xref: /petsc/src/mat/impls/baij/seq/baijsolvtrannat2.c (revision 9371c9d470a9602b6d10a8bf50c9b2280a79e45a)
12c733ed4SBarry Smith #include <../src/mat/impls/baij/seq/baij.h>
22c733ed4SBarry Smith 
3*9371c9d4SSatish Balay PetscErrorCode MatSolveTranspose_SeqBAIJ_2_NaturalOrdering_inplace(Mat A, Vec bb, Vec xx) {
42c733ed4SBarry Smith   Mat_SeqBAIJ     *a = (Mat_SeqBAIJ *)A->data;
52c733ed4SBarry Smith   PetscInt         i, nz, idx, idt, oidx;
62c733ed4SBarry Smith   const PetscInt  *diag = a->diag, *vi, n = a->mbs, *ai = a->i, *aj = a->j;
72c733ed4SBarry Smith   const MatScalar *aa = a->a, *v;
82c733ed4SBarry Smith   PetscScalar      s1, s2, x1, x2, *x;
92c733ed4SBarry Smith 
102c733ed4SBarry Smith   PetscFunctionBegin;
119566063dSJacob Faibussowitsch   PetscCall(VecCopy(bb, xx));
129566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx, &x));
132c733ed4SBarry Smith 
142c733ed4SBarry Smith   /* forward solve the U^T */
152c733ed4SBarry Smith   idx = 0;
162c733ed4SBarry Smith   for (i = 0; i < n; i++) {
172c733ed4SBarry Smith     v  = aa + 4 * diag[i];
182c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
19*9371c9d4SSatish Balay     x1 = x[idx];
20*9371c9d4SSatish Balay     x2 = x[1 + idx];
212c733ed4SBarry Smith     s1 = v[0] * x1 + v[1] * x2;
222c733ed4SBarry Smith     s2 = v[2] * x1 + v[3] * x2;
232c733ed4SBarry Smith     v += 4;
242c733ed4SBarry Smith 
252c733ed4SBarry Smith     vi = aj + diag[i] + 1;
262c733ed4SBarry Smith     nz = ai[i + 1] - diag[i] - 1;
272c733ed4SBarry Smith     while (nz--) {
282c733ed4SBarry Smith       oidx = 2 * (*vi++);
292c733ed4SBarry Smith       x[oidx] -= v[0] * s1 + v[1] * s2;
302c733ed4SBarry Smith       x[oidx + 1] -= v[2] * s1 + v[3] * s2;
312c733ed4SBarry Smith       v += 4;
322c733ed4SBarry Smith     }
33*9371c9d4SSatish Balay     x[idx]     = s1;
34*9371c9d4SSatish Balay     x[1 + idx] = s2;
352c733ed4SBarry Smith     idx += 2;
362c733ed4SBarry Smith   }
372c733ed4SBarry Smith   /* backward solve the L^T */
382c733ed4SBarry Smith   for (i = n - 1; i >= 0; i--) {
392c733ed4SBarry Smith     v   = aa + 4 * diag[i] - 4;
402c733ed4SBarry Smith     vi  = aj + diag[i] - 1;
412c733ed4SBarry Smith     nz  = diag[i] - ai[i];
422c733ed4SBarry Smith     idt = 2 * i;
43*9371c9d4SSatish Balay     s1  = x[idt];
44*9371c9d4SSatish Balay     s2  = x[1 + idt];
452c733ed4SBarry Smith     while (nz--) {
462c733ed4SBarry Smith       idx = 2 * (*vi--);
472c733ed4SBarry Smith       x[idx] -= v[0] * s1 + v[1] * s2;
482c733ed4SBarry Smith       x[idx + 1] -= v[2] * s1 + v[3] * s2;
492c733ed4SBarry Smith       v -= 4;
502c733ed4SBarry Smith     }
512c733ed4SBarry Smith   }
529566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx, &x));
539566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0 * 4.0 * (a->nz) - 2.0 * A->cmap->n));
542c733ed4SBarry Smith   PetscFunctionReturn(0);
