alias number = f32;

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

@group(0) @binding(0)
var<storage, read> l: MatrixBatch;
@group(0) @binding(1)
var<storage, read> r_or_result: MatrixBatch;
@group(0) @binding(2)
var<storage, read_write> result: MatrixBatch;

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

  var resi = (global_id.z * result.size.y + global_id.y) * result.size.x / 4 + global_id.x;
  result.data[resi] = l.data[resi] - r_or_result.data[resi];
}