xref: /petsc/src/mat/impls/baij/seq/dgefa2.c (revision 29bbc08cd5461c366aba6645924e8ff42acd1de0)
1*29bbc08cSBarry Smith /*$Id: dgefa2.c,v 1.6 2000/04/12 04:23:32 bsmith Exp bsmith $*/
284643e36SBarry Smith /*
384643e36SBarry Smith      Inverts 2 by 2 matrix using partial pivoting.
484643e36SBarry Smith 
584643e36SBarry Smith        Used by the sparse factorization routines in
684643e36SBarry Smith      src/mat/impls/baij/seq and src/mat/impls/bdiag/seq
784643e36SBarry Smith 
884643e36SBarry Smith        See also src/inline/ilu.h
984643e36SBarry Smith 
1084643e36SBarry Smith        This is a combination of the Linpack routines
1184643e36SBarry Smith     dgefa() and dgedi() specialized for a size of 2.
1284643e36SBarry Smith 
1384643e36SBarry Smith */
1484643e36SBarry Smith #include "petsc.h"
1584643e36SBarry Smith 
1684643e36SBarry Smith #undef __FUNC__
17b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"Kernel_A_gets_inverse_A_2"
1884643e36SBarry Smith int Kernel_A_gets_inverse_A_2(MatScalar *a)
1984643e36SBarry Smith {
202515f8d2SSatish Balay     int        i__2,i__3,kp1,j,k,l,ll,i,ipvt_l[2],*ipvt = ipvt_l-1,k3;
2184643e36SBarry Smith     int        k4,j3;
2284643e36SBarry Smith     MatScalar  *aa,*ax,*ay,work_l[4],*work = work_l-1,stmp;
23329f5518SBarry Smith     MatReal    tmp,max;
2484643e36SBarry Smith 
2584643e36SBarry Smith /*     gaussian elimination with partial pivoting */
2684643e36SBarry Smith 
2784643e36SBarry Smith     PetscFunctionBegin;
2884643e36SBarry Smith     /* Parameter adjustments */
2984643e36SBarry Smith     a       -= 3;
3084643e36SBarry Smith 
3184643e36SBarry Smith     /*for (k = 1; k <= 1; ++k) {*/
3284643e36SBarry Smith         k   = 1;
3384643e36SBarry Smith 	kp1 = k + 1;
3484643e36SBarry Smith         k3  = 2*k;
3584643e36SBarry Smith         k4  = k3 + k;
3684643e36SBarry Smith /*        find l = pivot index */
3784643e36SBarry Smith 
3884643e36SBarry Smith 	i__2 = 2 - k;
3984643e36SBarry Smith         aa = &a[k4];
4084643e36SBarry Smith         max = PetscAbsScalar(aa[0]);
4184643e36SBarry Smith         l = 1;
4284643e36SBarry Smith         for (ll=1; ll<i__2; ll++) {
4384643e36SBarry Smith           tmp = PetscAbsScalar(aa[ll]);
4484643e36SBarry Smith           if (tmp > max) { max = tmp; l = ll+1;}
4584643e36SBarry Smith         }
4684643e36SBarry Smith         l       += k - 1;
4784643e36SBarry Smith 	ipvt[k] = l;
4884643e36SBarry Smith 
4984643e36SBarry Smith 	if (a[l + k3] == 0.) {
50*29bbc08cSBarry Smith 	  SETERRQ(k,"Zero pivot");
5184643e36SBarry Smith 	}
5284643e36SBarry Smith 
5384643e36SBarry Smith /*           interchange if necessary */
5484643e36SBarry Smith 
5584643e36SBarry Smith 	if (l != k) {
5684643e36SBarry Smith 	  stmp      = a[l + k3];
5784643e36SBarry Smith 	  a[l + k3] = a[k4];
5884643e36SBarry Smith 	  a[k4]     = stmp;
5984643e36SBarry Smith         }
6084643e36SBarry Smith 
6184643e36SBarry Smith /*           compute multipliers */
6284643e36SBarry Smith 
6384643e36SBarry Smith 	stmp = -1. / a[k4];
6484643e36SBarry Smith 	i__2 = 2 - k;
6584643e36SBarry Smith         aa = &a[1 + k4];
6684643e36SBarry Smith         for (ll=0; ll<i__2; ll++) {
6784643e36SBarry Smith           aa[ll] *= stmp;
6884643e36SBarry Smith         }
6984643e36SBarry Smith 
7084643e36SBarry Smith /*           row elimination with column indexing */
7184643e36SBarry Smith 
7284643e36SBarry Smith 	ax = &a[k4+1];
7384643e36SBarry Smith         for (j = kp1; j <= 2; ++j) {
7484643e36SBarry Smith             j3   = 2*j;
