Fix AMD and OpenVINO

.claude/settings.local.json

@@ -81,7 +81,9 @@
       "Bash(cmake --build build --target ANSUtilities)",
       "Bash(ls -d /c/Projects/CLionProjects/ANSCORE/cmake-build-* /c/Projects/CLionProjects/ANSCORE/out/*)",
       "Bash(cmake --build /c/Projects/CLionProjects/ANSCORE/cmake-build-release --target ANSUtilities)",
-      "Bash(find /c/Projects/CLionProjects/ANSCORE -name *json* -o -name *Json*)"
+      "Bash(find /c/Projects/CLionProjects/ANSCORE -name *json* -o -name *Json*)",
+      "Bash(grep -n \"CreateANSALPRHandle\\\\|LoadANSALPREngineHandle\\\\|CreateANSRTSPHandle\\\\|ReleaseANSALPRHandle\\\\|ANSALPR_RunInference\" C:ProjectsCLionProjectsANSCOREsrc*)",
+      "Bash(find C:ProjectsCLionProjectsANSCORE -name ANSLibsLoader* -type f)"
     ]
   }
 }

core/ANSLibsLoader/EPLoader.cpp

@@ -179,9 +179,12 @@ namespace ANSCENTER {
     EngineType EPLoader::AutoDetect()
     {
         std::cout << "[EPLoader] Auto-detecting hardware..." << std::endl;
+        ANS_DBG("EPLoader", "AutoDetect: starting hardware detection");
         ANSLicenseHelper helper;
         EngineType detected = helper.CheckHardwareInformation();
         std::cout << "[EPLoader] Detected: " << EngineTypeName(detected) << std::endl;
+        ANS_DBG("EPLoader", "AutoDetect: result=%d (%s)",
+            static_cast<int>(detected), EngineTypeName(detected));
         return detected;
     }
 
@@ -217,6 +220,9 @@ namespace ANSCENTER {
         s_info.fromSubdir = (ep_dir != shared_dir);
         s_initialized = true;
 
+        ANS_DBG("EPLoader", "Initialize: EP=%d (%s) dir=%s fromSubdir=%d",
+            static_cast<int>(type), EngineTypeName(type),
+            ep_dir.c_str(), s_info.fromSubdir ? 1 : 0);
         std::cout << "[EPLoader] Ready. EP=" << EngineTypeName(type)
             << " dir=" << ep_dir << std::endl;
         return s_info;

core/ANSLicensingSystem/ANSLicense.h

@@ -8,7 +8,7 @@
 // Set to 0 for production builds to eliminate all debug output overhead.
 // ============================================================================
 #ifndef ANSCORE_DEBUGVIEW
-#define ANSCORE_DEBUGVIEW 1   // 1 = enabled (debug), 0 = disabled (production)
+#define ANSCORE_DEBUGVIEW 0   // 1 = enabled (debug), 0 = disabled (production)
 #endif
 
 // ANS_DBG: Debug logging macro for DebugView (OutputDebugStringA on Windows).

engines/ONNXEngine/ONNXEngine.cpp

@@ -122,36 +122,65 @@ namespace ANSCENTER {
         // Use AppendExecutionProvider_OpenVINO_V2 instead of the generic string API,
         // matching the pattern used in YOLOOD/YOLO12OD/ANSONNXCL etc.
         // Try device configs in priority order, falling back gracefully.
+        //
+        // NPU availability is probed once per process. If AUTO:NPU,GPU fails on
+        // the first call, we skip it for all subsequent models to avoid repeated
+        // "Failed to load shared library" errors cluttering the log.
+        static bool s_npuProbed = false;
+        static bool s_npuAvailable = false;
+
         const std::string precision = "FP16";
         const std::string numberOfThreads = "4";
         const std::string numberOfStreams = "4";
 
-        std::vector<std::unordered_map<std::string, std::string>> try_configs = {
+        auto makeConfig = [&](const std::string& device) {
-            { {"device_type","AUTO:NPU,GPU"}, {"precision",precision},
+            return std::unordered_map<std::string, std::string>{
-              {"num_of_threads",numberOfThreads}, {"num_streams",numberOfStreams},
+                {"device_type", device}, {"precision", precision},
-              {"enable_opencl_throttling","False"}, {"enable_qdq_optimizer","True"} },
+                {"num_of_threads", numberOfThreads}, {"num_streams", numberOfStreams},
-            { {"device_type","GPU.0"},        {"precision",precision},
+                {"enable_opencl_throttling", "False"}, {"enable_qdq_optimizer", "True"}
-              {"num_of_threads",numberOfThreads}, {"num_streams",numberOfStreams},
+            };
-              {"enable_opencl_throttling","False"}, {"enable_qdq_optimizer","True"} },
-            { {"device_type","GPU.1"},        {"precision",precision},
-              {"num_of_threads",numberOfThreads}, {"num_streams",numberOfStreams},
-              {"enable_opencl_throttling","False"}, {"enable_qdq_optimizer","True"} },
-            { {"device_type","AUTO:GPU,CPU"}, {"precision",precision},
-              {"num_of_threads",numberOfThreads}, {"num_streams",numberOfStreams},
-              {"enable_opencl_throttling","False"}, {"enable_qdq_optimizer","True"} }
         };
 
+        std::vector<std::unordered_map<std::string, std::string>> try_configs;
+
+        // Only try NPU if it hasn't been probed yet or was previously available
+        if (!s_npuProbed || s_npuAvailable) {
+            try_configs.push_back(makeConfig("AUTO:NPU,GPU"));
+        }
+        try_configs.push_back(makeConfig("GPU.0"));
+        try_configs.push_back(makeConfig("GPU.1"));
+        try_configs.push_back(makeConfig("AUTO:GPU,CPU"));
+
         for (const auto& config : try_configs) {
             try {
                 session_options.AppendExecutionProvider_OpenVINO_V2(config);
+                const auto& device = config.at("device_type");
                 std::cout << "[ORT] OpenVINO EP attached ("
-                    << config.at("device_type") << ", " << precision << ")." << std::endl;
+                    << device << ", " << precision << ")." << std::endl;
+                ANS_DBG("OrtHandler", "OpenVINO EP attached: %s", device.c_str());
+
+                // If NPU config succeeded, mark it available
+                if (device.find("NPU") != std::string::npos) {
+                    s_npuProbed = true;
+                    s_npuAvailable = true;
+                }
                 return true;
             }
             catch (const Ort::Exception& e) {
-                std::cerr << "[ORT] OpenVINO EP failed for device "
+                const auto& device = config.at("device_type");
-                    << config.at("device_type") << ": " << e.what() << std::endl;
+
-                // try next config
+                // If NPU config failed, remember so we skip it next time
+                if (device.find("NPU") != std::string::npos) {
+                    if (!s_npuProbed) {
+                        std::cout << "[ORT] NPU not available — skipping NPU configs for subsequent models." << std::endl;
+                        ANS_DBG("OrtHandler", "NPU not available, will skip in future");
+                    }
+                    s_npuProbed = true;
+                    s_npuAvailable = false;
+                } else {
+                    std::cerr << "[ORT] OpenVINO EP failed for device "
+                        << device << ": " << e.what() << std::endl;
+                }
             }
         }
         std::cerr << "[ORT] OpenVINO EP: all device configs failed." << std::endl;
@@ -164,7 +193,10 @@ namespace ANSCENTER {
 
     void BasicOrtHandler::initialize_handler()
     {
+        ANS_DBG("OrtHandler", "initialize_handler: m_engineType=%d", static_cast<int>(m_engineType));
         const auto& epInfo = EPLoader::Current();
+        ANS_DBG("OrtHandler", "initialize_handler: EPLoader type=%d dir=%s",
+            static_cast<int>(epInfo.type), epInfo.libraryDir.c_str());
         if (Ort::Global<void>::api_ == nullptr)
             Ort::InitApi(static_cast<const OrtApi*>(EPLoader::GetOrtApiRaw()));
 
@@ -172,6 +204,12 @@ namespace ANSCENTER {
 
         EngineType engine = (static_cast<int>(m_engineType) == -1)
             ? epInfo.type : m_engineType;
+        // Persist the resolved engine type so subclasses (e.g. ONNXYOLO)
+        // can branch on the actual EP at inference time (IoBinding for DML).
+        m_engineType = engine;
+        ANS_DBG("OrtHandler", "initialize_handler: resolved engine=%d (from %s)",
+            static_cast<int>(engine),
+            (static_cast<int>(m_engineType) == -1) ? "EPLoader" : "explicit");
         
         ort_env = new Ort::Env(ORT_LOGGING_LEVEL_ERROR, log_id);
         memory_info_handler = new Ort::MemoryInfo(
@@ -186,7 +224,17 @@ namespace ANSCENTER {
             GraphOptimizationLevel::ORT_ENABLE_ALL);
         session_options.SetLogSeverityLevel(4);
 
-  
+        // DirectML REQUIRES these two settings per ORT documentation:
+        //  - DisableMemPattern: DML manages its own memory; ORT's memory
+        //    pattern optimization conflicts with DML's D3D12 allocator.
+        //  - ORT_SEQUENTIAL: DML uses a single command queue and cannot
+        //    handle parallel execution mode — doing so causes deadlocks
+        //    when synchronizing GPU→CPU data transfers.
+        if (engine == EngineType::AMD_GPU) {
+            session_options.DisableMemPattern();
+            session_options.SetExecutionMode(ExecutionMode::ORT_SEQUENTIAL);
+            ANS_DBG("OrtHandler", "DirectML: DisableMemPattern + ORT_SEQUENTIAL set");
+        }
 
         std::vector<std::string> available = Ort::GetAvailableProviders();
         std::cout << "[ORT] Available providers: ";
@@ -206,41 +254,55 @@ namespace ANSCENTER {
         {
             // --------------------------------------------------------
         case EngineType::NVIDIA_GPU:
+            ANS_DBG("OrtHandler", "Trying CUDA EP...");
             if (hasProvider("CUDAExecutionProvider"))
                 epAttached = TryAppendCUDA(session_options);
-            if (!epAttached)
+            if (!epAttached) {
                 std::cerr << "[ORT] CUDA EP unavailable — falling back to CPU."
                 << std::endl;
+                ANS_DBG("OrtHandler", "CUDA EP FAILED — fallback to CPU");
+            }
             break;
 
             // --------------------------------------------------------
         case EngineType::AMD_GPU:
+            ANS_DBG("OrtHandler", "Trying DirectML EP...");
             if (hasProvider("DmlExecutionProvider"))
                 epAttached = TryAppendDirectML(session_options);
-            if (!epAttached)
+            if (!epAttached) {
                 std::cerr << "[ORT] DirectML EP unavailable — falling back to CPU."
                 << std::endl;
+                ANS_DBG("OrtHandler", "DirectML EP FAILED — fallback to CPU");
+            }
             break;
 
             // --------------------------------------------------------
         case EngineType::OPENVINO_GPU:
+            ANS_DBG("OrtHandler", "Trying OpenVINO EP...");
             if (hasProvider("OpenVINOExecutionProvider"))
                 epAttached = TryAppendOpenVINO(session_options);
-            if (!epAttached)
+            if (!epAttached) {
                 std::cerr << "[ORT] OpenVINO EP unavailable — falling back to CPU."
                 << std::endl;
+                ANS_DBG("OrtHandler", "OpenVINO EP FAILED — fallback to CPU");
+            }
             break;
 
             // --------------------------------------------------------
         case EngineType::CPU:
         default:
             std::cout << "[ORT] Using CPU EP." << std::endl;
+            ANS_DBG("OrtHandler", "Using CPU EP");
             epAttached = true;
             break;
         }
 
-        if (!epAttached)
+        if (!epAttached) {
             std::cout << "[ORT] Running on CPU EP (fallback)." << std::endl;
+            ANS_DBG("OrtHandler", "EP not attached — running on CPU fallback");
+        } else {
+            ANS_DBG("OrtHandler", "EP attached successfully");
+        }
 
         // ----------------------------------------------------------------
         // Create session
@@ -367,15 +429,19 @@ namespace ANSCENTER {
             std::cout << "[ORT] Session created OK (" << label << ")." << std::endl;
         };
 
+        ANS_DBG("OrtHandler", "Creating session for model: %ls", onnx_path);
         try {
             createSession(session_options, "primary EP");
+            ANS_DBG("OrtHandler", "Session created OK with primary EP");
         }
         catch (const Ort::Exception& e) {
+            ANS_DBG("OrtHandler", "Session FAILED with primary EP: %s", e.what());
             std::cerr << "[ORT] Session creation FAILED with primary EP: "
                       << e.what() << std::endl;
 
             // If we were using a GPU EP, fall back to CPU
             if (engine != EngineType::CPU && epAttached) {
+                ANS_DBG("OrtHandler", "Retrying with CPU fallback...");
                 std::cerr << "[ORT] Retrying with CPU EP (fallback)..." << std::endl;
 
