vyre-driver-cuda 0.6.1

CUDA/PTX backend for vyre through the CUDA driver API.
Documentation 
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

//! Parity tests for vyre-primitives reduce::{gather, scatter, histogram,
//! radix_sort, segment_reduce_sum}.

#![cfg(test)]

mod common;

use common::{bytes_u32, u32_bytes, with_live_backend};
use vyre::DispatchConfig;
use vyre_primitives::reduce::gather::{cpu_ref as gather_cpu, gather};
use vyre_primitives::reduce::histogram::{cpu_ref as hist_cpu, histogram};
use vyre_primitives::reduce::radix_sort::{cpu_ref as radix_cpu, radix_sort};
use vyre_primitives::reduce::scatter::{cpu_ref as scatter_cpu, scatter};
use vyre_primitives::reduce::segment_reduce::{cpu_ref as seg_cpu, segment_reduce_sum};

// ---------------------------------------------------------------------
// gather
// ---------------------------------------------------------------------

fn run_gather(src: &[u32], indices: &[u32]) -> Vec<u32> {
    let count = src.len() as u32;
    let program = gather("src", "indices", "dst", count);
    let inputs: Vec<Vec<u8>> = vec![
        u32_bytes(src),
        u32_bytes(indices),
        vec![0u8; count as usize * 4],
    ];
    let mut config = DispatchConfig::default();
    let workgroup_x = 256u32;
    let grid_x = ((count + workgroup_x - 1) / workgroup_x).max(1);
    config.grid_override = Some([grid_x, 1, 1]);
    let outputs = with_live_backend("gather primitive", |backend| {
        backend
            .dispatch(&program, &inputs, &config)
            .unwrap_or_else(|error| panic!("Fix: CUDA gather dispatch failed: {error}"))
    });
    let mut out = bytes_u32(&outputs[0]);
    out.truncate(count as usize);
    out
}

#[test]
fn cuda_gather_identity() {
    let src = vec![10u32, 20, 30, 40];
    let indices = vec![0u32, 1, 2, 3];
    let cpu = gather_cpu(&src, &indices);
    let gpu = run_gather(&src, &indices);
    assert_eq!(gpu, cpu);
    assert_eq!(gpu, src);
}

#[test]
fn cuda_gather_reverse() {
    let src = vec![10u32, 20, 30, 40];
    let indices = vec![3u32, 2, 1, 0];
    let cpu = gather_cpu(&src, &indices);
    let gpu = run_gather(&src, &indices);
    assert_eq!(gpu, cpu);
    assert_eq!(gpu, vec![40, 30, 20, 10]);
}

// ---------------------------------------------------------------------
// scatter
// ---------------------------------------------------------------------

fn run_scatter(src: &[u32], indices: &[u32]) -> Vec<u32> {
    let count = src.len() as u32;
    let program = scatter("src", "indices", "dst", count);
    let inputs: Vec<Vec<u8>> = vec![
        u32_bytes(src),
        u32_bytes(indices),
        vec![0u8; count as usize * 4],
    ];
    let mut config = DispatchConfig::default();
    let workgroup_x = 256u32;
    let grid_x = ((count + workgroup_x - 1) / workgroup_x).max(1);
    config.grid_override = Some([grid_x, 1, 1]);
    let outputs = with_live_backend("scatter primitive", |backend| {
        backend
            .dispatch(&program, &inputs, &config)
            .unwrap_or_else(|error| panic!("Fix: CUDA scatter dispatch failed: {error}"))
    });
    let mut out = bytes_u32(&outputs[0]);
    out.truncate(count as usize);
    out
}

#[test]
fn cuda_scatter_inverse_of_gather() {
    let src = vec![10u32, 20, 30, 40];
    let indices = vec![3u32, 2, 1, 0];
    let cpu = scatter_cpu(&src, &indices, src.len());
    let gpu = run_scatter(&src, &indices);
    assert_eq!(gpu, cpu);
    // Each src[i] is written to dst[indices[i]] → reversed source.
    assert_eq!(gpu, vec![40, 30, 20, 10]);
}

#[test]
fn cuda_scatter_identity() {
    let src = vec![5u32, 6, 7, 8];
    let indices = vec![0u32, 1, 2, 3];
    let cpu = scatter_cpu(&src, &indices, src.len());
    let gpu = run_scatter(&src, &indices);
    assert_eq!(gpu, cpu);
    assert_eq!(gpu, src);
}

// ---------------------------------------------------------------------
// histogram
// ---------------------------------------------------------------------

fn run_histogram(input: &[u32], num_bins: u32) -> Vec<u32> {
    let count = input.len() as u32;
    let program = histogram("input", "output", count, num_bins);
    let inputs: Vec<Vec<u8>> = vec![u32_bytes(input), vec![0u8; num_bins as usize * 4]];
    let mut config = DispatchConfig::default();
    let workgroup_x = 256u32;
    let grid_x = ((count + workgroup_x - 1) / workgroup_x).max(1);
    config.grid_override = Some([grid_x, 1, 1]);
    let outputs = with_live_backend("histogram primitive", |backend| {
        backend
            .dispatch(&program, &inputs, &config)
            .unwrap_or_else(|error| panic!("Fix: CUDA histogram dispatch failed: {error}"))
    });
    let mut out = bytes_u32(&outputs[0]);
    out.truncate(num_bins as usize);
    out
}

