PERF/CLN: see what we can use from offsets.pyx

[jreback]

xref #11205

https://github.com/pydata/pandas/blob/master/pandas/src/offsets.pyx

is in the repo but is not included in the build, nor updated in 2+ years.

Looks like their might be some cython code to speed up DateOffset apply in the non-vectorized cases.
Further might look to move some routines from tslib.pyx for these types of things (and obviously would need to be included in the build).

[jreback]

cc @chris-b1
cc @cancan101
cc @sinhrks

[sinhrks]

Also looks nice place to put some cythonized codes from offsets.py and frequencies.py.

[chris-b1]

I didn't find a whole lot useful in the existing code, it's a slightly different definition of an offset, basically stored as a day ordinal from what I can tell.

I did scratch out what the current DateOffset class might look like as a cython extension type - https://github.com/chris-b1/pandas/tree/cythonize-offset - it could improve performance quite a bit in the non-vectorized case (and remove the dual definition needed for the vectorized case). Also could generate ranges faster.

But it'd take quite a bit of work to actually port everything over in a feature-complete way, and would have worry about things like pickle compat. So maybe makes more sense to define some targeted helper cython functions and attach to the existing classes or something?

In [1]: from pandas.tseries.offsets import DateOffset, DateOffset2
   ...: old_do = DateOffset(months=1)
   ...: new_do =  DateOffset2(months=1)

In [2]: ts = pd.Timestamp('2014-1-1')

In [3]: %%timeit
   ...: for _ in xrange(10000):
   ...:     old_do.apply(ts)
10 loops, best of 3: 143 ms per loop

In [4]: %%timeit
   ...: for _ in xrange(10000):
   ...:     new_do.apply(ts)
100 loops, best of 3: 13.5 ms per loop

In [5]: dti = pd.date_range('1900-1-1', periods=10000)

In [6]: %timeit old_do.apply_index(dti)
1000 loops, best of 3: 1 ms per loop

In [7]: %timeit new_do.apply_index(dti)
1000 loops, best of 3: 1.02 ms per loop

[jreback]

@chris-b1 right, your helpers could be used in cases when advancing more than 1 date rather than doing it in a python loop.

so will repurpose this issue.

if you have a chance can you put up a list of things that should be targeted and I will put up checkboxes.