552c733ed4SBarry Smith }
562c733ed4SBarry Smith 
57*9371c9d4SSatish Balay PetscErrorCode MatSolveTranspose_SeqBAIJ_2_NaturalOrdering(Mat A, Vec bb, Vec xx) {
582c733ed4SBarry Smith   Mat_SeqBAIJ     *a = (Mat_SeqBAIJ *)A->data;
592c733ed4SBarry Smith   const PetscInt   n = a->mbs, *vi, *ai = a->i, *aj = a->j, *diag = a->diag;
602c733ed4SBarry Smith   PetscInt         nz, idx, idt, j, i, oidx;
612c733ed4SBarry Smith   const PetscInt   bs = A->rmap->bs, bs2 = a->bs2;
622c733ed4SBarry Smith   const MatScalar *aa = a->a, *v;
632c733ed4SBarry Smith   PetscScalar      s1, s2, x1, x2, *x;
642c733ed4SBarry Smith 
652c733ed4SBarry Smith   PetscFunctionBegin;
669566063dSJacob Faibussowitsch   PetscCall(VecCopy(bb, xx));
679566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx, &x));
682c733ed4SBarry Smith 
692c733ed4SBarry Smith   /* forward solve the U^T */
702c733ed4SBarry Smith   idx = 0;
712c733ed4SBarry Smith   for (i = 0; i < n; i++) {
722c733ed4SBarry Smith     v  = aa + bs2 * diag[i];
732c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
74*9371c9d4SSatish Balay     x1 = x[idx];
75*9371c9d4SSatish Balay     x2 = x[1 + idx];
762c733ed4SBarry Smith     s1 = v[0] * x1 + v[1] * x2;
772c733ed4SBarry Smith     s2 = v[2] * x1 + v[3] * x2;
782c733ed4SBarry Smith     v -= bs2;
792c733ed4SBarry Smith 
802c733ed4SBarry Smith     vi = aj + diag[i] - 1;
812c733ed4SBarry Smith     nz = diag[i] - diag[i + 1] - 1;
822c733ed4SBarry Smith     for (j = 0; j > -nz; j--) {
832c733ed4SBarry Smith       oidx = bs * vi[j];
842c733ed4SBarry Smith       x[oidx] -= v[0] * s1 + v[1] * s2;
852c733ed4SBarry Smith       x[oidx + 1] -= v[2] * s1 + v[3] * s2;
862c733ed4SBarry Smith       v -= bs2;
872c733ed4SBarry Smith     }
88*9371c9d4SSatish Balay     x[idx]     = s1;
89*9371c9d4SSatish Balay     x[1 + idx] = s2;
902c733ed4SBarry Smith     idx += bs;
912c733ed4SBarry Smith   }
922c733ed4SBarry Smith   /* backward solve the L^T */
932c733ed4SBarry Smith   for (i = n - 1; i >= 0; i--) {
942c733ed4SBarry Smith     v   = aa + bs2 * ai[i];
952c733ed4SBarry Smith     vi  = aj + ai[i];
962c733ed4SBarry Smith     nz  = ai[i + 1] - ai[i];
972c733ed4SBarry Smith     idt = bs * i;
98*9371c9d4SSatish Balay     s1  = x[idt];
99*9371c9d4SSatish Balay     s2  = x[1 + idt];
1002c733ed4SBarry Smith     for (j = 0; j < nz; j++) {
1012c733ed4SBarry Smith       idx = bs * vi[j];
1022c733ed4SBarry Smith       x[idx] -= v[0] * s1 + v[1] * s2;
1032c733ed4SBarry Smith       x[idx + 1] -= v[2] * s1 + v[3] * s2;
1042c733ed4SBarry Smith       v += bs2;
1052c733ed4SBarry Smith     }
1062c733ed4SBarry Smith   }
1079566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx, &x));
1089566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0 * bs2 * (a->nz) - bs * A->cmap->n));
1092c733ed4SBarry Smith   PetscFunctionReturn(0);
1102c733ed4SBarry Smith }
111