish Balay if (jmin < jmax) { 248*81278733SSatish Balay for (j=jmin; j<jmax; j++){ 249*81278733SSatish Balay vj = bj[j]; /* block col. index of U */ 250*81278733SSatish Balay u = ba + j*36; 251*81278733SSatish Balay wp = w + vj*36; 252*81278733SSatish Balay for (k1=0; k1<36; k1++){ 253*81278733SSatish Balay *u++ = *wp; 254*81278733SSatish Balay *wp++ = 0.0; 255*81278733SSatish Balay } 256*81278733SSatish Balay } 257*81278733SSatish Balay 258*81278733SSatish Balay /* ... add k to row list for first nonzero entry in k-th row */ 259*81278733SSatish Balay il[k] = jmin; 260*81278733SSatish Balay i = bj[jmin]; 261*81278733SSatish Balay jl[k] = jl[i]; jl[i] = k; 262*81278733SSatish Balay } 263*81278733SSatish Balay } 264*81278733SSatish Balay 265*81278733SSatish Balay ierr = PetscFree(w);CHKERRQ(ierr); 266*81278733SSatish Balay ierr = PetscFree(il);CHKERRQ(ierr); 267*81278733SSatish Balay ierr = PetscFree(dk);CHKERRQ(ierr); 268*81278733SSatish Balay if (a->permute){ 269*81278733SSatish Balay ierr = PetscFree(aa);CHKERRQ(ierr); 270*81278733SSatish Balay } 271*81278733SSatish Balay 272*81278733SSatish Balay ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 273*81278733SSatish Balay C->factor = FACTOR_CHOLESKY; 274*81278733SSatish Balay C->assembled = PETSC_TRUE; 275*81278733SSatish Balay C->preallocated = PETSC_TRUE; 276*81278733SSatish Balay PetscLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */ 277*81278733SSatish Balay PetscFunctionReturn(0); 278*81278733SSatish Balay } 279