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

@group(0) @binding(0)
var<storage, read> b: 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 = b.size;

  var vc = result.size.x / 4;
  var resi = (global_id.z * result.size.y + global_id.y) * vc + global_id.x;
  var sv = vec4(0f);
  var bi = (global_id.z * b.size.y + global_id.y) * vc;
  for (var i = 0u; i < vc; i++) {
    sv += b.data[bi];
    bi++;
  }
  var ss = sv.x + sv.y + sv.z + sv.w;
  result.data[resi] = b.data[resi] / vec4(ss);
}