Switch to requiring Pillow rather than having our own png wrapper?

[anntzer]

It looks like Pillow's png handling (which is hand-written https://github.com/python-pillow/Pillow/blob/master/src/PIL/PngImagePlugin.py, not based on libpng) is faster than matplotlib's, both for writing and for reading:

from time import perf_counter
import matplotlib as mpl; mpl.use("agg")
from matplotlib import pyplot as plt
import numpy as np
from PIL import Image

fig, ax = plt.subplots()
fig.canvas.draw()
t = np.asarray(fig.canvas.renderer._renderer)
for writer in ["mpl", "pil"]:
    print(f"write with {writer}")
    start = perf_counter()
    for _ in range(100):
        mpl.image.imsave("/tmp/test.png", t,
                         pil_kwargs={"mpl": None, "pil": {}}[writer])
    print("write", perf_counter() - start)
    start = perf_counter()
    for _ in range(100):
        mpl.image.imread("/tmp/test.png")
    print("mpl read", perf_counter() - start)
    start = perf_counter()
    for _ in range(100):
        with Image.open("/tmp/test.png") as image:
            mpl.image.pil_to_array(image)
    print("pil read", perf_counter() - start)

yields, for me:

write with mpl
write 1.5242545299988706
mpl read 0.9935761439992348
pil read 0.46841289500298444
write with pil
write 1.400021340996318
mpl read 0.9913013100012904
pil read 0.4612834110012045

#15160 reminded me that it's not always trivial to get C-level dependencies set up before building (though that specific issue is about freetype). We already vendor agg and qhull, and have an option to download freetype (via local_freetype/MPLLOCALFREETYPE), so the sole remaining dependency that always needs to be provided by the builder is libpng.

If we just depended (as an install_requires) on Pillow for png handling instead, we would be able to drop that dependency (together with a bunch of C code). Pillow has wheels for all three major OSes, supports PyPy, is actively developed, and has been around for a very long time.

Thoughts on making the switch?

[jklymak]

The idea seems reasonable. Is the performance at all dependent on the complexity of the plots, or the size of the pngs?

[anntzer]

I don't have the numbers yet, but can set up more benchmarking.

[tacaswell]

👍 if this is going to make both the build process simpler and get us a factor of 2 in writing.

[anntzer]

in reading, not in writing.

[tacaswell]

oh, derp, I apparently can not read, but even a small speed up, in conjunction with the easier install is still a win.

[anntzer]

Right now this is stuck on a few upstream issues (python-pillow/Pillow#3041, imageio/imageio#475); may end up needing to vendor some code from imageio with some more fixes to fix that.

On the other hand the (unpublished) draft patch I have does indeed make building on Windows much easier.