G1: JUCE engine SHM chain rong POC - probe pass, khong lag them

.gitignore

@@ -61,3 +61,4 @@ native_bridge/debug/shm_selfcheck.exe
 # Linux-only runtime artifacts (built via Docker; never committed)
 install/daw_vst_bridge
 install/libsfizz.so.1
+build_juce

.gitmodules

@@ -1,3 +1,8 @@
 [submodule "native_bridge/vst3sdk"]
 	path = native_bridge/vst3sdk
 	url = https://github.com/steinbergmedia/vst3sdk.git
+[submodule "native_bridge/JUCE"]
+	path = native_bridge/JUCE
+	url = https://github.com/juce-framework/JUCE
+	branch = 8.0.7
+	shallow = true

app/core/fx_realtime.py

@@ -162,7 +162,8 @@ class FxRealtimeSession:
         return out
 
     def _spawn(self):
-        exe = find_fx_bridge_exe()
+        juce_exe = os.getenv("SF_JUCE_FX_BRIDGE_PATH", "").strip()
+        exe = juce_exe or find_fx_bridge_exe()
         if not exe:
             raise RuntimeError("Không tìm thấy fx_vst_bridge.exe")
         job = {
@@ -174,8 +175,12 @@ class FxRealtimeSession:
         fd, self._job_path = tempfile.mkstemp(suffix=".json", prefix="fxrt_")
         with os.fdopen(fd, "w", encoding="utf-8") as f:
             json.dump(job, f)
-        cmd = [exe, "--realtime-fx", self._job_path, "--shm", self.name,
-               "--parent", str(os.getpid())]
+        if juce_exe:
+            cmd = [exe, "--juce-fx", self._job_path, "--shm", self.name,
+                   "--parent", str(os.getpid())]
+        else:
+            cmd = [exe, "--realtime-fx", self._job_path, "--shm", self.name,
+                   "--parent", str(os.getpid())]
         # FXRT_BRIDGE_LOG=<path>: bat stderr bridge (FxRTPerf) vao file — chan
         # doan outFull/drop; mac dinh DEVNULL nhu cu.
         # V34 diag: mac dinh ghi FxRTPerf vao fxrt_bridge.log de phan tich

native_bridge/include/FxShm.h

@@ -0,0 +1,50 @@
+// native_bridge/include/FxShm.h
+// openShm/closeShm dùng chung giữa RealtimeFxLoop.cpp (fx_vst_bridge
+// --realtime-fx) và juce_fx/JuceFxLoop.cpp (juce_fx_bridge --juce-fx).
+// Windows-only (FileMapping); mở mapping có sẵn do engine (Python) tạo.
+#pragma once
+#ifdef _WIN32
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+#include <string>
+
+namespace fxshm {
+
+struct ShmView {
+    void* map = nullptr;
+    void* view = nullptr;
+};
+
+inline ShmView* openShm(const std::string& name, size_t size) {
+#ifdef _WIN32
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, name.c_str(), -1, nullptr, 0);
+    std::wstring wname(wlen, L'\0');
+    MultiByteToWideChar(CP_UTF8, 0, name.c_str(), -1, &wname[0], wlen);
+    HANDLE map = OpenFileMappingW(FILE_MAP_ALL_ACCESS, FALSE, wname.c_str());
+    if (!map) return nullptr;
+    void* view = MapViewOfFile(map, FILE_MAP_ALL_ACCESS, 0, 0, size);
+    if (!view) { CloseHandle(map); return nullptr; }
+    ShmView* v = new ShmView();
+    v->map = map;
+    v->view = view;
+    return v;
+#else
+    (void)name; (void)size;
+    return nullptr;
+#endif
+}
+
+inline void closeShm(ShmView* v) {
+    if (!v) return;
+#ifdef _WIN32
+    if (v->view) UnmapViewOfFile(v->view);
+    if (v->map) CloseHandle((HANDLE)v->map);
+#endif
+    delete v;
+}
+
+}  // namespace fxshm

