xref: /petsc/src/mat/impls/baij/seq/baijsolvnat7.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_7_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, s6, s7, x1, x2, x3, x4, x5, x6, x7;
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];
22*9371c9d4SSatish Balay   x[5] = b[5 + idx];
232c733ed4SBarry Smith   x[6] = b[6 + idx];
242c733ed4SBarry Smith   for (i = 1; i < n; i++) {
252c733ed4SBarry Smith     v   = aa + 49 * ai[i];
262c733ed4SBarry Smith     vi  = aj + ai[i];
272c733ed4SBarry Smith     nz  = diag[i] - ai[i];
282c733ed4SBarry Smith     idx = 7 * i;
29*9371c9d4SSatish Balay     s1  = b[idx];
30*9371c9d4SSatish Balay     s2  = b[1 + idx];
31*9371c9d4SSatish Balay     s3  = b[2 + idx];
32*9371c9d4SSatish Balay     s4  = b[3 + idx];
33*9371c9d4SSatish Balay     s5  = b[4 + idx];
34*9371c9d4SSatish Balay     s6  = b[5 + idx];
352c733ed4SBarry Smith     s7  = b[6 + idx];
362c733ed4SBarry Smith     while (nz--) {
372c733ed4SBarry Smith       jdx = 7 * (*vi++);
38*9371c9d4SSatish Balay       x1  = x[jdx];
39*9371c9d4SSatish Balay       x2  = x[1 + jdx];
40*9371c9d4SSatish Balay       x3  = x[2 + jdx];
41*9371c9d4SSatish Balay       x4  = x[3 + jdx];
42*9371c9d4SSatish Balay       x5  = x[4 + jdx];
43*9371c9d4SSatish Balay       x6  = x[5 + jdx];
442c733ed4SBarry Smith       x7  = x[6 + jdx];
452c733ed4SBarry Smith       s1 -= v[0] * x1 + v[7] * x2 + v[14] * x3 + v[21] * x4 + v[28] * x5 + v[35] * x6 + v[42] * x7;
462c733ed4SBarry Smith       s2 -= v[1] * x1 + v[8] * x2 + v[15] * x3 + v[22] * x4 + v[29] * x5 + v[36] * x6 + v[43] * x7;
472c733ed4SBarry Smith       s3 -= v[2] * x1 + v[9] * x2 + v[16] * x3 + v[23] * x4 + v[30] * x5 + v[37] * x6 + v[44] * x7;
482c733ed4SBarry Smith       s4 -= v[3] * x1 + v[10] * x2 + v[17] * x3 + v[24] * x4 + v[31] * x5 + v[38] * x6 + v[45] * x7;
492c733ed4SBarry Smith       s5 -= v[4] * x1 + v[11] * x2 + v[18] * x3 + v[25] * x4 + v[32] * x5 + v[39] * x6 + v[46] * x7;
502c733ed4SBarry Smith       s6 -= v[5] * x1 + v[12] * x2 + v[19] * x3 + v[26] * x4 + v[33] * x5 + v[40] * x6 + v[47] * x7;
512c733ed4SBarry Smith       s7 -= v[6] * x1 + v[13] * x2 + v[20] * x3 + v[27] * x4 + v[34] * x5 + v[41] * x6 + v[48] * x7;
522c733ed4SBarry Smith       v += 49;
532c733ed4SBarry Smith     }
542c733ed4SBarry Smith     x[idx]     = s1;
552c733ed4SBarry Smith     x[1 + idx] = s2;
562c733ed4SBarry Smith     x[2 + idx] = s3;
572c733ed4SBarry Smith     x[3 + idx] = s4;
582c733ed4SBarry Smith     x[4 + idx] = s5;
592c733ed4SBarry Smith     x[5 + idx] = s6;
602c733ed4SBarry Smith     x[6 + idx] = s7;
612c733ed4SBarry Smith   }
