// SPDX-License-Identifier: AGPL-3.0-or-later
// local_response_norm.wgsl - Local Response Normalization (LRN) (f64 canonical)
//
// Normalizes activations within local neighborhoods
// Used in AlexNet and other early CNNs
//
// Formula: y_i = x_i / (k + alpha * sum(x_j^2) / size)^beta
struct Params {
batch_size: u32,
channels: u32,
height: u32,
width: u32,
size: u32, // Neighborhood size
alpha: f64, // Scaling parameter (typically 1e-4)
beta: f64, // Exponent (typically 0.75)
k: f64, // Bias (typically 1.0 or 2.0)
}
@group(0) @binding(0) var<storage, read> input: array<f64>; // [B, C, H, W]
@group(0) @binding(1) var<storage, read_write> output: array<f64>; // [B, C, H, W]
@group(0) @binding(2) var<uniform> params: Params;
@compute @workgroup_size(8, 8, 1)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
let b = global_id.z % params.batch_size;
let c = global_id.z / params.batch_size;
let h = global_id.y;
let w = global_id.x;
if (c >= params.channels || h >= params.height || w >= params.width) {
return;
}
let half_size = params.size / 2u;
// Calculate sum of squares in local neighborhood (across channels)
var sum_sq: f64 = 0.0;
let c_start = max(0, i32(c) - i32(half_size));
let c_end = min(i32(params.channels), i32(c) + i32(half_size) + 1);
for (var nc: i32 = c_start; nc < c_end; nc = nc + 1) {
let idx = b * params.channels * params.height * params.width +
u32(nc) * params.height * params.width +
h * params.width +
w;
let val = input[idx];
sum_sq = sum_sq + val * val;
}
// Compute normalization
let scale = pow_f64(params.k + params.alpha * sum_sq / f64(params.size), params.beta);
let in_idx = b * params.channels * params.height * params.width +
c * params.height * params.width +
h * params.width +
w;
output[in_idx] = input[in_idx] / scale;
}