Use WASM workers for even smaller size

API.md

@@ -71,6 +71,6 @@ cd Vosklet/src &&
 | Option | Description | Default value |
 |---|---|---|
 | INITIAL_MEMORY | Set inital memory, valid suffixes: kb, mb, gb, tb or none (bytes) | ```300mb``` as [recommended](https://alphacephei.com/vosk/models). This memory will grow if usage exceeds this value, but this may [affect performance](https://github.com/WebAssembly/design/issues/1271). |
-| MAX_THREADS | Set the max number of threads (>=1), this should be equal to the number of model and speaker model that is used in the program | ```1``` (1 recognizer, 1 model, no speaker model) |
+| MAX_THREADS | Set the max number of threads (>=1), this should be equal to the number of recognizers used in the program | ```1``` |
 | JOBS | Set the number of jobs (threads) when building | ```$(nproc)```   |
 | EMSDK | Set EMSDK's path (will install EMSDK in root folder if unset) | ```../emsdk``` |

README.md

@@ -1,19 +1,13 @@
 # Overview
-- A lightweight speech recognizer built on Vosk that can be run on the browser, inspired by [vosk-browser](https://github.com/ccoreilly/vosk-browser), but built from scratch and no code taken!
-- Designed with basic/nothrow exception safety
-- See the *examples* folder for examples on using the API.
-- See *API.md* for the API reference
-- See *test* for a developer build script for just the JS
+- A lightweight, up to date speech recognizer in the browser with total gzipped size of **under a megabyte**
+- Built from scratch, inspired by [vosk-browser](https://github.com/ccoreilly/vosk-browser)
 
-# Compared to vosk-browser
+# Vosklet ...
+- Is regularly maintained
 - Support multiple models
-- Has models' storage path management
-- Has models' ID management (for model updates)
-- Has smaller JS size (>3.1MB vs 900KB gzipped)
-- Has all related files (pthread worker, audio worklet processor,...) merged
-- Has faster processing time
-- Has shorter from-scratch build time
-- Has more Vosk functions exposed
+- Include model storage management
+- Include model ID management (for updates)
+- Wraps all Vosk's functionaly
 
 # Basic usage (microphone recognition in English)
 - Result are logged to the console.

src/Recognizer.cc

@@ -1,5 +1,5 @@
 #include "Recognizer.h"
-#include "emscripten/atomic.h"
+
 Recognizer::Recognizer(int index, float sampleRate, CommonModel* model) :
   index{index},
   rec{vosk_recognizer_new(std::get<VoskModel*>(model->mdl), sampleRate)}

