xref: /petsc/src/mat/impls/sbaij/seq/sbaijfact10.c (revision 81278733cfc5391c951b33ea559a85cfc96ebb21)
1*81278733SSatish Balay #include "sbaij.h"
2*81278733SSatish Balay #include "src/inline/ilu.h"
3*81278733SSatish Balay 
4*81278733SSatish Balay /*
5*81278733SSatish Balay       Version for when blocks are 6 by 6 Using natural ordering
6*81278733SSatish Balay */
7*81278733SSatish Balay #undef __FUNC__
8*81278733SSatish Balay #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering"
9*81278733SSatish Balay int MatCholeskyFactorNumeric_SeqSBAIJ_6_NaturalOrdering(Mat A,Mat *B)
10*81278733SSatish Balay {
11*81278733SSatish Balay   Mat                C = *B;
12*81278733SSatish Balay   Mat_SeqSBAIJ       *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
13*81278733SSatish Balay   int                ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j;
14*81278733SSatish Balay   int                *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili;
15*81278733SSatish Balay   MatScalar          *ba = b->a,*aa,*ap,*dk,*uik;
16*81278733SSatish Balay   MatScalar          *u,*d,*w,*wp;
17*81278733SSatish Balay 
18*81278733SSatish Balay   PetscFunctionBegin;
19*81278733SSatish Balay 
20*81278733SSatish Balay   /* initialization */
21*81278733SSatish Balay   ierr = PetscMalloc(36*mbs*sizeof(MatScalar),&w);CHKERRQ(ierr);
22*81278733SSatish Balay   ierr = PetscMemzero(w,36*mbs*sizeof(MatScalar));CHKERRQ(ierr);
23*81278733SSatish Balay   ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr);
24*81278733SSatish Balay   jl = il + mbs;
25*81278733SSatish Balay   for (i=0; i<mbs; i++) {
26*81278733SSatish Balay     jl[i] = mbs; il[0] = 0;
27*81278733SSatish Balay   }
28*81278733SSatish Balay   ierr = PetscMalloc(72*sizeof(MatScalar),&dk);CHKERRQ(ierr);
29*81278733SSatish Balay   uik   = dk + 36;
30*81278733SSatish Balay 
31*81278733SSatish Balay   ai = a->i; aj = a->j; aa = a->a;
32*81278733SSatish Balay 
33*81278733SSatish Balay   /* for each row k */
34*81278733SSatish Balay   for (k = 0; k<mbs; k++){
35*81278733SSatish Balay 
36*81278733SSatish Balay     /*initialize k-th row with elements nonzero in row k of A */
37*81278733SSatish Balay     jmin = ai[k]; jmax = ai[k+1];
38*81278733SSatish Balay     if (jmin < jmax) {
39*81278733SSatish Balay       ap = aa + jmin*36;
40*81278733SSatish Balay       for (j = jmin; j < jmax; j++){
41*81278733SSatish Balay         vj = aj[j];         /* block col. index */
42*81278733SSatish Balay         wp = w + vj*36;
43*81278733SSatish Balay         for (i=0; i<36; i++) *wp++ = *ap++;
44*81278733SSatish Balay       }
45*81278733SSatish Balay     }
46*81278733SSatish Balay 
47*81278733SSatish Balay     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
48*81278733SSatish Balay     ierr = PetscMemcpy(dk,w+k*36,36*sizeof(MatScalar));CHKERRQ(ierr);
49*81278733SSatish Balay     i = jl[k]; /* first row to be added to k_th row  */
50*81278733SSatish Balay 
51*81278733SSatish Balay     while (i < mbs){
52*81278733SSatish Balay       nexti = jl[i]; /* next row to be added to k_th row */
53*81278733SSatish Balay 
54*81278733SSatish Balay       /* compute multiplier */
55*81278733SSatish Balay       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
56*81278733SSatish Balay 
57*81278733SSatish Balay       /* uik = -inv(Di)*U_bar(i,k) */
58*81278733SSatish Balay       d = ba + i*36;
59*81278733SSatish Balay       u    = ba + ili*36;
60*81278733SSatish Balay 
61*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]);
62*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]);
63*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]);
64*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]);
65*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]);
66*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]);
67*81278733SSatish Balay 
68*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]);
69*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]);
70*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]);
71*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]);
72*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]);
73*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]);
74*81278733SSatish Balay 
75*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]);
76*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]);
77*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]);
78*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]);
79*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]);
80*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]);
81*81278733SSatish Balay 
82*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]);
83*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]);
84*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]);
85*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]);
86*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]);
87*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]);
88*81278733SSatish Balay 
89*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]);
90*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]);
91*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]);
92*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]);
93*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]);
94*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]);
95*81278733SSatish Balay 
96*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]);
97*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]);
98*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]);
99*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]);
100*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]);
101*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]);
102*81278733SSatish Balay 
103*81278733SSatish Balay       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
104*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];
105*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];
106*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];
107*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];
108*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];
109*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];
110*81278733SSatish Balay 
111*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];
112*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];
113*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];
114*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];
115*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];
116*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];
117*81278733SSatish Balay 
118*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];
119*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];
