Speed up max_len_string_array

[cpcloud]

Before:

In [5]: x = np.array(['abcd', 'abcde', 'abcdef', 'abcdefg'] * int(1e7), object)

In [2]: f = pd.lib.max_len_string_array

In [6]: %timeit f(x)
1 loops, best of 3: 501 ms per loop

After:

In [1]: x = np.array(['abcd', 'abcde', 'abcdef', 'abcdefg'] * int(1e7), object)

In [2]: f = pd.lib.max_len_string_array

In [3]: %timeit f(x)
10 loops, best of 3: 68.4 ms per loop

[jreback]

looks good. release note in perf section & squash. merge when ready.

[shoyer]

I was going to mention you could use a better name here :).

[jreback]

my naive quick look does show some improvements to to_hdf when you have object dtypes, order is 10-15% impv. (only used in to_stata/to_hdf).

[cpcloud]

@shoyer I didn't have a particular pandas use case for this. I'm going to start using it in some often called paths in odo and I wanted to see if I could squeeze out some more perf.

[cpcloud]

I think @jreback had some ideas about using cython memoryviews in some of re csv code similar to how I use them here. IIRC he said there are quite a few places where we don't take full advantage of what cython has to offer. For example if you type a variable as just ndarray you incur the overhead of the fully general get item c API whereas if you type it as ndarray[type] the getitem syntax goes directly to the underlying raw pointer array.

[cpcloud]

ok squashed. merging on pass