Actual source code: baijfact11.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 4 by 4
12: */
15: PetscErrorCode MatLUFactorNumeric_SeqBAIJ_4(Mat C,Mat A,const MatFactorInfo *info)
16: {
17: Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ *)C->data;
18: IS isrow = b->row,isicol = b->icol;
20: const PetscInt *r,*ic;
21: PetscInt i,j,n = a->mbs,*bi = b->i,*bj = b->j;
22: PetscInt *ajtmpold,*ajtmp,nz,row;
23: PetscInt *diag_offset = b->diag,idx,*ai=a->i,*aj=a->j,*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,x10,x11,x12,x13,x14,x15,x16;
27: MatScalar p10,p11,p12,p13,p14,p15,p16,m10,m11,m12;
28: MatScalar m13,m14,m15,m16;
29: MatScalar *ba = b->a,*aa = a->a;
30: PetscTruth pivotinblocks = b->pivotinblocks;
31: PetscReal shift = info->shiftinblocks;
34: ISGetIndices(isrow,&r);
35: ISGetIndices(isicol,&ic);
36: PetscMalloc(16*(n+1)*sizeof(MatScalar),&rtmp);
38: for (i=0; i<n; i++) {
39: nz = bi[i+1] - bi[i];
40: ajtmp = bj + bi[i];
41: for (j=0; j<nz; j++) {
42: x = rtmp+16*ajtmp[j];
43: x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0;
44: x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0;
45: }
46: /* load in initial (unfactored row) */
47: idx = r[i];
48: nz = ai[idx+1] - ai[idx];
49: ajtmpold = aj + ai[idx];
50: v = aa + 16*ai[idx];
51: for (j=0; j<nz; j++) {
52: x = rtmp+16*ic[ajtmpold[j]];
53: x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3];
54: x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8];
55: x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13];
56: x[14] = v[14]; x[15] = v[15];
57: v += 16;
58: }
59: row = *ajtmp++;
60: while (row < i) {
61: pc = rtmp + 16*row;
62: p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3];
63: p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8];
64: p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13];
65: p15 = pc[14]; p16 = pc[15];
66: if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 ||
67: p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 ||
68: p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0
69: || p16 != 0.0) {
70: pv = ba + 16*diag_offset[row];
71: pj = bj + diag_offset[row] + 1;
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: x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13];
75: x15 = pv[14]; x16 = pv[15];
76: pc[0] = m1 = p1*x1 + p5*x2 + p9*x3 + p13*x4;
77: pc[1] = m2 = p2*x1 + p6*x2 + p10*x3 + p14*x4;
78: pc[2] = m3 = p3*x1 + p7*x2 + p11*x3 + p15*x4;
79: pc[3] = m4 = p4*x1 + p8*x2 + p12*x3 + p16*x4;
81: pc[4] = m5 = p1*x5 + p5*x6 + p9*x7 + p13*x8;
82: pc[5] = m6 = p2*x5 + p6*x6 + p10*x7 + p14*x8;
83: pc[6] = m7 = p3*x5 + p7*x6 + p11*x7 + p15*x8;
84: pc[7] = m8 = p4*x5 + p8*x6 + p12*x7 + p16*x8;
86: pc[8] = m9 = p1*x9 + p5*x10 + p9*x11 + p13*x12;
87: pc[9] = m10 = p2*x9 + p6*x10 + p10*x11 + p14*x12;
88: pc[10] = m11 = p3*x9 + p7*x10 + p11*x11 + p15*x12;
89: pc[11] = m12 = p4*x9 + p8*x10 + p12*x11 + p16*x12;
91: pc[12] = m13 = p1*x13 + p5*x14 + p9*x15 + p13*x16;
92: pc[13] = m14 = p2*x13 + p6*x14 + p10*x15 + p14*x16;
93: pc[14] = m15 = p3*x13 + p7*x14 + p11*x15 + p15*x16;
94: pc[15] = m16 = p4*x13 + p8*x14 + p12*x15 + p16*x16;
96: nz = bi[row+1] - diag_offset[row] - 1;
97: pv += 16;
98: for (j=0; j<nz; j++) {
99: x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3];
100: x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8];
101: x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12];
102: x14 = pv[13]; x15 = pv[14]; x16 = pv[15];
103: x = rtmp + 16*pj[j];
104: x[0] -= m1*x1 + m5*x2 + m9*x3 + m13*x4;
105: x[1] -= m2*x1 + m6*x2 + m10*x3 + m14*x4;
106: x[2] -= m3*x1 + m7*x2 + m11*x3 + m15*x4;
107: x[3] -= m4*x1 + m8*x2 + m12*x3 + m16*x4;
109: x[4] -= m1*x5 + m5*x6 + m9*x7 + m13*x8;
110: x[5] -= m2*x5 + m6*x6 + m10*x7 + m14*x8;
111: x[6] -= m3*x5 + m7*x6 + m11*x7 + m15*x8;
112: x[7] -= m4*x5 + m8*x6 + m12*x7 + m16*x8;
114: x[8] -= m1*x9 + m5*x10 + m9*x11 + m13*x12;
115: x[9] -= m2*x9 + m6*x10 + m10*x11 + m14*x12;
116: x[10] -= m3*x9 + m7*x10 + m11*x11 + m15*x12;
117: x[11] -= m4*x9 + m8*x10 + m12*x11 + m16*x12;
119: x[12] -= m1*x13 + m5*x14 + m9*x15 + m13*x16;
120: x[13] -= m2*x13 + m6*x14 + m10*x15 + m14*x16;
121: x[14] -= m3*x13 + m7*x14 + m11*x15 + m15*x16;
122: x[15] -= m4*x13 + m8*x14 + m12*x15 + m16*x16;
124: pv += 16;
125: }
126: PetscLogFlops(128*nz+112);
127: }
128: row = *ajtmp++;
129: }
130: /* finished row so stick it into b->a */
131: pv = ba + 16*bi[i];
132: pj = bj + bi[i];
133: nz = bi[i+1] - bi[i];
134: for (j=0; j<nz; j++) {
135: x = rtmp+16*pj[j];
136: pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3];
137: pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8];
138: pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12];
139: pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15];
140: pv += 16;
141: }
142: /* invert diagonal block */
143: w = ba + 16*diag_offset[i];
144: if (pivotinblocks) {
145: Kernel_A_gets_inverse_A_4(w,shift);
146: } else {
147: Kernel_A_gets_inverse_A_4_nopivot(w);
148: }
149: }
151: PetscFree(rtmp);
152: ISRestoreIndices(isicol,&ic);
153: ISRestoreIndices(isrow,&r);
154: C->ops->solve = MatSolve_SeqBAIJ_4;
155: C->ops->solvetranspose = MatSolveTranspose_SeqBAIJ_4;
156: C->assembled = PETSC_TRUE;
157: PetscLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */
158: return(0);
159: }