native_bridge/juce_fx/CMakeLists.txt

@@ -0,0 +1,35 @@
+# native_bridge/juce_fx/CMakeLists.txt
+# G1 POC: juce_fx_bridge.exe — JUCE AudioProcessorGraph thay RealtimeFxChain
+# trong SHM realtime loop. Windows only. Build riêng (không chung với
+# daw_vst_bridge) để JUCE không kéo opengl/fluidsynth vào bridge cũ.
+cmake_minimum_required(VERSION 3.20)
+project(juce_fx_bridge LANGUAGES CXX)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+if(NOT EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../JUCE/CMakeLists.txt")
+  message(FATAL_ERROR "JUCE submodule missing — run: git submodule update --init native_bridge/JUCE")
+endif()
+
+# JUCE framework (submodule, tag 8.0.7). add_subdirectory tạo targets
+# juce::core, juce::audio_basics, juce::audio_processors, ...
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/../JUCE JUCE EXCLUDE_FROM_ALL)
+
+add_executable(juce_fx_bridge
+  main_juce_fx.cpp
+  JuceFxLoop.cpp
+  JuceFxEngine.cpp
+)
+target_include_directories(juce_fx_bridge PRIVATE
+  ${CMAKE_CURRENT_SOURCE_DIR}/../include
+)
+target_link_libraries(juce_fx_bridge PRIVATE
+  juce::juce_core
+  juce::juce_audio_basics
+  juce::juce_audio_processors
+)
+if(WIN32)
+  # timeBeginPeriod (winmm) — như fx_vst_bridge
+  target_link_libraries(juce_fx_bridge PRIVATE winmm)
+endif()

native_bridge/juce_fx/JuceFxEngine.cpp

@@ -0,0 +1,82 @@
+// native_bridge/juce_fx/JuceFxEngine.cpp
+// Implementation: AudioProcessorGraph với input/output node — chain rỗng G1.
+#include "JuceFxEngine.h"
+
+#include <juce_audio_processors/juce_audio_processors.h>
+
+#include <algorithm>
+#include <array>
+#include <vector>
+
+struct JuceFxEngine::Impl {
+    juce::AudioProcessorGraph graph;
+    juce::AudioBuffer<float> buf;
+    uint32_t sampleRate = 0;
+    uint32_t blockSize = 0;
+    bool prepared = false;
+};
+
+JuceFxEngine::JuceFxEngine() : impl_(std::make_unique<Impl>()) {}
+
+JuceFxEngine::~JuceFxEngine() { shutdown(); }
+
+void JuceFxEngine::shutdown() {
+    if (impl_->prepared) {
+        impl_->graph.releaseResources();
+        impl_->graph.clear();
+        impl_->prepared = false;
+        impl_->sampleRate = 0;
+        impl_->blockSize = 0;
+    }
+}
+
+void JuceFxEngine::prepare(uint32_t sampleRate, uint32_t blockSize) {
+    if (impl_->prepared && impl_->sampleRate == sampleRate &&
+        impl_->blockSize == blockSize)
+        return;  // không đổi — giữ graph
+    shutdown();
+
+    impl_->sampleRate = sampleRate;
+    impl_->blockSize = blockSize;
+    impl_->graph.setPlayConfigDetails(2, 2, (double)sampleRate, (int)blockSize);
+
+    // Chain rỗng G1: audio input -> audio output (passthrough). G2: chèn
+    // plugin node giữa hai node này theo fx_chain.
+    auto& g = impl_->graph;
+    const auto inNode = g.addNode(std::make_unique<
+        juce::AudioProcessorGraph::AudioGraphIOProcessor>(
+        juce::AudioProcessorGraph::AudioGraphIOProcessor::audioInputNode));
+    const auto outNode = g.addNode(std::make_unique<
+        juce::AudioProcessorGraph::AudioGraphIOProcessor>(
+        juce::AudioProcessorGraph::AudioGraphIOProcessor::audioOutputNode));
+    using NCh = juce::AudioProcessorGraph::NodeAndChannel;
+    using Conn = juce::AudioProcessorGraph::Connection;
+    for (int ch = 0; ch < 2; ++ch) {
+        NCh src; src.nodeID = inNode->nodeID; src.channelIndex = ch;
+        NCh dst; dst.nodeID = outNode->nodeID; dst.channelIndex = ch;
+        g.addConnection(Conn(src, dst));
+    }
+
+    g.prepareToPlay((double)sampleRate, (int)(blockSize * 4));  // batch tối đa 4 block
+    impl_->buf.setSize(2, (int)(blockSize * 4), false, false, true);
+    impl_->prepared = true;
+}
+
+void JuceFxEngine::process(float* inL, float* inR, uint32_t n) {
+    if (!impl_->prepared || n == 0) return;
+    auto& b = impl_->buf;
+    b.setSize(2, (int)n, false, false, true);
+    b.copyFrom(0, 0, inL, (int)n);
+    b.copyFrom(1, 0, inR, (int)n);
+    juce::MidiBuffer midi;
+    impl_->graph.processBlock(b, midi);
+    // G1 passthrough: copy kết quả về buffer đầu vào (in-place như contract).
+    const auto* outL = b.getReadPointer(0);
+    const auto* outR = b.getReadPointer(1);
+    std::copy(outL, outL + n, inL);
+    std::copy(outR, outR + n, inR);
+}
+
+uint32_t JuceFxEngine::latencySamples() const {
+    return impl_->prepared ? (uint32_t)impl_->graph.getLatencySamples() : 0u;
+}

