xref: /petsc/src/mat/impls/baij/seq/baijsolvtran4.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_4_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,s3,s4,x1,x2,x3,x4,*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      = 4*c[i];
272c733ed4SBarry Smith     t[ii]   = b[ic];
282c733ed4SBarry Smith     t[ii+1] = b[ic+1];
292c733ed4SBarry Smith     t[ii+2] = b[ic+2];
302c733ed4SBarry Smith     t[ii+3] = b[ic+3];
312c733ed4SBarry Smith     ii     += 4;
322c733ed4SBarry Smith   }
332c733ed4SBarry Smith 
342c733ed4SBarry Smith   /* forward solve the U^T */
352c733ed4SBarry Smith   idx = 0;
362c733ed4SBarry Smith   for (i=0; i<n; i++) {
372c733ed4SBarry Smith 
382c733ed4SBarry Smith     v = aa + 16*diag[i];
392c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
402c733ed4SBarry Smith     x1 = t[idx];   x2 = t[1+idx]; x3    = t[2+idx]; x4 = t[3+idx];
412c733ed4SBarry Smith     s1 = v[0]*x1  +  v[1]*x2 +  v[2]*x3 +  v[3]*x4;
422c733ed4SBarry Smith     s2 = v[4]*x1  +  v[5]*x2 +  v[6]*x3 +  v[7]*x4;
432c733ed4SBarry Smith     s3 = v[8]*x1  +  v[9]*x2 + v[10]*x3 + v[11]*x4;
442c733ed4SBarry Smith     s4 = v[12]*x1 + v[13]*x2 + v[14]*x3 + v[15]*x4;
452c733ed4SBarry Smith     v += 16;
462c733ed4SBarry Smith 
472c733ed4SBarry Smith     vi = aj + diag[i] + 1;
482c733ed4SBarry Smith     nz = ai[i+1] - diag[i] - 1;
492c733ed4SBarry Smith     while (nz--) {
502c733ed4SBarry Smith       oidx       = 4*(*vi++);
512c733ed4SBarry Smith       t[oidx]   -= v[0]*s1  +  v[1]*s2 +  v[2]*s3 +  v[3]*s4;
522c733ed4SBarry Smith       t[oidx+1] -= v[4]*s1  +  v[5]*s2 +  v[6]*s3 +  v[7]*s4;
532c733ed4SBarry Smith       t[oidx+2] -= v[8]*s1 + v[9]*s2 + v[10]*s3 + v[11]*s4;
542c733ed4SBarry Smith       t[oidx+3] -= v[12]*s1 + v[13]*s2 + v[14]*s3 + v[15]*s4;
552c733ed4SBarry Smith       v         += 16;
562c733ed4SBarry Smith     }
572c733ed4SBarry Smith     t[idx] = s1;t[1+idx] = s2; t[2+idx] = s3;t[3+idx] = s4;
582c733ed4SBarry Smith     idx   += 4;
592c733ed4SBarry Smith   }
602c733ed4SBarry Smith   /* backward solve the L^T */
612c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
622c733ed4SBarry Smith     v   = aa + 16*diag[i] - 16;
632c733ed4SBarry Smith     vi  = aj + diag[i] - 1;
642c733ed4SBarry Smith     nz  = diag[i] - ai[i];
652c733ed4SBarry Smith     idt = 4*i;
662c733ed4SBarry Smith     s1  = t[idt];  s2 = t[1+idt]; s3 = t[2+idt];s4 = t[3+idt];
672c733ed4SBarry Smith     while (nz--) {
682c733ed4SBarry Smith       idx       = 4*(*vi--);
692c733ed4SBarry Smith       t[idx]   -=  v[0]*s1 +  v[1]*s2 +  v[2]*s3 +  v[3]*s4;
702c733ed4SBarry Smith       t[idx+1] -=  v[4]*s1 +  v[5]*s2 +  v[6]*s3 +  v[7]*s4;
712c733ed4SBarry Smith       t[idx+2] -= v[8]*s1 + v[9]*s2 + v[10]*s3 + v[11]*s4;
722c733ed4SBarry Smith       t[idx+3] -= v[12]*s1 + v[13]*s2 + v[14]*s3 + v[15]*s4;
732c733ed4SBarry Smith       v        -= 16;
742c733ed4SBarry Smith     }
