ENH/API: offsets funcs now accepts datetime64

[sinhrks]

Even though CustomBusinessDay.apply can handle np.datetime64, most of other offsets cannot accept datetime64 and raises ApplyTypeError. The fix allows all offsets apply, rollforward and rollback to handle np.datetime64 properly.

import pandas as pd
import numpy as np
t = np.datetime64('2011-01-01 09:00Z')

cday = pd.offsets.CustomBusinessDay()
cday.apply(t)
#2011-01-02 09:00:00

day = pd.offsets.Day()
day.apply(t)
# pandas.tseries.offsets.ApplyTypeError: Unhandled type: datetime64

NOTE: CustomBusinessDay had separate logic for datetime and np.datetime64. Based on the comparison using current master, np.datetime64 logic looks slower. Thus I removed it.

import timeit
setup = """
import pandas as pd
import numpy as np

cday = pd.offsets.CustomBusinessDay()

np_dt64 = [np.datetime64('2014-05-{0:02} 09:00Z'.format(i)) for i in range(1, 31)]
timestamps = [pd.Timestamp('2014-05-{0:02} 09:00Z'.format(i)) for i in range(1, 31)]
"""

t = timeit.Timer('[cday.apply(d) for d in np_dt64]', setup)
print t.timeit(1000)
#1.6253619194
t = timeit.Timer('[cday.apply(d) for d in timestamps]', setup)
print t.timeit(1000)
#0.959406137466

[jreback]

hmm their is a vbench for this
can u add some of the other types to it as well (the other offsets)

I dont recall exactly the speed issue wrt to datetime/dagetime64 but their is a closed issue about it - can u zref the issue and make sure that nothing has changed

thxs

[jreback]

xref. #6592

[sinhrks]

Following is the vbench result. I understand #6592 added 2 separate logic for datetime and np.datetime64, and both are faster than previous one.

If we compare datetime logic and np.datetime64 logic, datetime logic is faster (thus removed np.datetime64 logic in this PR).

CC @bjonen

-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------
read_store_table                             |   5.0253 |   4.8374 |   1.0389 |
frame_to_csv_date_formatting                 |  36.4916 |  34.9390 |   1.0444 |
read_store_table_mixed                       |  18.4183 |  17.6213 |   1.0452 |
index_float64_boolean_indexer                |   7.6843 |   7.3157 |   1.0504 |
frame_loc_dups                               |   1.3700 |   1.2767 |   1.0731 |
dtype_infer_uint32                           |   2.6900 |   2.3301 |   1.1545 |
index_str_boolean_series_indexer             |  41.9977 |  35.5417 |   1.1816 |

[hayd]

related to #6318 ?

[jreback]

@sinhrks can you post the benches that relate to this?

[sinhrks]

@hayd Looks different, this is for datetime64, not timedelta.
@jreback Here it is.

-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------
timeseries_custom_bday_apply_dt64            |   0.0366 |   0.0540 |   0.6779 |
timeseries_custom_bmonthend_incr_n           |   0.2480 |   0.2427 |   1.0219 |
timeseries_custom_bday_incr_n                |   0.0547 |   0.0533 |   1.0253 |
timeseries_custom_bmonthend_incr             |   0.2210 |   0.2154 |   1.0262 |
timeseries_custom_bday_apply                 |   0.0337 |   0.0327 |   1.0316 |
timeseries_custom_bday_incr                  |   0.0363 |   0.0350 |   1.0386 |
-------------------------------------------------------------------------------