Skip to content

PERF: Defer tick materialization during Axes init/clear#31525

Open
eendebakpt wants to merge 7 commits intomatplotlib:mainfrom
eendebakpt:perf/lazy-axis-init
Open

PERF: Defer tick materialization during Axes init/clear#31525
eendebakpt wants to merge 7 commits intomatplotlib:mainfrom
eendebakpt:perf/lazy-axis-init

Conversation

@eendebakpt
Copy link
Copy Markdown
Contributor

@eendebakpt eendebakpt commented Apr 18, 2026

PR summary

The performance if matplotlibs ticks is a bottleneck in various plots. See for example the discussions and references in #5665, #31012, #29594.

In this PR we prevent materialization of the _LazyTickList when there are no ticks created yet. With the tick-materialization cascade gone from Axes.__clear, the spine transforms the cascade used to install as a side effect are installed explicitly at the end of __clear.

Benchmark results (updated):

nit_grid 8x8:                  [main] 177 ms ± 2 ms   -> [branch] 112 ms ± 24 ms:  1.58x faster
clear_grid 8x8:                 [main] 181 ms ± 2 ms   -> [branch] 139 ms ± 22 ms:  1.31x faster
reuse_axes 8x8:                 [main] 154 ms ± 1 ms   -> [branch] 53.8 ms ± 0.4 ms: 2.86x faster
fig100 clear+plot1000+legend:   [main] 9.45 ms ± 2.4 ms -> [branch] 3.28 ms ± 0.02 ms: 2.88x faster

Geometric mean: 2.03x faster
Benchmark script
# /// script
# requires-python = ">=3.10"
# dependencies = ['matplotlib', 'numpy', 'pyperf']
# ///
"""pyperf micro-benchmarks for matplotlib axis/tick init+clear cost.

"""
import pyperf

setup = """
import matplotlib
matplotlib.use("Agg")
# matplotlib.use("QtAgg")  # snap/glibc mismatch; use offscreen:
# import os; os.environ.setdefault("QT_QPA_PLATFORM", "offscreen")

import numpy as np
import matplotlib.pyplot as plt

GRID = 8
rng = np.random.default_rng(0)
x1000 = np.arange(1000)
y1000 = rng.standard_normal(1000)

# Pre-create reusable figure for clear_grid / reuse_axes cases.
_fig_clear = plt.figure()
_fig_grid, _axs_grid = plt.subplots(GRID, GRID)
_axs_flat = _axs_grid.ravel()

# Warmup — first call pays font-cache / backend-init costs.
plt.close(plt.subplots(GRID, GRID)[0])
"""

runner = pyperf.Runner()

# Fresh figure each iter — full Axes.__init__ cost for an 8x8 grid.
runner.timeit(
    name="init_grid 8x8",
    stmt="fig, axs = plt.subplots(GRID, GRID); plt.close(fig)",
    setup=setup,
)

# Reuse one Figure, clear + re-populate with an 8x8 grid.
runner.timeit(
    name="clear_grid 8x8",
    stmt="_fig_clear.clear(); _fig_clear.subplots(GRID, GRID)",
    setup=setup,
)

# Iterate ax.clear() across an existing 8x8 grid.
runner.timeit(
    name="reuse_axes 8x8",
    stmt="[ax.clear() for ax in _axs_flat]",
    setup=setup,
)

# Figure num=100: clear + plot 1000-point line + legend. Reuses the same
# numbered figure across iterations so only clear+plot+legend is measured.
runner.timeit(
    name="fig100 clear+plot1000+legend",
    stmt=(
        "fig = plt.figure(num=100);"
        " fig.clear();"
        " ax = fig.add_subplot();"
        " ax.plot(x1000, y1000, label='y');"
        " ax.legend()"
    ),
    setup=setup,
)

Closes #23771.

AI Disclosure

Claude was used in identifying performance bottlenecks related to tick creation. Initially the goal was to create tick collections (as described in one of the references), but this approach seems to be a small change with large impact.

