xref: /petsc/src/mat/impls/baij/seq/baijsolvtrannat5.c (revision 9566063d113dddea24716c546802770db7481bc0)
12c733ed4SBarry Smith #include <../src/mat/impls/baij/seq/baij.h>
22c733ed4SBarry Smith 
32c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_5_NaturalOrdering_inplace(Mat A,Vec bb,Vec xx)
42c733ed4SBarry Smith {
52c733ed4SBarry Smith   Mat_SeqBAIJ     *a=(Mat_SeqBAIJ*)A->data;
62c733ed4SBarry Smith   const PetscInt  *diag=a->diag,n=a->mbs,*vi,*ai=a->i,*aj=a->j;
72c733ed4SBarry Smith   PetscInt        i,nz,idx,idt,oidx;
82c733ed4SBarry Smith   const MatScalar *aa=a->a,*v;
92c733ed4SBarry Smith   PetscScalar     s1,s2,s3,s4,s5,x1,x2,x3,x4,x5,*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 + 25*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]; x4 = x[3+idx]; x5 = x[4+idx];
222c733ed4SBarry Smith     s1 = v[0]*x1  +  v[1]*x2 +  v[2]*x3 +  v[3]*x4 +  v[4]*x5;
232c733ed4SBarry Smith     s2 = v[5]*x1  +  v[6]*x2 +  v[7]*x3 +  v[8]*x4 +  v[9]*x5;
242c733ed4SBarry Smith     s3 = v[10]*x1 + v[11]*x2 + v[12]*x3 + v[13]*x4 + v[14]*x5;
252c733ed4SBarry Smith     s4 = v[15]*x1 + v[16]*x2 + v[17]*x3 + v[18]*x4 + v[19]*x5;
262c733ed4SBarry Smith     s5 = v[20]*x1 + v[21]*x2 + v[22]*x3 + v[23]*x4 + v[24]*x5;
272c733ed4SBarry Smith     v += 25;
282c733ed4SBarry Smith 
292c733ed4SBarry Smith     vi = aj + diag[i] + 1;
302c733ed4SBarry Smith     nz = ai[i+1] - diag[i] - 1;
312c733ed4SBarry Smith     while (nz--) {
322c733ed4SBarry Smith       oidx       = 5*(*vi++);
332c733ed4SBarry Smith       x[oidx]   -= v[0]*s1  +  v[1]*s2 +  v[2]*s3 +  v[3]*s4 +  v[4]*s5;
342c733ed4SBarry Smith       x[oidx+1] -= v[5]*s1  +  v[6]*s2 +  v[7]*s3 +  v[8]*s4 +  v[9]*s5;
352c733ed4SBarry Smith       x[oidx+2] -= v[10]*s1 + v[11]*s2 + v[12]*s3 + v[13]*s4 + v[14]*s5;
362c733ed4SBarry Smith       x[oidx+3] -= v[15]*s1 + v[16]*s2 + v[17]*s3 + v[18]*s4 + v[19]*s5;
372c733ed4SBarry Smith       x[oidx+4] -= v[20]*s1 + v[21]*s2 + v[22]*s3 + v[23]*s4 + v[24]*s5;
382c733ed4SBarry Smith       v         += 25;
392c733ed4SBarry Smith     }
402c733ed4SBarry Smith     x[idx] = s1;x[1+idx] = s2; x[2+idx] = s3;x[3+idx] = s4; x[4+idx] = s5;
412c733ed4SBarry Smith     idx   += 5;
422c733ed4SBarry Smith   }
432c733ed4SBarry Smith   /* backward solve the L^T */
442c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
452c733ed4SBarry Smith     v   = aa + 25*diag[i] - 25;
462c733ed4SBarry Smith     vi  = aj + diag[i] - 1;
472c733ed4SBarry Smith     nz  = diag[i] - ai[i];
482c733ed4SBarry Smith     idt = 5*i;
492c733ed4SBarry Smith     s1  = x[idt];  s2 = x[1+idt]; s3 = x[2+idt];s4 = x[3+idt]; s5 = x[4+idt];
502c733ed4SBarry Smith     while (nz--) {
512c733ed4SBarry Smith       idx       = 5*(*vi--);
522c733ed4SBarry Smith       x[idx]   -=  v[0]*s1 +  v[1]*s2 +  v[2]*s3 +  v[3]*s4 +  v[4]*s5;
532c733ed4SBarry Smith       x[idx+1] -=  v[5]*s1 +  v[6]*s2 +  v[7]*s3 +  v[8]*s4 +  v[9]*s5;
542c733ed4SBarry Smith       x[idx+2] -= v[10]*s1 + v[11]*s2 + v[12]*s3 + v[13]*s4 + v[14]*s5;
552c733ed4SBarry Smith       x[idx+3] -= v[15]*s1 + v[16]*s2 + v[17]*s3 + v[18]*s4 + v[19]*s5;
562c733ed4SBarry Smith       x[idx+4] -= v[20]*s1 + v[21]*s2 + v[22]*s3 + v[23]*s4 + v[24]*s5;
572c733ed4SBarry Smith       v        -= 25;
582c733ed4SBarry Smith     }
592c733ed4SBarry Smith   }
60*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
61*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*25*(a->nz) - 5.0*A->cmap->n));
622c733ed4SBarry Smith   PetscFunctionReturn(0);
632c733ed4SBarry Smith }
642c733ed4SBarry Smith 
652c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_5_NaturalOrdering(Mat A,Vec bb,Vec xx)
662c733ed4SBarry Smith {
672c733ed4SBarry Smith   Mat_SeqBAIJ     *a=(Mat_SeqBAIJ*)A->data;
