xref: /petsc/src/mat/impls/baij/seq/dgefa5.c (revision 98921bda46e76d7aaed9e0138c5ff9d0ce93f355)
1be1d678aSKris Buschelman 
24224c193SBarry Smith /*
3ec1892c8SHong Zhang       Inverts 5 by 5 matrix using gaussian elimination with partial pivoting.
471c5468dSBarry Smith 
571c5468dSBarry Smith        Used by the sparse factorization routines in
6dd882469SBarry Smith      src/mat/impls/baij/seq
771c5468dSBarry Smith 
871c5468dSBarry Smith        This is a combination of the Linpack routines
971c5468dSBarry Smith     dgefa() and dgedi() specialized for a size of 5.
1071c5468dSBarry Smith 
114224c193SBarry Smith */
12c6db04a5SJed Brown #include <petscsys.h>
134224c193SBarry Smith 
142e92ee13SHong Zhang PETSC_EXTERN PetscErrorCode PetscKernel_A_gets_inverse_A_5(MatScalar *a,PetscInt *ipvt,MatScalar *work,PetscReal shift,PetscBool allowzeropivot,PetscBool *zeropivotdetected)
154224c193SBarry Smith {
16766f9fbaSBarry Smith   PetscInt  i__2,i__3,kp1,j,k,l,ll,i,kb,k3;
17690b6cddSBarry Smith   PetscInt  k4,j3;
18766f9fbaSBarry Smith   MatScalar *aa,*ax,*ay,stmp;
19329f5518SBarry Smith   MatReal   tmp,max;
204224c193SBarry Smith 
213a40ed3dSBarry Smith   PetscFunctionBegin;
22c80103daSHong Zhang   if (zeropivotdetected) *zeropivotdetected = PETSC_FALSE;
23943c8ff5SShri Abhyankar   shift = .25*shift*(1.e-12 + PetscAbsScalar(a[0]) + PetscAbsScalar(a[6]) + PetscAbsScalar(a[12]) + PetscAbsScalar(a[18]) + PetscAbsScalar(a[24]));
24ec1892c8SHong Zhang 
254224c193SBarry Smith   /* Parameter adjustments */
268a36c062SBarry Smith   a -= 6;
274224c193SBarry Smith 
288a36c062SBarry Smith   for (k = 1; k <= 4; ++k) {
294224c193SBarry Smith     kp1 = k + 1;
308a36c062SBarry Smith     k3  = 5*k;
314224c193SBarry Smith     k4  = k3 + k;
324224c193SBarry Smith 
33ec1892c8SHong Zhang     /* find l = pivot index */
34ed33f8a5SSatish Balay     i__2 = 6 - k;
354224c193SBarry Smith     aa   = &a[k4];
364224c193SBarry Smith     max  = PetscAbsScalar(aa[0]);
374224c193SBarry Smith     l    = 1;
384224c193SBarry Smith     for (ll=1; ll<i__2; ll++) {
394224c193SBarry Smith       tmp = PetscAbsScalar(aa[ll]);
404224c193SBarry Smith       if (tmp > max) { max = tmp; l = ll+1;}
414224c193SBarry Smith     }
424224c193SBarry Smith     l        += k - 1;
43b48ee343SBarry Smith     ipvt[k-1] = l;
444224c193SBarry Smith 
45943c8ff5SShri Abhyankar     if (a[l + k3] == 0.0) {
46ec1892c8SHong Zhang       if (shift == 0.0) {
47ec1892c8SHong Zhang         if (allowzeropivot) {
48ec1892c8SHong Zhang           PetscErrorCode ierr;
49c0aa6a63SJacob Faibussowitsch           ierr = PetscInfo1(NULL,"Zero pivot, row %" PetscInt_FMT "\n",k-1);CHKERRQ(ierr);
50ec1892c8SHong Zhang           *zeropivotdetected = PETSC_TRUE;
51*98921bdaSJacob Faibussowitsch         } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %" PetscInt_FMT,k-1);
52ec1892c8SHong Zhang       } else {
53943c8ff5SShri Abhyankar         /* SHIFT is applied to SINGLE diagonal entry; does this make any sense? */
54943c8ff5SShri Abhyankar         a[l + k3] = shift;
55943c8ff5SShri Abhyankar       }
56943c8ff5SShri Abhyankar     }
574224c193SBarry Smith 
584224c193SBarry Smith     /* interchange if necessary */
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     stmp = -1. / a[k4];
678a36c062SBarry Smith     i__2 = 5 - k;
684224c193SBarry Smith     aa   = &a[1 + k4];
6926fbe8dcSKarl Rupp     for (ll=0; ll<i__2; ll++) aa[ll] *= stmp;
704224c193SBarry Smith 
714224c193SBarry Smith     /* row elimination with column indexing */
724224c193SBarry Smith     ax = &a[k4+1];
738a36c062SBarry Smith     for (j = kp1; j <= 5; ++j) {
748a36c062SBarry Smith       j3   = 5*j;