PR checklist

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@eendebakpt eendebakpt force-pushed the perf/lazy-axis-init branch from 5708dc8 to 3ac8224 Compare April 18, 2026 12:21
@eendebakpt eendebakpt marked this pull request as ready for review April 18, 2026 16:06
Copy link
Copy Markdown
Member

@timhoffm timhoffm left a comment

Choose a reason for hiding this comment

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

Thanks for the PR. The speedup is impressive, and the added complexity (rc caching) is bearable.

Strategically, I would like to move away from single-tick handling, but in the mean time this is a reasonable improvement.

Comment thread lib/matplotlib/axis.py
Comment thread lib/matplotlib/axis.py Outdated
Comment thread lib/matplotlib/axis.py Outdated
Comment thread lib/matplotlib/axis.py Outdated
@eendebakpt
Copy link
Copy Markdown
Contributor Author

Strategically, I would like to move away from single-tick handling, but in the mean time this is a reasonable improvement.

Having tick collections is indeed the way to go. This change is orthogonal as it avoids some tick operations altogether. (but maybe if ticks are really fast that would not matter)

Comment thread lib/matplotlib/axis.py Outdated
Comment thread lib/matplotlib/axis.py
Comment thread lib/matplotlib/axis.py Outdated
Comment thread lib/matplotlib/axis.py Outdated
Comment thread lib/matplotlib/axis.py Outdated
Comment thread lib/matplotlib/axis.py Outdated
Comment thread lib/matplotlib/axes/_base.py Outdated
eendebakpt and others added 3 commits April 23, 2026 21:36
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@eendebakpt eendebakpt force-pushed the perf/lazy-axis-init branch from a3f1be1 to d0ed04c Compare April 23, 2026 19:38
Comment thread lib/matplotlib/axis.py
Comment on lines +587 to +615
# instance._get_tick() can itself try to access the majorTicks
# attribute (e.g. in certain projection classes which override
# e.g. get_xaxis_text1_transform). To avoid infinite recursion,
# bind the attribute to an empty list before calling _get_tick().
# _get_tick() may also call reset_ticks(), which pops the attribute
# from the instance dict; the final setattr below re-binds the
# (now non-empty) list so subsequent accesses skip the descriptor.
attr = 'majorTicks' if self._major else 'minorTicks'
tick_list = []
setattr(instance, attr, tick_list)
# Build the Tick (and its sub-artists) under the rcParams captured
# at the last Axis.clear() so that a lazily-materialized Tick
# matches an eager (pre-lazy) Tick (see Axis._tick_rcParams).
with _rc_context_raw(instance._tick_rcParams):
tick = instance._get_tick(major=self._major)
# Re-apply any set_tick_params() overrides to the fresh Tick.
# Subclasses of Axis (e.g. the SkewXAxis in the skewt gallery
# example) sometimes override _get_tick() without forwarding
# _{major,minor}_tick_kw; calling _apply_params() here guarantees
# those overrides still take effect, matching the pre-lazy
# behaviour where the first tick was materialized eagerly and
# updated in place by set_tick_params().
tick_kw = (instance._major_tick_kw if self._major
else instance._minor_tick_kw)
if tick_kw:
tick._apply_params(**tick_kw)
tick._configure_for_axis(instance)
tick_list.append(tick)
setattr(instance, attr, tick_list)
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Suggested change
# instance._get_tick() can itself try to access the majorTicks
# attribute (e.g. in certain projection classes which override
# e.g. get_xaxis_text1_transform). To avoid infinite recursion,
# bind the attribute to an empty list before calling _get_tick().
# _get_tick() may also call reset_ticks(), which pops the attribute
# from the instance dict; the final setattr below re-binds the
# (now non-empty) list so subsequent accesses skip the descriptor.
attr = 'majorTicks' if self._major else 'minorTicks'
tick_list = []
setattr(instance, attr, tick_list)
# Build the Tick (and its sub-artists) under the rcParams captured
# at the last Axis.clear() so that a lazily-materialized Tick
# matches an eager (pre-lazy) Tick (see Axis._tick_rcParams).
with _rc_context_raw(instance._tick_rcParams):
tick = instance._get_tick(major=self._major)
# Re-apply any set_tick_params() overrides to the fresh Tick.
# Subclasses of Axis (e.g. the SkewXAxis in the skewt gallery
# example) sometimes override _get_tick() without forwarding
# _{major,minor}_tick_kw; calling _apply_params() here guarantees
# those overrides still take effect, matching the pre-lazy
# behaviour where the first tick was materialized eagerly and
# updated in place by set_tick_params().
tick_kw = (instance._major_tick_kw if self._major
else instance._minor_tick_kw)
if tick_kw:
tick._apply_params(**tick_kw)
tick._configure_for_axis(instance)
tick_list.append(tick)
setattr(instance, attr, tick_list)
# tick list materialization logic:
# 1. create a preliminary empty tick list on the instance
# 2. create the first tick via instance._get_tick().
# - instance._get_tick() can itself try to access the majorTicks attribute
# (e.g. in certain projection classes which override e.g.
# get_xaxis_text1_transform). Therefore step 1 is needed to avoid
# infinite recursion.
# - Creation is done under the rcParams captured at the last Axis.clear()
# so that properties relfect the state of Axis creation/clear.
# 3. apply set_tick_params() and axis config to the tick
# - Subclasses of Axis (e.g. the SkewXAxis in the skewt gallery example)
# sometimes override _get_tick() without forwarding
# _{major,minor}_tick_kw. Calling _apply_params() here guarantees
# those overrides still take effect, matching the pre-lazy
# behaviour where the first tick was materialized eagerly and
# updated in place by set_tick_params().
# 4. bind the final tick list to the instance
# - _get_tick() may also call reset_ticks(), which pops the attribute
# from the instance dict; the final setattr is the eventual assignment
# so that subsequent accesses skip the descriptor.
attr = 'majorTicks' if self._major else 'minorTicks'
setattr(instance, attr, []) # preliminary empty tick list
with _rc_context_raw(instance._tick_rcParams):
tick = instance._get_tick(major=self._major)
if tick_kw:
tick._apply_params(**tick_kw)
tick._configure_for_axis(instance)
tick_list = [tick]
setattr(instance, attr, tick_list)