752c733ed4SBarry Smith   }
762c733ed4SBarry Smith 
772c733ed4SBarry Smith   /* copy t into x according to permutation */
782c733ed4SBarry Smith   ii = 0;
792c733ed4SBarry Smith   for (i=0; i<n; i++) {
802c733ed4SBarry Smith     ir      = 4*r[i];
812c733ed4SBarry Smith     x[ir]   = t[ii];
822c733ed4SBarry Smith     x[ir+1] = t[ii+1];
832c733ed4SBarry Smith     x[ir+2] = t[ii+2];
842c733ed4SBarry Smith     x[ir+3] = t[ii+3];
852c733ed4SBarry Smith     ii     += 4;
862c733ed4SBarry Smith   }
872c733ed4SBarry Smith 
88*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(isrow,&rout));
89*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(iscol,&cout));
90*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb,&b));
91*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
92*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*16*(a->nz) - 4.0*A->cmap->n));
932c733ed4SBarry Smith   PetscFunctionReturn(0);
942c733ed4SBarry Smith }
952c733ed4SBarry Smith 
962c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_4(Mat A,Vec bb,Vec xx)
972c733ed4SBarry Smith {
982c733ed4SBarry Smith   Mat_SeqBAIJ       *a=(Mat_SeqBAIJ*)A->data;
992c733ed4SBarry Smith   IS                iscol=a->col,isrow=a->row;
1002c733ed4SBarry Smith   const PetscInt    n    =a->mbs,*vi,*ai=a->i,*aj=a->j,*diag=a->diag;
1012c733ed4SBarry Smith   const PetscInt    *r,*c,*rout,*cout;
1022c733ed4SBarry Smith   PetscInt          nz,idx,idt,j,i,oidx,ii,ic,ir;
1032c733ed4SBarry Smith   const PetscInt    bs =A->rmap->bs,bs2=a->bs2;
1042c733ed4SBarry Smith   const MatScalar   *aa=a->a,*v;
1052c733ed4SBarry Smith   PetscScalar       s1,s2,s3,s4,x1,x2,x3,x4,*x,*t;
1062c733ed4SBarry Smith   const PetscScalar *b;
1072c733ed4SBarry Smith 
1082c733ed4SBarry Smith   PetscFunctionBegin;
109*9566063dSJacob Faibussowitsch   PetscCall(VecGetArrayRead(bb,&b));
110*9566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx,&x));
1112c733ed4SBarry Smith   t    = a->solve_work;
1122c733ed4SBarry Smith 
113*9566063dSJacob Faibussowitsch   PetscCall(ISGetIndices(isrow,&rout)); r = rout;
114*9566063dSJacob Faibussowitsch   PetscCall(ISGetIndices(iscol,&cout)); c = cout;
1152c733ed4SBarry Smith 
1162c733ed4SBarry Smith   /* copy b into temp work space according to permutation */
1172c733ed4SBarry Smith   for (i=0; i<n; i++) {
1182c733ed4SBarry Smith     ii    = bs*i; ic = bs*c[i];
1192c733ed4SBarry Smith     t[ii] = b[ic]; t[ii+1] = b[ic+1]; t[ii+2] = b[ic+2]; t[ii+3] = b[ic+3];
1202c733ed4SBarry Smith   }
1212c733ed4SBarry Smith 
1222c733ed4SBarry Smith   /* forward solve the U^T */
1232c733ed4SBarry Smith   idx = 0;
1242c733ed4SBarry Smith   for (i=0; i<n; i++) {
1252c733ed4SBarry Smith     v = aa + bs2*diag[i];
1262c733ed4SBarry Smith     /* multiply by the inverse of the block diagonal */
1272c733ed4SBarry Smith     x1 = t[idx];   x2 = t[1+idx]; x3    = t[2+idx];  x4 = t[3+idx];
