// SPDX-License-Identifier: AGPL-3.0-or-later
// Clip Grad Norm - Gradient clipping by total norm (f64 canonical)
struct Params {
size: u32,
max_norm: f64,
_pad1: u32,
_pad2: u32,
}
@group(0) @binding(0) var<uniform> params: Params;
@group(0) @binding(1) var<storage, read> gradients: array<f64>;
@group(0) @binding(2) var<storage, read_write> norm_buffer: array<f64>;
@group(0) @binding(3) var<storage, read_write> output: array<f64>;
var<workgroup> shared_sq: array<f64, 256>;
@compute @workgroup_size(256)
fn compute_norm(
@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;
var grad_sq: f64 = 0.0;
if (gid < params.size) {
let g = gradients[gid];
grad_sq = g * g;
}
shared_sq[tid] = grad_sq;
workgroupBarrier();
for (var stride = 128u; stride > 0u; stride = stride / 2u) {
if (tid < stride) {
shared_sq[tid] = shared_sq[tid] + shared_sq[tid + stride];
}
workgroupBarrier();
}
if (tid == 0u) {
norm_buffer[workgroup_id.x] = shared_sq[0];
}
}
@compute @workgroup_size(256)
fn compute_norm_final(
@builtin(local_invocation_id) local_id: vec3<u32>,
) {
let tid = local_id.x;
let num_partials = (params.size + 255u) / 256u;
if (tid == 0u) {
var total: f64 = 0.0;
for (var i = 0u; i < num_partials; i = i + 1u) {
total = total + norm_buffer[i];
}
norm_buffer[0] = total;
}
}
@compute @workgroup_size(256)
fn clip_gradients(@builtin(global_invocation_id) global_id: vec3<u32>) {
let idx = global_id.x;
if (idx >= params.size) {
return;
}
let norm_sq = norm_buffer[0];
let total_norm = sqrt_f64(norm_sq);
var scale: f64 = 1.0;
if (total_norm > params.max_norm) {
scale = params.max_norm / (total_norm + 1e-8);
}
output[idx] = gradients[idx] * scale;
}