native_bridge/juce_fx/JuceFxEngine.h

@@ -0,0 +1,29 @@
+// native_bridge/juce_fx/JuceFxEngine.h
+// JUCE AudioProcessorGraph engine (G1): đọc input ring → graph (rỗng) →
+// ghi output ring. Tách khỏi loop để G2 chỉ thêm plugin vào graph.
+#pragma once
+#include <cstdint>
+#include <memory>
+
+class JuceFxEngine {
+public:
+    JuceFxEngine();
+    ~JuceFxEngine();
+    // Không copy.
+    JuceFxEngine(const JuceFxEngine&) = delete;
+    JuceFxEngine& operator=(const JuceFxEngine&) = delete;
+
+    // Tạo graph mới (input node -> output node, passthrough), prepareToPlay.
+    // sampleRate/blockSize đổi → teardown graph cũ + prepare lại (validate
+    // header G1). blockSize là kích thước mỗi SHM block (256).
+    void prepare(uint32_t sampleRate, uint32_t blockSize);
+    // Audio thread: xử lý n mẫu stereo in-place (n <= blockSize * FXRT_IN_SLOTS).
+    void process(float* inL, float* inR, uint32_t n);
+    // Graph latency samples (G1 chain rỗng = 0; G2 = getLatencySamples thật).
+    uint32_t latencySamples() const;
+    void shutdown();
+
+private:
+    struct Impl;
+    std::unique_ptr<Impl> impl_;
+};

