alias number = f32;

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

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

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

@compute @workgroup_size(1)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>, @builtin(num_workgroups) num_workgroups: vec3<u32>) {
  result.size = m.size;
  result.size.z = num_workgroups.z * 2;

  var vc = result.size.x / 4;
  var mi = global_id.y * vc + global_id.x;
  result.data[global_id.z * result.size.y * vc + mi] = m.data[mi];
}