xref: /petsc/src/mat/impls/baij/seq/dgefa7.c (revision 84643e36d7f84da5417ef27d7e86f8d8c86095d3)
1*84643e36SBarry Smith #ifdef PETSC_RCS_HEADER
2*84643e36SBarry Smith static char vcid[] = "$Id: dgefa7.c,v 1.7 1998/12/24 04:10:08 bsmith Exp bsmith $";
3*84643e36SBarry Smith #endif
4*84643e36SBarry Smith /*
5*84643e36SBarry Smith       Inverts 7 by 7 matrix using partial pivoting.
6*84643e36SBarry Smith 
7*84643e36SBarry Smith        Used by the sparse factorization routines in
8*84643e36SBarry Smith      src/mat/impls/baij/seq and src/mat/impls/bdiag/seq
9*84643e36SBarry Smith 
10*84643e36SBarry Smith        See also src/inline/ilu.h
11*84643e36SBarry Smith 
12*84643e36SBarry Smith        This is a combination of the Linpack routines
13*84643e36SBarry Smith     dgefa() and dgedi() specialized for a size of 7.
14*84643e36SBarry Smith 
15*84643e36SBarry Smith */
16*84643e36SBarry Smith #include "petsc.h"
17*84643e36SBarry Smith 
18*84643e36SBarry Smith #undef __FUNC__
19*84643e36SBarry Smith #define __FUNC__ "Kernel_A_gets_inverse_A_7"
20*84643e36SBarry Smith int Kernel_A_gets_inverse_A_7(MatScalar *a)
21*84643e36SBarry Smith {
22*84643e36SBarry Smith     int        i__2, i__3, kp1, j, k, l,ll,i,ipvt_l[7],*ipvt = ipvt_l-1,kb,k3;
23*84643e36SBarry Smith     int        k4,j3;
24*84643e36SBarry Smith     MatScalar  *aa,*ax,*ay,work_l[49],*work = work_l-1,stmp;
25*84643e36SBarry Smith     MatFloat   tmp,max;
26*84643e36SBarry Smith 
27*84643e36SBarry Smith /*     gaussian elimination with partial pivoting */
28*84643e36SBarry Smith 
29*84643e36SBarry Smith     PetscFunctionBegin;
30*84643e36SBarry Smith     /* Parameter adjustments */
31*84643e36SBarry Smith     a       -= 8;
32*84643e36SBarry Smith 
33*84643e36SBarry Smith     for (k = 1; k <= 6; ++k) {
34*84643e36SBarry Smith 	kp1 = k + 1;
35*84643e36SBarry Smith         k3  = 7*k;
36*84643e36SBarry Smith         k4  = k3 + k;
37*84643e36SBarry Smith /*        find l = pivot index */
38*84643e36SBarry Smith 
39*84643e36SBarry Smith 	i__2 = 7 - k;
40*84643e36SBarry Smith         aa = &a[k4];
41*84643e36SBarry Smith         max = PetscAbsScalar(aa[0]);
42*84643e36SBarry Smith         l = 1;
43*84643e36SBarry Smith         for ( ll=1; ll<i__2; ll++ ) {
44*84643e36SBarry Smith           tmp = PetscAbsScalar(aa[ll]);
45*84643e36SBarry Smith           if (tmp > max) { max = tmp; l = ll+1;}
46*84643e36SBarry Smith         }
47*84643e36SBarry Smith         l       += k - 1;
48*84643e36SBarry Smith 	ipvt[k] = l;
49*84643e36SBarry Smith 
50*84643e36SBarry Smith 	if (a[l + k3] == 0.) {
51*84643e36SBarry Smith 	  SETERRQ(k,0,"Zero pivot");
52*84643e36SBarry Smith 	}
53*84643e36SBarry Smith 
54*84643e36SBarry Smith /*           interchange if necessary */
55*84643e36SBarry Smith 
56*84643e36SBarry Smith 	if (l != k) {
57*84643e36SBarry Smith 	  stmp      = a[l + k3];
58*84643e36SBarry Smith 	  a[l + k3] = a[k4];
59*84643e36SBarry Smith 	  a[k4]     = stmp;
60*84643e36SBarry Smith         }
61*84643e36SBarry Smith 
62*84643e36SBarry Smith /*           compute multipliers */
63*84643e36SBarry Smith 
64*84643e36SBarry Smith 	stmp = -1. / a[k4];
65*84643e36SBarry Smith 	i__2 = 7 - k;
66*84643e36SBarry Smith         aa = &a[1 + k4];
67*84643e36SBarry Smith         for ( ll=0; ll<i__2; ll++ ) {
68*84643e36SBarry Smith           aa[ll] *= stmp;
69*84643e36SBarry Smith         }
70*84643e36SBarry Smith 
71*84643e36SBarry Smith /*           row elimination with column indexing */
72*84643e36SBarry Smith 
73*84643e36SBarry Smith 	ax = &a[k4+1];
74*84643e36SBarry Smith         for (j = kp1; j <= 7; ++j) {
75*84643e36SBarry Smith             j3   = 7*j;
76*84643e36SBarry Smith 	    stmp = a[l + j3];
77*84643e36SBarry Smith 	    if (l != k) {
78*84643e36SBarry Smith 	      a[l + j3] = a[k + j3];
79*84643e36SBarry Smith 	      a[k + j3] = stmp;
80*84643e36SBarry Smith             }
81*84643e36SBarry Smith 
82*84643e36SBarry Smith 	    i__3 = 7 - k;
