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
// SPDX-License-Identifier: AGPL-3.0-or-later
use super::*;
use crate::device::test_pool::get_test_device;
use crate::device::{Device, WorkloadHint};

async fn esn_with_pool(config: ESNConfig) -> ESN {
    let device = get_test_device().await;
    ESN::with_device(config, device).await.unwrap()
}

#[tokio::test]
async fn test_esn_creation() {
    let esn = esn_with_pool(ESNConfig::default()).await;
    assert!(!esn.is_trained());
    assert_eq!(esn.state().shape(), &[100, 1]);
}

#[tokio::test]
async fn test_esn_invalid_config() {
    let config = ESNConfig {
        input_size: 0,
        ..Default::default()
    };
    let device = get_test_device().await;
    assert!(ESN::with_device(config, device).await.is_err());
}

#[tokio::test]
async fn test_esn_device_preference() {
    let esn = esn_with_pool(ESNConfig::default()).await;
    let _esn_gpu = esn.prefer_device(Device::GPU);
}

#[tokio::test]
async fn test_esn_workload_hint() {
    let esn = esn_with_pool(ESNConfig::default()).await;
    let _esn_large = esn.with_hint(WorkloadHint::LargeMatrices);
}

#[tokio::test]
async fn test_esn_device_query() {
    let esn = esn_with_pool(ESNConfig::default()).await;
    let device = esn.query_device();
    assert!(matches!(device, Device::CPU | Device::GPU | Device::Auto));
}

#[tokio::test]
async fn test_esn_train_simple() {
    let config = ESNConfig {
        input_size: 1,
        reservoir_size: 20,
        output_size: 1,
        spectral_radius: 0.9,
        connectivity: 0.1,
        leak_rate: 0.3,
        regularization: 1e-6,
        seed: 42,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config).await;

    let inputs = vec![vec![0.0], vec![1.0], vec![2.0], vec![3.0], vec![4.0]];
    let targets = vec![vec![1.0], vec![2.0], vec![3.0], vec![4.0], vec![5.0]];

    let error = esn.train(&inputs, &targets).await.unwrap();
    assert!(error >= 0.0);
    assert!(esn.is_trained());
    assert!(esn.w_out.is_some());
}

#[tokio::test]
async fn test_esn_predict_after_train() {
    let config = ESNConfig {
        input_size: 1,
        reservoir_size: 30,
        output_size: 1,
        spectral_radius: 0.95,
        connectivity: 0.15,
        leak_rate: 0.3,
        regularization: 1e-5,
        seed: 42,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config).await;

    let inputs = vec![vec![0.0], vec![1.0], vec![2.0], vec![3.0]];
    let targets = vec![vec![2.0], vec![3.0], vec![4.0], vec![5.0]];

    esn.train(&inputs, &targets).await.unwrap();

    let prediction = esn.predict(&[10.0]).await.unwrap();
    assert_eq!(prediction.len(), 1);
    assert!(prediction[0] > 5.0 && prediction[0] < 20.0);
}

#[tokio::test]
async fn test_esn_train_mismatched_lengths() {
    let mut esn = esn_with_pool(ESNConfig::default()).await;

    let inputs = vec![vec![0.0], vec![1.0]];
    let targets = vec![vec![1.0]];

    let result = esn.train(&inputs, &targets).await;
    assert!(result.is_err());
    assert!(!esn.is_trained());
}

#[tokio::test]
async fn test_esn_train_empty_data() {
    let mut esn = esn_with_pool(ESNConfig::default()).await;

    let inputs: Vec<Vec<f32>> = vec![];
    let targets: Vec<Vec<f32>> = vec![];

    let result = esn.train(&inputs, &targets).await;
    assert!(result.is_err());
    assert!(!esn.is_trained());
}