                 // Build fresh session options — no GPU EP, no graph opt
@@ -404,6 +470,7 @@ namespace ANSCENTER {
             }
         }
         catch (const std::exception& e) {
+            ANS_DBG("OrtHandler", "Session FAILED (std::exception): %s", e.what());
             std::cerr << "[ORT] Session creation FAILED (std::exception): "
                       << e.what() << std::endl;
             throw;

modules/ANSFR/dllmain.cpp

@@ -514,6 +514,177 @@ extern "C" ANSFR_API int          InsertUser(ANSCENTER::ANSFacialRecognition** H
 		return -1;
 	}
 }
+
+// Helper: repair mixed-encoding LabVIEW LStrHandle to clean UTF-16LE.
+// LabVIEW text controls may produce a mix of UTF-16LE pairs, embedded UTF-8
+// multi-byte sequences, and lone space bytes (0x20 without 0x00 high byte).
+// This normalizes everything to proper UTF-16LE pairs.
+// Input: BOM-stripped raw bytes.  Output: clean UTF-16LE vector.
+static std::vector<unsigned char> RepairLabVIEWUTF16LE_Local(const unsigned char* data, int len) {
+    std::vector<unsigned char> repaired;
+    if (!data || len <= 0) return repaired;
+    repaired.reserve(len + 32);
+
+    auto emitU16 = [&](uint16_t cp) {
+        repaired.push_back(static_cast<unsigned char>(cp & 0xFF));
+        repaired.push_back(static_cast<unsigned char>((cp >> 8) & 0xFF));
+    };
+
+    for (int i = 0; i < len; ) {
+        unsigned char b = data[i];
+
+        // 1. Detect embedded UTF-8 multi-byte sequences
+        // 2-byte UTF-8: C2-DF followed by 80-BF
+        if (b >= 0xC2 && b <= 0xDF && i + 1 < len) {
+            unsigned char b1 = data[i + 1];
+            if ((b1 & 0xC0) == 0x80) {
+                uint32_t cp = ((b & 0x1F) << 6) | (b1 & 0x3F);
+                emitU16(static_cast<uint16_t>(cp));
+                i += 2; continue;
+            }
+        }
+        // 3-byte UTF-8: E0-EF followed by 80-BF 80-BF
+        if (b >= 0xE0 && b <= 0xEF && i + 2 < len) {
+            unsigned char b1 = data[i + 1], b2 = data[i + 2];
+            if ((b1 & 0xC0) == 0x80 && (b2 & 0xC0) == 0x80) {
+                uint32_t cp = ((b & 0x0F) << 12) | ((b1 & 0x3F) << 6) | (b2 & 0x3F);
+                if (cp >= 0x0800 && (cp < 0xD800 || cp > 0xDFFF)) {
+                    emitU16(static_cast<uint16_t>(cp));
+                    i += 3; continue;
+                }
+            }
+        }
+        // 4-byte UTF-8: F0-F4 followed by 80-BF 80-BF 80-BF
+        if (b >= 0xF0 && b <= 0xF4 && i + 3 < len) {
+            unsigned char b1 = data[i + 1], b2 = data[i + 2], b3 = data[i + 3];
+            if ((b1 & 0xC0) == 0x80 && (b2 & 0xC0) == 0x80 && (b3 & 0xC0) == 0x80) {
+                uint32_t cp = ((b & 0x07) << 18) | ((b1 & 0x3F) << 12)
+                            | ((b2 & 0x3F) << 6) | (b3 & 0x3F);
+                if (cp >= 0x10000 && cp <= 0x10FFFF) {
+                    cp -= 0x10000;
+                    emitU16(static_cast<uint16_t>(0xD800 + (cp >> 10)));
+                    emitU16(static_cast<uint16_t>(0xDC00 + (cp & 0x3FF)));
+                    i += 4; continue;
+                }
+            }
+        }
+
+        // 2. Normal UTF-16LE pair (low byte + 0x00 high byte)
+        if (i + 1 < len && data[i + 1] == 0x00) {
+            repaired.push_back(data[i]); repaired.push_back(0x00); i += 2;
+        }
+        // 3. Lone space byte — LabVIEW dropped the 0x00 high byte
+        else if (b == 0x20 && (i + 1 >= len || data[i + 1] != 0x00)) {
+            repaired.push_back(0x20); repaired.push_back(0x00); i += 1;
+        }
+        // 4. Non-ASCII UTF-16LE pair
+        else if (i + 1 < len) {
+            repaired.push_back(data[i]); repaired.push_back(data[i + 1]); i += 2;
+        }
+        // 5. Trailing odd byte — skip
+        else { i++; }
+    }
+    return repaired;
+}
+
+// Helper: convert LStrHandle (mixed UTF-8/UTF-16LE or system codepage) to UTF-8 string
+static std::string LStrHandleToUTF8(LStrHandle handle) {
+    if (!handle) return "";
+    int byteLen = (*handle)->cnt;
+    if (byteLen <= 0) return "";
+    const unsigned char* data = reinterpret_cast<const unsigned char*>((*handle)->str);
+
+    // Check for BOM or 0x00 bytes → UTF-16LE (possibly mixed with UTF-8)
+    bool isUtf16le = false;
+    if (byteLen >= 2 && data[0] == 0xFF && data[1] == 0xFE) isUtf16le = true;
+    if (!isUtf16le) {
+        for (int i = 0; i < byteLen; i++) {
+            if (data[i] == 0x00) { isUtf16le = true; break; }
+        }
+    }
+
+    if (isUtf16le) {
+        const unsigned char* convData = data;
+        int convLen = byteLen;
+        if (convLen >= 2 && convData[0] == 0xFF && convData[1] == 0xFE) { convData += 2; convLen -= 2; }
+        if (convLen <= 0) return "";
+
+        // Repair mixed encoding (UTF-8 islands, lone spaces) → clean UTF-16LE
+        auto repaired = RepairLabVIEWUTF16LE_Local(convData, convLen);
+
+#ifdef _WIN32
+        int wideLen = static_cast<int>(repaired.size()) / 2;
+        const wchar_t* wideStr = reinterpret_cast<const wchar_t*>(repaired.data());
+        int utf8Len = WideCharToMultiByte(CP_UTF8, 0, wideStr, wideLen, nullptr, 0, nullptr, nullptr);
+        if (utf8Len > 0) {
+            std::string utf8(utf8Len, 0);
+            WideCharToMultiByte(CP_UTF8, 0, wideStr, wideLen, &utf8[0], utf8Len, nullptr, nullptr);
+            return utf8;
+        }
+#endif
+        return std::string(reinterpret_cast<const char*>(repaired.data()), repaired.size());
+    } else {
+        // No 0x00 bytes — try UTF-8 first, fall back to system codepage.
+        // IsValidUTF8: check if bytes form valid UTF-8 with at least one multi-byte sequence.
+        auto IsValidUTF8 = [](const unsigned char* d, int l) -> bool {
+            bool hasMulti = false;
+            for (int j = 0; j < l; ) {
+                unsigned char c = d[j];
+                if (c <= 0x7F) { j++; }
+                else if (c >= 0xC2 && c <= 0xDF) {
+                    if (j + 1 >= l || (d[j + 1] & 0xC0) != 0x80) return false;
+                    hasMulti = true; j += 2;
+                } else if (c >= 0xE0 && c <= 0xEF) {
+                    if (j + 2 >= l || (d[j + 1] & 0xC0) != 0x80 || (d[j + 2] & 0xC0) != 0x80) return false;
+                    hasMulti = true; j += 3;
+                } else if (c >= 0xF0 && c <= 0xF4) {
+                    if (j + 3 >= l || (d[j + 1] & 0xC0) != 0x80 || (d[j + 2] & 0xC0) != 0x80 || (d[j + 3] & 0xC0) != 0x80) return false;
+                    hasMulti = true; j += 4;
+                } else { return false; }
+            }
+            return hasMulti;
+        };
+
+        if (IsValidUTF8(data, byteLen)) {
+            return std::string(reinterpret_cast<const char*>(data), byteLen);
+        }
+#ifdef _WIN32
+        int wideLen = MultiByteToWideChar(CP_ACP, 0, reinterpret_cast<const char*>(data), byteLen, nullptr, 0);
+        if (wideLen > 0) {
+            std::wstring wideStr(wideLen, 0);
+            MultiByteToWideChar(CP_ACP, 0, reinterpret_cast<const char*>(data), byteLen, &wideStr[0], wideLen);
+            int utf8Len = WideCharToMultiByte(CP_UTF8, 0, wideStr.c_str(), wideLen, nullptr, 0, nullptr, nullptr);
+            if (utf8Len > 0) {
+                std::string utf8(utf8Len, 0);
+                WideCharToMultiByte(CP_UTF8, 0, wideStr.c_str(), wideLen, &utf8[0], utf8Len, nullptr, nullptr);
+                return utf8;
+            }
+        }
+#endif
+        return std::string(reinterpret_cast<const char*>(data), byteLen);
+    }
+}
+
+extern "C" ANSFR_API int InsertUser_LV(ANSCENTER::ANSFacialRecognition** Handle, const char* userCode, LStrHandle userName) {
+    try {
+        if (!Handle || !*Handle || !userCode || !userName) return -1;
+        std::string utf8Name = LStrHandleToUTF8(userName);
+        if (utf8Name.empty()) return -1;
+        return (*Handle)->InsertUser(userCode, utf8Name);
+    }
+    catch (const std::exception& e) { return -1; }
+}
+
+extern "C" ANSFR_API int UpdateUser_LV(ANSCENTER::ANSFacialRecognition** Handle, int userId, const char* userCode, LStrHandle userName) {
+    try {
+        if (!Handle || !*Handle || !userCode || !userName) return -1;
+        std::string utf8Name = LStrHandleToUTF8(userName);
+        if (utf8Name.empty()) return -1;
+        return (*Handle)->UpdateUser(userId, userCode, utf8Name);
+    }
+    catch (const std::exception& e) { return -1; }
+}
+
 extern "C" ANSFR_API int          UpdateUser(ANSCENTER::ANSFacialRecognition** Handle, int userId, const char* userCode, const char* userName) {
     try {
         if (!Handle || !*Handle || !userCode || !userName) return -1;

modules/ANSLPR/ANSLPR_OD.cpp

@@ -963,7 +963,9 @@ namespace ANSCENTER {
 
 		// Run license plate detection
 		cv::Mat activeFrame = frame(detectedArea);
+		fprintf(stderr, "[ALPR] RunInference: calling lpd %dx%d cam=%s\n", activeFrame.cols, activeFrame.rows, cameraId.c_str());
 		std::vector<Object> lprOutput = _lpDetector->RunInference(activeFrame, cameraId);
+		fprintf(stderr, "[ALPR] RunInference: lpd done, %zu detections cam=%s\n", lprOutput.size(), cameraId.c_str());
 		for (size_t _di = 0; _di < lprOutput.size(); ++_di) {
 			ANS_DBG("ALPR_Track", "cam=%s det[%zu] tid=%d box=(%d,%d,%d,%d) conf=%.2f",
 				cameraId.c_str(), _di, lprOutput[_di].trackId,
@@ -1005,7 +1007,9 @@ namespace ANSCENTER {
 			cv::Mat alignedLPR = frame(lprPos);// .clone();
 
 			// OCR inference
+			fprintf(stderr, "[ALPR] RunInference: calling OCR on plate %dx%d cam=%s\n", alignedLPR.cols, alignedLPR.rows, cameraId.c_str());
 			std::string ocrText = DetectLicensePlateString(alignedLPR, cameraId);
+			fprintf(stderr, "[ALPR] RunInference: OCR done, text='%s' cam=%s\n", ocrText.c_str(), cameraId.c_str());
 
 			if (ocrText.empty()) {
 				continue;

modules/ANSODEngine/ANSONNXYOLO.cpp

@@ -335,7 +335,7 @@ namespace ANSCENTER {
                 // to distinguish OBB (angle values in [-pi, pi]) from detection
                 bool likelyOBB = false;
                 if (extra >= 2) {
-                    const float* rawOutput = outputTensors[0].GetTensorData<float>();
+                    const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
                     int numSamples = std::min(numBoxes, 100);
                     int angleCount = 0;
                     for (int s = 0; s < numSamples; ++s) {
@@ -371,13 +371,13 @@ namespace ANSCENTER {
     std::vector<Object> ONNXYOLO::postprocessEndToEnd(
         const cv::Size& originalImageSize,
         const cv::Size& resizedImageShape,
-        const std::vector<Ort::Value>& outputTensors,
+        std::vector<Ort::Value>& outputTensors,
         const std::vector<std::string>& classNames,
         float confThreshold)
     {
         if (outputTensors.empty()) return {};
 
-        const float* rawOutput = outputTensors[0].GetTensorData<float>();
+        const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
         const auto outputShape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
         if (outputShape.size() < 3) return {};
 
@@ -427,13 +427,13 @@ namespace ANSCENTER {
     std::vector<Object> ONNXYOLO::postprocessLegacy(
         const cv::Size& originalImageSize,
         const cv::Size& resizedImageShape,
-        const std::vector<Ort::Value>& outputTensors,
+        std::vector<Ort::Value>& outputTensors,
         const std::vector<std::string>& classNames,
         float confThreshold, float iouThreshold, int maxDet)
     {
         if (outputTensors.empty()) return {};
 
