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
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Sparse iterative solvers
//!
//! Provides iterative methods for solving sparse linear systems Ax = b.
//!
//! # Methods
//!
//! - **CG** (Conjugate Gradient): Best for symmetric positive definite (SPD)
//! - **`BiCGSTAB`**: For general non-symmetric systems
//! - **Jacobi**: Simple iteration, good for diagonally dominant systems
//! - **Gauss-Seidel**: Faster than Jacobi for many problems
//!
//! # Convergence
//!
//! All methods use a residual-based stopping criterion:
//! ‖b - Ax‖ / ‖b‖ < tolerance
//!
//! # Preconditioning
//!
//! For faster convergence, consider diagonal (Jacobi) preconditioning:
//! - Cheap: O(n) setup, O(nnz) per iteration
//! - Effective for diagonally dominant systems

use super::csr::CsrMatrix;
use crate::error::{BarracudaError, Result};

/// Solver configuration
#[derive(Debug, Clone)]
pub struct SolverConfig {
    /// Convergence tolerance (relative residual)
    pub tolerance: f64,
    /// Maximum iterations
    pub max_iterations: usize,
    /// Use diagonal preconditioning
    pub use_preconditioner: bool,
    /// Verbose output
    pub verbose: bool,
}

impl Default for SolverConfig {
    fn default() -> Self {
        Self {
            tolerance: 1e-10,
            max_iterations: 1000,
            use_preconditioner: true,
            verbose: false,
        }
    }
}

impl SolverConfig {
    /// Create with custom tolerance
    #[must_use]
    pub fn with_tolerance(mut self, tol: f64) -> Self {
        self.tolerance = tol;
        self
    }

    /// Create with custom max iterations
    #[must_use]
    pub fn with_max_iterations(mut self, max: usize) -> Self {
        self.max_iterations = max;
        self
    }

    /// Disable preconditioning
    #[must_use]
    pub fn no_preconditioner(mut self) -> Self {
        self.use_preconditioner = false;
        self
    }
}

/// Solver result
#[derive(Debug, Clone)]
pub struct SolverResult {
    /// Solution vector
    pub x: Vec<f64>,
    /// Number of iterations performed
    pub iterations: usize,
    /// Final relative residual
    pub residual: f64,
    /// Whether convergence was achieved
    pub converged: bool,
}

impl SolverResult {
    /// Check if solve was successful
    #[must_use]
    pub fn is_ok(&self) -> bool {
        self.converged
    }
}

/// Compute vector dot product
fn dot(a: &[f64], b: &[f64]) -> f64 {
    a.iter().zip(b.iter()).map(|(&x, &y)| x * y).sum()
}

/// Compute vector L2 norm
fn norm(v: &[f64]) -> f64 {
    dot(v, v).sqrt()
}

/// Vector: y = a*x + y
fn axpy(a: f64, x: &[f64], y: &mut [f64]) {
    for (yi, &xi) in y.iter_mut().zip(x.iter()) {
        *yi = a.mul_add(xi, *yi);
    }
}

/// Vector: y = x - y (y = x + (-1)*y)
fn sub_into(x: &[f64], y: &mut [f64]) {
    for (yi, &xi) in y.iter_mut().zip(x.iter()) {
        *yi = xi - *yi;
    }
}

/// Apply diagonal preconditioner: z = M⁻¹ * r
fn apply_preconditioner(diag: &[f64], r: &[f64], z: &mut [f64]) {
    for (i, (&ri, &di)) in r.iter().zip(diag.iter()).enumerate() {
        // Robust division - avoid very small divisors
        z[i] = if di.abs() > 1e-12 { ri / di } else { ri };
    }
}

