// SPDX-License-Identifier: AGPL-3.0-or-later
// pdist.wgsl - Pairwise Distances (all pairs) (f64 canonical)
//
// Computes distances between all pairs of row vectors
// Output is condensed distance matrix (upper triangle)
struct Params {
num_vectors: u32,
dim: u32,
p: f64, // p-norm
epsilon: f64,
}
@group(0) @binding(0) var<storage, read> input: array<f64>; // [num_vectors, dim]
@group(0) @binding(1) var<storage, read_write> output: array<f64>; // [num_pairs] where num_pairs = n*(n-1)/2
@group(0) @binding(2) var<uniform> params: Params;
@compute @workgroup_size(16, 16)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
let i = global_id.x;
let j = global_id.y;
if (i >= params.num_vectors || j >= params.num_vectors || i >= j) {
return;
}
var dist: f64 = 0.0;
for (var d: u32 = 0u; d < params.dim; d = d + 1u) {
let val_i = input[i * params.dim + d];
let val_j = input[j * params.dim + d];
let diff = abs(val_i - val_j);
if (params.p == 1.0) {
dist = dist + diff;
} else if (params.p == 2.0) {
dist = dist + diff * diff;
} else {
dist = dist + pow_f64(diff + params.epsilon, params.p);
}
}
if (params.p == 2.0) {
dist = sqrt_f64(dist);
} else if (params.p > 1.0 && params.p != 2.0) {
dist = pow_f64(dist + params.epsilon, 1.0 / params.p);
}
// Condensed index: row i, col j (i < j)
// Index = i * n - i*(i+1)/2 + (j - i - 1)
let out_idx = i * params.num_vectors - (i * (i + 1u)) / 2u + (j - i - 1u);
output[out_idx] = dist;
}