tenferro-gpu 0.1.0

CubeCL-backed CUDA and WebGPU provider backends for tenferro tensors.
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

use std::{fs, path::Path};

#[test]
fn logical_kernels_do_not_take_tensor_shapes_as_comptime_parameters() {
    let root = Path::new(env!("CARGO_MANIFEST_DIR"))
        .join("src")
        .join("kernels");
    let files = [
        root.join("diagonal.rs"),
        root.join("helpers.rs"),
        root.join("indexing.rs"),
        root.join("structural.rs"),
    ];
    let banned = [
        "#[comptime] batch_shape",
        "#[comptime] input_shape",
        "#[comptime] output_shape",
        "#[comptime] operand_shape",
        "#[comptime] updates_shape",
        "#[comptime] scatter_indices_shape",
        "#[comptime] start_indices_shape",
        "#[comptime] window_shape_updates",
        "#[comptime] shape: Sequence<usize>",
    ];

    let mut violations = Vec::new();
    for file in files {
        let source = fs::read_to_string(&file).expect("kernel source should be readable");
        for needle in banned {
            if source.contains(needle) {
                violations.push(format!("{} contains {needle}", file.display()));
            }
        }
    }

    assert!(
        violations.is_empty(),
        "logical CubeCL kernels must receive tensor shape/stride through TensorBinding, not comptime shape parameters:\n{}",
        violations.join("\n")
    );
}

#[test]
fn reduction_kernels_do_not_hide_unbounded_axis_work_in_one_worker() {
    let root = Path::new(env!("CARGO_MANIFEST_DIR"))
        .join("src")
        .join("kernels");
    let launch_source = fs::read_to_string(root.join("reduce").join("launch.rs"))
        .expect("reduction launch source should be readable");
    let kernels_source = fs::read_to_string(root.join("reduce").join("kernels.rs"))
        .expect("reduction kernel source should be readable");
    let checks = [
        (
            "reduce/launch.rs",
            launch_source.as_str(),
            "ReduceStrategy::Auto | ReduceStrategy::Unit",
        ),
        (
            "reduce/kernels.rs",
            kernels_source.as_str(),
            "for reduce_index in 1..reduce_len",
        ),
    ];

    let mut violations = Vec::new();
    for (name, source, needle) in checks {
        if source.contains(needle) {
            violations.push(format!("{name} contains {needle}"));
        }
    }

    assert!(
        violations.is_empty(),
        "CubeCL reductions must not route Auto to a per-output worker with unbounded serial axis work; use a parallel reduction strategy or an explicitly bounded fallback:\n{}",
        violations.join("\n")
    );
}

#[test]
fn scatter_kernels_are_not_single_thread_fallbacks() {
    let indexing_source = fs::read_to_string(
        Path::new(env!("CARGO_MANIFEST_DIR"))
            .join("src")
            .join("kernels")
            .join("indexing.rs"),
    )
    .expect("indexing kernel source should be readable");
    let banned = ["ABSOLUTE_POS == 0", "for pos in 0..out.len()"];

    let mut violations = Vec::new();
    for needle in banned {
        if indexing_source.contains(needle) {
            violations.push(format!("indexing.rs contains {needle}"));
        }
    }

    assert!(
        violations.is_empty(),
        "scatter CubeCL kernels must cover the output or update domain in parallel:\n{}",
        violations.join("\n")
    );
}