xref: /petsc/src/mat/impls/baij/seq/baijsolvtrann.c (revision 2c733ed45a445b0336611d812c866924feeaee2c)
1*2c733ed4SBarry Smith #include <../src/mat/impls/baij/seq/baij.h>
2*2c733ed4SBarry Smith #include <petsc/private/kernels/blockinvert.h>
3*2c733ed4SBarry Smith 
4*2c733ed4SBarry Smith /* ----------------------------------------------------------- */
5*2c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_N_inplace(Mat A,Vec bb,Vec xx)
6*2c733ed4SBarry Smith {
7*2c733ed4SBarry Smith   Mat_SeqBAIJ       *a   =(Mat_SeqBAIJ*)A->data;
8*2c733ed4SBarry Smith   IS                iscol=a->col,isrow=a->row;
9*2c733ed4SBarry Smith   PetscErrorCode    ierr;
10*2c733ed4SBarry Smith   const PetscInt    *r,*c,*rout,*cout,*ai=a->i,*aj=a->j,*vi;
11*2c733ed4SBarry Smith   PetscInt          i,nz,j;
12*2c733ed4SBarry Smith   const PetscInt    n  =a->mbs,bs=A->rmap->bs,bs2=a->bs2;
13*2c733ed4SBarry Smith   const MatScalar   *aa=a->a,*v;
14*2c733ed4SBarry Smith   PetscScalar       *x,*t,*ls;
15*2c733ed4SBarry Smith   const PetscScalar *b;
16*2c733ed4SBarry Smith 
17*2c733ed4SBarry Smith   PetscFunctionBegin;
18*2c733ed4SBarry Smith   ierr = VecGetArrayRead(bb,&b);CHKERRQ(ierr);
19*2c733ed4SBarry Smith   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
20*2c733ed4SBarry Smith   t    = a->solve_work;
21*2c733ed4SBarry Smith 
22*2c733ed4SBarry Smith   ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout;
23*2c733ed4SBarry Smith   ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout;
24*2c733ed4SBarry Smith 
25*2c733ed4SBarry Smith   /* copy the b into temp work space according to permutation */
26*2c733ed4SBarry Smith   for (i=0; i<n; i++) {
27*2c733ed4SBarry Smith     for (j=0; j<bs; j++) {
28*2c733ed4SBarry Smith       t[i*bs+j] = b[c[i]*bs+j];
29*2c733ed4SBarry Smith     }
30*2c733ed4SBarry Smith   }
31*2c733ed4SBarry Smith 
32*2c733ed4SBarry Smith 
33*2c733ed4SBarry Smith   /* forward solve the upper triangular transpose */
34*2c733ed4SBarry Smith   ls = a->solve_work + A->cmap->n;
35*2c733ed4SBarry Smith   for (i=0; i<n; i++) {
36*2c733ed4SBarry Smith     ierr = PetscMemcpy(ls,t+i*bs,bs*sizeof(PetscScalar));CHKERRQ(ierr);
37*2c733ed4SBarry Smith     PetscKernel_w_gets_transA_times_v(bs,ls,aa+bs2*a->diag[i],t+i*bs);
38*2c733ed4SBarry Smith     v  = aa + bs2*(a->diag[i] + 1);
39*2c733ed4SBarry Smith     vi = aj + a->diag[i] + 1;
40*2c733ed4SBarry Smith     nz = ai[i+1] - a->diag[i] - 1;
41*2c733ed4SBarry Smith     while (nz--) {
42*2c733ed4SBarry Smith       PetscKernel_v_gets_v_minus_transA_times_w(bs,t+bs*(*vi++),v,t+i*bs);
43*2c733ed4SBarry Smith       v += bs2;
44*2c733ed4SBarry Smith     }
45*2c733ed4SBarry Smith   }
46*2c733ed4SBarry Smith 
47*2c733ed4SBarry Smith   /* backward solve the lower triangular transpose */
48*2c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
49*2c733ed4SBarry Smith     v  = aa + bs2*ai[i];
50*2c733ed4SBarry Smith     vi = aj + ai[i];
51*2c733ed4SBarry Smith     nz = a->diag[i] - ai[i];
52*2c733ed4SBarry Smith     while (nz--) {
53*2c733ed4SBarry Smith       PetscKernel_v_gets_v_minus_transA_times_w(bs,t+bs*(*vi++),v,t+i*bs);
54*2c733ed4SBarry Smith       v += bs2;
55*2c733ed4SBarry Smith     }
56*2c733ed4SBarry Smith   }
57*2c733ed4SBarry Smith 
58*2c733ed4SBarry Smith   /* copy t into x according to permutation */
59*2c733ed4SBarry Smith   for (i=0; i<n; i++) {
60*2c733ed4SBarry Smith     for (j=0; j<bs; j++) {
61*2c733ed4SBarry Smith       x[bs*r[i]+j]   = t[bs*i+j];
62*2c733ed4SBarry Smith     }
63*2c733ed4SBarry Smith   }
64*2c733ed4SBarry Smith 
65*2c733ed4SBarry Smith   ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr);
66*2c733ed4SBarry Smith   ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr);
67*2c733ed4SBarry Smith   ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr);
68*2c733ed4SBarry Smith   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
69*2c733ed4SBarry Smith   ierr = PetscLogFlops(2.0*(a->bs2)*(a->nz) - A->rmap->bs*A->cmap->n);CHKERRQ(ierr);