/// Conjugate Gradient solver for symmetric positive definite matrices
///
/// Solves Ax = b where A is SPD.
///
/// # Arguments
///
/// * `a` - Symmetric positive definite CSR matrix
/// * `b` - Right-hand side vector
/// * `tol` - Convergence tolerance
/// * `max_iter` - Maximum iterations
///
/// # Example
///
/// ```
/// use barracuda::linalg::sparse::{CsrMatrix, cg_solve};
///
/// // Tridiagonal SPD matrix
/// let a = CsrMatrix::from_triplets(3, 3, &[
///     (0, 0, 4.0), (0, 1, -1.0),
///     (1, 0, -1.0), (1, 1, 4.0), (1, 2, -1.0),
///     (2, 1, -1.0), (2, 2, 4.0),
/// ]);
///
/// let b = vec![1.0, 2.0, 3.0];
/// let result = cg_solve(&a, &b, 1e-10, 100).unwrap();
///
/// assert!(result.converged);
/// ```
///
/// # Errors
///
/// Returns [`Err`] if the matrix is not square, if `b.len() != n`, or if
/// [`matvec`](super::csr::CsrMatrix::matvec) fails during iteration.
pub fn cg_solve(a: &CsrMatrix, b: &[f64], tol: f64, max_iter: usize) -> Result<SolverResult> {
    let config = SolverConfig::default()
        .with_tolerance(tol)
        .with_max_iterations(max_iter);
    cg_solve_with_config(a, b, &config)
}

/// Conjugate Gradient with configuration
///
/// # Errors
///
/// Returns [`Err`] if the matrix is not square, if `b.len() != n`, or if [`matvec`](super::csr::CsrMatrix::matvec)
/// fails during iteration.
pub fn cg_solve_with_config(
    a: &CsrMatrix,
    b: &[f64],
    config: &SolverConfig,
) -> Result<SolverResult> {
    let n = a.n_rows;
    if a.n_cols != n {
        return Err(BarracudaError::InvalidInput {
            message: "CG requires square matrix".to_string(),
        });
    }
    if b.len() != n {
        return Err(BarracudaError::InvalidInput {
            message: format!("Vector length {} doesn't match matrix size {}", b.len(), n),
        });
    }

    // Initial guess: x = 0
    let mut x = vec![0.0; n];

    // r = b - A*x = b (since x = 0)
    let mut r = b.to_vec();
    let b_norm = norm(b);

    if b_norm < 1e-14 {
        return Ok(SolverResult {
            x,
            iterations: 0,
            residual: 0.0,
            converged: true,
        });
    }

    // Diagonal preconditioner
    let diag = if config.use_preconditioner {
        a.diagonal()
    } else {
        vec![1.0; n]
    };

    // z = M⁻¹ * r
    let mut z = vec![0.0; n];
    apply_preconditioner(&diag, &r, &mut z);

    // p = z
    let mut p = z.clone();

    // rz = r·z
    let mut rz = dot(&r, &z);

    for iter in 0..config.max_iterations {
        // Ap = A * p
        let ap = a.matvec(&p)?;

        // α = rz / (p·Ap)
        let pap = dot(&p, &ap);
        if pap.abs() < 1e-30 {
            // Near-breakdown - return current solution
            let residual = norm(&r) / b_norm;
            return Ok(SolverResult {
                x,
                iterations: iter + 1,
                residual,
                converged: residual < config.tolerance,
            });
        }
        let alpha = rz / pap;

        // x = x + α*p
        axpy(alpha, &p, &mut x);

        // r = r - α*Ap
        axpy(-alpha, &ap, &mut r);

        // Check convergence
        let residual = norm(&r) / b_norm;
        if residual < config.tolerance {
            return Ok(SolverResult {
                x,
                iterations: iter + 1,
                residual,
                converged: true,
            });
        }

        // z = M⁻¹ * r
        apply_preconditioner(&diag, &r, &mut z);

        // β = (r·z)_new / (r·z)_old
        let rz_new = dot(&r, &z);
        let beta = rz_new / rz;
        rz = rz_new;

        // p = z + β*p
        for (pi, &zi) in p.iter_mut().zip(z.iter()) {
            *pi = zi + beta * *pi;
        }
    }

    // Did not converge
    Ok(SolverResult {
        x,
        iterations: config.max_iterations,
        residual: norm(&r) / b_norm,
        converged: false,
    })
}

