xref: /petsc/src/mat/impls/baij/seq/baijsolvtran1.c (revision 9566063d113dddea24716c546802770db7481bc0)
12c733ed4SBarry Smith #include <../src/mat/impls/baij/seq/baij.h>
22c733ed4SBarry Smith #include <petsc/private/kernels/blockinvert.h>
32c733ed4SBarry Smith 
42c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_1(Mat A,Vec bb,Vec xx)
52c733ed4SBarry Smith {
62c733ed4SBarry Smith   Mat_SeqBAIJ       *a    = (Mat_SeqBAIJ*)A->data;
72c733ed4SBarry Smith   IS                iscol = a->col,isrow = a->row;
82c733ed4SBarry Smith   const PetscInt    *rout,*cout,*r,*c,*adiag = a->diag,*ai = a->i,*aj = a->j,*vi;
92c733ed4SBarry Smith   PetscInt          i,n = a->mbs,j;
102c733ed4SBarry Smith   PetscInt          nz;
112c733ed4SBarry Smith   PetscScalar       *x,*tmp,s1;
122c733ed4SBarry Smith   const MatScalar   *aa = a->a,*v;
132c733ed4SBarry Smith   const PetscScalar *b;
142c733ed4SBarry Smith 
152c733ed4SBarry Smith   PetscFunctionBegin;
16*9566063dSJacob Faibussowitsch   PetscCall(VecGetArrayRead(bb,&b));
17*9566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx,&x));
182c733ed4SBarry Smith   tmp  = a->solve_work;
192c733ed4SBarry Smith 
20*9566063dSJacob Faibussowitsch   PetscCall(ISGetIndices(isrow,&rout)); r = rout;
21*9566063dSJacob Faibussowitsch   PetscCall(ISGetIndices(iscol,&cout)); c = cout;
222c733ed4SBarry Smith 
232c733ed4SBarry Smith   /* copy the b into temp work space according to permutation */
242c733ed4SBarry Smith   for (i=0; i<n; i++) tmp[i] = b[c[i]];
252c733ed4SBarry Smith 
262c733ed4SBarry Smith   /* forward solve the U^T */
272c733ed4SBarry Smith   for (i=0; i<n; i++) {
282c733ed4SBarry Smith     v   = aa + adiag[i+1] + 1;
292c733ed4SBarry Smith     vi  = aj + adiag[i+1] + 1;
302c733ed4SBarry Smith     nz  = adiag[i] - adiag[i+1] - 1;
312c733ed4SBarry Smith     s1  = tmp[i];
322c733ed4SBarry Smith     s1 *= v[nz];  /* multiply by inverse of diagonal entry */
332c733ed4SBarry Smith     for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j];
342c733ed4SBarry Smith     tmp[i] = s1;
352c733ed4SBarry Smith   }
362c733ed4SBarry Smith 
372c733ed4SBarry Smith   /* backward solve the L^T */
382c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
392c733ed4SBarry Smith     v  = aa + ai[i];
402c733ed4SBarry Smith     vi = aj + ai[i];
412c733ed4SBarry Smith     nz = ai[i+1] - ai[i];
422c733ed4SBarry Smith     s1 = tmp[i];
432c733ed4SBarry Smith     for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j];
442c733ed4SBarry Smith   }
452c733ed4SBarry Smith 
462c733ed4SBarry Smith   /* copy tmp into x according to permutation */
472c733ed4SBarry Smith   for (i=0; i<n; i++) x[r[i]] = tmp[i];
482c733ed4SBarry Smith 
49*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(isrow,&rout));
50*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(iscol,&cout));
51*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb,&b));
52*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
532c733ed4SBarry Smith 
54*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*a->nz-A->cmap->n));
552c733ed4SBarry Smith   PetscFunctionReturn(0);
562c733ed4SBarry Smith }
572c733ed4SBarry Smith 
582c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_1_inplace(Mat A,Vec bb,Vec xx)
592c733ed4SBarry Smith {
602c733ed4SBarry Smith   Mat_SeqBAIJ       *a   =(Mat_SeqBAIJ*)A->data;
612c733ed4SBarry Smith   IS                iscol=a->col,isrow=a->row;
622c733ed4SBarry Smith   const PetscInt    *r,*c,*rout,*cout;
632c733ed4SBarry Smith   const PetscInt    *diag=a->diag,n=a->mbs,*vi,*ai=a->i,*aj=a->j;
642c733ed4SBarry Smith   PetscInt          i,nz;
652c733ed4SBarry Smith   const MatScalar   *aa=a->a,*v;
662c733ed4SBarry Smith   PetscScalar       s1,*x,*t;
672c733ed4SBarry Smith   const PetscScalar *b;
682c733ed4SBarry Smith 
692c733ed4SBarry Smith   PetscFunctionBegin;
70*9566063dSJacob Faibussowitsch   PetscCall(VecGetArrayRead(bb,&b));
71*9566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx,&x));
722c733ed4SBarry Smith   t    = a->solve_work;
732c733ed4SBarry Smith 
74*9566063dSJacob Faibussowitsch   PetscCall(ISGetIndices(isrow,&rout)); r = rout;
75*9566063dSJacob Faibussowitsch   PetscCall(ISGetIndices(iscol,&cout)); c = cout;
762c733ed4SBarry Smith 
772c733ed4SBarry Smith   /* copy the b into temp work space according to permutation */
782c733ed4SBarry Smith   for (i=0; i<n; i++) t[i] = b[c[i]];
792c733ed4SBarry Smith 
802c733ed4SBarry Smith   /* forward solve the U^T */
812c733ed4SBarry Smith   for (i=0; i<n; i++) {
822c733ed4SBarry Smith 
832c733ed4SBarry Smith     v = aa + diag[i];
842c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
852c733ed4SBarry Smith     s1 = (*v++)*t[i];
862c733ed4SBarry Smith     vi = aj + diag[i] + 1;
872c733ed4SBarry Smith     nz = ai[i+1] - diag[i] - 1;
882c733ed4SBarry Smith     while (nz--) {
892c733ed4SBarry Smith       t[*vi++] -= (*v++)*s1;
902c733ed4SBarry Smith     }
912c733ed4SBarry Smith     t[i] = s1;
922c733ed4SBarry Smith   }
932c733ed4SBarry Smith   /* backward solve the L^T */
942c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
952c733ed4SBarry Smith     v  = aa + diag[i] - 1;
962c733ed4SBarry Smith     vi = aj + diag[i] - 1;
972c733ed4SBarry Smith     nz = diag[i] - ai[i];
982c733ed4SBarry Smith     s1 = t[i];
992c733ed4SBarry Smith     while (nz--) {
1002c733ed4SBarry Smith       t[*vi--] -=  (*v--)*s1;
1012c733ed4SBarry Smith     }
1022c733ed4SBarry Smith   }
1032c733ed4SBarry Smith 
1042c733ed4SBarry Smith   /* copy t into x according to permutation */
1052c733ed4SBarry Smith   for (i=0; i<n; i++) x[r[i]] = t[i];
1062c733ed4SBarry Smith 
107*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(isrow,&rout));
108*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(iscol,&cout));
109*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb,&b));
110*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
111*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*(a->nz) - A->cmap->n));
1122c733ed4SBarry Smith   PetscFunctionReturn(0);
1132c733ed4SBarry Smith }
114