-        const float* rawOutput = outputTensors[0].GetTensorData<float>();
+        const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
         const auto outputShape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
         if (outputShape.size() < 3) return {};
 
@@ -656,12 +656,12 @@ namespace ANSCENTER {
     std::vector<Object> ONNXYOLO::postprocessOBBEndToEnd(
         const cv::Size& originalImageSize,
         const cv::Size& resizedImageShape,
-        const std::vector<Ort::Value>& outputTensors,
+        std::vector<Ort::Value>& outputTensors,
         const std::vector<std::string>& classNames,
         float confThreshold)
     {
         if (outputTensors.empty()) return {};
-        const float* raw = outputTensors[0].GetTensorData<float>();
+        const float* raw = outputTensors[0].GetTensorMutableData<float>();
         const auto shape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
         if (shape.size() < 3) return {};
 
@@ -721,12 +721,12 @@ namespace ANSCENTER {
     std::vector<Object> ONNXYOLO::postprocessOBBLegacy(
         const cv::Size& originalImageSize,
         const cv::Size& resizedImageShape,
-        const std::vector<Ort::Value>& outputTensors,
+        std::vector<Ort::Value>& outputTensors,
         const std::vector<std::string>& classNames,
         float confThreshold, float iouThreshold, int maxDet)
     {
         if (outputTensors.empty()) return {};
-        const float* rawOutput = outputTensors[0].GetTensorData<float>();
+        const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
         const auto outputShape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
         if (outputShape.size() < 3) return {};
 
@@ -822,13 +822,13 @@ namespace ANSCENTER {
     std::vector<Object> ONNXYOLO::postprocessSegEndToEnd(
         const cv::Size& originalImageSize,
         const cv::Size& resizedImageShape,
-        const std::vector<Ort::Value>& outputTensors,
+        std::vector<Ort::Value>& outputTensors,
         const std::vector<std::string>& classNames,
         float confThreshold)
     {
         if (outputTensors.size() < 2) return {};
 
-        const float* raw = outputTensors[0].GetTensorData<float>();
+        const float* raw = outputTensors[0].GetTensorMutableData<float>();
         const auto shape0 = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
         const auto protoShape = outputTensors[1].GetTensorTypeAndShapeInfo().GetShape();
         if (shape0.size() < 3 || protoShape.size() < 4) return {};
@@ -884,7 +884,7 @@ namespace ANSCENTER {
 
         // Generate masks: coeffs @ protos → sigmoid → crop-in-proto → resize-to-box → threshold
         if (!objs.empty() && !maskCoeffs.empty()) {
-            const float* protoData = outputTensors[1].GetTensorData<float>();
+            const float* protoData = outputTensors[1].GetTensorMutableData<float>();
             cv::Mat protos(nm, protoH * protoW, CV_32F, const_cast<float*>(protoData));
             cv::Mat matmulRes = (maskCoeffs * protos).t();
 
@@ -951,13 +951,13 @@ namespace ANSCENTER {
     std::vector<Object> ONNXYOLO::postprocessSegLegacy(
         const cv::Size& originalImageSize,
         const cv::Size& resizedImageShape,
-        const std::vector<Ort::Value>& outputTensors,
+        std::vector<Ort::Value>& outputTensors,
         const std::vector<std::string>& classNames,
         float confThreshold, float iouThreshold, int maxDet)
     {
         if (outputTensors.size() < 2) return {};
 
-        const float* rawOutput = outputTensors[0].GetTensorData<float>();
+        const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
         const auto shape0 = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
         const auto protoShape = outputTensors[1].GetTensorTypeAndShapeInfo().GetShape();
         if (shape0.size() < 3 || protoShape.size() < 4) return {};
@@ -1035,7 +1035,7 @@ namespace ANSCENTER {
 
         // Generate masks
         if (!objs.empty() && !masks.empty()) {
-            const float* protoData = outputTensors[1].GetTensorData<float>();
+            const float* protoData = outputTensors[1].GetTensorMutableData<float>();
             cv::Mat protos(nm, protoH * protoW, CV_32F, const_cast<float*>(protoData));
             cv::Mat matmulRes = (masks * protos).t();
 
@@ -1106,12 +1106,12 @@ namespace ANSCENTER {
     std::vector<Object> ONNXYOLO::postprocessPoseEndToEnd(
         const cv::Size& originalImageSize,
         const cv::Size& resizedImageShape,
-        const std::vector<Ort::Value>& outputTensors,
+        std::vector<Ort::Value>& outputTensors,
         const std::vector<std::string>& classNames,
         float confThreshold, int numKPS)
     {
         if (outputTensors.empty()) return {};
-        const float* raw = outputTensors[0].GetTensorData<float>();
+        const float* raw = outputTensors[0].GetTensorMutableData<float>();
         const auto shape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
         if (shape.size() < 3) return {};
 
@@ -1172,12 +1172,12 @@ namespace ANSCENTER {
     std::vector<Object> ONNXYOLO::postprocessPoseLegacy(
         const cv::Size& originalImageSize,
         const cv::Size& resizedImageShape,
-        const std::vector<Ort::Value>& outputTensors,
+        std::vector<Ort::Value>& outputTensors,
         const std::vector<std::string>& classNames,
         float confThreshold, float iouThreshold, int numKPS, int maxDet)
     {
         if (outputTensors.empty()) return {};
-        const float* rawOutput = outputTensors[0].GetTensorData<float>();
+        const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
         const auto outputShape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
         if (outputShape.size() < 3) return {};
 
@@ -1273,12 +1273,12 @@ namespace ANSCENTER {
     // ====================================================================
 
     std::vector<Object> ONNXYOLO::postprocessClassify(
-        const std::vector<Ort::Value>& outputTensors,
+        std::vector<Ort::Value>& outputTensors,
         const std::vector<std::string>& classNames,
         const cv::Size& imageSize)
     {
         if (outputTensors.empty()) return {};
-        const float* raw = outputTensors[0].GetTensorData<float>();
+        const float* raw = outputTensors[0].GetTensorMutableData<float>();
         const auto shape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
         if (shape.size() < 2) return {};
 
@@ -1339,7 +1339,7 @@ namespace ANSCENTER {
     // ====================================================================
 
     /*static*/ Ort::Value ONNXYOLO::sliceBatchOutput(
-        const Ort::Value& batchTensor,
+        Ort::Value& batchTensor,
         int64_t batchIndex,
         const std::vector<int64_t>& fullShape,
         Ort::MemoryInfo& memInfo)
@@ -1349,8 +1349,8 @@ namespace ANSCENTER {
         for (size_t d = 1; d < fullShape.size(); ++d)
             elemsPerImage *= fullShape[d];
 
-        const float* batchData = batchTensor.GetTensorData<float>();
+        float* batchData = batchTensor.GetTensorMutableData<float>();
-        float* imageData = const_cast<float*>(batchData + batchIndex * elemsPerImage);
+        float* imageData = batchData + batchIndex * elemsPerImage;
 
         // Shape for single image: [1, D1, D2, ...]
         std::vector<int64_t> singleShape = fullShape;
@@ -1504,7 +1504,7 @@ namespace ANSCENTER {
                     // Class count mismatch — probe last channel for OBB angles
                     bool likelyOBB = false;
                     if (extra >= 2) {
-                        const float* rawOutput = perImageOutputs[0].GetTensorData<float>();
+                        const float* rawOutput = perImageOutputs[0].GetTensorMutableData<float>();
                         int numSamp = std::min(numBoxes, 100);
                         int angleCount = 0;
                         for (int s = 0; s < numSamp; ++s) {
@@ -1571,6 +1571,22 @@ namespace ANSCENTER {
         }
     }
 
+    bool ANSONNXYOLO::InitOrtEngine(ANSCENTER::EngineType engineType) {
+        try {
+            if (!FileExist(_modelFilePath)) {
+                _logger.LogError("ANSONNXYOLO::InitOrtEngine",
+                    "Model file does not exist: " + _modelFilePath, __FILE__, __LINE__);
+                return false;
+            }
+            m_ortEngine = std::make_unique<ONNXYOLO>(_modelFilePath, engineType);
+            return true;
+        }
+        catch (const std::exception& e) {
+            _logger.LogFatal("ANSONNXYOLO::InitOrtEngine", e.what(), __FILE__, __LINE__);
+            return false;
+        }
+    }
+
     bool ANSONNXYOLO::Initialize(std::string licenseKey, ModelConfig modelConfig,
                                    const std::string& modelZipFilePath,
                                    const std::string& modelZipPassword,
@@ -1807,9 +1823,12 @@ namespace ANSCENTER {
                                                       const std::string& camera_id)
     {
         try {
+            ANS_DBG("ONNXYOLO", "DetectObjects: cam=%s acquiring mutex...", camera_id.c_str());
             std::lock_guard<std::recursive_mutex> lock(_mutex);
+            ANS_DBG("ONNXYOLO", "DetectObjects: mutex acquired, cam=%s", camera_id.c_str());
             if (!m_ortEngine) {
                 _logger.LogError("ANSONNXYOLO::DetectObjects", "ORT engine is null", __FILE__, __LINE__);
+                ANS_DBG("ONNXYOLO", "DetectObjects: ORT engine is null!");
                 return {};
             }
 
@@ -1880,6 +1899,7 @@ namespace ANSCENTER {
             return results;
         }
         catch (const std::exception& e) {
+            ANS_DBG("ONNXYOLO", "DetectObjects EXCEPTION: %s cam=%s", e.what(), camera_id.c_str());
             _logger.LogFatal("ANSONNXYOLO::DetectObjects", e.what(), __FILE__, __LINE__);
             return {};
         }

modules/ANSODEngine/ANSONNXYOLO.h

@@ -83,55 +83,55 @@ namespace ANSCENTER {
         // ── Detection postprocess ───────────────────────────────────────
         std::vector<Object> postprocessEndToEnd(
             const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-            const std::vector<Ort::Value>& outputTensors,
+            std::vector<Ort::Value>& outputTensors,
             const std::vector<std::string>& classNames, float confThreshold);
 
         std::vector<Object> postprocessLegacy(
             const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-            const std::vector<Ort::Value>& outputTensors,
+            std::vector<Ort::Value>& outputTensors,
             const std::vector<std::string>& classNames,
             float confThreshold, float iouThreshold, int maxDet = 300);
 
         // ── OBB postprocess ─────────────────────────────────────────────
         std::vector<Object> postprocessOBBEndToEnd(
             const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-            const std::vector<Ort::Value>& outputTensors,
+            std::vector<Ort::Value>& outputTensors,
             const std::vector<std::string>& classNames, float confThreshold);
 
         std::vector<Object> postprocessOBBLegacy(
             const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-            const std::vector<Ort::Value>& outputTensors,
+            std::vector<Ort::Value>& outputTensors,
             const std::vector<std::string>& classNames,
             float confThreshold, float iouThreshold, int maxDet = 300);
 
         // ── Segmentation postprocess ────────────────────────────────────
         std::vector<Object> postprocessSegEndToEnd(
             const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-            const std::vector<Ort::Value>& outputTensors,
+            std::vector<Ort::Value>& outputTensors,
             const std::vector<std::string>& classNames, float confThreshold);
 
         std::vector<Object> postprocessSegLegacy(
             const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-            const std::vector<Ort::Value>& outputTensors,
+            std::vector<Ort::Value>& outputTensors,
             const std::vector<std::string>& classNames,
             float confThreshold, float iouThreshold, int maxDet = 300);
 
         // ── Pose postprocess ────────────────────────────────────────────
         std::vector<Object> postprocessPoseEndToEnd(
             const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-            const std::vector<Ort::Value>& outputTensors,
+            std::vector<Ort::Value>& outputTensors,
             const std::vector<std::string>& classNames,
             float confThreshold, int numKPS);
 
         std::vector<Object> postprocessPoseLegacy(
             const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-            const std::vector<Ort::Value>& outputTensors,
+            std::vector<Ort::Value>& outputTensors,
             const std::vector<std::string>& classNames,
             float confThreshold, float iouThreshold, int numKPS, int maxDet = 300);
 
         // ── Classification postprocess ──────────────────────────────────
         std::vector<Object> postprocessClassify(
-            const std::vector<Ort::Value>& outputTensors,
+            std::vector<Ort::Value>& outputTensors,
             const std::vector<std::string>& classNames,
             const cv::Size& imageSize);
 
@@ -154,7 +154,7 @@ namespace ANSCENTER {
 
         // ── Batch output slicing helper ────────────────────────────────
         static Ort::Value sliceBatchOutput(
-            const Ort::Value& batchTensor,
+            Ort::Value& batchTensor,
             int64_t batchIndex,
             const std::vector<int64_t>& fullShape,
             Ort::MemoryInfo& memInfo);
@@ -224,6 +224,9 @@ namespace ANSCENTER {
 
