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
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Gather ND - N-dimensional gather operation - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its multi-dimensional 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;
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
/// Gather ND operation - N-dimensional gather
pub struct GatherNd {
input: Tensor,
indices: Tensor, // Shape: [batch_size, num_indices, index_rank]
}
impl GatherNd {
/// Create a new gather ND operation
/// # Errors
/// Returns [`Err`] if indices have fewer than 2 dimensions, or if index rank exceeds input rank.
pub fn new(input: Tensor, indices: Tensor) -> Result<Self> {
let input_shape = input.shape();
let indices_shape = indices.shape();
// Validate indices shape: should be [batch_size, num_indices, index_rank]
// where index_rank <= input_shape.len()
if indices_shape.len() < 2 {
return Err(BarracudaError::invalid_op(
"GatherNd",
format!(
"Indices must have at least 2 dimensions, got {}",
indices_shape.len()
),
));
}
let index_rank = indices_shape[indices_shape.len() - 1];
if index_rank > input_shape.len() {
return Err(BarracudaError::invalid_op(
"GatherNd",
format!(
"Index rank {} exceeds input rank {}",
index_rank,
input_shape.len()
),
));
}
Ok(Self { input, 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/gather_nd_f64.wgsl").to_string()
});
&S
}
}
/// Execute the gather ND operation
/// # Errors
/// Returns [`Err`] if buffer allocation fails, GPU dispatch fails, buffer readback fails
/// (e.g. device lost), or indices reference out-of-bounds elements.
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let input_shape = self.input.shape();
let indices_shape = self.indices.shape();
let input_size: usize = input_shape.iter().product();
let indices_size: usize = indices_shape.iter().product();
// Calculate output shape
// Output shape: [batch_size, num_indices] + input_shape[index_rank..]
let index_rank = indices_shape[indices_shape.len() - 1];
let num_indices = if indices_shape.len() > 1 {
indices_shape[indices_shape.len() - 2]
} else {
1
};
let batch_size = if indices_shape.len() > 2 {
indices_shape[..indices_shape.len() - 2].iter().product()
} else {
1
};
let mut output_shape = vec![batch_size, num_indices];
if index_rank < input_shape.len() {
output_shape.extend_from_slice(&input_shape[index_rank..]);
}
let output_size = output_shape.iter().product::<usize>();
// Access buffers directly (zero-copy)
let input_buffer = self.input.buffer();
let indices_buffer = self.indices.buffer();
// Create output buffer
let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("GatherNd Output"),
size: (output_size * std::mem::size_of::<f32>()) as u64,
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
mapped_at_creation: false,
});
// Create buffers for shape information
let input_shape_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GatherNd Input Shape"),
contents: bytemuck::cast_slice(
&input_shape.iter().map(|&x| x as u32).collect::<Vec<_>>(),
),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let indices_shape_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GatherNd Indices Shape"),
contents: bytemuck::cast_slice(
&indices_shape.iter().map(|&x| x as u32).collect::<Vec<_>>(),
),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let output_shape_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GatherNd Output Shape"),
contents: bytemuck::cast_slice(
&output_shape.iter().map(|&x| x as u32).collect::<Vec<_>>(),
),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
// Create uniform buffer for parameters
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
input_size: u32,
indices_size: u32,
output_size: u32,
input_rank: u32,
indices_rank: u32,
index_rank: u32,
batch_size: u32,
num_indices: u32,
}
let params = Params {
input_size: input_size as u32,
indices_size: indices_size as u32,
output_size: output_size as u32,
input_rank: input_shape.len() as u32,
indices_rank: indices_shape.len() as u32,
index_rank: index_rank as u32,
batch_size: batch_size as u32,
num_indices: num_indices as u32,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GatherNd 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("GatherNd 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: 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::Storage { read_only: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 5,
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: 6,
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,
},
],
});
// Create bind group
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("GatherNd 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(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: input_shape_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 5,
resource: indices_shape_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 6,
resource: output_shape_buffer.as_entire_binding(),
},
],
});
// Compile shader
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("GatherNd Shader"));
// Create compute pipeline
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("GatherNd Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GatherNd 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("GatherNd Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GatherNd Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&compute_pipeline);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
// Dispatch using standard 1D shader workgroup size (256)
let caps = DeviceCapabilities::from_device(device);
let workgroups = caps.dispatch_1d(output_size as u32);
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, output_size)?;
Ok(Tensor::new(output_data, output_shape, device.clone()))
}
}
impl Tensor {
/// Gather elements using N-dimensional indices
/// # Arguments
/// * `indices` - N-dimensional indices tensor
/// # Errors
/// Returns [`Err`] if indices have fewer than 2 dimensions, index rank exceeds input rank,
/// buffer allocation fails, GPU dispatch fails, or buffer readback fails (e.g. device lost).
pub fn gather_nd(self, indices: Self) -> Result<Self> {
GatherNd::new(self, indices)?.execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_gather_nd_2d() {
let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
else {
return;
};
// Input: 3x3 matrix
let input = Tensor::new(
vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0],
vec![3, 3],
device.clone(),
);
// Indices: [[0,0], [2,1]]
let indices_data: Vec<f32> = vec![0.0, 0.0, 2.0, 1.0];
let indices = Tensor::new(indices_data, vec![2, 2], device);
let result = input.gather_nd(indices).unwrap();
let output_data = result.to_vec().unwrap();
// Expected: [1.0, 8.0] (input[0,0] and input[2,1])
assert_eq!(output_data[0], 1.0);
assert_eq!(output_data[1], 8.0);
}
}