/// `BiCGSTAB` solver for general non-symmetric matrices
///
/// Solves Ax = b where A is any square matrix.
///
/// # Arguments
///
/// * `a` - Square CSR matrix
/// * `b` - Right-hand side vector
/// * `tol` - Convergence tolerance
/// * `max_iter` - Maximum iterations
///
/// # Example
///
/// ```
/// use barracuda::linalg::sparse::{CsrMatrix, bicgstab_solve};
///
/// // Non-symmetric matrix
/// let a = CsrMatrix::from_triplets(3, 3, &[
///     (0, 0, 4.0), (0, 1, 1.0),
///     (1, 0, -1.0), (1, 1, 3.0), (1, 2, 1.0),
///     (2, 1, -1.0), (2, 2, 2.0),
/// ]);
///
/// let b = vec![1.0, 2.0, 3.0];
/// let result = bicgstab_solve(&a, &b, 1e-10, 100).unwrap();
///
/// assert!(result.converged);
/// ```
///
/// # Errors
///
/// Returns [`Err`] if the matrix is not square, if `b.len() != n`, if [`matvec`](super::csr::CsrMatrix::matvec)
/// fails, or if `BiCGSTAB` breaks down (rho = 0, `r_hat·v` = 0, or omega = 0).
pub fn bicgstab_solve(a: &CsrMatrix, b: &[f64], tol: f64, max_iter: usize) -> Result<SolverResult> {
    let config = SolverConfig::default()
        .with_tolerance(tol)
        .with_max_iterations(max_iter);
    bicgstab_solve_with_config(a, b, &config)
}

/// `BiCGSTAB` with configuration
///
/// # Errors
///
/// Returns [`Err`] if the matrix is not square, if `b.len() != n`, if [`matvec`](super::csr::CsrMatrix::matvec)
/// fails, or if `BiCGSTAB` breaks down (rho = 0, `r_hat·v` = 0, or omega = 0).
pub fn bicgstab_solve_with_config(
    a: &CsrMatrix,
    b: &[f64],
    config: &SolverConfig,
) -> Result<SolverResult> {
    let n = a.n_rows;
    if a.n_cols != n {
        return Err(BarracudaError::InvalidInput {
            message: "BiCGSTAB requires square matrix".to_string(),
        });
    }
    if b.len() != n {
        return Err(BarracudaError::InvalidInput {
            message: format!("Vector length {} doesn't match matrix size {}", b.len(), n),
        });
    }

    // Initial guess: x = 0
    let mut x = vec![0.0; n];

    // r = b - A*x = b
    let mut r = b.to_vec();
    let b_norm = norm(b);

    if b_norm < 1e-14 {
        return Ok(SolverResult {
            x,
            iterations: 0,
            residual: 0.0,
            converged: true,
        });
    }

    // r_hat = r (shadow residual, fixed)
    let r_hat = r.clone();

    // Initialize
    let mut rho = 1.0;
    let mut alpha = 1.0;
    let mut omega = 1.0;

    let mut v = vec![0.0; n];
    let mut p = vec![0.0; n];
    let mut s = vec![0.0; n];
    let mut t = vec![0.0; n];

    // Diagonal preconditioner
    let diag = if config.use_preconditioner {
        a.diagonal()
    } else {
        vec![1.0; n]
    };

    for iter in 0..config.max_iterations {
        // ρ = r_hat · r
        let rho_new = dot(&r_hat, &r);

        if rho_new.abs() < 1e-14 {
            return Err(BarracudaError::Internal(
                "BiCGSTAB breakdown: rho = 0".to_string(),
            ));
        }

        // β = (ρ_new / ρ) * (α / ω)
        let beta = (rho_new / rho) * (alpha / omega);
        rho = rho_new;

        // p = r + β * (p - ω * v)
        for i in 0..n {
            p[i] = r[i] + beta * (p[i] - omega * v[i]);
        }

        // Apply preconditioner to p
        let mut p_hat = vec![0.0; n];
        apply_preconditioner(&diag, &p, &mut p_hat);

        // v = A * p_hat
        v = a.matvec(&p_hat)?;

        // α = ρ / (r_hat · v)
        let rv = dot(&r_hat, &v);
        if rv.abs() < 1e-14 {
            return Err(BarracudaError::Internal(
                "BiCGSTAB breakdown: r_hat·v = 0".to_string(),
            ));
        }
        alpha = rho / rv;

        // s = r - α * v
        for i in 0..n {
            s[i] = r[i] - alpha * v[i];
        }

