Follow-up to https://github.com/GoogleChromeLabs/squoosh/pull/838.
I have no idea what the JS+Wasm diffs in the original PR even represented if they weren't proper rebuilds, but this time I just removed node_modules in all codecs to enforce a proper, clean rebuild for each C++ codec.
(Still think the speed-ups are worth it.)
Change the way AVIF finds AOM from default ([avif source]/ext/aom) to custom paths. This allows us to avoid unpacking same archives into duplicate folders, and instead make multiple builds from the same source.
See the discussion in https://github.com/AOMediaCodec/libavif/issues/254 where this was implemented.
This allows us to avoid using ERROR_ON_UNDEFINED_SYMBOLS and build a truly separate encoder/decoder libs.
Still not perfect due to usage of a static global, but this is much cleaner and more efficient thanks to proper blocking of Workers that wait for new messages instead of a manual spin-loop.
- Refactor to work around Chromium's issue with postMessage queuing. https://bugs.chromium.org/p/chromium/issues/detail?id=1075645
- Convert codec code to TypeScript.
- Make separate parallel and non-parallel builds.
- Switch to nightly Rust for OxiPNG to allow parallel builds (but also reuse it for regular builds to avoid installing two toolchains).
Now initialise all workers with module+memory separately, and then instead of using postMessage to send thread pointers, push them into a crossbeam-deque on the Rust side.
Rayon already depends on crossbeam-dequeue, so we're not even adding another dependency, and this model allows us to push "tasks" (thread pointers) on the main thread and pop them on worker threads in arbitrary order without sacrificing correctness.
