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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Bidirectional LSTM operation
//!
//! **Pure WGSL**: Single implementation via WebGPU shader
//! Processes sequence in both forward and backward directions
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 BiLSTMParams {
batch_size: u32,
seq_len: u32,
input_size: u32,
hidden_size: u32,
direction: u32, // 0 = forward, 1 = backward
}
/// Bidirectional LSTM: processes sequence in forward and backward directions.
pub struct BiLSTM {
input: Tensor,
weight_ih: Tensor,
weight_hh: Tensor,
bias_ih: Tensor,
bias_hh: Tensor,
direction: u32,
}
impl BiLSTM {
/// Create `BiLSTM` operation
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn new(
input: Tensor,
weight_ih: Tensor,
weight_hh: Tensor,
bias_ih: Tensor,
bias_hh: Tensor,
direction: u32,
) -> Result<Self> {
// Validate direction
if direction > 1 {
return Err(BarracudaError::invalid_op(
"BiLSTM",
format!("direction must be 0 (forward) or 1 (backward), got {direction}"),
));
}
Ok(Self {
input,
weight_ih,
weight_hh,
bias_ih,
bias_hh,
direction,
})
}
/// WGSL shader source (embedded at compile time)
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/rnn/bi_lstm_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
/// Execute `BiLSTM` on tensor
///
/// # 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 input_shape = self.input.shape();
if input_shape.len() != 3 {
return Err(BarracudaError::invalid_op(
"BiLSTM",
format!("input must be 3D [seq_len, batch, input_size], got shape {input_shape:?}"),
));
}
let seq_len = input_shape[0];
let batch_size = input_shape[1];
let input_size = input_shape[2];
let hidden_size = self.bias_ih.len() / 4; // 4 gates: i, f, g, o
// Create output buffer: [seq_len, batch, hidden_size]
let output_size = seq_len * batch_size * hidden_size;
let output_buffer = device.create_buffer_f32(output_size)?;
// Create h_state and c_state buffers: [batch, hidden_size]
let state_size = batch_size * hidden_size;
let h_state_buffer = device.create_buffer_f32(state_size)?;
let c_state_buffer = device.create_buffer_f32(state_size)?;
// Initialize states to zero
device.queue.write_buffer(
&h_state_buffer,
0,
bytemuck::cast_slice(&vec![0.0f32; state_size]),
);
device.queue.write_buffer(
&c_state_buffer,
0,
bytemuck::cast_slice(&vec![0.0f32; state_size]),
);
let params = BiLSTMParams {
batch_size: batch_size as u32,
seq_len: seq_len as u32,
input_size: input_size as u32,
hidden_size: hidden_size as u32,
direction: self.direction,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("BiLSTM Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
// Create bind group layout
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("BiLSTM 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: false },
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::Storage { read_only: false },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 8,
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("BiLSTM Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.weight_ih.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.weight_hh.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: self.bias_ih.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: self.bias_hh.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 5,
resource: h_state_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 6,
resource: c_state_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 7,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 8,
resource: params_buffer.as_entire_binding(),
},
],
});
// Compile shader
let shader = device.compile_shader(Self::wgsl_shader(), Some("BiLSTM"));
// Create pipeline
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("BiLSTM Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("BiLSTM Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
// Encode and execute
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("BiLSTM Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("BiLSTM Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
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 = (batch_size as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
// Create output tensor
Ok(Tensor::from_buffer(
output_buffer,
vec![seq_len, batch_size, hidden_size],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_bi_lstm_basic() {
let device = crate::device::test_pool::get_test_device().await;
// Create test tensors: [seq_len=2, batch=1, input_size=3]
let input = Tensor::from_vec_on(
vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0],
vec![2, 1, 3],
device.clone(),
)
.await
.unwrap();
let hidden_size = 4;
let weight_ih = Tensor::from_vec_on(
vec![0.1; 4 * hidden_size * 3], // [4*hidden, input]
vec![4 * hidden_size, 3],
device.clone(),
)
.await
.unwrap();
let weight_hh = Tensor::from_vec_on(
vec![0.1; 4 * hidden_size * hidden_size], // [4*hidden, hidden]
vec![4 * hidden_size, hidden_size],
device.clone(),
)
.await
.unwrap();
let bias_ih = Tensor::from_vec_on(
vec![0.0; 4 * hidden_size],
vec![4 * hidden_size],
device.clone(),
)
.await
.unwrap();
let bias_hh = Tensor::from_vec_on(
vec![0.0; 4 * hidden_size],
vec![4 * hidden_size],
device.clone(),
)
.await
.unwrap();
let result = BiLSTM::new(input, weight_ih, weight_hh, bias_ih, bias_hh, 0)
.unwrap()
.execute()
.unwrap();
assert_eq!(result.shape(), &[2, 1, 4]);
}
}