// SPDX-License-Identifier: AGPL-3.0-or-later
// pairwise_hamming_f64.wgsl — Pairwise Hamming Distance (f64 canonical)
//
// Each thread computes the Hamming distance (proportion of differing
// sites) between one pair of sequences. N sequences of length L
// produce N*(N-1)/2 pairwise distances.
//
// Provenance: neuralSpring metalForge (Feb 21, 2026) → ToadStool absorption
struct Params {
n_seqs: u32,
seq_len: u32,
}
@group(0) @binding(0) var<storage, read> sequences: array<u32>;
@group(0) @binding(1) var<storage, read_write> distances: array<f64>;
@group(0) @binding(2) var<uniform> params: Params;
@compute @workgroup_size(256)
fn pairwise_hamming(@builtin(global_invocation_id) gid: vec3<u32>) {
let pair_idx = gid.x;
let n = params.n_seqs;
let n_pairs = n * (n - 1u) / 2u;
if pair_idx >= n_pairs { return; }
// Decode pair_idx to (i, j) where i < j
var i: u32 = 0u;
var running: u32 = 0u;
for (var k: u32 = 0u; k < n - 1u; k = k + 1u) {
let count = n - 1u - k;
if running + count > pair_idx {
i = k;
break;
}
running = running + count;
}
let j = pair_idx - running + i + 1u;
var diff: u32 = 0u;
let offset_i = i * params.seq_len;
let offset_j = j * params.seq_len;
for (var s: u32 = 0u; s < params.seq_len; s = s + 1u) {
if sequences[offset_i + s] != sequences[offset_j + s] {
diff = diff + 1u;
}
}
distances[pair_idx] = f64(diff) / f64(params.seq_len);
}