xref: /petsc/src/mat/impls/baij/seq/dgefa5.c (revision 943c8ff524c3b3228b734140abd933ffca3f4d05)
1be1d678aSKris Buschelman #define PETSCMAT_DLL
2be1d678aSKris Buschelman 
34224c193SBarry Smith /*
48a36c062SBarry Smith       Inverts 5 by 5 matrix using partial pivoting.
571c5468dSBarry Smith 
671c5468dSBarry Smith        Used by the sparse factorization routines in
7dd882469SBarry Smith      src/mat/impls/baij/seq
871c5468dSBarry Smith 
971c5468dSBarry Smith        This is a combination of the Linpack routines
1071c5468dSBarry Smith     dgefa() and dgedi() specialized for a size of 5.
1171c5468dSBarry Smith 
124224c193SBarry Smith */
13d382aafbSBarry Smith #include "petscsys.h"
144224c193SBarry Smith 
154a2ae208SSatish Balay #undef __FUNCT__
164a2ae208SSatish Balay #define __FUNCT__ "Kernel_A_gets_inverse_A_5"
17766f9fbaSBarry Smith PetscErrorCode Kernel_A_gets_inverse_A_5(MatScalar *a,PetscInt *ipvt,MatScalar *work,PetscReal shift)
184224c193SBarry Smith {
19766f9fbaSBarry Smith     PetscInt   i__2,i__3,kp1,j,k,l,ll,i,kb,k3;
20690b6cddSBarry Smith     PetscInt   k4,j3;
21766f9fbaSBarry Smith     MatScalar  *aa,*ax,*ay,stmp;
22329f5518SBarry Smith     MatReal    tmp,max;
234224c193SBarry Smith 
244224c193SBarry Smith /*     gaussian elimination with partial pivoting */
254224c193SBarry Smith 
263a40ed3dSBarry Smith     PetscFunctionBegin;
27*943c8ff5SShri Abhyankar     shift = .25*shift*(1.e-12 + PetscAbsScalar(a[0]) + PetscAbsScalar(a[6]) + PetscAbsScalar(a[12]) + PetscAbsScalar(a[18]) + PetscAbsScalar(a[24]));
284224c193SBarry Smith     /* Parameter adjustments */
298a36c062SBarry Smith     a       -= 6;
304224c193SBarry Smith 
318a36c062SBarry Smith     for (k = 1; k <= 4; ++k) {
324224c193SBarry Smith 	kp1 = k + 1;
338a36c062SBarry Smith         k3  = 5*k;
344224c193SBarry Smith         k4  = k3 + k;
354224c193SBarry Smith /*        find l = pivot index */
364224c193SBarry Smith 
37ed33f8a5SSatish Balay 	i__2 = 6 - 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;
46b48ee343SBarry Smith 	ipvt[k-1] = l;
474224c193SBarry Smith 
48*943c8ff5SShri Abhyankar         if (a[l + k3] == 0.0) {
49*943c8ff5SShri Abhyankar           if (shift == 0.0) {
50*943c8ff5SShri Abhyankar 	    SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D",k-1);
51*943c8ff5SShri Abhyankar   	  } else {
52*943c8ff5SShri Abhyankar             /* SHIFT is applied to SINGLE diagonal entry; does this make any sense? */
53*943c8ff5SShri Abhyankar   	    a[l + k3] = shift;
54*943c8ff5SShri Abhyankar   	  }
55*943c8ff5SShri Abhyankar         }
564224c193SBarry Smith 
574224c193SBarry Smith /*           interchange if necessary */
584224c193SBarry Smith 
594224c193SBarry Smith 	if (l != k) {
604224c193SBarry Smith 	  stmp      = a[l + k3];
614224c193SBarry Smith 	  a[l + k3] = a[k4];
624224c193SBarry Smith 	  a[k4]     = stmp;
634224c193SBarry Smith         }
644224c193SBarry Smith 
654224c193SBarry Smith /*           compute multipliers */
664224c193SBarry Smith 
674224c193SBarry Smith 	stmp = -1. / a[k4];
688a36c062SBarry Smith 	i__2 = 5 - k;
694224c193SBarry Smith         aa = &a[1 + k4];
704224c193SBarry Smith         for (ll=0; ll<i__2; ll++) {
714224c193SBarry Smith           aa[ll] *= stmp;
724224c193SBarry Smith         }
734224c193SBarry Smith 
744224c193SBarry Smith /*           row elimination with column indexing */
754224c193SBarry Smith 
764224c193SBarry Smith 	ax = &a[k4+1];
778a36c062SBarry Smith         for (j = kp1; j <= 5; ++j) {
788a36c062SBarry Smith             j3   = 5*j;
794224c193SBarry Smith 	    stmp = a[l + j3];
804224c193SBarry Smith 	    if (l != k) {
