PERF: improve resample perf

[sinhrks]

Related to #7633. It gets better than the result attached #7633, but still slower more than 1.2 times compared to 1.4.0

Modified:

• Avoid every time module import in Index.max/min
• Avoid duplicated max call from resample/_get_time_bins and _get_range_edges.
• Optimize lib/generate_bins_dt64 and tslib/period_asfreq_arr.

Remaining bottlenecks are NaT masking performed in lib/generate_bins_dt64 and tslib/period_asfreq_arr. Is there any better way to do that?

-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------
dataframe_resample_mean_numpy                |   4.9963 |   3.7940 |   1.3169 |
dataframe_resample_mean_string               |   5.0424 |   3.8280 |   1.3172 |
dataframe_resample_max_numpy                 |   4.1796 |   3.0069 |   1.3900 |
dataframe_resample_min_numpy                 |   4.2127 |   2.9987 |   1.4049 |
dataframe_resample_min_string                |   4.1687 |   2.9490 |   1.4136 |
dataframe_resample_max_string                |   4.3443 |   2.9283 |   1.4835 |
timeseries_timestamp_downsample_mean         |  16.1959 |   8.6366 |   1.8753 |
timeseries_period_downsample_mean            |  47.6096 |  19.7030 |   2.4164 |
-------------------------------------------------------------------------------


Ratio < 1.0 means the target commit is faster then the baseline.
Seed used: 1234

Target [54fb875] : PERF: Improve index.min and max perf
Base   [da0f7ae] : RLS: 0.14.0 final

[jreback]

since nats are normally not present
you should check mask.any()
if this fails that you don't need to do self[mask] u can just use self

[sinhrks]

Thanks. Modified. Based on the cProfile, min/max no longer seems to be a critical.

dataframe_resample_mean_string               |   4.8401 |   3.6057 |   1.3423 |
dataframe_resample_max_string                |   3.9593 |   2.8160 |   1.4060 |
dataframe_resample_min_string                |   4.2106 |   2.8293 |   1.4882 |
dataframe_resample_mean_numpy                |   5.5540 |   3.6797 |   1.5094 |
dataframe_resample_max_numpy                 |   5.5197 |   3.4370 |   1.6059 |
dataframe_resample_min_numpy                 |   6.0006 |   3.1863 |   1.8832 |
timeseries_timestamp_downsample_mean         |  20.7520 |  10.1076 |   2.0531 |
timeseries_period_downsample_mean            |  45.5034 |  19.4256 |   2.3424 |

[jreback]

it's the re sampling this shouldn't be 2x slower

[sinhrks]

Period resample issue has been solved.

timeseries_period_downsample_mean            |  20.4194 |  19.6237 |   1.0405 |
dataframe_resample_min_string                |   3.9610 |   3.1497 |   1.2576 |
dataframe_resample_mean_numpy                |   4.8030 |   3.6697 |   1.3088 |
dataframe_resample_max_numpy                 |   4.0140 |   3.0653 |   1.3095 |
dataframe_resample_mean_string               |   4.7546 |   3.6027 |   1.3198 |
dataframe_resample_min_numpy                 |   4.0313 |   2.8987 |   1.3907 |
dataframe_resample_max_string                |   3.9481 |   2.7956 |   1.4122 |
timeseries_timestamp_downsample_mean         |  16.6246 |   8.6427 |   1.9235 |

[jreback]

ok, looks a lot better. ping me when green (and see if the perf diff in timeseries resample is fixable).
That is used quite a bit.

[jreback]

@sinhrks merged, but still havev a look at resample perf degredation