handle retina/hidpi displays with metadata

[xzackli]

This implements the handling of hidpi displays for PyPlot (#902) like IPython does. During display, it sends metadata with the PNG images that specify half-height and width, with these dimensions read from the PNG header. See the issue for links to how IPython does it.

To complete this feature, it might be nice to add a PyPlot.retina(true) command which doubles the inline DPI and sets IJulia.retina[].

Example Usage

using IJulia
IJulia.retina[] = true

using PyPlot
plt.plot( sin.(1:0.1:10) )

[stevengj]

I don't think we should throw an error here; we should let the front end decide what to do with malformed PNG data (i.e. just don't set any metadata).

[xzackli]

Makes sense, I'll make this change.

[stevengj]

I don't understand why this can't be done by the front-end display. It seems weird to do this in the kernel, because the kernel doesn't know anything about the display, and indeed the same kernel may be talking to multiple displays in theory.

[stevengj]

Is there a Jupyter/Jupyterlab issue for this as a front-end feature?

[xzackli]

Is there a Jupyter/Jupyterlab issue for this as a front-end feature?

I've made an issue at jupyterlab/jupyterlab#9440 to see what people think. Maybe it's historical.

[tfiers]

A short alternative, for if you only care about retina'ing PyPlot Figures, and don't mind that matplotlib's (figwidth, figheight) is not the exact same as the actual png's:

using IJulia, PyPlot

function IJulia.metadata(x::PyPlot.Figure)
    w, h = (x.get_figwidth(), x.get_figheight()) .* x.get_dpi()
    return Dict("image/png" => Dict("width" => w/2, "height" => h/2))
end