The very first update in the series is about GNOME Shell and Mutter. I’ve been increasingly involved with the development of those two core components of GNOME, and recently this has been the focus of my development time.
Fortunately, Endless allows me to use part of my work time to improve it. Naturally, I prioritize my upstream work considering what will impact Endless OS the most. So far, that lead to a series of very nice improvements to Mutter and GNOME Shell.
Most of my work time dedicated to GNOME Shell was oriented to performance and cleanup. At Endless, we have a modified GNOME Shell that constantly needs to be rebased. Since I’m taking care of these rebases now, it makes sense for me to also make myself familiar with the vanilla GNOME Shell codebase.
I’ll start with the work that makes me the proudest: removing the Shell.GenericContainer class.
First, a bit of history.
For that reason, GNOME Shell developers created ShellGenericContainer: an actor that sends signals for various virtual functions. Because GJS supports signals, that worked well.
There are a few problems with that approach though:
- Signals are slow, and should not be used on hot paths like layouting or rendering;
Thanks to the fantastic work by Jasper St. Pierre, GJS now supports overriding virtual functions. And that made Shell.GenericContainer obsolete. So I spent quite some time untangling it from GNOME Shell, and results were positive:
This is now merged and will be available with GNOME Shell 3.32, to be released on March 2019.
Improvements to the texture cache
After various investigations, another potential improvement that showed up was on StTextureCache. Textures (icons, image files, etc) are cached in GNOME Shell by StTextureCache, and that happened by keeping a ClutterTexture object alive.
That turned out to be a problem.
ClutterTexture is deprecated. Clutter has a new interface for drawing the contents of an actor: ClutterContent. It does not necessarily make the code faster, but it allows different actors to share a single ClutterContent without having to override ClutterActor.paint(). In other words, it is a nice and sane abstraction layer to control what an actor is drawing.
Unexpectedly, it made a small but noticeable difference! Icons are now slightly faster to load, but the most visible impact was in the startup animation.
I did not know how fun and exciting compositors could be. It definitely is a new passion of mine, working on Mutter! So much has happened that it’ll be hard to summarize.
During last year’s GUADEC, Jonas Ådahl worked on a Meson port of Mutter. After a series of reviews, and a few follow-up fixes, it reached almost complete feature parity with Autotools – the only exception being installed tests.
So I went ahead and added installed tests to the Meson build too.
And also removed Autotools.
Naturally, builds are much faster now. Saving us a few minutes per day.
Wayland vs X11
Another area that was interesting to work on was untangling X11-specific code from Wayland, and vice-versa. There are a handful of developers working on that already, and I had my fair share in better splitting X11 and Wayland code paths in Mutter.
Specifically, I worked on splitting X11-specific code from MetaWindowActor into subclasses. Mutter already handles different surfaces correctly; on X11 sessions, all surfaces are MetaSurfaceActorX11, and under Wayland, MetaSurfaceActorWayland.
MetaWindowActor has now the same split: Wayland windows have a MetaWindowActorWayland associated, while X11 windows have MetaWindowActorX11.
Interestingly, XWayland windows are X11 windows with a Wayland surface. You can check that using GNOME Shell’s Looking Glass:
There’s a lot more happening in this front, but I’ll spare the words for now. You’ll hear more about it in the future (and not necessarily from me).
More recently, I’ve been experimenting with the Cogl journal and ironing out a few bugs that are preventing a completely CPU-side picking implementation.
Picking is the process to figure out which element is beneath the cursor. There are two big approaches: geometry-based, and color-based. On games, the latter is the usual approach: each object in the scene is drawn with a plain color, and the final image is read to find out the color beneath a point. Geometry-based picking is what browsers usually do, and it’s basically math around rectangles.
Clutter uses color-based picking, but has a nice feature around that: a journal that tracks drawing operations and, under some conditions, hits an optimized path and does geometry-based picking. This is interesting for Mutter and GNOME Shell because it avoids sending draw operations to the GPU unecessarily when picking, reducing resource usage.
Unfortunately, due to various bugs and implementation details, we do not hit this optimization, causing GPU commands to be issued when they could be avoided.
There’s much more that happened, so I will probably do a part 2 of this article soon. But those are big points already, and the post is becoming lengthy.
Many of these experiments and investigations already landed, and will be available with GNOME 3.32. This is all valuable work that is partially sponsored by my employer, Endless, and I’m happy to keep working on it!