native_bridge/juce_fx/JuceFxLoop.cpp

@@ -0,0 +1,241 @@
+// native_bridge/juce_fx/JuceFxLoop.cpp
+// juce_fx_bridge --juce-fx <job.json> --shm <name> [--parent <pid>]
+// G1 POC: vòng lặp SHM giống RealtimeFxLoop.cpp nhưng xử lý qua JuceFxEngine
+// (JUCE AudioProcessorGraph chain rỗng) thay vì RealtimeFxChain. Job format
+// giữ nguyên {sample_rate, block_size, fx_chain} — G1 bỏ qua fx_chain.
+// Validate header mỗi iteration: sampleRate/blockSize đổi giữa chừng →
+// teardown + prepareToPlay lại + report latency mới qua FxLatReport.
+#include "JuceFxEngine.h"
+#include "FxRealtimeIPC.h"
+#include "FxShm.h"
+
+#ifdef _WIN32
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <mmsystem.h>
+#endif
+
+#include <algorithm>
+#include <chrono>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <iterator>
+#include <string>
+#include <thread>
+#include <vector>
+
+namespace {
+
+static bool parentAlive(uint32_t pid) {
+    if (pid == 0) return true;
+    HANDLE h = OpenProcess(PROCESS_QUERY_LIMITED_INFORMATION, FALSE, pid);
+    if (!h) return false;
+    CloseHandle(h);
+    return true;
+}
+
+static void sleepMs(uint32_t ms) { Sleep(ms); }
+
+std::string readFile(const std::string& path) {
+    std::ifstream f(path, std::ios::binary);
+    return std::string((std::istreambuf_iterator<char>(f)),
+                       std::istreambuf_iterator<char>());
+}
+
+double jsonNumber(const std::string& s, const std::string& key, double def) {
+    std::string k = "\"" + key + "\"";
+    size_t p = s.find(k);
+    if (p == std::string::npos) return def;
+    p += k.size();
+    while (p < s.size() && (s[p] == ':' || s[p] == ' ' || s[p] == '\t' ||
+                            s[p] == '\r' || s[p] == '\n'))
+        ++p;
+    size_t e = p;
+    while (e < s.size() && (std::isdigit((unsigned char)s[e]) || s[e] == '.' ||
+                            s[e] == '-' || s[e] == '+' || s[e] == 'e' ||
+                            s[e] == 'E'))
+        ++e;
+    if (e == p) return def;
+    try { return std::stod(s.substr(p, e - p)); } catch (...) { return def; }
+}
+
+}  // namespace
+
+int run_juce_fx_loop(const std::string& jobPath, const std::string& shmName,
+                     uint32_t parentPid) {
+    std::cerr << "[JuceFxLoop] start job=" << jobPath << " shm=" << shmName
+              << " parent=" << parentPid << std::endl;
+    const std::string job = readFile(jobPath);
+    if (job.empty()) {
+        std::cerr << "[JuceFxLoop] cannot read job file" << std::endl;
+        return 1;
+    }
+    const double srD = jsonNumber(job, "sample_rate", 44100.0);
+    const uint32_t sampleRate = (uint32_t)srD;
+    const double blkD = jsonNumber(job, "block_size", (double)FXRT_BLOCK);
+    const uint32_t block = (uint32_t)std::max<double>(32.0, std::min<double>(blkD, (double)FXRT_BLOCK));
+    // G1: fx_chain bỏ qua (chain rỗng). G2: đọc fx_chain -> build graph.
+
+    fxshm::ShmView* v = fxshm::openShm(shmName, sizeof(FxRealtimeIPC));
+    if (!v) {
+        std::cerr << "[JuceFxLoop] cannot open SHM: " << shmName
+                  << " (engine phải tạo trước)" << std::endl;
+        return 2;
+    }
+    auto* ipc = static_cast<FxRealtimeIPC*>(v->view);
+
+    for (int i = 0; i < 200 && ipc->h.magic != FXRT_MAGIC; ++i) sleepMs(10);
+    if (ipc->h.magic != FXRT_MAGIC) {
+        std::cerr << "[JuceFxLoop] SHM magic mismatch (engine chưa init?)" << std::endl;
+        fxshm::closeShm(v);
+        return 2;
+    }
+    if (ipc->h.inSlots != FXRT_IN_SLOTS || ipc->h.outSlots != FXRT_OUT_SLOTS) {
+        std::cerr << "[JuceFxLoop] slot count mismatch" << std::endl;
+        fxshm::closeShm(v);
+        return 2;
+    }
+
+#ifdef _WIN32
+    timeBeginPeriod(1);
+    SetThreadPriority(GetCurrentThread(), THREAD_PRIORITY_ABOVE_NORMAL);
+#endif
+
+    // Engine: prepare graph (chain rỗng). sr/block từ job; header là nguồn
+    // thật (engine ghi lúc tạo SHM) — nếu khác, lấy header.
+    uint32_t curSr = sampleRate;
+    uint32_t curBlock = block;
+    JuceFxEngine engine;
+    if (ipc->h.sampleRate) curSr = ipc->h.sampleRate;
+    if (ipc->h.blockSize) curBlock = ipc->h.blockSize;
+    engine.prepare(curSr, curBlock);
+    std::cerr << "[JuceFxLoop] prepared sr=" << curSr << " block=" << curBlock
+              << " latency=" << engine.latencySamples() << std::endl;
+
+    // Report latency ban đầu (slot 0 = tổng chain).
+    {
+        const uint32_t ls = ipc->h.latWrite & (FXRT_LAT_SLOTS - 1);
+        ipc->lat[ls].slot = 0;
+        ipc->lat[ls].samples = engine.latencySamples();
+        MemoryBarrier();
+        ipc->h.latWrite++;
+    }
+    ipc->h.state = FXRT_STATE_READY;
+    std::cerr << "[JuceFxLoop] ready" << std::endl;
+
+    std::thread hb([&]() {
+        while (ipc->h.running && ipc->h.state == FXRT_STATE_READY) {
+            ipc->h.heartbeat++;
+            sleepMs(100);
+        }
+    });
+
+    const uint32_t n = curBlock;
+    const uint32_t inMask = ipc->h.inSlots - 1;
+    const uint32_t outMask = ipc->h.outSlots - 1;
+    float L[FXRT_BLOCK * FXRT_IN_SLOTS], R[FXRT_BLOCK * FXRT_IN_SLOTS];
+    uint64_t processed = 0;
