This could be done much faster using a precomputed table of state transitions. Depending on what is the old i-th differential value, the updated i-th differential value, the old i+1-th differential value and whether the i+1-th position is in match or mismatch.
This is a graph with 5 states (old transition and updated transition):
[old]-[updated]
up-same
same-same
same-down
down-same
down-down

Each state has six possible transitions:
[match/mismatch]-[old i+1-th differential value]
match-up
match-same
(...)
mismatch-down
which yields an emission (the updated i+1-th differential value) and connects to one of the 5 states. This way a pattern of length m could be updated in m operations without reverting the differential encoding.