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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Sparse Matrix-Vector Product (CSR format) - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its computation
//! - Zero hardcoding: Hardware-agnostic implementation
//! - 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::Result;
use crate::tensor::Tensor;
/// GPU shader for f64 sparse matrix-vector product (CSR format).
///
/// Includes `SpMV`, axpy, dot product, scale, copy, diagonal preconditioner,
/// linear combination, and full CG solver kernels — all in f64.
///
/// Entry points: `spmv_f64`, `axpy_f64`, `dot_f64`, `scale_f64`, `copy_f64`,
/// `precond_f64`, `linear_comb_f64`, `final_reduce_f64`, `cg_update_xr`,
/// `cg_update_p`, `compute_alpha`, `compute_beta`.
pub const WGSL_SPARSE_MATVEC_F64: &str = include_str!("../shaders/misc/sparse_matvec_f64.wgsl");
/// Sparse matrix-vector product in CSR (Compressed Sparse Row) format.
pub struct SparseMatVec {
values: Tensor,
col_indices: Vec<u32>,
row_ptrs: Vec<u32>,
vector: Tensor,
}
impl SparseMatVec {
/// Create sparse matvec from CSR format (values, `col_indices`, `row_ptrs`) and dense vector.
#[must_use]
pub fn new(values: Tensor, col_indices: Vec<u32>, row_ptrs: Vec<u32>, vector: Tensor) -> Self {
Self {
values,
col_indices,
row_ptrs,
vector,
}
}
fn wgsl_shader() -> &'static str {
include_str!("../shaders/misc/sparse_matvec.wgsl")
}
/// Execute sparse matrix-vector product.
/// # 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.values.device();
let num_rows = self.row_ptrs.len() - 1;
if num_rows == 0 {
return Ok(Tensor::new(vec![], vec![0], device.clone()));
}
let output_buffer = device.create_buffer_f32(num_rows)?;
// Create buffers for CSR data
let col_indices_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("SparseMatVec col_indices"),
contents: bytemuck::cast_slice(&self.col_indices),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let row_ptrs_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("SparseMatVec row_ptrs"),
contents: bytemuck::cast_slice(&self.row_ptrs),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
num_rows: u32,
}
let params = Params {
num_rows: num_rows as u32,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("SparseMatVec Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("SparseMatVec Bind Group Layout"),
entries: &[
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,
},
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: true },
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: false },
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::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("SparseMatVec Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.values.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: col_indices_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: row_ptrs_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: self.vector.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 5,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("SparseMatVec"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("SparseMatVec Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("SparseMatVec 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("SparseMatVec Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("SparseMatVec Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(device);
let workgroups = caps.dispatch_1d(num_rows as u32);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
vec![num_rows],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_sparse_matvec() {
let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
else {
return;
};
// 2x3 matrix: [[1, 0, 2], [0, 3, 0]]
// values: 1, 2, 3
// col_indices: 0, 2, 1
// row_ptrs: 0, 2, 3
let values = vec![1.0f32, 2.0, 3.0];
let col_indices = vec![0u32, 2, 1];
let row_ptrs = vec![0u32, 2, 3];
let vector = vec![1.0f32, 2.0, 3.0];
let values_tensor = Tensor::new(values, vec![3], device.clone());
let vector_tensor = Tensor::new(vector, vec![3], device);
let output = SparseMatVec::new(values_tensor, col_indices, row_ptrs, vector_tensor)
.execute()
.unwrap();
let result = output.to_vec().unwrap();
// Row 0: 1*1 + 2*3 = 7
// Row 1: 3*2 = 6
assert_eq!(result.len(), 2);
assert!((result[0] - 7.0).abs() < 1e-5);
assert!((result[1] - 6.0).abs() < 1e-5);
}
}