Actual source code: baijfact14.c
1: /*$Id: baijfact14.c,v 1.4 2001/03/23 23:22:07 balay Exp $*/
2: /*
3: Factorization code for BAIJ format.
4: */
5: #include src/mat/impls/baij/seq/baij.h
6: #include src/vec/vecimpl.h
7: #include src/inline/ilu.h
8: /*
9: Version for when blocks are 3 by 3 Using natural ordering
10: */
11: int MatLUFactorNumeric_SeqBAIJ_3_NaturalOrdering(Mat A,Mat *B)
12: {
13: Mat C = *B;
14: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ *)C->data;
15: int ierr,i,j,n = a->mbs,*bi = b->i,*bj = b->j;
16: int *ajtmpold,*ajtmp,nz,row;
17: int *diag_offset = b->diag,*ai=a->i,*aj=a->j,*pj;
18: MatScalar *pv,*v,*rtmp,*pc,*w,*x;
19: MatScalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4;
20: MatScalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9;
21: MatScalar *ba = b->a,*aa = a->a;
24: PetscMalloc(9*(n+1)*sizeof(MatScalar),&rtmp);
26: for (i=0; i<n; i++) {
27: nz = bi[i+1] - bi[i];
28: ajtmp = bj + bi[i];
29: for (j=0; j<nz; j++) {
30: x = rtmp+9*ajtmp[j];
31: x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = 0.0;
32: }
33: /* load in initial (unfactored row) */
34: nz = ai[i+1] - ai[i];
35: ajtmpold = aj + ai[i];
36: v = aa + 9*ai[i];
37: for (j=0; j<nz; j++) {
38: x = rtmp+9*ajtmpold[j];
39: x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3];
40: x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8];
41: v += 9;
42: }
43: row = *ajtmp++;
44: while (row < i) {
45: pc = rtmp + 9*row;
46: p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3];
47: p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8];
48: if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 ||
49: p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) {
50: pv = ba + 9*diag_offset[row];
51: pj = bj + diag_offset[row] + 1;
52: x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3];
53: x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8];
54: pc[0] = m1 = p1*x1 + p4*x2 + p7*x3;
55: pc[1] = m2 = p2*x1 + p5*x2 + p8*x3;
56: pc[2] = m3 = p3*x1 + p6*x2 + p9*x3;
58: pc[3] = m4 = p1*x4 + p4*x5 + p7*x6;
59: pc[4] = m5 = p2*x4 + p5*x5 + p8*x6;
60: pc[5] = m6 = p3*x4 + p6*x5 + p9*x6;
62: pc[6] = m7 = p1*x7 + p4*x8 + p7*x9;
63: pc[7] = m8 = p2*x7 + p5*x8 + p8*x9;
64: pc[8] = m9 = p3*x7 + p6*x8 + p9*x9;
66: nz = bi[row+1] - diag_offset[row] - 1;
67: pv += 9;
68: for (j=0; j<nz; j++) {
69: x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3];
70: x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8];
71: x = rtmp + 9*pj[j];
72: x[0] -= m1*x1 + m4*x2 + m7*x3;
73: x[1] -= m2*x1 + m5*x2 + m8*x3;
74: x[2] -= m3*x1 + m6*x2 + m9*x3;
75:
76: x[3] -= m1*x4 + m4*x5 + m7*x6;
77: x[4] -= m2*x4 + m5*x5 + m8*x6;
78: x[5] -= m3*x4 + m6*x5 + m9*x6;
80: x[6] -= m1*x7 + m4*x8 + m7*x9;
81: x[7] -= m2*x7 + m5*x8 + m8*x9;
82: x[8] -= m3*x7 + m6*x8 + m9*x9;
83: pv += 9;
84: }
85: PetscLogFlops(54*nz+36);
86: }
87: row = *ajtmp++;
88: }
89: /* finished row so stick it into b->a */
90: pv = ba + 9*bi[i];
91: pj = bj + bi[i];
92: nz = bi[i+1] - bi[i];
93: for (j=0; j<nz; j++) {
94: x = rtmp+9*pj[j];
95: pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3];
96: pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8];
97: pv += 9;
98: }
99: /* invert diagonal block */
100: w = ba + 9*diag_offset[i];
101: Kernel_A_gets_inverse_A_3(w);
102: }
104: PetscFree(rtmp);
105: C->factor = FACTOR_LU;
106: C->assembled = PETSC_TRUE;
107: PetscLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */
108: return(0);
109: }