primal-barracuda 0.4.0

barraCuda — BARrier-free Rust Abstracted Cross-platform Unified Dimensional Algebra. Vendor-agnostic GPU compute via WGSL.
Documentation
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
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Non-Negative Matrix Factorization (NMF) — Lee & Seung multiplicative updates.
//!
//! Factorises a non-negative matrix **V** (m x n) into two non-negative factors
//! **W** (m x k) and **H** (k x n) such that V ≈ WH.
//!
//! Two objective functions are supported:
//! - **Euclidean** (Frobenius): minimise ‖V − `WH‖²_F`
//! - **KL divergence**: minimise `D_KL(V` ‖ WH)
//!
//! Absorbed from wetSpring bio module (Feb 2026). CPU-only implementation;
//! GPU WGSL generation is future work (OdeSystem-style trait pattern).
//!
//! # References
//!
//! Lee, D.D. & Seung, H.S. (1999). Learning the parts of objects by
//! non-negative matrix factorization. *Nature*, 401, 788–791.
//!
//! Lee, D.D. & Seung, H.S. (2000). Algorithms for Non-negative Matrix
//! Factorization. *`NeurIPS` 2000*.

use crate::error::BarracudaError;

/// Result of an NMF factorisation.
#[derive(Debug, Clone)]
pub struct NmfResult {
    /// Factor matrix W (m x k), stored row-major.
    pub w: Vec<f64>,
    /// Factor matrix H (k x n), stored row-major.
    pub h: Vec<f64>,
    /// Number of rows in V.
    pub m: usize,
    /// Factorisation rank.
    pub k: usize,
    /// Number of columns in V.
    pub n: usize,
    /// Reconstruction error at each iteration.
    pub errors: Vec<f64>,
}

/// Objective function for NMF.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum NmfObjective {
    /// Minimise ‖V − `WH‖²_F` (Frobenius norm).
    Euclidean,
    /// Minimise `D_KL(V` ‖ WH) (generalised Kullback-Leibler divergence).
    KlDivergence,
}

/// Configuration for NMF.
#[derive(Debug, Clone)]
pub struct NmfConfig {
    /// Number of latent factors.
    pub rank: usize,
    /// Maximum number of iterations.
    pub max_iter: usize,
    /// Convergence tolerance (relative change in error).
    pub tol: f64,
    /// Objective function to minimise.
    pub objective: NmfObjective,
    /// Random seed for initialisation.
    pub seed: u64,
}

impl Default for NmfConfig {
    fn default() -> Self {
        Self {
            rank: 10,
            max_iter: 200,
            tol: 1e-4,
            objective: NmfObjective::Euclidean,
            seed: 42,
        }
    }
}

/// Simple LCG PRNG for reproducible initialisation (no external dep).
struct Lcg(u64);

impl Lcg {
    fn new(seed: u64) -> Self {
        Self(seed.wrapping_add(1))
    }
    fn next_f64(&mut self) -> f64 {
        self.0 = self
            .0
            .wrapping_mul(6_364_136_223_846_793_005)
            .wrapping_add(1);
        let bits = (self.0 >> 11) | 0x3FF0_0000_0000_0000;
        f64::from_bits(bits) - 1.0
    }
}

/// Run NMF with Lee & Seung multiplicative update rules.
///
/// `v` is the input matrix (m x n) in row-major order. All entries must be >= 0.
///
/// # Errors
///
/// Returns `BarracudaError::InvalidShape` if dimensions are inconsistent or rank
/// is out of range.
pub fn nmf(v: &[f64], m: usize, n: usize, config: &NmfConfig) -> Result<NmfResult, BarracudaError> {
    let k = config.rank;
    if v.len() != m * n {
        return Err(BarracudaError::InvalidShape {
            expected: vec![m, n],
            actual: vec![v.len()],
        });
    }
    if k == 0 || k > m.min(n) {
        return Err(BarracudaError::InvalidShape {
            expected: vec![1, m.min(n)],
            actual: vec![k],
        });
    }

    let mut rng = Lcg::new(config.seed);

    let mut w = vec![0.0; m * k];
    let mut h = vec![0.0; k * n];
    for val in &mut w {
        *val = rng.next_f64().mul_add(0.1, 1e-10);
    }
    for val in &mut h {
        *val = rng.next_f64().mul_add(0.1, 1e-10);
    }

    let mut errors = Vec::with_capacity(config.max_iter);
    let eps = 1e-15;

    match config.objective {
        NmfObjective::Euclidean => {
            nmf_euclidean(v, &mut w, &mut h, m, n, k, config, &mut errors, eps);
        }
        NmfObjective::KlDivergence => {
            nmf_kl(v, &mut w, &mut h, m, n, k, config, &mut errors, eps);
        }
    }

