PERF: fix some of .clip() performance regression by using numpy arrays where possible #24735

qwhelan · 2019-01-11T22:17:21Z

A recent change to respect dtypes in .clip() (#24458) introduced a decent overhead of ~2ms to the call:

$ asv compare v0.23.4 v0.24.0rc1 --only-changed
       before           after         ratio
     [04095216]       [fdc4db25]
     <v0.23.4^0>       <v0.24.0rc1^0>
+         120±1μs      2.04±0.04ms    17.01  series_methods.Clip.time_clip(1000)
+        936±40μs      3.09±0.02ms     3.31  series_methods.Clip.time_clip(100000)
+         111±3μs      2.05±0.08ms    18.55  series_methods.Clip.time_clip(50)

This PR cuts the overhead from ~2ms to ~0.6ms by keeping subset as a numpy array; it's entirely boolean regardless of underlying dtype, so a DataFrame only adds overhead here:

$ asv compare v0.24.0rc1 HEAD --only-changed -s
       before           after         ratio
     [fdc4db25]       [63c47c58]
     <v0.24.0rc1^0>       <clip>
-     2.04±0.04ms         759±20μs     0.37  series_methods.Clip.time_clip(1000)
-     3.09±0.02ms      1.46±0.04ms     0.47  series_methods.Clip.time_clip(100000)
-     2.05±0.08ms         724±20μs     0.35  series_methods.Clip.time_clip(50)

closes #xxxx
tests added / passed
passes git diff upstream/master -u -- "*.py" | flake8 --diff
whatsnew entry

codecov · 2019-01-11T22:51:29Z

Codecov Report

Merging #24735 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master   #24735      +/-   ##
==========================================
+ Coverage   92.39%   92.39%   +<.01%     
==========================================
  Files         166      166              
  Lines       52358    52362       +4     
==========================================
+ Hits        48374    48378       +4     
  Misses       3984     3984

Flag	Coverage Δ
#multiple	`90.81% <100%> (ø)`	⬆️
#single	`43.06% <0%> (-0.03%)`	⬇️

Impacted Files	Coverage Δ
pandas/core/generic.py	`96.63% <100%> (ø)`	⬆️

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 fdc4db2...63c47c5. Read the comment docs.

codecov · 2019-01-11T22:51:30Z

Codecov Report

Merging #24735 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master   #24735      +/-   ##
==========================================
+ Coverage   92.39%   92.39%   +<.01%     
==========================================
  Files         166      166              
  Lines       52378    52382       +4     
==========================================
+ Hits        48393    48398       +5     
+ Misses       3985     3984       -1

Flag	Coverage Δ
#multiple	`90.81% <100%> (ø)`	⬆️
#single	`42.9% <0%> (-0.01%)`	⬇️

Impacted Files	Coverage Δ
pandas/core/generic.py	`96.63% <100%> (ø)`	⬆️
pandas/util/testing.py	`88.13% <0%> (+0.09%)`	⬆️

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 f4458c1...f372c9e. Read the comment docs.

mroeschke · 2019-01-12T06:32:41Z

pandas/core/generic.py

+
+        with np.errstate(all='ignore'):
+            if upper is not None:
+                subset = self.values <= upper


Curious, does this place nice when self is the new Int64 dtype?

It appears to - let me know if there's a more extensive test you have in mind:

In [2]: s = pd.Series(range(5)).astype('Int64') In [3]: s.clip(1, 3) Out[3]: 0 1 1 1 2 2 3 3 4 3 dtype: Int64

WillAyd · 2019-01-12T19:18:45Z

pandas/core/generic.py

-            result = result.where(subset, lower, axis=None, inplace=False)
+        mask = isna(self.values)
+
+        with np.errstate(all='ignore'):


What's the point of this?

This is simply reverting back to what this block used to do; it's needed in the event values <= upper would otherwise raise a type error.

WillAyd · 2019-01-12T19:20:09Z

pandas/core/generic.py

+
+        with np.errstate(all='ignore'):
+            if upper is not None:
+                subset = self.values <= upper


Idiomatic approach here would now be to_numpy instead of .values

jreback · 2019-01-16T02:17:53Z

@qwhelan can you update

jreback · 2019-01-19T21:56:24Z

@qwhelan can you merge master and ping when passing

…s where possible

jreback · 2019-01-20T16:05:11Z

thanks @qwhelan

…s where possible (pandas-dev#24735)

jschendel added Performance Memory or execution speed performance Dtype Conversions Unexpected or buggy dtype conversions labels Jan 12, 2019

mroeschke reviewed Jan 12, 2019

View reviewed changes

WillAyd requested changes Jan 12, 2019

View reviewed changes

qwhelan force-pushed the clip branch from 63c47c5 to c5611dd Compare January 17, 2019 00:12

PERF: fix some of .clip() performance regression by using numpy array…

f372c9e

…s where possible

qwhelan force-pushed the clip branch from c5611dd to f372c9e Compare January 20, 2019 09:59

jreback added this to the 0.24.0 milestone Jan 20, 2019

jreback merged commit 2d47349 into pandas-dev:master Jan 20, 2019

Pingviinituutti pushed a commit to Pingviinituutti/pandas that referenced this pull request Feb 28, 2019

PERF: fix some of .clip() performance regression by using numpy array…

7a6306c

…s where possible (pandas-dev#24735)

Pingviinituutti pushed a commit to Pingviinituutti/pandas that referenced this pull request Feb 28, 2019

PERF: fix some of .clip() performance regression by using numpy array…

c1e2bc5

…s where possible (pandas-dev#24735)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

PERF: fix some of .clip() performance regression by using numpy arrays where possible #24735

PERF: fix some of .clip() performance regression by using numpy arrays where possible #24735

qwhelan commented Jan 11, 2019

codecov bot commented Jan 11, 2019

codecov bot commented Jan 11, 2019 •

edited

Loading

mroeschke Jan 12, 2019

qwhelan Jan 17, 2019

WillAyd Jan 12, 2019

qwhelan Jan 16, 2019

WillAyd Jan 12, 2019

qwhelan Jan 17, 2019

jreback commented Jan 16, 2019

jreback commented Jan 19, 2019

jreback commented Jan 20, 2019

PERF: fix some of .clip() performance regression by using numpy arrays where possible #24735

PERF: fix some of .clip() performance regression by using numpy arrays where possible #24735

Conversation

qwhelan commented Jan 11, 2019

codecov bot commented Jan 11, 2019

Codecov Report

codecov bot commented Jan 11, 2019 • edited Loading

Codecov Report

mroeschke Jan 12, 2019

Choose a reason for hiding this comment

qwhelan Jan 17, 2019

Choose a reason for hiding this comment

WillAyd Jan 12, 2019

Choose a reason for hiding this comment

qwhelan Jan 16, 2019

Choose a reason for hiding this comment

WillAyd Jan 12, 2019

Choose a reason for hiding this comment

qwhelan Jan 17, 2019

Choose a reason for hiding this comment

jreback commented Jan 16, 2019

jreback commented Jan 19, 2019

jreback commented Jan 20, 2019

codecov bot commented Jan 11, 2019 •

edited

Loading