src/Util.cc

@@ -1,10 +1,21 @@
 #include "Util.h"
-#include <emscripten/em_asm.h>
-ThreadPool globalPool;
-void fireEv(int index, const char* content, const char* type) {
-  MAIN_THREAD_EM_ASM({
-    objs[$0].dispatchEvent(new CustomEvent($2 === 0 ? "0" : UTF8ToString($2), { "detail" : UTF8ToString($1) }));
-  }, index, content, type);
+#include <emscripten/em_js.h>
+#include <emscripten/wasm_worker.h>
+
+WorkerPool globalPool;
+EM_JS(void, _fireEv, (int index, int content, int type), {
+  objs[index].dispatchEvent(new CustomEvent(type === 0 ? "0" : UTF8ToString(type), { "detail" : UTF8ToString(content) }));
+})
+void fireEv(int index, const char* _content, const char* _type) {
+  int content = reinterpret_cast<int>(_content);
+  int type = reinterpret_cast<int>(_type);
+  switch(emscripten_wasm_worker_self_id()) {
+  case 0: [[unlikely]]
+    _fireEv(index, content, type);
+    break;
+  case 1: [[likely]]
+    emscripten_wasm_worker_post_function_viii(0, _fireEv, index, content, type);
+  }
 }
 int untar(unsigned char* tar, int tarSize, const std::string& storepath) {
   if(std::memcmp(tar + 257, "ustar", 5)) return IncorrectFormat;
@@ -46,40 +57,41 @@ int untar(unsigned char* tar, int tarSize, const std::string& storepath) {
   }
   return Successful;
 }
-void Thread::startup(ThreadPool* pool) {
-  while(!pool->done) {
+void Worker::startup(int _self, int _pool) {
+  Worker& self = *reinterpret_cast<Worker*>(_self);
+  WorkerPool& pool = *reinterpret_cast<WorkerPool*>(_pool);
+  while(!pool.done) {
     // Wait until unlocked
-    emscripten_atomic_wait_u32(&pool->qLock, true, -1);
-    if(pool->done) break;
+    emscripten_atomic_wait_u32(&pool.qLock, true, -1);
+    if(pool.done) break;
     // If there is no task then everyone has to wait until there is more
-    if(pool->taskQ.empty()) {
-      emscripten_atomic_store_u32(&pool->qLock, true);
+    if(pool.taskQ.empty()) {
+      emscripten_atomic_store_u32(&pool.qLock, true);
       continue;
     }
     // If this locks, the returned (loaded) value will be false, and we move on
-    if(emscripten_atomic_cas_u32(&pool->qLock, false, true)) continue;
-    fn = pool->taskQ.front();
-    pool->taskQ.pop();
+    if(emscripten_atomic_cas_u32(&pool.qLock, false, true)) continue;
+    self.fn = pool.taskQ.front();
+    pool.taskQ.pop();
     // Unlock
-    emscripten_atomic_store_u32(&pool->qLock, false);
-    emscripten_atomic_notify(&pool->qLock, 1);
-    fn();
+    emscripten_atomic_store_u32(&pool.qLock, false);
+    emscripten_atomic_notify(&pool.qLock, 1);
+    self.fn();
   }
 }
-ThreadPool::ThreadPool() {
-  for(Thread& thrd : threads) {
-    thrd.handle = std::thread{&Thread::startup, &thrd, this};
+WorkerPool::WorkerPool() {
+  for(int i = 0; i < workers.size(); i++) {
+    workers[i].handle = emscripten_create_wasm_worker(&stacks[i * workerStack], workerStack);
+    emscripten_wasm_worker_post_function_vii(workers[i].handle, Worker::startup, reinterpret_cast<int>(&workers[i]), reinterpret_cast<int>(this));
   }
 }
-ThreadPool::~ThreadPool() {
+WorkerPool::~WorkerPool() {
   done = true;
   emscripten_atomic_store_u32(&qLock, false);
   emscripten_atomic_notify(&qLock, -1);
-  for(Thread& thrd : threads) {
-    thrd.handle.detach();
-  }
+  emscripten_terminate_all_wasm_workers();
 }
-void ThreadPool::exec(std::function<void()> fn) {
+void WorkerPool::exec(std::function<void()> fn) {
   taskQ.emplace(fn);
   emscripten_atomic_store_u32(&qLock, false);
   emscripten_atomic_notify(&qLock, 1);

src/Util.h

@@ -1,10 +1,9 @@
 #pragma once
-
-#include <thread>
 #include <filesystem>
 #include <functional>
 #include <variant>
 #include <fstream>
+
 #include <emscripten/atomic.h>
 #include <emscripten/console.h>
 namespace fs = std::filesystem;
@@ -22,26 +21,28 @@ enum UntarStatus {
   FailedWrite,
   FailedClose
 };
-struct ThreadPool;
-struct Thread {
-  std::thread handle;
+struct WorkerPool;
+struct Worker {
+  int handle;
   std::function<void()> fn;
-  void startup(ThreadPool* pool);
+  static void startup(int _self, int _pool);
 };
-struct ThreadPool {
+#ifndef MAX_WORKERS
+#define MAX_WORKERS 1
+#endif
+static constexpr int workerStack{65536};
+static std::array<std::byte, MAX_WORKERS * workerStack> stacks;
+struct WorkerPool {
   bool qLock{true}; // True is locked, false is unlocked
   bool done{};
   std::queue<std::function<void()>> taskQ;
-#ifndef MAX_THREADS
-#define MAX_THREADS 1
-#endif
-  std::array<Thread, MAX_THREADS> threads;
-#undef MAX_THREADS
-  ThreadPool();
-  ~ThreadPool();
+  std::array<Worker, MAX_WORKERS> workers;
+#undef MAX_WORKERS
+  WorkerPool();
+  ~WorkerPool();
   void exec(std::function<void()> fn);
 };
 
