Actual source code: baijfact14.c
1: #define PETSCMAT_DLL
3: /*
4: Factorization code for BAIJ format.
5: */
6: #include src/mat/impls/baij/seq/baij.h
7: #include src/inline/ilu.h
8: /*
9: Version for when blocks are 3 by 3 Using natural ordering
10: */
13: PetscErrorCode MatLUFactorNumeric_SeqBAIJ_3_NaturalOrdering(Mat A,MatFactorInfo *info,Mat *B)
14: {
15: Mat C = *B;
16: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ *)C->data;
18: PetscInt i,j,n = a->mbs,*bi = b->i,*bj = b->j;
19: PetscInt *ajtmpold,*ajtmp,nz,row;
20: PetscInt *diag_offset = b->diag,*ai=a->i,*aj=a->j,*pj;
21: MatScalar *pv,*v,*rtmp,*pc,*w,*x;
22: MatScalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4;
23: MatScalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9;
24: MatScalar *ba = b->a,*aa = a->a;
27: PetscMalloc(9*(n+1)*sizeof(MatScalar),&rtmp);
29: for (i=0; i<n; i++) {
30: nz = bi[i+1] - bi[i];
31: ajtmp = bj + bi[i];
32: for (j=0; j<nz; j++) {
33: x = rtmp+9*ajtmp[j];
34: x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = 0.0;
35: }
36: /* load in initial (unfactored row) */
37: nz = ai[i+1] - ai[i];
38: ajtmpold = aj + ai[i];
39: v = aa + 9*ai[i];
40: for (j=0; j<nz; j++) {
41: x = rtmp+9*ajtmpold[j];
42: x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3];
43: x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8];
44: v += 9;
45: }
46: row = *ajtmp++;
47: while (row < i) {
48: pc = rtmp + 9*row;
49: p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3];
50: p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8];
51: if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 ||
52: p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) {
53: pv = ba + 9*diag_offset[row];
54: pj = bj + diag_offset[row] + 1;
55: x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3];
56: x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8];
57: pc[0] = m1 = p1*x1 + p4*x2 + p7*x3;
58: pc[1] = m2 = p2*x1 + p5*x2 + p8*x3;
59: pc[2] = m3 = p3*x1 + p6*x2 + p9*x3;
61: pc[3] = m4 = p1*x4 + p4*x5 + p7*x6;
62: pc[4] = m5 = p2*x4 + p5*x5 + p8*x6;
63: pc[5] = m6 = p3*x4 + p6*x5 + p9*x6;
65: pc[6] = m7 = p1*x7 + p4*x8 + p7*x9;
66: pc[7] = m8 = p2*x7 + p5*x8 + p8*x9;
67: pc[8] = m9 = p3*x7 + p6*x8 + p9*x9;
69: nz = bi[row+1] - diag_offset[row] - 1;
70: pv += 9;
71: for (j=0; j<nz; j++) {
72: x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3];
73: x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8];
74: x = rtmp + 9*pj[j];
75: x[0] -= m1*x1 + m4*x2 + m7*x3;
76: x[1] -= m2*x1 + m5*x2 + m8*x3;
77: x[2] -= m3*x1 + m6*x2 + m9*x3;
78:
79: x[3] -= m1*x4 + m4*x5 + m7*x6;
80: x[4] -= m2*x4 + m5*x5 + m8*x6;
81: x[5] -= m3*x4 + m6*x5 + m9*x6;
83: x[6] -= m1*x7 + m4*x8 + m7*x9;
84: x[7] -= m2*x7 + m5*x8 + m8*x9;
85: x[8] -= m3*x7 + m6*x8 + m9*x9;
86: pv += 9;
87: }
88: PetscLogFlops(54*nz+36);
89: }
90: row = *ajtmp++;
91: }
92: /* finished row so stick it into b->a */
93: pv = ba + 9*bi[i];
94: pj = bj + bi[i];
95: nz = bi[i+1] - bi[i];
96: for (j=0; j<nz; j++) {
97: x = rtmp+9*pj[j];
98: pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3];
99: pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8];
100: pv += 9;
101: }
102: /* invert diagonal block */
103: w = ba + 9*diag_offset[i];
104: Kernel_A_gets_inverse_A_3(w);
105: }
107: PetscFree(rtmp);
108: C->factor = FACTOR_LU;
109: C->assembled = PETSC_TRUE;
110: PetscLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */
111: return(0);
112: }