bboard

LZNIB can send three actions : literal run ("lrl") , match with offset ("normal match"), match with repeat offset ("rep match").

One of the key realizations for LZNIB is that even though there are three actions, only two are possible at any time.


After Literals :

- LRL should not occur (would have just been a longer LRL)
- match & rep match are possible

After Match or Rep Match :

- rep match should not occur (would have just been a longer match)
- normal match & lrl are possible

Because there are only two actions possible at each point, we can send this using a nibble with a decision threshold value :


Threshold T

- values in [0 , T-1] are action type 1
- values in [T , 15] are action type 2

So the core decode step of LZNIB is to grab the next nibble, do one branch, and then process that action. The values within your threshold group are the first part of your LRL or match len. There are at least two different thresholds, one for {match/rep-match} in the after-literals state, and one for {match/lrl} in the after-match state. In LZNIB we hard-coded the threshold for match/rep-match to 5 as this was found to be good on most data. The optimal {match/lrl} threshold is more dependent on the data.