ANSCORE/modules/ANSODEngine/dllmain.cpp

// dllmain.cpp : Defines the entry point for the DLL application.
#include "pch.h"
#include "ANSODEngine.h"
#include "NV12PreprocessHelper.h"   // tl_currentGpuFrame()
#include "ANSGpuFrameRegistry.h"    // gpu_frame_lookup(cv::Mat*)
#include "engine/TRTEngineCache.h"   // clearAll() on DLL_PROCESS_DETACH
#include "engine/EnginePoolManager.h" // clearAll() on DLL_PROCESS_DETACH
#include <climits>                    // INT_MIN
#include "ANSLicense.h"              // ANS_DBG macro for DebugView

// Process-wide flag: when true, all engines force single-GPU path (no pool, no idle timers).
// Defined here, declared extern in EngineBuildLoadNetwork.inl.
ANSODENGINE_API std::atomic<bool> g_forceNoPool{false};

// Canonical thread-local GpuFrameData* for NV12 zero-copy fast path.
// Exported so that ALL DLLs (ANSLPR, ANSFR, ANSOCR) share the SAME
// thread_local slot.  Without this, each DLL gets its own inline copy
// of tl_currentGpuFrame()'s thread_local — ANSLPR sets its copy, but
// ANSODEngine's tryNV12() reads ANSODEngine's copy (nullptr), silently
// disabling NV12 zero-copy for ALPR, FR, and OCR inference.
extern "C" ANSODENGINE_API GpuFrameData** ANSODEngine_GetTlsGpuFrame() {
    thread_local GpuFrameData* ptr = nullptr;
    return &ptr;
}

#include "ANSYOLOOD.h"
#include "ANSODHUB.h"
#include "ANSCUSTOMPY.h"
#include "ANSTENSORRTOD.h"
#include "ANSTENSORRTCL.h"
#include "ANSOPENVINOCL.h"
#include "ANSOPENVINOOD.h"
#include "ANSYOLOV10RTOD.h"
#include "ANSYOLOV10OVOD.h"
#include "ANSYOLO12OD.h"
#include "ANSONNXCL.h"
#include "ANSONNXPOSE.h"
#include "ANSTENSORRTPOSE.h"
#include "ANSMotionDetector.h"
#include "ANSCUSTOMDetector.h"
#include "ANSONNXSEG.h"	
#include "ANSTENSORRTSEG.h"
#include "ANSONNXOBB.h"
#include "ANSOVSEG.h"
#include "ANSFD.h"
#include "ANSANOMALIB.h"
#include "ANSPOSE.h"
#include "ANSSAM.h"
#include "Movienet.h"
#include "ANSSAM3.h"	
#include "ANSONNXSAM3.h"
#include "ANSRTYOLO.h"
#include "ANSONNXYOLO.h"
#include <pipelines/metadata.h>
#include <models/input_data.h>
#include "utils/visualizer.hpp"
#include <turbojpeg.h>
#include <vector>
#include <map>
#include <string>
#include <unordered_set>
#include <mutex>

// Handle registry with refcount — prevents use-after-free when
// ReleaseANSODHandle is called while inference is still running.
// refcount: starts at 1 on Register. AcquireODHandle increments,
// ReleaseODHandle decrements. Object is destroyed when refcount hits 0.
#include <unordered_map>
#include <condition_variable>
#include <atomic>
#include <cuda_runtime.h>

// ============================================================================
// Round-Robin GPU Assigner
//
// Each call to CreateANSODHandle assigns the next GPU in round-robin order.
// This distributes tasks evenly across all available GPUs:
//   Task 1 → GPU 0, Task 2 → GPU 1, Task 3 → GPU 0, Task 4 → GPU 1, ...
//
// The GPU count is queried once (lazily) and cached. The atomic counter
// ensures thread-safe assignment even when multiple tasks are created
// concurrently.
// ============================================================================
static std::atomic<int> g_gpuRoundRobinCounter{0};
static int g_numGPUs = -1;  // -1 = not yet queried
static std::mutex g_gpuCountMutex;

static int GetNumGPUs() {
	std::lock_guard<std::mutex> lk(g_gpuCountMutex);
	if (g_numGPUs < 0) {
		// Use yield mode before any CUDA call to avoid busy-wait spinning
		// that falsely reports 100% GPU utilization in nvidia-smi.
		cudaSetDeviceFlags(cudaDeviceScheduleYield);
		cudaGetDeviceCount(&g_numGPUs);
		if (g_numGPUs <= 0) g_numGPUs = 1;  // fallback to GPU 0
		std::cout << "Info [GPU]: Detected " << g_numGPUs << " CUDA GPU(s) for round-robin assignment" << std::endl;
	}
	return g_numGPUs;
}

// Determine maxSlotsPerGpu based on GPU topology:
//   1 GPU            → 1  (single slot, no round-robin needed)
//   >1 GPU, VRAM<24GB → 1  (round-robin: 1 slot per GPU)
//   >1 GPU, VRAM≥24GB → -1 (elastic: on-demand slot growth)
// Result is cached after the first query.
static int GetPoolMaxSlotsPerGpu() {
	static int s_result = INT_MIN;
	static std::mutex s_mutex;
	std::lock_guard<std::mutex> lk(s_mutex);
	if (s_result != INT_MIN) return s_result;
	const int n = GetNumGPUs();
	if (n <= 1) {
		s_result = 1;
		std::cout << "Info [GPU]: Single GPU — pool mode: 1 slot, no round-robin" << std::endl;
		return s_result;
	}
	// Multiple GPUs — check VRAM (GPUs are assumed same spec)
	constexpr size_t kLargeVramBytes = 24ULL * 1024 * 1024 * 1024;  // 24 GB
	size_t totalMem = 0, freeMem = 0;
	cudaSetDevice(0);
	cudaMemGetInfo(&freeMem, &totalMem);
	if (totalMem >= kLargeVramBytes) {
		s_result = -1;
		std::cout << "Info [GPU]: " << n << " GPUs, VRAM >= 24 GB — pool mode: elastic" << std::endl;
	} else {
		s_result = 1;
		std::cout << "Info [GPU]: " << n << " GPUs, VRAM < 24 GB — pool mode: round-robin" << std::endl;
	}
	return s_result;
}