622c733ed4SBarry Smith   /* backward solve the upper triangular */
632c733ed4SBarry Smith   for (i = n - 1; i >= 0; i--) {
642c733ed4SBarry Smith     v   = aa + 49 * diag[i] + 49;
652c733ed4SBarry Smith     vi  = aj + diag[i] + 1;
662c733ed4SBarry Smith     nz  = ai[i + 1] - diag[i] - 1;
672c733ed4SBarry Smith     idt = 7 * i;
68*9371c9d4SSatish Balay     s1  = x[idt];
69*9371c9d4SSatish Balay     s2  = x[1 + idt];
70*9371c9d4SSatish Balay     s3  = x[2 + idt];
71*9371c9d4SSatish Balay     s4  = x[3 + idt];
72*9371c9d4SSatish Balay     s5  = x[4 + idt];
73*9371c9d4SSatish Balay     s6  = x[5 + idt];
742c733ed4SBarry Smith     s7  = x[6 + idt];
752c733ed4SBarry Smith     while (nz--) {
762c733ed4SBarry Smith       idx = 7 * (*vi++);
77*9371c9d4SSatish Balay       x1  = x[idx];
78*9371c9d4SSatish Balay       x2  = x[1 + idx];
79*9371c9d4SSatish Balay       x3  = x[2 + idx];
80*9371c9d4SSatish Balay       x4  = x[3 + idx];
81*9371c9d4SSatish Balay       x5  = x[4 + idx];
82*9371c9d4SSatish Balay       x6  = x[5 + idx];
832c733ed4SBarry Smith       x7  = x[6 + idx];
842c733ed4SBarry Smith       s1 -= v[0] * x1 + v[7] * x2 + v[14] * x3 + v[21] * x4 + v[28] * x5 + v[35] * x6 + v[42] * x7;
852c733ed4SBarry Smith       s2 -= v[1] * x1 + v[8] * x2 + v[15] * x3 + v[22] * x4 + v[29] * x5 + v[36] * x6 + v[43] * x7;
862c733ed4SBarry Smith       s3 -= v[2] * x1 + v[9] * x2 + v[16] * x3 + v[23] * x4 + v[30] * x5 + v[37] * x6 + v[44] * x7;
872c733ed4SBarry Smith       s4 -= v[3] * x1 + v[10] * x2 + v[17] * x3 + v[24] * x4 + v[31] * x5 + v[38] * x6 + v[45] * x7;
882c733ed4SBarry Smith       s5 -= v[4] * x1 + v[11] * x2 + v[18] * x3 + v[25] * x4 + v[32] * x5 + v[39] * x6 + v[46] * x7;
892c733ed4SBarry Smith       s6 -= v[5] * x1 + v[12] * x2 + v[19] * x3 + v[26] * x4 + v[33] * x5 + v[40] * x6 + v[47] * x7;
902c733ed4SBarry Smith       s7 -= v[6] * x1 + v[13] * x2 + v[20] * x3 + v[27] * x4 + v[34] * x5 + v[41] * x6 + v[48] * x7;
912c733ed4SBarry Smith       v += 49;
922c733ed4SBarry Smith     }
932c733ed4SBarry Smith     v          = aa + 49 * diag[i];
94*9371c9d4SSatish Balay     x[idt]     = v[0] * s1 + v[7] * s2 + v[14] * s3 + v[21] * s4 + v[28] * s5 + v[35] * s6 + v[42] * s7;
95*9371c9d4SSatish Balay     x[1 + idt] = v[1] * s1 + v[8] * s2 + v[15] * s3 + v[22] * s4 + v[29] * s5 + v[36] * s6 + v[43] * s7;
96*9371c9d4SSatish Balay     x[2 + idt] = v[2] * s1 + v[9] * s2 + v[16] * s3 + v[23] * s4 + v[30] * s5 + v[37] * s6 + v[44] * s7;
97*9371c9d4SSatish Balay     x[3 + idt] = v[3] * s1 + v[10] * s2 + v[17] * s3 + v[24] * s4 + v[31] * s5 + v[38] * s6 + v[45] * s7;
