Wasm Workers API
The Wasm Workers API enables C/C++ code to leverage Web Workers and sharedWebAssembly.Memory (SharedArrayBuffer) to build multithreaded programsvia a direct web-like programming API.
Quick Example
#include<emscripten/wasm_worker.h>#include<stdio.h>voidrun_in_worker(){printf("Hello from Wasm Worker!\n");}intmain(){emscripten_wasm_worker_tworker=emscripten_malloc_wasm_worker(/*stackSize: */1024);emscripten_wasm_worker_post_function_v(worker,run_in_worker);}
Build the code by passing the Emscripten flag-sWASM_WORKERS at both compileand link steps. The example code creates a new Worker on the main browser thread,which shares the same WebAssembly.Module and WebAssembly.Memory object. Then apostMessage() is passed to the Worker to ask it to execute the functionrun_in_worker() to print a string.
To explicitly control the memory allocation placement when creating a worker,use theemscripten_create_wasm_worker() function. This function takes aregion of memory that must be large enough to hold both the stack and the TLSdata for the worker. You can use__builtin_wasm_tls_size() to find out atruntime how much space is required for the program’s TLS data.
Introduction
In WebAssembly programs, the Memory object that contains the application state can beshared across multiple Workers. This enables direct, high performance (and if explicitcare is not taken, racy!) access to synchronously share data state between multipleWorkers (shared state multithreading).
- Emscripten supports two multithreading APIs to leverage this web feature:
POSIX Threads (Pthreads) API, and
Wasm Workers API.
The Pthreads API has a long history with native C programming and the POSIX standard,while Wasm Workers API is unique to Emscripten compiler only.
These two APIs provide largely the same feature set, but have important differences,which this documentation seeks to explain to help decide which API one should target.
Pthreads vs Wasm Workers: Which One to Use?
The intended audience and use cases of these two multithreading APIs are slightlydifferent.
The focus on Pthreads API is on portability and cross-platform compatibility. This APIis best used in scenarios where portability is most important, e.g. when a codebase iscross-compiled to multiple platforms, like building both a native Linux x64 executable and anEmscripten WebAssembly based web site.
Pthreads API in Emscripten seeks to carefully emulate compatibility and the features thatthe native Pthreads platforms already provide. This helps porting large C/C++ codebasesover to WebAssembly.
Wasm Workers API on the other hand seeks to provide a “direct mapping” to the webmultithreading primitives as they exist on the web, and call it a day. If an applicationis only developed to target WebAssembly, and portability is not a concern, then using WasmWorkers can provide great benefits in the form of simpler compiled output, less complexity,smaller code size and possibly better performance.
However this benefit might not be an obvious win. The Pthreads API was designed to be usefulfrom the synchronous C/C++ language, whereas Web Workers are designed to be useful fromasynchronous JavaScript. WebAssembly C/C++ programs can find themselves somewhere inthe middle.
Pthreads and Wasm Workers share several similarities:
Both can use emscripten_atomic_* Atomics API,
Both can use GCC __sync_* Atomics API,
Both can use C11 and C++11 Atomics APIs,
Both types of threads have a local stack.
Both types of threads have thread-local storage (TLS) support via
thread_local(C++11),_Thread_local(C11) and__thread(GNU11) keywords.Both types of threads support TLS via explicitly linked in Wasm globals (see
test/wasm_worker/wasm_worker_tls_wasm_assembly.c/.Sfor example code)Both types of threads have a concept of a thread ID (
pthread_self()for pthreads,emscripten_wasm_worker_self_id()for Wasm Workers)Both types of threads can perform an event-based and an infinite loop programming model.
Both can use
EM_ASMandEM_JSAPI to execute JS code on the calling thread.Both can call out to JS library functions (linked in with
--js-librarydirective) toexecute JS code on the calling thread.Neither pthreads nor Wasm Workers can be used in conjunction with
-sSINGLE_FILElinker flag.
However, the differences are more notable.
Pthreads can proxy JS functions
Only pthreads can use theMAIN_THREAD_EM_ASM*() andMAIN_THREAD_ASYNC_EM_ASM() functions andthefoo__proxy:'sync'/'async' proxying directive in JS libraries.