         // Initialise ORT engine from the resolved model path
         bool InitOrtEngine();
+    public:
+        // Initialise ORT engine with explicit engine type override (e.g. CPU fallback for AMD iGPUs)
+        bool InitOrtEngine(ANSCENTER::EngineType engineType);
     };
 }
 #endif

modules/ANSODEngine/ANSYOLOOD.cpp

@@ -218,6 +218,12 @@ namespace ANSCENTER
 				std::min(6, static_cast<int>(std::thread::hardware_concurrency())));
 			sessionOptions.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);
 
+			// DirectML REQUIRES these two settings per ORT documentation
+			if (ep.type == ANSCENTER::EngineType::AMD_GPU) {
+				sessionOptions.DisableMemPattern();
+				sessionOptions.SetExecutionMode(ExecutionMode::ORT_SEQUENTIAL);
+			}
+
 			// ── Log available providers ─────────────────────────────────────────
 			std::vector<std::string> availableProviders = Ort::GetAvailableProviders();
 			std::cout << "Available Execution Providers:" << std::endl;
@@ -519,7 +525,7 @@ namespace ANSCENTER
 	{
 		try {
 			// Get raw output pointer (NO COPY!)
-			const float* rawOutput = outputTensors[0].GetTensorData<float>();
+			const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
 			std::vector<int64_t> outputShape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
 
 			const int numClasses = static_cast<int>(outputShape[2]) - 5;
@@ -647,11 +653,11 @@ namespace ANSCENTER
 		}
 		return result;
 	}
-	std::vector<Object> YOLOOD::postprocessv11(const cv::Size& originalImageSize,const cv::Size& resizedImageShape,const std::vector<Ort::Value>& outputTensors,float confThreshold,float iouThreshold)
+	std::vector<Object> YOLOOD::postprocessv11(const cv::Size& originalImageSize,const cv::Size& resizedImageShape,std::vector<Ort::Value>& outputTensors,float confThreshold,float iouThreshold)
 	{
 		try {
 			// Get raw output
-			const float* rawOutput = outputTensors[0].GetTensorData<float>();
+			const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
 			const std::vector<int64_t> outputShape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
 
 			const size_t numFeatures = outputShape[1];
@@ -1448,7 +1454,7 @@ namespace ANSCENTER
 			);
 
 			// Parse output
-			const float* rawOutput = outputTensors[0].GetTensorData<float>();
+			const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
 			const std::vector<int64_t> outputShape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
 
 			const int dimensions = static_cast<int>(outputShape[1]);  // 4 + num_classes

modules/ANSODEngine/ANSYOLOOD.h

@@ -44,7 +44,7 @@ namespace ANSCENTER {
         cv::Mat preprocessv11(const cv::Mat& image, std::vector<float>& blob, std::vector<int64_t>& inputTensorShape);
         std::vector<Object> postprocessing(const cv::Size& resizedImageShape,const cv::Size& originalImageShape,std::vector<Ort::Value>& outputTensors,
                                            const float& confThreshold, const float& iouThreshold);
-        std::vector<Object> postprocessv11(const cv::Size& originalImageSize,const cv::Size& resizedImageShape,const std::vector<Ort::Value>& outputTensors,float confThreshold,float iouThreshold);
+        std::vector<Object> postprocessv11(const cv::Size& originalImageSize,const cv::Size& resizedImageShape,std::vector<Ort::Value>& outputTensors,float confThreshold,float iouThreshold);
         BoundingBox scaleCoordsv11(const cv::Size& imageShape, BoundingBox coords,const cv::Size& imageOriginalShape, bool p_Clip);
         std::vector<const char*> inputNodeNames;
         std::vector<const char*> outputNodeNames;

modules/ANSODEngine/dllmain.cpp

@@ -355,6 +355,7 @@ extern "C" ANSODENGINE_API std::string  CreateANSODHandle(ANSCENTER::ANSODBase**
 	//	TEXTSCENSE = 6
 
 	//Force modelType to ANSONNXYOLO and ANSRTYOLO if detectionType is detection and modelType is TENSORRT or ONNX
+
 	if ((modelType == 4) || // TensorRT  
 		(modelType == 14)|| // TensorRT Yolov10	 
 		(modelType == 22)|| // TensorRT Pose
@@ -376,7 +377,6 @@ extern "C" ANSODENGINE_API std::string  CreateANSODHandle(ANSCENTER::ANSODBase**
 	}
 
 
-
 	switch (detectionType) {
 		case 0:
 			modelConfig.detectionType = ANSCENTER::DetectionType::CLASSIFICATION;

modules/ANSUtilities/dllmain.cpp

@@ -804,34 +804,54 @@ extern "C" ANSULT_API int ANSConvertUTF8ToUTF16LE(const char* utf8Str, LStrHandl
 		int len = (int)strlen(utf8Str);
 		if (len == 0) return 0;
 		const char bom[2] = { '\xFF', '\xFE' };
+
+		// Check if input contains \uXXXX escape sequences
 		bool hasUnicodeEscapes = false;
 		for (int i = 0; i + 1 < len; i++) {
 			if (utf8Str[i] == '\\' && utf8Str[i + 1] == 'u') { hasUnicodeEscapes = true; break; }
 		}
+
 		if (hasUnicodeEscapes) {
-			std::string utf16le;
+			// Two-pass approach: first decode \uXXXX escapes to UTF-8, then convert to UTF-16LE.
-			if (includeBOM) utf16le.assign(bom, 2);
+			// This correctly handles mixed input (raw UTF-8 + \uXXXX escapes) by producing
-			utf16le.reserve(len * 2 + 2);
+			// clean UTF-8 first, then using MultiByteToWideChar for proper UTF-16LE conversion.
+			std::string utf8Decoded;
+			utf8Decoded.reserve(len);
 			for (int i = 0; i < len; ) {
 				if (i + 5 < len && utf8Str[i] == '\\' && utf8Str[i + 1] == 'u') {
 					char hex[5] = { utf8Str[i + 2], utf8Str[i + 3], utf8Str[i + 4], utf8Str[i + 5], 0 };
-					uint16_t cp = (uint16_t)strtoul(hex, nullptr, 16);
+					uint32_t cp = (uint32_t)strtoul(hex, nullptr, 16);
-					utf16le += static_cast<char>(cp & 0xFF);
+					// Encode codepoint as UTF-8
-					utf16le += static_cast<char>((cp >> 8) & 0xFF);
+					if (cp <= 0x7F) {
+						utf8Decoded += static_cast<char>(cp);
+					} else if (cp <= 0x7FF) {
+						utf8Decoded += static_cast<char>(0xC0 | (cp >> 6));
+						utf8Decoded += static_cast<char>(0x80 | (cp & 0x3F));
+					} else {
+						utf8Decoded += static_cast<char>(0xE0 | (cp >> 12));
+						utf8Decoded += static_cast<char>(0x80 | ((cp >> 6) & 0x3F));
+						utf8Decoded += static_cast<char>(0x80 | (cp & 0x3F));
+					}
 					i += 6;
 				} else {
-					utf16le += utf8Str[i];
+					utf8Decoded += utf8Str[i];
-					utf16le += '\0';
 					i++;
 				}
 			}
-			int size = (int)utf16le.size();
+			// Now convert the clean UTF-8 to UTF-16LE
-			MgErr error = DSSetHandleSize(result, sizeof(int32) + size * sizeof(uChar));
+			std::string converted = ANSCENTER::ANSUtilities::ConvertUTF8ToUTF16LE(utf8Decoded);
+			if (converted.empty()) return 0;
+			int dataSize = static_cast<int>(converted.size());
+			int bomSize = includeBOM ? 2 : 0;
+			int totalSize = bomSize + dataSize;
+			MgErr error = DSSetHandleSize(result, sizeof(int32) + totalSize * sizeof(uChar));
 			if (error != noErr) return -2;
-			(*result)->cnt = size;
+			(*result)->cnt = totalSize;
-			memcpy((*result)->str, utf16le.data(), size);
+			if (includeBOM) memcpy((*result)->str, bom, 2);
+			memcpy((*result)->str + bomSize, converted.data(), dataSize);
 			return 1;
 		}
+
 		std::string converted = ANSCENTER::ANSUtilities::ConvertUTF8ToUTF16LE(utf8Str);
 		if (converted.empty()) return 0;
 		int dataSize = static_cast<int>(converted.size());
@@ -850,23 +870,31 @@ extern "C" ANSULT_API int ANSConvertUTF8ToUTF16LE(const char* utf8Str, LStrHandl
 extern "C" ANSULT_API int ANSConvertUTF16LEToUTF8(const unsigned char* utf16leBytes, int byteLen, LStrHandle result) {
 	try {
 		if (!utf16leBytes || byteLen <= 0 || !result) return -1;
-		bool isUtf16le = (byteLen >= 2 && byteLen % 2 == 0);
+		const unsigned char* data = utf16leBytes;
+		int dataLen = byteLen;
+		// Strip BOM (FF FE) if present
+		if (dataLen >= 2 && data[0] == 0xFF && data[1] == 0xFE) {
+			data += 2;
+			dataLen -= 2;
+		}
+		if (dataLen <= 0) return 0;
+		bool isUtf16le = (dataLen >= 2 && dataLen % 2 == 0);
 		if (isUtf16le) {
 			bool isAscii = true;
-			for (int i = 1; i < byteLen; i += 2) {
+			for (int i = 1; i < dataLen; i += 2) {
-				if (utf16leBytes[i] != 0x00) { isAscii = false; break; }
+				if (data[i] != 0x00) { isAscii = false; break; }
 			}
 			if (isAscii) {
-				int asciiLen = byteLen / 2;
+				int asciiLen = dataLen / 2;
 				MgErr error = DSSetHandleSize(result, sizeof(int32) + asciiLen * sizeof(uChar));
 				if (error != noErr) return -2;
 				(*result)->cnt = asciiLen;
-				for (int i = 0; i < asciiLen; i++) (*result)->str[i] = utf16leBytes[i * 2];
+				for (int i = 0; i < asciiLen; i++) (*result)->str[i] = data[i * 2];
 				return 1;
 			}
 		}
 		std::string converted = ANSCENTER::ANSUtilities::ConvertUTF16LEToUTF8(
-			reinterpret_cast<const char*>(utf16leBytes), byteLen);
+			reinterpret_cast<const char*>(data), dataLen);
 		if (converted.empty()) return 0;
 		int size = static_cast<int>(converted.size());
 		MgErr error = DSSetHandleSize(result, sizeof(int32) + size * sizeof(uChar));
@@ -909,6 +937,168 @@ extern "C" ANSULT_API int ANSConvertUTF16LEToUnicodeEscapes(const unsigned char*
 	catch (...) { return -1; }
 }
 
