1be1d678aSKris Buschelman 284643e36SBarry Smith /* 3*ec1892c8SHong Zhang Inverts 7 by 7 matrix using gaussian elimination with partial pivoting. 484643e36SBarry Smith 584643e36SBarry Smith Used by the sparse factorization routines in 6dd882469SBarry Smith src/mat/impls/baij/seq 784643e36SBarry Smith 884643e36SBarry Smith This is a combination of the Linpack routines 984643e36SBarry Smith dgefa() and dgedi() specialized for a size of 7. 1084643e36SBarry Smith 1184643e36SBarry Smith */ 12c6db04a5SJed Brown #include <petscsys.h> 1384643e36SBarry Smith 144a2ae208SSatish Balay #undef __FUNCT__ 1596b95a6bSBarry Smith #define __FUNCT__ "PetscKernel_A_gets_inverse_A_7" 162e92ee13SHong Zhang PETSC_EXTERN PetscErrorCode PetscKernel_A_gets_inverse_A_7(MatScalar *a,PetscReal shift,PetscBool allowzeropivot,PetscBool *zeropivotdetected) 1784643e36SBarry Smith { 18690b6cddSBarry Smith PetscInt i__2,i__3,kp1,j,k,l,ll,i,ipvt[7],kb,k3; 19690b6cddSBarry Smith PetscInt k4,j3; 201f1046a5SBarry Smith MatScalar *aa,*ax,*ay,work[49],stmp; 21329f5518SBarry Smith MatReal tmp,max; 2284643e36SBarry Smith 2384643e36SBarry Smith PetscFunctionBegin; 24c80103daSHong Zhang if (zeropivotdetected) *zeropivotdetected = PETSC_FALSE; 25943c8ff5SShri Abhyankar shift = .25*shift*(1.e-12 + PetscAbsScalar(a[0]) + PetscAbsScalar(a[8]) + PetscAbsScalar(a[16]) + PetscAbsScalar(a[24]) + PetscAbsScalar(a[32]) + PetscAbsScalar(a[40]) + PetscAbsScalar(a[48])); 26943c8ff5SShri Abhyankar 2784643e36SBarry Smith /* Parameter adjustments */ 2884643e36SBarry Smith a -= 8; 2984643e36SBarry Smith 3084643e36SBarry Smith for (k = 1; k <= 6; ++k) { 3184643e36SBarry Smith kp1 = k + 1; 3284643e36SBarry Smith k3 = 7*k; 3384643e36SBarry Smith k4 = k3 + k; 3484643e36SBarry Smith 35*ec1892c8SHong Zhang /* find l = pivot index */ 36ed33f8a5SSatish Balay i__2 = 8 - k; 3784643e36SBarry Smith aa = &a[k4]; 3884643e36SBarry Smith max = PetscAbsScalar(aa[0]); 3984643e36SBarry Smith l = 1; 4084643e36SBarry Smith for (ll=1; ll<i__2; ll++) { 4184643e36SBarry Smith tmp = PetscAbsScalar(aa[ll]); 4284643e36SBarry Smith if (tmp > max) { max = tmp; l = ll+1;} 4384643e36SBarry Smith } 4484643e36SBarry Smith l += k - 1; 451f1046a5SBarry Smith ipvt[k-1] = l; 4684643e36SBarry Smith 47943c8ff5SShri Abhyankar if (a[l + k3] == 0.0) { 48*ec1892c8SHong Zhang if (shift == 0.0) { 49*ec1892c8SHong Zhang if (allowzeropivot) { 50*ec1892c8SHong Zhang PetscErrorCode ierr; 51*ec1892c8SHong Zhang ierr = PetscInfo1(NULL,"Zero pivot, row %D\n",k-1);CHKERRQ(ierr); 52*ec1892c8SHong Zhang *zeropivotdetected = PETSC_TRUE; 53*ec1892c8SHong Zhang } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D",k-1); 54*ec1892c8SHong Zhang } else { 55943c8ff5SShri Abhyankar /* SHIFT is applied to SINGLE diagonal entry; does this make any sense? */ 56943c8ff5SShri Abhyankar a[l + k3] = shift; 57943c8ff5SShri Abhyankar } 58943c8ff5SShri Abhyankar } 5984643e36SBarry Smith 6084643e36SBarry Smith /* interchange if necessary */ 6184643e36SBarry Smith if (l != k) { 6284643e36SBarry Smith stmp = a[l + k3]; 6384643e36SBarry Smith a[l + k3] = a[k4]; 6484643e36SBarry Smith a[k4] = stmp; 6584643e36SBarry Smith } 6684643e36SBarry Smith 6784643e36SBarry Smith /* compute multipliers */ 6884643e36SBarry Smith stmp = -1. / a[k4]; 6984643e36SBarry Smith i__2 = 7 - k; 7084643e36SBarry Smith aa = &a[1 + k4]; 7126fbe8dcSKarl Rupp for (ll=0; ll<i__2; ll++) aa[ll] *= stmp; 7284643e36SBarry Smith 7384643e36SBarry Smith /* row elimination with column indexing */ 7484643e36SBarry Smith ax = &a[k4+1]; 7584643e36SBarry Smith for (j = kp1; j <= 7; ++j) { 7684643e36SBarry Smith j3 = 7*j; 7784643e36SBarry Smith stmp = a[l + j3]; 7884643e36SBarry Smith if (l != k) { 7984643e36SBarry Smith a[l + j3] = a[k + j3]; 8084643e36SBarry Smith a[k + j3] = stmp; 8184643e36SBarry Smith } 8284643e36SBarry Smith 8384643e36SBarry Smith i__3 = 7 - k; 8484643e36SBarry Smith ay = &a[1+k+j3]; 8526fbe8dcSKarl Rupp for (ll=0; ll<i__3; ll++) ay[ll] += stmp*ax[ll]; 8684643e36SBarry Smith } 8784643e36SBarry Smith } 881f1046a5SBarry Smith ipvt[6] = 7; 892e92ee13SHong Zhang if (a[56] == 0.0) { 902e92ee13SHong Zhang if (allowzeropivot) { 91*ec1892c8SHong Zhang PetscErrorCode ierr; 922e92ee13SHong Zhang ierr = PetscInfo1(NULL,"Zero pivot, row %D\n",6);CHKERRQ(ierr); 932e92ee13SHong Zhang *zeropivotdetected = PETSC_TRUE; 942e92ee13SHong Zhang } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D",6); 952e92ee13SHong Zhang } 9684643e36SBarry Smith 97*ec1892c8SHong Zhang /* Now form the inverse */ 9884643e36SBarry Smith /* compute inverse(u) */ 9984643e36SBarry Smith for (k = 1; k <= 7; ++k) { 10084643e36SBarry Smith k3 = 7*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 (7 < kp1) continue; 10984643e36SBarry Smith ax = aa; 11084643e36SBarry Smith for (j = kp1; j <= 7; ++j) { 11184643e36SBarry Smith j3 = 7*j; 11284643e36SBarry Smith stmp = a[k + j3]; 11384643e36SBarry Smith a[k + j3] = 0.0; 11484643e36SBarry Smith ay = &a[j3 + 1]; 11526fbe8dcSKarl Rupp for (ll=0; ll<k; ll++) ay[ll] += stmp*ax[ll]; 11684643e36SBarry Smith } 11784643e36SBarry Smith } 11884643e36SBarry Smith 11984643e36SBarry Smith /* form inverse(u)*inverse(l) */ 12084643e36SBarry Smith for (kb = 1; kb <= 6; ++kb) { 12184643e36SBarry Smith k = 7 - kb; 12284643e36SBarry Smith k3 = 7*k; 12384643e36SBarry Smith kp1 = k + 1; 12484643e36SBarry Smith aa = a + k3; 12584643e36SBarry Smith for (i = kp1; i <= 7; ++i) { 1261f1046a5SBarry Smith work[i-1] = aa[i]; 12784643e36SBarry Smith aa[i] = 0.0; 12884643e36SBarry Smith } 12984643e36SBarry Smith for (j = kp1; j <= 7; ++j) { 1301f1046a5SBarry Smith stmp = work[j-1]; 13184643e36SBarry Smith ax = &a[7*j + 1]; 13284643e36SBarry Smith ay = &a[k3 + 1]; 13384643e36SBarry Smith ay[0] += stmp*ax[0]; 13484643e36SBarry Smith ay[1] += stmp*ax[1]; 13584643e36SBarry Smith ay[2] += stmp*ax[2]; 13684643e36SBarry Smith ay[3] += stmp*ax[3]; 13784643e36SBarry Smith ay[4] += stmp*ax[4]; 13884643e36SBarry Smith ay[5] += stmp*ax[5]; 13984643e36SBarry Smith ay[6] += stmp*ax[6]; 14084643e36SBarry Smith } 1411f1046a5SBarry Smith l = ipvt[k-1]; 14284643e36SBarry Smith if (l != k) { 14384643e36SBarry Smith ax = &a[k3 + 1]; 14415091d37SBarry Smith ay = &a[7*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 stmp = ax[2]; ax[2] = ay[2]; ay[2] = stmp; 14884643e36SBarry Smith stmp = ax[3]; ax[3] = ay[3]; ay[3] = stmp; 14984643e36SBarry Smith stmp = ax[4]; ax[4] = ay[4]; ay[4] = stmp; 15084643e36SBarry Smith stmp = ax[5]; ax[5] = ay[5]; ay[5] = stmp; 15184643e36SBarry Smith stmp = ax[6]; ax[6] = ay[6]; ay[6] = stmp; 15284643e36SBarry Smith } 15384643e36SBarry Smith } 15484643e36SBarry Smith PetscFunctionReturn(0); 15584643e36SBarry Smith } 15684643e36SBarry Smith 15784643e36SBarry Smith 158