[_0ffh]

The trick is to make sure the recursive dependency stays linear, that's how you enable parallel training.