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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Chunk - Split tensor into chunks along dimension
//!
//! **Deep Debt Principles**:
//! - Complete implementation: Uses existing chunk.wgsl shader
//! - Zero hardcoding: All parameters configurable
//! - Self-knowledge: Validates chunk count and dimension
//! - Modern idiomatic Rust: Result<T, E>
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 ChunkParams {
start_offset: u32, // Start offset in the split dimension for this chunk
chunk_size: u32, // Size of this chunk along split dimension
split_dim: u32,
dim_size: u32,
inner_size: u32,
outer_size: u32,
output_size: u32,
_pad1: u32,
}
/// Split tensor into chunks along a dimension.
pub struct Chunk {
input: Tensor,
chunks: usize,
dim: usize,
}
impl Chunk {
/// Creates a new chunk operation. Splits into `chunks` chunks along `dim`.
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn new(input: Tensor, chunks: usize, dim: usize) -> Result<Self> {
if chunks == 0 {
return Err(BarracudaError::invalid_op(
"chunk",
"Cannot split into 0 chunks",
));
}
let shape = input.shape();
if dim >= shape.len() {
return Err(BarracudaError::invalid_op(
"chunk",
format!("dim {} exceeds tensor rank {}", dim, shape.len()),
));
}
// Note: PyTorch allows non-divisible chunks - first (dim_size % chunks) chunks
// get (dim_size // chunks) + 1 elements, rest get (dim_size // chunks) elements
Ok(Self { input, chunks, dim })
}
fn wgsl_shader() -> &'static str {
{
static S: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/tensor/chunk_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&S).as_str()
}
}
/// Executes chunking and returns a vector of chunk tensors.
///
/// # 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<Vec<Tensor>> {
let device = self.input.device();
let shape = self.input.shape();
let dim_size = shape[self.dim];
// PyTorch-style chunking: first (dim_size % chunks) chunks get +1 element
let base_chunk_size = dim_size / self.chunks;
let extra_chunks = dim_size % self.chunks;
// Compute sizes
let outer_size: usize = shape[..self.dim].iter().product();
let inner_size: usize = shape[self.dim + 1..].iter().product();
let mut results = Vec::with_capacity(self.chunks);
let mut start_offset = 0;
for chunk_idx in 0..self.chunks {
// Calculate chunk size: first extra_chunks get +1 element
let chunk_size = if chunk_idx < extra_chunks {
base_chunk_size + 1
} else {
base_chunk_size
};
let output_size = outer_size * chunk_size * inner_size;
let params = ChunkParams {
start_offset: start_offset as u32,
chunk_size: chunk_size as u32,
split_dim: self.dim as u32,
dim_size: dim_size as u32,
inner_size: inner_size as u32,
outer_size: outer_size as u32,
output_size: output_size as u32,
_pad1: 0,
};
start_offset += chunk_size;
let params_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Chunk Params"),
contents: bytemuck::bytes_of(¶ms),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let output_buffer = device.create_buffer_f32(output_size)?;
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Chunk Bind Group Layout"),
entries: &[
wgpu::BindGroupLayoutEntry {
binding: 0,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Uniform,
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,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("Chunk Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: output_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("Chunk"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Chunk Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Chunk 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("Chunk Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Chunk 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::ElementWise);
let workgroups = (output_size as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
// Compute output shape
let mut output_shape = shape.to_vec();
output_shape[self.dim] = chunk_size;
results.push(Tensor::from_buffer(
output_buffer,
output_shape,
device.clone(),
));
}
Ok(results)
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_chunk_basic() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_vec_on(
(0..12).map(|i| i as f32).collect(),
vec![3, 4],
device.clone(),
)
.await
.unwrap();
let chunks = Chunk::new(input, 2, 0).unwrap().execute().unwrap();
assert_eq!(chunks.len(), 2);
// PyTorch-style: first chunk gets +1 element when not divisible
// 3 elements into 2 chunks: first gets 2, second gets 1
assert_eq!(chunks[0].shape(), &[2, 4]);
assert_eq!(chunks[1].shape(), &[1, 4]);
}
#[tokio::test]
async fn test_chunk_along_dim() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_vec_on(
(0..12).map(|i| i as f32).collect(),
vec![2, 6],
device.clone(),
)
.await
.unwrap();
let chunks = Chunk::new(input, 3, 1).unwrap().execute().unwrap();
assert_eq!(chunks.len(), 3);
assert_eq!(chunks[0].shape(), &[2, 2]);
}
#[tokio::test]
async fn test_chunk_invalid() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_vec_on(vec![1.0, 2.0, 3.0], vec![3], device.clone())
.await
.unwrap();
assert!(Chunk::new(input.clone(), 0, 0).is_err());
// Non-divisible chunks are now allowed (PyTorch-style)
let chunks = Chunk::new(input, 2, 0).unwrap().execute().unwrap();
assert_eq!(chunks.len(), 2);
assert_eq!(chunks[0].shape(), &[2]);
assert_eq!(chunks[1].shape(), &[1]);
}
}