70*2c733ed4SBarry Smith   PetscFunctionReturn(0);
71*2c733ed4SBarry Smith }
72*2c733ed4SBarry Smith 
73*2c733ed4SBarry Smith PetscErrorCode MatSolveTranspose_SeqBAIJ_N(Mat A,Vec bb,Vec xx)
74*2c733ed4SBarry Smith {
75*2c733ed4SBarry Smith   Mat_SeqBAIJ       *a   =(Mat_SeqBAIJ*)A->data;
76*2c733ed4SBarry Smith   IS                iscol=a->col,isrow=a->row;
77*2c733ed4SBarry Smith   PetscErrorCode    ierr;
78*2c733ed4SBarry Smith   const PetscInt    *r,*c,*rout,*cout;
79*2c733ed4SBarry Smith   const PetscInt    n=a->mbs,*ai=a->i,*aj=a->j,*vi,*diag=a->diag;
80*2c733ed4SBarry Smith   PetscInt          i,j,nz;
81*2c733ed4SBarry Smith   const PetscInt    bs =A->rmap->bs,bs2=a->bs2;
82*2c733ed4SBarry Smith   const MatScalar   *aa=a->a,*v;
83*2c733ed4SBarry Smith   PetscScalar       *x,*t,*ls;
84*2c733ed4SBarry Smith   const PetscScalar *b;
85*2c733ed4SBarry Smith 
86*2c733ed4SBarry Smith   PetscFunctionBegin;
87*2c733ed4SBarry Smith   ierr = VecGetArrayRead(bb,&b);CHKERRQ(ierr);
88*2c733ed4SBarry Smith   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
89*2c733ed4SBarry Smith   t    = a->solve_work;
90*2c733ed4SBarry Smith 
91*2c733ed4SBarry Smith   ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout;
92*2c733ed4SBarry Smith   ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout;
93*2c733ed4SBarry Smith 
94*2c733ed4SBarry Smith   /* copy the b into temp work space according to permutation */
95*2c733ed4SBarry Smith   for (i=0; i<n; i++) {
96*2c733ed4SBarry Smith     for (j=0; j<bs; j++) {
97*2c733ed4SBarry Smith       t[i*bs+j] = b[c[i]*bs+j];
98*2c733ed4SBarry Smith     }
99*2c733ed4SBarry Smith   }
100*2c733ed4SBarry Smith 
101*2c733ed4SBarry Smith 
102*2c733ed4SBarry Smith   /* forward solve the upper triangular transpose */
103*2c733ed4SBarry Smith   ls = a->solve_work + A->cmap->n;
104*2c733ed4SBarry Smith   for (i=0; i<n; i++) {
105*2c733ed4SBarry Smith     ierr = PetscMemcpy(ls,t+i*bs,bs*sizeof(PetscScalar));CHKERRQ(ierr);
106*2c733ed4SBarry Smith     PetscKernel_w_gets_transA_times_v(bs,ls,aa+bs2*diag[i],t+i*bs);
107*2c733ed4SBarry Smith     v  = aa + bs2*(diag[i] - 1);
108*2c733ed4SBarry Smith     vi = aj + diag[i] - 1;
109*2c733ed4SBarry Smith     nz = diag[i] - diag[i+1] - 1;
110*2c733ed4SBarry Smith     for (j=0; j>-nz; j--) {
111*2c733ed4SBarry Smith       PetscKernel_v_gets_v_minus_transA_times_w(bs,t+bs*(vi[j]),v,t+i*bs);
112*2c733ed4SBarry Smith       v -= bs2;
113*2c733ed4SBarry Smith     }
114*2c733ed4SBarry Smith   }
115*2c733ed4SBarry Smith 
116*2c733ed4SBarry Smith   /* backward solve the lower triangular transpose */
117*2c733ed4SBarry Smith   for (i=n-1; i>=0; i--) {
118*2c733ed4SBarry Smith     v  = aa + bs2*ai[i];
119*2c733ed4SBarry Smith     vi = aj + ai[i];
120*2c733ed4SBarry Smith     nz = ai[i+1] - ai[i];
121*2c733ed4SBarry Smith     for (j=0; j<nz; j++) {
122*2c733ed4SBarry Smith       PetscKernel_v_gets_v_minus_transA_times_w(bs,t+bs*(vi[j]),v,t+i*bs);
123*2c733ed4SBarry Smith       v += bs2;
124*2c733ed4SBarry Smith     }
125*2c733ed4SBarry Smith   }
126*2c733ed4SBarry Smith 
127*2c733ed4SBarry Smith   /* copy t into x according to permutation */
128*2c733ed4SBarry Smith   for (i=0; i<n; i++) {
129*2c733ed4SBarry Smith     for (j=0; j<bs; j++) {
130*2c733ed4SBarry Smith       x[bs*r[i]+j]   = t[bs*i+j];
131*2c733ed4SBarry Smith     }
132*2c733ed4SBarry Smith   }
133*2c733ed4SBarry Smith 
134*2c733ed4SBarry Smith   ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr);
135*2c733ed4SBarry Smith   ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr);
136*2c733ed4SBarry Smith   ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr);
137*2c733ed4SBarry Smith   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
138*2c733ed4SBarry Smith   ierr = PetscLogFlops(2.0*(a->bs2)*(a->nz) - A->rmap->bs*A->cmap->n);CHKERRQ(ierr);
139*2c733ed4SBarry Smith   PetscFunctionReturn(0);
140*2c733ed4SBarry Smith }
141*2c733ed4SBarry Smith 
142