// Returns the next GPU index in round-robin order.
// Thread-safe: uses atomic fetch_add.
static int AssignNextGPU() {
	const int numGPUs = GetNumGPUs();
	const int idx = g_gpuRoundRobinCounter.fetch_add(1);
	const int gpuIndex = idx % numGPUs;
	std::cout << "Info [GPU]: Assigning task to GPU " << gpuIndex
		<< " (task #" << (idx + 1) << ", " << numGPUs << " GPU(s) available)" << std::endl;
	return gpuIndex;
}

// Check if a GPU has enough free VRAM for a new engine.
// Returns true if sufficient, false if not.
// minFreeBytes: minimum free VRAM required (default 512 MiB safety margin).
static bool CheckGPUVRAM(int gpuIndex, size_t minFreeBytes = 512ULL * 1024 * 1024) {
	int prevDevice = 0;
	cudaGetDevice(&prevDevice);
	cudaSetDevice(gpuIndex);

	size_t freeBytes = 0, totalBytes = 0;
	cudaMemGetInfo(&freeBytes, &totalBytes);

	// Restore previous device to avoid side-effects on caller's thread
	cudaSetDevice(prevDevice);

	if (freeBytes < minFreeBytes) {
		std::cout << "Warning [GPU]: GPU " << gpuIndex << " has only "
			<< (freeBytes / (1024 * 1024)) << " MiB free (need "
			<< (minFreeBytes / (1024 * 1024)) << " MiB). Task creation may fail." << std::endl;
		return false;
	}
	std::cout << "Info [GPU]: GPU " << gpuIndex << " has "
		<< (freeBytes / (1024 * 1024)) << " MiB free / "
		<< (totalBytes / (1024 * 1024)) << " MiB total" << std::endl;
	return true;
}

static std::unordered_map<ANSCENTER::ANSODBase*, int>& ODHandleRegistry() {
	static std::unordered_map<ANSCENTER::ANSODBase*, int> s;
	return s;
}
static std::mutex& ODHandleRegistryMutex() {
	static std::mutex m;
	return m;
}
static std::condition_variable& ODHandleRegistryCV() {
	static std::condition_variable cv;
	return cv;
}

static void RegisterODHandle(ANSCENTER::ANSODBase* h) {
	std::lock_guard<std::mutex> lk(ODHandleRegistryMutex());
	ODHandleRegistry()[h] = 1;  // refcount = 1
}

// Acquire a handle for use (increment refcount). Returns the handle
// if valid, nullptr if already released.
static ANSCENTER::ANSODBase* AcquireODHandle(ANSCENTER::ANSODBase* h) {
	std::lock_guard<std::mutex> lk(ODHandleRegistryMutex());
	auto it = ODHandleRegistry().find(h);
	if (it == ODHandleRegistry().end()) return nullptr;
	it->second++;
	return h;
}

// Release a use of the handle (decrement refcount).
// Returns true if this was the last reference (caller should destroy).
static bool ReleaseODHandleRef(ANSCENTER::ANSODBase* h) {
	std::lock_guard<std::mutex> lk(ODHandleRegistryMutex());
	auto it = ODHandleRegistry().find(h);
	if (it == ODHandleRegistry().end()) return false;
	it->second--;
	if (it->second <= 0) {
		ODHandleRegistry().erase(it);
		ODHandleRegistryCV().notify_all();
		return true;  // Caller should destroy
	}
	return false;
}

// Unregister and wait for all in-flight uses to finish.
// Decrements the creation refcount and blocks until refcount hits 0.
// Returns true if caller should destroy the object.
static bool UnregisterODHandle(ANSCENTER::ANSODBase* h) {
	std::unique_lock<std::mutex> lk(ODHandleRegistryMutex());
	auto it = ODHandleRegistry().find(h);
	if (it == ODHandleRegistry().end()) return false;
	it->second--;  // Remove creation ref
	// Wait for in-flight inferences to finish (30s timeout as safety net)
	bool ok = ODHandleRegistryCV().wait_for(lk, std::chrono::seconds(30), [&]() {
		auto it2 = ODHandleRegistry().find(h);
		return it2 == ODHandleRegistry().end() || it2->second <= 0;
	});
	if (!ok) {
		OutputDebugStringA("WARNING: UnregisterODHandle timed out waiting for in-flight inference\n");
	}
	ODHandleRegistry().erase(h);
	return true;  // Safe to destroy now
}

// RAII guard — ensures ReleaseODHandleRef is always called, preventing
// refcount leaks that would cause UnregisterODHandle to deadlock.
class ODHandleGuard {
	ANSCENTER::ANSODBase* engine;
public:
	explicit ODHandleGuard(ANSCENTER::ANSODBase* e) : engine(e) {}
	~ODHandleGuard() { if (engine) ReleaseODHandleRef(engine); }
	ANSCENTER::ANSODBase* get() const { return engine; }
	explicit operator bool() const { return engine != nullptr; }
	ODHandleGuard(const ODHandleGuard&) = delete;
	ODHandleGuard& operator=(const ODHandleGuard&) = delete;
};

