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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! `GatedConv2D` - Gated Convolution 2D
//!
//! **Deep Debt Principles**:
//! - ✅ Pure WGSL implementation
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Complete implementation (production-ready)
//! - ✅ Modern idiomatic Rust (no traits, direct impl)
//!
//! Convolution with multiplicative gating mechanism
//! Used in `PixelCNN`, `WaveNet`, and generative models
//!
//! Output = `tanh(W_f` * x) ⊙ `sigmoid(W_g` * x)
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 GatedConv2DParams {
batch_size: u32,
in_channels: u32,
out_channels: u32,
in_height: u32,
in_width: u32,
out_height: u32,
out_width: u32,
kernel_size: u32,
stride: u32,
padding: u32,
_padding: u32,
_padding2: u32,
}
/// Gated 2D convolution (PixelCNN/WaveNet style: `tanh(W_f`*x) ⊙ `sigmoid(W_g`*x)).
pub struct GatedConv2D {
input: Tensor,
weight_feature: Tensor,
weight_gate: Tensor,
bias_feature: Tensor,
bias_gate: Tensor,
kernel_size: usize,
stride: usize,
padding: usize,
}
impl GatedConv2D {
/// Creates a new gated conv2d. Input must be 4D [B, C, H, W].
/// # Errors
/// Returns [`Err`] if input is not 4D [B, C, H, W], or if `kernel_size` or stride is zero.
pub fn new(
input: Tensor,
weight_feature: Tensor,
weight_gate: Tensor,
bias_feature: Tensor,
bias_gate: Tensor,
kernel_size: usize,
stride: usize,
padding: usize,
) -> Result<Self> {
// Validate input shape: must be 4D [B, C, H, W]
let input_shape = input.shape();
if input_shape.len() != 4 {
return Err(BarracudaError::invalid_op(
"gated_conv2d",
"input must be 4D tensor [B, C, H, W]",
));
}
if kernel_size == 0 || stride == 0 {
return Err(BarracudaError::invalid_op(
"gated_conv2d",
"kernel_size and stride must be positive",
));
}
Ok(Self {
input,
weight_feature,
weight_gate,
bias_feature,
bias_gate,
kernel_size,
stride,
padding,
})
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/conv/gated_conv2d_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
/// Executes gated conv2d and returns the output tensor.
/// # Errors
/// Returns [`Err`] if buffer allocation fails, GPU dispatch fails, or the device is lost.
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let input_shape = self.input.shape();
let batch_size = input_shape[0];
let in_channels = input_shape[1];
let in_height = input_shape[2];
let in_width = input_shape[3];
let out_height = ((in_height + 2 * self.padding - self.kernel_size) / self.stride) + 1;
let out_width = ((in_width + 2 * self.padding - self.kernel_size) / self.stride) + 1;
let out_channels = self.bias_feature.shape()[0];
let output_size = batch_size * out_channels * out_height * out_width;
let output_buffer = device.create_buffer_f32(output_size)?;
let params = GatedConv2DParams {
batch_size: batch_size as u32,
in_channels: in_channels as u32,
out_channels: out_channels as u32,
in_height: in_height as u32,
in_width: in_width as u32,
out_height: out_height as u32,
out_width: out_width as u32,
kernel_size: self.kernel_size as u32,
stride: self.stride as u32,
padding: self.padding as u32,
_padding: 0,
_padding2: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("gated_conv2d_params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("gated_conv2d_shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("gated_conv2d_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::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("gated_conv2d_pipeline_layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("gated_conv2d_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("gated_conv2d_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_feature.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.weight_gate.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: self.bias_feature.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: self.bias_gate.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 5,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 6,
resource: params_buffer.as_entire_binding(),
},
],
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("gated_conv2d_encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("gated_conv2d_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::Convolution);
let workgroups_x = (out_width as u32).div_ceil(optimal_wg_size);
let workgroups_y = (out_height as u32).div_ceil(optimal_wg_size);
let workgroups_z = batch_size * out_channels;
compute_pass.dispatch_workgroups(workgroups_x, workgroups_y, workgroups_z as u32);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
vec![batch_size, out_channels, out_height, out_width],
device.clone(),
))
}
}
impl Tensor {
/// Apply gated convolution 2D
/// # Arguments
/// - `weight_feature`: Feature weight tensor [`C_out`, `C_in`, K, K]
/// - `weight_gate`: Gate weight tensor [`C_out`, `C_in`, K, K]
/// - `bias_feature`: Feature bias tensor [`C_out`]
/// - `bias_gate`: Gate bias tensor [`C_out`]
/// - `kernel_size`: Kernel size
/// - `stride`: Stride
/// - `padding`: Padding
/// # Errors
/// Returns [`Err`] if input is not 4D, `kernel_size/stride` is zero, buffer allocation fails,
/// GPU dispatch fails, or the device is lost.
pub fn gated_conv2d(
self,
weight_feature: Self,
weight_gate: Self,
bias_feature: Self,
bias_gate: Self,
kernel_size: usize,
stride: usize,
padding: usize,
) -> Result<Self> {
GatedConv2D::new(
self,
weight_feature,
weight_gate,
bias_feature,
bias_gate,
kernel_size,
stride,
padding,
)?
.execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_gated_conv2d_basic() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_vec_on(vec![1.0; 3 * 4 * 4], vec![1, 3, 4, 4], device.clone())
.await
.unwrap();
let weight_feature =
Tensor::from_vec_on(vec![0.1; 4 * 3 * 3 * 3], vec![4, 3, 3, 3], device.clone())
.await
.unwrap();
let weight_gate =
Tensor::from_vec_on(vec![0.1; 4 * 3 * 3 * 3], vec![4, 3, 3, 3], device.clone())
.await
.unwrap();
let bias_feature = Tensor::from_vec_on(vec![0.0; 4], vec![4], device.clone())
.await
.unwrap();
let bias_gate = Tensor::from_vec_on(vec![0.0; 4], vec![4], device.clone())
.await
.unwrap();
let output = input
.gated_conv2d(
weight_feature,
weight_gate,
bias_feature,
bias_gate,
3,
1,
1,
)
.unwrap();
let result = output.to_vec().unwrap();
assert_eq!(output.shape()[0], 1);
assert_eq!(output.shape()[1], 4);
assert!(result.iter().all(|&x| x.is_finite()));
}
}