120*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];
121*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];
122*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];
123*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];
124*81278733SSatish Balay 
125*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];
126*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];
127*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];
128*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];
129*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];
130*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];
131*81278733SSatish Balay 
132*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];
133*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];
134*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];
135*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];
136*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];
137*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];
138*81278733SSatish Balay 
139*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];
140*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];
141*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];
142*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];
143*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];
144*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];
145*81278733SSatish Balay 
146*81278733SSatish Balay       /* update -U(i,k) */
147*81278733SSatish Balay       ierr = PetscMemcpy(ba+ili*36,uik,36*sizeof(MatScalar));CHKERRQ(ierr);
148*81278733SSatish Balay 
149*81278733SSatish Balay       /* add multiple of row i to k-th row ... */
150*81278733SSatish Balay       jmin = ili + 1; jmax = bi[i+1];
151*81278733SSatish Balay       if (jmin < jmax){
152*81278733SSatish Balay         for (j=jmin; j<jmax; j++) {
153*81278733SSatish Balay           /* w += -U(i,k)^T * U_bar(i,j) */
154*81278733SSatish Balay           wp = w + bj[j]*36;
155*81278733SSatish Balay           u = ba + j*36;
156*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];
157*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];
158*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];
159*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];
160*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];
161*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];
162*81278733SSatish Balay 
163*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];
164*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];
165*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];
166*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];
167*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];
168*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];
169*81278733SSatish Balay 
170*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];
171*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];
172*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];
173*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];
174*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];
175*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];
176*81278733SSatish Balay 
177*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];
178*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];
179*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];
180*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];
181*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];
182*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];
183*81278733SSatish Balay 
184*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];
185*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];
186*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];
187*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];
188*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];
189*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];
190*81278733SSatish Balay 
191*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];
192*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];
193*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];
194*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];
195*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];
196*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];
197*81278733SSatish Balay         }
198*81278733SSatish Balay 
199*81278733SSatish Balay         /* ... add i to row list for next nonzero entry */
200*81278733SSatish Balay         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
201*81278733SSatish Balay         j     = bj[jmin];
202*81278733SSatish Balay         jl[i] = jl[j]; jl[j] = i; /* update jl */
203*81278733SSatish Balay       }
204*81278733SSatish Balay       i = nexti;
205*81278733SSatish Balay     }
206*81278733SSatish Balay 
207*81278733SSatish Balay     /* save nonzero entries in k-th row of U ... */
208*81278733SSatish Balay 
209*81278733SSatish Balay     /* invert diagonal block */
210*81278733SSatish Balay     d = ba+k*36;
211*81278733SSatish Balay     ierr = PetscMemcpy(d,dk,36*sizeof(MatScalar));CHKERRQ(ierr);
212*81278733SSatish Balay     ierr = Kernel_A_gets_inverse_A_6(d);CHKERRQ(ierr);
213*81278733SSatish Balay 
214*81278733SSatish Balay     jmin = bi[k]; jmax = bi[k+1];
215*81278733SSatish Balay     if (jmin < jmax) {
216*81278733SSatish Balay       for (j=jmin; j<jmax; j++){
217*81278733SSatish Balay          vj = bj[j];           /* block col. index of U */
218*81278733SSatish Balay          u   = ba + j*36;
219*81278733SSatish Balay          wp = w + vj*36;
220*81278733SSatish Balay          for (k1=0; k1<36; k1++){
221*81278733SSatish Balay            *u++        = *wp;
222*81278733SSatish Balay            *wp++ = 0.0;
223*81278733SSatish Balay          }
224*81278733SSatish Balay       }
225*81278733SSatish Balay 
226*81278733SSatish Balay       /* ... add k to row list for first nonzero entry in k-th row */
227*81278733SSatish Balay       il[k] = jmin;
228*81278733SSatish Balay       i     = bj[jmin];
229*81278733SSatish Balay       jl[k] = jl[i]; jl[i] = k;
230*81278733SSatish Balay     }
231*81278733SSatish Balay   }
232*81278733SSatish Balay 
233*81278733SSatish Balay   ierr = PetscFree(w);CHKERRQ(ierr);
234*81278733SSatish Balay   ierr = PetscFree(il);CHKERRQ(ierr);
235*81278733SSatish Balay   ierr = PetscFree(dk);CHKERRQ(ierr);
236*81278733SSatish Balay 
237*81278733SSatish Balay   C->factor    = FACTOR_CHOLESKY;
238*81278733SSatish Balay   C->assembled = PETSC_TRUE;
239*81278733SSatish Balay   C->preallocated = PETSC_TRUE;
240*81278733SSatish Balay   PetscLogFlops(1.3333*216*b->mbs); /* from inverting diagonal blocks */
241*81278733SSatish Balay   PetscFunctionReturn(0);
242*81278733SSatish Balay }
243