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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Grid Sample - Spatial transformer sampling - 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
//!
//! Samples input at arbitrary grid positions (Spatial Transformer Networks):
//! ```text
//! Input: [B, C, H_in, W_in] - Input tensor
//! Grid: [B, H_out, W_out, 2] - Sampling coordinates (normalized [-1, 1])
//! Output: [B, C, H_out, W_out] - Sampled output
//!
//! Uses bilinear interpolation for smooth sampling
//! ```
use crate::device::DeviceCapabilities;
use crate::error::Result;
use crate::tensor::Tensor;
/// Spatial transformer sampling: sample input at grid coordinates.
pub struct GridSample {
input: Tensor,
grid: Tensor,
}
impl GridSample {
/// Create a grid sample operation. Input [B,C,H,W], grid [`B,H_out,W_out,2`].
#[must_use]
pub fn new(input: Tensor, grid: Tensor) -> Self {
Self { input, grid }
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/misc/grid_sample_f64.wgsl").to_string()
});
&SHADER
}
/// Execute grid sampling on GPU with bilinear interpolation.
///
/// # 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 input_shape = self.input.shape();
let grid_shape = self.grid.shape();
// Expect input: [B, C, H, W]
if input_shape.len() != 4 {
return Err(crate::error::BarracudaError::InvalidShape {
expected: vec![0, 0, 0, 0],
actual: input_shape.to_vec(),
});
}
// Expect grid: [B, H_out, W_out, 2]
if grid_shape.len() != 4 || grid_shape[3] != 2 {
return Err(crate::error::BarracudaError::InvalidShape {
expected: vec![input_shape[0], 0, 0, 2],
actual: grid_shape.to_vec(),
});
}
let batch_size = input_shape[0];
let channels = input_shape[1];
let in_height = input_shape[2];
let in_width = input_shape[3];
let out_height = grid_shape[1];
let out_width = grid_shape[2];
if grid_shape[0] != batch_size {
return Err(crate::error::BarracudaError::InvalidShape {
expected: vec![batch_size, out_height, out_width, 2],
actual: grid_shape.to_vec(),
});
}
let output_size = batch_size * channels * out_height * out_width;
let output_buffer = device.create_buffer_f32(output_size)?;
// Create params buffer
let params_data = [
batch_size as u32,
channels as u32,
in_height as u32,
in_width as u32,
out_height as u32,
out_width 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("GridSample 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: 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: false },
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::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("GridSample BG"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.grid.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("GridSample"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("GridSample PL"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GridSample 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("GridSample Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GridSample Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
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);
pass.dispatch_workgroups(workgroups_x, workgroups_y, 1);
}
device.submit_commands(Some(encoder.finish()));
let output_data = crate::utils::read_buffer(device, &output_buffer, output_size)?;
Ok(Tensor::new(
output_data,
vec![batch_size, channels, out_height, out_width],
device.clone(),
))
}
}
impl Tensor {
/// Sample this tensor at grid coordinates. Uses bilinear interpolation.
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn grid_sample_wgsl(self, grid: Self) -> Result<Self> {
GridSample::new(self, grid).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_grid_sample_identity() {
let device = crate::device::test_pool::get_test_device().await;
// Simple 1x1x2x2 input
let input_data = vec![1.0, 2.0, 3.0, 4.0];
let input = Tensor::from_vec_on(input_data, vec![1, 1, 2, 2], device.clone())
.await
.unwrap();
// Identity grid (sample at original positions)
// Grid coordinates in normalized space [-1, 1]
let grid_data = vec![
-1.0, -1.0, // Top-left
1.0, -1.0, // Top-right
-1.0, 1.0, // Bottom-left
1.0, 1.0, // Bottom-right
];
let grid = Tensor::from_vec_on(grid_data, vec![1, 2, 2, 2], device)
.await
.unwrap();
let result = input.grid_sample_wgsl(grid).unwrap();
let output = result.to_vec().unwrap();
// Should be close to original (allowing for interpolation)
assert_eq!(output.len(), 4);
assert!((output[0] - 1.0).abs() < 0.2);
}
}