xref: /petsc/src/mat/impls/baij/seq/dgefa3.c (revision 96b95a6bf9ed541503cb02ce9899069044d5ddfa)
1be1d678aSKris Buschelman 
2ede5e988SBarry Smith /*
325783f72SBarry Smith      Inverts 3 by 3 matrix using partial pivoting.
471c5468dSBarry Smith 
571c5468dSBarry Smith        Used by the sparse factorization routines in
6dd882469SBarry Smith      src/mat/impls/baij/seq
771c5468dSBarry Smith 
871c5468dSBarry Smith 
971c5468dSBarry Smith        This is a combination of the Linpack routines
1071c5468dSBarry Smith     dgefa() and dgedi() specialized for a size of 3.
1171c5468dSBarry Smith 
12ede5e988SBarry Smith */
13c6db04a5SJed Brown #include <petscsys.h>
14ede5e988SBarry Smith 
154a2ae208SSatish Balay #undef __FUNCT__
16*96b95a6bSBarry Smith #define __FUNCT__ "PetscKernel_A_gets_inverse_A_3"
17*96b95a6bSBarry Smith PetscErrorCode PetscKernel_A_gets_inverse_A_3(MatScalar *a,PetscReal shift)
18ede5e988SBarry Smith {
19690b6cddSBarry Smith     PetscInt   i__2,i__3,kp1,j,k,l,ll,i,ipvt[3],kb,k3;
20690b6cddSBarry Smith     PetscInt   k4,j3;
21b48ee343SBarry Smith     MatScalar  *aa,*ax,*ay,work[9],stmp;
22329f5518SBarry Smith     MatReal    tmp,max;
23ede5e988SBarry Smith 
24ede5e988SBarry Smith /*     gaussian elimination with partial pivoting */
25ede5e988SBarry Smith 
263a40ed3dSBarry Smith     PetscFunctionBegin;
270daa89e9SHong Zhang     shift = .333*shift*(1.e-12 + PetscAbsScalar(a[0]) + PetscAbsScalar(a[4]) + PetscAbsScalar(a[8]));
28ede5e988SBarry Smith     /* Parameter adjustments */
29ede5e988SBarry Smith     a       -= 4;
30ede5e988SBarry Smith 
31ede5e988SBarry Smith     for (k = 1; k <= 2; ++k) {
32ede5e988SBarry Smith 	kp1 = k + 1;
3373d4a2d6SBarry Smith         k3  = 3*k;
3473d4a2d6SBarry Smith         k4  = k3 + k;
35ede5e988SBarry Smith /*        find l = pivot index */
36ede5e988SBarry Smith 
37ede5e988SBarry Smith 	i__2 = 4 - k;
3873d4a2d6SBarry Smith         aa = &a[k4];
39ede5e988SBarry Smith         max = PetscAbsScalar(aa[0]);
40ede5e988SBarry Smith         l = 1;
41ede5e988SBarry Smith         for (ll=1; ll<i__2; ll++) {
42ede5e988SBarry Smith           tmp = PetscAbsScalar(aa[ll]);
43ede5e988SBarry Smith           if (tmp > max) { max = tmp; l = ll+1;}
44ede5e988SBarry Smith         }
45ede5e988SBarry Smith         l       += k - 1;
46da10e913SBarry Smith 	ipvt[k-1] = l;
47ede5e988SBarry Smith 
48ab055debSShri Abhyankar 	if (a[l + k3] == 0.0) {
49ab055debSShri Abhyankar 	  if (shift == 0.0) {
50ab055debSShri Abhyankar 	    SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D",k-1);
51ab055debSShri Abhyankar 	  } else {
52ab055debSShri Abhyankar 	    /* Shift is applied to single diagonal entry */
53ab055debSShri Abhyankar 	    a[l + k3] = shift;
54ab055debSShri Abhyankar 	  }
55ab055debSShri Abhyankar 	}
56ede5e988SBarry Smith /*           interchange if necessary */
57ede5e988SBarry Smith 
58ede5e988SBarry Smith 	if (l != k) {
5939d66777SBarry Smith 	  stmp      = a[l + k3];
6073d4a2d6SBarry Smith 	  a[l + k3] = a[k4];
6139d66777SBarry Smith 	  a[k4]     = stmp;
62ede5e988SBarry Smith         }
63ede5e988SBarry Smith 
64ede5e988SBarry Smith /*           compute multipliers */
65ede5e988SBarry Smith 
6639d66777SBarry Smith 	stmp = -1. / a[k4];
67ede5e988SBarry Smith 	i__2 = 3 - k;
6873d4a2d6SBarry Smith         aa = &a[1 + k4];
69ede5e988SBarry Smith         for (ll=0; ll<i__2; ll++) {
7039d66777SBarry Smith           aa[ll] *= stmp;
71ede5e988SBarry Smith         }
72ede5e988SBarry Smith 
73ede5e988SBarry Smith /*           row elimination with column indexing */
74ede5e988SBarry Smith 
7573d4a2d6SBarry Smith 	ax = &a[k4+1];