814224c193SBarry Smith 	      a[l + j3] = a[k + j3];
824224c193SBarry Smith 	      a[k + j3] = stmp;
834224c193SBarry Smith             }
844224c193SBarry Smith 
858a36c062SBarry Smith 	    i__3 = 5 - k;
864224c193SBarry Smith             ay = &a[1+k+j3];
874224c193SBarry Smith             for (ll=0; ll<i__3; ll++) {
884224c193SBarry Smith               ay[ll] += stmp*ax[ll];
894224c193SBarry Smith             }
904224c193SBarry Smith 	}
914224c193SBarry Smith     }
92b48ee343SBarry Smith     ipvt[4] = 5;
9365e19b50SBarry Smith     if (a[30] == 0.0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D",4);
944224c193SBarry Smith 
954224c193SBarry Smith     /*
964224c193SBarry Smith          Now form the inverse
974224c193SBarry Smith     */
984224c193SBarry Smith 
994224c193SBarry Smith    /*     compute inverse(u) */
1004224c193SBarry Smith 
1018a36c062SBarry Smith     for (k = 1; k <= 5; ++k) {
1028a36c062SBarry Smith         k3    = 5*k;
1034224c193SBarry Smith         k4    = k3 + k;
1044224c193SBarry Smith 	a[k4] = 1.0 / a[k4];
1054224c193SBarry Smith 	stmp  = -a[k4];
1064224c193SBarry Smith 	i__2  = k - 1;
1074224c193SBarry Smith         aa    = &a[k3 + 1];
1084224c193SBarry Smith         for (ll=0; ll<i__2; ll++) aa[ll] *= stmp;
1094224c193SBarry Smith 	kp1 = k + 1;
1108a36c062SBarry Smith 	if (5 < kp1) continue;
1114224c193SBarry Smith         ax = aa;
1128a36c062SBarry Smith         for (j = kp1; j <= 5; ++j) {
1138a36c062SBarry Smith             j3        = 5*j;
1144224c193SBarry Smith 	    stmp      = a[k + j3];
1154224c193SBarry Smith 	    a[k + j3] = 0.0;
1164224c193SBarry Smith             ay        = &a[j3 + 1];
1174224c193SBarry Smith             for (ll=0; ll<k; ll++) {
1184224c193SBarry Smith               ay[ll] += stmp*ax[ll];
1194224c193SBarry Smith             }
1204224c193SBarry Smith 	}
1214224c193SBarry Smith     }
1224224c193SBarry Smith 
1234224c193SBarry Smith    /*    form inverse(u)*inverse(l) */
1244224c193SBarry Smith 
1258a36c062SBarry Smith     for (kb = 1; kb <= 4; ++kb) {
1268a36c062SBarry Smith 	k   = 5 - kb;
1278a36c062SBarry Smith         k3  = 5*k;
1284224c193SBarry Smith 	kp1 = k + 1;
1294224c193SBarry Smith         aa  = a + k3;
1308a36c062SBarry Smith 	for (i = kp1; i <= 5; ++i) {
131b48ee343SBarry Smith             work[i-1] = aa[i];
1324224c193SBarry Smith 	    aa[i]   = 0.0;
1334224c193SBarry Smith 	}
1348a36c062SBarry Smith 	for (j = kp1; j <= 5; ++j) {
135b48ee343SBarry Smith 	    stmp  = work[j-1];
1368a36c062SBarry Smith             ax    = &a[5*j + 1];
1374224c193SBarry Smith             ay    = &a[k3 + 1];
1384224c193SBarry Smith             ay[0] += stmp*ax[0];
1394224c193SBarry Smith             ay[1] += stmp*ax[1];
1404224c193SBarry Smith             ay[2] += stmp*ax[2];
1418a36c062SBarry Smith             ay[3] += stmp*ax[3];
1428a36c062SBarry Smith             ay[4] += stmp*ax[4];
1434224c193SBarry Smith 	}
144b48ee343SBarry Smith 	l = ipvt[k-1];
1454224c193SBarry Smith 	if (l != k) {
1464224c193SBarry Smith             ax = &a[k3 + 1];
1478a36c062SBarry Smith             ay = &a[5*l + 1];
1484224c193SBarry Smith             stmp = ax[0]; ax[0] = ay[0]; ay[0] = stmp;
1494224c193SBarry Smith             stmp = ax[1]; ax[1] = ay[1]; ay[1] = stmp;
1504224c193SBarry Smith             stmp = ax[2]; ax[2] = ay[2]; ay[2] = stmp;
1518a36c062SBarry Smith             stmp = ax[3]; ax[3] = ay[3]; ay[3] = stmp;
1528a36c062SBarry Smith             stmp = ax[4]; ax[4] = ay[4]; ay[4] = stmp;
1534224c193SBarry Smith 	}
1544224c193SBarry Smith     }
1553a40ed3dSBarry Smith     PetscFunctionReturn(0);
1564224c193SBarry Smith }
1574224c193SBarry Smith 
158