        // Check convergence after first half-step
        let s_norm = norm(&s);
        if s_norm / b_norm < config.tolerance {
            // x = x + α * p_hat
            axpy(alpha, &p_hat, &mut x);
            return Ok(SolverResult {
                x,
                iterations: iter + 1,
                residual: s_norm / b_norm,
                converged: true,
            });
        }

        // Apply preconditioner to s
        let mut s_hat = vec![0.0; n];
        apply_preconditioner(&diag, &s, &mut s_hat);

        // t = A * s_hat
        t = a.matvec(&s_hat)?;

        // ω = (t · s) / (t · t)
        let tt = dot(&t, &t);
        if tt.abs() < 1e-14 {
            omega = 0.0;
        } else {
            omega = dot(&t, &s) / tt;
        }

        // x = x + α * p_hat + ω * s_hat
        axpy(alpha, &p_hat, &mut x);
        axpy(omega, &s_hat, &mut x);

        // r = s - ω * t
        for i in 0..n {
            r[i] = s[i] - omega * t[i];
        }

        // Check convergence
        let residual = norm(&r) / b_norm;
        if residual < config.tolerance {
            return Ok(SolverResult {
                x,
                iterations: iter + 1,
                residual,
                converged: true,
            });
        }

        if omega.abs() < 1e-14 {
            return Err(BarracudaError::Internal(
                "BiCGSTAB breakdown: omega = 0".to_string(),
            ));
        }
    }

    // Did not converge
    Ok(SolverResult {
        x,
        iterations: config.max_iterations,
        residual: norm(&r) / b_norm,
        converged: false,
    })
}

/// Jacobi iteration solver
///
/// Simple iterative method for diagonally dominant systems.
///
/// # Arguments
///
/// * `a` - Square CSR matrix (should be diagonally dominant)
/// * `b` - Right-hand side vector
/// * `tol` - Convergence tolerance
/// * `max_iter` - Maximum iterations
///
/// # Errors
///
/// Returns [`Err`] if the matrix is not square or `b.len() != n` (dimension mismatch), if any
/// diagonal element is zero, or if [`matvec`](super::csr::CsrMatrix::matvec) fails during iteration.
pub fn jacobi_solve(a: &CsrMatrix, b: &[f64], tol: f64, max_iter: usize) -> Result<SolverResult> {
    let n = a.n_rows;
    if a.n_cols != n || b.len() != n {
        return Err(BarracudaError::InvalidInput {
            message: "Dimension mismatch".to_string(),
        });
    }

    let b_norm = norm(b);
    if b_norm < 1e-14 {
        return Ok(SolverResult {
            x: vec![0.0; n],
            iterations: 0,
            residual: 0.0,
            converged: true,
        });
    }

    // Get diagonal
    let diag = a.diagonal();

    // Check diagonal is non-zero
    for (i, d) in diag.iter().enumerate() {
        if d.abs() < 1e-14 {
            return Err(BarracudaError::InvalidInput {
                message: format!("Zero diagonal at row {i}"),
            });
        }
    }

    let mut x = vec![0.0; n];
    let mut x_new = vec![0.0; n];

    for iter in 0..max_iter {
        // x_new[i] = (b[i] - Σ_{j≠i} A[i,j] * x[j]) / A[i,i]
        for i in 0..n {
            let row_start = a.row_ptr[i];
            let row_end = a.row_ptr[i + 1];

            let mut sum = b[i];
            for k in row_start..row_end {
                let j = a.col_indices[k];
                if j != i {
                    sum = a.values[k].mul_add(-x[j], sum);
                }
            }
            x_new[i] = sum / diag[i];
        }

        // Check convergence
        let ax = a.matvec(&x_new)?;
        let mut r = ax;
        sub_into(b, &mut r);
        let residual = norm(&r) / b_norm;

        if residual < tol {
            return Ok(SolverResult {
                x: x_new,
                iterations: iter + 1,
                residual,
                converged: true,
            });
        }

        std::mem::swap(&mut x, &mut x_new);
    }

    // Did not converge
    let ax = a.matvec(&x)?;
    let mut r = ax;
    sub_into(b, &mut r);

    Ok(SolverResult {
        x,
        iterations: max_iter,
        residual: norm(&r) / b_norm,
        converged: false,
    })
}

#[cfg(test)]
#[path = "solvers_tests.rs"]
mod tests;