xref: /petsc/src/mat/impls/baij/seq/dgefa4.c (revision 8a36c062a42118aeb7a8c1f4af6ee2dc0f5436fa)
1*8a36c062SBarry Smith 
24224c193SBarry Smith #ifndef lint
3*8a36c062SBarry Smith static char vcid[] = "$Id: dgefa4.c,v 1.1 1997/06/10 03:18:18 bsmith Exp bsmith $";
44224c193SBarry Smith #endif
54224c193SBarry Smith /*
6*8a36c062SBarry Smith     Inverts 4 by 4 matrix using partial pivoting.
74224c193SBarry Smith */
84224c193SBarry Smith #include "petsc.h"
94224c193SBarry Smith 
104224c193SBarry Smith #undef __FUNC__
11*8a36c062SBarry Smith #define __FUNC__ "Kernel_A_gets_inverse_A_4"
12*8a36c062SBarry Smith int Kernel_A_gets_inverse_A_4(Scalar *a)
134224c193SBarry Smith {
14*8a36c062SBarry Smith     int     i__2, i__3, kp1, j, k, l,ll,i,ipvt_l[4],*ipvt = ipvt_l-1,kb,k3;
154224c193SBarry Smith     int     k4,j3;
16*8a36c062SBarry Smith     Scalar  *aa,*ax,*ay,work_l[16],*work = work_l-1,stmp;
174224c193SBarry Smith     double  tmp,max;
184224c193SBarry Smith 
194224c193SBarry Smith /*     gaussian elimination with partial pivoting */
204224c193SBarry Smith 
214224c193SBarry Smith     /* Parameter adjustments */
22*8a36c062SBarry Smith     a       -= 5;
234224c193SBarry Smith 
24*8a36c062SBarry Smith     for (k = 1; k <= 3; ++k) {
254224c193SBarry Smith 	kp1 = k + 1;
26*8a36c062SBarry Smith         k3  = 4*k;
274224c193SBarry Smith         k4  = k3 + k;
284224c193SBarry Smith /*        find l = pivot index */
294224c193SBarry Smith 
304224c193SBarry Smith 	i__2 = 4 - k;
314224c193SBarry Smith         aa = &a[k4];
324224c193SBarry Smith         max = PetscAbsScalar(aa[0]);
334224c193SBarry Smith         l = 1;
344224c193SBarry Smith         for ( ll=1; ll<i__2; ll++ ) {
354224c193SBarry Smith           tmp = PetscAbsScalar(aa[ll]);
364224c193SBarry Smith           if (tmp > max) { max = tmp; l = ll+1;}
374224c193SBarry Smith         }
384224c193SBarry Smith         l       += k - 1;
394224c193SBarry Smith 	ipvt[k] = l;
404224c193SBarry Smith 
414224c193SBarry Smith 	if (a[l + k3] == 0.) {
424224c193SBarry Smith 	  SETERRQ(k,0,"Zero pivot");
434224c193SBarry Smith 	}
444224c193SBarry Smith 
454224c193SBarry Smith /*           interchange if necessary */
464224c193SBarry Smith 
474224c193SBarry Smith 	if (l != k) {
484224c193SBarry Smith 	  stmp      = a[l + k3];
494224c193SBarry Smith 	  a[l + k3] = a[k4];
504224c193SBarry Smith 	  a[k4]     = stmp;
514224c193SBarry Smith         }
524224c193SBarry Smith 
534224c193SBarry Smith /*           compute multipliers */
544224c193SBarry Smith 
554224c193SBarry Smith 	stmp = -1. / a[k4];
56*8a36c062SBarry Smith 	i__2 = 4 - k;
574224c193SBarry Smith         aa = &a[1 + k4];
584224c193SBarry Smith         for ( ll=0; ll<i__2; ll++ ) {
594224c193SBarry Smith           aa[ll] *= stmp;
604224c193SBarry Smith         }
614224c193SBarry Smith 
624224c193SBarry Smith /*           row elimination with column indexing */
634224c193SBarry Smith 
644224c193SBarry Smith 	ax = &a[k4+1];
65*8a36c062SBarry Smith         for (j = kp1; j <= 4; ++j) {
66*8a36c062SBarry Smith             j3   = 4*j;
674224c193SBarry Smith 	    stmp = a[l + j3];
684224c193SBarry Smith 	    if (l != k) {
694224c193SBarry Smith 	      a[l + j3] = a[k + j3];
704224c193SBarry Smith 	      a[k + j3] = stmp;
714224c193SBarry Smith             }
724224c193SBarry Smith 
73*8a36c062SBarry Smith 	    i__3 = 4 - k;
