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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! f32 triangular solve (Tensor-based) and Tensor extension
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
/// Triangular solve operation
///
/// Solves L·x = b (forward) or Uᵀ·x = b (backward)
pub struct TriangularSolve {
matrix: Tensor, // Triangular matrix (L or U)
rhs: Tensor, // Right-hand side vector b
lower: bool, // true for lower triangular (forward), false for upper (backward)
}
impl TriangularSolve {
/// Create new triangular solve operation
/// # Arguments
/// * `matrix` - Triangular matrix [N, N]
/// * `rhs` - Right-hand side vector [N]
/// * `lower` - true for lower triangular (forward substitution)
/// # Deep Debt Compliance
/// - No hardcoded sizes (runtime N)
/// - No unsafe blocks
/// - Agnostic design (works with any triangular system)
#[must_use]
pub fn new(matrix: Tensor, rhs: Tensor, lower: bool) -> Self {
Self { matrix, rhs, lower }
}
/// Create forward substitution: L·x = b
#[must_use]
pub fn forward(matrix: Tensor, rhs: Tensor) -> Self {
Self::new(matrix, rhs, true)
}
/// Create backward substitution: Uᵀ·x = b
#[must_use]
pub fn backward(matrix: Tensor, rhs: Tensor) -> Self {
Self::new(matrix, rhs, false)
}
fn wgsl_shader() -> &'static str {
include_str!("../../../shaders/linalg/triangular_solve.wgsl")
}
/// Shared with f64 module for transpose solve
pub(crate) fn wgsl_shader_f64() -> &'static str {
include_str!("../../../shaders/linalg/triangular_solve_f64.wgsl")
}
/// Execute triangular solve on GPU
/// # Returns
/// Solution vector x
/// # Errors
/// Returns [`Err`] if matrix is not square, rhs size does not match matrix dimension,
/// buffer allocation fails, GPU dispatch fails, or buffer readback fails (e.g. device lost).
/// # Deep Debt Compliance
/// - Pure WGSL execution (no CPU fallback)
/// - Capability-based workgroup dispatch
/// - Safe buffer management
pub fn execute(self) -> Result<Tensor> {
let device = self.matrix.device();
let matrix_shape = self.matrix.shape();
let rhs_shape = self.rhs.shape();
// Validate square matrix
if matrix_shape.len() != 2 || matrix_shape[0] != matrix_shape[1] {
return Err(BarracudaError::InvalidShape {
expected: vec![0, 0],
actual: matrix_shape.to_vec(),
});
}
let n = matrix_shape[0];
// Validate rhs is a vector of length n
if rhs_shape.len() != 1 || rhs_shape[0] != n {
return Err(BarracudaError::InvalidShape {
expected: vec![n],
actual: rhs_shape.to_vec(),
});
}
// Create output buffer for solution vector x
let solution_buffer = device.create_buffer_f32(n)?;
// Create params buffer with matrix size and substitution type
let is_lower = u32::from(self.lower);
let params_buffer =
device.create_uniform_buffer("TriangularSolve Params", &[n as u32, is_lower]);
// Create bind group layout
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("TriangularSolve BGL"),
entries: &[
// Triangular matrix (L or U)
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,
},
// Right-hand side vector b
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,
},
// Solution vector x
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,
},
// Parameters (n, is_lower)
wgpu::BindGroupLayoutEntry {
binding: 3,
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("TriangularSolve BG"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.matrix.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.rhs.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: solution_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: params_buffer.as_entire_binding(),
},
],
});
// Compile shader
let shader = device.compile_shader(Self::wgsl_shader(), Some("TriangularSolve"));
// Create pipeline
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("TriangularSolve PL"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("TriangularSolve Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
// Create command encoder
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("TriangularSolve Encoder"),
});
// Execute compute pass
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("TriangularSolve Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
// Deep Debt Evolution: Sequential algorithm (dependency chain)
// Single-threaded execution (workgroup_size=1 in shader)
pass.dispatch_workgroups(1, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
let output_data = crate::utils::read_buffer(device, &solution_buffer, n)?;
Ok(Tensor::new(output_data, vec![n], device.clone()))
}
}
/// Tensor extension for triangular solve
impl Tensor {
/// Solve L·x = b (forward substitution)
/// # Arguments
/// * `rhs` - Right-hand side vector b
/// # Returns
/// Solution vector x
/// # Example
/// ```ignore
/// let l = tensor.cholesky()?; // Get lower triangular L
/// let x = l.solve_triangular_forward(&b)?;
/// ```
/// # Errors
/// Returns [`Err`] if matrix is not square, rhs size does not match, buffer allocation fails,
/// GPU dispatch fails, or buffer readback fails (e.g. device lost).
pub fn solve_triangular_forward(&self, rhs: &Self) -> Result<Self> {
TriangularSolve::forward(self.clone(), rhs.clone()).execute()
}
/// Solve Uᵀ·x = b (backward substitution)
/// # Arguments
/// * `rhs` - Right-hand side vector b
/// # Returns
/// Solution vector x
/// # Errors
/// Returns [`Err`] if matrix is not square, rhs size does not match, buffer allocation fails,
/// GPU dispatch fails, or buffer readback fails (e.g. device lost).
pub fn solve_triangular_backward(&self, rhs: &Self) -> Result<Self> {
TriangularSolve::backward(self.clone(), rhs.clone()).execute()
}
/// Solve triangular system L·x = b or Uᵀ·x = b
/// # Arguments
/// * `rhs` - Right-hand side vector b
/// * `lower` - true for lower triangular, false for upper
/// # Errors
/// Returns [`Err`] if matrix is not square, rhs size does not match, buffer allocation fails,
/// GPU dispatch fails, or buffer readback fails (e.g. device lost).
pub fn solve_triangular(&self, rhs: &Self, lower: bool) -> Result<Self> {
TriangularSolve::new(self.clone(), rhs.clone(), lower).execute()
}
}