PERF: cache _get_cython_function in groupby ops

[jorisvandenbossche]

Looking up the exact cython function is relatively costly, and is also something that is done repeatedly for a limited number of possible cases, so a good target for caching.

Using the GroupManyLabels benchmark we have:

pandas/asv_bench/benchmarks/groupby.py

Lines 133 to 145 in b835ca2

	class GroupManyLabels:
	params = [1, 1000]
	param_names = ["ncols"]
	def setup(self, ncols):
	N = 1000
	data = np.random.randn(N, ncols)
	self.labels = np.random.randint(0, 100, size=N)
	self.df = DataFrame(data)
	def time_sum(self, ncols):
	self.df.groupby(self.labels).sum()

ncols = 1000
N = 1000
data = np.random.randn(N, ncols)
labels = np.random.randint(0, 100, size=N)
df = pd.DataFrame(data)
df_am = df._as_manager('array')

I get

In [2]: %timeit df_am.groupby(labels).sum()
72.6 ms ± 1.24 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) <-- master
66.4 ms ± 1.43 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) <-- PR

(the difference is not huge, but I am going to do several PRs with small changes that give a few ms improvement. And with all those together I hope to get the total time down to around 20ms, and in that case 5ms is significant)

[jorisvandenbossche]

This function is a straight cut&paste of the existing method. But moving it to a standalone function instead of method to be able to use lru_cache on it.

[jreback]

lgtm.

[jbrockmendel]

LGTM

[jbrockmendel]

any idea how many cases make it past here? might be we can make some of this unnecessary on the cython side

[jorisvandenbossche]

Quite some I think. Eg for sum, we have group_add_.. which only has type-specific functions exposed (the same for mean, prod, var). Not directly sure why we expose some functions with numeric fused type, and others we expose the type specific ones (because those are also coded with fused types). That's a bigger change for another PR, though ;)

(now, an expensive part of this function is dtype.name, which is above this line anyway, so won't change that it's useful to cache this I think)