BOOL APIENTRY DllMain( HMODULE hModule,
                       DWORD  ul_reason_for_call,
                       LPVOID lpReserved
                     )
{
    switch (ul_reason_for_call)
    {
    case DLL_PROCESS_ATTACH:
        // Pin the DLL so it is never unmapped while idle-timer or CUDA threads
        // are still running.  During LabVIEW shutdown the CLR/COM teardown can
        // unload DLLs before all threads exit → crash at unmapped code.
        {
            HMODULE hSelf = nullptr;
            GetModuleHandleExW(
                GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS |
                GET_MODULE_HANDLE_EX_FLAG_PIN,
                reinterpret_cast<LPCWSTR>(&DllMain),
                &hSelf);
        }
        break;
    case DLL_THREAD_ATTACH:
    case DLL_THREAD_DETACH:
        break;
    case DLL_PROCESS_DETACH:
        // When lpReserved != NULL, the process is terminating via ExitProcess.
        // The OS has already killed all worker threads (idle timers, CUDA
        // threads, etc.).  Calling ~Engine() → stopIdleTimer() → thread::join()
        // on a dead thread causes undefined behavior → std::terminate → abort.
        // Set the global flag so atexit destructors (which still run after
        // DllMain returns) skip thread joins and CUDA/TRT cleanup.
        if (lpReserved != nullptr) {
            g_processExiting().store(true, std::memory_order_relaxed);
            break;
        }

        // Dynamic FreeLibrary — threads are still alive, safe to clean up.
        // Without this, idle-timer threads keep the process alive indefinitely.
        try {
            std::vector<ANSCENTER::ANSODBase*> leakedHandles;
            {
                std::lock_guard<std::mutex> lk(ODHandleRegistryMutex());
                for (auto& [h, _] : ODHandleRegistry())
                    leakedHandles.push_back(h);
                ODHandleRegistry().clear();
            }
            for (auto* h : leakedHandles) {
                try { h->Destroy(); delete h; } catch (...) {}
            }
            try { EnginePoolManager<float>::instance().clearAll(); } catch (...) {}
            try { TRTEngineCache::instance().clearAll(); } catch (...) {}
        } catch (...) {}
        break;
    }
    return TRUE;
}

// CLASSIFICATION = 0,
//	DETECTION = 1,
//	SEGMENTATION = 2,
//	FACEDETECTOR = 3,
//	FACERECOGNIZER = 4,
//	LICENSEPLATE = 5,
//	TEXTSCENSE = 6
// External APIs
extern "C" ANSODENGINE_API std::string  CreateANSODHandle(ANSCENTER::ANSODBase** Handle,
	const char* licenseKey,
	const char* modelFilePath,
	const char* modelFileZipPassword,
	float detectionScoreThreshold,
	float modelConfThreshold,
	float modelMNSThreshold,
	int autoDetectEngine,//-1: CPU, 0: GPU; 1 auto detection
	int modelType,
	int detectionType,
	int loadEngineOnCreation)
{
	if (Handle == nullptr) return "";

	// NOTE: We intentionally do NOT destroy any existing *Handle here.
	// LabVIEW reuses DLL parameter buffer addresses, so *Handle may point
	// to ANOTHER task's live engine — not this task's old engine.
	// The caller must use ReleaseANSODHandle() explicitly to clean up.
	// (See ansod_debug.log analysis: Handle** addresses are recycled by
	// LabVIEW across different tasks, causing the old cleanup code to
	// destroy active engines belonging to other tasks.)

	std::string labelMap = "";
	bool _loadEngineOnCreation = false;
	if (loadEngineOnCreation == 1) {
		_loadEngineOnCreation = true;
	}
	else {
		_loadEngineOnCreation = false;
	}
	labelMap.clear();
	ANSCENTER::ModelConfig modelConfig;

	if (detectionScoreThreshold <= 0)modelConfig.detectionScoreThreshold = 0.5;
	else modelConfig.detectionScoreThreshold = detectionScoreThreshold;

	if (modelConfThreshold <= 0)modelConfig.modelConfThreshold = 0.5;
	else modelConfig.modelConfThreshold = modelConfThreshold;

	if (modelMNSThreshold <= 0)modelConfig.modelMNSThreshold = 0.45;
	else modelConfig.modelMNSThreshold = modelMNSThreshold;

	modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
	if (autoDetectEngine == 1)modelConfig.autoGPUDetection = true;
	else modelConfig.autoGPUDetection = false;

	ANSCENTER::EngineType engineType = ANSCENTER::ANSLicenseHelper::CheckHardwareInformation();
	if (autoDetectEngine==-1)engineType=ANSCENTER::EngineType::CPU;// We force to use CPU

	//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
		(modelType == 24))  // TensorRT Segmentation
	{
		if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) modelType = 31; // RTYOLO
		else modelType=30;// ONNXYOLO
	}
	else if ((modelType == 3) ||	// YoloV8/YoloV11 (Object Detection)
		     (modelType == 17)||	// YOLO V12
			 (modelType == 20) ||	// ONNX Classification	
			 (modelType == 21) ||	// ONNX Pose
			 (modelType == 23) ||	// ONNX Segmentation	
			 (modelType == 25))	    // OBB Segmentation
	{
		modelType = 30; // ONNXYOLO
	}
	else {
		// do nothing, use the modelType specified by user
	}

