xref: /petsc/src/mat/impls/sbaij/seq/sbaijfact8.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 5 by 5 Using natural ordering
6*81278733SSatish Balay */
7*81278733SSatish Balay #undef __FUNC__
8*81278733SSatish Balay #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering"
9*81278733SSatish Balay int MatCholeskyFactorNumeric_SeqSBAIJ_5_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,*rtmp,*rtmp_ptr;
17*81278733SSatish Balay 
18*81278733SSatish Balay   PetscFunctionBegin;
19*81278733SSatish Balay 
20*81278733SSatish Balay   /* initialization */
21*81278733SSatish Balay   ierr = PetscMalloc(25*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
22*81278733SSatish Balay   ierr = PetscMemzero(rtmp,25*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(50*sizeof(MatScalar),&dk);CHKERRQ(ierr);
29*81278733SSatish Balay   uik  = dk + 25;
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*25;
40*81278733SSatish Balay       for (j = jmin; j < jmax; j++){
41*81278733SSatish Balay         vj = aj[j];         /* block col. index */
42*81278733SSatish Balay         rtmp_ptr = rtmp + vj*25;
43*81278733SSatish Balay         for (i=0; i<25; i++) *rtmp_ptr++ = *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,rtmp+k*25,25*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*25;
59*81278733SSatish Balay       u    = ba + ili*25;
60*81278733SSatish Balay 
61*81278733SSatish Balay       uik[0] = -(d[0]*u[0] + d[5]*u[1] + d[10]*u[2] + d[15]*u[3] + d[20]*u[4]);
62*81278733SSatish Balay       uik[1] = -(d[1]*u[0] + d[6]*u[1] + d[11]*u[2] + d[16]*u[3] + d[21]*u[4]);
63*81278733SSatish Balay       uik[2] = -(d[2]*u[0] + d[7]*u[1] + d[12]*u[2] + d[17]*u[3] + d[22]*u[4]);
64*81278733SSatish Balay       uik[3] = -(d[3]*u[0] + d[8]*u[1] + d[13]*u[2] + d[18]*u[3] + d[23]*u[4]);
65*81278733SSatish Balay       uik[4] = -(d[4]*u[0] + d[9]*u[1] + d[14]*u[2] + d[19]*u[3] + d[24]*u[4]);
66*81278733SSatish Balay 
67*81278733SSatish Balay       uik[5] = -(d[0]*u[5] + d[5]*u[6] + d[10]*u[7] + d[15]*u[8] + d[20]*u[9]);
68*81278733SSatish Balay       uik[6] = -(d[1]*u[5] + d[6]*u[6] + d[11]*u[7] + d[16]*u[8] + d[21]*u[9]);
69*81278733SSatish Balay       uik[7] = -(d[2]*u[5] + d[7]*u[6] + d[12]*u[7] + d[17]*u[8] + d[22]*u[9]);
70*81278733SSatish Balay       uik[8] = -(d[3]*u[5] + d[8]*u[6] + d[13]*u[7] + d[18]*u[8] + d[23]*u[9]);
71*81278733SSatish Balay       uik[9] = -(d[4]*u[5] + d[9]*u[6] + d[14]*u[7] + d[19]*u[8] + d[24]*u[9]);
72*81278733SSatish Balay 
73*81278733SSatish Balay       uik[10]= -(d[0]*u[10] + d[5]*u[11] + d[10]*u[12] + d[15]*u[13] + d[20]*u[14]);
74*81278733SSatish Balay       uik[11]= -(d[1]*u[10] + d[6]*u[11] + d[11]*u[12] + d[16]*u[13] + d[21]*u[14]);
75*81278733SSatish Balay       uik[12]= -(d[2]*u[10] + d[7]*u[11] + d[12]*u[12] + d[17]*u[13] + d[22]*u[14]);
76*81278733SSatish Balay       uik[13]= -(d[3]*u[10] + d[8]*u[11] + d[13]*u[12] + d[18]*u[13] + d[23]*u[14]);
77*81278733SSatish Balay       uik[14]= -(d[4]*u[10] + d[9]*u[11] + d[14]*u[12] + d[19]*u[13] + d[24]*u[14]);
78*81278733SSatish Balay 
79*81278733SSatish Balay       uik[15]= -(d[0]*u[15] + d[5]*u[16] + d[10]*u[17] + d[15]*u[18] + d[20]*u[19]);
80*81278733SSatish Balay       uik[16]= -(d[1]*u[15] + d[6]*u[16] + d[11]*u[17] + d[16]*u[18] + d[21]*u[19]);
