←back to thread

579 points seanisom | 2 comments | | HN request time: 0.471s | source

I used to work at Adobe on the infrastructure powering big applications like Photoshop and Acrobat. One of our worst headaches was making these really powerful codebases work on desktop, web, mobile, and the cloud without having to completely rewrite them. For example, to get Lightroom and Photoshop working on the web we took a winding path through JavaScript, Google’s PNaCl, asm.js, and finally WebAssembly, all while having to rethink our GPU architecture around these devices. We even had to get single-threaded builds working and rebuild the UI around Web Components. Today the web builds work great, but it was a decade-long journey to get there!

The graphics stack continues to be one of the biggest bottlenecks in portability. One day I realized that WebAssembly (Wasm) actually held the solution to the madness. It’s runnable anywhere, embeddable into anything, and performant enough for real-time graphics. So I quit my job and dove into the adventure of creating a portable, embeddable WASM-based graphics framework from the ground up: high-level enough for app developers to easily make whatever graphics they want, and low-level enough to take full advantage of the GPU and everything else needed for a high-performance application.

I call it Renderlet to emphasize the embeddable aspect — you can make self-contained graphics modules that do just what you want, connect them together, and make them run on anything or in anything with trivial interop.

If you think of how Unity made it easy for devs to build cross-platform games, the idea is to do the same thing for all visual applications.

Somewhere along the way I got into YC as a solo founder (!) but mostly I’ve been heads-down building this thing for the last 6 months. It’s not quite ready for an open alpha release, but it’s close—close enough that I’m ready to write about it, show it off, and start getting feedback. This is the thing I dreamed of as an application developer, and I want to know what you think!

When Rive open-sourced their 2D vector engine and made a splash on HN a couple weeks ago (https://news.ycombinator.com/item?id=39766893), I was intrigued. Rive’s renderer is built as a higher-level 2D API similar to SVG, whereas the Wander renderer (the open-source runtime part of Renderlet) exposes a lower-level 3D API over the GPU. Could Renderlet use its GPU backend to run the Rive Renderer library, enabling any 3D app to have a 2D vector backend? Yes it can - I implemented it!

You can see it working here: https://vimeo.com/929416955 and there’s a deep technical dive here: https://github.com/renderlet/wander/wiki/Using-renderlet-wit.... The code for my runtime Wasm Renderer (a.k.a. Wander) is here: https://github.com/renderlet/wander.

I’ll come back and do a proper Show HN or Launch HN when the compiler is ready for anyone to use and I have the integration working on all platforms, but I hope this is interesting enough to take a look at now. I want to hear what you think of this!

Show context
zengid ◴[] No.39911269[source]
This is super neat and I am very interested!

I'm in a rush so I can't look to closely now but I have a few questions (and please forgive any stupid questions, I'm not a graphics dev, just a hobbyist):

What's the runtime like? Is there an event loop driving the rendering? (who calls the `render` on each frame? are there hooks into that? ) FFI story? Who owns the window pointer?

I'm interested in audio plugins, and VSTs (etc) have a lot of constrains on what can be done around event loops and window management. JUCE is pretty much the de-facto solution there, but it's pretty old and feels crufty.

replies(3): >>39911396 #>>39911899 #>>39913843 #
1. PaulDavisThe1st ◴[] No.39911899[source]
The limits on what audio plugins can do is not a function of the drawing toolkit, but the fact that they do not own the event loop if the GUI is run in-process with the host. And as long as they do, they will never own the evelop loop. In addition (and mostly related to this) the top level window they appear in is owned by the host, which also inherently limits the plugin's role in window management.

If you want more, use the capability built into LV2, AU and VST3 for out-of-process GUIs for a plugin (LV2 has had this for more than a decade). CLAP has, I think, abandoned plans to support this based on lack of uptake elsewhere.

I'd hardly call JUCE "pretty old", but then I'm a lot older than JUCE. And it's likely only crufty if you're more used to other styles of GUI toolkits; in terms of the "regular" desktop GUI toolkits, it's really bad at all.

replies(1): >>39913718 #
2. zengid ◴[] No.39913718[source]
Hey Paul, thanks for sharing!

Yes I think JUCE is great, It's very well made, but it drives you into a very narrow path of either using everything in the library, or leaving you to fend for yourself (which I admit may be a normal experience for C++ devs). For instance, the ValueTrees frequently used for UI state are very powerful, but they're not very type safe (or thread safe), and they feel clunky compared to more contemporary reactive state management patterns like signals.

I'm sure folks who use ValueTrees are happy, but I don't see much advancement to that pattern being shared in the JUCE forums. If y'all have some better tricks over in the Ardour project I'd love to know! (BTW, I'm a fan of y'all's work. I really enjoyed reading some of the development resources, like the essay on handling time [0]).

[0] https://ardour.org/timing.html