cartopy: contourf method significantly slower than basemap

Description

I’m migrating my codes from Basemap to Cartopy. Generally Cartopy is always faster, however I find contourf method can take me seconds. Here’s my example code.

import time

import cartopy
import matplotlib.pyplot as plt
import numpy as np
import scipy.ndimage as sn
from mpl_toolkits.basemap import Basemap


#-------------------
#  Create data
#-------------------
x, y = np.mgrid[60:140:0.25, 0:45:0.25]
r = np.random.randint(0, 255, size=(320, 180), dtype='uint8')
r = sn.filters.gaussian_filter(r, (3, 3), mode='constant')

#-------------------
#  Cartopy
#-------------------
ax = plt.axes(projection=cartopy.crs.LambertConformal(
    central_longitude=100.0, central_latitude=35.0))
ax.set_extent((60,140,0,45))
for step in (20, 10, 5):
    tic = time.time()
    col = ax.contourf(x, y, r, transform=cartopy.crs.PlateCarree(),
        levels=np.arange(100,161,step))
    toc = time.time()
    points = sum(len(e.get_paths()) for e in col.collections)
    print('[Cartopy] Step: {} Points: {}  Time: {:.0f}ms'.format(
        step, points, (toc - tic) * 1000))


#-------------------
#  Basemap
#-------------------
_map = Basemap(projection='lcc', resolution=None,
    lon_0=100.0, lat_0=35.0, llcrnrlat=0,
    urcrnrlat=45, llcrnrlon=60, urcrnrlon=140)
for step in (20, 10, 5):
    tic = time.time()
    col = _map.contourf(x, y, r, levels=np.arange(100,161,step))
    toc = time.time()
    points = sum(len(e.get_paths()) for e in col.collections)
    print('[Basemap] Step: {} Points: {}  Time: {:.0f}ms'.format(
        step, points, (toc - tic) * 1000))

The result:

[Cartopy] Step: 20 Points: 123  Time: 729ms
[Cartopy] Step: 10 Points: 273  Time: 1433ms
[Cartopy] Step: 5 Points: 528  Time: 2887ms
[Basemap] Step: 20 Points: 123  Time: 118ms
[Basemap] Step: 10 Points: 273  Time: 151ms
[Basemap] Step: 5 Points: 528  Time: 179ms

which means Cartopy can be up to 16x slower on my old laptop. After some profiling work I find the culprit is https://github.com/SciTools/cartopy/blob/master/lib/cartopy/mpl/geoaxes.py#L1505, it seems this step is to calculate new datalim and autoscale the extent. This is how far I can get to since I don’t have enough time to dig in.

• Why does this step take so much time?
• As a quick workaround, can we just check if the extent is already set (which is the most scenario imo) before autoscaling?

Environment

• OS: Windows 10
• cartopy: 0.17.0
• basemap: 1.2.0
• matplotlib: 3.0.2
• shapely: 1.6.4.post1

Thanks!