xref: /petsc/src/mat/impls/baij/seq/baijsolvtrannat3.c (revision 9566063d113dddea24716c546802770db7481bc0)
12c733ed4SBarry Smith #include <../src/mat/impls/baij/seq/baij.h>
22c733ed4SBarry Smith 
32c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_3_NaturalOrdering_inplace(Mat A,Vec bb,Vec xx)
42c733ed4SBarry Smith {
52c733ed4SBarry Smith   Mat_SeqBAIJ     *a=(Mat_SeqBAIJ*)A->data;
62c733ed4SBarry Smith   const PetscInt  n=a->mbs,*vi,*ai=a->i,*aj=a->j,*diag=a->diag;
72c733ed4SBarry Smith   PetscInt        i,nz,idx,idt,oidx;
82c733ed4SBarry Smith   const MatScalar *aa=a->a,*v;
92c733ed4SBarry Smith   PetscScalar     s1,s2,s3,x1,x2,x3,*x;
102c733ed4SBarry Smith 
112c733ed4SBarry Smith   PetscFunctionBegin;
12*9566063dSJacob Faibussowitsch   PetscCall(VecCopy(bb,xx));
13*9566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx,&x));
142c733ed4SBarry Smith 
152c733ed4SBarry Smith   /* forward solve the U^T */
162c733ed4SBarry Smith   idx = 0;
172c733ed4SBarry Smith   for (i=0; i<n; i++) {
182c733ed4SBarry Smith 
192c733ed4SBarry Smith     v = aa + 9*diag[i];
202c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
212c733ed4SBarry Smith     x1 = x[idx];   x2 = x[1+idx]; x3    = x[2+idx];
222c733ed4SBarry Smith     s1 = v[0]*x1  +  v[1]*x2 +  v[2]*x3;
232c733ed4SBarry Smith     s2 = v[3]*x1  +  v[4]*x2 +  v[5]*x3;
242c733ed4SBarry Smith     s3 = v[6]*x1  +  v[7]*x2 + v[8]*x3;
252c733ed4SBarry Smith     v += 9;
262c733ed4SBarry Smith 
272c733ed4SBarry Smith     vi = aj + diag[i] + 1;
282c733ed4SBarry Smith     nz = ai[i+1] - diag[i] - 1;
292c733ed4SBarry Smith     while (nz--) {
302c733ed4SBarry Smith       oidx       = 3*(*vi++);
312c733ed4SBarry Smith       x[oidx]   -= v[0]*s1  +  v[1]*s2 +  v[2]*s3;
322c733ed4SBarry Smith       x[oidx+1] -= v[3]*s1  +  v[4]*s2 +  v[5]*s3;
332c733ed4SBarry Smith       x[oidx+2] -= v[6]*s1 + v[7]*s2 + v[8]*s3;
342c733ed4SBarry Smith       v         += 9;
352c733ed4SBarry Smith     }
362c733ed4SBarry Smith     x[idx] = s1;x[1+idx] = s2; x[2+idx] = s3;
372c733ed4SBarry Smith     idx   += 3;
382c733ed4SBarry Smith   }
392c733ed4SBarry Smith   /* backward solve the L^T */
402c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
412c733ed4SBarry Smith     v   = aa + 9*diag[i] - 9;
422c733ed4SBarry Smith     vi  = aj + diag[i] - 1;
432c733ed4SBarry Smith     nz  = diag[i] - ai[i];
442c733ed4SBarry Smith     idt = 3*i;
452c733ed4SBarry Smith     s1  = x[idt];  s2 = x[1+idt]; s3 = x[2+idt];
462c733ed4SBarry Smith     while (nz--) {
472c733ed4SBarry Smith       idx       = 3*(*vi--);
482c733ed4SBarry Smith       x[idx]   -=  v[0]*s1 +  v[1]*s2 +  v[2]*s3;
492c733ed4SBarry Smith       x[idx+1] -=  v[3]*s1 +  v[4]*s2 +  v[5]*s3;
502c733ed4SBarry Smith       x[idx+2] -= v[6]*s1 + v[7]*s2 + v[8]*s3;
512c733ed4SBarry Smith       v        -= 9;
522c733ed4SBarry Smith     }
532c733ed4SBarry Smith   }
54*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
55*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*9.0*(a->nz) - 3.0*A->cmap->n));
562c733ed4SBarry Smith   PetscFunctionReturn(0);