7584643e36SBarry Smith 	    stmp = a[l + j3];
7684643e36SBarry Smith 	    if (l != k) {
7784643e36SBarry Smith 	      a[l + j3] = a[k + j3];
7884643e36SBarry Smith 	      a[k + j3] = stmp;
7984643e36SBarry Smith             }
8084643e36SBarry Smith 
8184643e36SBarry Smith 	    i__3 = 2 - k;
8284643e36SBarry Smith             ay = &a[1+k+j3];
8384643e36SBarry Smith             for (ll=0; ll<i__3; ll++) {
8484643e36SBarry Smith               ay[ll] += stmp*ax[ll];
8584643e36SBarry Smith             }
8684643e36SBarry Smith 	}
8784643e36SBarry Smith     /*}*/
8884643e36SBarry Smith     ipvt[2] = 2;
8984643e36SBarry Smith     if (a[6] == 0.) {
90*29bbc08cSBarry Smith 	SETERRQ(3,"Zero pivot,final row");
9184643e36SBarry Smith     }
9284643e36SBarry Smith 
9384643e36SBarry Smith     /*
9484643e36SBarry Smith          Now form the inverse
9584643e36SBarry Smith     */
9684643e36SBarry Smith 
9784643e36SBarry Smith    /*     compute inverse(u) */
9884643e36SBarry Smith 
9984643e36SBarry Smith     for (k = 1; k <= 2; ++k) {
10084643e36SBarry Smith         k3    = 2*k;
10184643e36SBarry Smith         k4    = k3 + k;
10284643e36SBarry Smith 	a[k4] = 1.0 / a[k4];
10384643e36SBarry Smith 	stmp  = -a[k4];
10484643e36SBarry Smith 	i__2  = k - 1;
10584643e36SBarry Smith         aa    = &a[k3 + 1];
10684643e36SBarry Smith         for (ll=0; ll<i__2; ll++) aa[ll] *= stmp;
10784643e36SBarry Smith 	kp1 = k + 1;
10884643e36SBarry Smith 	if (2 < kp1) continue;
10984643e36SBarry Smith         ax = aa;
11084643e36SBarry Smith         for (j = kp1; j <= 2; ++j) {
11184643e36SBarry Smith             j3        = 2*j;
11284643e36SBarry Smith 	    stmp      = a[k + j3];
11384643e36SBarry Smith 	    a[k + j3] = 0.0;
11484643e36SBarry Smith             ay        = &a[j3 + 1];
11584643e36SBarry Smith             for (ll=0; ll<k; ll++) {
11684643e36SBarry Smith               ay[ll] += stmp*ax[ll];
11784643e36SBarry Smith             }
11884643e36SBarry Smith 	}
11984643e36SBarry Smith     }
12084643e36SBarry Smith 
12184643e36SBarry Smith    /*    form inverse(u)*inverse(l) */
12284643e36SBarry Smith 
12384643e36SBarry Smith     /*for (kb = 1; kb <= 1; ++kb) {*/
12484643e36SBarry Smith 
12584643e36SBarry Smith 	k   = 1;
12684643e36SBarry Smith         k3  = 2*k;
12784643e36SBarry Smith 	kp1 = k + 1;
12884643e36SBarry Smith         aa  = a + k3;
12984643e36SBarry Smith 	for (i = kp1; i <= 2; ++i) {
13084643e36SBarry Smith             work_l[i-1] = aa[i];
13184643e36SBarry Smith             /* work[i] = aa[i]; Fix for -O3 error on Origin 2000 */
13284643e36SBarry Smith 	    aa[i]   = 0.0;
13384643e36SBarry Smith 	}
13484643e36SBarry Smith 	for (j = kp1; j <= 2; ++j) {
13584643e36SBarry Smith 	    stmp  = work[j];
13684643e36SBarry Smith             ax    = &a[2*j + 1];
13784643e36SBarry Smith             ay    = &a[k3 + 1];
13884643e36SBarry Smith             ay[0] += stmp*ax[0];
13984643e36SBarry Smith             ay[1] += stmp*ax[1];
14084643e36SBarry Smith 	}
14184643e36SBarry Smith 	l = ipvt[k];
14284643e36SBarry Smith 	if (l != k) {
14384643e36SBarry Smith             ax = &a[k3 + 1];
14484643e36SBarry Smith             ay = &a[2*l + 1];
14584643e36SBarry Smith             stmp = ax[0]; ax[0] = ay[0]; ay[0] = stmp;
14684643e36SBarry Smith             stmp = ax[1]; ax[1] = ay[1]; ay[1] = stmp;
14784643e36SBarry Smith 	}
14884643e36SBarry Smith 
14984643e36SBarry Smith     PetscFunctionReturn(0);
15084643e36SBarry Smith }
15184643e36SBarry Smith 
15284643e36SBarry Smith 
153