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
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
// SPDX-License-Identifier: AGPL-3.0-or-later
//! MAE Loss - GPU-accelerated Mean Absolute Error Loss
//!
//! **Deep Debt Principles**:
//! - ✅ Pure WGSL implementation (existing shader evolved)
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Complete implementation (production-ready)
//! - ✅ Modern idiomatic Rust (no traits, direct impl)
//!
//! ## Algorithm
//!
//! ```text
//! MAE = (1/n) * Σ |y_pred - y_true|
//! ```
//!
//! **Key Properties**:
//! - Less sensitive to outliers than MSE
//! - Linear penalty for errors
//! - Robust loss function
//! - Used in regression tasks
//!
//! **Used By**: Robust regression, forecasting, time series
//!
//! ## Usage
//!
//! ```rust,ignore
//! use barracuda::tensor::Tensor;
//!
//! let predictions = Tensor::randn(vec![1000]).await?;
//! let targets = Tensor::randn(vec![1000]).await?;
//!
//! let loss = predictions.mae_loss(&targets)?;
//! ```
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct MAELossParams {
reduction_mode: u32,
size: u32,
_padding: [u32; 2],
}
/// Mean absolute error loss for robust regression.
pub struct MAELoss {
predictions: Tensor,
targets: Tensor,
}
impl MAELoss {
/// Creates a new MAE loss. Shapes must match.
/// # Errors
/// Returns [`Err`] if prediction and target shapes do not match.
pub fn new(predictions: Tensor, targets: Tensor) -> Result<Self> {
// Validate shapes match
if predictions.shape() != targets.shape() {
return Err(BarracudaError::shape_mismatch(
predictions.shape().to_vec(),
targets.shape().to_vec(),
));
}
Ok(Self {
predictions,
targets,
})
}
fn wgsl_shader() -> &'static str {
include_str!("../shaders/loss/mae_loss.wgsl")
}
/// f64 MAE loss (tree reduction for accumulation accuracy).
#[must_use]
pub fn wgsl_shader_f64() -> &'static str {
include_str!("../shaders/loss/mae_loss_f64.wgsl")
}
/// Executes MAE loss and returns a scalar loss tensor.
/// # Errors
/// Returns [`Err`] if buffer allocation, GPU dispatch, or device submission fails (e.g. device lost).
pub fn execute(self) -> Result<Tensor> {
let device = self.predictions.device();
let size = self.predictions.shape().iter().product::<usize>();
let params = MAELossParams {
reduction_mode: 0, // mean
size: size as u32,
_padding: [0; 2],
};
let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("mae_loss_output"),
size: (size * std::mem::size_of::<f32>()) as u64,
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
mapped_at_creation: false,
});
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("mae_loss_params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("mae_loss_shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("mae_loss_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: 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 pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("mae_loss_pipeline_layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("mae_loss_pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("mae_loss_bind_group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.predictions.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.targets.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 mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("mae_loss_encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("mae_loss_pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&pipeline);
compute_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 = (size as u32).div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
self.predictions.shape().to_vec(),
device.clone(),
))
}
}
// ═══════════════════════════════════════════════════════════════
// TENSOR API INTEGRATION (MODERN IDIOMATIC RUST)
// ═══════════════════════════════════════════════════════════════
impl Tensor {
/// MAE Loss (Mean Absolute Error) - robust regression loss
/// **Deep Debt**: Essential for robust regression tasks
/// # Arguments
/// - `targets`: Target tensor [same shape as predictions]
/// # Returns
/// - Loss tensor [same shape as input]
/// # Example
/// ```rust,ignore
/// // Regression
/// let loss = predictions.mae_loss(&targets)?;
/// ```
/// # Note
/// - Less sensitive to outliers than MSE
/// - Linear penalty for errors
/// - Used in robust regression
/// # Errors
/// Returns [`Err`] if shapes do not match, or if buffer allocation, GPU dispatch, or device submission fails (e.g. device lost).
pub fn mae_loss(self, targets: &Self) -> Result<Self> {
MAELoss::new(self, targets.clone())?.execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_mae_loss_basic() {
let device = crate::device::test_pool::get_test_device().await;
let predictions = Tensor::from_vec_on(vec![1.0, 2.0, 3.0, 4.0], vec![4], device.clone())
.await
.unwrap();
let targets = Tensor::from_vec_on(vec![1.5, 2.5, 3.5, 4.5], vec![4], device.clone())
.await
.unwrap();
let loss = predictions.mae_loss(&targets).unwrap();
let data = loss.to_vec().unwrap();
assert_eq!(data.len(), 4);
assert!(data.iter().all(|&x| x.is_finite()));
assert!(data.iter().all(|&x| x >= 0.0)); // MAE is always non-negative
}
#[tokio::test]
async fn test_mae_loss_perfect() {
let device = crate::device::test_pool::get_test_device().await;
// Perfect predictions should have zero loss
let predictions = Tensor::from_vec_on(vec![1.0, 2.0, 3.0], vec![3], device.clone())
.await
.unwrap();
let targets = Tensor::from_vec_on(vec![1.0, 2.0, 3.0], vec![3], device.clone())
.await
.unwrap();
let loss = predictions.mae_loss(&targets).unwrap();
let data = loss.to_vec().unwrap();
assert!(data.iter().all(|&x| x.abs() < 1e-5));
}
#[tokio::test]
async fn test_mae_loss_validation() {
let device = crate::device::test_pool::get_test_device().await;
// Shape mismatch
let predictions = Tensor::from_vec_on(vec![1.0; 10], vec![10], device.clone())
.await
.unwrap();
let targets = Tensor::from_vec_on(vec![1.0; 5], vec![5], device.clone())
.await
.unwrap();
assert!(predictions.mae_loss(&targets).is_err());
}
#[tokio::test]
async fn test_mae_loss_large_batch() {
let device = crate::device::test_pool::get_test_device().await;
let size = 1000;
let predictions = Tensor::from_vec_on(vec![1.0; size], vec![size], device.clone())
.await
.unwrap();
let targets = Tensor::from_vec_on(vec![2.0; size], vec![size], device.clone())
.await
.unwrap();
let loss = predictions.mae_loss(&targets).unwrap();
let data = loss.to_vec().unwrap();
assert_eq!(data.len(), size);
assert!(data.iter().all(|&x| x.is_finite()));
// Should be close to 1.0 (absolute difference)
assert!((data[0] - 1.0).abs() < 0.1);
}
}