High-DPI displays and Linux

I do t know about the "some", but that is what that Apple engineer says:

Which leads to the problem of centering. I wish to center a bitmap image within a button's border. The image is 101 logical points high, and the border is 200 logical points high. With a 2x scale factor, I can center absolutely, and still be aligned to device pixels. With a 1x, 1.25, 1.33, etc. scale factor, centering will align me on a partial pixel, which looks like crap. So I have to round. Which way? If the goal is "make it look good," then the answer is "whichever way looks good," which depends on the visual style of the bezel, i.e. whether the top or bottom has more visual weight. So now we need hinting.

And that's where things start to get really nasty. In order to make things look good at arbitrary resolutions, we want to round to device pixels. But the rounding direction is not a local question! Consider what happens if we have two visual elements abutting in logical coordinates, and they round in opposite directions: now there's a device pixel between them. That's very visible: you get a pixel crack! So you have to coordinate rounding.

Question: even with hinting, is it actually possible to make a font where two glyphs align perfectly without white space between them?

WPF is a good example of a framework that attempted resolution independence and encountered this problem. Initially it has the "SnapsToDevicePixels" property, which triggers rounding behavior at draw time. But draw time is too late, because of the "abutting elements rounding in opposite directions" problem. So they introduced the "UseLayoutRounding" property, which does...something. And the guidance is basically "turn it on and see if it helps, if not, disable it." Great.

I think a way out of this is way higher DPI with some blurring as the last step. That's what print does. At 1000 or 2500 DPI nobody will notice aliasing anymore, and he blurring would take care of small cracks between objects that supposedly are abutting, make cases where two objects that shouldn't accidentally do overlap less obnoxious, etc.

Another interesting angle is that it’s not only application software we need to change, but also the hardware drivers are not quite there yet:

I have a Dell UP2414Q (3840x2160 resolution, driven via DisplayPort 1.2) connected to a nVidia GTX 660 card, which was one of the cheapest ones that support DP 1.2.

With the proprietary nvidia driver, I need to manually edit the xorg configuration file to have the correct modes and most importantly disable XRandR in favor of Xinerama.

This in turn breaks e.g. GNOME shell on Fedora 20 (without RandR, you’ll just get an exception in your syslog), and in general prevents plenty of use-cases (e.g. redshift for controlling display brightness, or changing rotation settings without restarting X11).

The reason for having to disable RandR is that there is not currently a standard way on how to represent multi-stream transport (MST) connections, and 4K displays require 2 streams (1920x1080 each) at the same time. With RandR enabled, what you’ll see is 2 connected outputs, and all applications will treat them as such, even though you have only one monitor connected.

Fixing this requires changes in RandR (i.e. the X server) and each driver. AFAIK, on the intel driver this should work, on nouveau there’s work under way, no clue about the proprietary nvidia driver.

Not bad catch, I remember the picture, certainly nineties.