REF: dont use numexpr in places where it doesnt help.

[jbrockmendel]

DataFrame ops will still use it when dispatching to series/block ops.

Perf neutral:

In [3]: arr = np.arange(10**4).reshape(100, 100)                                                                                                                                    
In [4]: df = pd.DataFrame(arr)                                                                                                                                                      
In [5]: df[4] = df[4].astype(float)                                                                                                                                                 
In [6]: ser = df[0]                     
                                                                                                                                            
In [8]: %timeit res = df+ser                                                                                                                                                        
19.8 ms ± 317 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # <-- PR
19.8 ms ± 267 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # <-- master

In [10]: %timeit res2 = df.add(ser, axis=0)                                                                                                                                         
19.2 ms ± 276 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # <-- PR
19.2 ms ± 301 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # <-- master

In [13]: %timeit res3 = df + df                                                                                                                                                     
23.6 ms ± 382 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # <-- PR
24 ms ± 551 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # <-- master

[jreback]

hmm, the original and entire purpose IS to use numexpr for comparisions. I agree consistency is important, but this is turned off for DataFrames? (for regular dtypes)????

[jbrockmendel]

yah, the right place to do that is in array_ops (which will cause is to do it for Series too)

I've got a branch that implements dataframe-with-dataframe ops block-wise, but the diff for that is pretty big, so im splitting pieces off. the numexpr-for-series-comparisons part will be next

[jreback]

ok great

yeah comparison ops should. r a pretty big speedup as it’s parallle
computation

[jbrockmendel]

ok, im in the process of splitting off the relevant parts of that branch. That may end up affecting the tests here, but the core.ops edits here will be correct regardless.

I think we need to improve the used_numexpr check in the tests to not just check whether numexpr was called, but whether it actually was used effectively

[jbrockmendel]

also xref #31297 will smooth this out

[jbrockmendel]

closing in favor of #32047.