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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! `AdaGrad` Optimizer - GPU-accelerated Adaptive Gradient Algorithm
//!
//! **Deep Debt Principles**:
//! - ✅ Pure WGSL implementation (existing shader evolved)
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Complete implementation (production-ready)
//! - ✅ Modern idiomatic Rust (no traits, direct impl)
//! - ✅ Capability-based dispatch (vendor-optimized workgroups)
//!
//! ## Algorithm
//!
//! ```text
//! G_t = G_{t-1} + g_t²
//! w_t = w_{t-1} - (lr / sqrt(G_t + ε)) * g_t
//! ```
//!
//! **Key Properties**:
//! - Adapts learning rate per parameter based on gradient history
//! - Accumulates squared gradients
//! - Good for sparse gradients
//! - Learning rate monotonically decreases (limitation)
//!
//! **Parameters**:
//! - `learning_rate`: Initial step size, typically 0.01
//! - `epsilon` (ε): Numerical stability constant, typically 1e-8
//!
//! **Used By**: Sparse gradient problems, NLP tasks
//!
//! **Note**: `RMSprop` and Adam address `AdaGrad`'s monotonic decrease issue
//!
//! ## Usage
//!
//! ```rust,ignore
//! use barracuda::tensor::Tensor;
//!
//! let weights = Tensor::randn(vec![1000]).await?;
//! let gradients = Tensor::randn(vec![1000]).await?;
//!
//! // First step
//! let (w1, acc1) = weights.adagrad_step(&gradients, 0.01, None)?;
//!
//! // Subsequent steps
//! let (w2, acc2) = w1.adagrad_step(&gradients, 0.01, Some(&acc1))?;
//! ```

use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;

#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct AdaGradParams {
    learning_rate: f32,
    epsilon: f32,
    weight_decay: f32,
    _padding: u32,
}

/// `AdaGrad` optimizer — per-parameter adaptive learning rates.
pub struct AdaGrad {
    weights: Tensor,
    gradients: Tensor,
    accumulated: Option<Tensor>,
    learning_rate: f32,
}

impl AdaGrad {
    /// Create `AdaGrad` optimizer. Good for sparse gradients.
    /// # Errors
    /// Returns [`Err`] if weights and gradients shapes differ, `learning_rate` is not positive,
    /// or accumulated (if provided) shape differs from weights.
    pub fn new(
        weights: Tensor,
        gradients: Tensor,
        learning_rate: f32,
        accumulated: Option<Tensor>,
    ) -> Result<Self> {
        // Validate shapes match
        if weights.shape() != gradients.shape() {
            return Err(BarracudaError::shape_mismatch(
                weights.shape().to_vec(),
                gradients.shape().to_vec(),
            ));
        }

        // Validate learning rate is positive
        if learning_rate <= 0.0 {
            return Err(BarracudaError::invalid_op(
                "adagrad",
                "learning_rate must be positive",
            ));
        }

        // Validate accumulated shape if provided
        if let Some(ref acc) = accumulated
            && acc.shape() != weights.shape()
        {
            return Err(BarracudaError::shape_mismatch(
                acc.shape().to_vec(),
                weights.shape().to_vec(),
            ));
        }

        Ok(Self {
            weights,
            gradients,
            accumulated,
            learning_rate,
        })
    }

    fn wgsl_shader() -> &'static str {
        static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
            include_str!("../shaders/optimizer/adagrad_f64.wgsl").to_string()
        });
        std::sync::LazyLock::force(&SHADER).as_str()
    }

    /// Execute `AdaGrad` step. Returns (`updated_weights`, accumulated).
    /// # Errors
    /// Returns [`Err`] if buffer allocation fails, GPU dispatch fails, or the device is lost.
    pub fn execute(self) -> Result<(Tensor, Tensor)> {
        let device = self.weights.device();
        let size = self.weights.shape().iter().product::<usize>();

        let params = AdaGradParams {
            learning_rate: self.learning_rate,
            epsilon: 1e-8,
            weight_decay: 0.0,
            _padding: 0,
        };

        // Create accumulated buffer if not provided
        let zeros = vec![0.0f32; size];
        let accumulated_in = if let Some(ref acc_tensor) = self.accumulated {
            acc_tensor.buffer()
        } else {
            &device
                .device
                .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                    label: Some("adagrad_acc_zeros"),
                    contents: bytemuck::cast_slice(&zeros),
                    usage: wgpu::BufferUsages::STORAGE,
                })
        };

        let weights_out_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
            label: Some("adagrad_weights_out"),
            size: (size * std::mem::size_of::<f32>()) as u64,
            usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
            mapped_at_creation: false,
        });

        let accumulated_out_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
            label: Some("adagrad_accumulated_out"),
            size: (size * std::mem::size_of::<f32>()) as u64,
            usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
            mapped_at_creation: false,
        });

        let params_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("adagrad_params"),
                contents: bytemuck::cast_slice(&[params]),
                usage: wgpu::BufferUsages::UNIFORM,
            });

        let shader = device.compile_shader(Self::wgsl_shader(), Some("adagrad_shader"));