682c733ed4SBarry Smith   const PetscInt  n=a->mbs,*vi,*ai=a->i,*aj=a->j,*diag=a->diag;
692c733ed4SBarry Smith   PetscInt        nz,idx,idt,j,i,oidx;
702c733ed4SBarry Smith   const PetscInt  bs =A->rmap->bs,bs2=a->bs2;
712c733ed4SBarry Smith   const MatScalar *aa=a->a,*v;
722c733ed4SBarry Smith   PetscScalar     s1,s2,s3,s4,s5,x1,x2,x3,x4,x5,*x;
732c733ed4SBarry Smith 
742c733ed4SBarry Smith   PetscFunctionBegin;
75*9566063dSJacob Faibussowitsch   PetscCall(VecCopy(bb,xx));
76*9566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx,&x));
772c733ed4SBarry Smith 
782c733ed4SBarry Smith   /* forward solve the U^T */
792c733ed4SBarry Smith   idx = 0;
802c733ed4SBarry Smith   for (i=0; i<n; i++) {
812c733ed4SBarry Smith     v = aa + bs2*diag[i];
822c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
832c733ed4SBarry Smith     x1 = x[idx];   x2 = x[1+idx];  x3 = x[2+idx];  x4 = x[3+idx];
842c733ed4SBarry Smith     x5 = x[4+idx];
852c733ed4SBarry Smith     s1 =  v[0]*x1   +  v[1]*x2   + v[2]*x3   + v[3]*x4   + v[4]*x5;
862c733ed4SBarry Smith     s2 =  v[5]*x1   +  v[6]*x2   + v[7]*x3   + v[8]*x4   + v[9]*x5;
872c733ed4SBarry Smith     s3 =  v[10]*x1  +  v[11]*x2  + v[12]*x3  + v[13]*x4  + v[14]*x5;
882c733ed4SBarry Smith     s4 =  v[15]*x1  +  v[16]*x2  + v[17]*x3  + v[18]*x4  + v[19]*x5;
892c733ed4SBarry Smith     s5 =  v[20]*x1  +  v[21]*x2  + v[22]*x3  + v[23]*x4   + v[24]*x5;
902c733ed4SBarry Smith     v -= bs2;
912c733ed4SBarry Smith 
922c733ed4SBarry Smith     vi = aj + diag[i] - 1;
932c733ed4SBarry Smith     nz = diag[i] - diag[i+1] - 1;
942c733ed4SBarry Smith     for (j=0; j>-nz; j--) {
952c733ed4SBarry Smith       oidx       = bs*vi[j];
962c733ed4SBarry Smith       x[oidx]   -=  v[0]*s1   +  v[1]*s2   + v[2]*s3   + v[3]*s4   + v[4]*s5;
972c733ed4SBarry Smith       x[oidx+1] -=  v[5]*s1   +  v[6]*s2   + v[7]*s3   + v[8]*s4   + v[9]*s5;
982c733ed4SBarry Smith       x[oidx+2] -=  v[10]*s1  +  v[11]*s2  + v[12]*s3  + v[13]*s4  + v[14]*s5;
992c733ed4SBarry Smith       x[oidx+3] -=  v[15]*s1  +  v[16]*s2  + v[17]*s3  + v[18]*s4  + v[19]*s5;
1002c733ed4SBarry Smith       x[oidx+4] -=  v[20]*s1  +  v[21]*s2  + v[22]*s3  + v[23]*s4   + v[24]*s5;
1012c733ed4SBarry Smith       v         -= bs2;
1022c733ed4SBarry Smith     }
1032c733ed4SBarry Smith     x[idx] = s1;x[1+idx] = s2;  x[2+idx] = s3;  x[3+idx] = s4; x[4+idx] = s5;
1042c733ed4SBarry Smith     idx   += bs;
1052c733ed4SBarry Smith   }
1062c733ed4SBarry Smith   /* backward solve the L^T */
1072c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
1082c733ed4SBarry Smith     v   = aa + bs2*ai[i];
1092c733ed4SBarry Smith     vi  = aj + ai[i];
1102c733ed4SBarry Smith     nz  = ai[i+1] - ai[i];
1112c733ed4SBarry Smith     idt = bs*i;
1122c733ed4SBarry Smith     s1  = x[idt];  s2 = x[1+idt];  s3 = x[2+idt];  s4 = x[3+idt];  s5 = x[4+idt];
1132c733ed4SBarry Smith     for (j=0; j<nz; j++) {
1142c733ed4SBarry Smith       idx       = bs*vi[j];
1152c733ed4SBarry Smith       x[idx]   -=  v[0]*s1   +  v[1]*s2   + v[2]*s3   + v[3]*s4   + v[4]*s5;
1162c733ed4SBarry Smith       x[idx+1] -=  v[5]*s1   +  v[6]*s2   + v[7]*s3   + v[8]*s4   + v[9]*s5;
1172c733ed4SBarry Smith       x[idx+2] -=  v[10]*s1  +  v[11]*s2  + v[12]*s3  + v[13]*s4  + v[14]*s5;
1182c733ed4SBarry Smith       x[idx+3] -=  v[15]*s1  +  v[16]*s2  + v[17]*s3  + v[18]*s4  + v[19]*s5;
1192c733ed4SBarry Smith       x[idx+4] -=  v[20]*s1  +  v[21]*s2  + v[22]*s3  + v[23]*s4   + v[24]*s5;
1202c733ed4SBarry Smith       v        += bs2;
1212c733ed4SBarry Smith     }
1222c733ed4SBarry Smith   }
123*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
124*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*bs2*(a->nz) - bs*A->cmap->n));
1252c733ed4SBarry Smith   PetscFunctionReturn(0);
1262c733ed4SBarry Smith }
127