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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Channel Shuffle - Rearrange CNN channels - Pure WGSL
//!
//! **Deep Debt Principles**:
//! - ✅ Pure WGSL implementation (GPU-optimized)
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Complete implementation (production-ready)
//!
//! ## Algorithm
//!
//! Rearranges channels for efficient group convolutions (ShuffleNet):
//! ```text
//! Input: [B, C, H, W] with G groups
//! Output: [B, C, H, W] with shuffled channels
//!
//! Shuffle pattern: c = g*cpg + i → c' = i*G + g
//! where g = group, i = index in group, cpg = channels per group
//! ```
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
/// Shuffle channels for group convolutions (`ShuffleNet`).
pub struct ChannelShuffle {
input: Tensor,
groups: usize,
}
impl ChannelShuffle {
/// Create a channel shuffle. Channels must be divisible by groups.
#[must_use]
pub fn new(input: Tensor, groups: usize) -> Self {
Self { input, groups }
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/tensor/channel_shuffle_f64.wgsl").to_string()
});
&SHADER
}
/// Execute channel shuffle on GPU.
///
/// # 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 shape = self.input.shape();
// Expect 4D tensor [batch, channels, height, width]
if shape.len() != 4 {
return Err(crate::error::BarracudaError::InvalidInput {
message: format!(
"ChannelShuffle requires 4D tensor [N, C, H, W], got {}D {:?}",
shape.len(),
shape
),
});
}
let batch_size = shape[0];
let channels = shape[1];
let height = shape[2];
let width = shape[3];
if !channels.is_multiple_of(self.groups) {
return Err(crate::error::BarracudaError::InvalidShape {
expected: vec![batch_size, channels, height, width],
actual: shape.to_vec(),
});
}
let channels_per_group = channels / self.groups;
let total_size = self.input.len();
let output_buffer = device.create_buffer_f32(total_size)?;
// Create params buffer
let params_data = [
batch_size as u32,
channels as u32,
height as u32,
width as u32,
self.groups as u32,
channels_per_group as u32,
];
let params_buffer = device.create_uniform_buffer("Params", ¶ms_data);
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("ChannelShuffle BGL"),
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: false },
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::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("ChannelShuffle BG"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("ChannelShuffle"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("ChannelShuffle PL"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("ChannelShuffle Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("ChannelShuffle Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("ChannelShuffle 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 = (total_size as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
shape.to_vec(),
device.clone(),
))
}
}
impl Tensor {
/// Shuffle channels into groups. Expects [B, C, H, W], C divisible by groups.
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn channel_shuffle_wgsl(self, groups: usize) -> Result<Self> {
ChannelShuffle::new(self, groups).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_channel_shuffle_simple() {
let device = crate::device::test_pool::get_test_device().await;
// Shape: [1, 4, 2, 2] with 2 groups
// Channels 0,1 in group 0, channels 2,3 in group 1
let input_data = vec![
// Channel 0
1.0, 2.0, 3.0, 4.0, // Channel 1
5.0, 6.0, 7.0, 8.0, // Channel 2
9.0, 10.0, 11.0, 12.0, // Channel 3
13.0, 14.0, 15.0, 16.0,
];
let input = Tensor::from_vec_on(input_data, vec![1, 4, 2, 2], device)
.await
.unwrap();
let result = input.channel_shuffle_wgsl(2).unwrap();
let output = result.to_vec().unwrap();
// After shuffle with 2 groups (cpg=2):
// c=0 (g=0,i=0) → c'=0*2+0=0
// c=1 (g=0,i=1) → c'=1*2+0=2
// c=2 (g=1,i=0) → c'=0*2+1=1
// c=3 (g=1,i=1) → c'=1*2+1=3
// New order: [0, 2, 1, 3] → [ch0, ch2, ch1, ch3]
// Verify first spatial position (0,0) of each channel
assert_eq!(output[0], 1.0); // ch0[0,0]
assert_eq!(output[4], 9.0); // ch2[0,0]
assert_eq!(output[8], 5.0); // ch1[0,0]
assert_eq!(output[12], 13.0); // ch3[0,0]
}
}