+// Helper: copy a std::string into a LabVIEW LStrHandle.
+static int CopyStringToLStrHandle(LStrHandle handle, const std::string& str) {
+	if (str.empty()) return 0;
+	int size = static_cast<int>(str.size());
+	MgErr error = DSSetHandleSize(handle, sizeof(int32) + size * sizeof(uChar));
+	if (error != noErr) return -2;
+	(*handle)->cnt = size;
+	memcpy((*handle)->str, str.data(), size);
+	return 1;
+}
+
+// Helper: copy raw bytes into a LabVIEW LStrHandle.
+static int CopyBytesToLStrHandle(LStrHandle handle, const unsigned char* data, int len) {
+	if (!data || len <= 0) return 0;
+	MgErr error = DSSetHandleSize(handle, sizeof(int32) + len * sizeof(uChar));
+	if (error != noErr) return -2;
+	(*handle)->cnt = len;
+	memcpy((*handle)->str, data, len);
+	return 1;
+}
+
+// Helper: detect if LabVIEW LStrHandle contains UTF-16LE (BOM or 0x00 bytes).
+static bool DetectUTF16LE(const unsigned char* data, int byteLen) {
+	if (byteLen >= 2 && data[0] == 0xFF && data[1] == 0xFE) return true;
+	for (int i = 0; i < byteLen; i++) {
+		if (data[i] == 0x00) return true;
+	}
+	return false;
+}
+
+// Helper: strip BOM from UTF-16LE data. Returns pointer and adjusts length.
+static const unsigned char* StripBOM(const unsigned char* data, int& len) {
+	if (len >= 2 && data[0] == 0xFF && data[1] == 0xFE) { data += 2; len -= 2; }
+	return data;
+}
+
+// LStrHandle-safe version: reads raw bytes from LabVIEW LStrHandle directly.
+// Two paths:
+//   1. Pure UTF-8 (no BOM, no 0x00 bytes, valid UTF-8) → pass through to output as-is
+//   2. Contains UTF-16LE (BOM or 0x00 bytes) → RepairLabVIEWUTF16LE (normalizes
+//      mixed UTF-8/UTF-16LE + lone spaces to clean UTF-16LE) → convert to UTF-8
+extern "C" ANSULT_API int ANSConvertUTF16LEToUTF8_LV(LStrHandle input, LStrHandle result) {
+	try {
+		if (!input || !result) return -1;
+		int byteLen = (*input)->cnt;
+		if (byteLen <= 0) return 0;
+
+		// Copy input data first — input and result may be the same LStrHandle
+		std::vector<unsigned char> inputCopy(byteLen);
+		memcpy(inputCopy.data(), (*input)->str, byteLen);
+		const unsigned char* data = inputCopy.data();
+
+		if (DetectUTF16LE(data, byteLen)) {
+			// Path 2: UTF-16LE detected — repair mixed encoding, then convert to UTF-8
+			int convLen = byteLen;
+			const unsigned char* convData = StripBOM(data, convLen);
+			if (convLen <= 0) return 0;
+
+			auto repaired = ANSCENTER::ANSUtilities::RepairLabVIEWUTF16LE(convData, convLen);
+			std::string converted = ANSCENTER::ANSUtilities::ConvertUTF16LEToUTF8(
+				reinterpret_cast<const char*>(repaired.data()), static_cast<int>(repaired.size()));
+			return CopyStringToLStrHandle(result, converted);
+		}
+
+		if (ANSCENTER::ANSUtilities::IsValidUTF8(data, byteLen)) {
+			// Path 1: Pure UTF-8 — pass through as-is
+			return CopyBytesToLStrHandle(result, data, byteLen);
+		}
+
+		// Fallback: not UTF-16LE, not valid UTF-8 — assume system codepage
+#ifdef _WIN32
+		int wideLen = MultiByteToWideChar(CP_ACP, 0,
+			reinterpret_cast<const char*>(data), byteLen, nullptr, 0);
+		if (wideLen > 0) {
+			std::wstring wideStr(wideLen, 0);
+			MultiByteToWideChar(CP_ACP, 0,
+				reinterpret_cast<const char*>(data), byteLen, &wideStr[0], wideLen);
+			int utf8Len = WideCharToMultiByte(CP_UTF8, 0,
+				wideStr.c_str(), wideLen, nullptr, 0, nullptr, nullptr);
+			if (utf8Len > 0) {
+				std::string utf8Str(utf8Len, 0);
+				WideCharToMultiByte(CP_UTF8, 0,
+					wideStr.c_str(), wideLen, &utf8Str[0], utf8Len, nullptr, nullptr);
+				return CopyStringToLStrHandle(result, utf8Str);
+			}
+		}
+#endif
+		return CopyBytesToLStrHandle(result, data, byteLen);
+	}
+	catch (...) { return -1; }
+}
+
+// LStrHandle-safe version with auto-detection.
+// Two paths:
+//   1. Pure UTF-8 → convert UTF-8 to Unicode escapes (\uXXXX)
+//   2. Contains UTF-16LE → RepairLabVIEWUTF16LE → convert to Unicode escapes
+extern "C" ANSULT_API int ANSConvertUTF16LEToUnicodeEscapes_LV(LStrHandle input, LStrHandle result) {
+	try {
+		if (!input || !result) return -1;
+		int byteLen = (*input)->cnt;
+		if (byteLen <= 0) return 0;
+
+		// Copy input data first — input and result may be the same LStrHandle
+		std::vector<unsigned char> inputCopy(byteLen);
+		memcpy(inputCopy.data(), (*input)->str, byteLen);
+		const unsigned char* data = inputCopy.data();
+
+		std::string escaped;
+
+		if (DetectUTF16LE(data, byteLen)) {
+			// Path 2: UTF-16LE detected — repair mixed encoding, then convert to escapes
+			int convLen = byteLen;
+			const unsigned char* convData = StripBOM(data, convLen);
+			if (convLen <= 0) return 0;
+
+			auto repaired = ANSCENTER::ANSUtilities::RepairLabVIEWUTF16LE(convData, convLen);
+
+			// Re-add BOM for ConvertUTF16LEToUnicodeEscapes (it expects optional BOM)
+			std::vector<unsigned char> withBom;
+			withBom.reserve(2 + repaired.size());
+			withBom.push_back(0xFF);
+			withBom.push_back(0xFE);
+			withBom.insert(withBom.end(), repaired.begin(), repaired.end());
+
+			escaped = ANSCENTER::ANSUtilities::ConvertUTF16LEToUnicodeEscapes(
+				reinterpret_cast<const char*>(withBom.data()), static_cast<int>(withBom.size()));
+		}
+		else {
+			// Path 1: No UTF-16LE — get UTF-8, then convert to Unicode escapes
+			std::string utf8Str;
+			if (ANSCENTER::ANSUtilities::IsValidUTF8(data, byteLen)) {
+				utf8Str.assign(reinterpret_cast<const char*>(data), byteLen);
+			}
+#ifdef _WIN32
+			else {
+				int wideLen = MultiByteToWideChar(CP_ACP, 0,
+					reinterpret_cast<const char*>(data), byteLen, nullptr, 0);
+				if (wideLen > 0) {
+					std::wstring wideStr(wideLen, 0);
+					MultiByteToWideChar(CP_ACP, 0,
+						reinterpret_cast<const char*>(data), byteLen, &wideStr[0], wideLen);
+					int utf8Len = WideCharToMultiByte(CP_UTF8, 0,
+						wideStr.c_str(), wideLen, nullptr, 0, nullptr, nullptr);
+					if (utf8Len > 0) {
+						utf8Str.resize(utf8Len);
+						WideCharToMultiByte(CP_UTF8, 0,
+							wideStr.c_str(), wideLen, &utf8Str[0], utf8Len, nullptr, nullptr);
+					}
+				}
+			}
+#endif
+			if (utf8Str.empty()) {
+				utf8Str.assign(reinterpret_cast<const char*>(data), byteLen);
+			}
+			escaped = ANSCENTER::ANSUtilities::ConvertUTF8ToUnicodeEscapes(utf8Str);
+		}
+
+		return CopyStringToLStrHandle(result, escaped);
+	}
+	catch (...) { return -1; }
+}
+
 extern "C" ANSULT_API int ANSConvertUnicodeEscapesToUTF8(const char* escapedStr, LStrHandle result) {
 	try {
 		if (!escapedStr || !result) return -1;

tests/ANSLPR-UnitTest/ANSLPR-UnitTest.cpp

@@ -29,6 +29,7 @@
 #include <set>
 #include <map>
 #include <cuda_runtime.h>
+#include "EPLoader.h"
 
 template<typename T>
 T GetOptionalValue(const boost::property_tree::ptree& pt, std::string attribute, T defaultValue) {
@@ -664,9 +665,21 @@ struct GpuSnapshot {
     size_t      usedMiB  = 0;
 };
 
+// Safe check: is CUDA runtime available? (prevents crash on CPU-only PCs)
+static bool IsCudaAvailable() {
+    static int cached = -1;
+    if (cached < 0) {
+        HMODULE h = LoadLibraryA("nvcuda.dll");
+        cached = (h != nullptr) ? 1 : 0;
+        if (h) FreeLibrary(h);
+    }
+    return cached == 1;
+}
+
 // Query current GPU VRAM usage for all devices
 static std::vector<GpuSnapshot> QueryGpuVram() {
     std::vector<GpuSnapshot> snapshots;
+    if (!IsCudaAvailable()) return snapshots;
     int deviceCount = 0;
     if (cudaGetDeviceCount(&deviceCount) != cudaSuccess) return snapshots;
     for (int i = 0; i < deviceCount; i++) {
@@ -693,6 +706,7 @@ static std::vector<GpuSnapshot> QueryGpuVram() {
 // Measure per-GPU free VRAM (returns array indexed by device)
 static std::vector<size_t> GetPerGpuFreeMiB() {
     std::vector<size_t> result;
+    if (!IsCudaAvailable()) return result;
     int deviceCount = 0;
     if (cudaGetDeviceCount(&deviceCount) != cudaSuccess) return result;
     int prevDevice;
@@ -712,6 +726,11 @@ static ThreadSafeLog     g_log;
 
 // Log GPU info using CUDA runtime
 static void LogGpuInfo() {
+    if (!IsCudaAvailable()) {
+        g_log.add("No NVIDIA GPU detected — running in CPU mode");
+        printf("[GPU] No NVIDIA GPU detected — running in CPU mode\n");
+        return;
+    }
     int deviceCount = 0;
     cudaError_t err = cudaGetDeviceCount(&deviceCount);
     if (err != cudaSuccess) {
@@ -749,6 +768,12 @@ static void LogGpuInfo() {
     printf("============================================================\n");
 }
 
+// Global inference mutex: serializes inference on non-NVIDIA GPUs (DirectML/OpenVINO).
+// DirectML is not thread-safe when multiple ORT sessions run concurrently on the
+// same integrated GPU — causes access violations on 4K frames.
+// On NVIDIA, each task has its own CUDA context so no serialization needed.
+static std::mutex g_inferenceMutex;
+
 // Worker thread: reads RTSP frames and runs ALPR inference
 // RTSP client and ALPR engine are pre-created on the main thread to avoid
 // race conditions in CreateANSRTSPHandle / CreateANSALPRHandle.
@@ -845,12 +870,18 @@ static void ALPRWorkerThread(int taskId,
         if (grabMs > maxGrabMs) maxGrabMs = grabMs;
 
         // Run ALPR inference
+        bool isNvidia = (ANSCENTER::EPLoader::Current().type == ANSCENTER::EngineType::NVIDIA_GPU);
+        fprintf(stderr, "[Worker T%d] frame %d: calling inference %dx%d...\n",
+            taskId, state.frameCount + 1, framePtr->cols, framePtr->rows);
         auto infStart = std::chrono::steady_clock::now();
         std::string lpnResult, jpegImage;
-        // Pass framePtr directly — NOT a copy. ANSGpuFrameRegistry::lookup()
+        {
-        // matches by cv::Mat* pointer, so `new cv::Mat(*framePtr)` would create
+            std::unique_lock<std::mutex> infLock(g_inferenceMutex, std::defer_lock);
-        // a different pointer the registry doesn't know, breaking NV12 zero-copy.
+            if (!isNvidia) infLock.lock();
-        ANSALPR_RunInferenceComplete_CPP(&alprHandle, &framePtr, cameraId.c_str(), 0, 0, lpnResult, jpegImage);
+            ANSALPR_RunInferenceComplete_CPP(&alprHandle, &framePtr, cameraId.c_str(), 0, 0, lpnResult, jpegImage);
+        }
+        fprintf(stderr, "[Worker T%d] frame %d: inference done, result len=%zu\n",
+            taskId, state.frameCount + 1, lpnResult.size());
 
         // Release stream lock — inference is done, CHAOS can now safely destroy.
         streamLock.unlock();
@@ -950,25 +981,454 @@ static void ALPRWorkerThread(int taskId,
     g_log.add(prefix + " Worker loop exited");
 }
 