76ede5e988SBarry Smith         for (j = kp1; j <= 3; ++j) {
7773d4a2d6SBarry Smith             j3   = 3*j;
7839d66777SBarry Smith 	    stmp = a[l + j3];
79ede5e988SBarry Smith 	    if (l != k) {
8073d4a2d6SBarry Smith 	      a[l + j3] = a[k + j3];
8139d66777SBarry Smith 	      a[k + j3] = stmp;
82ede5e988SBarry Smith             }
83ede5e988SBarry Smith 
84ede5e988SBarry Smith 	    i__3 = 3 - k;
8539d66777SBarry Smith             ay = &a[1+k+j3];
86ede5e988SBarry Smith             for (ll=0; ll<i__3; ll++) {
8739d66777SBarry Smith               ay[ll] += stmp*ax[ll];
88ede5e988SBarry Smith             }
89ede5e988SBarry Smith 	}
90ede5e988SBarry Smith     }
91da10e913SBarry Smith     ipvt[2] = 3;
9265e19b50SBarry Smith     if (a[12] == 0.0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D",2);
93ede5e988SBarry Smith 
94ede5e988SBarry Smith     /*
9525783f72SBarry Smith          Now form the inverse
96ede5e988SBarry Smith     */
97ede5e988SBarry Smith 
98ede5e988SBarry Smith    /*     compute inverse(u) */
99ede5e988SBarry Smith 
10025783f72SBarry Smith     for (k = 1; k <= 3; ++k) {
10173d4a2d6SBarry Smith         k3    = 3*k;
10273d4a2d6SBarry Smith         k4    = k3 + k;
10373d4a2d6SBarry Smith 	a[k4] = 1.0 / a[k4];
10439d66777SBarry Smith 	stmp  = -a[k4];
105ede5e988SBarry Smith 	i__2  = k - 1;
10673d4a2d6SBarry Smith         aa    = &a[k3 + 1];
10739d66777SBarry Smith         for (ll=0; ll<i__2; ll++) aa[ll] *= stmp;
108ede5e988SBarry Smith 	kp1 = k + 1;
10925783f72SBarry Smith 	if (3 < kp1) continue;
110ede5e988SBarry Smith         ax = aa;
11125783f72SBarry Smith         for (j = kp1; j <= 3; ++j) {
11273d4a2d6SBarry Smith             j3        = 3*j;
11339d66777SBarry Smith 	    stmp      = a[k + j3];
11439d66777SBarry Smith 	    a[k + j3] = 0.0;
11573d4a2d6SBarry Smith             ay        = &a[j3 + 1];
116ede5e988SBarry Smith             for (ll=0; ll<k; ll++) {
11739d66777SBarry Smith               ay[ll] += stmp*ax[ll];
118ede5e988SBarry Smith             }
119ede5e988SBarry Smith 	}
120ede5e988SBarry Smith     }
121ede5e988SBarry Smith 
122ede5e988SBarry Smith    /*    form inverse(u)*inverse(l) */
123ede5e988SBarry Smith 
12425783f72SBarry Smith     for (kb = 1; kb <= 2; ++kb) {
12525783f72SBarry Smith 	k   = 3 - kb;
12673d4a2d6SBarry Smith         k3  = 3*k;
127ede5e988SBarry Smith 	kp1 = k + 1;
12873d4a2d6SBarry Smith         aa  = a + k3;
12925783f72SBarry Smith 	for (i = kp1; i <= 3; ++i) {
130b48ee343SBarry Smith             work[i-1] = aa[i];
13173d4a2d6SBarry Smith 	    aa[i]   = 0.0;
132ede5e988SBarry Smith 	}
13325783f72SBarry Smith 	for (j = kp1; j <= 3; ++j) {
134b48ee343SBarry Smith 	    stmp  = work[j-1];
13525783f72SBarry Smith             ax    = &a[3*j + 1];
13673d4a2d6SBarry Smith             ay    = &a[k3 + 1];
13739d66777SBarry Smith             ay[0] += stmp*ax[0];
13839d66777SBarry Smith             ay[1] += stmp*ax[1];
13939d66777SBarry Smith             ay[2] += stmp*ax[2];
140ede5e988SBarry Smith 	}
141da10e913SBarry Smith 	l = ipvt[k-1];
142ede5e988SBarry Smith 	if (l != k) {
14373d4a2d6SBarry Smith             ax = &a[k3 + 1];
14425783f72SBarry Smith             ay = &a[3*l + 1];
14573d4a2d6SBarry Smith             stmp = ax[0]; ax[0] = ay[0]; ay[0] = stmp;
14673d4a2d6SBarry Smith             stmp = ax[1]; ax[1] = ay[1]; ay[1] = stmp;
14773d4a2d6SBarry Smith             stmp = ax[2]; ax[2] = ay[2]; ay[2] = stmp;
148ede5e988SBarry Smith 	}
149ede5e988SBarry Smith     }
1503a40ed3dSBarry Smith     PetscFunctionReturn(0);
151ede5e988SBarry Smith }
152ede5e988SBarry Smith 
15371c5468dSBarry Smith 
154