-extern ThreadPool globalPool;
-void fireEv(int index, const char* content, const char* type = nullptr);
+extern WorkerPool globalPool;
+void fireEv(int index, const char* _content, const char* _type = nullptr);
 int untar(unsigned char* tar, int tarSize, const std::string& storepath);

src/Wrapper.js

@@ -147,7 +147,4 @@ Module.createRecognizerWithSpkModel = (model, sampleRate, spkModel) => {
 
 Module.createRecognizerWithGrm = (model, sampleRate, grammar) => {
   return Recognizer.create(model.obj, sampleRate, 3, grammar, null)
-}
-
-// See Emscripten issue #21937
-if (!ENVIRONMENT_IS_PTHREAD) Module['mainScriptUrlOrBlob'] = new Blob([`importScripts('${_scriptName}')`], { type : "text/javascript" })
+}

src/make

@@ -73,12 +73,12 @@ if [ ! -d "$VOSK" ]; then
 fi
 
 cd "$SRC" &&
-em++ Util.cc CommonModel.cc Recognizer.cc Bindings.cc -O3 -Wno-pthreads-mem-growth -DEMSCRIPTEN_HAS_UNBOUND_TYPE_NAMES=0 -fno-rtti -DMAX_THREADS="$MAX_THREADS" -sWASMFS -sWASM_BIGINT -sMODULARIZE -sEMBIND_STD_STRING_IS_UTF8 -sPTHREAD_POOL_DELAY_LOAD -sTEXTDECODER=2 -sPTHREAD_POOL_SIZE_STRICT=2 -sINITIAL_MEMORY="$INITIAL_MEMORY" -sALLOW_MEMORY_GROWTH -sPTHREAD_POOL_SIZE="$MAX_THREADS" -sPOLYFILL=0 -sEXIT_RUNTIME=0 -sINVOKE_RUN=0 -sSUPPORT_LONGJMP=0 -sALLOW_BLOCKING_ON_MAIN_THREAD=0 -sEXPORTED_FUNCTIONS=_malloc -sEXPORT_NAME=loadVosklet -sMALLOC=emmalloc -sEXPORTED_RUNTIME_METHODS=UTF8ToString,stringToUTF8OnStack -sENVIRONMENT=web,worker -I. -I"$VOSK"/src -L"$KALDI"/src -l:online2/kaldi-online2.a -l:decoder/kaldi-decoder.a -l:ivector/kaldi-ivector.a -l:gmm/kaldi-gmm.a -l:tree/kaldi-tree.a -l:feat/kaldi-feat.a -l:cudamatrix/kaldi-cudamatrix.a -l:lat/kaldi-lat.a -l:lm/kaldi-lm.a -l:rnnlm/kaldi-rnnlm.a -l:hmm/kaldi-hmm.a -l:nnet3/kaldi-nnet3.a -l:transform/kaldi-transform.a -l:matrix/kaldi-matrix.a -l:fstext/kaldi-fstext.a -l:util/kaldi-util.a -l:base/kaldi-base.a -L"$OPENFST"/lib -l:libfst.a -l:libfstngram.a -L"$CLAPACK_WASM" -l:CBLAS/lib/cblas.a -l:CLAPACK-3.2.1/lapack.a -l:CLAPACK-3.2.1/libcblaswr.a -l:f2c_BLAS-3.8.0/blas.a -l:libf2c/libf2c.a -L"$VOSK"/src -l:vosk.a -lembind -pthread -flto -msimd128 -mreference-types -mnontrapping-fptoint -mextended-const -msign-ext -mmutable-globals --pre-js Wrapper.js -o ../Vosklet.js &&
+em++ Util.cc CommonModel.cc Recognizer.cc Bindings.cc -O3 -Wall -Werror -DEMSCRIPTEN_HAS_UNBOUND_TYPE_NAMES=0 -fno-rtti -DMAX_WORKERS="$MAX_THREADS" -sWASMFS -sWASM_BIGINT -sMODULARIZE -sTEXTDECODER=2 -sWASM_WORKERS=2 -sEVAL_CTORS=2 -sINITIAL_MEMORY="$INITIAL_MEMORY" -sALLOW_MEMORY_GROWTH -sPOLYFILL=0 -sEXIT_RUNTIME=0 -sINVOKE_RUN=0 -sSUPPORT_LONGJMP=0 -sEXPORTED_FUNCTIONS=_malloc -sEXPORT_NAME=loadVosklet -sMALLOC=emmalloc -sEXPORTED_RUNTIME_METHODS=UTF8ToString,stringToUTF8OnStack -sENVIRONMENT=web,worker -I. -I"$VOSK"/src -L"$KALDI"/src -l:online2/kaldi-online2.a -l:decoder/kaldi-decoder.a -l:ivector/kaldi-ivector.a -l:gmm/kaldi-gmm.a -l:tree/kaldi-tree.a -l:feat/kaldi-feat.a -l:cudamatrix/kaldi-cudamatrix.a -l:lat/kaldi-lat.a -l:lm/kaldi-lm.a -l:rnnlm/kaldi-rnnlm.a -l:hmm/kaldi-hmm.a -l:nnet3/kaldi-nnet3.a -l:transform/kaldi-transform.a -l:matrix/kaldi-matrix.a -l:fstext/kaldi-fstext.a -l:util/kaldi-util.a -l:base/kaldi-base.a -L"$OPENFST"/lib -l:libfst.a -l:libfstngram.a -L"$CLAPACK_WASM" -l:CBLAS/lib/cblas.a -l:CLAPACK-3.2.1/lapack.a -l:CLAPACK-3.2.1/libcblaswr.a -l:f2c_BLAS-3.8.0/blas.a -l:libf2c/libf2c.a -L"$VOSK"/src -l:vosk.a -lembind -flto -msimd128 -mreference-types -mnontrapping-fptoint -mextended-const -msign-ext -mmutable-globals --pre-js Wrapper.js -o ../Vosklet.js
 cd .. &&
 rm -f Vosklet.worker.js
 
