/* One workgroup normalizes each row of residual = input + skip, with an
 * optional bias. */
{% if usesF16 %}
enable f16;
{% endif %}
{{ env.wgsl.resourceDeclarations }}

const HIDDEN: u32 = {{ hiddenSize }}u;
const WG: u32 = {{ workgroupSize }}u;

var<workgroup> partial: array<f32, WG>;
{% macro wgsl_tree_reduce_f32(name, mode, buffer="partial", wg="WG", trailingBarrier=true) %}
fn {{ name }}(value: f32, tid: u32) -> f32 {
  {{ buffer }}[tid] = value;
  workgroupBarrier();
  // Ceil-halving keeps every lane when the workgroup size is not a power of
  // two. For even n this matches the power-of-two tree order; for odd n, lanes
  // [0, n-half) fold the upper tail while the middle lane carries forward.
  var n: u32 = {{ wg }};
  loop {
    let half = (n + 1u) / 2u;
    if (tid < n - half) {
{% if mode == "max" %}
      {{ buffer }}[tid] = max({{ buffer }}[tid], {{ buffer }}[tid + half]);
{% else %}
      {{ buffer }}[tid] = {{ buffer }}[tid] + {{ buffer }}[tid + half];
{% endif %}
    }
    workgroupBarrier();
    n = half;
    if (n == 1u) {
      break;
    }
  }
  // The default trailing barrier makes this helper safe for back-to-back calls: every lane reads
  // slot 0 here, so the next call's first store must not run until all lanes have read it.
  // `trailingBarrier=false` is safe only when the buffer is never written again before kernel exit.
  let reduced = {{ buffer }}[0];
{% if trailingBarrier %}
  workgroupBarrier();
{% endif %}
  return reduced;
}
{% endmacro %}

{{ wgsl_tree_reduce_f32("reduce_sum", "add", "partial", "WG") }}
var<workgroup> row_inv: f32;

fn residual_value(row: u32, d: u32) -> f32 {
  let index = row * HIDDEN + d;
  var value = f32(input[index]) + f32(skip[index]);
{% if hasBias %}
  value = value + f32(bias[d]);
{% endif %}
  return value;
}

@compute @workgroup_size(WG, 1, 1)
fn main(
  @builtin(workgroup_id) wg: vec3<u32>,
  @builtin(local_invocation_id) lid: vec3<u32>) {
  // 2D-folded row index: wg.y carries the high bits past the per-axis dispatch fold width.
  // Reduces to wg.x when the dispatch does not fold;
  // the row >= params.rows guard drops the over-dispatched tail.
  let row = wg.x + wg.y * {{ DISPATCH_FOLD_WIDTH }}u;
  if (row >= params.rows) {
    return;
  }
  let tid = lid.x;

  // RMS normalization uses one sum-of-squares sweep, without a mean or beta.

  var local_sq = 0.0;
  for (var d: u32 = tid; d < HIDDEN; d = d + WG) {
    let value = residual_value(row, d);
    local_sq = local_sq + value * value;
  }
  let sq = reduce_sum(local_sq, tid);
  if (tid == 0u) {
    row_inv = inverseSqrt(sq / f32(HIDDEN) + params.epsilon);
  }
  workgroupBarrier();

  for (var d: u32 = tid; d < HIDDEN; d = d + WG) {
    let index = row * HIDDEN + d;
    let residual = residual_value(row, d);
{% if writeResidualSum %}
    input_skip_bias_sum[index] = {{ scalar }}(residual);
{% endif %}
    output[index] = {{ scalar }}(residual * row_inv * f32(gamma[d]));
  }
}