vyre-driver-cuda 0.7.2

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

use super::*;

#[test]
fn planner_emits_passes_in_row_child_group_order() {
    let view = synthetic_view(3, 2, 2);
    let plan = plan_cuda_egraph_kernel_work(
        view,
        CudaEGraphKernelLaunchConfig {
            threads_per_block: 4,
            max_blocks_per_launch: 2,
        },
    )
    .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - valid egraph kernel plan");

    assert_eq!(plan.waves.len(), 3);
    assert_eq!(plan.waves[0].pass, CudaEGraphKernelPass::RowScan);
    assert_eq!(plan.waves[0].item_count, 3);
    assert_eq!(plan.waves[0].blocks, 1);
    assert_eq!(plan.waves[1].pass, CudaEGraphKernelPass::ChildEdgeScan);
    assert_eq!(plan.waves[1].item_count, 2);
    assert_eq!(plan.waves[2].pass, CudaEGraphKernelPass::EclassGroupScan);
    assert_eq!(plan.waves[2].item_count, 2);
    assert_eq!(plan.total_items, 7);
    assert_eq!(plan.total_blocks, 3);
}

#[test]
fn planner_splits_large_passes_into_bounded_waves() {
    let view = synthetic_view(19, 0, 0);
    let plan = plan_cuda_egraph_kernel_work(
        view,
        CudaEGraphKernelLaunchConfig {
            threads_per_block: 4,
            max_blocks_per_launch: 2,
        },
    )
    .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - valid split egraph kernel plan");

    let items = plan
        .waves
        .iter()
        .map(|wave| (wave.first_item, wave.item_count, wave.blocks))
        .collect::<Vec<_>>();
    assert_eq!(
        items,
        vec![
            (0, 8, 2),
            (8, 8, 2),
            (16, 3, 1),
            (0, 8, 2),
            (8, 8, 2),
            (16, 3, 1),
        ]
    );
    assert_eq!(plan.total_items, 38);
    assert_eq!(plan.total_blocks, 10);
}

#[test]
fn planner_rejects_zero_launch_dimensions() {
    let view = synthetic_view(1, 0, 0);
    assert_eq!(
        plan_cuda_egraph_kernel_work(
            view,
            CudaEGraphKernelLaunchConfig {
                threads_per_block: 0,
                max_blocks_per_launch: 1,
            },
        )
        .expect_err("zero threads must be rejected"),
        CudaEGraphKernelPlanError::ZeroThreadsPerBlock
    );
    assert_eq!(
        plan_cuda_egraph_kernel_work(
            view,
            CudaEGraphKernelLaunchConfig {
                threads_per_block: 1,
                max_blocks_per_launch: 0,
            },
        )
        .expect_err("zero max blocks must be rejected"),
        CudaEGraphKernelPlanError::ZeroMaxBlocksPerLaunch
    );
}

#[test]
fn signature_bucket_planner_groups_only_candidate_duplicate_rows() {
    let snapshot = GpuEGraphSnapshot::build([
        (0u32, "lit", &[][..]),
        (1u32, "lit", &[][..]),
        (2u32, "add", &[0u32, 1u32][..]),
        (3u32, "add", &[0u32, 1u32][..]),
        (4u32, "add", &[1u32, 0u32][..]),
        (5u32, "mul", &[0u32, 1u32][..]),
    ]);
    let image = snapshot
        .try_pack_device_image()
        .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - valid egraph image must pack");
    let plan = plan_cuda_egraph_signature_buckets(
        &image,
        view_for_image(&image),
        CudaEGraphKernelLaunchConfig {
            threads_per_block: 8,
            max_blocks_per_launch: 1,
        },
    )
    .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - signature bucket plan must build");

    let grouped_rows = plan
        .buckets
        .iter()
        .map(|bucket| {
            let start = bucket.first_bucket_row as usize;
            let end = start + bucket.row_count as usize;
            plan.bucket_rows[start..end].to_vec()
        })
        .collect::<Vec<_>>();

    assert_eq!(grouped_rows.len(), 2);
    assert!(grouped_rows.contains(&vec![0, 1]));
    assert!(grouped_rows.contains(&vec![2, 3]));
    assert_eq!(plan.candidate_pair_count, 2);
    assert_eq!(plan.pair_waves.len(), 2);
    assert!(plan
        .pair_waves
        .iter()
        .all(|wave| wave.pair_count == 1 && wave.blocks == 1));
}

#[test]
fn structural_equivalence_planner_rejects_divergent_language_op_ids() {
    let snapshot = GpuEGraphSnapshot::build([(10u32, "lit", &[][..]), (20u32, "opaque", &[][..])]);
    let image = snapshot
        .try_pack_device_image()
        .expect("Fix: valid divergent-op egraph image must pack");

    assert_ne!(
        image.row_language_op_ids()[0],
        image.row_language_op_ids()[1]
    );

    let plan = plan_cuda_egraph_structural_equivalences(
        &image,
        view_for_image(&image),
        CudaEGraphKernelLaunchConfig {
            threads_per_block: 8,
            max_blocks_per_launch: 1,
        },
    )
    .expect("Fix: divergent-op egraph image must remain plannable");

    assert!(plan.signature_plan.buckets.is_empty());
    assert_eq!(plan.signature_plan.candidate_pair_count, 0);
    assert!(plan.equivalences.is_empty());
    assert_eq!(plan.exact_pair_count, 0);
    assert_eq!(plan.rejected_candidate_pair_count, 0);
}