Potential Optimisation: Most of round 1007 can be cached when X doesn't change.

[phire]

From what I can tell, most of round 1007 is static when Y doesn't change. As data_next, X only impacts 5 words of the state.

So we could implement a truncated round 1007 that only calculates state[10:11] and state[13-15]. Round 1008 is already terminates early and only calculates state[0:9]. Interestingly, state[12] is actually fixed.