1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
//! Incomplete Cholesky CG solver
use crate::fea::SparseMatrix;
use crate::fea::solver::{LinearSolver, SolverCapabilities, SolverError};
/// Incomplete Cholesky CG solver
pub struct IccgSolver {
n: usize,
max_iter: usize,
tol: f64,
l_diag: Vec<f64>,
l_lower: Vec<f64>,
l_row_ptr: Vec<usize>,
l_col_idx: Vec<usize>,
matrix: Option<SparseMatrix>,
}
impl IccgSolver {
pub fn new() -> Self {
Self {
n: 0,
max_iter: 10000,
tol: 1e-10,
l_diag: Vec::new(),
l_lower: Vec::new(),
l_row_ptr: Vec::new(),
l_col_idx: Vec::new(),
matrix: None,
}
}
pub fn with_params(mut self, max_iter: usize, tol: f64) -> Self {
self.max_iter = max_iter;
self.tol = tol;
self
}
}
impl LinearSolver for IccgSolver {
fn name(&self) -> &'static str {
"iccg"
}
fn capabilities(&self) -> SolverCapabilities {
SolverCapabilities::iccg()
}
fn factor(&mut self, matrix: &SparseMatrix) -> Result<(), SolverError> {
// Invalidate any previous factorization first, so a failed factor()
// cannot leave stale preconditioner/factors usable by a later solve().
self.matrix = None;
self.n = 0;
let n = matrix.n;
if n == 0 {
return Err(SolverError::invalid_input("Empty matrix"));
}
// Check symmetry
if !matrix.is_symmetric(1e-12) {
return Err(SolverError::not_symmetric());
}
// Incomplete Cholesky factorization (IC(0))
let mut l_diag = vec![0.0f64; n];
let mut l_lower = Vec::new();
let mut l_col_idx = Vec::new();
let mut l_row_ptr = vec![0usize; n + 1];
// Count nonzeros in lower triangle
for i in 0..n {
for k in matrix.row_ptr[i]..matrix.row_ptr[i + 1] {
let j = matrix.csr_cols[k];
if j <= i {
l_row_ptr[i + 1] += 1;
}
}
}
for i in 1..=n {
l_row_ptr[i] += l_row_ptr[i - 1];
}
l_lower.resize(l_row_ptr[n], 0.0);
l_col_idx.resize(l_row_ptr[n], 0);
// Fill column indices
let mut pos = l_row_ptr.clone();
for i in 0..n {
for k in matrix.row_ptr[i]..matrix.row_ptr[i + 1] {
let j = matrix.csr_cols[k];
if j <= i {
let idx = pos[i];
l_col_idx[idx] = j;
pos[i] += 1;
}
}
}
// IC(0) factorization
for i in 0..n {
let _diag_pos = l_row_ptr[i + 1] - 1; // Diagonal is last in row
// Compute diagonal
let mut sum = 0.0;
for k in matrix.row_ptr[i]..matrix.row_ptr[i + 1] {
let j = matrix.csr_cols[k];
let val = matrix.csr_vals[k];
if j == i {
sum = val;
} else if j < i {
// Find L_ij in our structure
for idx in l_row_ptr[j]..l_row_ptr[j + 1] {
if l_col_idx[idx] == i {
sum -= l_lower[idx] * l_lower[idx] * l_diag[j];
break;
}
}
}
}
if sum <= 0.0 {
// Modified IC: add small positive value
l_diag[i] = 1e-12;
} else {
l_diag[i] = sum.sqrt();
}
// Compute off-diagonals
for idx in l_row_ptr[i]..l_row_ptr[i + 1] - 1 {
let j = l_col_idx[idx];
// Find A_ij
let mut a_ij = 0.0;
for k in matrix.row_ptr[i]..matrix.row_ptr[i + 1] {
if matrix.csr_cols[k] == j {
a_ij = matrix.csr_vals[k];
break;
}
}
let mut sum = a_ij;
// Subtract contributions
for idx2 in l_row_ptr[j]..l_row_ptr[j + 1] {
let k = l_col_idx[idx2];
if k < j {
// Find L_ik
for idx3 in l_row_ptr[i]..l_row_ptr[i + 1] {
if l_col_idx[idx3] == k {
sum -= l_lower[idx3] * l_lower[idx2] * l_diag[k];
break;
}
}
}
}
l_lower[idx] = sum / l_diag[j];
}
}
self.n = n;
self.l_diag = l_diag;
self.l_lower = l_lower;
self.l_row_ptr = l_row_ptr;
self.l_col_idx = l_col_idx;
self.matrix = Some(matrix.clone());
Ok(())
}
fn solve(&self, rhs: &[f64]) -> Result<Vec<f64>, SolverError> {
let matrix = self
.matrix
.as_ref()
.ok_or_else(|| SolverError::not_factorized())?;
let n = self.n;
if rhs.len() != n {
return Err(SolverError::dimension_mismatch(n, rhs.len()));
}
// Check for NaN/Inf
for (i, &val) in rhs.iter().enumerate() {
if !val.is_finite() {
return Err(SolverError::non_finite(i));
}
}
// Preconditioned CG with IC(0)
let mut x = vec![0.0f64; n];
let mut r = rhs.to_vec();
// Apply preconditioner: M = L L^T
let mut z = self.apply_preconditioner(&r)?;
let mut p = z.clone();
let mut rz_old = r.iter().zip(z.iter()).map(|(ri, zi)| ri * zi).sum::<f64>();
if rz_old.sqrt() < self.tol {
return Ok(x);
}
for _iter in 0..self.max_iter {
// A * p
let mut ap = vec![0.0f64; n];
for i in 0..n {
let mut sum = 0.0;
for k in matrix.row_ptr[i]..matrix.row_ptr[i + 1] {
let j = matrix.csr_cols[k];
sum += matrix.csr_vals[k] * p[j];
}
ap[i] = sum;
}
let p_ap: f64 = p.iter().zip(ap.iter()).map(|(pi, api)| pi * api).sum();
if p_ap <= 0.0 {
return Err(SolverError::singular(
"Matrix not positive definite (p^T A p <= 0)".to_string(),
));
}
let alpha = rz_old / p_ap;
for i in 0..n {
x[i] += alpha * p[i];
r[i] -= alpha * ap[i];
}
z = self.apply_preconditioner(&r)?;
let rz_new = r.iter().zip(z.iter()).map(|(ri, zi)| ri * zi).sum::<f64>();
if rz_new.sqrt() < self.tol {
return Ok(x);
}
let beta = rz_new / rz_old;
for i in 0..n {
p[i] = z[i] + beta * p[i];
}
rz_old = rz_new;
}
Err(SolverError::not_converged())
}
}
impl IccgSolver {
fn apply_preconditioner(&self, r: &[f64]) -> Result<Vec<f64>, SolverError> {
let n = self.n;
let mut y = vec![0.0f64; n];
let mut z = vec![0.0f64; n];
// Forward: L y = r
for i in 0..n {
let mut sum = r[i];
for idx in self.l_row_ptr[i]..self.l_row_ptr[i + 1] - 1 {
let j = self.l_col_idx[idx];
sum -= self.l_lower[idx] * y[j];
}
y[i] = sum / self.l_diag[i];
}
// Backward: L^T z = y
for i in (0..n).rev() {
let mut sum = y[i];
for idx in self.l_row_ptr[i]..self.l_row_ptr[i + 1] - 1 {
let j = self.l_col_idx[idx];
sum -= self.l_lower[idx] * z[j];
}
z[i] = sum / self.l_diag[i];
}
Ok(z)
}
}