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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! RNN Cell - Basic recurrent neural network cell
//!
//! **Canonical `BarraCuda` Pattern**: Struct with new/execute
//!
//! ## Algorithm
//!
//! ```text
//! h_t = tanh(W_ih * x_t + b_ih + W_hh * h_{t-1} + b_hh)
//! ```

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

/// RNN cell weight matrices and biases.
#[derive(Clone)]
pub struct RNNWeights {
    /// Input-to-hidden weights (`hidden_size` × `input_size`).
    pub w_ih: Vec<f32>,
    /// Hidden-to-hidden weights (`hidden_size` × `hidden_size`).
    pub w_hh: Vec<f32>,
    /// Input bias (`hidden_size`).
    pub b_ih: Vec<f32>,
    /// Hidden bias (`hidden_size`).
    pub b_hh: Vec<f32>,
}

/// RNN cell operation (`h_t` = `tanh(W_ih` x + `b_ih` + `W_hh` h + `b_hh`)).
pub struct RNNCell {
    input: Tensor,
    prev_hidden: Tensor,
    weights: RNNWeights,
    batch_size: usize,
    input_size: usize,
    hidden_size: usize,
}

impl RNNCell {
    /// Create a new RNN cell operation.
    /// # Errors
    /// Returns [`Err`] if input or hidden shapes are invalid, or if weight/bias
    /// dimensions do not match the expected sizes.
    pub fn new(
        input: Tensor,
        prev_hidden: Tensor,
        weights: RNNWeights,
        batch_size: usize,
        input_size: usize,
        hidden_size: usize,
    ) -> Result<Self> {
        // Validate input shapes
        let input_shape = input.shape();
        let hidden_shape = prev_hidden.shape();

        if input_shape.len() != 2 || input_shape[0] != batch_size || input_shape[1] != input_size {
            return Err(BarracudaError::InvalidShape {
                expected: vec![batch_size, input_size],
                actual: input_shape.to_vec(),
            });
        }

        if hidden_shape.len() != 2
            || hidden_shape[0] != batch_size
            || hidden_shape[1] != hidden_size
        {
            return Err(BarracudaError::InvalidShape {
                expected: vec![batch_size, hidden_size],
                actual: hidden_shape.to_vec(),
            });
        }

        // Validate weight dimensions
        let expected_w_ih_size = hidden_size * input_size;
        let expected_w_hh_size = hidden_size * hidden_size;

        if weights.w_ih.len() != expected_w_ih_size {
            return Err(BarracudaError::InvalidInput {
                message: format!(
                    "w_ih size mismatch: expected {}, got {}",
                    expected_w_ih_size,
                    weights.w_ih.len()
                ),
            });
        }

        if weights.w_hh.len() != expected_w_hh_size {
            return Err(BarracudaError::InvalidInput {
                message: format!(
                    "w_hh size mismatch: expected {}, got {}",
                    expected_w_hh_size,
                    weights.w_hh.len()
                ),
            });
        }

        if weights.b_ih.len() != hidden_size {
            return Err(BarracudaError::InvalidInput {
                message: format!(
                    "b_ih size mismatch: expected {}, got {}",
                    hidden_size,
                    weights.b_ih.len()
                ),
            });
        }

        if weights.b_hh.len() != hidden_size {
            return Err(BarracudaError::InvalidInput {
                message: format!(
                    "b_hh size mismatch: expected {}, got {}",
                    hidden_size,
                    weights.b_hh.len()
                ),
            });
        }

        Ok(Self {
            input,
            prev_hidden,
            weights,
            batch_size,
            input_size,
            hidden_size,
        })
    }

    /// Get the WGSL shader source
    fn wgsl_shader() -> &'static str {
        static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
            include_str!("../shaders/rnn/rnn_cell_f64.wgsl").to_string()
        });
        &SHADER
    }

    /// Execute the RNN cell and return the new hidden state.
    /// # Errors
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn execute(self) -> Result<Tensor> {
        let device = self.input.device();
        let output_size = self.batch_size * self.hidden_size;

        // Create buffers for weights and biases
        let w_ih_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("RNN w_ih"),
                contents: bytemuck::cast_slice(&self.weights.w_ih),
                usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
            });

        let w_hh_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("RNN w_hh"),
                contents: bytemuck::cast_slice(&self.weights.w_hh),
                usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
            });

        let b_ih_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("RNN b_ih"),
                contents: bytemuck::cast_slice(&self.weights.b_ih),
                usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
            });

        let b_hh_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("RNN b_hh"),
                contents: bytemuck::cast_slice(&self.weights.b_hh),
                usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
            });

        // Create output buffer
        let output_buffer = device.create_buffer_f32(output_size)?;

        // Create uniform buffer for parameters
        #[repr(C)]
        #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
        struct Params {
            batch_size: u32,
            input_size: u32,
            hidden_size: u32,
            _padding: u32,
        }

        let params = Params {
            batch_size: self.batch_size as u32,
            input_size: self.input_size as u32,
            hidden_size: self.hidden_size as u32,
            _padding: 0,
        };

