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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Tril - Complete triangular (GPU implementation)
//!
//! **Deep Debt Principles**:
//! - Complete GPU implementation: OVERWRITE existing CPU version
//! - No CPU fallbacks: All computation on GPU
//! - Self-knowledge: Validates matrix dimensions
//! - Modern idiomatic Rust: Result<T, E>
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
/// f64 is the canonical source — math is universal, precision is silicon.
const SHADER_F64: &str = include_str!("../shaders/linalg/tril_f64.wgsl");
/// f32 variant derived from f64 via precision downcast.
static SHADER_F32: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| SHADER_F64.to_string());
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct TrilParams {
rows: u32,
cols: u32,
diagonal: i32,
_pad1: u32,
}
/// Lower triangular matrix extraction operation.
pub struct Tril {
input: Tensor,
diagonal: i32,
}
impl Tril {
/// Creates a new tril operation. `diagonal` selects which diagonal to include (0 = main).
/// # 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, diagonal: i32) -> Result<Self> {
let shape = input.shape();
if shape.len() < 2 {
return Err(BarracudaError::invalid_op(
"tril",
"Requires at least 2D tensor",
));
}
Ok(Self { input, diagonal })
}
fn wgsl_shader() -> &'static str {
&SHADER_F32
}
/// Executes the tril operation and returns the lower triangular result.
/// # 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();
let rows = shape[shape.len() - 2];
let cols = shape[shape.len() - 1];
let matrix_size = rows * cols;
let output_buffer = device.create_buffer_f32(matrix_size)?;
let params = TrilParams {
rows: rows as u32,
cols: cols as u32,
diagonal: self.diagonal,
_pad1: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Tril Params"),
contents: bytemuck::bytes_of(¶ms),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Tril 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("Tril 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("Tril"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Tril Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Tril 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("Tril Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Tril Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let workgroups_x = (cols as u32).div_ceil(16);
let workgroups_y = (rows as u32).div_ceil(16);
pass.dispatch_workgroups(workgroups_x, workgroups_y, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
shape.to_vec(),
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_tril_basic() {
let device = crate::device::test_pool::get_test_device().await;
let matrix = Tensor::from_vec_on(
vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0],
vec![3, 3],
device.clone(),
)
.await
.unwrap();
let result = Tril::new(matrix, 0).unwrap().execute().unwrap();
let output = result.to_vec().unwrap();
assert_eq!(output[0], 1.0);
assert_eq!(output[1], 0.0); // Above diagonal
assert_eq!(output[3], 4.0);
}
}