        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("adagrad_bind_group_layout"),
                    entries: &[
                        wgpu::BindGroupLayoutEntry {
                            binding: 0,
                            visibility: wgpu::ShaderStages::COMPUTE,
                            ty: wgpu::BindingType::Buffer {
                                ty: wgpu::BufferBindingType::Storage { read_only: true },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                        wgpu::BindGroupLayoutEntry {
                            binding: 1,
                            visibility: wgpu::ShaderStages::COMPUTE,
                            ty: wgpu::BindingType::Buffer {
                                ty: wgpu::BufferBindingType::Storage { read_only: true },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                        wgpu::BindGroupLayoutEntry {
                            binding: 2,
                            visibility: wgpu::ShaderStages::COMPUTE,
                            ty: wgpu::BindingType::Buffer {
                                ty: wgpu::BufferBindingType::Storage { read_only: true },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                        wgpu::BindGroupLayoutEntry {
                            binding: 3,
                            visibility: wgpu::ShaderStages::COMPUTE,
                            ty: wgpu::BindingType::Buffer {
                                ty: wgpu::BufferBindingType::Storage { read_only: false },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                        wgpu::BindGroupLayoutEntry {
                            binding: 4,
                            visibility: wgpu::ShaderStages::COMPUTE,
                            ty: wgpu::BindingType::Buffer {
                                ty: wgpu::BufferBindingType::Storage { read_only: false },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                        wgpu::BindGroupLayoutEntry {
                            binding: 5,
                            visibility: wgpu::ShaderStages::COMPUTE,
                            ty: wgpu::BindingType::Buffer {
                                ty: wgpu::BufferBindingType::Uniform,
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                    ],
                });

        let pipeline_layout =
            device
                .device
                .create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
                    label: Some("adagrad_pipeline_layout"),
                    bind_group_layouts: &[&bind_group_layout],
                    immediate_size: 0,
                });

        let pipeline = device
            .device
            .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                label: Some("adagrad_pipeline"),
                layout: Some(&pipeline_layout),
                module: &shader,
                entry_point: Some("main"),
                cache: None,
                compilation_options: Default::default(),
            });

        let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("adagrad_bind_group"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: self.weights.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: self.gradients.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: accumulated_in.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: weights_out_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: accumulated_out_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 5,
                    resource: params_buffer.as_entire_binding(),
                },
            ],
        });

        let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("adagrad_encoder"),
        });

        {
            let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("adagrad_pass"),
                timestamp_writes: None,
            });
            compute_pass.set_pipeline(&pipeline);
            compute_pass.set_bind_group(0, Some(&bind_group), &[]);

            // Deep Debt Evolution: Capability-based dispatch
            let caps = DeviceCapabilities::from_device(device);
            let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
            let workgroups = (size as u32).div_ceil(optimal_wg_size);
            compute_pass.dispatch_workgroups(workgroups, 1, 1);
        }

        device.submit_commands(Some(encoder.finish()));

        let updated_weights = Tensor::from_buffer(
            weights_out_buffer,
            self.weights.shape().to_vec(),
            device.clone(),
        );

        let updated_accumulated = Tensor::from_buffer(
            accumulated_out_buffer,
            self.weights.shape().to_vec(),
            device.clone(),
        );

        Ok((updated_weights, updated_accumulated))
    }
}

// ═══════════════════════════════════════════════════════════════
// TENSOR API INTEGRATION (MODERN IDIOMATIC RUST)
// ═══════════════════════════════════════════════════════════════

impl Tensor {
    /// `AdaGrad` optimizer step - adaptive learning rate for sparse gradients
    /// **Deep Debt**: Historical optimizer, foundation for Adam/RMSprop
    /// # Arguments
    /// - `gradients`: Gradient tensor [same shape as weights]
    /// - `learning_rate`: Initial step size, typically 0.01
    /// - `accumulated`: Accumulated squared gradients (None for first step)
    /// # Returns
    /// - Tuple: (`updated_weights`, `updated_accumulated`)
    /// # Example
    /// ```rust,ignore
    /// // First step
    /// let (w1, acc1) = weights.adagrad_step(&grads, 0.01, None)?;
    /// // Subsequent steps
    /// let (w2, acc2) = w1.adagrad_step(&grads, 0.01, Some(&acc1))?;
    /// ```
    /// # Note
    /// - Good for sparse gradients
    /// - Learning rate monotonically decreases
    /// - `RMSprop` and Adam address this limitation
    /// - `learning_rate` must be positive
    /// # Errors
    /// Returns [`Err`] if weights and gradients shapes differ, `learning_rate` is not positive,
    /// accumulated shape differs, buffer allocation fails, GPU dispatch fails, or the device is lost.
    pub fn adagrad_step(
        self,
        gradients: &Self,
        learning_rate: f32,
        accumulated: Option<&Self>,
    ) -> Result<(Self, Self)> {
        AdaGrad::new(self, gradients.clone(), learning_rate, accumulated.cloned())?.execute()
    }
}

