xref: /petsc/src/mat/impls/baij/seq/baijsolvnat5.c (revision 9371c9d470a9602b6d10a8bf50c9b2280a79e45a)
12c733ed4SBarry Smith #include <../src/mat/impls/baij/seq/baij.h>
22c733ed4SBarry Smith #include <petsc/private/kernels/blockinvert.h>
32c733ed4SBarry Smith 
4*9371c9d4SSatish Balay PetscErrorCode MatSolve_SeqBAIJ_5_NaturalOrdering_inplace(Mat A, Vec bb, Vec xx) {
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, jdx;
82c733ed4SBarry Smith   const MatScalar   *aa = a->a, *v;
92c733ed4SBarry Smith   PetscScalar       *x, s1, s2, s3, s4, s5, x1, x2, x3, x4, x5;
102c733ed4SBarry Smith   const PetscScalar *b;
112c733ed4SBarry Smith 
122c733ed4SBarry Smith   PetscFunctionBegin;
139566063dSJacob Faibussowitsch   PetscCall(VecGetArrayRead(bb, &b));
149566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx, &x));
152c733ed4SBarry Smith   /* forward solve the lower triangular */
162c733ed4SBarry Smith   idx  = 0;
17*9371c9d4SSatish Balay   x[0] = b[idx];
18*9371c9d4SSatish Balay   x[1] = b[1 + idx];
19*9371c9d4SSatish Balay   x[2] = b[2 + idx];
20*9371c9d4SSatish Balay   x[3] = b[3 + idx];
21*9371c9d4SSatish Balay   x[4] = b[4 + idx];
222c733ed4SBarry Smith   for (i = 1; i < n; i++) {
232c733ed4SBarry Smith     v   = aa + 25 * ai[i];
242c733ed4SBarry Smith     vi  = aj + ai[i];
252c733ed4SBarry Smith     nz  = diag[i] - ai[i];
262c733ed4SBarry Smith     idx = 5 * i;
27*9371c9d4SSatish Balay     s1  = b[idx];
28*9371c9d4SSatish Balay     s2  = b[1 + idx];
29*9371c9d4SSatish Balay     s3  = b[2 + idx];
30*9371c9d4SSatish Balay     s4  = b[3 + idx];
31*9371c9d4SSatish Balay     s5  = b[4 + idx];
322c733ed4SBarry Smith     while (nz--) {
332c733ed4SBarry Smith       jdx = 5 * (*vi++);
34*9371c9d4SSatish Balay       x1  = x[jdx];
35*9371c9d4SSatish Balay       x2  = x[1 + jdx];
36*9371c9d4SSatish Balay       x3  = x[2 + jdx];
37*9371c9d4SSatish Balay       x4  = x[3 + jdx];
38*9371c9d4SSatish Balay       x5  = x[4 + jdx];
392c733ed4SBarry Smith       s1 -= v[0] * x1 + v[5] * x2 + v[10] * x3 + v[15] * x4 + v[20] * x5;
402c733ed4SBarry Smith       s2 -= v[1] * x1 + v[6] * x2 + v[11] * x3 + v[16] * x4 + v[21] * x5;
412c733ed4SBarry Smith       s3 -= v[2] * x1 + v[7] * x2 + v[12] * x3 + v[17] * x4 + v[22] * x5;
422c733ed4SBarry Smith       s4 -= v[3] * x1 + v[8] * x2 + v[13] * x3 + v[18] * x4 + v[23] * x5;
432c733ed4SBarry Smith       s5 -= v[4] * x1 + v[9] * x2 + v[14] * x3 + v[19] * x4 + v[24] * x5;
442c733ed4SBarry Smith       v += 25;
452c733ed4SBarry Smith     }
462c733ed4SBarry Smith     x[idx]     = s1;
472c733ed4SBarry Smith     x[1 + idx] = s2;
482c733ed4SBarry Smith     x[2 + idx] = s3;
492c733ed4SBarry Smith     x[3 + idx] = s4;
502c733ed4SBarry Smith     x[4 + idx] = s5;
512c733ed4SBarry Smith   }
522c733ed4SBarry Smith   /* backward solve the upper triangular */
532c733ed4SBarry Smith   for (i = n - 1; i >= 0; i--) {
542c733ed4SBarry Smith     v   = aa + 25 * diag[i] + 25;
552c733ed4SBarry Smith     vi  = aj + diag[i] + 1;
562c733ed4SBarry Smith     nz  = ai[i + 1] - diag[i] - 1;
572c733ed4SBarry Smith     idt = 5 * i;
58*9371c9d4SSatish Balay     s1  = x[idt];
59*9371c9d4SSatish Balay     s2  = x[1 + idt];
60*9371c9d4SSatish Balay     s3  = x[2 + idt];
61*9371c9d4SSatish Balay     s4  = x[3 + idt];
62*9371c9d4SSatish Balay     s5  = x[4 + idt];
632c733ed4SBarry Smith     while (nz--) {
642c733ed4SBarry Smith       idx = 5 * (*vi++);
65*9371c9d4SSatish Balay       x1  = x[idx];
