I'm hoping this is something that improved in 3.0.0, because using Audacity is becoming increasingly impractical. I was just trying to listen to 10 tracks one by one yesterday and even just muting/soloing a track would take about a second; it was horrible. Shrinking the window helps, but it's silly to have a tiny window and scroll around when you have a 4K display.
I kind of suspect this to be a windowing toolkit issue, and originally I thought it was macOS-specific (since I first noticed when I tried to recommend it to friends who use that), but ever since I moved to a 4K screen on Linux I noticed it's bad here too...
The Mac story is here: https://forum.audacityteam.org/viewtopic.php?f=47&t=111105
The workaround is, well, not using HiDPI mode which is silly. But Audacity doesn't even seem to support HiDPI mode on Linux properly to begin with (the entire UI is small), so that's not it per se... I guess wxGTK just struggles when it has to draw 4x as many pixels?
Apparently it's bad on Windows too: https://forum.audacityteam.org/viewtopic.php?t=111857
This is because the state machine that drives the draw context is synchronous. Every line, pixels (or anything, really) takes some CPU cycles.
Newer toolkits (QtQuick2, GTK4, Skia-based-ones) are hardware accelerated. They are slower of standard DPI, but much faster on HiDPI.
So, yes, they cannot be as snappy on HiDPI, the number of pixel gets bottlenecked by the single thread performance of the CPU. Those barely grow year-over-year, while the number of pixel went up a cliff overnight.
> which tells me this isn't a simple pixel pushing problem but rather something more profound in Audacity.
That's probably because of the cursor animation (when you play). Audacity has to redraw the entire track widget each time the needle moves unless they implement special double buffering + special damage area slices. By default, Qt/GTK/wx are too dumb to understand layers properly. They will redraw the entire rectangle if you try to stack things with any transparency or non-rectangle clips. Even with proper double buffering, combining both frame buffers is still done on the CPU. That's 2x4x{width}x{height}x{fps} bytes up load to the GPU per second. for a 4k ARGB8888 video frame @ 60fps, that's like 4GB/s of raw transfer if you rasterize each frames in the CPU.
(disclaimer: former KDE dev)
edit: If you run KWin, you can enable the redraw compositor plugin and see for yourself which area of the window are repainted.
Yes, it's poor UI.
I've noticed that the Audacity issue seems proportional to the number of tracks (not how much of the screen they take up) and improves when I zoom in a lot (to the sample level), which tells me this isn't a simple pixel pushing problem but rather something more profound in Audacity. There is obviously something horribly inefficient about how waveforms are being drawn at lower zoom levels.
For comparison, I can run full-screen Ardour with an order of magnitude more tracks and complexity than Audacity and it still performs much better, even if it isn't perfectly smooth.
It's a potential project killer, but it can be done. The trick is to wrap all your objects in QObject classes. You keep the old UI code as-is as long as possible.
The first Jedi mind trick is to reverse the event loop. Qt supports GLib2 loops, but only as long as they are created by Qt. So you create a dummy QCoreApplication, start the event loop. Then you coerce GTK into using the glib event loop Qt created.
After that, you populate your QObjects I mentionned above. The `Q_PROPERTY` stuff is the important part. It *has* to be fully implemented for QML to ever work. For Qt5, you need to use Qt containers. This means porting your lists and other C++ types (or glib types) to Qt. For Qt6, I think you can get away with some of that, but not Qt5.
Then, make all the signals work. It is important that all properties have signals. The signals must be reliable. QML heavily relies on events for literally everything. Without events, your UX gets out of sync and crash.
The next part is models. Turn all list with selections to models. Don't try to render stuff using for-loops in QML, this is a terrible idea and will explode in your face. Qt Models are annoying to write (and get repetitive very fast), but reliable list models really greatly simplify the QML code.
When you have all those QObject to reflect your current codebase, you make a companion app. This is a small window (or LAN/Network smartphone UI using QtRemoteObjects).
Now you are ready for the UI. Add little features, one by one. Don't try to do everything at once. Again, as long as you can, keep the wx/GTK window alive and fully functional. Your app will look wierd with 1 wx window and 1 QtQuick window. However, having the fully functional old UI let you press buttons there and see the results being updated in QML. That let you polish/debug your signal/q_property code step by step.
Once you are "done" with your port, strip the UI code from the old codebase. Then refactor/clean the logic and turn it into a library. Then start merging your QObject shims and logic classes into a single class.
If you get there, congradulation, your port is complete. Time to annoy your users with a bunch of changes.
With Qt 4/5, there is an environment variable, to play with. It sill runs like crap and is now deprecated.
edit: Of course, GTK4 have "real" backends for both Vulkan and OpenGL. They implement scene graphs like QtQuick rather than building a big bitmap like GTK2, GTK3 and QtWidgets.
On KDE HiDPI, the audio input/output device selector positions the dropdown boxes according to 100% scaling, but they're sized according to 125% scaling, so they overlap!
Or you get ancient looking Ui
Or you get platform spesific stuff!
Actually Java applications are kinda reasonable compromise
As are .NET applications. Yeah, most .NET desktop apps in the wild are technically chock full of Windowsisms, I've found that said apps more often than not Just Work(TM) on e.g. Linux w/ Mono, and alternative toolkits like Avalonia help further.