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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Leave-One-Out Cross-Validation for kernel methods - 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;
use crate::error::Result;
use crate::tensor::Tensor;
/// Leave-one-out cross-validation residuals for kernel methods.
/// `LOO_i` = (`y_i` - `pred_i`) / (1 - `H_ii`)
pub struct LooCv {
hat_matrix: Tensor,
y: Tensor,
predictions: Tensor,
}
impl LooCv {
/// Create LOO-CV residuals from hat matrix, targets, and predictions.
#[must_use]
pub fn new(hat_matrix: Tensor, y: Tensor, predictions: Tensor) -> Self {
Self {
hat_matrix,
y,
predictions,
}
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/interpolation/loo_cv_f64.wgsl").to_string()
});
SHADER.as_str()
}
/// Compute leave-one-out cross-validation residuals.
/// # 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.hat_matrix.device();
let n: usize = self.y.shape().iter().product();
if n == 0 {
return Ok(Tensor::new(vec![], vec![0], device.clone()));
}
let output_buffer = device.create_buffer_f32(n)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
n: u32,
}
let params = Params { n: n as u32 };
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("LOO-CV Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("LOO-CV 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,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("LOO-CV Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.hat_matrix.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.y.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.predictions.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("LOO-CV"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("LOO-CV Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("LOO-CV 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("LOO-CV Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("LOO-CV Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(device);
let workgroups = caps.dispatch_1d(n as u32);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(output_buffer, vec![n], device.clone()))
}
}
impl Tensor {
/// Compute LOO-CV residuals: (y - pred) / (1 - diag(H))
/// # Errors
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn loo_cv(self, y: Self, predictions: Self) -> Result<Self> {
LooCv::new(self, y, predictions).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_loo_cv() {
let device = crate::device::test_pool::get_test_device().await;
// Simple 2x2 case: H = [[0.5, 0.5], [0.5, 0.5]], y = [1, 2], pred = [0.8, 1.7]
// LOO_0 = (1 - 0.8) / (1 - 0.5) = 0.2/0.5 = 0.4
// LOO_1 = (2 - 1.7) / (1 - 0.5) = 0.3/0.5 = 0.6
let hat_matrix = vec![0.5f32, 0.5, 0.5, 0.5];
let y = vec![1.0f32, 2.0];
let pred = vec![0.8f32, 1.7];
let hat = Tensor::new(hat_matrix, vec![2, 2], device.clone());
let y_t = Tensor::new(y, vec![2], device.clone());
let pred_t = Tensor::new(pred, vec![2], device);
let output = hat.loo_cv(y_t, pred_t).unwrap();
let result = output.to_vec().unwrap();
assert!((result[0] - 0.4).abs() < 1e-5);
assert!((result[1] - 0.6).abs() < 1e-5);
}
}