66*9371c9d4SSatish Balay       x2  = x[1 + idx];
67*9371c9d4SSatish Balay       x3  = x[2 + idx];
68*9371c9d4SSatish Balay       x4  = x[3 + idx];
69*9371c9d4SSatish Balay       x5  = x[4 + idx];
702c733ed4SBarry Smith       s1 -= v[0] * x1 + v[5] * x2 + v[10] * x3 + v[15] * x4 + v[20] * x5;
712c733ed4SBarry Smith       s2 -= v[1] * x1 + v[6] * x2 + v[11] * x3 + v[16] * x4 + v[21] * x5;
722c733ed4SBarry Smith       s3 -= v[2] * x1 + v[7] * x2 + v[12] * x3 + v[17] * x4 + v[22] * x5;
732c733ed4SBarry Smith       s4 -= v[3] * x1 + v[8] * x2 + v[13] * x3 + v[18] * x4 + v[23] * x5;
742c733ed4SBarry Smith       s5 -= v[4] * x1 + v[9] * x2 + v[14] * x3 + v[19] * x4 + v[24] * x5;
752c733ed4SBarry Smith       v += 25;
762c733ed4SBarry Smith     }
772c733ed4SBarry Smith     v          = aa + 25 * diag[i];
782c733ed4SBarry Smith     x[idt]     = v[0] * s1 + v[5] * s2 + v[10] * s3 + v[15] * s4 + v[20] * s5;
792c733ed4SBarry Smith     x[1 + idt] = v[1] * s1 + v[6] * s2 + v[11] * s3 + v[16] * s4 + v[21] * s5;
802c733ed4SBarry Smith     x[2 + idt] = v[2] * s1 + v[7] * s2 + v[12] * s3 + v[17] * s4 + v[22] * s5;
812c733ed4SBarry Smith     x[3 + idt] = v[3] * s1 + v[8] * s2 + v[13] * s3 + v[18] * s4 + v[23] * s5;
822c733ed4SBarry Smith     x[4 + idt] = v[4] * s1 + v[9] * s2 + v[14] * s3 + v[19] * s4 + v[24] * s5;
832c733ed4SBarry Smith   }
842c733ed4SBarry Smith 
859566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb, &b));
869566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx, &x));
879566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0 * 25 * (a->nz) - 5.0 * A->cmap->n));
882c733ed4SBarry Smith   PetscFunctionReturn(0);
892c733ed4SBarry Smith }
902c733ed4SBarry Smith 
91*9371c9d4SSatish Balay PetscErrorCode MatSolve_SeqBAIJ_5_NaturalOrdering(Mat A, Vec bb, Vec xx) {
922c733ed4SBarry Smith   Mat_SeqBAIJ       *a = (Mat_SeqBAIJ *)A->data;
932c733ed4SBarry Smith   const PetscInt     n = a->mbs, *vi, *ai = a->i, *aj = a->j, *adiag = a->diag;
942c733ed4SBarry Smith   PetscInt           i, k, nz, idx, idt, jdx;
952c733ed4SBarry Smith   const MatScalar   *aa = a->a, *v;
962c733ed4SBarry Smith   PetscScalar       *x, s1, s2, s3, s4, s5, x1, x2, x3, x4, x5;
972c733ed4SBarry Smith   const PetscScalar *b;
982c733ed4SBarry Smith 
992c733ed4SBarry Smith   PetscFunctionBegin;
1009566063dSJacob Faibussowitsch   PetscCall(VecGetArrayRead(bb, &b));
1019566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx, &x));
1022c733ed4SBarry Smith   /* forward solve the lower triangular */
1032c733ed4SBarry Smith   idx  = 0;
104*9371c9d4SSatish Balay   x[0] = b[idx];
105*9371c9d4SSatish Balay   x[1] = b[1 + idx];
106*9371c9d4SSatish Balay   x[2] = b[2 + idx];
107*9371c9d4SSatish Balay   x[3] = b[3 + idx];
108*9371c9d4SSatish Balay   x[4] = b[4 + idx];
1092c733ed4SBarry Smith   for (i = 1; i < n; i++) {
1102c733ed4SBarry Smith     v   = aa + 25 * ai[i];
1112c733ed4SBarry Smith     vi  = aj + ai[i];
1122c733ed4SBarry Smith     nz  = ai[i + 1] - ai[i];
1132c733ed4SBarry Smith     idx = 5 * i;
114*9371c9d4SSatish Balay     s1  = b[idx];
115*9371c9d4SSatish Balay     s2  = b[1 + idx];
116*9371c9d4SSatish Balay     s3  = b[2 + idx];
117*9371c9d4SSatish Balay     s4  = b[3 + idx];
118*9371c9d4SSatish Balay     s5  = b[4 + idx];
1192c733ed4SBarry Smith     for (k = 0; k < nz; k++) {
1202c733ed4SBarry Smith       jdx = 5 * vi[k];
121*9371c9d4SSatish Balay       x1  = x[jdx];
122*9371c9d4SSatish Balay       x2  = x[1 + jdx];