83*84643e36SBarry Smith             ay = &a[1+k+j3];
84*84643e36SBarry Smith             for ( ll=0; ll<i__3; ll++ ) {
85*84643e36SBarry Smith               ay[ll] += stmp*ax[ll];
86*84643e36SBarry Smith             }
87*84643e36SBarry Smith 	}
88*84643e36SBarry Smith     }
89*84643e36SBarry Smith     ipvt[7] = 7;
90*84643e36SBarry Smith     if (a[56] == 0.) {
91*84643e36SBarry Smith 	SETERRQ(3,0,"Zero pivot,final row");
92*84643e36SBarry Smith     }
93*84643e36SBarry Smith 
94*84643e36SBarry Smith     /*
95*84643e36SBarry Smith          Now form the inverse
96*84643e36SBarry Smith     */
97*84643e36SBarry Smith 
98*84643e36SBarry Smith    /*     compute inverse(u) */
99*84643e36SBarry Smith 
100*84643e36SBarry Smith     for (k = 1; k <= 7; ++k) {
101*84643e36SBarry Smith         k3    = 7*k;
102*84643e36SBarry Smith         k4    = k3 + k;
103*84643e36SBarry Smith 	a[k4] = 1.0 / a[k4];
104*84643e36SBarry Smith 	stmp  = -a[k4];
105*84643e36SBarry Smith 	i__2  = k - 1;
106*84643e36SBarry Smith         aa    = &a[k3 + 1];
107*84643e36SBarry Smith         for ( ll=0; ll<i__2; ll++ ) aa[ll] *= stmp;
108*84643e36SBarry Smith 	kp1 = k + 1;
109*84643e36SBarry Smith 	if (7 < kp1) continue;
110*84643e36SBarry Smith         ax = aa;
111*84643e36SBarry Smith         for (j = kp1; j <= 7; ++j) {
112*84643e36SBarry Smith             j3        = 7*j;
113*84643e36SBarry Smith 	    stmp      = a[k + j3];
114*84643e36SBarry Smith 	    a[k + j3] = 0.0;
115*84643e36SBarry Smith             ay        = &a[j3 + 1];
116*84643e36SBarry Smith             for ( ll=0; ll<k; ll++ ) {
117*84643e36SBarry Smith               ay[ll] += stmp*ax[ll];
118*84643e36SBarry Smith             }
119*84643e36SBarry Smith 	}
120*84643e36SBarry Smith     }
121*84643e36SBarry Smith 
122*84643e36SBarry Smith    /*    form inverse(u)*inverse(l) */
123*84643e36SBarry Smith 
124*84643e36SBarry Smith     for (kb = 1; kb <= 6; ++kb) {
125*84643e36SBarry Smith 	k   = 7 - kb;
126*84643e36SBarry Smith         k3  = 7*k;
127*84643e36SBarry Smith 	kp1 = k + 1;
128*84643e36SBarry Smith         aa  = a + k3;
129*84643e36SBarry Smith 	for (i = kp1; i <= 7; ++i) {
130*84643e36SBarry Smith             work_l[i-1] = aa[i];
131*84643e36SBarry Smith             /* work[i] = aa[i]; Fix for -O3 error on Origin 2000 */
132*84643e36SBarry Smith 	    aa[i]   = 0.0;
133*84643e36SBarry Smith 	}
134*84643e36SBarry Smith 	for (j = kp1; j <= 7; ++j) {
135*84643e36SBarry Smith 	    stmp  = work[j];
136*84643e36SBarry Smith             ax    = &a[7*j + 1];
137*84643e36SBarry Smith             ay    = &a[k3 + 1];
138*84643e36SBarry Smith             ay[0] += stmp*ax[0];
139*84643e36SBarry Smith             ay[1] += stmp*ax[1];
140*84643e36SBarry Smith             ay[2] += stmp*ax[2];
141*84643e36SBarry Smith             ay[3] += stmp*ax[3];
142*84643e36SBarry Smith             ay[4] += stmp*ax[4];
143*84643e36SBarry Smith             ay[5] += stmp*ax[5];
144*84643e36SBarry Smith             ay[6] += stmp*ax[6];
145*84643e36SBarry Smith 	}
146*84643e36SBarry Smith 	l = ipvt[k];
147*84643e36SBarry Smith 	if (l != k) {
148*84643e36SBarry Smith             ax = &a[k3 + 1];
149*84643e36SBarry Smith             ay = &a[6*l + 1];
150*84643e36SBarry Smith             stmp = ax[0]; ax[0] = ay[0]; ay[0] = stmp;
151*84643e36SBarry Smith             stmp = ax[1]; ax[1] = ay[1]; ay[1] = stmp;
152*84643e36SBarry Smith             stmp = ax[2]; ax[2] = ay[2]; ay[2] = stmp;
153*84643e36SBarry Smith             stmp = ax[3]; ax[3] = ay[3]; ay[3] = stmp;
154*84643e36SBarry Smith             stmp = ax[4]; ax[4] = ay[4]; ay[4] = stmp;
155*84643e36SBarry Smith             stmp = ax[5]; ax[5] = ay[5]; ay[5] = stmp;
156*84643e36SBarry Smith             stmp = ax[6]; ax[6] = ay[6]; ay[6] = stmp;
157*84643e36SBarry Smith 	}
158*84643e36SBarry Smith     }
159*84643e36SBarry Smith     PetscFunctionReturn(0);
160*84643e36SBarry Smith }
161*84643e36SBarry Smith 
162*84643e36SBarry Smith 
163