#[test]
fn cuda_histogram_simple() {
    let input = vec![0u32, 1, 0, 2, 1, 0];
    let num_bins = 4u32;
    let cpu = hist_cpu(&input, num_bins);
    let gpu = run_histogram(&input, num_bins);
    assert_eq!(gpu, cpu);
    assert_eq!(gpu, vec![3, 2, 1, 0]);
}

#[test]
fn cuda_histogram_skips_out_of_range() {
    // bin index 5 exceeds num_bins=4 → cpu_ref skips it.
    let input = vec![0u32, 5, 1, 5, 2];
    let num_bins = 4u32;
    let cpu = hist_cpu(&input, num_bins);
    let gpu = run_histogram(&input, num_bins);
    assert_eq!(gpu, cpu);
    assert_eq!(gpu, vec![1, 1, 1, 0]);
}

// ---------------------------------------------------------------------
// radix_sort
// ---------------------------------------------------------------------

fn run_radix_sort(input: &[u32], bits: u32) -> Vec<u32> {
    let count = input.len() as u32;
    let program = radix_sort("input", "output", count, bits);
    let inputs: Vec<Vec<u8>> = vec![u32_bytes(input), vec![0u8; count as usize * 4]];
    let mut config = DispatchConfig::default();
    let workgroup_x = 256u32;
    let grid_x = ((count + workgroup_x - 1) / workgroup_x).max(1);
    config.grid_override = Some([grid_x, 1, 1]);
    let outputs = with_live_backend("radix sort primitive", |backend| {
        backend
            .dispatch(&program, &inputs, &config)
            .unwrap_or_else(|error| panic!("Fix: CUDA radix-sort dispatch failed: {error}"))
    });
    let mut out = bytes_u32(&outputs[0]);
    out.truncate(count as usize);
    out
}

#[test]
fn cuda_radix_sort_already_sorted() {
    let v = vec![1u32, 2, 3, 4, 5];
    let cpu = radix_cpu(&v, 8);
    let gpu = run_radix_sort(&v, 8);
    assert_eq!(gpu, cpu);
    assert_eq!(gpu, v);
}

#[test]
fn cuda_radix_sort_reverse() {
    let v = vec![5u32, 4, 3, 2, 1];
    let cpu = radix_cpu(&v, 8);
    let gpu = run_radix_sort(&v, 8);
    assert_eq!(gpu, cpu);
    assert_eq!(gpu, vec![1, 2, 3, 4, 5]);
}

#[test]
fn cuda_radix_sort_with_duplicates() {
    let v = vec![3u32, 1, 4, 1, 5, 9, 2, 6, 5, 3];
    let cpu = radix_cpu(&v, 8);
    let gpu = run_radix_sort(&v, 8);
    assert_eq!(gpu, cpu);
    let mut expected = v.clone();
    expected.sort_unstable();
    assert_eq!(gpu, expected);
}

// ---------------------------------------------------------------------
// segment_reduce_sum
// ---------------------------------------------------------------------

fn run_segment_reduce(input: &[u32], segment_offsets: &[u32]) -> Vec<u32> {
    let num_segments = (segment_offsets.len() - 1) as u32;
    let program = segment_reduce_sum("input", "segments", "output", num_segments);
    let inputs: Vec<Vec<u8>> = vec![
        u32_bytes(input),
        u32_bytes(segment_offsets),
        vec![0u8; num_segments as usize * 4],
    ];
    let mut config = DispatchConfig::default();
    let workgroup_x = 256u32;
    let grid_x = ((num_segments + workgroup_x - 1) / workgroup_x).max(1);
    config.grid_override = Some([grid_x, 1, 1]);
    let outputs = with_live_backend("segment reduce primitive", |backend| {
        backend
            .dispatch(&program, &inputs, &config)
            .unwrap_or_else(|error| panic!("Fix: CUDA segment-reduce dispatch failed: {error}"))
    });
    let mut out = bytes_u32(&outputs[0]);
    out.truncate(num_segments as usize);
    out
}

#[test]
fn cuda_segment_reduce_uniform_segments() {
    let input = vec![1u32, 2, 3, 4, 5, 6];
    // Segments [0..2), [2..4), [4..6).
    let segments = vec![0u32, 2, 4, 6];
    let cpu = seg_cpu(&input, &segments);
    let gpu = run_segment_reduce(&input, &segments);
    assert_eq!(gpu, cpu);
    assert_eq!(gpu, vec![3, 7, 11]);
}

#[test]
fn cuda_segment_reduce_uneven_segments() {
    let input = vec![10u32, 20, 30, 40, 50];
    let segments = vec![0u32, 1, 4, 5];
    let cpu = seg_cpu(&input, &segments);
    let gpu = run_segment_reduce(&input, &segments);
    assert_eq!(gpu, cpu);
    assert_eq!(gpu, vec![10, 90, 50]);
}