+    uint64_t perfIter = 0, perfProc = 0, perfIdle = 0, perfOutFull = 0;
+    double perfProcSum = 0.0, perfProcMax = 0.0;
+    double perfLoopSum = 0.0, perfLoopMax = 0.0;
+    uint32_t perfTakeSum = 0;
+    auto perfT0 = std::chrono::steady_clock::now();
+    auto perfRunStart = perfT0;
+
+    while (ipc->h.running) {
+        if (parentPid && !parentAlive(parentPid)) {
+            std::cerr << "[JuceFxLoop] parent gone — exiting" << std::endl;
+            break;
+        }
+        // Validate header mỗi iteration: sampleRate/blockSize đổi giữa chừng
+        // (đổi thiết bị audio / session mới khác rate) → teardown graph +
+        // prepareToPlay lại + report latency mới. WebAudio sampleRate bất
+        // biến — check này phòng header bị ghi lại.
+        if (ipc->h.sampleRate && (ipc->h.sampleRate != curSr ||
+                                  ipc->h.blockSize != curBlock)) {
+            curSr = ipc->h.sampleRate;
+            curBlock = ipc->h.blockSize;
+            engine.prepare(curSr, curBlock);
+            const uint32_t ls = ipc->h.latWrite & (FXRT_LAT_SLOTS - 1);
+            ipc->lat[ls].slot = 0;
+            ipc->lat[ls].samples = engine.latencySamples();
+            MemoryBarrier();
+            ipc->h.latWrite++;
+            std::cerr << "[JuceFxLoop] re-prepared sr=" << curSr
+                      << " block=" << curBlock
+                      << " latency=" << engine.latencySamples() << std::endl;
+        }
+        // OUT-ring backpressure (giữ nguyên cơ chế RealtimeFxLoop — pointer
+        // collision → torn frame nếu ghi đè slot chưa đọc).
+        const uint32_t outFree = ipc->h.outSlots - (ipc->h.outWrite - ipc->h.outRead);
+        if (outFree < std::min<uint32_t>(ipc->h.inWrite - ipc->h.inRead, FXRT_IN_SLOTS)) {
+            ++perfOutFull;
+            std::this_thread::sleep_until(std::chrono::steady_clock::now() +
+                                          std::chrono::milliseconds(1));
+            continue;
+        }
+        const uint32_t avail = ipc->h.inWrite - ipc->h.inRead;
+        if (avail == 0) {
+            ++perfIdle;
+            std::this_thread::sleep_until(std::chrono::steady_clock::now() +
+                                          std::chrono::milliseconds(1));
+            continue;
+        }
+        const uint32_t take = std::min<uint32_t>(avail, FXRT_IN_SLOTS);
+        uint32_t off = 0;
+        for (uint32_t i = 0; i < take; ++i) {
+            const uint32_t slot = ipc->h.inRead & inMask;
+            std::memcpy(L + off, ipc->inL[slot], n * sizeof(float));
+            std::memcpy(R + off, ipc->inR[slot], n * sizeof(float));
+            ipc->h.inRead++;
+            off += n;
+        }
+        const auto perfP0 = std::chrono::steady_clock::now();
+        engine.process(L, R, off);  // chain rỗng = passthrough
+        const auto perfP1 = std::chrono::steady_clock::now();
+        const double perfProcMs = std::chrono::duration<double, std::milli>(perfP1 - perfP0).count();
+        ++perfProc; perfProcSum += perfProcMs; perfTakeSum += take;
+        if (perfProcMs > perfProcMax) perfProcMax = perfProcMs;
+        for (uint32_t i = 0, o = 0; i < take; ++i, o += n) {
+            const uint32_t oslot = ipc->h.outWrite & outMask;
+            std::memcpy(ipc->outL[oslot], L + o, n * sizeof(float));
+            std::memcpy(ipc->outR[oslot], R + o, n * sizeof(float));
+            MemoryBarrier();
+            ipc->h.outWrite++;
+            ++processed;
+        }
+        const auto perfIterT1 = std::chrono::steady_clock::now();
+        const double perfLoopMs = std::chrono::duration<double, std::milli>(perfIterT1 - perfT0).count();
+        perfT0 = perfIterT1; ++perfIter;
+        perfLoopSum += perfLoopMs; if (perfLoopMs > perfLoopMax) perfLoopMax = perfLoopMs;
+        if (perfIter % 100 == 0) {
+            const double runSec = std::chrono::duration<double>(std::chrono::steady_clock::now() - perfRunStart).count();
+            std::cerr << "[JuceFxPerf] iter=" << perfIter
+                      << " procN=" << perfProc
+                      << " procAvg=" << (perfProc ? perfProcSum / perfProc : 0.0) << "ms"
+                      << " procMax=" << perfProcMax << "ms"
+                      << " loopAvg=" << (perfIter ? perfLoopSum / perfIter : 0.0) << "ms"
+                      << " loopMax=" << perfLoopMax << "ms"
+                      << " idle=" << perfIdle
+                      << " outFull=" << perfOutFull
+                      << " takeAvg=" << (perfProc ? (double)perfTakeSum / perfProc : 0.0)
+                      << " procBlocks=" << processed
+                      << " rate=" << (runSec > 0 ? processed / runSec : 0.0) << "blk/s"
+                      << " outDepth=" << (ipc->h.outWrite - ipc->h.outRead)
+                      << std::endl;
+            perfProc = 0; perfProcSum = 0.0; perfTakeSum = 0; perfIdle = 0; perfOutFull = 0;
+        }
+    }
+    ipc->h.state = FXRT_STATE_STARTING;
+    hb.join();
+    engine.shutdown();
+    fxshm::closeShm(v);
+#ifdef _WIN32
+    timeEndPeriod(1);
+#endif
+    std::cerr << "[JuceFxLoop] exit processed=" << processed << std::endl;
+    return 0;
+}