+// =============================================================================
+//  ANSLPR_SingleTask_Test — 1 stream, 1 AI task. For isolating DirectML/ORT
+//  issues on non-NVIDIA GPUs. If this works but 2-task crashes, it's concurrency.
+// =============================================================================
+int ANSLPR_SingleTask_Test() {
+    ANSCENTER::ANSOPENCV::InitCameraNetwork();
+    g_log.init();
+
+    printf("\n");
+    printf("============================================================\n");
+    printf("  ANSLPR Single-Task Test — 1 Stream, 1 AI Task\n");
+    printf("  Press ESC to stop\n");
+    printf("  Log file: %s\n", LOG_FILE_PATH);
+    printf("============================================================\n\n");
+
+    g_log.add("============================================================");
+    g_log.add("  ANSLPR Single-Task Test — 1 Stream, 1 AI Task");
+    g_log.add("============================================================");
+
+    const std::string streamUrl = "rtsp://admin:admin123@103.156.0.133:8010/cam/realmonitor?channel=1&subtype=0";
+    g_log.add("Stream: " + streamUrl);
+
+    // --- Create RTSP client ---
+    ANSCENTER::ANSRTSPClient* rtspClient = nullptr;
+    printf("[Stream0] Creating RTSP handle...\n");
+    int rtspResult = CreateANSRTSPHandle(&rtspClient, "", "", "", streamUrl.c_str());
+    if (rtspResult != 1 || rtspClient == nullptr) {
+        printf("[Stream0] FAILED to create RTSP handle\n");
+        ANSCENTER::ANSOPENCV::DeinitCameraNetwork();
+        return -1;
+    }
+    SetRTSPImageQuality(&rtspClient, 0);
+    SetRTSPHWDecoding(&rtspClient, -1);  // Force software decoding
+    StartRTSP(&rtspClient);
+    g_log.add("[Stream0] RTSP started (software decode)");
+
+    // --- Create single ALPR engine ---
+    ANSCENTER::ANSALPR* alprHandle = nullptr;
+    std::string modelZipFile = "C:\\ProgramData\\ANSCENTER\\ANSVIS Server\\ANSALPR\\ANS_ALPR_v1.2.zip";
+    printf("[Task0] Creating ALPR handle...\n");
+    auto engineStart = std::chrono::steady_clock::now();
+    int createResult = CreateANSALPRHandle(&alprHandle, "", modelZipFile.c_str(), "",
+                                            1, 0.5, 0.5, 0.5);
+    if (createResult != 1 || alprHandle == nullptr) {
+        printf("[Task0] FAILED to create ALPR handle (result=%d)\n", createResult);
+        StopRTSP(&rtspClient); ReleaseANSRTSPHandle(&rtspClient);
+        ANSCENTER::ANSOPENCV::DeinitCameraNetwork();
+        return -1;
+    }
+
+    printf("[Task0] Loading ALPR engine...\n");
+    int loadResult = LoadANSALPREngineHandle(&alprHandle);
+    auto engineEnd = std::chrono::steady_clock::now();
+    double loadMs = std::chrono::duration<double, std::milli>(engineEnd - engineStart).count();
+    if (loadResult != 1) {
+        printf("[Task0] FAILED to load ALPR engine (result=%d)\n", loadResult);
+        ReleaseANSALPRHandle(&alprHandle);
+        StopRTSP(&rtspClient); ReleaseANSRTSPHandle(&rtspClient);
+        ANSCENTER::ANSOPENCV::DeinitCameraNetwork();
+        return -1;
+    }
+    printf("[Task0] Engine loaded in %.0f ms\n", loadMs);
+    g_log.add("[Task0] Engine loaded in " + std::to_string((int)loadMs) + " ms");
+
+    // --- Single-task worker + display ---
+    TaskState state;
+    state.engineLoaded = true;
+    state.streamOk = true;
+    state.statusMsg = "Running";
+
+    std::mutex streamGuard;
+    std::thread worker(ALPRWorkerThread, 0, &rtspClient, &streamGuard, alprHandle, std::ref(state));
+
+    const int cellW = 800, cellH = 600;
+    const int logPanelH = 80;
+    std::string windowTitle = "ANSLPR Single-Task Test";
+    cv::namedWindow(windowTitle, cv::WINDOW_NORMAL);
+    cv::resizeWindow(windowTitle, cellW, cellH + logPanelH);
+
+    auto testStart = std::chrono::steady_clock::now();
+
+    while (g_running.load()) {
+        cv::Mat canvas(cellH + logPanelH, cellW, CV_8UC3, cv::Scalar(30, 30, 30));
+
+        cv::Mat cell;
+        double fps = 0, infMs = 0;
+        int fCount = 0, dCount = 0;
+        std::string lastPlate;
+        {
+            std::lock_guard<std::mutex> lk(state.mtx);
+            if (!state.displayFrame.empty())
+                cv::resize(state.displayFrame, cell, cv::Size(cellW, cellH));
+            fps = state.fps;
+            infMs = state.inferenceMs;
+            fCount = state.frameCount;
+            dCount = state.detectionCount;
+            lastPlate = state.lastPlate;
+        }
+
+        if (cell.empty())
+            cell = cv::Mat(cellH, cellW, CV_8UC3, cv::Scalar(40, 40, 40));
+
+        cv::rectangle(cell, cv::Rect(0, cellH - 40, cellW, 40), cv::Scalar(0, 0, 0), cv::FILLED);
+        char bar[256];
+        snprintf(bar, sizeof(bar), "T0 | %.1f FPS | %.0fms | F:%d | D:%d | %s",
+                 fps, infMs, fCount, dCount, lastPlate.empty() ? "-" : lastPlate.c_str());
+        cv::putText(cell, bar, cv::Point(5, cellH - 12),
+                    cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 255, 0), 1);
+        cell.copyTo(canvas(cv::Rect(0, 0, cellW, cellH)));
+
+        cv::Mat logPanel = canvas(cv::Rect(0, cellH, cellW, logPanelH));
+        logPanel.setTo(cv::Scalar(20, 20, 20));
+        auto elapsed = std::chrono::duration<double>(std::chrono::steady_clock::now() - testStart).count();
+        char header[256];
+        snprintf(header, sizeof(header), "Elapsed: %.0fs | 1 camera, 1 AI task | %.1f FPS | Press ESC to stop",
+                 elapsed, fps);
+        cv::putText(logPanel, header, cv::Point(10, 20),
+                    cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(200, 200, 0), 1);
+
+        cv::imshow(windowTitle, canvas);
+        if (cv::waitKey(30) == 27) {
+            g_log.add("ESC pressed — stopping...");
+            printf("\nESC pressed — stopping...\n");
+            g_running.store(false);
+        }
+    }
+
+    if (worker.joinable()) worker.join();
+
+    printf("\n============================================================\n");
+    printf("  FINAL SUMMARY\n");
+    printf("  Frames: %d | Detections: %d | FPS: %.1f | InfMs: %.0f\n",
+        state.frameCount, state.detectionCount, state.fps, state.inferenceMs);
+    printf("============================================================\n");
+
+    ReleaseANSALPRHandle(&alprHandle);
+    StopRTSP(&rtspClient);
+    ReleaseANSRTSPHandle(&rtspClient);
+    g_log.close();
+    cv::destroyAllWindows();
+    ANSCENTER::ANSOPENCV::DeinitCameraNetwork();
+    return 0;
+}
+
+// =============================================================================
+//  ANSLPR_CPU_StressTest — Lightweight 2-task stress test for CPU-only PCs
+//  Uses ANSALPR_OD (engineType=1) which auto-falls-back to ONNX Runtime on CPU.
+//  No VRAM tracking, no NVDEC alignment, no chaos thread.
+// =============================================================================
+int ANSLPR_CPU_StressTest() {
+    ANSCENTER::ANSOPENCV::InitCameraNetwork();
+    g_log.init();
+
+    const int NUM_STREAMS = 2;
+    const int NUM_TASKS   = 2;
+
+    printf("\n");
+    printf("============================================================\n");
+    printf("  ANSLPR CPU Stress Test — %d Parallel ALPR Tasks\n", NUM_TASKS);
+    printf("  Press ESC to stop\n");
+    printf("  Log file: %s\n", LOG_FILE_PATH);
+    printf("============================================================\n\n");
+
+    g_log.add("============================================================");
+    g_log.add("  ANSLPR CPU Stress Test — " + std::to_string(NUM_TASKS) + " Tasks");
+    g_log.add("============================================================");
+
+    // --- RTSP URLs (2 camera streams) ---
+    const std::string streamUrls[NUM_STREAMS] = {
+        "rtsp://admin:admin123@103.156.0.133:8010/cam/realmonitor?channel=1&subtype=0",
+        "rtsp://nhathuocngoclinh.zapto.org:600/rtsp/streaming?channel=01&subtype=0"
+    };
+    const int taskStreamMap[NUM_TASKS] = { 0, 1 };
+
+    for (int i = 0; i < NUM_STREAMS; i++)
+        g_log.add("Stream " + std::to_string(i) + ": " + streamUrls[i]);
+
+    // --- Task states ---
+    TaskState taskStates[NUM_TASKS];
+
+    // --- Create RTSP clients (software decoding) ---
+    ANSCENTER::ANSRTSPClient* rtspClients[NUM_STREAMS] = {};
+    for (int s = 0; s < NUM_STREAMS; s++) {
+        printf("[Stream%d] Creating RTSP handle...\n", s);
+        int result = CreateANSRTSPHandle(&rtspClients[s], "", "", "", streamUrls[s].c_str());
+        if (result != 1 || rtspClients[s] == nullptr) {
+            printf("[Stream%d] FAILED to create RTSP handle\n", s);
+            g_log.add("[Stream" + std::to_string(s) + "] RTSP create FAILED");
+            rtspClients[s] = nullptr;
+            continue;
+        }
+        SetRTSPImageQuality(&rtspClients[s], 0);
+        SetRTSPHWDecoding(&rtspClients[s], -1);  // HW_DECODING_DISABLE: force software decoding
+        StartRTSP(&rtspClients[s]);
+        g_log.add("[Stream" + std::to_string(s) + "] RTSP started (software decode)");
+    }
+
+    // --- Create ALPR engines (engineType=1 → ANSALPR_OD, auto CPU/GPU) ---
+    ANSCENTER::ANSALPR* alprHandles[NUM_TASKS] = {};
+    std::string modelZipFile = "C:\\ProgramData\\ANSCENTER\\ANSVIS Server\\ANSALPR\\ANS_ALPR_v1.2.zip";
+    int engineType = 1; // ANSALPR_OD: auto CPU/GPU
+    double detThresh = 0.5, ocrThresh = 0.5, colThresh = 0.5;
+
+    for (int i = 0; i < NUM_TASKS; i++) {
+        char tag[32];
+        snprintf(tag, sizeof(tag), "[Task%d]", i);
+        int streamIdx = taskStreamMap[i];
+        if (rtspClients[streamIdx] == nullptr) {
+            printf("%s Skipped — Stream%d not available\n", tag, streamIdx);
+            continue;
+        }
+
+        {
+            std::lock_guard<std::mutex> lk(taskStates[i].mtx);
+            taskStates[i].streamOk = true;
+            taskStates[i].statusMsg = "Loading ALPR engine...";
+        }
+
+        printf("%s Creating ALPR handle...\n", tag);
+        auto engineStart = std::chrono::steady_clock::now();
+        int createResult = CreateANSALPRHandle(&alprHandles[i], "", modelZipFile.c_str(), "",
+                                                engineType, detThresh, ocrThresh, colThresh);
+        if (createResult != 1 || alprHandles[i] == nullptr) {
+            printf("%s FAILED to create ALPR handle (result=%d)\n", tag, createResult);
+            g_log.add(std::string(tag) + " ALPR create FAILED");
+            continue;
+        }
+
+        printf("%s Loading ALPR engine...\n", tag);
+        int loadResult = LoadANSALPREngineHandle(&alprHandles[i]);
+        auto engineEnd = std::chrono::steady_clock::now();
+        double loadMs = std::chrono::duration<double, std::milli>(engineEnd - engineStart).count();
+
+        if (loadResult != 1) {
+            printf("%s FAILED to load ALPR engine (result=%d)\n", tag, loadResult);
+            g_log.add(std::string(tag) + " Engine load FAILED");
+            ReleaseANSALPRHandle(&alprHandles[i]);
+            alprHandles[i] = nullptr;
+            continue;
+        }
+
+        char buf[256];
+        snprintf(buf, sizeof(buf), "%s Engine loaded in %.0f ms (Stream%d)", tag, loadMs, streamIdx);
+        printf("%s\n", buf);
+        g_log.add(buf);
+
+        {
+            std::lock_guard<std::mutex> lk(taskStates[i].mtx);
+            taskStates[i].engineLoaded = true;
+            taskStates[i].statusMsg = "Running";
+        }
+    }
+
+    // --- Launch worker threads ---
+    std::mutex streamGuards[NUM_STREAMS];
+    std::thread workers[NUM_TASKS];
+    for (int i = 0; i < NUM_TASKS; i++) {
+        int streamIdx = taskStreamMap[i];
+        if (rtspClients[streamIdx] && alprHandles[i]) {
+            workers[i] = std::thread(ALPRWorkerThread, i,
+                                     &rtspClients[streamIdx],
+                                     &streamGuards[streamIdx],
+                                     alprHandles[i],
+                                     std::ref(taskStates[i]));
+        }
+    }
+
+    // --- Display loop ---
+    const int cellW = 640, cellH = 480;
+    const int logPanelH = 120;
+    const int gridCols = 2, gridRows = 1;
+    std::string windowTitle = "ANSLPR CPU Stress Test";
+    cv::namedWindow(windowTitle, cv::WINDOW_NORMAL);
+    cv::resizeWindow(windowTitle, cellW * gridCols, cellH * gridRows + logPanelH);
+
+    auto testStart = std::chrono::steady_clock::now();
+
+    while (g_running.load()) {
+        cv::Mat canvas(cellH * gridRows + logPanelH, cellW * gridCols, CV_8UC3, cv::Scalar(30, 30, 30));
+
+        for (int i = 0; i < NUM_TASKS; i++) {
+            int col = i % gridCols, row = i / gridCols;
+            cv::Rect roi(col * cellW, row * cellH, cellW, cellH);
+
+            cv::Mat cell;
+            double fps = 0, infMs = 0;
+            int fCount = 0, dCount = 0;
+            std::string statusMsg, lastPlate;
+            bool engineLoaded = false;
+            {
+                std::lock_guard<std::mutex> lk(taskStates[i].mtx);
+                if (!taskStates[i].displayFrame.empty())
+                    cv::resize(taskStates[i].displayFrame, cell, cv::Size(cellW, cellH));
+                fps = taskStates[i].fps;
+                infMs = taskStates[i].inferenceMs;
+                fCount = taskStates[i].frameCount;
+                dCount = taskStates[i].detectionCount;
+                statusMsg = taskStates[i].statusMsg;
+                lastPlate = taskStates[i].lastPlate;
+                engineLoaded = taskStates[i].engineLoaded;
+            }
+
+            if (cell.empty()) {
+                cell = cv::Mat(cellH, cellW, CV_8UC3, cv::Scalar(40, 40, 40));
+                cv::putText(cell, "Task " + std::to_string(i) + ": " + statusMsg,
+                            cv::Point(20, cellH / 2),
+                            cv::FONT_HERSHEY_SIMPLEX, 0.8, cv::Scalar(100, 100, 255), 2);
+            }
+
+            // Status bar
+            cv::rectangle(cell, cv::Rect(0, cellH - 40, cellW, 40), cv::Scalar(0, 0, 0), cv::FILLED);
+            char bar[256];
+            snprintf(bar, sizeof(bar), "T%d(S%d) | %.1f FPS | %.0fms | F:%d | D:%d | %s",
+                     i, taskStreamMap[i], fps, infMs, fCount, dCount,
+                     lastPlate.empty() ? "-" : lastPlate.c_str());
+            cv::Scalar barColor = engineLoaded ? cv::Scalar(0, 255, 0) : cv::Scalar(0, 100, 255);
+            cv::putText(cell, bar, cv::Point(5, cellH - 12),
+                        cv::FONT_HERSHEY_SIMPLEX, 0.45, barColor, 1);
+            cell.copyTo(canvas(roi));
+        }
+
+        // Grid line
+        if (gridCols > 1)
+            cv::line(canvas, cv::Point(cellW, 0), cv::Point(cellW, cellH * gridRows),
+                     cv::Scalar(100, 100, 100), 1);
+
+        // Log panel
+        cv::Rect logRoi(0, cellH * gridRows, cellW * gridCols, logPanelH);
+        cv::Mat logPanel = canvas(logRoi);
+        logPanel.setTo(cv::Scalar(20, 20, 20));
+
+        auto elapsed = std::chrono::duration<double>(std::chrono::steady_clock::now() - testStart).count();
+        double totalFps = 0;
+        for (int i = 0; i < NUM_TASKS; i++) {
+            std::lock_guard<std::mutex> lk(taskStates[i].mtx);
+            totalFps += taskStates[i].fps;
+        }
+        char header[256];
+        snprintf(header, sizeof(header),
+                 "Elapsed: %.0fs | %d cameras, %d AI tasks | Total: %.1f FPS | Press ESC to stop",
+                 elapsed, NUM_STREAMS, NUM_TASKS, totalFps);
+        cv::putText(logPanel, header, cv::Point(10, 20),
+                    cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(200, 200, 0), 1);
+
+        // Per-task summary
+        for (int i = 0; i < NUM_TASKS; i++) {
+            std::lock_guard<std::mutex> lk(taskStates[i].mtx);
+            char tLine[256];
+            snprintf(tLine, sizeof(tLine),
+                "T%d(S%d): FPS=%.1f Inf=%.0fms Frames=%d Det=%d",
+                i, taskStreamMap[i], taskStates[i].fps, taskStates[i].inferenceMs,
+                taskStates[i].frameCount, taskStates[i].detectionCount);
+            cv::putText(logPanel, tLine, cv::Point(10, 42 + i * 18),
+                        cv::FONT_HERSHEY_SIMPLEX, 0.4, cv::Scalar(200, 200, 200), 1);
+        }
+
+        // Recent log
+        auto recentLogs = g_log.getRecent(3);
+        int logY = 42 + NUM_TASKS * 18 + 5;
+        for (const auto& line : recentLogs) {
+            if (logY > logPanelH - 5) break;
+            std::string display = (line.size() > 120) ? line.substr(0, 117) + "..." : line;
+            cv::putText(logPanel, display, cv::Point(10, logY),
+                        cv::FONT_HERSHEY_PLAIN, 1.0, cv::Scalar(140, 140, 140), 1);
+            logY += 15;
+        }
+
+        cv::imshow(windowTitle, canvas);
+        int key = cv::waitKey(30);
+        if (key == 27) {
+            g_log.add("ESC pressed — stopping...");
+            printf("\nESC pressed — stopping...\n");
+            g_running.store(false);
+        }
+    }
+
+    // --- Wait for workers ---
+    for (int i = 0; i < NUM_TASKS; i++) {
+        if (workers[i].joinable()) workers[i].join();
+    }
+
+    // --- Final summary ---
+    double totalElapsed = std::chrono::duration<double>(
+        std::chrono::steady_clock::now() - testStart).count();
+    printf("\n============================================================\n");
+    printf("  FINAL SUMMARY (runtime: %.0fs)\n", totalElapsed);
+    printf("============================================================\n");
+    double totalFpsFinal = 0;
+    for (int i = 0; i < NUM_TASKS; i++) {
+        char buf[256];
+        snprintf(buf, sizeof(buf), "  Task %d (Stream %d): %d frames, %d detections, FPS=%.1f, InfMs=%.0f",
+            i, taskStreamMap[i], taskStates[i].frameCount, taskStates[i].detectionCount,
+            taskStates[i].fps, taskStates[i].inferenceMs);
+        printf("%s\n", buf);
+        g_log.add(buf);
+        totalFpsFinal += taskStates[i].fps;
+    }
+    printf("  Total throughput: %.1f FPS\n", totalFpsFinal);
+    printf("============================================================\n");
+
+    // --- Cleanup ---
+    for (int i = 0; i < NUM_TASKS; i++) {
+        if (alprHandles[i]) ReleaseANSALPRHandle(&alprHandles[i]);
+    }
+    for (int s = 0; s < NUM_STREAMS; s++) {
+        if (rtspClients[s]) {
+            StopRTSP(&rtspClients[s]);
+            ReleaseANSRTSPHandle(&rtspClients[s]);
+        }
+    }
+
+    g_log.close();
+    cv::destroyAllWindows();
+    ANSCENTER::ANSOPENCV::DeinitCameraNetwork();
+    return 0;
+}
+
 int ANSLPR_MultiGPU_StressTest() {
     ANSCENTER::ANSOPENCV::InitCameraNetwork();
 
