* Fix race condition in Metal background shader compiler
Setting an atomic variable outside of a std::condition_variable's corresponding mutex and then notifying the condition_variable has a subtle race condition, and the ordering of operations looks like this:
Thread A: acquire mutex
Thread A: check m_shouldQuit // it's false
Thread B: m_shouldQuit = true
Thread B: m_workAddedCondition.notify_all() // nothing is currently waiting on it so this does nothing
Thread A: m_workAddedCondition.wait() // now it's waiting, having just missed the above notification
Now thread A will wait forever because no more notifies to the condition_variable are coming and, in the case of this code, the call to m_compilerThread.join() will also wait forever (since the thread will never quit), so shutting down the renderer will never finish.
Instead, m_shouldQuit should only be set to true while in the mutex (because the check of its value and the subsequent .wait call also happen within the mutex, so this prevents them interleaving). This change fixes this.
* Minor formatting adjustment to make the github workflow tests re-run
We need to prevent percent margin from being set on Artboards because percent margin relies on the layout parent's dimensions. Since an Artboard has no parent, it results in an error within Yoga and a crash.
Adds support for Triggers in Custom Property Groups. This will allow triggers to be keyframed, which can also be bound to ViewModel triggers. Currently an event has to be fired on the timeline in order to fire a ViewModel trigger via a listener.
* Update unit test scripts to use build_rive.sh
The unit test test.bat and test.sh were doing their own ad-hoc building instead of using build_rive. This updates test.sh to call build_rive.sh for the build. This script could probably be pared down further but I did not want to break any of the existing command line arguments that people are probably using.
Additionally, test.bat did not have up-to-date build flags in it, so now it does a similar Windows development environment setup to build_rive.bat and then calls directly into test.sh, so it won't get out of sync with the main script anymore.
* Additional test.sh changes
- There was a block of code that I could have removed but didn't (made redundant by the call to build_rive)
- Add build_rive's parent directory to the path before calling it (will fix the test failing on Linux)
- Removed the `sh` before build_rive.sh as it does not need to be shell invoked.
* The pr_unit_tests.yaml execution for Mac needed to do a clean furing the asan build
because it had already done a build without asan before, the change to build_rive means that we now catch the premake parameter change (adding asan), and so clean needs to happen for that build to be correct.
* For now, move the build_rive.sh call into the two branches that previously had the build calls (linux, for instance, does not build, which seems incorrect given the github workflow is trying to run it.
* Adding a note to test.sh to describe why the build_rive script is called in the way it is (vs. the obvious way)
* Additional unit test script fixes (unifying windows build further, and some warnings)
- Add the ability to specify a toolset on the command line (i.e. "--toolset=msc") that will get passed along to build_rive (thus making the 'clang' toolset in the windows build only a default)
- Update the unit_test_windows and unit_test_windows_msvc entries in the github unit test workflows to call test.bat now (which required combining the build/run steps, but this matches all the other configurations)
- Also added some additional parameter validation, mainly warnings on ignored/unsupported parameters
Currently, when items are inside a scroll view, when the scroll view is dragged and released, the item's click event will still trigger its listener if the same item is hovered. This adds a way to disable pointer events. In this implementation, when a scroll drag begins, we set the GestureClickPhase to disabled which prevents clicks from being captured.
Updating premake introduced a couple issues. Fix them and add a windows
webgpu builder to CI. Also move wagyu extension wrangling down into the
core webgpu renderer so individual clients don't have to think about it.
chore: Update audio_preview_generator to use build_rive.sh
Let's keep our premake setup & building all in one place.
Also update peon_audio_worker Dockerfile.
It looks like the premake5 "v5.0.0-beta3" tag isn't compatible with
Sequoia 15.4.1. Bump our tag to v5.0.0-beta7 and update the deprecated
features we had been using.
Also fix the script to rebuild if the premake5 binary doesn't exist.
Before it only checked if the parent "premake-core" directory existed,
so if a build had failed previously, build_rive.sh would fail forever
without ever attempting to build premake5 again.
* feature: detection of implemented type arguments
* fixing highlight of type union
* feature: custom properties
* feature: working on scripted properties refactor
* feature: working on inputs and outputs
* chore: working on crash
* fix: fixes crash issue with luau
* feat: working on inputs
* chore: allow using old solver
* chore: missed json file
* chore: missing files & cleanup
* chore: missed sriv file
* chore: fixing unused var
* chore: fix read not supported with old solver, and core collision
* fix: use different script update flag
It was a cool idea to download the repos as a zip, so that premake
didn't have to rely on any external dependencies, but github was
frequently giving us a successful download of an empty zip file. This
was causing daily flakes on CI.
Instead, just invoke 'git clone --depth=1 --branch ...' from premake.
This should hopefully be more reliable, and it seems safe enough to
depend on git being installed in the environment we're on.
Also:
* Update the dependencies that were using raw branch SHAs to point to
official tags. 'git clone --branch' requires an official tag, and this
will be more stable anyway.
* Rename the dependency directories to "owner_project_tag" rather than a
hash. Sometimes we need to poke around in those directories to debug
issues, and it's easier to figure out what's what with descriptive
names.
Description
We can already publish view model images, lists, list indices, artboards. Now we just need to handle them on the host file client side, which is what this PR does.
Details
The main part of this PR is that it changes the runtime_exporter. I've bit the bullet and sorted the file assets instead of using recursion. I've always wanted to export them iteratively instead of recursively but now I'm forced to add a 'preprocess' stage to exporting, so there's finally a good reason to do it.
About preprocessing. Imagine databinding Artboards. Any library file needs to be able to bind to any host/library Artboard. This means, the library file needs to export a vm instance whose value can be any library's Artboard, without necessarily needing to have that library in an asset. In preprocessing, we first calculate all the Artboards' indices that should exist in runtime (without writing them). Then at write time, the vm instance value can be accurately mapped to any Artboard's ID.
Same idea goes for fileAsset indices in preprocessing. Tho honestly, we probably just need to preprocess the host's file assets, since the UI only allows you to bind to the host's assets.
Same idea for lists in preprocessing. At write time, an instance's list property value need to refer to any view model+instance from any file, regardless on if it depends on that file as a LibraryAsset.
Demo: https://2dimensions.slack.com/archives/C07M7DQL4F2/p1753414266031969
* chore: explicit vec2D.xy and origin
* chore: reworking color api
* chore: refactor mat2d
* chore: add paint.with and paint.new
* feature: working on exposing builtin definitions
* chore: cleanup
* fix: removing unused var
* fix: bad api call
* chore: missed file