81*81278733SSatish Balay       uik[17]= -(d[2]*u[15] + d[7]*u[16] + d[12]*u[17] + d[17]*u[18] + d[22]*u[19]);
82*81278733SSatish Balay       uik[18]= -(d[3]*u[15] + d[8]*u[16] + d[13]*u[17] + d[18]*u[18] + d[23]*u[19]);
83*81278733SSatish Balay       uik[19]= -(d[4]*u[15] + d[9]*u[16] + d[14]*u[17] + d[19]*u[18] + d[24]*u[19]);
84*81278733SSatish Balay 
85*81278733SSatish Balay       uik[20]= -(d[0]*u[20] + d[5]*u[21] + d[10]*u[22] + d[15]*u[23] + d[20]*u[24]);
86*81278733SSatish Balay       uik[21]= -(d[1]*u[20] + d[6]*u[21] + d[11]*u[22] + d[16]*u[23] + d[21]*u[24]);
87*81278733SSatish Balay       uik[22]= -(d[2]*u[20] + d[7]*u[21] + d[12]*u[22] + d[17]*u[23] + d[22]*u[24]);
88*81278733SSatish Balay       uik[23]= -(d[3]*u[20] + d[8]*u[21] + d[13]*u[22] + d[18]*u[23] + d[23]*u[24]);
89*81278733SSatish Balay       uik[24]= -(d[4]*u[20] + d[9]*u[21] + d[14]*u[22] + d[19]*u[23] + d[24]*u[24]);
90*81278733SSatish Balay 
91*81278733SSatish Balay 
92*81278733SSatish Balay       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
93*81278733SSatish Balay       dk[0] +=  uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4];
94*81278733SSatish Balay       dk[1] +=  uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4];
95*81278733SSatish Balay       dk[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4];
96*81278733SSatish Balay       dk[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4];
97*81278733SSatish Balay       dk[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4];
98*81278733SSatish Balay 
99*81278733SSatish Balay       dk[5] +=  uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9];
100*81278733SSatish Balay       dk[6] +=  uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9];
101*81278733SSatish Balay       dk[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9];
102*81278733SSatish Balay       dk[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9];
103*81278733SSatish Balay       dk[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9];
104*81278733SSatish Balay 
105*81278733SSatish Balay       dk[10] +=  uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14];
106*81278733SSatish Balay       dk[11] +=  uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14];
107*81278733SSatish Balay       dk[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14];
108*81278733SSatish Balay       dk[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14];
109*81278733SSatish Balay       dk[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14];
110*81278733SSatish Balay 
111*81278733SSatish Balay       dk[15] +=  uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19];
112*81278733SSatish Balay       dk[16] +=  uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19];
113*81278733SSatish Balay       dk[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19];
114*81278733SSatish Balay       dk[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19];
115*81278733SSatish Balay       dk[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19];
116*81278733SSatish Balay 
117*81278733SSatish Balay       dk[20] +=  uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24];
118*81278733SSatish Balay       dk[21] +=  uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24];
119*81278733SSatish Balay       dk[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24];