	switch (detectionType) {
		case 0:
			modelConfig.detectionType = ANSCENTER::DetectionType::CLASSIFICATION;
			break;
		case 1:
			modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
			break;
		case 2:
			modelConfig.detectionType = ANSCENTER::DetectionType::SEGMENTATION;
			break;
		case 3:
			modelConfig.detectionType = ANSCENTER::DetectionType::FACEDETECTOR;
			break;
		case 4:
			modelConfig.detectionType = ANSCENTER::DetectionType::FACERECOGNIZER;
			break;
		case 5:
			modelConfig.detectionType = ANSCENTER::DetectionType::LICENSEPLATE;
			break;
		case 6:
			modelConfig.detectionType = ANSCENTER::DetectionType::TEXTSCENSE;
			break;
		case 7:
			modelConfig.detectionType = ANSCENTER::DetectionType::KEYPOINT;
			break;
		default:
			modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
			break;
	}
	switch (modelType) {
		case 0: //TENSORFLOW =0
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::TENSORFLOW;
			break;
		case 1: //YOLOV4 = 1
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV4;
			break;
		case 2://YOLOV5 = 2
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV5;
			break;
		case 3: //YOLOV8 = 3,
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
			break;
		case 4: //TENSORRT = 4,
			if (modelConfig.detectionType == ANSCENTER::DetectionType::CLASSIFICATION) {
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTCL();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					(*Handle) = new ANSCENTER::ANSONNXCL();
					modelConfig.modelType = ANSCENTER::ModelType::ONNXCL;
				}
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::DETECTION) {
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTOD();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					(*Handle) = new ANSCENTER::YOLOOD();
					modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
				}
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::SEGMENTATION) {// Segmentation
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTSEG();
					modelConfig.modelType = ANSCENTER::ModelType::RTSEG;
				}
				else {
					(*Handle) = new ANSCENTER::ANSONNXSEG();
					modelConfig.modelType = ANSCENTER::ModelType::ONNXSEG;
				}
				break;
			}
			else {// default is detection
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTOD();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					(*Handle) = new ANSCENTER::YOLOOD();
					modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
				}
				break;
			}
		case 5: //OPENVINO = 5
			if (modelConfig.detectionType == ANSCENTER::DetectionType::CLASSIFICATION) {
				(*Handle) = new ANSCENTER::OPENVINOCL();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::DETECTION) {
				(*Handle) = new ANSCENTER::OPENVINOOD();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::SEGMENTATION) {// Segmentation
				(*Handle) = new ANSCENTER::ANSOVSEG();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
			else {
				(*Handle) = new ANSCENTER::OPENVINOOD();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
		case 6: //FACEDETECT = 6
			(*Handle) = new ANSCENTER::ANSFD();
			modelConfig.modelType = ANSCENTER::ModelType::FACEDETECT;
			break;
		case 10: //ANOMALIB=10
			(*Handle) = new ANSCENTER::ANSANOMALIB();
			modelConfig.modelType = ANSCENTER::ModelType::ANOMALIB;
			break;
		case 11: //OPENPOSE=11
			(*Handle) = new ANSCENTER::ANSPOSE();
			modelConfig.modelType = ANSCENTER::ModelType::POSE;
			break;
		case 12: //SAM=12
			(*Handle) = new ANSCENTER::ANSSAM();
			modelConfig.modelType = ANSCENTER::ModelType::SAM;
			break;
		case 13: //ODHUBMODEL=13
			(*Handle) = new ANSCENTER::ODHUBAPI();
			modelConfig.modelType = ANSCENTER::ModelType::ODHUBMODEL;
			break;
		case 14: //TensorRT for Object Detection Yolov10
			(*Handle) = new ANSCENTER::ANSYOLOV10RTOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV10RTOD;
			break;
		case 15: //OpenVino for Object Detection Yolov10
			(*Handle) = new ANSCENTER::ANSOYOLOV10OVOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV10OVOD;
			break;
		case 16: //Custom detector
			(*Handle) = new ANSCENTER::ANSCUSTOMDETECTOR();
			modelConfig.modelType = ANSCENTER::ModelType::CUSTOMDETECTOR;
			break;
		case 17: //Yolo V12
			(*Handle) = new ANSCENTER::YOLO12OD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV12;
			break;
		case 18: //Custom script model
			(*Handle) = new ANSCENTER::ANSCUSTOMPY();
			modelConfig.modelType = ANSCENTER::ModelType::CUSTOMPY;
			break;
		case	19: //Motion Detector
			(*Handle) = new ANSCENTER::ANSMOTIONDETECTOR();
			modelConfig.modelType = ANSCENTER::ModelType::MOTIONDETECTOR;
			break;
		case 20: //ONNXCL
			(*Handle) = new ANSCENTER::ANSONNXCL();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXCL;
			break;
		case 21: //ONNXPOSE
			(*Handle) = new ANSCENTER::ANSONNXPOSE();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXPOSE;
			break;
		case 22: //TENSORRTPOSE
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::ANSTENSORRTPOSE();
				modelConfig.modelType = ANSCENTER::ModelType::RTPOSE;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXPOSE();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXPOSE;
			}
			break;
		case 23: //ONNXSEG
			(*Handle) = new ANSCENTER::ANSONNXSEG();	
			modelConfig.modelType = ANSCENTER::ModelType::ONNXSEG;
			break;
		case 24: //RTSEG
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::TENSORRTSEG();
				modelConfig.modelType = ANSCENTER::ModelType::RTSEG;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXSEG();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXSEG;
			}
			break;
		case 25: //ONNXOBB
			(*Handle) = new ANSCENTER::ANSONNXOBB();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXOBB;
			break;

