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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! `GraphNorm` - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its computation
//! - Zero hardcoding: Hardware-agnostic implementation
//! - 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::{BarracudaError, Result};
use crate::tensor::Tensor;
/// Graph Normalization operation
pub struct GraphNorm {
input: Tensor,
gamma: Tensor,
beta: Tensor,
num_nodes: usize,
num_features: usize,
epsilon: f32,
}
impl GraphNorm {
/// Create a new graph normalization 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, gamma: Tensor, beta: Tensor, epsilon: Option<f32>) -> Result<Self> {
let input_shape = input.shape();
let num_nodes = input_shape[0];
let num_features = input_shape[1..].iter().product::<usize>();
let gamma_size = gamma.shape().iter().product::<usize>();
if gamma_size != num_features {
return Err(BarracudaError::invalid_op(
"graph_norm",
"gamma must have num_features elements",
));
}
let beta_size = beta.shape().iter().product::<usize>();
if beta_size != num_features {
return Err(BarracudaError::invalid_op(
"graph_norm",
"beta must have num_features elements",
));
}
Ok(Self {
input,
gamma,
beta,
num_nodes,
num_features,
epsilon: epsilon.unwrap_or(1e-5),
})
}
/// Get the WGSL shader source
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/norm/graph_norm_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
/// Execute the graph 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();
// Create output buffer
let output_size = self.num_nodes * self.num_features;
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 {
num_nodes: u32,
num_features: u32,
epsilon: f32,
_padding: u32,
}
let params = Params {
num_nodes: self.num_nodes as u32,
num_features: self.num_features as u32,
epsilon: self.epsilon,
_padding: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GraphNorm 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("GraphNorm Shader"));
// Create bind group layout
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("GraphNorm 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: false },
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::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("GraphNorm Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.gamma.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.beta.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: params_buffer.as_entire_binding(),
},
],
});
// Create pipeline
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("GraphNorm Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GraphNorm Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
// Encode and execute
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("GraphNorm Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GraphNorm 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::Reduction);
let workgroups = (self.num_features 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![self.num_nodes, self.num_features],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_graph_norm_basic() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 3;
let num_features = 4;
let input = Tensor::from_vec_on(
vec![1.0; num_nodes * num_features],
vec![num_nodes, num_features],
device.clone(),
)
.await
.unwrap();
let gamma =
Tensor::from_vec_on(vec![1.0; num_features], vec![num_features], device.clone())
.await
.unwrap();
let beta = Tensor::from_vec_on(vec![0.0; num_features], vec![num_features], device.clone())
.await
.unwrap();
let output = GraphNorm::new(input, gamma, beta, None)
.unwrap()
.execute()
.unwrap();
assert_eq!(output.shape(), &[num_nodes, num_features]);
}
}