Float64Index is very slow in some condition.

[ruoyu0088]

The following code is very slow:

import pandas as pd
import numpy as np

dt = 4.8000000418824129e-08
data = np.random.rand(1000000, 4)
df = pd.DataFrame(data, columns=list("ABCD"))
df.index *= dt
print(df.loc[0:0.001].shape)

after debug it, I found Float64Engine.get_loc() is slow. Here is a demo:

import pandas as pd
import numpy as np

a = np.arange(1000000)
ind1 = pd.Float64Index(a * 4.8e-08)
ind2 = pd.Float64Index(a * 4.8000000418824129e-08)

%time ind1._engine.get_loc(0)
%time ind2._engine.get_loc(0)

outputs:

Wall time: 295 ms
Wall time: 9.9 s

[jreback]

This might have been true on older versions of pandas (maybe < 0.16.0, I don't recall the exact version) as these were object based. but these moved to true float based hashtables (typed).

In [1]: pd.__version__
Out[1]: u'0.18.1'

In [2]: a = np.arange(1000000)

In [3]: ind1 = pd.Float64Index(a * 4.8e-08)

In [4]: ind2 = pd.Float64Index(a * 4.8000000418824129e-08)

In [5]: %timeit -n 1 -r 1 ind1._engine.get_loc(0)
1 loop, best of 1: 116 ms per loop

In [6]: %timeit -n 1 -r 1 ind2._engine.get_loc(0)
1 loop, best of 1: 107 ms per loop

closing, but pls show your versions

note you have to time these with a single iteration as these build the hash tables which are then cached (so of course that tells you the lookup time, but you want the build time as well)

[ruoyu0088]

@jreback

I am using pandas 0.18.1, python 3.5 64bit, I confirmed this problem both on Linux and Windows 7.
The pd.hashtable.Float64HashTable is slow for ind2 on my system. I think this is due to the khash library.

INSTALLED VERSIONS
------------------
commit: None
python: 3.5.1.final.0
python-bits: 64
OS: Linux
OS-release: 3.10.18
machine: x86_64
processor: Intel(R) Celeron(R) CPU N2840 @ 2.16GHz
byteorder: little
LC_ALL: en_US.utf8
LANG: None

[ruoyu0088]

@jreback It seems that Python 3.5 has the problem, but Python 2.7 has no problem. Can you reopen this issue? You can confirm this on https://try.jupyter.org/.

[jreback]

hmm interesting

so you want to try profiling the cython ?

[ruoyu0088]

I think here is the problem, but I don't know why:

khash_python.h

#define kh_float64_hash_func _Py_HashDouble

I changed the line to following code, it view the double value as int64 and use the same formula as kh_int64_hash_func:

inline khint64_t asint64(double key)
{
  return *(khint64_t *)(&key);
}

#define kh_float64_hash_func(key) (khint32_t)((asint64(key))>>33^(asint64(key))^(asint64(key))<<11)

The %time result is almost the same, and it's even 2x faster for ind0.

[jreback]

does that break any tests? can you run the asv suite as well (and add benchmark for this).

[jreback]

https://github.com/python/cpython/blob/master/Python/pyhash.c#L85 is the existing PyHash_double.

Its probably generating 'better' hashes that your change, but in the end of the day I don't see why that's preferable.