 cp Vosklet.js Examples/Vosklet.js &&
-cp Vosklet.wasm Example/Vosklet.wasm &&
+cp Vosklet.wasm Examples/Vosklet.wasm &&
 
 # Can't serve files from raw.githubusercontent with Content-Encoding: gzip header so the browser won't decompress automatically. Manually decompressing instead.
 sed -i 's/res.body/new Response(res.body.pipeThrough(new DecompressionStream("gzip"))).body/' Examples/Vosklet.js &&

test

@@ -39,11 +39,9 @@ cd src &&
 MODE=1 && # 0: Ultra debug info, 1: Optimized release, else custom
 echo "Mode = $MODE" &&
 if [ "$MODE" = 0 ]; then
-  em++ Util.cc CommonModel.cc Recognizer.cc Bindings.cc -O0 -Wno-pthreads-mem-growth -DEMSCRIPTEN_HAS_UNBOUND_TYPE_NAMES=0 -fno-rtti -DMAX_THREADS="$MAX_THREADS" -Wall -Werror -Wno-pthreads-mem-growth -sWASMFS -sWASM_BIGINT -sMODULARIZE -sEMBIND_STD_STRING_IS_UTF8 -sPTHREAD_POOL_DELAY_LOAD -sRUNTIME_DEBUG -sALLOW_MEMORY_GROWTH -sSTACK_OVERFLOW_CHECK=2 -sTEXTDECODER=2 -sPTHREAD_POOL_SIZE_STRICT=2 -sASSERTIONS=2 -sINITIAL_MEMORY="$INITIAL_MEMORY" -sPTHREAD_POOL_SIZE="$MAX_THREADS" -sDISABLE_EXCEPTION_CATCHING=0 -sEXIT_RUNTIME=0 -sINVOKE_RUN=0 -sPOLYFILL=0 -sALLOW_BLOCKING_ON_MAIN_THREAD=0 -sEXPORTED_FUNCTIONS=_malloc -sEXPORT_NAME=loadVosklet -sMALLOC=emmalloc -sEXPORTED_RUNTIME_METHODS=UTF8ToString,stringToUTF8OnStack -sENVIRONMENT=web,worker -I. -I"$VOSK"/src -L"$KALDI"/src -l:online2/kaldi-online2.a -l:decoder/kaldi-decoder.a -l:ivector/kaldi-ivector.a -l:gmm/kaldi-gmm.a -l:tree/kaldi-tree.a -l:feat/kaldi-feat.a -l:cudamatrix/kaldi-cudamatrix.a -l:lat/kaldi-lat.a -l:lm/kaldi-lm.a -l:rnnlm/kaldi-rnnlm.a -l:hmm/kaldi-hmm.a -l:nnet3/kaldi-nnet3.a -l:transform/kaldi-transform.a -l:matrix/kaldi-matrix.a -l:fstext/kaldi-fstext.a -l:util/kaldi-util.a -l:base/kaldi-base.a -L"$OPENFST"/lib -l:libfst.a -l:libfstngram.a -L"$CLAPACK_WASM" -l:CBLAS/lib/cblas.a -l:CLAPACK-3.2.1/lapack.a -l:CLAPACK-3.2.1/libcblaswr.a -l:f2c_BLAS-3.8.0/blas.a -l:libf2c/libf2c.a -L"$VOSK"/src -l:vosk.a -lembind -pthread -flto -fsanitize=undefined -fsanitize=address -fsanitize=leak -msimd128 -mreference-types -mnontrapping-fptoint -mextended-const -msign-ext -mmutable-globals -g3 --pre-js Wrapper.js -o ../test.js
