// SPDX-License-Identifier: AGPL-3.0-or-later
// weight_norm.wgsl - Weight Normalization (f64 canonical)
//
// Reparameterizes weights as: w = g * (v / ||v||)
// Reference: "Weight Normalization" by Salimans & Kingma (2016)
//
// Improves training speed and convergence
struct Params {
num_weights: u32,
dim: u32, // Dimension to normalize over (0 = all)
_padding: vec2<u32>,
}
@group(0) @binding(0) var<storage, read> v: array<f64>; // Direction vectors
@group(0) @binding(1) var<storage, read> g: array<f64>; // Magnitude scalars
@group(0) @binding(2) var<storage, read_write> output: array<f64>; // Normalized weights
@group(0) @binding(3) var<uniform> params: Params;
@compute @workgroup_size(256)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
let idx = global_id.x;
if (idx >= params.num_weights) {
return;
}
if (params.dim == 0u) {
// Normalize all weights together
if (idx == 0u) {
// Compute norm
var norm_sq: f64 = 0.0;
for (var i: u32 = 0u; i < params.num_weights; i = i + 1u) {
norm_sq = norm_sq + v[i] * v[i];
}
let norm = sqrt_f64(norm_sq + 1e-12);
// Apply: w = g * (v / ||v||)
for (var i: u32 = 0u; i < params.num_weights; i = i + 1u) {
output[i] = g[0] * v[i] / norm;
}
}
} else {
// Per-dimension normalization (simplified - would need grouping)
output[idx] = v[idx];
}
}