		//case 26: //RTOBB
		//	(*Handle) = new ANSCENTER::ANSTENSORRTOBB();
		//	modelConfig.modelType = ANSCENTER::ModelType::RTOBB;
		//	break;
		case 27: //MOVIENET
				(*Handle) = new ANSCENTER::ANSMOVIENET();
				modelConfig.modelType = ANSCENTER::ModelType::MOVIENET;
				break;
		case 28: //ONNXSAM3
				(*Handle) = new ANSCENTER::ANSONNXSAM3();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXSAM3;
				break;
		case 29: //RTSAM3
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::ANSSAM3();
					modelConfig.modelType = ANSCENTER::ModelType::RTSAM3;
				}
				else {
					(*Handle) = new ANSCENTER::ANSONNXSAM3();
					modelConfig.modelType = ANSCENTER::ModelType::ONNXSAM3;
				}
				break;
		case 30: //ONNXYOLO
			(*Handle) = new ANSCENTER::ANSONNXYOLO();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXYOLO;
			break;
		case 31: //RTYOLO
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::ANSRTYOLO();
				modelConfig.modelType = ANSCENTER::ModelType::RTYOLO;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXYOLO();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXYOLO;
			}
			break;
		default:
			(*Handle) = new ANSCENTER::ANSFD();
			modelConfig.modelType = ANSCENTER::ModelType::FACEDETECT;
			break;
	}
	if (*Handle == nullptr) {
		return labelMap;
	}
	else {
		// CUDA round-robin + VRAM check — only relevant for NVIDIA GPUs.
		// On AMD/DirectML and OpenVINO these calls hit stub CUDA APIs that
		// return bogus 0-byte VRAM and pollute the log with false warnings.
		if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
			const int assignedGPU = AssignNextGPU();
			modelConfig.gpuDeviceIndex = assignedGPU;
			CheckGPUVRAM(assignedGPU);
			(*Handle)->SetMaxSlotsPerGpu(GetPoolMaxSlotsPerGpu());
		}

		RegisterODHandle(*Handle);
		(*Handle)->SetLoadEngineOnCreation(_loadEngineOnCreation); //Set force to load the engine immediately
		bool loadResult = (*Handle)->Initialize(licenseKey, modelConfig, modelFilePath, modelFileZipPassword, labelMap);
		return labelMap;
	}
}
extern "C" ANSODENGINE_API int  CreateANSODHandleEx(ANSCENTER::ANSODBase** Handle,
															const char* licenseKey,
															const char* modelFilePath,
															const char* modelFileZipPassword,
															float detectionScoreThreshold,
															float modelConfThreshold,
															float modelMNSThreshold,
															int autoDetectEngine,
															int modelType,
															int detectionType,
															std::string& labelMap,
															int loadEngineOnCreation)
{
	if (Handle == nullptr) return -1; // invalid modelType return

	bool _loadEngineOnCreation = false;
	if (loadEngineOnCreation == 1) {
		_loadEngineOnCreation = true;
	}
	else {
		_loadEngineOnCreation = false;
	}
	labelMap.clear();
	ANSCENTER::ModelConfig modelConfig;

	if (detectionScoreThreshold <= 0)modelConfig.detectionScoreThreshold = 0.5;
	else modelConfig.detectionScoreThreshold = detectionScoreThreshold;

	if (modelConfThreshold <= 0)modelConfig.modelConfThreshold = 0.5;
	else modelConfig.modelConfThreshold = modelConfThreshold;

	if (modelMNSThreshold <= 0)modelConfig.modelMNSThreshold = 0.45;
	else modelConfig.modelMNSThreshold = modelMNSThreshold;

	modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
	if (autoDetectEngine == 1)modelConfig.autoGPUDetection = true;
	else modelConfig.autoGPUDetection = false;

	ANSCENTER::EngineType engineType = ANSCENTER::ANSLicenseHelper::CheckHardwareInformation();
	if (autoDetectEngine==-1)engineType=ANSCENTER::EngineType::CPU;// We force to use CPU

	//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
		(modelType == 24))  // TensorRT Segmentation
	{
		if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) modelType = 31; // RTYOLO
		else modelType=30;// ONNXYOLO
	}
	else if ((modelType == 3) ||	// YoloV8/YoloV11 (Object Detection)
		     (modelType == 17)||	// YOLO V12
			 (modelType == 20) ||	// ONNX Classification
			 (modelType == 21) ||	// ONNX Pose
			 (modelType == 23) ||	// ONNX Segmentation
			 (modelType == 25))	    // OBB Segmentation
	{
		modelType = 30; // ONNXYOLO
	}
	else {
		// do nothing, use the modelType specified by user
	}
	// returnModelType will be set after the switch to reflect the actual
	// model class that was instantiated (e.g. RTYOLO→ONNXYOLO on AMD).
	int returnModelType = modelType;