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

        // Compile shader
        let shader_module = device.compile_shader(Self::wgsl_shader(), Some("RNN Cell Shader"));

        // Create bind group layout
        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("RNN Cell 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: true },
                                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: true },
                                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::Storage { read_only: true },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                        wgpu::BindGroupLayoutEntry {
                            binding: 6,
                            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: 7,
                            visibility: wgpu::ShaderStages::COMPUTE,
                            ty: wgpu::BindingType::Buffer {
                                ty: wgpu::BufferBindingType::Uniform,
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                    ],
                });

        // Create bind group
        let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("RNN Cell Bind Group"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: self.input.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: w_ih_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: w_hh_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: b_ih_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: b_hh_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 5,
                    resource: self.prev_hidden.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 6,
                    resource: output_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 7,
                    resource: params_buffer.as_entire_binding(),
                },
            ],
        });

        // Create compute pipeline
        let pipeline_layout =
            device
                .device
                .create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
                    label: Some("RNN Cell Pipeline Layout"),
                    bind_group_layouts: &[&bind_group_layout],
                    immediate_size: 0,
                });

        let compute_pipeline =
            device
                .device
                .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                    label: Some("RNN Cell Pipeline"),
                    layout: Some(&pipeline_layout),
                    module: &shader_module,
                    entry_point: Some("main"),
                    cache: None,
                    compilation_options: Default::default(),
                });

        // Execute compute shader
        let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("RNN Cell Encoder"),
        });

        {
            let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("RNN Cell Pass"),
                timestamp_writes: None,
            });
            compute_pass.set_pipeline(&compute_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 = (self.batch_size as u32).div_ceil(optimal_wg_size);
            compute_pass.dispatch_workgroups(workgroups, 1, 1);
        }

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

        // Output shape: [batch_size, hidden_size]
        let output_shape = vec![self.batch_size, self.hidden_size];

        // Return tensor without reading back (zero-copy)
        Ok(Tensor::from_buffer(
            output_buffer,
            output_shape,
            device.clone(),
        ))
    }
}

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

    #[tokio::test]
    async fn test_rnn_cell_basic() {
        let device = crate::device::test_pool::get_test_device().await;
        let input = Tensor::from_vec_on(vec![0.5; 2 * 4], vec![2, 4], device.clone())
            .await
            .unwrap();
        let prev_hidden = Tensor::from_vec_on(vec![0.0; 2 * 8], vec![2, 8], device.clone())
            .await
            .unwrap();
        let weights = RNNWeights {
            w_ih: vec![0.01; 8 * 4],
            w_hh: vec![0.01; 8 * 8],
            b_ih: vec![0.0; 8],
            b_hh: vec![0.0; 8],
        };
        let hidden = RNNCell::new(input, prev_hidden, weights, 2, 4, 8)
            .unwrap()
            .execute()
            .unwrap();
        assert_eq!(hidden.shape(), &[2, 8]);
        let result = hidden.to_vec().unwrap();
        assert_eq!(result.len(), 16);
        assert!(result.iter().all(|&x| x.is_finite()));
    }

    #[tokio::test]
    async fn test_rnn_cell_edge_cases() {
        let device = crate::device::test_pool::get_test_device().await;
        // Single batch
        let input = Tensor::from_vec_on(vec![1.0; 3], vec![1, 3], device.clone())
            .await
            .unwrap();
        let prev_hidden = Tensor::from_vec_on(vec![0.0; 4], vec![1, 4], device.clone())
            .await
            .unwrap();
        let weights = RNNWeights {
            w_ih: vec![0.1; 4 * 3],
            w_hh: vec![0.1; 4 * 4],
            b_ih: vec![0.0; 4],
            b_hh: vec![0.0; 4],
        };
        let hidden = RNNCell::new(input, prev_hidden, weights, 1, 3, 4)
            .unwrap()
            .execute()
            .unwrap();
        assert_eq!(hidden.shape(), &[1, 4]);
    }

    #[tokio::test]
    async fn test_rnn_cell_boundary() {
        let device = crate::device::test_pool::get_test_device().await;
        // Non-zero previous hidden state
        let input = Tensor::from_vec_on(vec![0.5; 4], vec![1, 4], device.clone())
            .await
            .unwrap();
        let prev_hidden = Tensor::from_vec_on(vec![0.5; 8], vec![1, 8], device.clone())
            .await
            .unwrap();
        let weights = RNNWeights {
            w_ih: vec![0.1; 8 * 4],
            w_hh: vec![0.1; 8 * 8],
            b_ih: vec![0.1; 8],
            b_hh: vec![0.1; 8],
        };
        let hidden = RNNCell::new(input, prev_hidden, weights, 1, 4, 8)
            .unwrap()
            .execute()
            .unwrap();
        let result = hidden.to_vec().unwrap();
        assert!(result.iter().all(|&x| x.is_finite()));
        // tanh bounds: -1 < x < 1
        assert!(result.iter().all(|&x| x > -1.0 && x < 1.0));
    }
}