Wasm Workers on the other hand do not provide a built-in JS function proxying facility. Proxying a JSfunction with Wasm Workers can be done by explicitly passing the address of that function to theemscripten_wasm_worker_post_function_* API.
If you need to synchronously wait for the posted function to finish from within a Worker, use one oftheemscripten_wasm_worker_*() thread synchronization functions to sleep the calling thread untilthe callee has finished the operation.
Note that Wasm Workers cannot
Pthreads have cancellation points
At the expense of performance and code size, pthreads implement a notion ofPOSIX cancellationpoints (pthread_cancel(),pthread_testcancel()).
Wasm Workers are more lightweight and performant by not enabling that concept.
Pthreads may start up synchronously - Wasm Workers always start asynchronously
Creating new Workers can be slow. Spawning a Worker in JavaScript is an asynchronous operation. In orderto support synchronous pthread startup (for applications that need it) and to improve thread startupperformance, pthreads are hosted in a cached Emscripten runtime managed Worker pool.
Wasm Workers omit this concept, and as result Wasm Workers will always start up asynchronously.If you need to detect when a Wasm Worker has started up, post a ping-pong function and reply pairmanually between the Worker and its creator. If you need to spin up new threads quickly, considermanaging a pool of Wasm Workers yourself.
Pthread topology is flat - Wasm Workers are hierarchical
On the web, if a Worker spawns a child Worker of its own, it will create a nested Worker hierarchythat the main thread cannot directly access. To sidestep portability issues stemming from this kindof topology, pthreads flatten the Worker creation chain under the hood so that only the main browser threadever spawns threads.
Wasm Workers do not implement this kind of topology flattening, and creating a Wasm Worker in aWasm Worker will produce a nested Worker hierarchy. If you need to create Wasm Workers from withina Wasm Worker, consider which type of hierarchy you would like, and if necessary, flatten thehierarchy manually by posting the Worker creation over to the main thread yourself.
Note that support for nested Workers varies across browsers. As of 02/2022, nested Workers arenotsupported in Safari. Seehere for a polyfill.
Pthreads can use the Wasm Worker synchronization API, but not vice versa
The multithreading synchronization primitives offered inemscripten/wasm_worker.h(emscripten_lock_*,emscripten_semaphore_*,emscripten_condvar_*) can be freely invokedfrom within pthreads if one so wishes, but Wasm Workers cannot utilize any of the synchronizationfunctionality in the Pthread API (pthread_mutex_*,pthread_cond_,pthread_rwlock_*, etc),since they lack the needed pthread runtime.
Pthreads have a “thread main” function and atexit handlers
The startup/execution model of pthreads is to start up executing a given thread entry point function.When that function exits, the pthread will also (by default) quit, and the Worker hosting that pthreadwill return to the Worker pool to wait for another thread to be created on it.
Wasm Workers instead implement the direct web-like model, where a newly created Worker sits idle in itsevent loop, waiting for functions to be posted to it. When those functions finish, the Worker willreturn to its event loop, waiting to receive more functions (or worker scope web events) to execute.A Wasm Worker will only quit with a call toemscripten_terminate_wasm_worker(worker_id) oremscripten_terminate_all_wasm_workers().
Pthreads allow one to register thread exit handlers viapthread_atexit, which will be called whenthe thread quits. Wasm Workers do not have this concept.
Pthreads have a per-thread incoming proxy message queue, Wasm Workers do not
In order to enable flexible synchronous execution of code on other threads, and to implement supportAPIs for example for MEMFS filesystem and Offscreen Framebuffer (WebGL emulated from a Worker) features,main browser thread and each pthread have a system-backed “proxy message queue” to receive messages.
This enables user code to call API functions,emscripten_sync_run_in_main_runtime_thread(),emscripten_async_run_in_main_runtime_thread(),emscripten_dispatch_to_thread(), etc. fromemscripten/threading.h to perform proxied calls.
Wasm Workers do not provide this functionality. If needed, such messaging should be implemented manuallyby users via regular multithreaded synchronized programming techniques (mutexes, futexes, semaphores, etc.)
Pthreads synchronize wallclock times
Another portability aiding emulation feature that Pthreads provide is that the time values returned byemscripten_get_now() are synchronized to a common time base across all threads.
