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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Generalized Eigenvalue Decomposition (`gen_eigh`) - GPU-Accelerated Implementation (f64)
//!
//! Solves the generalized eigenvalue problem: Ax = λBx
//! where A is symmetric and B is symmetric positive definite.
//!
//! **Deep Debt Principles**:
//! - ✅ GPU-accelerated eigenvalue decomposition (dominant cost)
//! - ✅ Full f64 precision via SPIR-V/Vulkan
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Hybrid CPU/GPU strategy for optimal performance
//!
//! ## Algorithm
//!
//! Uses Cholesky-based reduction to standard form:
//! ```text
//! Input:  A [N × N] symmetric matrix
//!         B [N × N] symmetric positive definite matrix
//! Output: eigenvalues [N]
//!         eigenvectors [N × N]
//!
//! Steps:
//! 1. Cholesky: B = LLᵀ (CPU - O(n³/3))
//! 2. Transform: C = L⁻¹ A L⁻ᵀ (CPU - O(n³))
//! 3. Eigensolve: Cy = λy (GPU - dominant cost)
//! 4. Back-transform: x = L⁻ᵀ y (CPU - O(n²))
//! ```
//!
//! ## Performance Rationale
//!
//! The eigenvalue decomposition is the bottleneck, requiring O(n³) with a high
//! constant factor due to many Jacobi sweeps. CPU Cholesky and triangular solves
//! are fast enough that GPU overhead isn't worth it for typical sizes.
//!
//! ## Applications
//!
//! - Hartree-Fock-Bogoliubov (HFB) nuclear structure
//! - Vibration analysis in structural mechanics
//! - Quantum chemistry (Roothaan equations)
//!
//! ## References
//!
//! - Golub & Van Loan, "Matrix Computations", Section 8.7
//! - LAPACK DSYGV routine

use crate::device::WgpuDevice;
use crate::error::{BarracudaError, Result};
use crate::linalg::cholesky::cholesky_f64;
use crate::ops::linalg::BatchedEighGpu;
use std::sync::Arc;

/// Generalized eigenvalue decomposition result (GPU)
#[derive(Debug, Clone)]
pub struct GenEighDecompositionGpu {
    /// Eigenvalues (λ) in ascending order
    pub eigenvalues: Vec<f64>,
    /// Eigenvectors (x) as row-major n×n matrix
    /// The i-th eigenvector is column i (accessed as row*n + col)
    pub eigenvectors: Vec<f64>,
    /// Matrix dimension
    pub n: usize,
}

impl GenEighDecompositionGpu {
    /// Get the i-th eigenvector
    #[must_use]
    pub fn eigenvector(&self, i: usize) -> Option<Vec<f64>> {
        if i >= self.n {
            return None;
        }
        let mut v = vec![0.0; self.n];
        for (row, v_row) in v.iter_mut().enumerate() {
            *v_row = self.eigenvectors[row * self.n + i];
        }
        Some(v)
    }

    /// Sort eigenvalues and eigenvectors in descending order
    pub fn sort_descending(&mut self) {
        let n = self.n;
        let mut indices: Vec<usize> = (0..n).collect();
        indices.sort_by(|&a, &b| {
            self.eigenvalues[b]
                .partial_cmp(&self.eigenvalues[a])
                .unwrap_or(std::cmp::Ordering::Equal)
        });

        let old_eigenvalues = self.eigenvalues.clone();
        let old_eigenvectors = self.eigenvectors.clone();

        for (new_idx, &old_idx) in indices.iter().enumerate() {
            self.eigenvalues[new_idx] = old_eigenvalues[old_idx];
            for row in 0..n {
                self.eigenvectors[row * n + new_idx] = old_eigenvectors[row * n + old_idx];
            }
        }
    }

    /// Verify that Ax = λBx for all eigenpairs
    #[must_use]
    pub fn verify(&self, a: &[f64], b: &[f64]) -> f64 {
        let n = self.n;
        let mut max_residual: f64 = 0.0;

        for i in 0..n {
            let lambda = self.eigenvalues[i];
            let x: Vec<f64> = (0..n).map(|row| self.eigenvectors[row * n + i]).collect();

