Incorrect lineheight on high-resolution non-quartz devices

[cpsievert]

In the example below, note that without showtext enabled, all three plots below produce plots with consistent lineheight, but with showtext, the last two produce a much smaller lineheight

library(showtext)
library(ggplot2)

showtext::showtext_auto()

# lineheight appears ok
tmp <- tempfile(fileext = ".png")
png(tmp, type = "quartz", res = 220, width = 600, height = 400, units = "px")
print(qplot(1:10))
dev.off()
file.show(tmp)

# lineheight is much smaller (with showtext enabled)
tmp <- tempfile(fileext = ".png")
png(tmp, type = "cairo", res = 220, width = 600, height = 400, units = "px")
print(qplot(1:10))
dev.off()
file.show(tmp)

# lineheight is much smaller (with showtext enabled)
tmp <- tempfile(fileext = ".png")
Cairo::CairoPNG(tmp, res = 220, width = 600, height = 400, units = "px")
print(qplot(1:10))
dev.off()
file.show(tmp)

[yixuan]

The second is easy to fix. You just need to call showtext::showtext_opts(dpi = 220) before plotting.

The CairoPNG device is a bit complicated, and I'm thinking of ways to fix it.

[cpsievert]

Great, thank you. Should knitr be setting showtext_opts(dpi = opts_current$get("dpi")) (when fig.showtext is TRUE)?

(cc @yihui)

[cpsievert]

Thanks again for your response @yixuan! Just curious, what are the prospects of a fix for CairoPNG (i.e., do you think it'll ever happen)? Totally ok to be honest if it's unlikely, just want to know what to expect.

[yixuan]

@cpsievert This is a bit more complicated than I thought. What showtext needs to know is whether a specific graphics device is bitmap-based (e.g. PNG) or vector-based (PDF), and the current implementation of showtext uses a guess based on the device name. But for Cairo devices, they share the same device name "Cairo", so showtext does not know whether it is a PNG or PDF or something else.

One possible solution is to ask the user to provide such information. This makes dealing with fonts easier, but is just "less automatic". Or I need to read the C code of the Cairo package to see if there is anything hidden in the internals that can help me distinguish different types.

[cpsievert]

Oh, good to know. And yea, maybe it's worth contributing to Cairo itself to add a filepath attribute to .Device (then you could try to infer based on the file extension), which is what grDevices does, for example:

> png("foo.png")
> .Device
[1] "quartz_off_screen"
attr(,"filepath")
[1] "foo.png"

I think we could just add that information here (I'm not very familiar with C, so I don't immediately see how that'd be done, but perhaps you'd know how to?)

[yixuan]

I think I have fixed this issue in the master branch. Can you give it a try?

[cpsievert]

Yes, CairoPNG looks much better now with showtext_opts(dip = 220), thank you!

I also noticed in 39bcd28 that you've added ragg awareness, but I'm noticing some degradation in the font rendering. It seems like it'd be better to do nothing for ragg as it has its own API for rendering custom fonts (via systemfonts::register_font()).

library(showtext)
library(ggplot2)
showtext::showtext_auto()
showtext::showtext_opts(dpi = 220)
tmp <- tempfile(fileext = ".png")
ragg::agg_png(tmp, res = 220, width = 600, height = 400, units = "px")
print(qplot(1:10))
dev.off()
file.show(tmp)

[yixuan]

I think this is a bug of ragg. You can find that if you change the bit size to be 16 in agg_png, the output image is correct.

library(showtext)
library(ggplot2)
showtext::showtext_auto()
showtext::showtext_opts(dpi = 220)
tmp <- tempfile(fileext = ".png")
ragg::agg_png(tmp, res = 220, width = 600, height = 400, units = "px", bitsize = 16)
print(qplot(1:10))
dev.off()
file.show(tmp)

I'll create an example and report there.

[cpsievert]

Thanks again for the CairoPNG fix (and for submitting that fix to ragg)! 🍻

Looks as though this is working in all the cases I care about now, so I'm happy to close.

Just curious, do you have a timeline for a new showtext CRAN release?

[yihui]

@cpsievert I just made the change in knitr to pass the dpi option to showtext_opts(). Thanks!