Skip to content

PERF: Sticky edges speedup#31244

Open
scottshambaugh wants to merge 5 commits intomatplotlib:mainfrom
scottshambaugh:sticky_edges_speedup
Open

PERF: Sticky edges speedup#31244
scottshambaugh wants to merge 5 commits intomatplotlib:mainfrom
scottshambaugh:sticky_edges_speedup

Conversation

@scottshambaugh
Copy link
Contributor

@scottshambaugh scottshambaugh commented Mar 6, 2026 •
edited
Loading

PR summary

This speeds up the sticky edges handling in autoscale_view. Most artists do not have sticky edges, so:

  • Sticky edges are now lazily constructed
  • Skip non-sticky children for checks, like ticks, spines, figure text, etc
  • We do not add empty sticky edges to x_sticky_lists and y_sticky_lists, which saves us from having to concatenate a large number of empty arrays.

This completely removes the time spent in the circled code paths here, about 60% of _unstale_view_lim and 15% of the total run time in this test script.
image

Profiling test script:

import numpy as np
import matplotlib.pyplot as plt

fig, ax = plt.subplots()

n = 1000
rng = np.random.default_rng(42)
x = rng.random(n)
y = rng.random(n)
for i in range(n):
    ax.scatter(x[i], y[i])

fig.canvas.draw()
plt.close(fig)

AI Disclosure

PR checklist

@scottshambaugh scottshambaugh marked this pull request as draft March 6, 2026 06:04
@scottshambaugh scottshambaugh marked this pull request as ready for review March 6, 2026 06:17
timhoffm
timhoffm reviewed Mar 6, 2026
View reviewed changes
lib/matplotlib/axes/_base.py Outdated
Comment on lines +3019 to +3020
# Use ._children and ._sticky_edges directly, because most extra
# artists in .get_children() (spines, titles, etc.) never have
Copy link
Member

@timhoffm timhoffm Mar 6, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We should formalize the concept of different categories of children. I would distinguish

  • data children: basically _children
  • auxiliary children: all supporing artists that are needed to draw the Axes itself.
    Names and exact definitions to be bikeshed.

For a start here, I would create Axes._get_data_children() as a private method, so that we can better document what that is and dont have to go into the private _childrenof the shared axes. That would also make the comment part on_children` unnecessary, because what you are writing there is expressed through the code concept.

Copy link
Contributor Author

@scottshambaugh scottshambaugh Mar 6, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I like that, updated

lib/matplotlib/axes/_base.py
x_sticky_lists = []
for ax in x_shared:
for artist in (*ax._children, *ax._axis_map.values()):
sticky_edges = artist._sticky_edges
Copy link
Member

@timhoffm timhoffm Mar 6, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Slighty worried that this is unnecessary micro-optimization.

Image

Do 20ns per artist really make a difference? If not, I'd rather stay within the public API bundaries

Copy link
Contributor Author

@scottshambaugh scottshambaugh Mar 6, 2026 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It does show up in the results - just the .sticky_edges property accesses are the circled bits in the baseline profiling here. Because autoscale_view is called for each scatter call in the script, this inner loop is O(n^2) with the number of artists. We could defer autoscaling or be smarter about which artists are on the boundary, but that would be a much bigger change.

image

timhoffm
timhoffm approved these changes Mar 6, 2026
View reviewed changes
@scottshambaugh
Copy link
Contributor Author

scottshambaugh commented Mar 6, 2026 •
edited
Loading

Rebased, hadn't realized I already made a portion of these changes in #31004

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@timhoffm timhoffm timhoffm approved these changes

Assignees

No one assigned

Labels

Performance topic: axes

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@scottshambaugh @timhoffm