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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Bray-Curtis Distance (f64) — GPU-Accelerated All-Pairs Distance Matrix
//!
//! ABSORBED from wetSpring (Feb 16, 2026) — general-purpose distance metric
//!
//! Computes the condensed Bray-Curtis distance matrix for N samples, each with
//! D features. Bray-Curtis dissimilarity is commonly used for non-negative
//! abundance/count data in ecology and metagenomics.
//!
//! # Formula
//!
//! ```text
//! BC(a, b) = Σ|aₖ - bₖ| / Σ(aₖ + bₖ)
//! ```
//!
//! # Output Format
//!
//! Condensed distance matrix: N*(N-1)/2 values in order:
//! (1,0), (2,0), (2,1), (3,0), (3,1), (3,2), ...
//!
//! # Performance
//!
//! - O(N² × D) work, embarrassingly parallel across N*(N-1)/2 pairs
//! - For N=1000, D=2000: 500K independent tasks × 2000 features each
//! - GPU threshold: N ≥ 32 (GPU overhead dominates for smaller inputs)
//!
//! # Example
//!
//! ```rust,ignore
//! use barracuda::ops::bray_curtis_f64::BrayCurtisF64;
//!
//! let bc = BrayCurtisF64::new(device.clone())?;
//!
//! // 4 samples, 3 features each
//! let samples = vec![
//! 1.0, 2.0, 3.0, // sample 0
//! 4.0, 5.0, 6.0, // sample 1
//! 1.0, 1.0, 1.0, // sample 2
//! 7.0, 8.0, 9.0, // sample 3
//! ];
//!
//! let distances = bc.condensed_distance_matrix(&samples, 4, 3)?;
//! // distances.len() = 4*(4-1)/2 = 6
//! // distances[0] = BC(sample1, sample0)
//! // distances[1] = BC(sample2, sample0)
//! // ...
//! ```
use crate::device::WgpuDevice;
use crate::device::capabilities::WORKGROUP_SIZE_1D;
use crate::error::{BarracudaError, Result};
use bytemuck::{Pod, Zeroable};
use std::sync::Arc;
/// Parameters for Bray-Curtis shader
#[repr(C)]
#[derive(Clone, Copy, Pod, Zeroable)]
struct BcParams {
n_samples: u32,
n_features: u32,
n_pairs: u32,
_pad: u32,
}
/// GPU-accelerated Bray-Curtis distance computation
pub struct BrayCurtisF64 {
device: Arc<WgpuDevice>,
pipeline: wgpu::ComputePipeline,
}
#[expect(
dead_code,
reason = "public API — exercised by tests, available to downstream consumers"
)]
impl BrayCurtisF64 {
/// Create a new Bray-Curtis GPU operator
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn new(device: Arc<WgpuDevice>) -> Result<Self> {
let shader_source = include_str!("../shaders/math/bray_curtis_f64.wgsl");
let shader_module = device.compile_shader_f64(shader_source, Some("BrayCurtisF64 Shader"));
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("BrayCurtisF64 Pipeline"),
layout: None,
module: &shader_module,
entry_point: Some("bray_curtis_pairs"),
cache: None,
compilation_options: Default::default(),
});
Ok(Self { device, pipeline })
}
/// Compute condensed Bray-Curtis distance matrix
///
/// # Arguments
/// * `samples` - Flattened sample matrix [`n_samples` × `n_features`], row-major
/// * `n_samples` - Number of samples (N)
/// * `n_features` - Number of features per sample (D)
///
/// # Returns
/// Condensed distance matrix [N*(N-1)/2]
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn condensed_distance_matrix(
&self,
samples: &[f64],
n_samples: usize,
n_features: usize,
) -> Result<Vec<f64>> {
let expected_len = n_samples * n_features;
if samples.len() < expected_len {
return Err(BarracudaError::InvalidInput {
message: format!(
"Expected {} elements ({} samples × {} features), got {}",
expected_len,
n_samples,
n_features,
samples.len()
),
});
}
if n_samples < 2 {
return Ok(Vec::new());
}
let n_pairs = n_samples * (n_samples - 1) / 2;
// Create input buffer
let input_buffer =
self.device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("BC Input"),
contents: bytemuck::cast_slice(samples),
usage: wgpu::BufferUsages::STORAGE,
});
// Create output buffer
let output_buffer = self.device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("BC Output"),
size: (n_pairs * 8) as u64, // f64 = 8 bytes
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
mapped_at_creation: false,
});
// Create params buffer
let params = BcParams {
n_samples: n_samples as u32,
n_features: n_features as u32,
n_pairs: n_pairs as u32,
_pad: 0,
};
let params_buffer =
self.device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("BC Params"),
contents: bytemuck::bytes_of(¶ms),
usage: wgpu::BufferUsages::UNIFORM,
});
// Create bind group
let bind_group_layout = self.pipeline.get_bind_group_layout(0);
let bind_group = self
.device
.device