Wasm Workers omit this concept, and it is recommended to use the functionemscripten_performance_now()for high performance timing in a Wasm Worker, and avoid comparing resulting values across Workers, ormanually synchronize them.
Input events API backproxies only to pthreads
The multithreaded input API provided inemscripten/html5.h only works with the pthread API. Whencalling any of the functionsemscripten_set_*_callback_on_thread(), one can choose the targetpthread to be the recipient of the received events.
With Wasm Workers, if desired, “backproxying” events from the main browser thread to a Wasm Workershould be implemented manually e.g. by using theemscripten_wasm_worker_post_function_*() API family.
However note that backproxying input events has a drawback that it prevents security sensitive operations,like fullscreen requests, pointer locking and audio playback resuming, since handling the input eventis detached from the event callback context executing the initial operation.
Pthread vs emscripten_lock implementation differences
The mutex implementation frompthread_mutex_* has a few different creation options, one being a“recursive” mutex.
The lock implemented byemscripten_lock_* API is not recursive (and does not provide an option).
Pthreads also offer a programming guard against a programming error that one thread would not releasea lock that is owned by another thread.emscripten_lock_* API does not track lock ownership.
Memory requirements
Pthreads have a fixed dependency to dynamic memory allocation, and perform calls tomalloc andfreeto allocate thread specific data, stacks and TLS slots.
With the exception of the helper functionemscripten_malloc_wasm_worker(), Wasm Workers are not dependenton a dynamic memory allocator. Memory allocation needs are met by the caller at Worker creation time, andcan be statically placed if desired.
Generated code size
The disk size overhead from pthreads is on the order of a few hundred KBs. Wasm Workers runtime on the otherhand is optimized for tiny deployments, just a few hundred bytes on disk.
API Differences
To further understand the different APIs available between Pthreads and Wasm Workers, refer to the followingtable.
| Feature | Pthreads | Wasm Workers |
| Thread termination | Thread callspthread_exit(status)or main thread calls pthread_kill(code) | Worker cannot terminate itself, parent thread terminates by callingemscripten_terminate_wasm_worker(worker) |
| Thread stack | Specify in pthread_attr_t structure. | Manage thread stack area explicitly withemscripten_create_wasm_worker_*_tls()functions, or automatically allocate stack+TLS area with emscripten_malloc_wasm_worker()API. |
| Thread Local Storage (TLS) | Supported transparently. | Supported either explicitly withemscripten_create_wasm_worker_*_tls()functions, or automatically via emscripten_malloc_wasm_worker()API. |
| Thread ID | Creating a pthread obtains its ID. Callpthread_self()to acquire ID of calling thread. | Creating a Worker obtains its ID. Callemscripten_wasm_worker_self_id()acquire ID of calling thread. |
| High resolution timer | ``emscripten_get_now()`` | ``emscripten_performance_now()`` |
| Synchronous blocking on main thread | Synchronization primitives internally fall back to busy spin loops. | Explicit spin vs sleep synchronization primitives. |
| Futex API | emscripten_futex_wait emscripten_futex_wakein emscripten/threading.h | emscripten_atomic_wait_u32 emscripten_atomic_wait_u64 emscripten_atomic_notifyin emscripten/atomic.h |
| Asynchronous futex wait | N/A | emscripten_atomic_wait_async() emscripten_*_async_acquire()However these are a difficult footgun, readWebAssembly/threads issue #176 |
| C/C++ Function Proxying | emscripten/threading.h API for proxying function calls to other threads. | Use emscripten_wasm_worker_post_function_*() API to message functions to other threads. These messages follow event queue semantics rather than proxy queue semantics. |
| Build flags | Compile and link with -pthread | Compile and link with -sWASM_WORKERS |
| Preprocessor directives | __EMSCRIPTEN_SHARED_MEMORY__=1 and __EMSCRIPTEN_PTHREADS__=1 are active | __EMSCRIPTEN_SHARED_MEMORY__=1 and __EMSCRIPTEN_WASM_WORKERS__=1 are active |
| JS library directives | USE_PTHREADS and SHARED_MEMORY are active | USE_PTHREADS, SHARED_MEMORY and WASM_WORKER are active |
| Atomics API | Supported, use any of__atomic_* API,__sync_* API orC++11 std::atomic API. | |
| Nonrecursive mutex | pthread_mutex_* | emscripten_lock_* |
| Recursive mutex | pthread_mutex_* | N/A |
| Semaphores | N/A | emscripten_semaphore_* |
| Condition Variables | pthread_cond_* | emscripten_condvar_* |
| Read-Write locks | pthread_rwlock_* | N/A |
| Spinlocks | pthread_spin_* | emscripten_lock_busyspin* |
| WebGL Offscreen Framebuffer | Supported with -sOFFSCREEN_FRAMEBUFFER | Not supported. |
Wasm Workers stack size considerations
When instantiating a Wasm Worker, one has to create a memory array for the LLVMdata stack for the created Worker. This data stack will generally consist onlyof local variables that have been “spilled” by LLVM into memory, e.g. to containlarge arrays, structs, or other variables that are referenced by a memoryaddress. This stack will not contain control flow information.