native_bridge/juce_fx/JuceFxLoop.h

@@ -0,0 +1,8 @@
+// native_bridge/juce_fx/JuceFxLoop.h
+#pragma once
+#include <cstdint>
+#include <string>
+
+// juce_fx_bridge --juce-fx <job.json> --shm <name> [--parent <pid>]
+int run_juce_fx_loop(const std::string& jobPath, const std::string& shmName,
+                     uint32_t parentPid);

native_bridge/juce_fx/main_juce_fx.cpp

@@ -0,0 +1,52 @@
+// native_bridge/juce_fx/main_juce_fx.cpp
+// Entry point của juce_fx_bridge.exe — G1 POC: JUCE AudioProcessorGraph
+// thay RealtimeFxChain trong SHM realtime loop. Windows only.
+//   --juce-fx <job.json> --shm <name> [--parent <pid>]
+#include "JuceFxLoop.h"
+
+#ifdef _WIN32
+#include <windows.h>
+#include <cstdio>
+#include <cstdlib>
+#endif
+
+#include <cstring>
+#include <iostream>
+#include <string>
+
+#ifdef _WIN32
+static LONG WINAPI JuceCrashDumpHandler(EXCEPTION_POINTERS* ep) {
+    static LONG once = 0;
+    if (InterlockedCompareExchange(&once, 1, 0)) return EXCEPTION_CONTINUE_SEARCH;
+    PEXCEPTION_RECORD er = ep ? ep->ExceptionRecord : nullptr;
+    std::cerr << "[CRASH] juce_fx_bridge exception code=0x" << std::hex
+              << (er ? er->ExceptionCode : 0) << " addr=0x"
+              << (er ? (void*)er->ExceptionAddress : nullptr) << std::dec << std::endl;
+    return EXCEPTION_CONTINUE_SEARCH;
+}
+#endif
+
+int main(int argc, char** argv) {
+#ifdef _WIN32
+    SetUnhandledExceptionFilter(JuceCrashDumpHandler);
+#endif
+    std::string mode, jobPath, shmName;
+    uint32_t parentPid = 0;
+    for (int i = 1; i < argc; ++i) {
+        const std::string a = argv[i];
+        if (a == "--juce-fx" && i + 1 < argc) {
+            mode = a;
+            jobPath = argv[++i];
+        } else if (a == "--shm" && i + 1 < argc) {
+            shmName = argv[++i];
+        } else if (a == "--parent" && i + 1 < argc) {
+            parentPid = (uint32_t)std::strtoul(argv[++i], nullptr, 10);
+        }
+    }
+    if (mode != "--juce-fx" || jobPath.empty() || shmName.empty()) {
+        std::cerr << "usage: juce_fx_bridge --juce-fx <job.json> --shm <name> "
+                     "[--parent <pid>]" << std::endl;
+        return 1;
+    }
+    return run_juce_fx_loop(jobPath, shmName, parentPid);
+}