Rewritten to make the complex logic as clear as possible. Motivation

  • explain the complete logic first
    • structured by main steps and indentend additional explanation
  • have the code all together - if you know the background from the comment, seeing and understanding the actual code in one go is easier than distributed commands with long comment in between

Note: I've reverted from having a tick_list all around. I had missed that it has to be re-bound anyway. Under that condition it's IMHO logically simpler to have a temporary (unnamed) empty list, and create and bind the actual list at the end.

# layout or draw. Spine.__init__ installs self.axes.transData as
# a placeholder; the real blended transform is set by
# Spine.set_position via _ensure_position_is_set(). Historically
# this fired as a side effect of tick materialization during
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Suggested change
# this fired as a side effect of tick materialization during
# the spine position was set as a side effect of tick materialization during

wording

Comment on lines +1439 to +1440
# layout or draw. Spine.__init__ installs self.axes.transData as
# a placeholder; the real blended transform is set by
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

This is some spine-internal arcane knowledge. Axes should not need to know about this. Can we add some sort of spine._is_positioned flag?

I'm also not fully following the logic: It seems sping._ensure_position_is_set() does not care about spine._transform, but only checks spine._position is that a shortcoming on _ensure_position_is_set()?

Comment thread lib/matplotlib/axis.py
Comment on lines +609 to +613
tick_kw = (instance._major_tick_kw if self._major
else instance._minor_tick_kw)
if tick_kw:
tick._apply_params(**tick_kw)
tick._configure_for_axis(instance)
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Optional: We could move the whole tick configuration into the function

Suggested change
tick_kw = (instance._major_tick_kw if self._major
else instance._minor_tick_kw)
if tick_kw:
tick._apply_params(**tick_kw)
tick._configure_for_axis(instance)
tick._configure_for_axis(instance, self._major)

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

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[Bug]: matplotlib.pyplot.clf is very slow

3 participants