PERF: fastpath on Period construction from PeriodIndex

[jreback]

vbench already in timeseries/period_setitem
very inefficient because on the boxing of the PeriodIndex (to an object index), it recomputes the freq for each Period (even though it is passed), need a fastpath on this type of construction

In [1]: rng = period_range('1/1/1990', freq='S', periods=20000)

In [2]: df = DataFrame(index=range(len(rng)))

In [3]: def f():
   ...:     df['col'] = rng
   ...:     
   Ordered by: internal time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
    40000    0.055    0.000    0.095    0.000 abc.py:128(__instancecheck__)
    20000    0.043    0.000    0.345    0.000 period.py:66(__init__)
    20000    0.037    0.000    0.381    0.000 period.py:683(<lambda>)
    80031    0.037    0.000    0.131    0.000 {isinstance}
    20000    0.034    0.000    0.223    0.000 frequencies.py:76(get_freq_code)
    60003    0.029    0.000    0.029    0.000 _weakrefset.py:68(__contains__)
    20000    0.028    0.000    0.063    0.000 frequencies.py:464(_base_and_stride)
    20000    0.028    0.000    0.028    0.000 {method 'match' of '_sre.SRE_Pattern' objects}
    20000    0.027    0.000    0.038    0.000 frequencies.py:767(_period_str_to_code)
    40000    0.023    0.000    0.144    0.000 common.py:1730(is_integer)
        1    0.016    0.016    0.398    0.398 {pandas.lib.map_infer}
    40007    0.010    0.000    0.010    0.000 {getattr}
    20000    0.009    0.000    0.012    0.000 frequencies.py:112(_get_freq_str)
    60000    0.008    0.000    0.008    0.000 {method 'get' of 'dict' objects}
    40000    0.006    0.000    0.006    0.000 {method 'group' of '_sre.SRE_Match' objects}
    20000    0.004    0.000    0.004    0.000 {method 'lower' of 'str' objects}
    20000    0.002    0.000    0.002    0.000 {method 'upper' of 'str' objects}
    20015    0.002    0.000    0.002    0.000 {len}

[jtratner]

@jreback did you resolve this already? I might have to deal with this soon anyways, as I'm about to tackle Period and DatetimeIndex fixups.

[jreback]

nope
it shouldn't be too hard; Period needs a fast path where it just sets stuff

it works now but does lots of work when u have a long PeriodIndex

[jreback]

IOW it reinterprets the freq code on every Period construction even when u r passing valid values (but it doesn't know that)

[jtratner]

You should also make sure you're fastpathing asobject (not sure if it matters for PI though)

[jreback]

not planning in this till 0.14

[jtratner]

got it

On Mon, Oct 14, 2013 at 11:10 PM, jreback [email protected] wrote:

not planning in this till 0.14

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/5155#issuecomment-26306475
.

[jreback]

revisit after #5148

[nehalecky]

Any relevance: #2949?

[jreback]

that's a separate issue...this is a pretty easy fix

[nehalecky]

Thanks, just wanted to throw it in the mix. I love (conceptually) PeriodIndexs, but they need a lot of tlc soon to be all they can be. :)

[jreback]

@nehalecky turns out this was pretty trivial, see #7891

Period constructor has lots of inference going on. So when you know that you simply want to wrap an ordinal/freq its trivial to construct. (you can usually only do this AFTER its constructed, but that's usually when you pay the cost, eg. when passing to matplotlib or setting actual periods in a data frame).