struct MatrixBatch {
  size: vec4<u32>,
  data: array<vec4<f32>>,
}

@group(0) @binding(0)
var<storage, read> c: MatrixBatch;
@group(0) @binding(1)
var<storage, read_write> result: MatrixBatch;

@compute @workgroup_size(1)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
  result.size = c.size;

  let vc = c.size.x / 4;
  let ro = (global_id.z * c.size.y + global_id.y) * vc;

  var s = vec4(0f);
  for (var x = 0u; x < vc; x++) {
    s += c.data[ro + x] * c.data[ro + x];
  }

  result.data[ro + global_id.x] = c.data[ro + global_id.x] * vec4(inverseSqrt((s.x + s.y + s.z + s.w) / f32(c.size.x) + 1e-5f));
}