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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Group Normalization - Normalize within groups - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its group parameters
//! - Zero hardcoding: All parameters passed at runtime
//! - Modern idiomatic Rust: Safe, zero unsafe code
//! - Complete implementation: Production-ready, no mocks
//! - Hardware-agnostic: Pure WGSL for universal compute
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
/// Group normalization operation
pub struct GroupNorm {
input: Tensor,
num_groups: usize,
epsilon: f32,
}
impl GroupNorm {
/// Create a new group normalization operation
#[must_use]
pub fn new(input: Tensor, num_groups: usize, epsilon: f32) -> Self {
Self {
input,
num_groups,
epsilon,
}
}
/// Get the WGSL shader source
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/norm/group_norm_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
/// Execute the group normalization operation
///
/// # 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();
let size: usize = shape.iter().product();
// Assume NCHW format: [batch, channels, height, width]
let batch = shape[0];
let channels = shape[1];
let spatial_size: usize = shape[2..].iter().product();
let group_size = channels / self.num_groups;
// Create buffers
// Access input buffer directly (zero-copy)
let input_buffer = self.input.buffer();
let output_buffer = device.create_buffer_f32(size)?;
// Create uniform buffer for parameters
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
batch: u32,
channels: u32,
num_groups: u32,
group_size: u32,
spatial_size: u32,
epsilon: f32,
}
let params = Params {
batch: batch as u32,
channels: channels as u32,
num_groups: self.num_groups as u32,
group_size: group_size as u32,
spatial_size: spatial_size as u32,
epsilon: self.epsilon,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GroupNorm Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
// Create bind group layout
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("GroupNorm 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: 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,
},
],
});
// Create bind group
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("GroupNorm Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: 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(),
},
],
});
// Create compute pipeline
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("Shader"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("GroupNorm Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GroupNorm 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("GroupNorm Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GroupNorm 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 (vendor-optimized workgroups)
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
let total_groups = (batch * self.num_groups) as u32;
let workgroups = total_groups.div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
// Read back results
let output_data = crate::utils::read_buffer(device, &output_buffer, size)?;
Ok(Tensor::new(output_data, shape.to_vec(), device.clone()))
}
}
impl Tensor {
/// Apply group normalization (NCHW format)
///
/// # Arguments
///
/// * `num_groups` - Number of groups to divide channels into
/// * `epsilon` - Small constant for numerical stability (default: 1e-5)
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn group_norm_wgsl(self, num_groups: usize, epsilon: f32) -> Result<Self> {
GroupNorm::new(self, num_groups, epsilon).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_group_norm_simple() {
let device = crate::device::test_pool::get_test_device().await;
// 1 batch, 4 channels, 1x1 spatial, 2 groups
let data = vec![1.0, 2.0, 3.0, 4.0];
let input = Tensor::new(data, vec![1, 4, 1, 1], device);
let output = input.group_norm_wgsl(2, 1e-5).unwrap();
assert_eq!(output.shape(), &[1, 4, 1, 1]);
// Each group should be normalized independently
let result = output.to_vec().unwrap();
// Group 1: [1.0, 2.0] -> mean ~0, std ~1
// Group 2: [3.0, 4.0] -> mean ~0, std ~1
assert_eq!(result.len(), 4);
}
#[tokio::test]
async fn test_group_norm_batch() {
let device = crate::device::test_pool::get_test_device().await;
// 2 batches, 2 channels, 1x1 spatial, 1 group
let data = vec![
1.0, 2.0, // batch 0
3.0, 4.0, // batch 1
];
let input = Tensor::new(data, vec![2, 2, 1, 1], device);
let output = input.group_norm_wgsl(1, 1e-5).unwrap();
assert_eq!(output.shape(), &[2, 2, 1, 1]);
assert_eq!(output.to_vec().unwrap().len(), 4);
}
}