     // --- Initialize log file ---
     g_log.init();
 
+    printf("\n");
+    // --- Auto-detect GPU availability (safe on CPU-only PCs without CUDA runtime) ---
+    int gpuCount = 0;
+    bool hasGpu = false;
+    if (IsCudaAvailable()) {
+        cudaGetDeviceCount(&gpuCount);
+        hasGpu = (gpuCount > 0);
+    }
+    const char* modeStr = hasGpu ? "GPU (NVIDIA CUDA)" : "CPU (Software Decoding)";
+
     printf("\n");
     printf("============================================================\n");
-    printf("  ANSLPR Multi-GPU Stress Test — 5 Parallel ALPR Tasks\n");
+    printf("  ANSLPR Multi-Engine Stress Test — 5 Parallel ALPR Tasks\n");
+    printf("  Mode: %s\n", modeStr);
     printf("  (4 cameras, 5 AI tasks — Task 4 shares Stream 2)\n");
     printf("  Press ESC to stop\n");
     printf("  Log file: %s\n", LOG_FILE_PATH);
     printf("============================================================\n\n");
 
     g_log.add("============================================================");
-    g_log.add("  ANSLPR Multi-GPU Stress Test — 5 Parallel ALPR Tasks");
+    g_log.add("  ANSLPR Multi-Engine Stress Test — 5 Parallel ALPR Tasks");
+    g_log.add("  Mode: " + std::string(modeStr));
     g_log.add("============================================================");
 
-    // --- Log GPU info for diagnostics ---
+    // --- Log GPU info for diagnostics (safe on CPU — prints "no GPU found") ---
     LogGpuInfo();
 
     // --- RTSP URLs (4 independent camera streams) ---
@@ -1027,7 +1487,7 @@ int ANSLPR_MultiGPU_StressTest() {
             continue;
         }
         SetRTSPImageQuality(&rtspClients[s], 0);
-        SetRTSPHWDecoding(&rtspClients[s], 7);  // HW_DECODING_CUDA: force CUDA/NVDEC zero-copy path
+        if (hasGpu) SetRTSPHWDecoding(&rtspClients[s], 7);  // CUDA HW decode only with GPU
         StartRTSP(&rtspClients[s]);
         g_log.add("[Stream" + std::to_string(s) + "] RTSP started");
     }
@@ -1040,7 +1500,7 @@ int ANSLPR_MultiGPU_StressTest() {
     // =========================================================================
     ANSCENTER::ANSALPR* alprHandles[NUM_TASKS] = {};
     std::string modelZipFile = "C:\\ProgramData\\ANSCENTER\\ANSVIS Server\\ANSALPR\\ANS_ALPR_v1.2.zip";
-    int engineType = 1; // NVIDIA_GPU
+    int engineType = 1; // ANSALPR_OD: auto-detects GPU/CPU, uses ONNX Runtime on CPU
     double detThresh = 0.5, ocrThresh = 0.5, colThresh = 0.5;
 
     for (int i = 0; i < NUM_TASKS; i++) {
@@ -1074,11 +1534,12 @@ int ANSLPR_MultiGPU_StressTest() {
             continue;
         }
 
-        printf("%s Loading ALPR engine (TensorRT)...\n", tag);
+        printf("%s Loading ALPR engine (%s)...\n", tag, hasGpu ? "TensorRT" : "CPU");
         g_log.add(std::string(tag) + " Loading ALPR engine...");
 
-        // Snapshot VRAM before engine load to measure consumption
+        // Snapshot VRAM before engine load to measure consumption (GPU only)
-        auto vramBefore = GetPerGpuFreeMiB();
+        std::vector<size_t> vramBefore;
+        if (hasGpu) vramBefore = GetPerGpuFreeMiB();
 
         int loadResult = LoadANSALPREngineHandle(&alprHandles[i]);
         auto engineEnd = std::chrono::steady_clock::now();
@@ -1094,40 +1555,47 @@ int ANSLPR_MultiGPU_StressTest() {
             continue;
         }
 
-        // Snapshot VRAM after engine load — find which GPU lost the most VRAM
+        int bestGpu = -1;
-        auto vramAfter = GetPerGpuFreeMiB();
-        int bestGpu = 0;
         size_t maxDelta = 0;
-        size_t gpuCount = vramBefore.size() < vramAfter.size() ? vramBefore.size() : vramAfter.size();
+
-        for (size_t g = 0; g < gpuCount; g++) {
+        if (hasGpu) {
-            size_t delta = (vramBefore[g] > vramAfter[g]) ? (vramBefore[g] - vramAfter[g]) : 0;
+            // Snapshot VRAM after engine load — find which GPU lost the most VRAM
-            if (delta > maxDelta) {
+            auto vramAfter = GetPerGpuFreeMiB();
-                maxDelta = delta;
+            size_t gpuCnt = vramBefore.size() < vramAfter.size() ? vramBefore.size() : vramAfter.size();
-                bestGpu = (int)g;
+            bestGpu = 0;
+            for (size_t g = 0; g < gpuCnt; g++) {
+                size_t delta = (vramBefore[g] > vramAfter[g]) ? (vramBefore[g] - vramAfter[g]) : 0;
+                if (delta > maxDelta) {
+                    maxDelta = delta;
+                    bestGpu = (int)g;
+                }
             }
-        }
 
-        char buf[512];
+            char buf[512];
-        snprintf(buf, sizeof(buf),
+            snprintf(buf, sizeof(buf),
-            "%s Engine loaded in %.0f ms | GPU[%d] | VRAM used: %zu MiB (Stream%d)",
+                "%s Engine loaded in %.0f ms | GPU[%d] | VRAM used: %zu MiB (Stream%d)",
-            tag, loadMs, bestGpu, maxDelta, streamIdx);
+                tag, loadMs, bestGpu, maxDelta, streamIdx);
-        printf("%s\n", buf);
+            printf("%s\n", buf);
-        g_log.add(buf);
+            g_log.add(buf);
 
-        // Log per-GPU VRAM state after this engine load
+            // Log per-GPU VRAM state after this engine load
-        for (size_t g = 0; g < vramAfter.size(); g++) {
+            for (size_t g = 0; g < vramAfter.size(); g++) {
-            size_t total = 0;
+                size_t total = 0;
-            if (g < vramBefore.size()) {
-                // Compute total from free + used
                 auto gpus = QueryGpuVram();
                 if (g < gpus.size()) total = gpus[g].totalMiB;
+                char vbuf[256];
+                snprintf(vbuf, sizeof(vbuf),
+                    "  GPU[%zu] VRAM: %zu MiB free (of %zu MiB)",
+                    g, vramAfter[g], total);
+                printf("%s\n", vbuf);
+                g_log.add(vbuf);
             }
-            char vbuf[256];
+        } else {
-            snprintf(vbuf, sizeof(vbuf),
+            char buf[256];
-                "  GPU[%zu] VRAM: %zu MiB free (of %zu MiB)",
+            snprintf(buf, sizeof(buf), "%s Engine loaded in %.0f ms (CPU mode, Stream%d)",
-                g, vramAfter[g], total);
+                tag, loadMs, streamIdx);
-            printf("%s\n", vbuf);
+            printf("%s\n", buf);
-            g_log.add(vbuf);
+            g_log.add(buf);
         }
 