    Ok(NmfResult {
        w,
        h,
        m,
        k,
        n,
        errors,
    })
}

/// Euclidean (Frobenius) NMF — multiplicative updates.
fn nmf_euclidean(
    v: &[f64],
    w: &mut [f64],
    h: &mut [f64],
    m: usize,
    n: usize,
    k: usize,
    config: &NmfConfig,
    errors: &mut Vec<f64>,
    eps: f64,
) {
    let mut wh = vec![0.0; m * n];
    let mut v_ht = vec![0.0; m * k];
    let mut w_h_ht = vec![0.0; m * k];
    let mut wt_v = vec![0.0; k * n];
    let mut wt_w_h = vec![0.0; k * n];
    let mut h_ht = vec![0.0; k * k];
    let mut wt_w = vec![0.0; k * k];

    for iter in 0..config.max_iter {
        matmul(w, h, &mut wh, m, k, n);
        let err = frobenius_error(v, &wh);
        errors.push(err);
        if iter > 0 && (errors[iter - 1] - err).abs() / (errors[iter - 1] + eps) < config.tol {
            break;
        }

        matmul_at_b(w, v, &mut wt_v, m, k, n);
        matmul_at_b(w, w, &mut wt_w, m, k, k);
        matmul(&wt_w, h, &mut wt_w_h, k, k, n);
        for i in 0..k * n {
            h[i] *= wt_v[i] / (wt_w_h[i] + eps);
        }

        matmul_a_bt(v, h, &mut v_ht, m, n, k);
        matmul_a_bt(h, h, &mut h_ht, k, n, k);
        matmul(w, &h_ht, &mut w_h_ht, m, k, k);
        for i in 0..m * k {
            w[i] *= v_ht[i] / (w_h_ht[i] + eps);
        }
    }
}

/// KL-divergence NMF — multiplicative updates.
fn nmf_kl(
    v: &[f64],
    w: &mut [f64],
    h: &mut [f64],
    m: usize,
    n: usize,
    k: usize,
    config: &NmfConfig,
    errors: &mut Vec<f64>,
    eps: f64,
) {
    let mut wh = vec![0.0; m * n];
    let mut ratio = vec![0.0; m * n];
    let mut wt_ratio = vec![0.0; k * n];
    let mut ratio_ht = vec![0.0; m * k];

    for iter in 0..config.max_iter {
        matmul(w, h, &mut wh, m, k, n);
        let err = kl_divergence(v, &wh, eps);
        errors.push(err);
        if iter > 0 && (errors[iter - 1] - err).abs() / (errors[iter - 1] + eps) < config.tol {
            break;
        }

        for i in 0..m * n {
            ratio[i] = v[i] / (wh[i] + eps);
        }

        matmul_at_b(w, &ratio, &mut wt_ratio, m, k, n);
        let w_col_sums = col_sums_of_transposed(w, m, k);
        for j in 0..n {
            for kk in 0..k {
                h[kk * n + j] *= wt_ratio[kk * n + j] / (w_col_sums[kk] + eps);
            }
        }

        matmul(w, h, &mut wh, m, k, n);
        for i in 0..m * n {
            ratio[i] = v[i] / (wh[i] + eps);
        }

        matmul_a_bt(&ratio, h, &mut ratio_ht, m, n, k);
        let h_row_sums = row_sums(h, k, n);
        for i in 0..m {
            for kk in 0..k {
                w[i * k + kk] *= ratio_ht[i * k + kk] / (h_row_sums[kk] + eps);
            }
        }
    }
}