754224c193SBarry Smith       stmp = a[l + j3];
764224c193SBarry Smith       if (l != k) {
774224c193SBarry Smith         a[l + j3] = a[k + j3];
784224c193SBarry Smith         a[k + j3] = stmp;
794224c193SBarry Smith       }
804224c193SBarry Smith 
818a36c062SBarry Smith       i__3 = 5 - k;
824224c193SBarry Smith       ay   = &a[1+k+j3];
8326fbe8dcSKarl Rupp       for (ll=0; ll<i__3; ll++) ay[ll] += stmp*ax[ll];
844224c193SBarry Smith     }
854224c193SBarry Smith   }
86b48ee343SBarry Smith   ipvt[4] = 5;
872e92ee13SHong Zhang   if (a[30] == 0.0) {
88c0aa6a63SJacob Faibussowitsch     if (PetscLikely(allowzeropivot)) {
89ec1892c8SHong Zhang       PetscErrorCode ierr;
90c0aa6a63SJacob Faibussowitsch       ierr = PetscInfo(NULL,"Zero pivot, row 4\n");CHKERRQ(ierr);
912e92ee13SHong Zhang       *zeropivotdetected = PETSC_TRUE;
92c0aa6a63SJacob Faibussowitsch     } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row 4");
932e92ee13SHong Zhang   }
944224c193SBarry Smith 
95ec1892c8SHong Zhang   /* Now form the inverse */
964224c193SBarry Smith   /* compute inverse(u) */
978a36c062SBarry Smith   for (k = 1; k <= 5; ++k) {
988a36c062SBarry Smith     k3    = 5*k;
994224c193SBarry Smith     k4    = k3 + k;
1004224c193SBarry Smith     a[k4] = 1.0 / a[k4];
1014224c193SBarry Smith     stmp  = -a[k4];
1024224c193SBarry Smith     i__2  = k - 1;
1034224c193SBarry Smith     aa    = &a[k3 + 1];
1044224c193SBarry Smith     for (ll=0; ll<i__2; ll++) aa[ll] *= stmp;
1054224c193SBarry Smith     kp1 = k + 1;
1068a36c062SBarry Smith     if (5 < kp1) continue;
1074224c193SBarry Smith     ax = aa;
1088a36c062SBarry Smith     for (j = kp1; j <= 5; ++j) {
1098a36c062SBarry Smith       j3        = 5*j;
1104224c193SBarry Smith       stmp      = a[k + j3];
1114224c193SBarry Smith       a[k + j3] = 0.0;
1124224c193SBarry Smith       ay        = &a[j3 + 1];
11326fbe8dcSKarl Rupp       for (ll=0; ll<k; ll++) ay[ll] += stmp*ax[ll];
1144224c193SBarry Smith     }
1154224c193SBarry Smith   }
1164224c193SBarry Smith 
1174224c193SBarry Smith   /* form inverse(u)*inverse(l) */
1188a36c062SBarry Smith   for (kb = 1; kb <= 4; ++kb) {
1198a36c062SBarry Smith     k   = 5 - kb;
1208a36c062SBarry Smith     k3  = 5*k;
1214224c193SBarry Smith     kp1 = k + 1;
1224224c193SBarry Smith     aa  = a + k3;
1238a36c062SBarry Smith     for (i = kp1; i <= 5; ++i) {
124b48ee343SBarry Smith       work[i-1] = aa[i];
1254224c193SBarry Smith       aa[i]     = 0.0;
1264224c193SBarry Smith     }
1278a36c062SBarry Smith     for (j = kp1; j <= 5; ++j) {
128b48ee343SBarry Smith       stmp   = work[j-1];
1298a36c062SBarry Smith       ax     = &a[5*j + 1];
1304224c193SBarry Smith       ay     = &a[k3 + 1];
1314224c193SBarry Smith       ay[0] += stmp*ax[0];
1324224c193SBarry Smith       ay[1] += stmp*ax[1];
1334224c193SBarry Smith       ay[2] += stmp*ax[2];
1348a36c062SBarry Smith       ay[3] += stmp*ax[3];
1358a36c062SBarry Smith       ay[4] += stmp*ax[4];
1364224c193SBarry Smith     }
137b48ee343SBarry Smith     l = ipvt[k-1];
1384224c193SBarry Smith     if (l != k) {
1394224c193SBarry Smith       ax   = &a[k3 + 1];
1408a36c062SBarry Smith       ay   = &a[5*l + 1];
1414224c193SBarry Smith       stmp = ax[0]; ax[0] = ay[0]; ay[0] = stmp;
1424224c193SBarry Smith       stmp = ax[1]; ax[1] = ay[1]; ay[1] = stmp;
1434224c193SBarry Smith       stmp = ax[2]; ax[2] = ay[2]; ay[2] = stmp;
1448a36c062SBarry Smith       stmp = ax[3]; ax[3] = ay[3]; ay[3] = stmp;
1458a36c062SBarry Smith       stmp = ax[4]; ax[4] = ay[4]; ay[4] = stmp;
1464224c193SBarry Smith     }
1474224c193SBarry Smith   }
1483a40ed3dSBarry Smith   PetscFunctionReturn(0);
1494224c193SBarry Smith }
1504224c193SBarry Smith 
151