         {
@@ -1140,6 +1608,8 @@ int ANSLPR_MultiGPU_StressTest() {
     }
 
     // --- Align NVDEC decode GPU with inference GPU for NV12 zero-copy ---
+    // (GPU only — software decoding on CPU doesn't use NVDEC)
+    if (hasGpu)
     // Each stream should decode on the same GPU as its inference engine to enable
     // direct NVDEC→TensorRT zero-copy (0.5ms vs 17ms preprocess per frame).
     //
@@ -1343,7 +1813,7 @@ int ANSLPR_MultiGPU_StressTest() {
                     streamUrls[streamIdx].c_str());
                 if (result == 1 && rtspClients[streamIdx]) {
                     SetRTSPImageQuality(&rtspClients[streamIdx], 0);
-                    SetRTSPHWDecoding(&rtspClients[streamIdx], 7);
+                    if (hasGpu) SetRTSPHWDecoding(&rtspClients[streamIdx], 7);
                     StartRTSP(&rtspClients[streamIdx]);
 
                     auto chaosEnd = std::chrono::steady_clock::now();
@@ -1368,8 +1838,9 @@ int ANSLPR_MultiGPU_StressTest() {
     const int cellW = 480, cellH = 360;  // Smaller cells for 3-column layout
     const int logPanelH = 220;
     const int gridCols = 3, gridRows = 2;
-    cv::namedWindow("ANSLPR Multi-GPU Stress Test", cv::WINDOW_NORMAL);
+    std::string windowTitle = hasGpu ? "ANSLPR Multi-GPU Stress Test" : "ANSLPR CPU Stress Test";
-    cv::resizeWindow("ANSLPR Multi-GPU Stress Test", cellW * gridCols, cellH * gridRows + logPanelH);
+    cv::namedWindow(windowTitle, cv::WINDOW_NORMAL);
+    cv::resizeWindow(windowTitle, cellW * gridCols, cellH * gridRows + logPanelH);
 
     auto testStart = std::chrono::steady_clock::now();
     auto lastGpuSnapshot = std::chrono::steady_clock::now();
@@ -1468,7 +1939,9 @@ int ANSLPR_MultiGPU_StressTest() {
             snprintf(bar1, sizeof(bar1), "T%d(S%d) | %.1f FPS | %.0fms | F:%d | D:%d | %s",
                      i, taskStreamMap[i], fps, infMs, fCount, dCount,
                      lastPlate.empty() ? "-" : lastPlate.c_str());
-            if (gpuId >= 0) {
+            if (!hasGpu) {
+                snprintf(bar2, sizeof(bar2), "CPU mode (software decoding)");
+            } else if (gpuId >= 0) {
                 snprintf(bar2, sizeof(bar2), "GPU[%d] | VRAM: %zu MiB", gpuId, vramMiB);
             } else {
                 snprintf(bar2, sizeof(bar2), "GPU: N/A");
@@ -1572,7 +2045,7 @@ int ANSLPR_MultiGPU_StressTest() {
             gpuLineY += 15;
         }
 
-        cv::imshow("ANSLPR Multi-GPU Stress Test", canvas);
+        cv::imshow(windowTitle, canvas);
         int key = cv::waitKey(30);
         if (key == 27) { // ESC
             g_log.add("ESC pressed — stopping all tasks...");
@@ -2930,6 +3403,136 @@ int ANSLPR_MultiGPU_StressTest_FilePlayer() {
     return 0;
 }
 
+// ANSLPR_OD_CPU_VideoTest — Uses ANSALPR_OD (engineType=1) on Intel CPU/iGPU.
+// ANSALPR_OD auto-detects hardware (OpenVINO on Intel, DirectML on AMD, etc.)
+// No CUDA calls — safe on non-NVIDIA systems.
+int ANSLPR_OD_CPU_VideoTest() {
+    std::cout << "\n============================================================" << std::endl;
+    std::cout << "  ANSLPR CPU/iGPU Test (ANSALPR_OD with auto-detect)" << std::endl;
+    std::cout << "============================================================\n" << std::endl;
+
+    std::string modelZipFile = "C:\\ProgramData\\ANSCENTER\\ANSVIS Server\\ANSALPR\\ANS_ALPR_v1.2.zip";
+    std::string videoFilePath = "C:\\ProgramData\\ANSCENTER\\Shared\\classroom.mp4";
+
+    std::cout << "Model: " << modelZipFile << std::endl;
+    std::cout << "Video: " << videoFilePath << std::endl;
+
+    ANSCENTER::ANSALPR* infHandle = nullptr;
+    int engineType = 1; // ANSALPR_OD (auto-detects HW internally)
+    double detThresh = 0.5, ocrThresh = 0.5, colThresh = 0.5;
+
+    // Step 1: Create handle
+    std::cout << "[LPR-CPU] Step 1: Creating handle..." << std::endl;
+    int createResult = CreateANSALPRHandle(&infHandle, "", modelZipFile.c_str(), "",
+                                            engineType, detThresh, ocrThresh, colThresh);
+    std::cout << "[LPR-CPU] CreateANSALPRHandle result: " << createResult << std::endl;
+    if (createResult != 1 || infHandle == nullptr) {
+        std::cerr << "[LPR-CPU] FAILED: CreateANSALPRHandle returned " << createResult << std::endl;
+        return -1;
+    }
+
+    // Step 2: Load engine
+    std::cout << "[LPR-CPU] Step 2: Loading engine..." << std::endl;
+    int loadResult = LoadANSALPREngineHandle(&infHandle);
+    std::cout << "[LPR-CPU] LoadANSALPREngineHandle result: " << loadResult << std::endl;
+    if (loadResult != 1) {
+        std::cerr << "[LPR-CPU] FAILED: LoadANSALPREngineHandle returned " << loadResult << std::endl;
+        ReleaseANSALPRHandle(&infHandle);
+        return -2;
+    }
+
+    // Step 3: Open video
+    std::cout << "[LPR-CPU] Step 3: Opening video..." << std::endl;
+    cv::VideoCapture capture(videoFilePath);
+    if (!capture.isOpened()) {
+        std::cerr << "[LPR-CPU] FAILED: Could not open video: " << videoFilePath << std::endl;
+        ReleaseANSALPRHandle(&infHandle);
+        return -3;
+    }
+
+    int totalFrames = static_cast<int>(capture.get(cv::CAP_PROP_FRAME_COUNT));
+    std::cout << "[LPR-CPU] Video opened: " << totalFrames << " frames" << std::endl;
+
+    // Step 4: Run inference
+    std::cout << "[LPR-CPU] Step 4: Running inference..." << std::endl;
+    boost::property_tree::ptree pt;
+    int frameIndex = 0;
+    int totalDetections = 0;
+    double totalInferenceMs = 0.0;
+    int maxFrames = 200;
+
+    while (frameIndex < maxFrames) {
+        cv::Mat frame;
+        if (!capture.read(frame)) {
+            std::cout << "[LPR-CPU] End of video at frame " << frameIndex << std::endl;
+            break;
+        }
+        frameIndex++;
+
+        unsigned int bufferLength = 0;
+        unsigned char* jpeg_bytes = CVMatToBytes(frame, bufferLength);
+        int height = frame.rows;
+        int width = frame.cols;
+
+        auto start = std::chrono::system_clock::now();
+        std::string detectionResult = ANSALPR_RunInferenceBinary(&infHandle, jpeg_bytes, width, height);
+        auto end = std::chrono::system_clock::now();
+        auto elapsed = std::chrono::duration_cast<std::chrono::milliseconds>(end - start);
+        totalInferenceMs += static_cast<double>(elapsed.count());
+
+        delete[] jpeg_bytes;
+
+        if (!detectionResult.empty()) {
+            try {
+                pt.clear();
+                std::stringstream ss;
+                ss << detectionResult;
+                boost::property_tree::read_json(ss, pt);
+                int detCount = 0;
+                BOOST_FOREACH(const boost::property_tree::ptree::value_type& child, pt.get_child("results")) {
+                    const boost::property_tree::ptree& r = child.second;
+                    const auto class_name = GetData<std::string>(r, "class_name");
+                    const auto x = GetData<float>(r, "x");
+                    const auto y = GetData<float>(r, "y");
+                    const auto w = GetData<float>(r, "width");
+                    const auto h = GetData<float>(r, "height");
+                    detCount++;
+                    cv::rectangle(frame, cv::Rect(x, y, w, h), cv::Scalar(0, 255, 0), 2);
+                    cv::putText(frame, class_name, cv::Point(x, y - 5),
+                        0, 0.6, cv::Scalar(0, 0, 255), 1, cv::LINE_AA);
+                }
+                totalDetections += detCount;
+            }
+            catch (...) {}
+        }
+
+        if (frameIndex % 10 == 0) {
+            double avgSoFar = totalInferenceMs / frameIndex;
+            std::cout << "[LPR-CPU] Frame " << frameIndex << "/" << maxFrames
+                << " | Time: " << elapsed.count() << "ms"
+                << " | Avg: " << static_cast<int>(avgSoFar) << "ms"
+                << " | Detections: " << totalDetections << std::endl;
+        }
+
+        cv::imshow("ANSLPR CPU Test", frame);
+        if (cv::waitKey(1) == 27) break;
+    }
+
+    // Summary
+    double avgMs = (frameIndex > 0) ? (totalInferenceMs / frameIndex) : 0.0;
+    std::cout << "\n=== LPR CPU Test Summary ===" << std::endl;
+    std::cout << "Frames processed: " << frameIndex << std::endl;
+    std::cout << "Total detections: " << totalDetections << std::endl;
+    std::cout << "Avg inference:    " << avgMs << " ms/frame" << std::endl;
+    std::cout << "Total time:       " << totalInferenceMs << " ms" << std::endl;
+    std::cout << (frameIndex > 0 ? "[LPR-CPU] PASSED" : "[LPR-CPU] FAILED") << std::endl;
+
+    capture.release();
+    cv::destroyAllWindows();
+    ReleaseANSALPRHandle(&infHandle);
+    return (frameIndex > 0) ? 0 : -4;
+}
+
 int main()
 {
    // ANSLPR_OD_INDOInferences_FileTest();
@@ -2940,9 +3543,12 @@ int main()
     //for (int i = 0; i < 100; i++) {
     //    ANSLPR_CPU_Inferences_FileTest();
     //}
-    ANSLPR_MultiGPU_StressTest();
+    //ANSLPR_SingleTask_Test();
+    ANSLPR_CPU_StressTest();
+    //ANSLPR_MultiGPU_StressTest();
     //ANSLPR_MultiGPU_StressTest_SimulatedCam();
    // ANSLPR_MultiGPU_StressTest_FilePlayer();
+    //ANSLPR_OD_CPU_VideoTest();
     return 0;
 
 }

tests/ANSLPR-UnitTest/CMakeLists.txt

@@ -7,6 +7,7 @@ target_include_directories(ANSLPR-UnitTest PRIVATE
     ${CMAKE_SOURCE_DIR}/modules/ANSLPR
     ${CMAKE_SOURCE_DIR}/modules/ANSLPR/include
     ${CMAKE_SOURCE_DIR}/modules/ANSODEngine
+    ${CMAKE_SOURCE_DIR}/core/ANSLibsLoader/include
     ${CMAKE_SOURCE_DIR}/modules/ANSCV
     ${CMAKE_SOURCE_DIR}/MediaClient
     ${CMAKE_SOURCE_DIR}/MediaClient/media
@@ -36,6 +37,7 @@ target_link_libraries(ANSLPR-UnitTest
     PRIVATE ANSODEngine
     PRIVATE ANSCV
     PRIVATE ANSLicensingSystem
+    PRIVATE ANSLibsLoader
     PRIVATE anslicensing
     PRIVATE ANSMOT
     PRIVATE opencv

tests/ANSODEngine-UnitTest/ANSODEngine-UnitTest.cpp

@@ -1449,8 +1449,8 @@ int YOLO26ODYolo12Test() {
 }
 int YOLO26ODYolo11Test() {
     std::string modelFilePath = "C:\\Projects\\ANSVIS\\Models\\ANS_VehicleDetection_v2.0.zip";
-    std::string videoFile = "E:\\Programs\\DemoAssets\\Videos\\road.mp4";
+    std::string videoFile = "C:\\ProgramData\\ANSCENTER\\Shared\\road.mp4";
-    int modelType = 31; // ONNX YOLO (30) RT YOLO (31)
+    int modelType = 3; // ONNX YOLO (30) RT YOLO (31)
     VideoDetectorEngine(modelFilePath, videoFile, modelType);
     return 0;
 }
@@ -1861,12 +1861,12 @@ int main()
     //YOLO26POSEYolo11Test();
     //YOLO26CLYolo11Test();
      //YOLO26ODYolo12Test();
-    //YOLO26ODYolo11Test();
+    YOLO26ODYolo11Test();
     //YOLO26ODYolo10Test();
     //YOLO26OBBYolo11Test();
     //SAM3ONNX_ImageTest();   // ORT reference — runs first, prints decoder input stats
     //SAM3TRT_ImageTest();    // TRT under test — compare decoder input stats with above
-    CustomModel_StressTest_FilePlayer();  // Multi-task stress test (LabVIEW flow)
+    //CustomModel_StressTest_FilePlayer();  // Multi-task stress test (LabVIEW flow)
     //SAM3TRT_UnitTest();  // TensorRT SAM3 test (in ANSSAM3-UnitTest.cpp)
     //TensorRT10Test();
     //FireNSmokeCustomDetection();