            // Compute Ax
            let mut ax = vec![0.0; n];
            for row in 0..n {
                for col in 0..n {
                    ax[row] = a[row * n + col].mul_add(x[col], ax[row]);
                }
            }

            // Compute λBx
            let mut lbx = vec![0.0; n];
            for row in 0..n {
                for col in 0..n {
                    lbx[row] = (lambda * b[row * n + col]).mul_add(x[col], lbx[row]);
                }
            }

            let residual_norm: f64 = ax
                .iter()
                .zip(lbx.iter())
                .map(|(a, b)| (a - b).powi(2))
                .sum::<f64>()
                .sqrt();

            let x_norm: f64 = x.iter().map(|v| v.powi(2)).sum::<f64>().sqrt();
            let relative_residual = residual_norm / x_norm.max(1e-15);
            max_residual = max_residual.max(relative_residual);
        }

        max_residual
    }
}

/// GPU-accelerated generalized eigenvalue decomposition
///
/// Solves Ax = λBx where A is symmetric and B is symmetric positive definite.
pub struct GenEighGpu;

impl GenEighGpu {
    /// Execute generalized eigenvalue decomposition with GPU-accelerated eigensolve
    /// # Arguments
    /// * `device` - `WgpuDevice` to execute on
    /// * `a` - Symmetric matrix A (row-major, n×n) as f64
    /// * `b` - Symmetric positive definite matrix B (row-major, n×n) as f64
    /// * `n` - Matrix dimension
    /// * `max_sweeps` - Maximum Jacobi sweeps for eigensolve (default: 30)
    /// # Returns
    /// `GenEighDecompositionGpu` with eigenvalues and eigenvectors
    /// # Errors
    /// - If B is not positive definite (Cholesky fails)
    /// - If matrix dimensions are invalid
    /// - If GPU execution fails
    /// # Example
    /// ```ignore
    /// use barracuda::ops::linalg::GenEighGpu;
    /// // Generalized eigenvalue problem: Ax = λBx
    /// let a = vec![4.0, 2.0, 2.0, 3.0]; // [[4, 2], [2, 3]]
    /// let b = vec![2.0, 1.0, 1.0, 2.0]; // [[2, 1], [1, 2]]
    /// let result = GenEighGpu::execute_f64(device, &a, &b, 2, 30)?;
    /// ```
    /// # Errors
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    /// Also returns [`Err`] if matrix A or B has invalid size.
    pub fn execute_f64(
        device: Arc<WgpuDevice>,
        a: &[f64],
        b: &[f64],
        n: usize,
        max_sweeps: u32,
    ) -> Result<GenEighDecompositionGpu> {
        let expected_len = n * n;
        if a.len() != expected_len {
            return Err(BarracudaError::InvalidInput {
                message: format!(
                    "Matrix A has wrong size: expected {}, got {}",
                    expected_len,
                    a.len()
                ),
            });
        }
        if b.len() != expected_len {
            return Err(BarracudaError::InvalidInput {
                message: format!(
                    "Matrix B has wrong size: expected {}, got {}",
                    expected_len,
                    b.len()
                ),
            });
        }

        // Step 1: Cholesky decomposition of B = LLᵀ (GPU)
        let chol = cholesky_f64(device.clone(), b, n)?;
        let l = &chol.l;

        // Step 2: Transform A to standard form: C = L⁻¹ A L⁻ᵀ (CPU)
        // First compute L⁻¹ A by solving L Y = A column by column
        let l_inv_a = solve_lower_triangular_matrix(l, a, n);

        // Then compute C = (L⁻¹ A) L⁻ᵀ
        let c = compute_symmetric_transform(l, &l_inv_a, n);

        // Step 3: GPU eigensolve on transformed matrix C
        // Use batch_size=1 for single matrix
        let (eigenvalues, eigenvectors_c) = BatchedEighGpu::execute_f64(
            device, &c, n, 1, // batch_size = 1
            max_sweeps,
        )?;

