constrained-layout gives wrong results when explicitly equal width ratios are set

[anntzer]

Bug report

Bug summary

When passing width_ratios = [1, 1, 1] instead of not passing it at all (which should be synonymous), constrained_layout sometimes gives incorrect results.

Code for reproduction

import matplotlib as mpl
from matplotlib import pyplot as plt, gridspec, testing

mpl.rcdefaults()
mpl.style.use(["classic", "_classic_test_patch"])
mpl.testing.set_font_settings_for_testing()

def example_plot(ax, fontsize=12):
    ax.plot([1, 2])
    ax.locator_params(nbins=3)
    ax.set_xlabel('x-label', fontsize=fontsize)
    ax.set_ylabel('y-label', fontsize=fontsize)
    ax.set_title('Title', fontsize=fontsize)

fig = plt.figure(constrained_layout=True)
gs = gridspec.GridSpec(3, 3, figure=fig)

ax1 = fig.add_subplot(gs[0, 0])
ax2 = fig.add_subplot(gs[0, 1:])
ax3 = fig.add_subplot(gs[1:, 0:2])
ax4 = fig.add_subplot(gs[1:, -1])

for ax in fig.axes:
    example_plot(ax)

fig.savefig("/tmp/cl1.png")

fig = plt.figure(constrained_layout=True)
gs = gridspec.GridSpec(3, 3, figure=fig,
                       width_ratios=[1, 1, 1], height_ratios=[1, 1, 1])

ax1 = fig.add_subplot(gs[0, 0])
ax2 = fig.add_subplot(gs[0, 1:])
ax3 = fig.add_subplot(gs[1:, 0:2])
ax4 = fig.add_subplot(gs[1:, -1])

for ax in fig.axes:
    example_plot(ax)

fig.savefig("/tmp/cl2.png")

plt.show()

Actual outcome

cl1.png

cl2.png

Clearly in cl2.png the widths are not in the ratios 1:1:1 but (approximately) 1:2:1.

This is likely a floating point error; it is extremely sensitive to conditions, in particular the call to set_font_settings_for_reproducibility (which just sets some font-related rcs) is necessary to observe this.

Expected outcome

twice the same image

Matplotlib version

• Operating system: archlinux
• Matplotlib version: master
• Matplotlib backend (print(matplotlib.get_backend())): agg
• Python version: 38
• Jupyter version (if applicable): no
• Other libraries:

[jklymak]

Yes agreed. The constraints are probably a bit flakey here and find the wrong solution sometimes. Since this is pretty reproducible I should be able to track down.

Note this is a pretty nasty subplot layout. I wouldn’t block any releases on this use case. If you wanted to set the tolerance really high for that test for the other PR that would probably be acceptable.

[anntzer]

I don't think the other PR is particularly urgent either, I'm fine just waiting for you on this :-)