xref: /petsc/src/mat/impls/baij/seq/baijfact13.c (revision 5d0c19d75c660d4fec594a5399ec8d8ba29c54a8)
1be1d678aSKris Buschelman #define PETSCMAT_DLL
2be1d678aSKris Buschelman 
383287d42SBarry Smith /*
483287d42SBarry Smith     Factorization code for BAIJ format.
583287d42SBarry Smith */
683287d42SBarry Smith #include "src/mat/impls/baij/seq/baij.h"
783287d42SBarry Smith #include "src/inline/ilu.h"
883287d42SBarry Smith 
983287d42SBarry Smith /* ------------------------------------------------------------*/
1083287d42SBarry Smith /*
1183287d42SBarry Smith       Version for when blocks are 3 by 3
1283287d42SBarry Smith */
134a2ae208SSatish Balay #undef __FUNCT__
144a2ae208SSatish Balay #define __FUNCT__ "MatLUFactorNumeric_SeqBAIJ_3"
15719d5645SBarry Smith PetscErrorCode MatLUFactorNumeric_SeqBAIJ_3(Mat C,Mat A,MatFactorInfo *info)
1683287d42SBarry Smith {
1783287d42SBarry Smith   Mat_SeqBAIJ    *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ *)C->data;
1883287d42SBarry Smith   IS             isrow = b->row,isicol = b->icol;
196849ba73SBarry Smith   PetscErrorCode ierr;
20*5d0c19d7SBarry Smith   const PetscInt *r,*ic;
21*5d0c19d7SBarry Smith   PetscInt       i,j,n = a->mbs,*bi = b->i,*bj = b->j;
22690b6cddSBarry Smith   PetscInt       *ajtmpold,*ajtmp,nz,row,*ai=a->i,*aj=a->j;
23690b6cddSBarry Smith   PetscInt       *diag_offset = b->diag,idx,*pj;
2483287d42SBarry Smith   MatScalar      *pv,*v,*rtmp,*pc,*w,*x;
2583287d42SBarry Smith   MatScalar      p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4;
2683287d42SBarry Smith   MatScalar      p5,p6,p7,p8,p9,x5,x6,x7,x8,x9;
2783287d42SBarry Smith   MatScalar      *ba = b->a,*aa = a->a;
2862bba022SBarry Smith   PetscReal      shift = info->shiftinblocks;
2983287d42SBarry Smith 
3083287d42SBarry Smith   PetscFunctionBegin;
3183287d42SBarry Smith   ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr);
3283287d42SBarry Smith   ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr);
33b0a32e0cSBarry Smith   ierr = PetscMalloc(9*(n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
3483287d42SBarry Smith 
3583287d42SBarry Smith   for (i=0; i<n; i++) {
3683287d42SBarry Smith     nz    = bi[i+1] - bi[i];
3783287d42SBarry Smith     ajtmp = bj + bi[i];
3883287d42SBarry Smith     for  (j=0; j<nz; j++) {
3983287d42SBarry Smith       x = rtmp + 9*ajtmp[j];
4083287d42SBarry Smith       x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = 0.0;
4183287d42SBarry Smith     }
4283287d42SBarry Smith     /* load in initial (unfactored row) */
4383287d42SBarry Smith     idx      = r[i];
4483287d42SBarry Smith     nz       = ai[idx+1] - ai[idx];
4583287d42SBarry Smith     ajtmpold = aj + ai[idx];
4683287d42SBarry Smith     v        = aa + 9*ai[idx];
4783287d42SBarry Smith     for (j=0; j<nz; j++) {
4883287d42SBarry Smith       x    = rtmp + 9*ic[ajtmpold[j]];
4983287d42SBarry Smith       x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3];
5083287d42SBarry Smith       x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8];
5183287d42SBarry Smith       v    += 9;
5283287d42SBarry Smith     }
5383287d42SBarry Smith     row = *ajtmp++;
5483287d42SBarry Smith     while (row < i) {
5583287d42SBarry Smith       pc = rtmp + 9*row;
5683287d42SBarry Smith       p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3];
5783287d42SBarry Smith       p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8];
5883287d42SBarry Smith       if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 ||
5983287d42SBarry Smith           p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) {
6083287d42SBarry Smith         pv = ba + 9*diag_offset[row];