// ── Dense linear algebra helpers ─────────────────────────────────

fn matmul(a: &[f64], b: &[f64], c: &mut [f64], m: usize, k: usize, n: usize) {
    c.fill(0.0);
    for i in 0..m {
        for p in 0..k {
            let a_ip = a[i * k + p];
            for j in 0..n {
                c[i * n + j] = a_ip.mul_add(b[p * n + j], c[i * n + j]);
            }
        }
    }
}

fn matmul_at_b(a: &[f64], b: &[f64], c: &mut [f64], m: usize, k: usize, n: usize) {
    c.fill(0.0);
    for i in 0..m {
        for p in 0..k {
            let a_ip = a[i * k + p];
            for j in 0..n {
                c[p * n + j] = a_ip.mul_add(b[i * n + j], c[p * n + j]);
            }
        }
    }
}

fn matmul_a_bt(a: &[f64], b: &[f64], c: &mut [f64], m: usize, n: usize, k: usize) {
    c.fill(0.0);
    for i in 0..m {
        for p in 0..n {
            let a_ip = a[i * n + p];
            for j in 0..k {
                c[i * k + j] = a_ip.mul_add(b[j * n + p], c[i * k + j]);
            }
        }
    }
}

fn frobenius_error(a: &[f64], b: &[f64]) -> f64 {
    a.iter()
        .zip(b.iter())
        .map(|(ai, bi)| (ai - bi).powi(2))
        .sum::<f64>()
        .sqrt()
}

fn kl_divergence(v: &[f64], wh: &[f64], eps: f64) -> f64 {
    v.iter()
        .zip(wh.iter())
        .map(|(vi, whi)| {
            if *vi > eps {
                vi * (vi / (whi + eps)).ln() - vi + whi
            } else {
                *whi
            }
        })
        .sum()
}

fn col_sums_of_transposed(a: &[f64], m: usize, k: usize) -> Vec<f64> {
    let mut sums = vec![0.0; k];
    for i in 0..m {
        for j in 0..k {
            sums[j] += a[i * k + j];
        }
    }
    sums
}

fn row_sums(a: &[f64], k: usize, n: usize) -> Vec<f64> {
    let mut sums = vec![0.0; k];
    for i in 0..k {
        for j in 0..n {
            sums[i] += a[i * n + j];
        }
    }
    sums
}

// ── Scoring helpers ──────────────────────────────────────────────

/// Cosine similarity between two vectors.
/// # Panics
/// Panics if `a.len() != b.len()`.
#[must_use]
pub fn cosine_similarity(a: &[f64], b: &[f64]) -> f64 {
    assert_eq!(a.len(), b.len());
    let dot: f64 = a.iter().zip(b.iter()).map(|(x, y)| x * y).sum();
    let na: f64 = a.iter().map(|x| x * x).sum::<f64>().sqrt();
    let nb: f64 = b.iter().map(|x| x * x).sum::<f64>().sqrt();
    if na < 1e-15 || nb < 1e-15 {
        return 0.0;
    }
    dot / (na * nb)
}

/// Score all row-column pairs by reconstructing V ≈ WH and return
/// the top-K pairs sorted by descending score.
///
/// Returns `Vec<(row_idx, col_idx, score)>`.
/// # Panics
/// Panics if `partial_cmp` returns `None` during sort (e.g. NaN values).
#[must_use]
pub fn top_k_predictions(result: &NmfResult, top_k: usize) -> Vec<(usize, usize, f64)> {
    let m = result.m;
    let n = result.n;
    let k = result.k;

    let mut v_hat = vec![0.0; m * n];
    matmul(&result.w, &result.h, &mut v_hat, m, k, n);

    let mut pairs: Vec<(usize, usize, f64)> = Vec::with_capacity(m * n);
    for i in 0..m {
        for j in 0..n {
            pairs.push((i, j, v_hat[i * n + j]));
        }
    }

    pairs.sort_by(|a, b| b.2.partial_cmp(&a.2).unwrap_or(std::cmp::Ordering::Equal));
    pairs.truncate(top_k);
    pairs
}

/// Reconstruction error: ‖V − WH‖_F / ‖V‖_F (relative).
#[must_use]
pub fn relative_reconstruction_error(v: &[f64], result: &NmfResult) -> f64 {
    let mut wh = vec![0.0; result.m * result.n];
    matmul(&result.w, &result.h, &mut wh, result.m, result.k, result.n);
    let err = frobenius_error(v, &wh);
    let v_norm: f64 = v.iter().map(|x| x * x).sum::<f64>().sqrt();
    if v_norm < 1e-15 {
        return 0.0;
    }
    err / v_norm
}