98*9371c9d4SSatish Balay     x[4 + idt] = v[4] * s1 + v[11] * s2 + v[18] * s3 + v[25] * s4 + v[32] * s5 + v[39] * s6 + v[46] * s7;
99*9371c9d4SSatish Balay     x[5 + idt] = v[5] * s1 + v[12] * s2 + v[19] * s3 + v[26] * s4 + v[33] * s5 + v[40] * s6 + v[47] * s7;
100*9371c9d4SSatish Balay     x[6 + idt] = v[6] * s1 + v[13] * s2 + v[20] * s3 + v[27] * s4 + v[34] * s5 + v[41] * s6 + v[48] * s7;
1012c733ed4SBarry Smith   }
1022c733ed4SBarry Smith 
1039566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb, &b));
1049566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx, &x));
1059566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0 * 36 * (a->nz) - 6.0 * A->cmap->n));
1062c733ed4SBarry Smith   PetscFunctionReturn(0);
1072c733ed4SBarry Smith }
1082c733ed4SBarry Smith 
109*9371c9d4SSatish Balay PetscErrorCode MatSolve_SeqBAIJ_7_NaturalOrdering(Mat A, Vec bb, Vec xx) {
1102c733ed4SBarry Smith   Mat_SeqBAIJ       *a = (Mat_SeqBAIJ *)A->data;
1112c733ed4SBarry Smith   const PetscInt     n = a->mbs, *vi, *ai = a->i, *aj = a->j, *adiag = a->diag;
1122c733ed4SBarry Smith   PetscInt           i, k, nz, idx, jdx, idt;
1132c733ed4SBarry Smith   const PetscInt     bs = A->rmap->bs, bs2 = a->bs2;
1142c733ed4SBarry Smith   const MatScalar   *aa = a->a, *v;
1152c733ed4SBarry Smith   PetscScalar       *x;
1162c733ed4SBarry Smith   const PetscScalar *b;
1172c733ed4SBarry Smith   PetscScalar        s1, s2, s3, s4, s5, s6, s7, x1, x2, x3, x4, x5, x6, x7;
1182c733ed4SBarry Smith 
1192c733ed4SBarry Smith   PetscFunctionBegin;
1209566063dSJacob Faibussowitsch   PetscCall(VecGetArrayRead(bb, &b));
1219566063dSJacob Faibussowitsch   PetscCall(VecGetArray(xx, &x));
1222c733ed4SBarry Smith   /* forward solve the lower triangular */
1232c733ed4SBarry Smith   idx  = 0;
124*9371c9d4SSatish Balay   x[0] = b[idx];
125*9371c9d4SSatish Balay   x[1] = b[1 + idx];
126*9371c9d4SSatish Balay   x[2] = b[2 + idx];
127*9371c9d4SSatish Balay   x[3] = b[3 + idx];
128*9371c9d4SSatish Balay   x[4] = b[4 + idx];
129*9371c9d4SSatish Balay   x[5] = b[5 + idx];
130*9371c9d4SSatish Balay   x[6] = b[6 + idx];
1312c733ed4SBarry Smith   for (i = 1; i < n; i++) {
1322c733ed4SBarry Smith     v   = aa + bs2 * ai[i];
1332c733ed4SBarry Smith     vi  = aj + ai[i];
1342c733ed4SBarry Smith     nz  = ai[i + 1] - ai[i];
1352c733ed4SBarry Smith     idx = bs * i;
136*9371c9d4SSatish Balay     s1  = b[idx];
137*9371c9d4SSatish Balay     s2  = b[1 + idx];
138*9371c9d4SSatish Balay     s3  = b[2 + idx];
139*9371c9d4SSatish Balay     s4  = b[3 + idx];
140*9371c9d4SSatish Balay     s5  = b[4 + idx];
141*9371c9d4SSatish Balay     s6  = b[5 + idx];
