Actual source code: baijfact13.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
9: /* ------------------------------------------------------------*/
10: /*
11: Version for when blocks are 3 by 3
12: */
15: PetscErrorCode MatLUFactorNumeric_SeqBAIJ_3(Mat A,MatFactorInfo *info,Mat *B)
16: {
17: Mat C = *B;
18: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ *)C->data;
19: IS isrow = b->row,isicol = b->icol;
21: PetscInt *r,*ic,i,j,n = a->mbs,*bi = b->i,*bj = b->j;
22: PetscInt *ajtmpold,*ajtmp,nz,row,*ai=a->i,*aj=a->j;
23: PetscInt *diag_offset = b->diag,idx,*pj;
24: MatScalar *pv,*v,*rtmp,*pc,*w,*x;
25: MatScalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4;
26: MatScalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9;
27: MatScalar *ba = b->a,*aa = a->a;
30: ISGetIndices(isrow,&r);
31: ISGetIndices(isicol,&ic);
32: PetscMalloc(9*(n+1)*sizeof(MatScalar),&rtmp);
34: for (i=0; i<n; i++) {
35: nz = bi[i+1] - bi[i];
36: ajtmp = bj + bi[i];
37: for (j=0; j<nz; j++) {
38: x = rtmp + 9*ajtmp[j];
39: x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = 0.0;
40: }
41: /* load in initial (unfactored row) */
42: idx = r[i];
43: nz = ai[idx+1] - ai[idx];
44: ajtmpold = aj + ai[idx];
45: v = aa + 9*ai[idx];
46: for (j=0; j<nz; j++) {
47: x = rtmp + 9*ic[ajtmpold[j]];
48: x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3];
49: x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8];
50: v += 9;
51: }
52: row = *ajtmp++;
53: while (row < i) {
54: pc = rtmp + 9*row;
55: p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3];
56: p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8];
57: if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 ||
58: p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0) {
59: pv = ba + 9*diag_offset[row];
60: pj = bj + diag_offset[row] + 1;
61: x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3];
62: x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8];
63: pc[0] = m1 = p1*x1 + p4*x2 + p7*x3;
64: pc[1] = m2 = p2*x1 + p5*x2 + p8*x3;
65: pc[2] = m3 = p3*x1 + p6*x2 + p9*x3;
67: pc[3] = m4 = p1*x4 + p4*x5 + p7*x6;
68: pc[4] = m5 = p2*x4 + p5*x5 + p8*x6;
69: pc[5] = m6 = p3*x4 + p6*x5 + p9*x6;
71: pc[6] = m7 = p1*x7 + p4*x8 + p7*x9;
72: pc[7] = m8 = p2*x7 + p5*x8 + p8*x9;
73: pc[8] = m9 = p3*x7 + p6*x8 + p9*x9;
74: nz = bi[row+1] - diag_offset[row] - 1;
75: pv += 9;
76: for (j=0; j<nz; j++) {
77: x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3];
78: x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8];
79: x = rtmp + 9*pj[j];
80: x[0] -= m1*x1 + m4*x2 + m7*x3;
81: x[1] -= m2*x1 + m5*x2 + m8*x3;
82: x[2] -= m3*x1 + m6*x2 + m9*x3;
83:
84: x[3] -= m1*x4 + m4*x5 + m7*x6;
85: x[4] -= m2*x4 + m5*x5 + m8*x6;
86: x[5] -= m3*x4 + m6*x5 + m9*x6;
88: x[6] -= m1*x7 + m4*x8 + m7*x9;
89: x[7] -= m2*x7 + m5*x8 + m8*x9;
90: x[8] -= m3*x7 + m6*x8 + m9*x9;
91: pv += 9;
92: }
93: PetscLogFlops(54*nz+36);
94: }
95: row = *ajtmp++;
96: }
97: /* finished row so stick it into b->a */
98: pv = ba + 9*bi[i];
99: pj = bj + bi[i];
100: nz = bi[i+1] - bi[i];
101: for (j=0; j<nz; j++) {
102: x = rtmp + 9*pj[j];
103: pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3];
104: pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8];
105: pv += 9;
106: }
107: /* invert diagonal block */
108: w = ba + 9*diag_offset[i];
109: Kernel_A_gets_inverse_A_3(w);
110: }
112: PetscFree(rtmp);
113: ISRestoreIndices(isicol,&ic);
114: ISRestoreIndices(isrow,&r);
115: C->factor = FACTOR_LU;
116: C->assembled = PETSC_TRUE;
117: PetscLogFlops(1.3333*27*b->mbs); /* from inverting diagonal blocks */
118: return(0);
119: }