#[tokio::test]
async fn test_esn_predict_untrained() {
    let mut esn = esn_with_pool(ESNConfig::default()).await;

    let result = esn.predict(&[1.0]).await;
    assert!(result.is_err());
}

#[tokio::test]
async fn test_esn_predict_wrong_input_size() {
    let config = ESNConfig {
        input_size: 2,
        reservoir_size: 20,
        output_size: 1,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config).await;

    let inputs = vec![vec![0.0, 0.0], vec![1.0, 1.0]];
    let targets = vec![vec![1.0], vec![2.0]];
    esn.train(&inputs, &targets).await.unwrap();

    let result = esn.predict(&[1.0]).await;
    assert!(result.is_err());
}

#[tokio::test]
async fn test_esn_multiple_outputs() {
    let config = ESNConfig {
        input_size: 2,
        reservoir_size: 40,
        output_size: 3,
        spectral_radius: 0.9,
        connectivity: 0.1,
        leak_rate: 0.3,
        regularization: 1e-5,
        seed: 42,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config).await;

    let inputs = vec![
        vec![0.0, 0.0],
        vec![1.0, 0.0],
        vec![0.0, 1.0],
        vec![1.0, 1.0],
    ];
    let targets = vec![
        vec![0.0, 0.0, 0.0],
        vec![1.0, 0.0, 1.0],
        vec![0.0, 1.0, 1.0],
        vec![1.0, 1.0, 2.0],
    ];

    let error = esn.train(&inputs, &targets).await.unwrap();
    assert!(error >= 0.0);
    assert!(esn.is_trained());

    let prediction = esn.predict(&[0.5, 0.5]).await.unwrap();
    assert_eq!(prediction.len(), 3);
}

#[tokio::test]
async fn test_esn_predict_return_state() {
    let config = ESNConfig {
        input_size: 1,
        reservoir_size: 20,
        output_size: 1,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config).await;

    let inputs = vec![vec![1.0], vec![2.0], vec![3.0]];
    let targets = vec![vec![2.0], vec![3.0], vec![4.0]];
    esn.train(&inputs, &targets).await.unwrap();

    let (output, state) = esn.predict_return_state(&[5.0]).await.unwrap();
    assert_eq!(output.len(), 1);
    assert_eq!(state.len(), 20);
    assert!(
        state.iter().any(|&v| v != 0.0),
        "State should be non-zero after update"
    );
}

#[tokio::test]
async fn test_esn_set_readout_weights() {
    let config = ESNConfig {
        input_size: 1,
        reservoir_size: 20,
        output_size: 1,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config).await;

    let inputs = vec![vec![1.0], vec![2.0]];
    let targets = vec![vec![2.0], vec![3.0]];
    esn.train(&inputs, &targets).await.unwrap();

    let original = esn.predict(&[5.0]).await.unwrap();

    let new_weights = Tensor::zeros_on(vec![20, 1], esn.device.clone())
        .await
        .unwrap();
    esn.set_readout_weights(new_weights).unwrap();

    let zeroed = esn.predict(&[5.0]).await.unwrap();
    assert!(
        (zeroed[0]).abs() < 1e-5,
        "Zero readout should produce near-zero output"
    );
    assert_ne!(
        original, zeroed,
        "Different readout weights should produce different output"
    );
}

#[tokio::test]
async fn test_esn_state_persistence() {
    let config = ESNConfig {
        input_size: 1,
        reservoir_size: 20,
        output_size: 1,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config).await;

    let inputs = vec![vec![1.0], vec![2.0], vec![3.0]];
    let targets = vec![vec![2.0], vec![3.0], vec![4.0]];
    esn.train(&inputs, &targets).await.unwrap();

    let pred1 = esn.predict(&[5.0]).await.unwrap();
    let pred2 = esn.predict(&[5.0]).await.unwrap();

    assert_ne!(pred1, pred2, "State should evolve between predictions");
}

#[tokio::test]
async fn test_esn_to_npu_weights() {
    let config = ESNConfig {
        input_size: 1,
        reservoir_size: 20,
        output_size: 1,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config).await;
    let inputs = vec![vec![1.0], vec![2.0], vec![3.0]];
    let targets = vec![vec![2.0], vec![3.0], vec![4.0]];
    esn.train(&inputs, &targets).await.unwrap();

    let npu = esn.to_npu_weights().unwrap();
    assert_eq!(npu.input_dim, 20);
    assert_eq!(npu.output_dim, 1);
    assert_eq!(npu.weights_i8.len(), 20);
    assert!(npu.scale > 0.0);
}

#[tokio::test]
async fn test_esn_to_npu_weights_untrained() {
    let esn = esn_with_pool(ESNConfig::default()).await;
    assert!(esn.to_npu_weights().is_err());
}

