xref: /petsc/src/mat/impls/baij/seq/baijsolvtrannat2.c (revision 9566063d113dddea24716c546802770db7481bc0)
12c733ed4SBarry Smith #include <../src/mat/impls/baij/seq/baij.h>
22c733ed4SBarry Smith 
32c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_2_NaturalOrdering_inplace(Mat A,Vec bb,Vec xx)
42c733ed4SBarry Smith {
52c733ed4SBarry Smith   Mat_SeqBAIJ     *a=(Mat_SeqBAIJ*)A->data;
62c733ed4SBarry Smith   PetscInt        i,nz,idx,idt,oidx;
72c733ed4SBarry Smith   const PetscInt  *diag = a->diag,*vi,n=a->mbs,*ai=a->i,*aj=a->j;
82c733ed4SBarry Smith   const MatScalar *aa   =a->a,*v;
92c733ed4SBarry Smith   PetscScalar     s1,s2,x1,x2,*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 + 4*diag[i];
202c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
212c733ed4SBarry Smith     x1 = x[idx];   x2 = x[1+idx];
222c733ed4SBarry Smith     s1 = v[0]*x1  +  v[1]*x2;
232c733ed4SBarry Smith     s2 = v[2]*x1  +  v[3]*x2;
242c733ed4SBarry Smith     v += 4;
252c733ed4SBarry Smith 
262c733ed4SBarry Smith     vi = aj + diag[i] + 1;
272c733ed4SBarry Smith     nz = ai[i+1] - diag[i] - 1;
282c733ed4SBarry Smith     while (nz--) {
292c733ed4SBarry Smith       oidx       = 2*(*vi++);
302c733ed4SBarry Smith       x[oidx]   -= v[0]*s1  +  v[1]*s2;
312c733ed4SBarry Smith       x[oidx+1] -= v[2]*s1  +  v[3]*s2;
322c733ed4SBarry Smith       v         += 4;
332c733ed4SBarry Smith     }
342c733ed4SBarry Smith     x[idx] = s1;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;
432c733ed4SBarry Smith     s1  = x[idt];  s2 = x[1+idt];
442c733ed4SBarry Smith     while (nz--) {
452c733ed4SBarry Smith       idx       = 2*(*vi--);
462c733ed4SBarry Smith       x[idx]   -=  v[0]*s1 +  v[1]*s2;
472c733ed4SBarry Smith       x[idx+1] -=  v[2]*s1 +  v[3]*s2;
482c733ed4SBarry Smith       v        -= 4;
492c733ed4SBarry Smith     }
502c733ed4SBarry Smith   }
51*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
52*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*4.0*(a->nz) - 2.0*A->cmap->n));
532c733ed4SBarry Smith   PetscFunctionReturn(0);
542c733ed4SBarry Smith }
552c733ed4SBarry Smith 
562c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_2_NaturalOrdering(Mat A,Vec bb,Vec xx)
572c733ed4SBarry Smith {
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;
66*9566063dSJacob Faibussowitsch   PetscCall(VecCopy(bb,xx));
67*9566063dSJacob 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 */
742c733ed4SBarry Smith     x1 = x[idx];   x2 = x[1+idx];
752c733ed4SBarry Smith     s1 = v[0]*x1  +  v[1]*x2;
762c733ed4SBarry Smith     s2 = v[2]*x1  +  v[3]*x2;
772c733ed4SBarry Smith     v -= bs2;
782c733ed4SBarry Smith 
792c733ed4SBarry Smith     vi = aj + diag[i] - 1;
802c733ed4SBarry Smith     nz = diag[i] - diag[i+1] - 1;
812c733ed4SBarry Smith     for (j=0; j>-nz; j--) {
822c733ed4SBarry Smith       oidx       = bs*vi[j];
832c733ed4SBarry Smith       x[oidx]   -= v[0]*s1  +  v[1]*s2;
842c733ed4SBarry Smith       x[oidx+1] -= v[2]*s1  +  v[3]*s2;
852c733ed4SBarry Smith       v         -= bs2;
862c733ed4SBarry Smith     }
872c733ed4SBarry Smith     x[idx] = s1;x[1+idx] = s2;
882c733ed4SBarry Smith     idx   += bs;
892c733ed4SBarry Smith   }
902c733ed4SBarry Smith   /* backward solve the L^T */
912c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
922c733ed4SBarry Smith     v   = aa + bs2*ai[i];
932c733ed4SBarry Smith     vi  = aj + ai[i];
942c733ed4SBarry Smith     nz  = ai[i+1] - ai[i];
952c733ed4SBarry Smith     idt = bs*i;
962c733ed4SBarry Smith     s1  = x[idt];  s2 = x[1+idt];
972c733ed4SBarry Smith     for (j=0; j<nz; j++) {
982c733ed4SBarry Smith       idx       = bs*vi[j];
992c733ed4SBarry Smith       x[idx]   -=  v[0]*s1 +  v[1]*s2;
1002c733ed4SBarry Smith       x[idx+1] -=  v[2]*s1 +  v[3]*s2;
1012c733ed4SBarry Smith       v        += bs2;
1022c733ed4SBarry Smith     }
1032c733ed4SBarry Smith   }
104*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
105*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*bs2*(a->nz) - bs*A->cmap->n));
1062c733ed4SBarry Smith   PetscFunctionReturn(0);
1072c733ed4SBarry Smith }
108