// SPDX-License-Identifier: AGPL-3.0-or-later
// Renorm - Renormalize with max norm constraint (f64 canonical)
// Clamps L2 norm of vectors to maximum value
// Used in training for gradient clipping alternative
//
// Algorithm:
// 1. Compute L2 norm: ||x||_2 = sqrt(Σ x_i^2)
// 2. If norm > max_norm: scale down: x = x * (max_norm / norm)
// 3. Else: keep unchanged
struct Params {
outer: u32, // Product of dimensions before norm_dim
dim_size: u32, // Size of dimension to normalize over
inner: u32, // Product of dimensions after norm_dim
max_norm: f64, // Maximum allowed norm
}
@group(0) @binding(0) var<uniform> params: Params;
@group(0) @binding(1) var<storage, read> input: array<f64>;
@group(0) @binding(2) var<storage, read_write> output: array<f64>;
@compute @workgroup_size(256)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
let idx = global_id.x;
let total = params.outer * params.inner;
if (idx >= total) {
return;
}
let o = idx / params.inner;
let i = idx % params.inner;
// Compute L2 norm
var norm_sq: f64 = 0.0;
for (var d = 0u; d < params.dim_size; d++) {
let val_idx = o * params.dim_size * params.inner + d * params.inner + i;
let val = input[val_idx];
norm_sq = norm_sq + val * val;
}
let norm = sqrt_f64(norm_sq);
// Compute scale factor (clamp norm to max_norm)
var scale: f64 = 1.0;
if (norm > params.max_norm) {
scale = params.max_norm / norm;
}
// Apply renormalization
for (var d = 0u; d < params.dim_size; d++) {
let val_idx = o * params.dim_size * params.inner + d * params.inner + i;
output[val_idx] = input[val_idx] * scale;
}
}