vyre 0.4.0

GPU compute intermediate representation with a standard operation library
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

// Portable GPU kernel for string_matching.kmp_find.
// This parallel verifier preserves KMP's first-offset contract.
@group(0) @binding(1) var<storage, read> haystack_words: array<u32>;
@group(0) @binding(2) var<storage, read> needle_words: array<u32>;
@group(0) @binding(3) var<storage, read_write> result: atomic<u32>;
fn packed_byte(words: ptr<storage, array<u32>, read>, index: u32) -> u32 {
    return ((*words)[index >> 2u] >> ((index & 3u) << 3u)) & 0xffu;
}
fn matches_at(start: u32) -> bool {
    var i = 0u;
    loop {
        if (i >= params.len_b) { break; }
        if (packed_byte(&haystack_words, start + i) != packed_byte(&needle_words, i)) { return false; }
        i = i + 1u;
    }
    return true;
}
@compute @workgroup_size(256, 1, 1)
fn string_matching_kmp_find(@builtin(global_invocation_id) id: vec3<u32>) {
    if (params.len_b == 0u) { atomicMin(&result, 0u); return; }
    if (params.len_b > params.len_a) { return; }
    let start = id.x;
    if (start <= params.len_a - params.len_b && start < atomicLoad(&result) && matches_at(start)) {
        atomicMin(&result, start);
    }
}