120*81278733SSatish Balay       dk[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24];
121*81278733SSatish Balay       dk[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24];
122*81278733SSatish Balay 
123*81278733SSatish Balay       /* update -U(i,k) */
124*81278733SSatish Balay       ierr = PetscMemcpy(ba+ili*25,uik,25*sizeof(MatScalar));CHKERRQ(ierr);
125*81278733SSatish Balay 
126*81278733SSatish Balay       /* add multiple of row i to k-th row ... */
127*81278733SSatish Balay       jmin = ili + 1; jmax = bi[i+1];
128*81278733SSatish Balay       if (jmin < jmax){
129*81278733SSatish Balay         for (j=jmin; j<jmax; j++) {
130*81278733SSatish Balay           /* rtmp += -U(i,k)^T * U_bar(i,j) */
131*81278733SSatish Balay           rtmp_ptr = rtmp + bj[j]*25;
132*81278733SSatish Balay           u = ba + j*25;
133*81278733SSatish Balay           rtmp_ptr[0] +=  uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3] + uik[4]*u[4];
134*81278733SSatish Balay           rtmp_ptr[1] +=  uik[5]*u[0] + uik[6]*u[1] + uik[7]*u[2] + uik[8]*u[3] + uik[9]*u[4];
135*81278733SSatish Balay           rtmp_ptr[2] += uik[10]*u[0]+ uik[11]*u[1]+ uik[12]*u[2]+ uik[13]*u[3]+ uik[14]*u[4];
136*81278733SSatish Balay           rtmp_ptr[3] += uik[15]*u[0]+ uik[16]*u[1]+ uik[17]*u[2]+ uik[18]*u[3]+ uik[19]*u[4];
137*81278733SSatish Balay           rtmp_ptr[4] += uik[20]*u[0]+ uik[21]*u[1]+ uik[22]*u[2]+ uik[23]*u[3]+ uik[24]*u[4];
138*81278733SSatish Balay 
139*81278733SSatish Balay           rtmp_ptr[5] +=  uik[0]*u[5] + uik[1]*u[6] + uik[2]*u[7] + uik[3]*u[8] + uik[4]*u[9];
140*81278733SSatish Balay           rtmp_ptr[6] +=  uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7] + uik[8]*u[8] + uik[9]*u[9];
141*81278733SSatish Balay           rtmp_ptr[7] += uik[10]*u[5]+ uik[11]*u[6]+ uik[12]*u[7]+ uik[13]*u[8]+ uik[14]*u[9];
142*81278733SSatish Balay           rtmp_ptr[8] += uik[15]*u[5]+ uik[16]*u[6]+ uik[17]*u[7]+ uik[18]*u[8]+ uik[19]*u[9];
143*81278733SSatish Balay           rtmp_ptr[9] += uik[20]*u[5]+ uik[21]*u[6]+ uik[22]*u[7]+ uik[23]*u[8]+ uik[24]*u[9];
144*81278733SSatish Balay 
145*81278733SSatish Balay           rtmp_ptr[10] +=  uik[0]*u[10] + uik[1]*u[11] + uik[2]*u[12] + uik[3]*u[13] + uik[4]*u[14];
146*81278733SSatish Balay           rtmp_ptr[11] +=  uik[5]*u[10] + uik[6]*u[11] + uik[7]*u[12] + uik[8]*u[13] + uik[9]*u[14];
147*81278733SSatish Balay           rtmp_ptr[12] += uik[10]*u[10]+ uik[11]*u[11]+ uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14];
148*81278733SSatish Balay           rtmp_ptr[13] += uik[15]*u[10]+ uik[16]*u[11]+ uik[17]*u[12]+ uik[18]*u[13]+ uik[19]*u[14];
149*81278733SSatish Balay           rtmp_ptr[14] += uik[20]*u[10]+ uik[21]*u[11]+ uik[22]*u[12]+ uik[23]*u[13]+ uik[24]*u[14];
150*81278733SSatish Balay 
151*81278733SSatish Balay           rtmp_ptr[15] +=  uik[0]*u[15] + uik[1]*u[16] + uik[2]*u[17] + uik[3]*u[18] + uik[4]*u[19];
152*81278733SSatish Balay           rtmp_ptr[16] +=  uik[5]*u[15] + uik[6]*u[16] + uik[7]*u[17] + uik[8]*u[18] + uik[9]*u[19];
153*81278733SSatish Balay           rtmp_ptr[17] += uik[10]*u[15]+ uik[11]*u[16]+ uik[12]*u[17]+ uik[13]*u[18]+ uik[14]*u[19];
154*81278733SSatish Balay           rtmp_ptr[18] += uik[15]*u[15]+ uik[16]*u[16]+ uik[17]*u[17]+ uik[18]*u[18]+ uik[19]*u[19];