744224c193SBarry Smith             ay = &a[1+k+j3];
754224c193SBarry Smith             for ( ll=0; ll<i__3; ll++ ) {
764224c193SBarry Smith               ay[ll] += stmp*ax[ll];
774224c193SBarry Smith             }
784224c193SBarry Smith 	}
794224c193SBarry Smith     }
80*8a36c062SBarry Smith     ipvt[4] = 4;
81*8a36c062SBarry Smith     if (a[20] == 0.) {
824224c193SBarry Smith 	SETERRQ(3,0,"Zero pivot,final row");
834224c193SBarry Smith     }
844224c193SBarry Smith 
854224c193SBarry Smith     /*
864224c193SBarry Smith          Now form the inverse
874224c193SBarry Smith     */
884224c193SBarry Smith 
894224c193SBarry Smith    /*     compute inverse(u) */
904224c193SBarry Smith 
91*8a36c062SBarry Smith     for (k = 1; k <= 4; ++k) {
92*8a36c062SBarry Smith         k3    = 4*k;
934224c193SBarry Smith         k4    = k3 + k;
944224c193SBarry Smith 	a[k4] = 1.0 / a[k4];
954224c193SBarry Smith 	stmp  = -a[k4];
964224c193SBarry Smith 	i__2  = k - 1;
974224c193SBarry Smith         aa    = &a[k3 + 1];
984224c193SBarry Smith         for ( ll=0; ll<i__2; ll++ ) aa[ll] *= stmp;
994224c193SBarry Smith 	kp1 = k + 1;
100*8a36c062SBarry Smith 	if (4 < kp1) continue;
1014224c193SBarry Smith         ax = aa;
102*8a36c062SBarry Smith         for (j = kp1; j <= 4; ++j) {
103*8a36c062SBarry Smith             j3        = 4*j;
1044224c193SBarry Smith 	    stmp      = a[k + j3];
1054224c193SBarry Smith 	    a[k + j3] = 0.0;
1064224c193SBarry Smith             ay        = &a[j3 + 1];
1074224c193SBarry Smith             for ( ll=0; ll<k; ll++ ) {
1084224c193SBarry Smith               ay[ll] += stmp*ax[ll];
1094224c193SBarry Smith             }
1104224c193SBarry Smith 	}
1114224c193SBarry Smith     }
1124224c193SBarry Smith 
1134224c193SBarry Smith    /*    form inverse(u)*inverse(l) */
1144224c193SBarry Smith 
115*8a36c062SBarry Smith     for (kb = 1; kb <= 3; ++kb) {
116*8a36c062SBarry Smith 	k   = 4 - kb;
117*8a36c062SBarry Smith         k3  = 4*k;
1184224c193SBarry Smith 	kp1 = k + 1;
1194224c193SBarry Smith         aa  = a + k3;
120*8a36c062SBarry Smith 	for (i = kp1; i <= 4; ++i) {
1214224c193SBarry Smith 	    work[i] = aa[i];
1224224c193SBarry Smith 	    aa[i]   = 0.0;
1234224c193SBarry Smith 	}
124*8a36c062SBarry Smith 	for (j = kp1; j <= 4; ++j) {
1254224c193SBarry Smith 	    stmp  = work[j];
126*8a36c062SBarry Smith             ax    = &a[4*j + 1];
1274224c193SBarry Smith             ay    = &a[k3 + 1];
1284224c193SBarry Smith             ay[0] += stmp*ax[0];
1294224c193SBarry Smith             ay[1] += stmp*ax[1];
1304224c193SBarry Smith             ay[2] += stmp*ax[2];
131*8a36c062SBarry Smith             ay[3] += stmp*ax[3];
1324224c193SBarry Smith 	}
1334224c193SBarry Smith 	l = ipvt[k];
1344224c193SBarry Smith 	if (l != k) {
1354224c193SBarry Smith             ax = &a[k3 + 1];
136*8a36c062SBarry Smith             ay = &a[4*l + 1];
1374224c193SBarry Smith             stmp = ax[0]; ax[0] = ay[0]; ay[0] = stmp;
1384224c193SBarry Smith             stmp = ax[1]; ax[1] = ay[1]; ay[1] = stmp;
1394224c193SBarry Smith             stmp = ax[2]; ax[2] = ay[2]; ay[2] = stmp;
140*8a36c062SBarry Smith             stmp = ax[3]; ax[3] = ay[3]; ay[3] = stmp;
1414224c193SBarry Smith 	}
1424224c193SBarry Smith     }
1434224c193SBarry Smith     return 0;
1444224c193SBarry Smith }
1454224c193SBarry Smith 
146