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

    #[test]
    fn euclidean_converges() {
        let v = vec![
            1.0, 2.0, 0.0, 1.0, 0.5, 0.0, 1.0, 3.0, 0.0, 2.0, 2.0, 0.0, 1.0, 2.0, 0.0, 1.0, 1.0,
            0.0, 3.0, 1.0,
        ];
        let config = NmfConfig {
            rank: 2,
            max_iter: 100,
            tol: 1e-6,
            objective: NmfObjective::Euclidean,
            seed: 42,
        };
        let result = nmf(&v, 4, 5, &config).expect("nmf");
        assert_eq!(result.m, 4);
        assert_eq!(result.n, 5);
        assert_eq!(result.k, 2);
        assert!(!result.errors.is_empty());
        for pair in result.errors.windows(2) {
            assert!(
                pair[1] <= pair[0] + 1e-10,
                "error not monotonically decreasing"
            );
        }
        let rel_err = relative_reconstruction_error(&v, &result);
        assert!(rel_err < 0.5, "relative error {rel_err} too high");
    }

    #[test]
    fn kl_divergence_converges() {
        let v = vec![
            1.0, 2.0, 0.5, 1.0, 0.5, 0.1, 1.0, 3.0, 0.1, 2.0, 2.0, 0.1, 1.0, 2.0, 0.1, 1.0, 1.0,
            0.1, 3.0, 1.0,
        ];
        let config = NmfConfig {
            rank: 2,
            max_iter: 100,
            tol: 1e-6,
            objective: NmfObjective::KlDivergence,
            seed: 123,
        };
        let result = nmf(&v, 4, 5, &config).expect("nmf");
        assert!(!result.errors.is_empty());
        let last = *result.errors.last().unwrap();
        assert!(last < result.errors[0], "KL error should decrease");
    }

    #[test]
    fn cosine_orthogonal() {
        let a = [1.0, 0.0, 0.0];
        let b = [0.0, 1.0, 0.0];
        assert!(cosine_similarity(&a, &b).abs() < 1e-10);
    }

    #[test]
    fn cosine_45_degrees() {
        let a = [1.0, 0.0, 0.0];
        let c = [1.0, 1.0, 0.0];
        let expected = 1.0 / 2.0_f64.sqrt();
        assert!((cosine_similarity(&a, &c) - expected).abs() < 1e-10);
    }

    #[test]
    fn top_k_diagonal_dominant() {
        let v = vec![5.0, 0.0, 0.0, 0.0, 5.0, 0.0, 0.0, 0.0, 5.0];
        let config = NmfConfig {
            rank: 2,
            max_iter: 200,
            tol: 1e-8,
            objective: NmfObjective::Euclidean,
            seed: 42,
        };
        let result = nmf(&v, 3, 3, &config).expect("nmf");
        let top = top_k_predictions(&result, 3);
        assert_eq!(top.len(), 3);
        assert!(top[0].2 > 1.0);
    }

    #[test]
    fn invalid_shape_rejected() {
        let v = vec![1.0, 2.0, 3.0];
        let config = NmfConfig {
            rank: 2,
            ..NmfConfig::default()
        };
        assert!(nmf(&v, 4, 5, &config).is_err());
    }

    #[test]
    fn zero_rank_rejected() {
        let v = vec![1.0; 12];
        let config = NmfConfig {
            rank: 0,
            ..NmfConfig::default()
        };
        assert!(nmf(&v, 3, 4, &config).is_err());
    }

    #[test]
    fn factors_are_nonnegative() {
        let v = vec![1.0, 2.0, 0.0, 1.0, 0.0, 1.0, 3.0, 0.0, 2.0, 0.0, 1.0, 2.0];
        let config = NmfConfig {
            rank: 2,
            max_iter: 50,
            ..NmfConfig::default()
        };
        let result = nmf(&v, 3, 4, &config).expect("nmf");
        assert!(result.w.iter().all(|x| *x >= 0.0), "W must be non-negative");
        assert!(result.h.iter().all(|x| *x >= 0.0), "H must be non-negative");
    }
}