1282c733ed4SBarry Smith     s1 = v[0]*x1  +  v[1]*x2 +  v[2]*x3 +  v[3]*x4;
1292c733ed4SBarry Smith     s2 = v[4]*x1  +  v[5]*x2 +  v[6]*x3 +  v[7]*x4;
1302c733ed4SBarry Smith     s3 = v[8]*x1  +  v[9]*x2 + v[10]*x3 + v[11]*x4;
1312c733ed4SBarry Smith     s4 = v[12]*x1 + v[13]*x2 + v[14]*x3 + v[15]*x4;
1322c733ed4SBarry Smith     v -= bs2;
1332c733ed4SBarry Smith 
1342c733ed4SBarry Smith     vi = aj + diag[i] - 1;
1352c733ed4SBarry Smith     nz = diag[i] - diag[i+1] - 1;
1362c733ed4SBarry Smith     for (j=0; j>-nz; j--) {
1372c733ed4SBarry Smith       oidx       = bs*vi[j];
1382c733ed4SBarry Smith       t[oidx]   -= v[0]*s1  +  v[1]*s2 +  v[2]*s3 +  v[3]*s4;
1392c733ed4SBarry Smith       t[oidx+1] -= v[4]*s1  +  v[5]*s2 +  v[6]*s3 +  v[7]*s4;
1402c733ed4SBarry Smith       t[oidx+2] -= v[8]*s1 + v[9]*s2 + v[10]*s3 + v[11]*s4;
1412c733ed4SBarry Smith       t[oidx+3] -= v[12]*s1 + v[13]*s2 + v[14]*s3 + v[15]*s4;
1422c733ed4SBarry Smith       v         -= bs2;
1432c733ed4SBarry Smith     }
1442c733ed4SBarry Smith     t[idx] = s1;t[1+idx] = s2;  t[2+idx] = s3;  t[3+idx] = s4;
1452c733ed4SBarry Smith     idx   += bs;
1462c733ed4SBarry Smith   }
1472c733ed4SBarry Smith   /* backward solve the L^T */
1482c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
1492c733ed4SBarry Smith     v   = aa + bs2*ai[i];
1502c733ed4SBarry Smith     vi  = aj + ai[i];
1512c733ed4SBarry Smith     nz  = ai[i+1] - ai[i];
1522c733ed4SBarry Smith     idt = bs*i;
1532c733ed4SBarry Smith     s1  = t[idt];  s2 = t[1+idt];  s3 = t[2+idt];  s4 = t[3+idt];
1542c733ed4SBarry Smith     for (j=0; j<nz; j++) {
1552c733ed4SBarry Smith       idx       = bs*vi[j];
1562c733ed4SBarry Smith       t[idx]   -=  v[0]*s1 +  v[1]*s2 +  v[2]*s3  +  v[3]*s4;
1572c733ed4SBarry Smith       t[idx+1] -=  v[4]*s1 +  v[5]*s2 +  v[6]*s3  +  v[7]*s4;
1582c733ed4SBarry Smith       t[idx+2] -=  v[8]*s1 +  v[9]*s2 +  v[10]*s3 + v[11]*s4;
1592c733ed4SBarry Smith       t[idx+3] -= v[12]*s1 +  v[13]*s2 + v[14]*s3 + v[15]*s4;
1602c733ed4SBarry Smith       v        += bs2;
1612c733ed4SBarry Smith     }
1622c733ed4SBarry Smith   }
1632c733ed4SBarry Smith 
1642c733ed4SBarry Smith   /* copy t into x according to permutation */
1652c733ed4SBarry Smith   for (i=0; i<n; i++) {
1662c733ed4SBarry Smith     ii    = bs*i;  ir = bs*r[i];
1672c733ed4SBarry Smith     x[ir] = t[ii];  x[ir+1] = t[ii+1]; x[ir+2] = t[ii+2];  x[ir+3] = t[ii+3];
1682c733ed4SBarry Smith   }
1692c733ed4SBarry Smith 
170*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(isrow,&rout));
171*9566063dSJacob Faibussowitsch   PetscCall(ISRestoreIndices(iscol,&cout));
172*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb,&b));
173*9566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx,&x));
174*9566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0*bs2*(a->nz) - bs*A->cmap->n));
1752c733ed4SBarry Smith   PetscFunctionReturn(0);
1762c733ed4SBarry Smith }
177