	switch (detectionType) {
		case 0:
			modelConfig.detectionType = ANSCENTER::DetectionType::CLASSIFICATION;
			break;
		case 1:
			modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
			break;
		case 2:
			modelConfig.detectionType = ANSCENTER::DetectionType::SEGMENTATION;
			break;
		case 3:
			modelConfig.detectionType = ANSCENTER::DetectionType::FACEDETECTOR;
			break;
		case 4:
			modelConfig.detectionType = ANSCENTER::DetectionType::FACERECOGNIZER;
			break;
		case 5:
			modelConfig.detectionType = ANSCENTER::DetectionType::LICENSEPLATE;
			break;
		case 6:
			modelConfig.detectionType = ANSCENTER::DetectionType::TEXTSCENSE;
			break;
		case 7:
			modelConfig.detectionType = ANSCENTER::DetectionType::KEYPOINT;
			break;
		default:
			modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
			break;
	}
	switch (modelType) {
		case 0: //TENSORFLOW =0
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::TENSORFLOW;
			break;
		case 1: //YOLOV4 = 1
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV4;
			break;
		case 2://YOLOV5 = 2
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV5;
			break;
		case 3: //YOLOV8 = 3,
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
			break;
		case 4: //TENSORRT = 4,
			if (modelConfig.detectionType == ANSCENTER::DetectionType::CLASSIFICATION) {
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTCL();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					(*Handle) = new ANSCENTER::ANSONNXCL();
					modelConfig.modelType = ANSCENTER::ModelType::ONNXCL;
				}
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::DETECTION) {
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTOD();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					(*Handle) = new ANSCENTER::YOLOOD();
					modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
				}
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::SEGMENTATION) {// Segmentation
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTSEG();
					modelConfig.modelType = ANSCENTER::ModelType::RTSEG;
				}
				else {
					(*Handle) = new ANSCENTER::ANSONNXSEG();
					modelConfig.modelType = ANSCENTER::ModelType::ONNXSEG;
				}
				break;
			}
			else {// default is detection
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTOD();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					(*Handle) = new ANSCENTER::YOLOOD();
					modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
				}
				break;
			}
		case 5: //OPENVINO = 5
			if (modelConfig.detectionType == ANSCENTER::DetectionType::CLASSIFICATION) {
				(*Handle) = new ANSCENTER::OPENVINOCL();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::DETECTION) {
				(*Handle) = new ANSCENTER::OPENVINOOD();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::SEGMENTATION) {// Segmentation
				(*Handle) = new ANSCENTER::ANSOVSEG();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
			else {
				(*Handle) = new ANSCENTER::OPENVINOOD();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
		case 6: //FACEDETECT = 6
			(*Handle) = new ANSCENTER::ANSFD();
			modelConfig.modelType = ANSCENTER::ModelType::FACEDETECT;
			break;
		case 10: //ANOMALIB=10
			(*Handle) = new ANSCENTER::ANSANOMALIB();
			modelConfig.modelType = ANSCENTER::ModelType::ANOMALIB;
			break;
		case 11: //OPENPOSE=11
			(*Handle) = new ANSCENTER::ANSPOSE();
			modelConfig.modelType = ANSCENTER::ModelType::POSE;
			break;
		case 12: //SAM=12
			(*Handle) = new ANSCENTER::ANSSAM();
			modelConfig.modelType = ANSCENTER::ModelType::SAM;
			break;
		case 13: //ODHUBMODEL=13
			(*Handle) = new ANSCENTER::ODHUBAPI();
			modelConfig.modelType = ANSCENTER::ModelType::ODHUBMODEL;
			break;
		case 14: //TensorRT for Object Detection Yolov10
			(*Handle) = new ANSCENTER::ANSYOLOV10RTOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV10RTOD;
			break;
		case 15: //OpenVino for Object Detection Yolov10
			(*Handle) = new ANSCENTER::ANSOYOLOV10OVOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV10OVOD;
			break;
		case 16: //Custom detector
			(*Handle) = new ANSCENTER::ANSCUSTOMDETECTOR();
			modelConfig.modelType = ANSCENTER::ModelType::CUSTOMDETECTOR;
			break;
		case 17: //Yolo V12
			(*Handle) = new ANSCENTER::YOLO12OD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV12;
			break;
		case 18: //Custom script model
			(*Handle) = new ANSCENTER::ANSCUSTOMPY();
			modelConfig.modelType = ANSCENTER::ModelType::CUSTOMPY;
			break;
		case	19: //Motion Detector
			(*Handle) = new ANSCENTER::ANSMOTIONDETECTOR();
			modelConfig.modelType = ANSCENTER::ModelType::MOTIONDETECTOR;
			break;
		case 20: //ONNXCL
			(*Handle) = new ANSCENTER::ANSONNXCL();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXCL;
			break;
		case 21: //ONNXPOSE
			(*Handle) = new ANSCENTER::ANSONNXPOSE();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXPOSE;
			break;
		case 22: //TENSORRTPOSE
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::ANSTENSORRTPOSE();
				modelConfig.modelType = ANSCENTER::ModelType::RTPOSE;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXPOSE();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXPOSE;
			}
			break;
		case 23: //ONNXSEG
			(*Handle) = new ANSCENTER::ANSONNXSEG();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXSEG;
			break;
		case 24: //RTSEG
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::TENSORRTSEG();
				modelConfig.modelType = ANSCENTER::ModelType::RTSEG;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXSEG();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXSEG;
			}
			break;
		case 25: //ONNXOBB
			(*Handle) = new ANSCENTER::ANSONNXOBB();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXOBB;
			break;

		//case 26: //RTOBB
		//	(*Handle) = new ANSCENTER::ANSTENSORRTOBB();
		//	modelConfig.modelType = ANSCENTER::ModelType::RTOBB;
		//	break;
		case 27: //MOVIENET
				(*Handle) = new ANSCENTER::ANSMOVIENET();
				modelConfig.modelType = ANSCENTER::ModelType::MOVIENET;
				break;
		case 28: //ONNXSAM3
				(*Handle) = new ANSCENTER::ANSONNXSAM3();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXSAM3;
				break;
		case 29: //RTSAM3
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::ANSSAM3();
					modelConfig.modelType = ANSCENTER::ModelType::RTSAM3;
				}
				else {
					(*Handle) = new ANSCENTER::ANSONNXSAM3();
					modelConfig.modelType = ANSCENTER::ModelType::ONNXSAM3;
				}
				break;
		case 30: //ONNXYOLO
			(*Handle) = new ANSCENTER::ANSONNXYOLO();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXYOLO;
			break;
		case 31: //RTYOLO
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::ANSRTYOLO();
				modelConfig.modelType = ANSCENTER::ModelType::RTYOLO;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXYOLO();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXYOLO;
			}
			break;
		default:
			(*Handle) = new ANSCENTER::ANSFD();
			modelConfig.modelType = ANSCENTER::ModelType::FACEDETECT;
			break;
	}
	// Update returnModelType to reflect the actual class that was created.
	// The switch may have fallen back (e.g. RTYOLO→ONNXYOLO on non-NVIDIA).
	returnModelType = static_cast<int>(modelConfig.modelType);

	if (*Handle == nullptr) {
		labelMap ="";
		return returnModelType;
	}
	else {
		// CUDA round-robin + VRAM check — only relevant for NVIDIA GPUs.
		// On AMD/DirectML and OpenVINO these calls hit stub CUDA APIs that
		// return bogus 0-byte VRAM and pollute the log with false warnings.
		if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
			const int assignedGPU = AssignNextGPU();
			modelConfig.gpuDeviceIndex = assignedGPU;
			CheckGPUVRAM(assignedGPU);
			(*Handle)->SetMaxSlotsPerGpu(GetPoolMaxSlotsPerGpu());
		}