.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("BC Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: input_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buffer.as_entire_binding(),
},
],
});
// Execute
let mut encoder = self
.device
.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("BC Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("BC Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&self.pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
// Dispatch: ceil(n_pairs / 256) workgroups
let n_workgroups = n_pairs.div_ceil(WORKGROUP_SIZE_1D as usize);
pass.dispatch_workgroups(n_workgroups as u32, 1, 1);
}
self.device.submit_commands(Some(encoder.finish()));
// Read results
self.read_results(&output_buffer, n_pairs)
}
/// Read results from GPU buffer
fn read_results(&self, buffer: &wgpu::Buffer, count: usize) -> Result<Vec<f64>> {
let staging = self.device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("BC Staging"),
size: (count * 8) as u64,
usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::MAP_READ,
mapped_at_creation: false,
});
let mut encoder = self
.device
.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("BC Copy Encoder"),
});
encoder.copy_buffer_to_buffer(buffer, 0, &staging, 0, (count * 8) as u64);
self.device.submit_commands(Some(encoder.finish()));
let results: Vec<f64> = self.device.map_staging_buffer(&staging, count)?;
Ok(results)
}
/// CPU fallback for small inputs
fn condensed_distance_matrix_cpu(
&self,
samples: &[f64],
n_samples: usize,
n_features: usize,
) -> Result<Vec<f64>> {
let n_pairs = n_samples * (n_samples - 1) / 2;
let mut distances = Vec::with_capacity(n_pairs);
for i in 1..n_samples {
for j in 0..i {
let bc = bray_curtis_cpu(samples, i, j, n_features);
distances.push(bc);
}
}
Ok(distances)
}
/// Convert condensed matrix index to (i, j) pair
///
/// Useful for interpreting results.
#[must_use]
pub fn condensed_index_to_pair(idx: usize) -> (usize, usize) {
// i = floor((1 + sqrt(1 + 8*idx)) / 2)
let k = idx as f64;
let i = f64::midpoint(1.0, 8.0f64.mul_add(k, 1.0).sqrt()).floor() as usize;
let j = idx - i * (i - 1) / 2;
(i, j)
}
/// Convert (i, j) pair to condensed matrix index
///
/// Requires i > j.
#[must_use]
pub fn pair_to_condensed_index(i: usize, j: usize) -> usize {
debug_assert!(i > j, "Requires i > j for condensed indexing");
i * (i - 1) / 2 + j
}
}
/// CPU reference: Bray-Curtis distance between samples i and j
fn bray_curtis_cpu(samples: &[f64], i: usize, j: usize, n_features: usize) -> f64 {
let base_i = i * n_features;
let base_j = j * n_features;
let mut sum_diff = 0.0;
let mut sum_sum = 0.0;
for k in 0..n_features {
let a = samples[base_i + k];
let b = samples[base_j + k];
sum_diff += (a - b).abs();
sum_sum += a + b;
}
if sum_sum > 0.0 {
sum_diff / sum_sum
} else {
0.0
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_bray_curtis_cpu_identical() {
// Identical samples → distance = 0
let samples = vec![1.0, 2.0, 3.0, 1.0, 2.0, 3.0];
let bc = bray_curtis_cpu(&samples, 1, 0, 3);
assert!((bc - 0.0).abs() < 1e-10);
}
#[test]
fn test_bray_curtis_cpu_disjoint() {
// Completely disjoint: [1,0,0] vs [0,1,0] → BC = 1.0
let samples = vec![1.0, 0.0, 0.0, 0.0, 1.0, 0.0];
let bc = bray_curtis_cpu(&samples, 1, 0, 3);
assert!((bc - 1.0).abs() < 1e-10);
}
#[test]
fn test_bray_curtis_cpu_known_value() {
// [1, 2, 3] vs [4, 5, 6]
// sum_diff = |1-4| + |2-5| + |3-6| = 9
// sum_sum = (1+4) + (2+5) + (3+6) = 21
// BC = 9/21 = 3/7 ≈ 0.4286
let samples = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0];
let bc = bray_curtis_cpu(&samples, 1, 0, 3);
let expected = 9.0 / 21.0;
assert!(
(bc - expected).abs() < 1e-10,
"Got {bc}, expected {expected}"
);
}
#[test]
fn test_condensed_index_conversion() {
// Test round-trip conversion
for i in 1..10 {
for j in 0..i {
let idx = BrayCurtisF64::pair_to_condensed_index(i, j);
let (i2, j2) = BrayCurtisF64::condensed_index_to_pair(idx);
assert_eq!((i, j), (i2, j2), "Failed for ({i}, {j})");
}
}
}
#[test]
fn test_condensed_matrix_order() {
// Verify condensed order: (1,0), (2,0), (2,1), (3,0), ...
assert_eq!(BrayCurtisF64::pair_to_condensed_index(1, 0), 0);
assert_eq!(BrayCurtisF64::pair_to_condensed_index(2, 0), 1);
assert_eq!(BrayCurtisF64::pair_to_condensed_index(2, 1), 2);
assert_eq!(BrayCurtisF64::pair_to_condensed_index(3, 0), 3);
assert_eq!(BrayCurtisF64::pair_to_condensed_index(3, 1), 4);
assert_eq!(BrayCurtisF64::pair_to_condensed_index(3, 2), 5);
}
}