Egui commit: Implement accessibility APIs via AccessKit
github.com
github.com
Disclosure: Google Fonts funded Matt's work and I was involved in the decision to do so.
For the live area, do desktop APIs support something like the HTML 'role' attribute? https://developer.mozilla.org/en-US/docs/Web/Accessibility/A...
As an app developer it would be helpful to be able to denote the live region in your example a 'status' (the user can read it but it's not read automatically on change).
And there may be other times when the same area should be an 'alert' (changes trigger 'assertive' messages to inform the user that their change results in an error or issue).
I’ve personally built relatively complex UIs with egui. Both a Chrome extension and a VSCode extension for https://membrane.io and couldn’t be happier with it.
Some examples: text rendering is awful on some common browser configurations (mine included—Firefox on Linux/Wayland, rendered at 2× scaling and downsampled to 1.5×; the text is extraordinarily fuzzy, and it’s rendering in its own font because I suppose it’s bundling a complete font, shaper and renderer—since I force my own font selections so only the generic font-families work, and I also block all web fonts at fetch time); the font rendering technique means no fallback which means things like emoji would only render if they provided a font that supported that (at present they don’t, and it’s just a nasty technique all round, you want to be able to use local fonts); links are fake, so various interactions can never work correctly (e.g. Ctrl+click and middle-click should mostly open a background tab, and right-clicking should give you the normal context menu); scrolling can’t be done properly and will thus inevitably be extremely obnoxious on most devices; mobile-style pinch zoom cannot be supported.
⸻And I’m just mentioning some of the unfixable problems with this approach; there are more things that are wonky or painful apart from that: things that could be fixed, but where it’s not surprising that they’re broken because that’s a common problem when you throw away almost all of the stuff the platform gives you and try to implement it all from scratch.
I very strongly recommend against using this general approach at all on the web. The web platform just doesn’t provide the primitives required to make it anything other than miserable.
(I’m not speaking against egui for platforms other than web, only for web. The fact of the matter is that functionally you’re fine building a native app with native tech, a native app with web tech, or a web app with web tech, but that building a web app with native tech is a terrible idea that will always be inferior, in important, unfixable ways.)
When building DOM-based applications that require frequent updates I've always ended up working around browser limitations in awkward ways, e.g. minimizing re-rendering by using redux's relesect.
I decided to use egui because I get to use Rust for the entire stack, but also because it lets me ship a native app without an entire (slow to init) browser.
Google Docs I have just now tried. It isn’t like egui.
egui throws away everything the browser gives it, other than a text input element and a canvas element, and does all the rendering itself and dumps it to the canvas as a bitmap.
Google Docs is full DOM for everything other than rendering the document, and it still uses the browser’s text layout and painting facilities. (I say this because I have my browser set to only use my chosen fonts—so for English text, there are only the default font, serif, sans-serif and monospace—and I observe that it renders and shapes everything in my chosen default font, and am mildly disappointed it doesn’t even do a sane fallback, e.g. “Courier New” to monospace, and “Sans Serif” to sans-serif.) Almost all of my criticisms of the egui approach are avoided. Scrolling works fine, because it’s using the DOM for that, so it’s just normal scrolling. Links work fine, because it uses normal links (these are in tooltip-like things when you move the cursor to the link in the normal text flow).
(I retract my remark about mobile-style pinch zoom not being possible in canvas-only, because Google Docs does support it. I can now imagine how it’s done and I’m not sure why I didn’t see it before.)
But I would also remark about Google Docs since I’ve now tried it in edit mode for the first time since canvas rendering: ugh it’s slow. Low throughput, high latency, high jitter. This on a very fast (Ryzen 5800HS), near-idling laptop with a 165 Hz screen and 2× devicePixelRatio. Just try holding a letter down (I have a 40ms repeat rate), and compare the result to the same actions on contenteditable: the difference is stark. It also gets keyboard caret navigation behaviour wrong for my platform (Linux, Firefox), and its behaviour doesn’t match my vague recollection of Windows’ either. (I mean things like where Ctrl+Left and Ctrl+Right stop—whether at the end of one word or the start of the next, and whether to consider punctuation as part of another word or a word on its own; that kind of thing.)
At what cost?
Google Docs is run a by a trillion-dollar corporation, they have the resources to rebuild it in canvas.