		RegisterODHandle(*Handle);
		(*Handle)->SetLoadEngineOnCreation(_loadEngineOnCreation); //Set force to load the engine immediately
		bool loadResult = (*Handle)->Initialize(licenseKey, modelConfig, modelFilePath, modelFileZipPassword, labelMap);
		return returnModelType;
	}

}
//// For LabVIEW API
extern "C" ANSODENGINE_API int			CreateANSODHandle_LV(ANSCENTER::ANSODBase** Handle, const char* licenseKey, const char* modelFilePath, const char* modelFileZipPassword, float modelThreshold, float modelConfThreshold, float modelNMSThreshold, int autoDetectEngine, int modelType, int detectorType, int loadEngineOnCreation, LStrHandle labelMap) {
	try {

		std::string lbMap;

		int returnModelType = CreateANSODHandleEx(Handle, licenseKey, modelFilePath, modelFileZipPassword, modelThreshold, modelConfThreshold, modelNMSThreshold, autoDetectEngine, modelType, detectorType, lbMap, loadEngineOnCreation);

		// CreateANSODHandleEx returns -1 only when Handle is nullptr.
		// Check that instead of lbMap.empty() — labelMap can be legitimately
		// empty when loadEngineOnCreation==0 or the model has no class file.
		if (returnModelType < 0 || Handle == nullptr || *Handle == nullptr) return -1;

		int size = static_cast<int>(lbMap.length());
		if (size > 0) {
			MgErr error = DSSetHandleSize(labelMap, sizeof(int32) + size * sizeof(uChar));
			if (error == noErr) {
				(*labelMap)->cnt = size;
				memcpy((*labelMap)->str, lbMap.c_str(), size);
			}
			else return -1;
		}
		else {
			// Empty label map — set LabVIEW string to empty
			MgErr error = DSSetHandleSize(labelMap, sizeof(int32));
			if (error == noErr) (*labelMap)->cnt = 0;
		}
		return returnModelType;
	}
	catch (...) {
		return -1;
	}
}

extern "C" __declspec(dllexport) int LoadModelFromFolder(ANSCENTER::ANSODBase** Handle, const char* licenseKey,
	const char* modelName,
	const char* className,
	float detectionScoreThreshold,
	float modelConfThreshold,
	float modelMNSThreshold,
	int autoDetectEngine,
	int modelType,
	int detectionType,
	int loadEngineOnCreation,
	const char* modelFolder,
	std::string& labelMap)
{
	try
	{
		if (Handle == nullptr) return 0;
		labelMap.clear();
		ANSCENTER::ModelConfig modelConfig;
		bool _loadEngineOnCreation = false;
		if (loadEngineOnCreation == 1) {
			_loadEngineOnCreation = true;
		}
		else {
			_loadEngineOnCreation = false;
		}
		if (detectionScoreThreshold <= 0)modelConfig.detectionScoreThreshold = 0.5;
		else modelConfig.detectionScoreThreshold = detectionScoreThreshold;

		if (modelConfThreshold <= 0)modelConfig.modelConfThreshold = 0.5;
		else modelConfig.modelConfThreshold = modelConfThreshold;

		if (modelMNSThreshold <= 0)modelConfig.modelMNSThreshold = 0.45;
		else modelConfig.modelMNSThreshold = modelMNSThreshold;

		ANSCENTER::EngineType engineType = ANSCENTER::ANSLicenseHelper::CheckHardwareInformation();

		modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
		if (autoDetectEngine == 1)modelConfig.autoGPUDetection = true;
		else modelConfig.autoGPUDetection = false;

		if (autoDetectEngine==-1)engineType=ANSCENTER::EngineType::CPU;// We force to use CPU


		//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
			(modelType == 24))  // TensorRT Segmentation
		{
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU)modelType = 31; // RTYOLO
			else modelType = 30;// ONNXYOLO
		}
		else if ((modelType == 3) ||	// YoloV8/YoloV11 (Object Detection)
			(modelType == 17) ||	// YOLO V12
			(modelType == 20) ||	// ONNX Classification	
			(modelType == 21) ||	// ONNX Pose
			(modelType == 23) ||	// ONNX Segmentation	
			(modelType == 25))	    // OBB Segmentation
		{
			modelType = 30; // ONNXYOLO
		}
		else {
			// do nothing, use the modelType specified by user
		}
		// NOTE: We intentionally do NOT destroy any existing *Handle here.
		// LabVIEW reuses DLL parameter buffer addresses, so *Handle may point
		// to ANOTHER task's live engine — not this task's old engine.
		// The caller must use ReleaseANSODHandle() explicitly to clean up.

		switch (detectionType) {
		case 0:
			modelConfig.detectionType = ANSCENTER::DetectionType::CLASSIFICATION;
			break;
		case 1:
			modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
			break;
		case 2:
			modelConfig.detectionType = ANSCENTER::DetectionType::SEGMENTATION;
			break;
		case 3:
			modelConfig.detectionType = ANSCENTER::DetectionType::FACEDETECTOR;
			break;
		case 4:
			modelConfig.detectionType = ANSCENTER::DetectionType::FACERECOGNIZER;
			break;
		case 5:
			modelConfig.detectionType = ANSCENTER::DetectionType::LICENSEPLATE;
			break;
		case 6:
			modelConfig.detectionType = ANSCENTER::DetectionType::TEXTSCENSE;
			break;
		case 7:
			modelConfig.detectionType = ANSCENTER::DetectionType::KEYPOINT;
			break;
		default:
			modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
			break;
		}
		switch (modelType) {
		case 0: //TENSORFLOW =0
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::TENSORFLOW;
			break;
		case 1: //YOLOV4 = 1
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV4;
			break;
		case 2://YOLOV5 = 2
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV5;
			break;
		case 3: //YOLOV8 = 3,
			(*Handle) = new ANSCENTER::YOLOOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
			break;
		case 4: //TENSORRT = 4,
			if (modelConfig.detectionType == ANSCENTER::DetectionType::CLASSIFICATION) {
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTCL();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					(*Handle) = new ANSCENTER::ANSONNXCL();
					modelConfig.modelType = ANSCENTER::ModelType::ONNXCL;
				}
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::DETECTION) {
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTOD();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					(*Handle) = new ANSCENTER::YOLOOD();
					modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
				}

