Parallel Raytracing
View full source code or view the compiled example online
This is an example of using threads with WebAssembly, Rust, and wasm-bindgen
,
culminating in a parallel raytracer demo. There's a number of moving pieces to
this demo and it's unfortunately not the easiest thing to wrangle, but it's
hoped that this'll give you a bit of a taste of what it's like to use threads
and wasm with Rust on the web.
Building the demo
One of the major gotchas with threaded WebAssembly is that Rust does not ship a
precompiled target (e.g. standard library) which has threading support enabled.
This means that you'll need to recompile the standard library with the
appropriate rustc flags, namely
-C target-feature=+atomics,+bulk-memory,+mutable-globals
.
Note that this requires a nightly Rust toolchain.
To do this you can use the RUSTFLAGS
environment variable that Cargo reads:
export RUSTFLAGS='-C target-feature=+atomics,+bulk-memory,+mutable-globals'
To recompile the standard library it's recommended to use Cargo's
-Zbuild-std
feature:
cargo build --target wasm32-unknown-unknown -Z build-std=panic_abort,std
Note that you can also configure this via .cargo/config.toml
:
[unstable]
build-std = ['std', 'panic_abort']
[build]
target = "wasm32-unknown-unknown"
rustflags = '-Ctarget-feature=+atomics,+bulk-memory,+mutable-globals'
After this cargo build
should produce a WebAssembly file with threading
enabled, and the standard library will be appropriately compiled as well.
The final step in this is to run wasm-bindgen
as usual, and wasm-bindgen
needs no extra configuration to work with threads. You can continue to run it
through wasm-pack
, for example.
Running the demo
Currently it's required to use the --target no-modules
or --target web
flag
with wasm-bindgen
to run threaded code. This is because the WebAssembly file
imports memory instead of exporting it, so we need to hook initialization of the
wasm module at this time to provide the appropriate memory object. This demo
uses --target no-modules
, because Firefox does not support modules in workers.
With --target no-modules
you'll be able to use importScripts
inside of each
web worker to import the shim JS generated by wasm-bindgen
as well as calling
the wasm_bindgen
initialization function with the shared memory instance from
the main thread. The expected usage is that WebAssembly on the main thread will
post its memory object to all other threads to get instantiated with.
Caveats
Unfortunately at this time running wasm on the web with threads has a number of
caveats, although some are specific to just wasm-bindgen
. These are some
pieces to consider and watch out for, although we're always looking for
improvements to be made so if you have an idea please file an issue!
-
The main thread in a browser cannot block. This means that if you run WebAssembly code on the main thread you can never block, meaning you can't do so much as acquire a mutex. This is an extremely difficult limitation to work with on the web, although one workaround is to run wasm exclusively in web workers and run JS on the main thread. It is possible to run the same wasm across all threads, but you need to be extremely vigilant about synchronization with the main thread.
-
Setting up a threaded environment is a bit wonky and doesn't feel smooth today. For example
--target bundler
is unsupported and very specific shims are required on both the main thread and worker threads. These are possible to work with but are somewhat brittle since there's no standard way to spin up web workers as wasm threads. -
There is no standard notion of a "thread". For example the standard library has no viable route to implement the
std::thread
module. As a consequence there is no concept of thread exit and TLS destructors will never run. We do expose a helper,__wbindgen_thread_destroy
, that deallocates the thread stack and TLS. If you invoke it, it must be the last function you invoke from the wasm module for a given thread. -
Any thread launched after the first one might attempt to block implicitly in its initialization routine. This is a constraint introduced by the way we set up the space for thread stacks and TLS. This means that if you attempt to run a wasm module in the main thread after you are already running it in a worker, it might fail.
-
Web Workers executing WebAssembly code cannot receive events from JS. A Web Worker has to fully return back to the browser (and ideally should do so occasionally) to receive JS messages and such. This means that common paradigms like a rayon thread pool do not apply straightforward-ly to the web. The intention of the web is that all long-term blocking happens in the browser itself, not in each thread, but many crates in the ecosystem leveraging threading are not necessarily engineered this way.
These caveats are all largely inherited from the web platform itself, and they're important to consider when designing an application for threading. It's highly unlikely that you can pull a crate off the shelf and "just use it" due to these limitations. You'll need to be sure to carefully plan ahead and ensure that gotchas such as these don't cause issues in the future. As mentioned before though we're always trying to actively develop this support so if folks have ideas about how to improve, or if web standards change, we'll try to update this documentation!
Browser Requirements
This demo should work in the latest Firefox and Chrome versions at this time,
and other browsers are likely to follow suit. Note that threads and
SharedArrayBuffer
require HTTP headers to be set to work correctly. For more
information see the documentation on
MDN
under "Security requirements" as well as Firefox's rollout blog
post. This
means that during local development you'll need to configure your web server
appropriately or enable a workaround in your browser.