6183287d42SBarry Smith         pj = bj + diag_offset[row] + 1;
6283287d42SBarry Smith         x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3];
6383287d42SBarry Smith         x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8];
6483287d42SBarry Smith         pc[0] = m1 = p1*x1 + p4*x2 + p7*x3;
6583287d42SBarry Smith         pc[1] = m2 = p2*x1 + p5*x2 + p8*x3;
6683287d42SBarry Smith         pc[2] = m3 = p3*x1 + p6*x2 + p9*x3;
6783287d42SBarry Smith 
6883287d42SBarry Smith         pc[3] = m4 = p1*x4 + p4*x5 + p7*x6;
6983287d42SBarry Smith         pc[4] = m5 = p2*x4 + p5*x5 + p8*x6;
7083287d42SBarry Smith         pc[5] = m6 = p3*x4 + p6*x5 + p9*x6;
7183287d42SBarry Smith 
7283287d42SBarry Smith         pc[6] = m7 = p1*x7 + p4*x8 + p7*x9;
7383287d42SBarry Smith         pc[7] = m8 = p2*x7 + p5*x8 + p8*x9;
7483287d42SBarry Smith         pc[8] = m9 = p3*x7 + p6*x8 + p9*x9;
7583287d42SBarry Smith         nz = bi[row+1] - diag_offset[row] - 1;
7683287d42SBarry Smith         pv += 9;
7783287d42SBarry Smith         for (j=0; j<nz; j++) {
7883287d42SBarry Smith           x1   = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3];
7983287d42SBarry Smith           x5   = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8];
8083287d42SBarry Smith           x    = rtmp + 9*pj[j];
8183287d42SBarry Smith           x[0] -= m1*x1 + m4*x2 + m7*x3;
8283287d42SBarry Smith           x[1] -= m2*x1 + m5*x2 + m8*x3;
8383287d42SBarry Smith           x[2] -= m3*x1 + m6*x2 + m9*x3;
8483287d42SBarry Smith 
8583287d42SBarry Smith           x[3] -= m1*x4 + m4*x5 + m7*x6;
8683287d42SBarry Smith           x[4] -= m2*x4 + m5*x5 + m8*x6;
8783287d42SBarry Smith           x[5] -= m3*x4 + m6*x5 + m9*x6;
8883287d42SBarry Smith 
8983287d42SBarry Smith           x[6] -= m1*x7 + m4*x8 + m7*x9;
9083287d42SBarry Smith           x[7] -= m2*x7 + m5*x8 + m8*x9;
9183287d42SBarry Smith           x[8] -= m3*x7 + m6*x8 + m9*x9;
9283287d42SBarry Smith           pv   += 9;
9383287d42SBarry Smith         }
94efee365bSSatish Balay         ierr = PetscLogFlops(54*nz+36);CHKERRQ(ierr);
9583287d42SBarry Smith       }
9683287d42SBarry Smith       row = *ajtmp++;
9783287d42SBarry Smith     }
9883287d42SBarry Smith     /* finished row so stick it into b->a */
9983287d42SBarry Smith     pv = ba + 9*bi[i];
10083287d42SBarry Smith     pj = bj + bi[i];
10183287d42SBarry Smith     nz = bi[i+1] - bi[i];
10283287d42SBarry Smith     for (j=0; j<nz; j++) {
10383287d42SBarry Smith       x     = rtmp + 9*pj[j];
10483287d42SBarry Smith       pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3];
10583287d42SBarry Smith       pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8];
10683287d42SBarry Smith       pv   += 9;
10783287d42SBarry Smith     }
10883287d42SBarry Smith     /* invert diagonal block */
10983287d42SBarry Smith     w = ba + 9*diag_offset[i];
11062bba022SBarry Smith     ierr = Kernel_A_gets_inverse_A_3(w,shift);CHKERRQ(ierr);
11183287d42SBarry Smith   }
11283287d42SBarry Smith 
11383287d42SBarry Smith   ierr = PetscFree(rtmp);CHKERRQ(ierr);
11483287d42SBarry Smith   ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr);
11583287d42SBarry Smith   ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr);
116db4efbfdSBarry Smith   C->ops->solve          = MatSolve_SeqBAIJ_3;
117db4efbfdSBarry Smith   C->ops->solvetranspose = MatSolveTranspose_SeqBAIJ_3;
11883287d42SBarry Smith   C->assembled = PETSC_TRUE;
119efee365bSSatish Balay   ierr = PetscLogFlops(1.3333*27*b->mbs);CHKERRQ(ierr); /* from inverting diagonal blocks */
12083287d42SBarry Smith   PetscFunctionReturn(0);
12183287d42SBarry Smith }
122