123*9371c9d4SSatish Balay       x3  = x[2 + jdx];
124*9371c9d4SSatish Balay       x4  = x[3 + jdx];
125*9371c9d4SSatish Balay       x5  = x[4 + jdx];
1262c733ed4SBarry Smith       s1 -= v[0] * x1 + v[5] * x2 + v[10] * x3 + v[15] * x4 + v[20] * x5;
1272c733ed4SBarry Smith       s2 -= v[1] * x1 + v[6] * x2 + v[11] * x3 + v[16] * x4 + v[21] * x5;
1282c733ed4SBarry Smith       s3 -= v[2] * x1 + v[7] * x2 + v[12] * x3 + v[17] * x4 + v[22] * x5;
1292c733ed4SBarry Smith       s4 -= v[3] * x1 + v[8] * x2 + v[13] * x3 + v[18] * x4 + v[23] * x5;
1302c733ed4SBarry Smith       s5 -= v[4] * x1 + v[9] * x2 + v[14] * x3 + v[19] * x4 + v[24] * x5;
1312c733ed4SBarry Smith       v += 25;
1322c733ed4SBarry Smith     }
1332c733ed4SBarry Smith     x[idx]     = s1;
1342c733ed4SBarry Smith     x[1 + idx] = s2;
1352c733ed4SBarry Smith     x[2 + idx] = s3;
1362c733ed4SBarry Smith     x[3 + idx] = s4;
1372c733ed4SBarry Smith     x[4 + idx] = s5;
1382c733ed4SBarry Smith   }
1392c733ed4SBarry Smith 
1402c733ed4SBarry Smith   /* backward solve the upper triangular */
1412c733ed4SBarry Smith   for (i = n - 1; i >= 0; i--) {
1422c733ed4SBarry Smith     v   = aa + 25 * (adiag[i + 1] + 1);
1432c733ed4SBarry Smith     vi  = aj + adiag[i + 1] + 1;
1442c733ed4SBarry Smith     nz  = adiag[i] - adiag[i + 1] - 1;
1452c733ed4SBarry Smith     idt = 5 * i;
146*9371c9d4SSatish Balay     s1  = x[idt];
147*9371c9d4SSatish Balay     s2  = x[1 + idt];
148*9371c9d4SSatish Balay     s3  = x[2 + idt];
149*9371c9d4SSatish Balay     s4  = x[3 + idt];
150*9371c9d4SSatish Balay     s5  = x[4 + idt];
1512c733ed4SBarry Smith     for (k = 0; k < nz; k++) {
1522c733ed4SBarry Smith       idx = 5 * vi[k];
153*9371c9d4SSatish Balay       x1  = x[idx];
154*9371c9d4SSatish Balay       x2  = x[1 + idx];
155*9371c9d4SSatish Balay       x3  = x[2 + idx];
156*9371c9d4SSatish Balay       x4  = x[3 + idx];
157*9371c9d4SSatish Balay       x5  = x[4 + idx];
1582c733ed4SBarry Smith       s1 -= v[0] * x1 + v[5] * x2 + v[10] * x3 + v[15] * x4 + v[20] * x5;
1592c733ed4SBarry Smith       s2 -= v[1] * x1 + v[6] * x2 + v[11] * x3 + v[16] * x4 + v[21] * x5;
1602c733ed4SBarry Smith       s3 -= v[2] * x1 + v[7] * x2 + v[12] * x3 + v[17] * x4 + v[22] * x5;
1612c733ed4SBarry Smith       s4 -= v[3] * x1 + v[8] * x2 + v[13] * x3 + v[18] * x4 + v[23] * x5;
1622c733ed4SBarry Smith       s5 -= v[4] * x1 + v[9] * x2 + v[14] * x3 + v[19] * x4 + v[24] * x5;
1632c733ed4SBarry Smith       v += 25;
1642c733ed4SBarry Smith     }
1652c733ed4SBarry Smith     /* x = inv_diagonal*x */
1662c733ed4SBarry Smith     x[idt]     = v[0] * s1 + v[5] * s2 + v[10] * s3 + v[15] * s4 + v[20] * s5;
1672c733ed4SBarry Smith     x[1 + idt] = v[1] * s1 + v[6] * s2 + v[11] * s3 + v[16] * s4 + v[21] * s5;
1682c733ed4SBarry Smith     x[2 + idt] = v[2] * s1 + v[7] * s2 + v[12] * s3 + v[17] * s4 + v[22] * s5;
1692c733ed4SBarry Smith     x[3 + idt] = v[3] * s1 + v[8] * s2 + v[13] * s3 + v[18] * s4 + v[23] * s5;
1702c733ed4SBarry Smith     x[4 + idt] = v[4] * s1 + v[9] * s2 + v[14] * s3 + v[19] * s4 + v[24] * s5;
1712c733ed4SBarry Smith   }
1722c733ed4SBarry Smith 
1739566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb, &b));
1749566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx, &x));
1759566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0 * 25 * (a->nz) - 5.0 * A->cmap->n));
1762c733ed4SBarry Smith   PetscFunctionReturn(0);
1772c733ed4SBarry Smith }
178