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

    #[tokio::test]
    async fn test_adagrad_basic() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let weights = Tensor::from_vec_on(vec![1.0, 2.0, 3.0, 4.0], vec![4], device.clone())
            .await
            .unwrap();

        let gradients = Tensor::from_vec_on(vec![0.1, 0.2, 0.3, 0.4], vec![4], device.clone())
            .await
            .unwrap();

        let (updated_weights, updated_acc) = weights.adagrad_step(&gradients, 0.01, None).unwrap();

        let result = updated_weights.to_vec().unwrap();
        let acc = updated_acc.to_vec().unwrap();

        assert_eq!(result.len(), 4);
        assert!(result.iter().all(|&x| x.is_finite()));
        assert!(acc.iter().all(|&x| x >= 0.0));
        assert!(result[0] < 1.0, "Expected descent, got {}", result[0]);
    }

    #[tokio::test]
    async fn test_adagrad_accumulation() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let weights = Tensor::from_vec_on(vec![1.0; 4], vec![4], device.clone())
            .await
            .unwrap();

        let gradients = Tensor::from_vec_on(vec![0.1; 4], vec![4], device.clone())
            .await
            .unwrap();

        // Step 1
        let (weights1, acc1) = weights.adagrad_step(&gradients, 0.01, None).unwrap();
        let acc_data1 = acc1.to_vec().unwrap();
        assert!(acc_data1.iter().all(|&x| x > 0.0));

        // Step 2 with accumulated state
        let (weights2, acc2) = weights1
            .adagrad_step(&gradients, 0.01, Some(&acc1))
            .unwrap();
        let result = weights2.to_vec().unwrap();
        let acc_data2 = acc2.to_vec().unwrap();

        assert!(result.iter().all(|&x| x.is_finite()));
        // Accumulated should increase
        assert!(acc_data2.iter().zip(&acc_data1).all(|(&a2, &a1)| a2 > a1));
    }

    #[tokio::test]
    async fn test_adagrad_validation() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let weights = Tensor::from_vec_on(vec![1.0; 10], vec![10], device.clone())
            .await
            .unwrap();
        let gradients = Tensor::from_vec_on(vec![0.1; 5], vec![5], device.clone())
            .await
            .unwrap();
        let grads_correct = Tensor::from_vec_on(vec![0.1; 10], vec![10], device.clone())
            .await
            .unwrap();

        // Shape mismatch
        assert!(
            weights
                .clone()
                .adagrad_step(&gradients, 0.01, None)
                .is_err()
        );

        // Invalid learning rate
        assert!(
            weights
                .clone()
                .adagrad_step(&grads_correct, -0.01, None)
                .is_err()
        );
        assert!(weights.adagrad_step(&grads_correct, 0.0, None).is_err());
    }

    #[tokio::test]
    async fn test_adagrad_large_batch() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let size = 128;
        let weights = Tensor::from_vec_on(vec![1.0; size], vec![size], device.clone())
            .await
            .unwrap();

        let gradients = Tensor::from_vec_on(vec![0.01; size], vec![size], device.clone())
            .await
            .unwrap();

        let (updated_weights, updated_acc) = weights.adagrad_step(&gradients, 0.01, None).unwrap();

        let result = updated_weights.to_vec().unwrap();
        let acc = updated_acc.to_vec().unwrap();

        assert_eq!(result.len(), size);
        assert_eq!(acc.len(), size);
        assert!(result.iter().all(|&x| x.is_finite()));
    }

    #[tokio::test]
    async fn test_adagrad_multi_step() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let weights = Tensor::from_vec_on(vec![10.0, 20.0], vec![2], device.clone())
            .await
            .unwrap();

        let gradients = Tensor::from_vec_on(vec![1.0, 2.0], vec![2], device.clone())
            .await
            .unwrap();

        // Step 1
        let (weights1, acc1) = weights.adagrad_step(&gradients, 0.1, None).unwrap();
        let result1 = weights1.to_vec().unwrap();

        assert!(result1[0] < 10.0, "Expected descent, got {}", result1[0]);
        assert!(result1[1] < 20.0, "Expected descent, got {}", result1[1]);

        // Step 2 with accumulated state
        let (weights2, _acc2) = weights1.adagrad_step(&gradients, 0.1, Some(&acc1)).unwrap();
        let result2 = weights2.to_vec().unwrap();

        // Should continue descending
        assert!(result2[0] < result1[0]);
        assert!(result2[1] < result1[1]);
    }
}