xref: /petsc/src/mat/impls/baij/seq/dgefa4.c (revision 29bbc08cd5461c366aba6645924e8ff42acd1de0)
1*29bbc08cSBarry Smith /*$Id: dgefa4.c,v 1.11 2000/04/12 04:23:32 bsmith Exp bsmith $*/
24224c193SBarry Smith /*
38a36c062SBarry Smith        Inverts 4 by 4 matrix using partial pivoting.
471c5468dSBarry Smith 
571c5468dSBarry Smith        Used by the sparse factorization routines in
671c5468dSBarry Smith      src/mat/impls/baij/seq and src/mat/impls/bdiag/seq
771c5468dSBarry Smith 
871c5468dSBarry Smith        See also src/inline/ilu.h
971c5468dSBarry Smith 
1071c5468dSBarry Smith        This is a combination of the Linpack routines
1171c5468dSBarry Smith     dgefa() and dgedi() specialized for a size of 4.
1271c5468dSBarry Smith 
134224c193SBarry Smith */
144224c193SBarry Smith #include "petsc.h"
154224c193SBarry Smith 
164224c193SBarry Smith #undef __FUNC__
17b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"Kernel_A_gets_inverse_A_4"
183f1db9ecSBarry Smith int Kernel_A_gets_inverse_A_4(MatScalar *a)
194224c193SBarry Smith {
208a36c062SBarry Smith     int        i__2,i__3,kp1,j,k,l,ll,i,ipvt_l[4],*ipvt = ipvt_l-1,kb,k3;
214224c193SBarry Smith     int        k4,j3;
223f1db9ecSBarry Smith     MatScalar  *aa,*ax,*ay,work_l[16],*work = work_l-1,stmp;
23329f5518SBarry Smith     MatReal    tmp,max;
244224c193SBarry Smith 
254224c193SBarry Smith /*     gaussian elimination with partial pivoting */
264224c193SBarry Smith 
273a40ed3dSBarry Smith     PetscFunctionBegin;
284224c193SBarry Smith     /* Parameter adjustments */
298a36c062SBarry Smith     a       -= 5;
304224c193SBarry Smith 
318a36c062SBarry Smith     for (k = 1; k <= 3; ++k) {
324224c193SBarry Smith 	kp1 = k + 1;
338a36c062SBarry Smith         k3  = 4*k;
344224c193SBarry Smith         k4  = k3 + k;
354224c193SBarry Smith /*        find l = pivot index */
364224c193SBarry Smith 
374224c193SBarry Smith 	i__2 = 4 - k;
384224c193SBarry Smith         aa = &a[k4];
394224c193SBarry Smith         max = PetscAbsScalar(aa[0]);
404224c193SBarry Smith         l = 1;
414224c193SBarry Smith         for (ll=1; ll<i__2; ll++) {
424224c193SBarry Smith           tmp = PetscAbsScalar(aa[ll]);
434224c193SBarry Smith           if (tmp > max) { max = tmp; l = ll+1;}
444224c193SBarry Smith         }
454224c193SBarry Smith         l       += k - 1;
464224c193SBarry Smith 	ipvt[k] = l;
474224c193SBarry Smith 
484224c193SBarry Smith 	if (a[l + k3] == 0.) {
49*29bbc08cSBarry Smith 	  SETERRQ(k,"Zero pivot");
504224c193SBarry Smith 	}
514224c193SBarry Smith 
524224c193SBarry Smith /*           interchange if necessary */
534224c193SBarry Smith 
544224c193SBarry Smith 	if (l != k) {
554224c193SBarry Smith 	  stmp      = a[l + k3];
564224c193SBarry Smith 	  a[l + k3] = a[k4];
574224c193SBarry Smith 	  a[k4]     = stmp;
584224c193SBarry Smith         }
594224c193SBarry Smith 
604224c193SBarry Smith /*           compute multipliers */
614224c193SBarry Smith 
624224c193SBarry Smith 	stmp = -1. / a[k4];
638a36c062SBarry Smith 	i__2 = 4 - k;
644224c193SBarry Smith         aa = &a[1 + k4];
654224c193SBarry Smith         for (ll=0; ll<i__2; ll++) {
664224c193SBarry Smith           aa[ll] *= stmp;
674224c193SBarry Smith         }
684224c193SBarry Smith 
694224c193SBarry Smith /*           row elimination with column indexing */
704224c193SBarry Smith 
714224c193SBarry Smith 	ax = &a[k4+1];
728a36c062SBarry Smith         for (j = kp1; j <= 4; ++j) {
738a36c062SBarry Smith             j3   = 4*j;
744224c193SBarry Smith 	    stmp = a[l + j3];
754224c193SBarry Smith 	    if (l != k) {
764224c193SBarry Smith 	      a[l + j3] = a[k + j3];
