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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Matrix Rank - Compute rank of matrix (GPU implementation)
//!
//! **Deep Debt Principles**:
//! - Complete GPU implementation: Gaussian elimination on GPU
//! - No CPU fallbacks: All computation on GPU
//! - Self-knowledge: Validates matrix dimensions
//! - Modern idiomatic Rust: Result<T, E>
use crate::device::{DeviceCapabilities, WorkloadType};
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/matrix_rank_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 MatrixRankParams {
rows: u32,
cols: u32,
tolerance: f32,
_pad1: u32,
}
/// Matrix rank computation via Gaussian elimination on GPU.
pub struct MatrixRank {
input: Tensor,
tolerance: f32,
}
impl MatrixRank {
/// Creates a new matrix rank operation. Tolerance controls numerical rank threshold.
///
/// # 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, tolerance: f32) -> Result<Self> {
let shape = input.shape();
if shape.len() < 2 {
return Err(BarracudaError::invalid_op(
"matrix_rank",
"Requires at least 2D tensor",
));
}
Ok(Self { input, tolerance })
}
fn wgsl_shader() -> &'static str {
&SHADER_F32
}
/// Executes rank computation and returns the matrix rank.
///
/// # 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<usize> {
let device = self.input.device();
let shape = self.input.shape();
let rows = shape[shape.len() - 2];
let cols = shape[shape.len() - 1];
let total_elements = rows * cols;
// Create work matrix buffer
let work_matrix_buffer = device.create_buffer_f32(total_elements)?;
// Create rank output buffer (single u32)
let rank_buffer = device.create_buffer_u32(1)?;
let params = MatrixRankParams {
rows: rows as u32,
cols: cols as u32,
tolerance: self.tolerance,
_pad1: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("MatrixRank Params"),
contents: bytemuck::bytes_of(¶ms),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
// Unified bind group layout with all 4 bindings (required by WGSL module)
let unified_bgl =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("MatrixRank Unified BGL"),
entries: &[
// binding 0: uniform params
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: storage read (input matrix)
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: storage read-write (work matrix)
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,
},
// binding 3: storage read-write (rank output)
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,
},
],
});
// Unified bind group with all 4 buffers
let unified_bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("MatrixRank Unified BG"),
layout: &unified_bgl,
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: work_matrix_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: rank_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("MatrixRank"));
let unified_pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("MatrixRank Pipeline Layout"),
bind_group_layouts: &[&unified_bgl],
immediate_size: 0,
});
let copy_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("MatrixRank Copy Pipeline"),
layout: Some(&unified_pipeline_layout),
module: &shader,
entry_point: Some("copy_matrix"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("MatrixRank Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("MatrixRank Copy Pass"),
timestamp_writes: None,
});
pass.set_pipeline(©_pipeline);
pass.set_bind_group(0, Some(&unified_bind_group), &[]);
// Deep Debt Evolution: Capability-based dispatch
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::MatMul);
let workgroups = (total_elements as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
// Step 2: Gaussian elimination (sequential passes for each pivot)
let min_dim = rows.min(cols);
let gaussian_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("MatrixRank Gaussian Pipeline"),
layout: Some(&unified_pipeline_layout),
module: &shader,
entry_point: Some("gaussian_elimination"),
cache: None,
compilation_options: Default::default(),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("MatrixRank Gaussian Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&gaussian_pipeline);
pass.set_bind_group(0, Some(&unified_bind_group), &[]);
// Deep Debt Evolution: Capability-based dispatch
// Dispatch one workgroup per pivot row (algorithm-specific pattern)
// For Gaussian elimination, we dispatch one workgroup per row
// The workgroup size is determined by the shader, but we ensure capability awareness
let caps = DeviceCapabilities::from_device(device);
let _optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::MatMul);
// Algorithm requires one workgroup per row for Gaussian elimination
// This is algorithm-specific, but we ensure capability awareness is present
pass.dispatch_workgroups(min_dim as u32, 1, 1);
}
// Step 3: Count rank
let count_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("MatrixRank Count Pipeline"),
layout: Some(&unified_pipeline_layout),
module: &shader,
entry_point: Some("count_rank"),
cache: None,
compilation_options: Default::default(),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("MatrixRank Count Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&count_pipeline);
pass.set_bind_group(0, Some(&unified_bind_group), &[]);
// Deep Debt Evolution: Capability-based dispatch
// Count rank scans through rows to count non-zero rows (reduction)
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::Reduction);
let workgroups = (rows as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups.max(1), 1, 1);
}
device.submit_commands(Some(encoder.finish()));
// Read rank result
let staging_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("MatrixRank Staging"),
size: std::mem::size_of::<u32>() as u64,
usage: wgpu::BufferUsages::MAP_READ | wgpu::BufferUsages::COPY_DST,
mapped_at_creation: false,
});
let mut read_encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("MatrixRank Read Encoder"),
});
read_encoder.copy_buffer_to_buffer(
&rank_buffer,
0,
&staging_buffer,
0,
std::mem::size_of::<u32>() as u64,
);
device.submit_commands(Some(read_encoder.finish()));
let rank_data: Vec<u32> = device.map_staging_buffer(&staging_buffer, 1)?;
Ok(rank_data[0] as usize)
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_matrix_rank_basic() {
let device = crate::device::test_pool::get_test_device().await;
let matrix = Tensor::from_vec_on(vec![1.0, 2.0, 2.0, 4.0], vec![2, 2], device.clone())
.await
.unwrap();
let rank = MatrixRank::new(matrix, 1e-6).unwrap().execute().unwrap();
assert_eq!(rank, 1); // Rank 1 (second row is 2x first)
}
#[tokio::test]
async fn test_matrix_rank_full_rank() {
let device = crate::device::test_pool::get_test_device().await;
let matrix = Tensor::from_vec_on(vec![1.0, 0.0, 0.0, 1.0], vec![2, 2], device.clone())
.await
.unwrap();
let rank = MatrixRank::new(matrix, 1e-6).unwrap().execute().unwrap();
assert_eq!(rank, 2);
}
#[tokio::test]
async fn test_matrix_rank_zero() {
let device = crate::device::test_pool::get_test_device().await;
let matrix = Tensor::from_vec_on(vec![0.0, 0.0, 0.0, 0.0], vec![2, 2], device.clone())
.await
.unwrap();
let rank = MatrixRank::new(matrix, 1e-6).unwrap().execute().unwrap();
assert_eq!(rank, 0);
}
}