        // Step 4: Back-transform eigenvectors: x = L⁻ᵀ y (CPU)
        // Solve Lᵀ x = y for each eigenvector
        let eigenvectors = back_transform_eigenvectors(l, &eigenvectors_c, n);

        Ok(GenEighDecompositionGpu {
            eigenvalues,
            eigenvectors,
            n,
        })
    }

    /// Execute for a batch of generalized eigenvalue problems
    /// # Arguments
    /// * `device` - `WgpuDevice` to execute on
    /// * `a_batch` - Packed A matrices [`batch_size` × n × n]
    /// * `b_batch` - Packed B matrices [`batch_size` × n × n]
    /// * `n` - Matrix dimension (same for all)
    /// * `batch_size` - Number of matrix pairs
    /// * `max_sweeps` - Maximum Jacobi sweeps
    /// # Returns
    /// Vector of `GenEighDecompositionGpu`, one per matrix pair
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn execute_batch_f64(
        device: Arc<WgpuDevice>,
        a_batch: &[f64],
        b_batch: &[f64],
        n: usize,
        batch_size: usize,
        max_sweeps: u32,
    ) -> Result<Vec<GenEighDecompositionGpu>> {
        let expected_len = batch_size * n * n;
        if a_batch.len() != expected_len {
            return Err(BarracudaError::InvalidInput {
                message: format!(
                    "A batch has wrong size: expected {}, got {}",
                    expected_len,
                    a_batch.len()
                ),
            });
        }
        if b_batch.len() != expected_len {
            return Err(BarracudaError::InvalidInput {
                message: format!(
                    "B batch has wrong size: expected {}, got {}",
                    expected_len,
                    b_batch.len()
                ),
            });
        }

        // Transform all matrices to standard form
        let mut c_batch = vec![0.0; expected_len];
        let mut l_batch = vec![0.0; expected_len]; // Store L for back-transform

        for i in 0..batch_size {
            let offset = i * n * n;
            let a = &a_batch[offset..offset + n * n];
            let b = &b_batch[offset..offset + n * n];

            // Cholesky
            let chol = cholesky_f64(device.clone(), b, n)?;
            let l = &chol.l;

            // Store L for back-transform
            l_batch[offset..offset + n * n].copy_from_slice(l);

            // Transform
            let l_inv_a = solve_lower_triangular_matrix(l, a, n);
            let c = compute_symmetric_transform(l, &l_inv_a, n);
            c_batch[offset..offset + n * n].copy_from_slice(&c);
        }

        // GPU eigensolve on all transformed matrices
        let (eigenvalues_batch, eigenvectors_c_batch) =
            BatchedEighGpu::execute_f64(device, &c_batch, n, batch_size, max_sweeps)?;

        // Back-transform all eigenvectors
        let mut results = Vec::with_capacity(batch_size);
        for i in 0..batch_size {
            let offset = i * n * n;
            let eig_offset = i * n;

            let l = &l_batch[offset..offset + n * n];
            let eigenvectors_c = &eigenvectors_c_batch[offset..offset + n * n];
            let eigenvalues: Vec<f64> = eigenvalues_batch[eig_offset..eig_offset + n].to_vec();

            let eigenvectors = back_transform_eigenvectors(l, eigenvectors_c, n);

            results.push(GenEighDecompositionGpu {
                eigenvalues,
                eigenvectors,
                n,
            });
        }

        Ok(results)
    }
}

/// Solve LY = A where L is lower triangular, column by column
fn solve_lower_triangular_matrix(l: &[f64], a: &[f64], n: usize) -> Vec<f64> {
    let mut result = vec![0.0; n * n];

    for col in 0..n {
        // Extract column of A
        let mut y = vec![0.0; n];
        for row in 0..n {
            y[row] = a[row * n + col];
        }

        // Forward substitution: Ly = a_col
        for i in 0..n {
            let mut sum = y[i];
            for j in 0..i {
                sum = l[i * n + j].mul_add(-y[j], sum);
            }
            let diag = l[i * n + i];
            y[i] = if diag.abs() > 1e-15 { sum / diag } else { 0.0 };
        }