774224c193SBarry Smith 	      a[k + j3] = stmp;
784224c193SBarry Smith             }
794224c193SBarry Smith 
808a36c062SBarry Smith 	    i__3 = 4 - k;
814224c193SBarry Smith             ay = &a[1+k+j3];
824224c193SBarry Smith             for (ll=0; ll<i__3; ll++) {
834224c193SBarry Smith               ay[ll] += stmp*ax[ll];
844224c193SBarry Smith             }
854224c193SBarry Smith 	}
864224c193SBarry Smith     }
878a36c062SBarry Smith     ipvt[4] = 4;
888a36c062SBarry Smith     if (a[20] == 0.) {
89*29bbc08cSBarry Smith 	SETERRQ(3,"Zero pivot,final row");
904224c193SBarry Smith     }
914224c193SBarry Smith 
924224c193SBarry Smith     /*
934224c193SBarry Smith          Now form the inverse
944224c193SBarry Smith     */
954224c193SBarry Smith 
964224c193SBarry Smith    /*     compute inverse(u) */
974224c193SBarry Smith 
988a36c062SBarry Smith     for (k = 1; k <= 4; ++k) {
998a36c062SBarry Smith         k3    = 4*k;
1004224c193SBarry Smith         k4    = k3 + k;
1014224c193SBarry Smith 	a[k4] = 1.0 / a[k4];
1024224c193SBarry Smith 	stmp  = -a[k4];
1034224c193SBarry Smith 	i__2  = k - 1;
1044224c193SBarry Smith         aa    = &a[k3 + 1];
1054224c193SBarry Smith         for (ll=0; ll<i__2; ll++) aa[ll] *= stmp;
1064224c193SBarry Smith 	kp1 = k + 1;
1078a36c062SBarry Smith 	if (4 < kp1) continue;
1084224c193SBarry Smith         ax = aa;
1098a36c062SBarry Smith         for (j = kp1; j <= 4; ++j) {
1108a36c062SBarry Smith             j3        = 4*j;
1114224c193SBarry Smith 	    stmp      = a[k + j3];
1124224c193SBarry Smith 	    a[k + j3] = 0.0;
1134224c193SBarry Smith             ay        = &a[j3 + 1];
1144224c193SBarry Smith             for (ll=0; ll<k; ll++) {
1154224c193SBarry Smith               ay[ll] += stmp*ax[ll];
1164224c193SBarry Smith             }
1174224c193SBarry Smith 	}
1184224c193SBarry Smith     }
1194224c193SBarry Smith 
1204224c193SBarry Smith    /*    form inverse(u)*inverse(l) */
1214224c193SBarry Smith 
1228a36c062SBarry Smith     for (kb = 1; kb <= 3; ++kb) {
1238a36c062SBarry Smith 	k   = 4 - kb;
1248a36c062SBarry Smith         k3  = 4*k;
1254224c193SBarry Smith 	kp1 = k + 1;
1264224c193SBarry Smith         aa  = a + k3;
1278a36c062SBarry Smith 	for (i = kp1; i <= 4; ++i) {
128f3f07278SSatish Balay             work_l[i-1] = aa[i];
129f3f07278SSatish Balay             /* work[i] = aa[i]; Fix for -O3 error on Origin 2000 */
1304224c193SBarry Smith 	    aa[i]   = 0.0;
1314224c193SBarry Smith 	}
1328a36c062SBarry Smith 	for (j = kp1; j <= 4; ++j) {
1334224c193SBarry Smith 	    stmp  = work[j];
1348a36c062SBarry Smith             ax    = &a[4*j + 1];
1354224c193SBarry Smith             ay    = &a[k3 + 1];
1364224c193SBarry Smith             ay[0] += stmp*ax[0];
1374224c193SBarry Smith             ay[1] += stmp*ax[1];
1384224c193SBarry Smith             ay[2] += stmp*ax[2];
1398a36c062SBarry Smith             ay[3] += stmp*ax[3];
1404224c193SBarry Smith 	}
1414224c193SBarry Smith 	l = ipvt[k];
1424224c193SBarry Smith 	if (l != k) {
1434224c193SBarry Smith             ax = &a[k3 + 1];
1448a36c062SBarry Smith             ay = &a[4*l + 1];
1454224c193SBarry Smith             stmp = ax[0]; ax[0] = ay[0]; ay[0] = stmp;
1464224c193SBarry Smith             stmp = ax[1]; ax[1] = ay[1]; ay[1] = stmp;
1474224c193SBarry Smith             stmp = ax[2]; ax[2] = ay[2]; ay[2] = stmp;
1488a36c062SBarry Smith             stmp = ax[3]; ax[3] = ay[3]; ay[3] = stmp;
1494224c193SBarry Smith 	}
1504224c193SBarry Smith     }
1513a40ed3dSBarry Smith     PetscFunctionReturn(0);
1524224c193SBarry Smith }
1534224c193SBarry Smith 
154