native_bridge/src/RealtimeFxLoop.cpp

@@ -10,6 +10,7 @@
 //   nhận JSON mảng {path,bypass,preset_b64} (type/name bỏ qua).
 #include "RenderFxJob.h"
 #include "FxRealtimeIPC.h"
+#include "FxShm.h"
 
 #ifdef _WIN32
 #ifndef NOMINMAX
@@ -31,32 +32,6 @@
 
 namespace {
 
-struct ShmView {
-    void* map = nullptr;
-    void* view = nullptr;
-};
-
-ShmView* openShm(const std::string& name, size_t size) {
-    int wlen = MultiByteToWideChar(CP_UTF8, 0, name.c_str(), -1, nullptr, 0);
-    std::wstring wname(wlen, L'\0');
-    MultiByteToWideChar(CP_UTF8, 0, name.c_str(), -1, &wname[0], wlen);
-    HANDLE map = OpenFileMappingW(FILE_MAP_ALL_ACCESS, FALSE, wname.c_str());
-    if (!map) return nullptr;
-    void* view = MapViewOfFile(map, FILE_MAP_ALL_ACCESS, 0, 0, size);
-    if (!view) { CloseHandle(map); return nullptr; }
-    ShmView* v = new ShmView();
-    v->map = map;
-    v->view = view;
-    return v;
-}
-
-void closeShm(ShmView* v) {
-    if (!v) return;
-    if (v->view) UnmapViewOfFile(v->view);
-    if (v->map) CloseHandle((HANDLE)v->map);
-    delete v;
-}
-
 static bool parentAlive(uint32_t pid) {
     if (pid == 0) return true;
     HANDLE h = OpenProcess(PROCESS_QUERY_LIMITED_INFORMATION, FALSE, pid);
@@ -137,7 +112,7 @@ int run_realtime_fx_loop(const std::string& jobPath, const std::string& shmName,
     RealtimeFxChain chain;
     chain.setChain(chainJson, sampleRate, (int32_t)batchN);
 
-    ShmView* v = openShm(shmName, sizeof(FxRealtimeIPC));
+    fxshm::ShmView* v = fxshm::openShm(shmName, sizeof(FxRealtimeIPC));
     if (!v) {
         std::cerr << "[RealtimeFxLoop] cannot open SHM: " << shmName
                   << " (engine phải tạo trước)" << std::endl;
@@ -149,12 +124,12 @@ int run_realtime_fx_loop(const std::string& jobPath, const std::string& shmName,
     for (int i = 0; i < 200 && ipc->h.magic != FXRT_MAGIC; ++i) sleepMs(10);
     if (ipc->h.magic != FXRT_MAGIC) {
         std::cerr << "[RealtimeFxLoop] SHM magic mismatch (engine chưa init?)" << std::endl;
-        closeShm(v);
+        fxshm::closeShm(v);
         return 2;
     }
     if (ipc->h.inSlots != FXRT_IN_SLOTS || ipc->h.outSlots != FXRT_OUT_SLOTS) {
         std::cerr << "[RealtimeFxLoop] slot count mismatch" << std::endl;
-        closeShm(v);
+        fxshm::closeShm(v);
         return 2;
     }
     // Windows timer resolution 1ms — bỏ granularity Sleep 15.6ms cho toàn
@@ -359,7 +334,7 @@ int run_realtime_fx_loop(const std::string& jobPath, const std::string& shmName,
     }
     ipc->h.state = FXRT_STATE_STARTING;  // đã dừng
     hb.join();
-    closeShm(v);
+    fxshm::closeShm(v);
 #ifdef _WIN32
     timeEndPeriod(1);
 #endif