        // Store result column
        for row in 0..n {
            result[row * n + col] = y[row];
        }
    }

    result
}

/// Compute C = (L⁻¹ A) L⁻ᵀ
fn compute_symmetric_transform(l: &[f64], l_inv_a: &[f64], n: usize) -> Vec<f64> {
    // First compute L⁻¹
    let mut l_inv = vec![0.0; n * n];
    for i in 0..n {
        let mut x = vec![0.0; n];
        x[i] = 1.0;

        for row in 0..n {
            let mut sum = if row == i { 1.0 } else { 0.0 };
            for k in 0..row {
                sum = l[row * n + k].mul_add(-x[k], sum);
            }
            let diag = l[row * n + row];
            x[row] = if diag.abs() > 1e-15 { sum / diag } else { 0.0 };
        }

        for row in 0..n {
            l_inv[row * n + i] = x[row];
        }
    }

    // C = Y * L⁻ᵀ where Y = L⁻¹ A
    let mut c = vec![0.0; n * n];
    for i in 0..n {
        for j in 0..n {
            let mut sum = 0.0;
            for k in 0..n {
                // Y[i,k] * L⁻ᵀ[k,j] = Y[i,k] * L⁻¹[j,k]
                sum = l_inv_a[i * n + k].mul_add(l_inv[j * n + k], sum);
            }
            c[i * n + j] = sum;
        }
    }

    c
}

/// Back-transform eigenvectors: x = L⁻ᵀ y = solve(Lᵀ, y)
fn back_transform_eigenvectors(l: &[f64], eigenvectors_c: &[f64], n: usize) -> Vec<f64> {
    let mut result = vec![0.0; n * n];

    for col in 0..n {
        // Extract eigenvector col
        let mut y = vec![0.0; n];
        for row in 0..n {
            y[row] = eigenvectors_c[row * n + col];
        }

        // Backward substitution: Lᵀ x = y
        // Process from row n-1 down to 0
        let mut x = vec![0.0; n];
        for i in (0..n).rev() {
            let mut sum = y[i];
            for j in (i + 1)..n {
                // Lᵀ[i,j] = L[j,i]
                sum = l[j * n + i].mul_add(-x[j], sum);
            }
            let diag = l[i * n + i];
            x[i] = if diag.abs() > 1e-15 { sum / diag } else { 0.0 };
        }

        // Store result column
        for row in 0..n {
            result[row * n + col] = x[row];
        }
    }

    result
}

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

    #[test]
    fn test_solve_lower_triangular() {
        // L = [[2, 0], [1, 3]]
        let l = vec![2.0, 0.0, 1.0, 3.0];
        // A = [[1, 0], [0, 1]] (identity)
        let a = vec![1.0, 0.0, 0.0, 1.0];
        let n = 2;

        let result = solve_lower_triangular_matrix(&l, &a, n);

        // L⁻¹ = [[0.5, 0], [-1/6, 1/3]]
        // L⁻¹ * I = L⁻¹
        assert!((result[0] - 0.5).abs() < 1e-10);
        assert!((result[1] - 0.0).abs() < 1e-10);
        assert!((result[2] - (-1.0 / 6.0)).abs() < 1e-10);
        assert!((result[3] - (1.0 / 3.0)).abs() < 1e-10);
    }

    #[test]
    fn test_back_transform() {
        // L = [[2, 0], [0, 2]]
        let l = vec![2.0, 0.0, 0.0, 2.0];
        // eigenvectors = [[1, 0], [0, 1]]
        let eigenvectors = vec![1.0, 0.0, 0.0, 1.0];
        let n = 2;

        let result = back_transform_eigenvectors(&l, &eigenvectors, n);

        // L⁻ᵀ = [[0.5, 0], [0, 0.5]]
        // Result should be scaled by 0.5
        assert!((result[0] - 0.5).abs() < 1e-10);
        assert!((result[1] - 0.0).abs() < 1e-10);
        assert!((result[2] - 0.0).abs() < 1e-10);
        assert!((result[3] - 0.5).abs() < 1e-10);
    }
}