Since WebAssembly does not support virtual memory, the size of the LLVM datastack that is defined both for Wasm Workers but also the main thread will not bepossible to grow at runtime. So if the Worker (or the main thread) runs out ofstack space, the program behavior will be undefined. Use the Emscripten linkerflag -sSTACK_OVERFLOW_CHECK=2 to emit runtime stack overflow checks into theprogram code to detect these situations during development.
Note that to avoid the need to perform two separate allocations, the TLS memoryfor the Wasm Worker will be located at the bottom end (low memory address) ofthe Wasm Worker stack space.
Wasm Workers vs the earlier Emscripten Worker API
Emscripten provides a second Worker API as part of the emscripten.h header. This Worker API predates the advent of SharedArrayBuffer, and is quite distinct from Wasm Workers API, just the naming of these two APIs is similar due to historical reasons.
Both APIs allow one to spawn Web Workers from the main thread, though the semantics are different.
With the Worker API, the user will be able to spawn a Web Worker from a custom URL. This URL can point to a completely separate JS file that was not compiled with Emscripten, to load up Workers from arbitrary URLs. With Wasm Workers, a custom URL is not specified: Wasm Workers will always spawn a Web Worker that computes in the same WebAssembly+JavaScript context as the main program.
The Worker API does not integrate with SharedArrayBuffer, so interaction with the loaded Worker will always be asynchronous. Wasm Workers howerer is built on top of SharedArrayBuffer, and each Wasm Worker shares and computes in the same WebAssembly Memory address space of the main thread.
Both the Worker API and Wasm Workers API provide the user with ability to postMessage() function calls to the Worker. In Worker API, this message posting is restricted to need to originate/initiate from the main thread towards the Worker (using the APIemscripten_call_worker() andemscripten_worker_respond() in<emscripten.h>). With Wasm Workers however one can also postMessage() function calls to their parent (owning) thread.
If posting function calls with the Emscripten Worker API, it is required that the target Worker URL points to an Emscripten compiled program (so it has theModule structure to locate function names). Only functions that have been exported to theModule object are callable. With Wasm Workers, any C/C++ function may be posted, and does not need to be exported.
- Use the Emscripten Worker API when:
you want to easily spawn a Worker from a JS file that was not built using Emscripten
you want to spawn as Worker a single separate compiled program than the main thread program represents, and the main thread and Worker programs do not share common code
you do not want to require the use of SharedArrayBuffer, or setting up COOP+COEP headers
you only need to communicate with the Worker asynchronously using postMessage() function calls
- Use the Wasm Workers API when:
you want to create one or more new threads that synchronously compute in the same Wasm Module context
you want to spawn multiple Workers from the same codebase and save memory by sharing the WebAssembly Module (object code) and Memory (address space) across the Workers
you want to synchronously coordinate communication between threads by using atomic primitives and locks
your web server has been configured with the needed COOP+COEP headers to enable SharedArrayBuffer capabilities on the site
Limitations
The following build options are not supported at the moment with Wasm Workers:
-sSINGLE_FILE
Dynamic linking (-sLINKABLE, -sMAIN_MODULE, -sSIDE_MODULE)
-sPROXY_TO_PTHREAD
Example Code
See the directorytest/wasm_worker/ for code examples on different Wasm Workers API functionality.
