CLN/PERF: Simplify argmin/argmax

[rhshadrach]

• closes #xxxx (Replace xxxx with the GitHub issue number)
• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added type annotations to new arguments/methods/functions.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.

Ref: #57971 (comment)

ASVs against main:

| Change   | Before [b63ae8c7]    | After [46bbd5eb] <cln_argmin_argmax>   |   Ratio | Benchmark (Parameter)                                         |
|----------|----------------------|----------------------------------------|---------|---------------------------------------------------------------|
| -        | 184±1μs              | 138±3μs                                |    0.75 | series_methods.NanOps.time_func('argmax', 1000000, 'int32')   |
| -        | 14.4±0.1μs           | 10.5±0.2μs                             |    0.73 | series_methods.NanOps.time_func('argmax', 1000, 'float64')    |
| -        | 1.23±0.06ms          | 863±20μs                               |    0.7  | series_methods.NanOps.time_func('argmax', 1000000, 'float64') |
| -        | 77.9±0.6μs           | 40.8±0.7μs                             |    0.52 | series_methods.NanOps.time_func('argmax', 1000000, 'int8')    |
| -        | 7.42±0.2μs           | 2.44±0.06μs                            |    0.33 | series_methods.NanOps.time_func('argmax', 1000, 'int64')      |
| -        | 7.18±0.2μs           | 2.29±0.02μs                            |    0.32 | series_methods.NanOps.time_func('argmax', 1000, 'int32')      |
| -        | 7.25±0.1μs           | 2.29±0.03μs                            |    0.32 | series_methods.NanOps.time_func('argmax', 1000, 'int8')       |

ASVs against 2.2.x show no perf change.

[WillAyd]

Overall this looks really nice

[WillAyd]

Shouldn't we still be hitting this with the axis == 1 case?

[rhshadrach]

When axis==1 we run into all NAs for a single row, but when skipna=False we still raise the error message "Encountered an NA value with skipna=False" for performance - see the bottom half of #57971 (comment)

[rhshadrach]

Without using the delegate here and the nanops.nanargmax branch below, I'm seeing a 50x perf regression in some ASVs, e.g. series_methods.NanOps.time_func('argmax', 1000000, 'int8'):

N = 1000000
dtype = 'int8'
s = Series(np.ones(N), dtype=dtype)
s.argmax()

@jbrockmendel - perhaps that branch should be deeper in the code where the argmin/argmax implementation is for EAs?

[jbrockmendel]

Do you mean if you do return return self.array.argmax(skipna=skipna) without the EA check?

It looks like sorting.nanargminmax is just a lot slower than the nanops.arg(min|max). Best case would be to improve sorting.nanargminmax, but failing that I think making NumpyExtensionArray.argmax call the faster version would be viable (need to double-check they are behaviorally equivalent)

[rhshadrach]

Do you mean if you do return return self.array.argmax(skipna=skipna) without the EA check?

Correct.

Best case would be to improve sorting.nanargminmax, but failing that I think making NumpyExtensionArray.argmax call the faster version would be viable

Thanks - will give these a shot.

[rhshadrach]

Looked into this, and I don't think either is viable. The main reason nanops.nanargmax is faster than sorting.nanargminmax is that the former only works for numeric dtypes by filling in NA values where appropriate (typically with -np.inf) whereas for EAs (e.g. StringArray) we need to perform indexing operations to remove NA values from any consideration since there is no equivalent of -np.inf I think.

[jbrockmendel]

bummer. thanks for takinga look

[mroeschke]

Thanks @rhshadrach