142*9371c9d4SSatish Balay     s7  = b[6 + idx];
1432c733ed4SBarry Smith     for (k = 0; k < nz; k++) {
1442c733ed4SBarry Smith       jdx = bs * vi[k];
145*9371c9d4SSatish Balay       x1  = x[jdx];
146*9371c9d4SSatish Balay       x2  = x[1 + jdx];
147*9371c9d4SSatish Balay       x3  = x[2 + jdx];
148*9371c9d4SSatish Balay       x4  = x[3 + jdx];
149*9371c9d4SSatish Balay       x5  = x[4 + jdx];
150*9371c9d4SSatish Balay       x6  = x[5 + jdx];
151*9371c9d4SSatish Balay       x7  = x[6 + jdx];
1522c733ed4SBarry Smith       s1 -= v[0] * x1 + v[7] * x2 + v[14] * x3 + v[21] * x4 + v[28] * x5 + v[35] * x6 + v[42] * x7;
1532c733ed4SBarry Smith       s2 -= v[1] * x1 + v[8] * x2 + v[15] * x3 + v[22] * x4 + v[29] * x5 + v[36] * x6 + v[43] * x7;
1542c733ed4SBarry Smith       s3 -= v[2] * x1 + v[9] * x2 + v[16] * x3 + v[23] * x4 + v[30] * x5 + v[37] * x6 + v[44] * x7;
1552c733ed4SBarry Smith       s4 -= v[3] * x1 + v[10] * x2 + v[17] * x3 + v[24] * x4 + v[31] * x5 + v[38] * x6 + v[45] * x7;
1562c733ed4SBarry Smith       s5 -= v[4] * x1 + v[11] * x2 + v[18] * x3 + v[25] * x4 + v[32] * x5 + v[39] * x6 + v[46] * x7;
1572c733ed4SBarry Smith       s6 -= v[5] * x1 + v[12] * x2 + v[19] * x3 + v[26] * x4 + v[33] * x5 + v[40] * x6 + v[47] * x7;
1582c733ed4SBarry Smith       s7 -= v[6] * x1 + v[13] * x2 + v[20] * x3 + v[27] * x4 + v[34] * x5 + v[41] * x6 + v[48] * x7;
1592c733ed4SBarry Smith       v += bs2;
1602c733ed4SBarry Smith     }
1612c733ed4SBarry Smith 
1622c733ed4SBarry Smith     x[idx]     = s1;
1632c733ed4SBarry Smith     x[1 + idx] = s2;
1642c733ed4SBarry Smith     x[2 + idx] = s3;
1652c733ed4SBarry Smith     x[3 + idx] = s4;
1662c733ed4SBarry Smith     x[4 + idx] = s5;
1672c733ed4SBarry Smith     x[5 + idx] = s6;
1682c733ed4SBarry Smith     x[6 + idx] = s7;
1692c733ed4SBarry Smith   }
1702c733ed4SBarry Smith 
1712c733ed4SBarry Smith   /* backward solve the upper triangular */
1722c733ed4SBarry Smith   for (i = n - 1; i >= 0; i--) {
1732c733ed4SBarry Smith     v   = aa + bs2 * (adiag[i + 1] + 1);
1742c733ed4SBarry Smith     vi  = aj + adiag[i + 1] + 1;
1752c733ed4SBarry Smith     nz  = adiag[i] - adiag[i + 1] - 1;
1762c733ed4SBarry Smith     idt = bs * i;
177*9371c9d4SSatish Balay     s1  = x[idt];
178*9371c9d4SSatish Balay     s2  = x[1 + idt];
179*9371c9d4SSatish Balay     s3  = x[2 + idt];
180*9371c9d4SSatish Balay     s4  = x[3 + idt];
181*9371c9d4SSatish Balay     s5  = x[4 + idt];
182*9371c9d4SSatish Balay     s6  = x[5 + idt];
183*9371c9d4SSatish Balay     s7  = x[6 + idt];
1842c733ed4SBarry Smith     for (k = 0; k < nz; k++) {
1852c733ed4SBarry Smith       idx = bs * vi[k];
