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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Circular Pad - Pad with circular wrapping - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its padding 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;
use crate::error::Result;
use crate::tensor::Tensor;
/// Circular pad operation
pub struct CircularPad {
input: Tensor,
padding: (usize, usize, usize, usize), // (left, right, top, bottom)
}
impl CircularPad {
/// Create a new circular pad operation
#[must_use]
pub fn new(input: Tensor, padding: (usize, usize, usize, usize)) -> Self {
Self { input, padding }
}
/// Get the WGSL shader source
fn wgsl_shader() -> &'static str {
{
static S: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/tensor/circular_pad_f64.wgsl").to_string()
});
&S
}
}
/// Execute the circular pad 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();
// Assume NCHW format
let batch = shape[0];
let channels = shape[1];
let in_height = shape[2];
let in_width = shape[3];
let (pad_left, pad_right, pad_top, pad_bottom) = self.padding;
let out_height = in_height + pad_top + pad_bottom;
let out_width = in_width + pad_left + pad_right;
let output_size = batch * channels * out_height * out_width;
// Create buffers
// Access input buffer directly (zero-copy)
let input_buffer = self.input.buffer();
let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("CircularPad Output"),
size: (output_size * std::mem::size_of::<f32>()) as u64,
usage: wgpu::BufferUsages::STORAGE
| wgpu::BufferUsages::COPY_SRC
| wgpu::BufferUsages::COPY_DST,
mapped_at_creation: false,
});
// Create uniform buffer for parameters
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
batch: u32,
channels: u32,
in_height: u32,
in_width: u32,
out_height: u32,
out_width: u32,
pad_top: u32,
pad_left: u32,
}
let params = Params {
batch: batch as u32,
channels: 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,
pad_top: pad_top as u32,
pad_left: pad_left as u32,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("CircularPad 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("CircularPad 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("CircularPad 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("CircularPad Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("CircularPad 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("CircularPad Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("CircularPad Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&compute_pipeline);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
// Dispatch using standard 2D shader workgroup size (16, 16)
let caps = DeviceCapabilities::from_device(device);
let (workgroups_x, workgroups_y) =
caps.dispatch_2d(out_width as u32, out_height as u32);
compute_pass.dispatch_workgroups(workgroups_x, workgroups_y, 1);
}
device.submit_commands(Some(encoder.finish()));
// Read back results
let output_data = crate::utils::read_buffer(device, &output_buffer, output_size)?;
Ok(Tensor::new(
output_data,
vec![batch, channels, out_height, out_width],
device.clone(),
))
}
}
impl Tensor {
/// Apply circular padding (NCHW format)
///
/// # Arguments
///
/// * `padding` - (left, right, top, bottom) padding amounts
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn circular_pad_wgsl(self, padding: (usize, usize, usize, usize)) -> Result<Self> {
CircularPad::new(self, padding).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_circular_pad_simple() {
let device = crate::device::test_pool::get_test_device().await;
// 1x1x2x2 input
let data = vec![1.0, 2.0, 3.0, 4.0];
let input = Tensor::new(data, vec![1, 1, 2, 2], device);
// Pad by 1 on all sides
let output = input.circular_pad_wgsl((1, 1, 1, 1)).unwrap();
assert_eq!(output.shape(), &[1, 1, 4, 4]);
assert_eq!(output.to_vec().unwrap().len(), 16);
}
#[tokio::test]
async fn test_circular_pad_wrapping() {
let device = crate::device::test_pool::get_test_device().await;
// 1x1x3x3 input (NCHW: row 0 = [1,2,3], row 1 = [4,5,6], row 2 = [7,8,9])
let data = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0];
let input = Tensor::from_data(&data, vec![1, 1, 3, 3], device).unwrap();
// Pad by 1 on left (should wrap from right edge)
let output = input.circular_pad_wgsl((1, 0, 0, 0)).unwrap();
assert_eq!(output.shape(), &[1, 1, 3, 4]);
assert_eq!(output.to_vec().unwrap().len(), 12);
// First column should be wrapped from the last column
let result = output.to_vec().unwrap();
// Row 0: [3, 1, 2, 3]
assert_eq!(result[0], 3.0); // Wrapped from right
assert_eq!(result[1], 1.0);
}
}