572c733ed4SBarry Smith }
582c733ed4SBarry Smith 
592c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_3_NaturalOrdering(Mat A,Vec bb,Vec xx)
602c733ed4SBarry Smith {
612c733ed4SBarry Smith   Mat_SeqBAIJ     *a=(Mat_SeqBAIJ*)A->data;
622c733ed4SBarry Smith   const PetscInt  n=a->mbs,*vi,*ai=a->i,*aj=a->j,*diag=a->diag;
632c733ed4SBarry Smith   PetscInt        nz,idx,idt,j,i,oidx;
642c733ed4SBarry Smith   const PetscInt  bs =A->rmap->bs,bs2=a->bs2;
652c733ed4SBarry Smith   const MatScalar *aa=a->a,*v;
662c733ed4SBarry Smith   PetscScalar     s1,s2,s3,x1,x2,x3,*x;
672c733ed4SBarry Smith 
682c733ed4SBarry Smith   PetscFunctionBegin;
69*9566063dSJacob Faibussowitsch   PetscCall(VecCopy(bb,xx));
70*9566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx,&x));
712c733ed4SBarry Smith 
722c733ed4SBarry Smith   /* forward solve the U^T */
732c733ed4SBarry Smith   idx = 0;
742c733ed4SBarry Smith   for (i=0; i<n; i++) {
752c733ed4SBarry Smith     v = aa + bs2*diag[i];
762c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
772c733ed4SBarry Smith     x1 = x[idx];   x2 = x[1+idx];  x3 = x[2+idx];
782c733ed4SBarry Smith     s1 = v[0]*x1  +  v[1]*x2  + v[2]*x3;
792c733ed4SBarry Smith     s2 = v[3]*x1  +  v[4]*x2  + v[5]*x3;
802c733ed4SBarry Smith     s3 = v[6]*x1  +  v[7]*x2  + v[8]*x3;
812c733ed4SBarry Smith     v -= bs2;
822c733ed4SBarry Smith 
832c733ed4SBarry Smith     vi = aj + diag[i] - 1;
842c733ed4SBarry Smith     nz = diag[i] - diag[i+1] - 1;
852c733ed4SBarry Smith     for (j=0; j>-nz; j--) {
862c733ed4SBarry Smith       oidx       = bs*vi[j];
872c733ed4SBarry Smith       x[oidx]   -= v[0]*s1  +  v[1]*s2  + v[2]*s3;
882c733ed4SBarry Smith       x[oidx+1] -= v[3]*s1  +  v[4]*s2  + v[5]*s3;
892c733ed4SBarry Smith       x[oidx+2] -= v[6]*s1  +  v[7]*s2  + v[8]*s3;
902c733ed4SBarry Smith       v         -= bs2;
912c733ed4SBarry Smith     }
922c733ed4SBarry Smith     x[idx] = s1;x[1+idx] = s2;  x[2+idx] = s3;
932c733ed4SBarry Smith     idx   += bs;
942c733ed4SBarry Smith   }
952c733ed4SBarry Smith   /* backward solve the L^T */
962c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
972c733ed4SBarry Smith     v   = aa + bs2*ai[i];
982c733ed4SBarry Smith     vi  = aj + ai[i];
992c733ed4SBarry Smith     nz  = ai[i+1] - ai[i];
1002c733ed4SBarry Smith     idt = bs*i;
1012c733ed4SBarry Smith     s1  = x[idt];  s2 = x[1+idt];  s3 = x[2+idt];
1022c733ed4SBarry Smith     for (j=0; j<nz; j++) {
1032c733ed4SBarry Smith       idx       = bs*vi[j];
1042c733ed4SBarry Smith       x[idx]   -= v[0]*s1  +  v[1]*s2  + v[2]*s3;
1052c733ed4SBarry Smith       x[idx+1] -= v[3]*s1  +  v[4]*s2  + v[5]*s3;
1062c733ed4SBarry Smith       x[idx+2] -= v[6]*s1  +  v[7]*s2  + v[8]*s3;
1072c733ed4SBarry Smith       v        += bs2;
1082c733ed4SBarry Smith     }
1092c733ed4SBarry Smith   }
110*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
111*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*bs2*(a->nz) - bs*A->cmap->n));
1122c733ed4SBarry Smith   PetscFunctionReturn(0);
1132c733ed4SBarry Smith }
114