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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Max Pool 1D - Temporal max pooling - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its own requirements (kernel size, stride)
//! - 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, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
/// Max Pool 1D operation - Temporal max pooling
///
/// Applies 1D max pooling over an input signal (batch, channels, length).
pub struct MaxPool1D {
input: Tensor,
kernel_size: usize,
stride: usize,
}
impl MaxPool1D {
/// Create a new `MaxPool1D` operation
#[must_use]
pub fn new(input: Tensor, kernel_size: usize, stride: usize) -> Self {
Self {
input,
kernel_size,
stride,
}
}
/// Get the WGSL shader source
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/pooling/max_pool1d_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
/// Execute the max pool 1D 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();
// Validate input shape (batch, channels, length)
if shape.len() != 3 {
return Err(BarracudaError::invalid_shape(
vec![0, 0, 0], // Expected: 3D tensor
shape.to_vec(),
));
}
let batch_size = shape[0];
let channels = shape[1];
let input_size = shape[2];
// Calculate output size
let output_size = (input_size - self.kernel_size) / self.stride + 1;
let total_elements = batch_size * channels * output_size;
// 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("MaxPool1D Output"),
size: (total_elements * std::mem::size_of::<f32>()) as u64,
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
mapped_at_creation: false,
});
// Create uniform buffer for parameters
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
input_size: u32,
output_size: u32,
channels: u32,
batch_size: u32,
kernel_size: u32,
stride: u32,
}
let params = Params {
input_size: input_size as u32,
output_size: output_size as u32,
channels: channels as u32,
batch_size: batch_size as u32,
kernel_size: self.kernel_size as u32,
stride: self.stride as u32,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("MaxPool1D 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("MaxPool1D 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("MaxPool1D 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("MaxPool1D Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("MaxPool1D 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("MaxPool1D Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("MaxPool1D Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&compute_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::Convolution);
let workgroups = (total_elements as u32).div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
// Read back results
let output_data = crate::utils::read_buffer(device, &output_buffer, total_elements)?;
Ok(Tensor::new(
output_data,
vec![batch_size, channels, output_size],
device.clone(),
))
}
}
impl Tensor {
/// Apply 1D max pooling over the tensor
///
/// # Arguments
///
/// * `kernel_size` - Size of the pooling window
/// * `stride` - Stride of the pooling operation
///
/// # Returns
///
/// Pooled tensor with shape (batch, channels, `output_length`)
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn max_pool1d_wgsl(self, kernel_size: usize, stride: usize) -> Result<Self> {
MaxPool1D::new(self, kernel_size, stride).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_max_pool1d_basic() {
let device = crate::device::test_pool::get_test_device().await;
// Input: [1, 1, 8] - single batch, single channel, 8 elements
let data = vec![1.0, 3.0, 2.0, 4.0, 5.0, 1.0, 6.0, 2.0];
let input = Tensor::new(data, vec![1, 1, 8], device);
let output = input.max_pool1d_wgsl(2, 2).unwrap();
assert_eq!(output.shape(), &[1, 1, 4]);
let result = output.to_vec().unwrap();
// Max of [1,3]=3, [2,4]=4, [5,1]=5, [6,2]=6
assert_eq!(result[0], 3.0);
assert_eq!(result[1], 4.0);
assert_eq!(result[2], 5.0);
assert_eq!(result[3], 6.0);
}
#[tokio::test]
async fn test_max_pool1d_multi_channel() {
let device = crate::device::test_pool::get_test_device().await;
// Input: [1, 2, 4] - single batch, 2 channels, 4 elements each
let data = vec![
1.0, 2.0, 3.0, 4.0, // Channel 0
4.0, 3.0, 2.0, 1.0, // Channel 1
];
let input = Tensor::new(data, vec![1, 2, 4], device);
let output = input.max_pool1d_wgsl(2, 2).unwrap();
assert_eq!(output.shape(), &[1, 2, 2]);
let result = output.to_vec().unwrap();
// Channel 0: max([1,2])=2, max([3,4])=4
// Channel 1: max([4,3])=4, max([2,1])=2
assert_eq!(result[0], 2.0);
assert_eq!(result[1], 4.0);
assert_eq!(result[2], 4.0);
assert_eq!(result[3], 2.0);
}
#[tokio::test]
async fn test_max_pool1d_stride_one() {
let device = crate::device::test_pool::get_test_device().await;
// Input: [1, 1, 5] - overlapping windows
let data = vec![1.0, 5.0, 2.0, 3.0, 4.0];
let input = Tensor::new(data, vec![1, 1, 5], device);
let output = input.max_pool1d_wgsl(3, 1).unwrap();
assert_eq!(output.shape(), &[1, 1, 3]);
let result = output.to_vec().unwrap();
// max([1,5,2])=5, max([5,2,3])=5, max([2,3,4])=4
assert_eq!(result[0], 5.0);
assert_eq!(result[1], 5.0);
assert_eq!(result[2], 4.0);
}
}