alias number = f32;

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

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

@compute @workgroup_size(1)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
  out.size.x = in.size.y;
  out.size.y = in.size.x;
  out.size.z = in.size.z;
  var l = global_id.z * out.size.y * out.size.x;
  out.data[l + global_id.y * out.size.x + global_id.x] = in.data[l + global_id.x * in.size.x + global_id.y];
}