Figma isn't based on canvas, it's based on WebGL. And here's what Figma had to say about it: "Pulling this off was really hard; we’ve basically ended up building a browser inside a browser... We’ve implemented our own rendering engine" [1]
Good luck getting text to render properly on your own.
Just because there are two great examples made at great cost doesn't mean it's scalable to all other apps.
[1] https://www.figma.com/blog/building-a-professional-design-to...
Aren't they the same thing?
And egui also talks about challenges with rendering on browsers using WebGL and WASM: https://github.com/emilk/egui/tree/master/crates/eframe#prob...
Ah. In my mind Canvas refers to the 2D API, https://developer.mozilla.org/en-US/docs/Web/API/Canvas_API
Which is different and separate from WebGL, https://developer.mozilla.org/en-US/docs/Web/API/WebGL_API
(There will be a third different and separate API, WebGPU: https://www.w3.org/TR/webgpu/)
> And egui also talks about challenges with rendering on browsers using WebGL and WASM
Ah. It's good they are aware of this.
I agree that using a canvas instead of leveraging existing accessibility supplied by HTML/CSS/JS and the DOM is an accessibility nightmare.
However, I disagree that canvas accessibility issues are "unfixable". Difficult? Yes. But not unfixable. In my view, the <canvas> element works best in partnership with its wider environment (HTML/CSS/JS and the DOM), not as a replacement for it. With that in mind, we can start to tackle the accessibility issues you raise - fonts, links, interactions, etc.[1][2]
I have an ambition to one day become intelligent enough to understand/code in Rust, and I'm really glad to see that people are thinking about accessibility as a fundamental part of UIs being developed in Rust.
[1] - Which is what my JS 2d canvas library tries to do: https://scrawl-v8.rikweb.org.uk/
[2] - My thoughts on <canvas> accessibility, and how I try to fix them using my library: https://scrawl-v8.rikweb.org.uk/learn/eleventh-lesson/
I’ll focus on the two most obviously significant issues:
① Scrolling. The only complete solution is to draw within scrolling elements. To treat scrolling content distinctly, and allow the browser to act as the compositor. This would require rebuilding egui from the ground up; its current philosophy is just completely incompatible.
② Links. The only complete solution is to use actual link elements. My complaints about scrawl-canvas’s handling of links <https://news.ycombinator.com/item?id=22887235> when you submitted it here 2½ years ago still apply: to get it right, you’ve got to use actual links, placed in the right place, so that that’s what the cursor is actually interacting with the whole time. The difference here from scrolling is that they can be transparent elements on top of everything else. On reflection, I think it might be theoretically possible for AccessKit to achieve this itself in the simplest cases, but that would certainly break egui’s conception of how things should work, and it can’t work generally because of event handling, the possibility of a non-rectangular hit box, and the fact that egui is a compositor (you could have a window completely covering your link).
And so I say: these are unfixable issues in egui’s design.
(I also report strongly negative experiences with every canvas-plus-DOM approach I’ve encountered. They never get everything right, and the duplication required to get things as close to right as possible almost always leaves me wondering why you bothered with the canvas parts in the first place.)
I'm concerned about the performance hit of an immediate-mode GUI when not tieing it into a 3D render (ie the CPU access from repeatedly re-drawing), but haven't compared.
I have to admit that I've only tested it on high-end laptops/phones though. Not sure about mid/low-end ones yet.
Whereas the DOM uses "retained mode": it continually modifies a representation of the layout (DOM elements) and re-renders when something _actually_ changes.
You can click on the "Backend" button at the top, and "Mean CPU usage" will tell you how long it takes for each frame to be produced. It'll vary depending on how complex the layout is: e.g. how many windows are open. But, contrary to retained mode, it doesn't matter at all if everything is animating simultaneously.
More about immediate vs. retained: https://github.com/emilk/egui#why-immediate-mode
Coming from a retained mode world it's so much easier to read, write and reason about.
Don't find the wasm/web compilation as useful though, it does seem to hit performance quite badly, not sure if that's the nature of the beast in general or something else to do with optimisations for the library.
Egui is an awesome tool for building quick-and-dirty GUIs: we use it at work for our game's developer tools.
I want to provide Python bindings at some point. That would enable Kivy to implement accessibility, for example. Just haven't gotten to it yet. If anyone wants to step up and work on that, let me know.