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
//! Index Select - Gather elements along a dimension - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its own dimension and indices
//! - 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;
/// Index Select operation - Gather elements along a dimension
///
/// Selects values from input tensor along a dimension using provided indices.
pub struct IndexSelect {
input: Tensor,
dim: usize,
indices: Vec<usize>,
}
impl IndexSelect {
/// Create a new `IndexSelect` operation
#[must_use]
pub fn new(input: Tensor, dim: usize, indices: Vec<usize>) -> Self {
Self {
input,
dim,
indices,
}
}
/// Get the WGSL shader source
fn wgsl_shader() -> &'static str {
{
static S: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/tensor/index_select_f64.wgsl").to_string()
});
&S
}
}
/// Execute the index select 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 dimension
if self.dim >= shape.len() {
let mut expected_shape = shape.to_vec();
expected_shape.push(self.dim); // Add dim as a hint
return Err(BarracudaError::invalid_shape(
expected_shape,
shape.to_vec(),
));
}
// Calculate output shape
let mut output_shape = shape.to_vec();
output_shape[self.dim] = self.indices.len();
// Calculate strides
let dim_size = shape[self.dim];
let outer_size: usize = shape[..self.dim].iter().product();
let inner_size: usize = shape[self.dim + 1..].iter().product();
let total_size = outer_size * self.indices.len() * inner_size;
// Convert indices to u32
let indices_u32: Vec<u32> = self.indices.iter().map(|&i| i as u32).collect();
// Create buffers
// Access input buffer directly (zero-copy)
let input_buffer = self.input.buffer();
let indices_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("IndexSelect Indices"),
contents: bytemuck::cast_slice(&indices_u32),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("IndexSelect Output"),
size: (total_size * 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 {
total_size: u32,
dim_size: u32,
outer_size: u32,
inner_size: u32,
num_indices: u32,
_pad0: u32,
_pad1: u32,
_pad2: u32,
}
let params = Params {
total_size: total_size as u32,
dim_size: dim_size as u32,
outer_size: outer_size as u32,
inner_size: inner_size as u32,
num_indices: self.indices.len() as u32,
_pad0: 0,
_pad1: 0,
_pad2: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("IndexSelect 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("IndexSelect Bind Group Layout"),
entries: &[
// binding 0: uniform params (matches WGSL)
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,
},
// binding 1: input storage (read)
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,
},
// binding 2: indices storage (read)
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,
},
// binding 3: output storage (read-write)
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,
},
],
});
// Create bind group
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("IndexSelect Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: input_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: indices_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: output_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("IndexSelect Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("IndexSelect 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("IndexSelect Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("IndexSelect 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::ElementWise);
let workgroups = (total_size 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_size)?;
Ok(Tensor::new(output_data, output_shape, device.clone()))
}
}
impl Tensor {
/// Select elements along a dimension using indices
///
/// # Arguments
///
/// * `dim` - Dimension to select from
/// * `indices` - Indices to select
///
/// # Returns
///
/// Tensor with selected elements
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn index_select_wgsl(self, dim: usize, indices: Vec<usize>) -> Result<Self> {
IndexSelect::new(self, dim, indices).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_index_select_1d() {
let device = crate::device::test_pool::get_test_device().await;
// Input: [5] = [0, 1, 2, 3, 4]
let data = vec![0.0, 1.0, 2.0, 3.0, 4.0];
let input = Tensor::new(data, vec![5], device);
let output = input.index_select_wgsl(0, vec![1, 3, 4]).unwrap();
assert_eq!(output.shape(), &[3]);
let result = output.to_vec().unwrap();
assert_eq!(result[0], 1.0);
assert_eq!(result[1], 3.0);
assert_eq!(result[2], 4.0);
}
#[tokio::test]
async fn test_index_select_2d_rows() {
let device = crate::device::test_pool::get_test_device().await;
// Input: [3, 2] = [[0,1], [2,3], [4,5]]
let data = vec![0.0, 1.0, 2.0, 3.0, 4.0, 5.0];
let input = Tensor::new(data, vec![3, 2], device);
let output = input.index_select_wgsl(0, vec![2, 0]).unwrap();
assert_eq!(output.shape(), &[2, 2]);
let result = output.to_vec().unwrap();
// Row 2: [4, 5], Row 0: [0, 1]
assert_eq!(result[0], 4.0);
assert_eq!(result[1], 5.0);
assert_eq!(result[2], 0.0);
assert_eq!(result[3], 1.0);
}
#[tokio::test]
async fn test_index_select_2d_cols() {
let device = crate::device::test_pool::get_test_device().await;
// Input: [2, 3] = [[0,1,2], [3,4,5]]
let data = vec![0.0, 1.0, 2.0, 3.0, 4.0, 5.0];
let input = Tensor::new(data, vec![2, 3], device);
let output = input.index_select_wgsl(1, vec![2, 0]).unwrap();
assert_eq!(output.shape(), &[2, 2]);
let result = output.to_vec().unwrap();
// Row 0: [2, 0], Row 1: [5, 3]
assert_eq!(result[0], 2.0);
assert_eq!(result[1], 0.0);
assert_eq!(result[2], 5.0);
assert_eq!(result[3], 3.0);
}
}