PERF: improved clip performance

[jreback]

closes #15400
In [1]: np.random.seed(1234)

In [2]: s = pd.Series(np.random.randn(50))

master

In [3]: %timeit s.clip(0, 1)
1.65 ms ± 48.2 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

PR

In [3]: %timeit s.clip(0, 1)
124 µs ± 2.79 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

prob as good as can do for now as we still have 2 where ops (numpy does this in a single loop), and we have a mask check and fill (and final construction).

but about 15x better

[codecov]

Codecov Report

Merging #16364 into master will decrease coverage by 0.01%.
The diff coverage is 94.11%.

@@            Coverage Diff             @@
##           master   #16364      +/-   ##
==========================================
- Coverage   90.38%   90.36%   -0.02%     
==========================================
  Files         161      161              
  Lines       50916    50933      +17     
==========================================
+ Hits        46021    46028       +7     
- Misses       4895     4905      +10

Flag	Coverage Δ
#multiple	88.14% <94.11%> (ø)	⬆️
#single	40.21% <5.88%> (-0.12%)	⬇️

Impacted Files	Coverage Δ
pandas/core/generic.py	91.96% <94.11%> (+0.01%)	⬆️
pandas/io/gbq.py	25% <0%> (-58.34%)	⬇️
pandas/core/frame.py	97.68% <0%> (-0.1%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d92f06a...62843f8. Read the comment docs.

[jorisvandenbossche]

There is a slight change in behaviour in that the new implementation does not preserve heterogeneous data types (eg int/float).
Not that this should hold back these perf improvements, but we might consider to keep it for 0.21 (this is also no regression)

[jorisvandenbossche]

I think this needs a with np.errstate, as we are working with raw array

In [8]: pd.Series([0, np.nan, 2]).clip(0, 1)
/home/joris/scipy/pandas/pandas/core/generic.py:4117: RuntimeWarning: invalid value encountered in greater_equal
  result = np.where(result >= upper, upper, result)
/home/joris/scipy/pandas/pandas/core/generic.py:4119: RuntimeWarning: invalid value encountered in less_equal
  result = np.where(result <= lower, lower, result)
Out[8]: 
0    0.0
1    NaN
2    1.0
dtype: float64

[jorisvandenbossche]

Are the np.any needed here? As lower/upper are already confirmed to be a scalar?

[jorisvandenbossche]

new implementation does not preserve heterogeneous data types

In principle could add a check for that (at the if statement to decide to take this path or not), but not sure that is worth it ..

[jreback]

yeah this is a limitation of the current methodology

let me see what i can do