// SPDX-License-Identifier: AGPL-3.0-or-later
// LayerNorm: Layer normalization with Welford's algorithm (f64 canonical)
// CUDA equivalent: cudnn::LayerNormalization
// Formula: output = (input - mean) / sqrt(variance + epsilon) * gamma + beta
// Use cases: Transformer normalization, stabilizing training
@group(0) @binding(0) var<storage, read> input: array<f64>;
@group(0) @binding(1) var<storage, read> gamma: array<f64>; // Scale (optional, can be all 1s)
@group(0) @binding(2) var<storage, read> beta: array<f64>; // Shift (optional, can be all 0s)
@group(0) @binding(3) var<storage, read_write> output: array<f64>;
@group(0) @binding(4) var<storage, read_write> stats: array<f64>; // [mean, variance]
struct Params {
size: u32,
epsilon: f64,
}
@group(0) @binding(5) var<uniform> params: Params;
var<workgroup> shared_sum: array<f64, 256>;
var<workgroup> shared_sum_sq: array<f64, 256>;
// Pass 1: Compute mean and variance using Welford's online algorithm
@compute @workgroup_size(256)
fn compute_stats(
@builtin(global_invocation_id) global_id: vec3<u32>,
@builtin(local_invocation_id) local_id: vec3<u32>,
@builtin(workgroup_id) workgroup_id: vec3<u32>,
) {
let tid = local_id.x;
let gid = global_id.x;
// Load and compute local statistics
var value: f64 = 0.0;
var value_sq: f64 = 0.0;
if (gid < params.size) {
value = input[gid];
value_sq = value * value;
}
shared_sum[tid] = value;
shared_sum_sq[tid] = value_sq;
workgroupBarrier();
// Tree reduction
for (var stride = 128u; stride > 0u; stride = stride / 2u) {
if (tid < stride) {
shared_sum[tid] = shared_sum[tid] + shared_sum[tid + stride];
shared_sum_sq[tid] = shared_sum_sq[tid] + shared_sum_sq[tid + stride];
}
workgroupBarrier();
}
if (tid == 0u) {
stats[workgroup_id.x * 2u] = shared_sum[0]; // Partial sum
stats[workgroup_id.x * 2u + 1u] = shared_sum_sq[0]; // Partial sum of squares
}
}
// Pass 2: Finalize statistics (reduce partial sums to final mean/variance)
@compute @workgroup_size(256)
fn finalize_stats(
@builtin(global_invocation_id) global_id: vec3<u32>,
@builtin(local_invocation_id) local_id: vec3<u32>,
) {
let tid = local_id.x;
// Load partial sums from multiple workgroups
var local_sum: f64 = 0.0;
var local_sum_sq: f64 = 0.0;
// Each thread loads one partial sum
let num_partials = arrayLength(&stats) / 2u;
if (tid < num_partials) {
local_sum = stats[tid * 2u];
local_sum_sq = stats[tid * 2u + 1u];
}
shared_sum[tid] = local_sum;
shared_sum_sq[tid] = local_sum_sq;
workgroupBarrier();
// Tree reduction to sum all partials
for (var stride = 128u; stride > 0u; stride = stride / 2u) {
if (tid < stride && (tid + stride) < num_partials) {
shared_sum[tid] = shared_sum[tid] + shared_sum[tid + stride];
shared_sum_sq[tid] = shared_sum_sq[tid] + shared_sum_sq[tid + stride];
}
workgroupBarrier();
}
// Thread 0 computes final mean and variance
if (tid == 0u) {
let total_sum = shared_sum[0];
let total_sum_sq = shared_sum_sq[0];
let n = f64(params.size);
let mean = total_sum / n;
let variance = (total_sum_sq / n) - (mean * mean);
// Store final stats at beginning of buffer
stats[0] = mean;
stats[1] = variance;
}
}
// Pass 3: Normalize
@compute @workgroup_size(256)
fn normalize(
@builtin(global_invocation_id) global_id: vec3<u32>,
) {
let gid = global_id.x;
if (gid < params.size) {
// Final mean and variance are now in stats[0] and stats[1]
let mean = stats[0];
let variance = stats[1];
let normalized = (input[gid] - mean) / sqrt_f64(variance + params.epsilon);
output[gid] = normalized * gamma[gid] + beta[gid];
}
}