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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Interpolate Nearest - Nearest neighbor interpolation - 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
//!
//! Resizes images using nearest neighbor sampling:
//! ```text
//! Input: [B, C, H_in, W_in]
//! Output: [B, C, H_out, W_out]
//!
//! For each output pixel, finds nearest input pixel:
//! in_y = floor(out_y * scale_h)
//! in_x = floor(out_x * scale_w)
//! ```
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
/// Nearest-neighbor interpolation for 4D tensors [batch, channels, height, width].
pub struct InterpolateNearest {
input: Tensor,
output_size: (usize, usize), // (height, width)
}
impl InterpolateNearest {
/// Create interpolator with target (height, width).
#[must_use]
pub fn new(input: Tensor, output_size: (usize, usize)) -> Self {
Self { input, output_size }
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/misc/interpolate_nearest_f64.wgsl").to_string()
});
&SHADER
}
/// Execute nearest-neighbor interpolation 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::InvalidShape {
expected: vec![0, 0, 0, 0],
actual: shape.to_vec(),
});
}
let batch_size = shape[0];
let channels = shape[1];
let in_height = shape[2];
let in_width = shape[3];
let (out_height, out_width) = self.output_size;
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("InterpolateNearest 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("InterpolateNearest 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("InterpolateNearest"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("InterpolateNearest PL"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("InterpolateNearest 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("InterpolateNearest Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("InterpolateNearest 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::Convolution);
let workgroups_x = (out_width as u32).div_ceil(optimal_wg_size);
let workgroups_y = (out_height as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups_x, workgroups_y, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
vec![batch_size, channels, out_height, out_width],
device.clone(),
))
}
}
impl Tensor {
/// Resize image using nearest-neighbor sampling. Expects [B, C, H, W].
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn interpolate_nearest_wgsl(self, output_size: (usize, usize)) -> Result<Self> {
InterpolateNearest::new(self, output_size).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_interpolate_nearest_upsample() {
let device = crate::device::test_pool::get_test_device().await;
// 1x1x2x2 image, upsample to 4x4
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)
.await
.unwrap();
let result = input.interpolate_nearest_wgsl((4, 4)).unwrap();
let output = result.to_vec().unwrap();
// Each input pixel should be replicated 2x2
assert_eq!(output.len(), 16);
// Top-left quadrant should all be 1.0
assert!((output[0] - 1.0).abs() < 1e-5);
}
}