186*9371c9d4SSatish Balay       x1  = x[idx];
187*9371c9d4SSatish Balay       x2  = x[1 + idx];
188*9371c9d4SSatish Balay       x3  = x[2 + idx];
189*9371c9d4SSatish Balay       x4  = x[3 + idx];
190*9371c9d4SSatish Balay       x5  = x[4 + idx];
191*9371c9d4SSatish Balay       x6  = x[5 + idx];
192*9371c9d4SSatish Balay       x7  = x[6 + idx];
1932c733ed4SBarry Smith       s1 -= v[0] * x1 + v[7] * x2 + v[14] * x3 + v[21] * x4 + v[28] * x5 + v[35] * x6 + v[42] * x7;
1942c733ed4SBarry Smith       s2 -= v[1] * x1 + v[8] * x2 + v[15] * x3 + v[22] * x4 + v[29] * x5 + v[36] * x6 + v[43] * x7;
1952c733ed4SBarry Smith       s3 -= v[2] * x1 + v[9] * x2 + v[16] * x3 + v[23] * x4 + v[30] * x5 + v[37] * x6 + v[44] * x7;
1962c733ed4SBarry Smith       s4 -= v[3] * x1 + v[10] * x2 + v[17] * x3 + v[24] * x4 + v[31] * x5 + v[38] * x6 + v[45] * x7;
1972c733ed4SBarry Smith       s5 -= v[4] * x1 + v[11] * x2 + v[18] * x3 + v[25] * x4 + v[32] * x5 + v[39] * x6 + v[46] * x7;
1982c733ed4SBarry Smith       s6 -= v[5] * x1 + v[12] * x2 + v[19] * x3 + v[26] * x4 + v[33] * x5 + v[40] * x6 + v[47] * x7;
1992c733ed4SBarry Smith       s7 -= v[6] * x1 + v[13] * x2 + v[20] * x3 + v[27] * x4 + v[34] * x5 + v[41] * x6 + v[48] * x7;
2002c733ed4SBarry Smith       v += bs2;
2012c733ed4SBarry Smith     }
2022c733ed4SBarry Smith     /* x = inv_diagonal*x */
2032c733ed4SBarry Smith     x[idt]     = v[0] * s1 + v[7] * s2 + v[14] * s3 + v[21] * s4 + v[28] * s5 + v[35] * s6 + v[42] * s7;
2042c733ed4SBarry Smith     x[1 + idt] = v[1] * s1 + v[8] * s2 + v[15] * s3 + v[22] * s4 + v[29] * s5 + v[36] * s6 + v[43] * s7;
2052c733ed4SBarry Smith     x[2 + idt] = v[2] * s1 + v[9] * s2 + v[16] * s3 + v[23] * s4 + v[30] * s5 + v[37] * s6 + v[44] * s7;
2062c733ed4SBarry Smith     x[3 + idt] = v[3] * s1 + v[10] * s2 + v[17] * s3 + v[24] * s4 + v[31] * s5 + v[38] * s6 + v[45] * s7;
2072c733ed4SBarry Smith     x[4 + idt] = v[4] * s1 + v[11] * s2 + v[18] * s3 + v[25] * s4 + v[32] * s5 + v[39] * s6 + v[46] * s7;
2082c733ed4SBarry Smith     x[5 + idt] = v[5] * s1 + v[12] * s2 + v[19] * s3 + v[26] * s4 + v[33] * s5 + v[40] * s6 + v[47] * s7;
2092c733ed4SBarry Smith     x[6 + idt] = v[6] * s1 + v[13] * s2 + v[20] * s3 + v[27] * s4 + v[34] * s5 + v[41] * s6 + v[48] * s7;
2102c733ed4SBarry Smith   }
2112c733ed4SBarry Smith 
2129566063dSJacob Faibussowitsch   PetscCall(VecRestoreArrayRead(bb, &b));
2139566063dSJacob Faibussowitsch   PetscCall(VecRestoreArray(xx, &x));
2149566063dSJacob Faibussowitsch   PetscCall(PetscLogFlops(2.0 * bs2 * (a->nz) - bs * A->cmap->n));
2152c733ed4SBarry Smith   PetscFunctionReturn(0);
2162c733ed4SBarry Smith }
217