+  em++ Util.cc CommonModel.cc Recognizer.cc Bindings.cc -O0 -DEMSCRIPTEN_HAS_UNBOUND_TYPE_NAMES=0 -fno-rtti -DMAX_WORKERS="$MAX_THREADS" -sWASMFS -sWASM_BIGINT -sMODULARIZE -sTEXTDECODER=2 -sWASM_WORKERS=2 -sEVAL_CTORS=2 -sSTACK_OVERFLOW_CHECK=2 -sASSERTIONS=2 -sINITIAL_MEMORY="$INITIAL_MEMORY" -sALLOW_MEMORY_GROWTH -sRUNTIME_DEBUG -sPOLYFILL=0 -sEXIT_RUNTIME=0 -sINVOKE_RUN=0  -sDISABLE_EXCEPTION_CATCHING=0 -sEXPORTED_FUNCTIONS=_malloc -sEXPORT_NAME=loadVosklet -sMALLOC=emmalloc -sEXPORTED_RUNTIME_METHODS=UTF8ToString -sENVIRONMENT=web,worker -I. -I"$VOSK"/src -L"$KALDI"/src -l:online2/kaldi-online2.a -l:decoder/kaldi-decoder.a -l:ivector/kaldi-ivector.a -l:gmm/kaldi-gmm.a -l:tree/kaldi-tree.a -l:feat/kaldi-feat.a -l:cudamatrix/kaldi-cudamatrix.a -l:lat/kaldi-lat.a -l:lm/kaldi-lm.a -l:rnnlm/kaldi-rnnlm.a -l:hmm/kaldi-hmm.a -l:nnet3/kaldi-nnet3.a -l:transform/kaldi-transform.a -l:matrix/kaldi-matrix.a -l:fstext/kaldi-fstext.a -l:util/kaldi-util.a -l:base/kaldi-base.a -L"$OPENFST"/lib -l:libfst.a -l:libfstngram.a -L"$CLAPACK_WASM" -l:CBLAS/lib/cblas.a -l:CLAPACK-3.2.1/lapack.a -l:CLAPACK-3.2.1/libcblaswr.a -l:f2c_BLAS-3.8.0/blas.a -l:libf2c/libf2c.a -L"$VOSK"/src -l:vosk.a -lembind -flto -msimd128 -mreference-types -mnontrapping-fptoint -mextended-const -msign-ext -mmutable-globals --pre-js Wrapper.js -o ../test.js
 elif [ "$MODE" = 1 ]; then
-  em++ Util.cc CommonModel.cc Recognizer.cc Bindings.cc -O3 -Wno-pthreads-mem-growth -DEMSCRIPTEN_HAS_UNBOUND_TYPE_NAMES=0 -fno-rtti -DMAX_THREADS="$MAX_THREADS" -sWASMFS -sWASM_BIGINT -sMODULARIZE -sEMBIND_STD_STRING_IS_UTF8 -sPTHREAD_POOL_DELAY_LOAD -sTEXTDECODER=2 -sPTHREAD_POOL_SIZE_STRICT=2 -sINITIAL_MEMORY="$INITIAL_MEMORY" -sALLOW_MEMORY_GROWTH -sPTHREAD_POOL_SIZE="$MAX_THREADS" -sPOLYFILL=0 -sEXIT_RUNTIME=0 -sINVOKE_RUN=0 -sSUPPORT_LONGJMP=0 -sALLOW_BLOCKING_ON_MAIN_THREAD=0 -sEXPORTED_FUNCTIONS=_malloc -sEXPORT_NAME=loadVosklet -sMALLOC=emmalloc -sEXPORTED_RUNTIME_METHODS=UTF8ToString,stringToUTF8OnStack -sENVIRONMENT=web,worker -I. -I"$VOSK"/src -L"$KALDI"/src -l:online2/kaldi-online2.a -l:decoder/kaldi-decoder.a -l:ivector/kaldi-ivector.a -l:gmm/kaldi-gmm.a -l:tree/kaldi-tree.a -l:feat/kaldi-feat.a -l:cudamatrix/kaldi-cudamatrix.a -l:lat/kaldi-lat.a -l:lm/kaldi-lm.a -l:rnnlm/kaldi-rnnlm.a -l:hmm/kaldi-hmm.a -l:nnet3/kaldi-nnet3.a -l:transform/kaldi-transform.a -l:matrix/kaldi-matrix.a -l:fstext/kaldi-fstext.a -l:util/kaldi-util.a -l:base/kaldi-base.a -L"$OPENFST"/lib -l:libfst.a -l:libfstngram.a -L"$CLAPACK_WASM" -l:CBLAS/lib/cblas.a -l:CLAPACK-3.2.1/lapack.a -l:CLAPACK-3.2.1/libcblaswr.a -l:f2c_BLAS-3.8.0/blas.a -l:libf2c/libf2c.a -L"$VOSK"/src -l:vosk.a -lembind -pthread -flto -msimd128 -mreference-types -mnontrapping-fptoint -mextended-const -msign-ext -mmutable-globals --pre-js Wrapper.js -o ../test.js
-else
-  :
+  em++ Util.cc CommonModel.cc Recognizer.cc Bindings.cc -O3 -Wall -Werror -DEMSCRIPTEN_HAS_UNBOUND_TYPE_NAMES=0 -fno-rtti -DMAX_WORKERS="$MAX_THREADS" -sWASMFS -sWASM_BIGINT -sMODULARIZE -sTEXTDECODER=2 -sWASM_WORKERS=2 -sEVAL_CTORS=2 -sINITIAL_MEMORY="$INITIAL_MEMORY" -sALLOW_MEMORY_GROWTH -sPOLYFILL=0 -sEXIT_RUNTIME=0 -sINVOKE_RUN=0 -sSUPPORT_LONGJMP=0 -sEXPORTED_FUNCTIONS=_malloc -sEXPORT_NAME=loadVosklet -sMALLOC=emmalloc -sEXPORTED_RUNTIME_METHODS=UTF8ToString,stringToUTF8OnStack -sENVIRONMENT=web,worker -I. -I"$VOSK"/src -L"$KALDI"/src -l:online2/kaldi-online2.a -l:decoder/kaldi-decoder.a -l:ivector/kaldi-ivector.a -l:gmm/kaldi-gmm.a -l:tree/kaldi-tree.a -l:feat/kaldi-feat.a -l:cudamatrix/kaldi-cudamatrix.a -l:lat/kaldi-lat.a -l:lm/kaldi-lm.a -l:rnnlm/kaldi-rnnlm.a -l:hmm/kaldi-hmm.a -l:nnet3/kaldi-nnet3.a -l:transform/kaldi-transform.a -l:matrix/kaldi-matrix.a -l:fstext/kaldi-fstext.a -l:util/kaldi-util.a -l:base/kaldi-base.a -L"$OPENFST"/lib -l:libfst.a -l:libfstngram.a -L"$CLAPACK_WASM" -l:CBLAS/lib/cblas.a -l:CLAPACK-3.2.1/lapack.a -l:CLAPACK-3.2.1/libcblaswr.a -l:f2c_BLAS-3.8.0/blas.a -l:libf2c/libf2c.a -L"$VOSK"/src -l:vosk.a -lembind -flto -msimd128 -mreference-types -mnontrapping-fptoint -mextended-const -msign-ext -mmutable-globals --pre-js Wrapper.js -o ../test.js
 fi
 sed -i 's/res.body/new Response(res.body.pipeThrough(new DecompressionStream("gzip"))).body/' ../test.js
 rm -f ../test.worker.js