155*81278733SSatish Balay           rtmp_ptr[19] += uik[20]*u[15]+ uik[21]*u[16]+ uik[22]*u[17]+ uik[23]*u[18]+ uik[24]*u[19];
156*81278733SSatish Balay 
157*81278733SSatish Balay           rtmp_ptr[20] +=  uik[0]*u[20] + uik[1]*u[21] + uik[2]*u[22] + uik[3]*u[23] + uik[4]*u[24];
158*81278733SSatish Balay           rtmp_ptr[21] +=  uik[5]*u[20] + uik[6]*u[21] + uik[7]*u[22] + uik[8]*u[23] + uik[9]*u[24];
159*81278733SSatish Balay           rtmp_ptr[22] += uik[10]*u[20]+ uik[11]*u[21]+ uik[12]*u[22]+ uik[13]*u[23]+ uik[14]*u[24];
160*81278733SSatish Balay           rtmp_ptr[23] += uik[15]*u[20]+ uik[16]*u[21]+ uik[17]*u[22]+ uik[18]*u[23]+ uik[19]*u[24];
161*81278733SSatish Balay           rtmp_ptr[24] += uik[20]*u[20]+ uik[21]*u[21]+ uik[22]*u[22]+ uik[23]*u[23]+ uik[24]*u[24];
162*81278733SSatish Balay         }
163*81278733SSatish Balay 
164*81278733SSatish Balay         /* ... add i to row list for next nonzero entry */
165*81278733SSatish Balay         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
166*81278733SSatish Balay         j     = bj[jmin];
167*81278733SSatish Balay         jl[i] = jl[j]; jl[j] = i; /* update jl */
168*81278733SSatish Balay       }
169*81278733SSatish Balay       i = nexti;
170*81278733SSatish Balay     }
171*81278733SSatish Balay 
172*81278733SSatish Balay     /* save nonzero entries in k-th row of U ... */
173*81278733SSatish Balay 
174*81278733SSatish Balay     /* invert diagonal block */
175*81278733SSatish Balay     d = ba+k*25;
176*81278733SSatish Balay     ierr = PetscMemcpy(d,dk,25*sizeof(MatScalar));CHKERRQ(ierr);
177*81278733SSatish Balay     ierr = Kernel_A_gets_inverse_A_5(d);CHKERRQ(ierr);
178*81278733SSatish Balay 
179*81278733SSatish Balay     jmin = bi[k]; jmax = bi[k+1];
180*81278733SSatish Balay     if (jmin < jmax) {
181*81278733SSatish Balay       for (j=jmin; j<jmax; j++){
182*81278733SSatish Balay          vj = bj[j];           /* block col. index of U */
183*81278733SSatish Balay          u   = ba + j*25;
184*81278733SSatish Balay          rtmp_ptr = rtmp + vj*25;
185*81278733SSatish Balay          for (k1=0; k1<25; k1++){
186*81278733SSatish Balay            *u++        = *rtmp_ptr;
187*81278733SSatish Balay            *rtmp_ptr++ = 0.0;
188*81278733SSatish Balay          }
189*81278733SSatish Balay       }
190*81278733SSatish Balay 
191*81278733SSatish Balay       /* ... add k to row list for first nonzero entry in k-th row */
192*81278733SSatish Balay       il[k] = jmin;
193*81278733SSatish Balay       i     = bj[jmin];
194*81278733SSatish Balay       jl[k] = jl[i]; jl[i] = k;
195*81278733SSatish Balay     }
196*81278733SSatish Balay   }
197*81278733SSatish Balay 
198*81278733SSatish Balay   ierr = PetscFree(rtmp);CHKERRQ(ierr);
199*81278733SSatish Balay   ierr = PetscFree(il);CHKERRQ(ierr);
200*81278733SSatish Balay   ierr = PetscFree(dk);CHKERRQ(ierr);
201*81278733SSatish Balay 
202*81278733SSatish Balay   C->factor    = FACTOR_CHOLESKY;
203*81278733SSatish Balay   C->assembled = PETSC_TRUE;
204*81278733SSatish Balay   C->preallocated = PETSC_TRUE;
205*81278733SSatish Balay   PetscLogFlops(1.3333*125*b->mbs); /* from inverting diagonal blocks */
206*81278733SSatish Balay   PetscFunctionReturn(0);
207*81278733SSatish Balay }
208