{{ env.wgsl.resourceDeclarations }}

{% macro flat_index_2d(workgroupSize, name="i", bound="params.count", guardInline=false) %}
{% set wgTerm = workgroupSize ~ "u" if workgroupSize is number else workgroupSize %}
  // 2D-folded flat index: gid.y carries the high bits past the dispatch's
  // per-axis workgroup fold width.
{% if bound == "" %}
  let {{ name }} = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ wgTerm }};{% else %}
  let {{ name }} = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ wgTerm }};
  if ({{ name }} >= {{ bound }}) {
    return;
  }{% endif %}{% endmacro %}
{% if vec4PerThread > 1 %}
const ITEMS: u32 = {{ vec4PerThread }}u;

{% endif %}
@compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
{% if vec4PerThread > 1 %}
  // Each invocation walks ITEMS vec4 groups a span apart. Consecutive lanes
  // access consecutive words on every step, while each lane can keep several
  // independent loads in flight.
{{ flat_index_2d(tunables.WORKGROUP_SIZE, "tid", "") }}
  let span = (params.count + ITEMS - 1u) / ITEMS;
  // Lanes at or past span exist only because the dispatch rounds up to whole
  // workgroups. Lane span + k would start on lane k's second group and rewrite
  // up to ITEMS - 1 groups another lane already stored.
  if (tid >= span) {
    return;
  }
  for (var j = 0u; j < ITEMS; j = j + 1u) {
    let i = tid + j * span;
    if (i >= params.count) {
      break;
    }
{% else %}
{{ flat_index_2d(tunables.WORKGROUP_SIZE) }}
{% endif %}
{% set scalarOperand = scalarOperand if scalarOperand is defined else "" %}
{% if scalarOperand == "a" %}
  // One-element operand: read once and splat across the vector.
  let av = {{ vectorScalar }}(a[0]);
{% else %}
  let av = a[i];
{% endif %}
{% if scalarOperand == "b" %}
  let bv = {{ vectorScalar }}(b[0]);
{% else %}
  let bv = b[i];
{% endif %}
{% if scalar == "i32" or scalar == "u32" %}
  let r = av + bv;
// Narrow integer operations wrap modulo the logical dtype width; int8/uint8
// use i32/u32 storage.
{% if cDtype == "int8" %}
  c[i] = (r << vec4<u32>(24u)) >> vec4<u32>(24u);
{% elif cDtype == "uint8" %}
  c[i] = r & vec4<u32>(0xFFu);
{% else %}
  c[i] = r;
{% endif %}
{% elif scalar == "f16" %}
  c[i] = vec4<f16>(vec4<f32>(av) + vec4<f32>(bv));
{% else %}
  c[i] = av + bv;
{% endif %}
{% if vec4PerThread > 1 %}
  }
{% endif %}
}