#[tokio::test]
async fn test_esn_train_ridge_regression_linear() {
    let config = ESNConfig {
        input_size: 1,
        reservoir_size: 4,
        output_size: 1,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config).await;

    let n_samples = 10;
    let mut states = vec![0.0; 4 * n_samples];
    let mut targets = vec![0.0; n_samples];
    for k in 0..n_samples {
        let x = k as f64 * 0.5;
        states[k] = 1.0;
        states[n_samples + k] = x;
        states[2 * n_samples + k] = x * x;
        states[3 * n_samples + k] = x * x * x;
        targets[k] = 3.0f64.mul_add(x, 2.0);
    }

    esn.train_ridge_regression(&states, &targets, 1e-6).unwrap();
    assert!(esn.is_trained());

    let w = esn.w_out.as_ref().unwrap().to_vec().unwrap();
    assert_eq!(w.len(), 4);
    assert!((w[0] - 2.0).abs() < 0.1);
    assert!((w[1] - 3.0).abs() < 0.1);
}

#[tokio::test]
async fn test_esn_train_ridge_regression_regularization() {
    let config = ESNConfig {
        input_size: 1,
        reservoir_size: 5,
        output_size: 1,
        ..Default::default()
    };

    let mut esn_small = esn_with_pool(config.clone()).await;
    let mut esn_large = esn_with_pool(config).await;

    let n_samples = 8;
    let mut states = vec![0.0; 5 * n_samples];
    let mut targets = vec![0.0; n_samples];
    for k in 0..n_samples {
        for i in 0..5 {
            states[i * n_samples + k] = (k as f64 + i as f64) * 0.1;
        }
        targets[k] = (k as f64) * 0.2;
    }

    esn_small
        .train_ridge_regression(&states, &targets, 1e-6)
        .unwrap();
    esn_large
        .train_ridge_regression(&states, &targets, 10.0)
        .unwrap();

    let w_small = esn_small.w_out.as_ref().unwrap().to_vec().unwrap();
    let w_large = esn_large.w_out.as_ref().unwrap().to_vec().unwrap();

    let norm_small: f32 = w_small.iter().map(|x| x * x).sum::<f32>().sqrt();
    let norm_large: f32 = w_large.iter().map(|x| x * x).sum::<f32>().sqrt();

    assert!(
        norm_large < norm_small,
        "Larger lambda should produce smaller weights: {norm_large} < {norm_small}"
    );
}

#[tokio::test]
async fn test_esn_export_import_weights() {
    let config = ESNConfig {
        input_size: 1,
        reservoir_size: 20,
        output_size: 1,
        ..Default::default()
    };

    let mut esn = esn_with_pool(config.clone()).await;
    let inputs = vec![vec![1.0], vec![2.0], vec![3.0]];
    let targets = vec![vec![2.0], vec![3.0], vec![4.0]];
    esn.train(&inputs, &targets).await.unwrap();

    let exported = esn.export_weights().unwrap();
    assert_eq!(exported.w_in.len(), 20);
    assert_eq!(exported.w_res.len(), 400);
    assert!(exported.w_out.is_some());
    assert_eq!(exported.w_out.as_ref().unwrap().len(), 20);

    let mut esn2 = esn_with_pool(config.clone()).await;
    esn2.import_weights(&exported.w_in, &exported.w_res, exported.w_out.as_deref())
        .unwrap();
    assert!(esn2.is_trained());
}

#[tokio::test]
async fn test_esn_exported_weights_migrate_to_multi_head() {
    let exported = super::ExportedWeights {
        w_in: vec![0.0; 20],
        w_res: vec![0.0; 400],
        w_out: Some(vec![1.0; 20]),
        input_size: 1,
        reservoir_size: 20,
        output_size: 1,
        leak_rate: 0.3,
        head_labels: Vec::new(),
    };

    let migrated = exported.migrate_to_multi_head(20, 11).unwrap();
    assert_eq!(migrated.w_out.as_ref().unwrap().len(), 11 * 20);
    assert_eq!(migrated.output_size, 11);
}

#[tokio::test]
async fn test_esn_reservoir_shape() {
    let config = ESNConfig {
        input_size: 3,
        reservoir_size: 16,
        output_size: 2,
        spectral_radius: 0.95,
        connectivity: 0.1,
        leak_rate: 0.2,
        regularization: 1e-4,
        seed: 42,
        ..Default::default()
    };

    let esn = esn_with_pool(config).await;
    assert_eq!(esn.state().shape(), &[16, 1]);
}