xref: /petsc/src/mat/impls/baij/seq/baijsolvtran2.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_2_inplace(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    *r,*c,*rout,*cout;
92c733ed4SBarry Smith   const PetscInt    *diag=a->diag,n=a->mbs,*vi,*ai=a->i,*aj=a->j;
102c733ed4SBarry Smith   PetscInt          i,nz,idx,idt,ii,ic,ir,oidx;
112c733ed4SBarry Smith   const MatScalar   *aa=a->a,*v;
122c733ed4SBarry Smith   PetscScalar       s1,s2,x1,x2,*x,*t;
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   t    = 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   ii = 0;
252c733ed4SBarry Smith   for (i=0; i<n; i++) {
262c733ed4SBarry Smith     ic      = 2*c[i];
272c733ed4SBarry Smith     t[ii]   = b[ic];
282c733ed4SBarry Smith     t[ii+1] = b[ic+1];
292c733ed4SBarry Smith     ii     += 2;
302c733ed4SBarry Smith   }
312c733ed4SBarry Smith 
322c733ed4SBarry Smith   /* forward solve the U^T */
332c733ed4SBarry Smith   idx = 0;
342c733ed4SBarry Smith   for (i=0; i<n; i++) {
352c733ed4SBarry Smith 
362c733ed4SBarry Smith     v = aa + 4*diag[i];
372c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
382c733ed4SBarry Smith     x1 = t[idx];   x2 = t[1+idx];
392c733ed4SBarry Smith     s1 = v[0]*x1  +  v[1]*x2;
402c733ed4SBarry Smith     s2 = v[2]*x1  +  v[3]*x2;
412c733ed4SBarry Smith     v += 4;
422c733ed4SBarry Smith 
432c733ed4SBarry Smith     vi = aj + diag[i] + 1;
442c733ed4SBarry Smith     nz = ai[i+1] - diag[i] - 1;
452c733ed4SBarry Smith     while (nz--) {
462c733ed4SBarry Smith       oidx       = 2*(*vi++);
472c733ed4SBarry Smith       t[oidx]   -= v[0]*s1  +  v[1]*s2;
482c733ed4SBarry Smith       t[oidx+1] -= v[2]*s1  +  v[3]*s2;
492c733ed4SBarry Smith       v         += 4;
502c733ed4SBarry Smith     }
512c733ed4SBarry Smith     t[idx] = s1;t[1+idx] = s2;
522c733ed4SBarry Smith     idx   += 2;
532c733ed4SBarry Smith   }
542c733ed4SBarry Smith   /* backward solve the L^T */
552c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
562c733ed4SBarry Smith     v   = aa + 4*diag[i] - 4;
572c733ed4SBarry Smith     vi  = aj + diag[i] - 1;
582c733ed4SBarry Smith     nz  = diag[i] - ai[i];
592c733ed4SBarry Smith     idt = 2*i;
602c733ed4SBarry Smith     s1  = t[idt];  s2 = t[1+idt];
612c733ed4SBarry Smith     while (nz--) {
622c733ed4SBarry Smith       idx       = 2*(*vi--);
632c733ed4SBarry Smith       t[idx]   -=  v[0]*s1 +  v[1]*s2;
642c733ed4SBarry Smith       t[idx+1] -=  v[2]*s1 +  v[3]*s2;
652c733ed4SBarry Smith       v        -= 4;
662c733ed4SBarry Smith     }
672c733ed4SBarry Smith   }
682c733ed4SBarry Smith 
692c733ed4SBarry Smith   /* copy t into x according to permutation */
702c733ed4SBarry Smith   ii = 0;
712c733ed4SBarry Smith   for (i=0; i<n; i++) {
722c733ed4SBarry Smith     ir      = 2*r[i];
732c733ed4SBarry Smith     x[ir]   = t[ii];
742c733ed4SBarry Smith     x[ir+1] = t[ii+1];
752c733ed4SBarry Smith     ii     += 2;
762c733ed4SBarry Smith   }
772c733ed4SBarry Smith 
78*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(isrow,&rout));
79*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(iscol,&cout));
80*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb,&b));
81*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
82*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*4*(a->nz) - 2.0*A->cmap->n));
832c733ed4SBarry Smith   PetscFunctionReturn(0);
842c733ed4SBarry Smith }
852c733ed4SBarry Smith 
862c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_2(Mat A,Vec bb,Vec xx)
872c733ed4SBarry Smith {
882c733ed4SBarry Smith   Mat_SeqBAIJ       *a=(Mat_SeqBAIJ*)A->data;
892c733ed4SBarry Smith   IS                iscol=a->col,isrow=a->row;
