kryst 3.2.1

Krylov subspace and preconditioned iterative solvers for dense and sparse linear systems, with shared and distributed memory parallelism.
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

#[allow(unused_imports)]
use crate::algebra::blas::{dot_conj, nrm2};
#[allow(unused_imports)]
use crate::algebra::prelude::*;

use super::{NgSymmetry, rap_ops::CsrPattern};

#[derive(Clone, Copy)]
pub(crate) struct NgRowFilter {
    pub tau_abs: f64,
    pub tau_rel: f64,
    pub k_max: usize,
    pub lump_diag: bool,
}

pub(crate) fn non_galerkin_filter_coarse(
    pat: &CsrPattern,
    vals: &[f64],
    symmetry: NgSymmetry,
    rf: NgRowFilter,
) -> (CsrPattern, Vec<f64>, Vec<Option<usize>>) {
    let m = pat.nrows;
    let pr = &pat.row_ptr;
    let pc = &pat.col_idx;
    let nnz = pc.len();

    let mut keep = vec![false; nnz];

    // First pass: row-wise filtering
    for i in 0..m {
        let rs = pr[i];
        let re = pr[i + 1];
        if rs == re {
            continue;
        }
        let mut idx: Vec<usize> = (rs..re).collect();
        // absolute and relative drops
        if rf.tau_abs > 0.0 || rf.tau_rel > 0.0 || (rf.k_max > 0 && re - rs > rf.k_max) {
            // sort indices by |v| asc then col
            idx.sort_unstable_by(|&u, &v| {
                let au = vals[u].abs();
                let av = vals[v].abs();
                au.total_cmp(&av).then_with(|| pc[u].cmp(&pc[v]))
            });
            let mut drop_mask = vec![false; re - rs];
            let mut dropped_sum = 0.0f64;
            let mut l1 = 0.0f64;
            for t in rs..re {
                if pc[t] != i {
                    l1 += vals[t].abs();
                }
            }
            for &t in idx.iter() {
                let col = pc[t];
                let v = vals[t];
                if col == i {
                    continue;
                }
                let by_abs = v.abs() < rf.tau_abs;
                let allow = rf.tau_rel * l1;
                let by_rel = dropped_sum + v.abs() <= allow + 1e-300;
                if by_abs || (rf.tau_rel > 0.0 && by_rel) {
                    drop_mask[t - rs] = true;
                    dropped_sum += v.abs();
                }
            }
            // cap
            if rf.k_max > 0 {
                let mut order_keep: Vec<usize> = (rs..re).collect();
                order_keep.sort_unstable_by(|&u, &v| {
                    let au = vals[u].abs();
                    let av = vals[v].abs();
                    av.total_cmp(&au).then_with(|| pc[u].cmp(&pc[v]))
                });
                let mut kept_off = 0usize;
                for &t in &order_keep {
                    if pc[t] == i {
                        keep[t] = true;
                        continue;
                    }
                    if kept_off < rf.k_max && !drop_mask[t - rs] {
                        keep[t] = true;
                        kept_off += 1;
                    }
                }
                // mark remaining non-dropped as keep
                for t in rs..re {
                    if pc[t] == i {
                        keep[t] = true;
                    } else if !drop_mask[t - rs] && !keep[t] && kept_off < rf.k_max {
                        keep[t] = true;
                        kept_off += 1;
                    }
                }
            } else {
                for t in rs..re {
                    if pc[t] == i || !drop_mask[t - rs] {
                        keep[t] = true;
                    }
                }
            }
        } else {
            for t in rs..re {
                keep[t] = true;
            }
        }

        // safety: ensure at least diag and one off-diagonal if row had any off-diagonals
        let kept_off = (rs..re).filter(|&t| keep[t] && pc[t] != i).count();
        let had_off = (rs..re).any(|t| pc[t] != i);
        if kept_off == 0 && had_off {
            // keep largest off-diagonal
            let mut best = rs;
            let mut best_mag = 0.0;
            for t in rs..re {
                if pc[t] != i {
                    let mag = vals[t].abs();
                    if mag > best_mag {
                        best = t;
                        best_mag = mag;
                    }
                }
            }
            keep[best] = true;
        }
        // always keep diagonal if present
        for t in rs..re {
            if pc[t] == i {
                keep[t] = true;
                break;
            }
        }
    }

    // symmetry enforcement
    if let NgSymmetry::Symmetric = symmetry {
        for i in 0..m {
            let rs = pr[i];
            let re = pr[i + 1];
            for t in rs..re {
                let j = pc[t];
                if i == j {
                    continue;
                }
                let rjs = pr[j];
                let rje = pr[j + 1];
                if let Ok(pos) = pc[rjs..rje].binary_search(&i) {
                    let tj = rjs + pos;
                    if keep[t] ^ keep[tj] {
                        keep[t] = true;
                        keep[tj] = true;
                    }
                }
            }
        }
    }

    // build filtered pattern and values
    let mut ng_row_ptr = Vec::with_capacity(m + 1);
    let mut ng_col_idx = Vec::new();
    let mut ng_vals = Vec::new();
    let mut full2ng = vec![None; nnz];
    ng_row_ptr.push(0);

    for i in 0..m {
        let rs = pr[i];
        let re = pr[i + 1];
        let mut diag_add = 0.0;
        for t in rs..re {
            let j = pc[t];
            if keep[t] {
                full2ng[t] = Some(ng_col_idx.len());
                ng_col_idx.push(j);
                ng_vals.push(vals[t]);
            } else if rf.lump_diag && j != i {
                diag_add += vals[t];
            }
        }
        if rf.lump_diag && diag_add != 0.0 {
            // find diagonal position
            let row_start = ng_row_ptr.last().copied().unwrap();
            if let Ok(pos) = ng_col_idx[row_start..].binary_search(&i) {
                let idx = row_start + pos;
                ng_vals[idx] += diag_add;
            } else {
                // insert diag
                let pos = match ng_col_idx[row_start..].binary_search(&i) {
                    Ok(p) => row_start + p,
                    Err(p) => row_start + p,
                };
                ng_col_idx.insert(pos, i);
                ng_vals.insert(pos, diag_add);
            }
        }
        ng_row_ptr.push(ng_col_idx.len());
    }

    let ng_pat = CsrPattern {
        nrows: pat.nrows,
        ncols: pat.ncols,
        row_ptr: ng_row_ptr,
        col_idx: ng_col_idx,
    };

    (ng_pat, ng_vals, full2ng)
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn basic_drop() {
        // 3x3 dense pattern
        let pat = CsrPattern {
            nrows: 3,
            ncols: 3,
            row_ptr: vec![0, 3, 6, 9],
            col_idx: vec![0, 1, 2, 0, 1, 2, 0, 1, 2],
        };
        let vals = vec![4.0, -0.1, 0.05, -0.1, 5.0, 0.02, 0.05, 0.02, 6.0];
        let rf = NgRowFilter {
            tau_abs: 0.1,
            tau_rel: 0.0,
            k_max: 0,
            lump_diag: true,
        };
        let (ng_pat, ng_vals, _) =
            non_galerkin_filter_coarse(&pat, &vals, NgSymmetry::Symmetric, rf);
        assert_eq!(ng_pat.row_ptr, vec![0, 3, 5, 7]);
        assert_eq!(ng_pat.col_idx, vec![0, 1, 2, 0, 1, 0, 2]);
        let expected_vals = vec![4.0, -0.1, 0.05, -0.1, 5.02, 0.05, 6.02];
        assert_eq!(ng_vals, expected_vals);
    }
}