902c733ed4SBarry Smith   const PetscInt    n    =a->mbs,*vi,*ai=a->i,*aj=a->j,*diag=a->diag;
912c733ed4SBarry Smith   const PetscInt    *r,*c,*rout,*cout;
922c733ed4SBarry Smith   PetscInt          nz,idx,idt,j,i,oidx,ii,ic,ir;
932c733ed4SBarry Smith   const PetscInt    bs =A->rmap->bs,bs2=a->bs2;
942c733ed4SBarry Smith   const MatScalar   *aa=a->a,*v;
952c733ed4SBarry Smith   PetscScalar       s1,s2,x1,x2,*x,*t;
962c733ed4SBarry Smith   const PetscScalar *b;
972c733ed4SBarry Smith 
982c733ed4SBarry Smith   PetscFunctionBegin;
99*9566063dSJacob Faibussowitsch   PetscCall(VecGetArrayRead(bb,&b));
100*9566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx,&x));
1012c733ed4SBarry Smith   t    = a->solve_work;
1022c733ed4SBarry Smith 
103*9566063dSJacob Faibussowitsch   PetscCall(ISGetIndices(isrow,&rout)); r = rout;
104*9566063dSJacob Faibussowitsch   PetscCall(ISGetIndices(iscol,&cout)); c = cout;
1052c733ed4SBarry Smith 
1062c733ed4SBarry Smith   /* copy b into temp work space according to permutation */
1072c733ed4SBarry Smith   for (i=0; i<n; i++) {
1082c733ed4SBarry Smith     ii    = bs*i; ic = bs*c[i];
1092c733ed4SBarry Smith     t[ii] = b[ic]; t[ii+1] = b[ic+1];
1102c733ed4SBarry Smith   }
1112c733ed4SBarry Smith 
1122c733ed4SBarry Smith   /* forward solve the U^T */
1132c733ed4SBarry Smith   idx = 0;
1142c733ed4SBarry Smith   for (i=0; i<n; i++) {
1152c733ed4SBarry Smith     v = aa + bs2*diag[i];
1162c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
1172c733ed4SBarry Smith     x1 = t[idx];   x2 = t[1+idx];
1182c733ed4SBarry Smith     s1 = v[0]*x1  +  v[1]*x2;
1192c733ed4SBarry Smith     s2 = v[2]*x1  +  v[3]*x2;
1202c733ed4SBarry Smith     v -= bs2;
1212c733ed4SBarry Smith 
1222c733ed4SBarry Smith     vi = aj + diag[i] - 1;
1232c733ed4SBarry Smith     nz = diag[i] - diag[i+1] - 1;
1242c733ed4SBarry Smith     for (j=0; j>-nz; j--) {
1252c733ed4SBarry Smith       oidx       = bs*vi[j];
1262c733ed4SBarry Smith       t[oidx]   -= v[0]*s1  +  v[1]*s2;
1272c733ed4SBarry Smith       t[oidx+1] -= v[2]*s1  +  v[3]*s2;
1282c733ed4SBarry Smith       v         -= bs2;
1292c733ed4SBarry Smith     }
1302c733ed4SBarry Smith     t[idx] = s1;t[1+idx] = s2;
1312c733ed4SBarry Smith     idx   += bs;
1322c733ed4SBarry Smith   }
1332c733ed4SBarry Smith   /* backward solve the L^T */
1342c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
1352c733ed4SBarry Smith     v   = aa + bs2*ai[i];
1362c733ed4SBarry Smith     vi  = aj + ai[i];
1372c733ed4SBarry Smith     nz  = ai[i+1] - ai[i];
1382c733ed4SBarry Smith     idt = bs*i;
1392c733ed4SBarry Smith     s1  = t[idt];  s2 = t[1+idt];
1402c733ed4SBarry Smith     for (j=0; j<nz; j++) {
1412c733ed4SBarry Smith       idx       = bs*vi[j];
1422c733ed4SBarry Smith       t[idx]   -=  v[0]*s1 +  v[1]*s2;
1432c733ed4SBarry Smith       t[idx+1] -=  v[2]*s1 +  v[3]*s2;
1442c733ed4SBarry Smith       v        += bs2;
1452c733ed4SBarry Smith     }
1462c733ed4SBarry Smith   }
1472c733ed4SBarry Smith 
1482c733ed4SBarry Smith   /* copy t into x according to permutation */
1492c733ed4SBarry Smith   for (i=0; i<n; i++) {
1502c733ed4SBarry Smith     ii    = bs*i;  ir = bs*r[i];
1512c733ed4SBarry Smith     x[ir] = t[ii];  x[ir+1] = t[ii+1];
1522c733ed4SBarry Smith   }
1532c733ed4SBarry Smith 
154*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(isrow,&rout));
155*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(iscol,&cout));
156*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb,&b));
157*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
158*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*bs2*(a->nz) - bs*A->cmap->n));
1592c733ed4SBarry Smith   PetscFunctionReturn(0);
1602c733ed4SBarry Smith }
161