				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::SEGMENTATION) {// Segmentation
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTSEG();
					modelConfig.modelType = ANSCENTER::ModelType::RTSEG;
				}
				break;
			}
			else {// default is detection
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					(*Handle) = new ANSCENTER::TENSORRTOD();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					(*Handle) = new ANSCENTER::YOLOOD();
					modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
				}
				break;
			}
		case 5: //OPENVINO = 5
			if (modelConfig.detectionType == ANSCENTER::DetectionType::CLASSIFICATION) {
				(*Handle) = new ANSCENTER::OPENVINOCL();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::DETECTION) {
				(*Handle) = new ANSCENTER::OPENVINOOD();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
			else if (modelConfig.detectionType == ANSCENTER::DetectionType::SEGMENTATION) {// Segmentation
				(*Handle) = new ANSCENTER::ANSOVSEG();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}
			else {
				(*Handle) = new ANSCENTER::OPENVINOOD();
				modelConfig.modelType = ANSCENTER::ModelType::OPENVINO;
				break;
			}

		case 6: //FACEDETECT = 6
			(*Handle) = new ANSCENTER::ANSFD();
			modelConfig.modelType = ANSCENTER::ModelType::FACEDETECT;
			break;
		case 10: //ANOMALIB=10
			(*Handle) = new ANSCENTER::ANSANOMALIB();
			modelConfig.modelType = ANSCENTER::ModelType::ANOMALIB;
			break;
		case 11: //OPENPOSE=11
			(*Handle) = new ANSCENTER::ANSPOSE();
			modelConfig.modelType = ANSCENTER::ModelType::POSE;
			break;
		case 12: //SAM=12
			(*Handle) = new ANSCENTER::ANSSAM();
			modelConfig.modelType = ANSCENTER::ModelType::SAM;
			break;
		case 13: //ODHUBMODEL=13
			(*Handle) = new ANSCENTER::ODHUBAPI();
			modelConfig.modelType = ANSCENTER::ModelType::ODHUBMODEL;
			break;
		case 14: //TensorRT for Object Detection Yolov10
			(*Handle) = new ANSCENTER::ANSYOLOV10RTOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV10RTOD;
			break;
		case 15: //OpenVino for Object Detection Yolov10
			(*Handle) = new ANSCENTER::ANSOYOLOV10OVOD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV10OVOD;
			break;
		case 16: //Custom detector
			(*Handle) = new ANSCENTER::ANSCUSTOMDETECTOR();
			modelConfig.modelType = ANSCENTER::ModelType::CUSTOMDETECTOR;
			break;
		case 17: //Yolo V12
			(*Handle) = new ANSCENTER::YOLO12OD();
			modelConfig.modelType = ANSCENTER::ModelType::YOLOV12;
			break;
		case 18: //Custom Python
			(*Handle) = new ANSCENTER::ANSCUSTOMPY();
			modelConfig.modelType = ANSCENTER::ModelType::CUSTOMPY;
			break;
		case 19: //Motion Detector
			(*Handle) = new ANSCENTER::ANSMOTIONDETECTOR();
			modelConfig.modelType = ANSCENTER::ModelType::MOTIONDETECTOR;
			break;
		case 20: //ANSONNXCL
			(*Handle) = new ANSCENTER::ANSONNXCL();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXCL;
			break;
		case 21: //ANSONNXPOSE
			(*Handle) = new ANSCENTER::ANSONNXPOSE();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXPOSE;
			break;
		case 22: //ANSTENSORRTPOSE
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::ANSTENSORRTPOSE();
				modelConfig.modelType = ANSCENTER::ModelType::RTPOSE;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXPOSE();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXPOSE;
			}
			break;
		case 23: //ONNXSEG
			(*Handle) = new ANSCENTER::ANSONNXSEG();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXSEG;
			break;
		case 24: //RTSEG
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::TENSORRTSEG();
				modelConfig.modelType = ANSCENTER::ModelType::RTSEG;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXSEG();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXSEG;
			}
			break;
		case 25: //ONNXOBB
			(*Handle) = new ANSCENTER::ANSONNXOBB();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXOBB;
			break;
		case 27: //MOVIENET
			(*Handle) = new ANSCENTER::ANSMOVIENET();
			modelConfig.modelType = ANSCENTER::ModelType::MOVIENET;
			break;
		case 28: //ANSONNXSAM3
			(*Handle) = new ANSCENTER::ANSONNXSAM3();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXSAM3;
			break;
		case 29: //ANSSAM3
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::ANSSAM3();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXSAM3;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXSAM3();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXSAM3;
			}
			break;
		case 30: //ONNXYOLO
			(*Handle) = new ANSCENTER::ANSONNXYOLO();
			modelConfig.modelType = ANSCENTER::ModelType::ONNXYOLO;
			break;
		case 31: //RTYOLO
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				(*Handle) = new ANSCENTER::ANSRTYOLO();
				modelConfig.modelType = ANSCENTER::ModelType::RTYOLO;
			}
			else {
				(*Handle) = new ANSCENTER::ANSONNXYOLO();
				modelConfig.modelType = ANSCENTER::ModelType::ONNXYOLO;
			}
			break;
		default:
			(*Handle) = new ANSCENTER::ANSFD();
			modelConfig.modelType = ANSCENTER::ModelType::FACEDETECT;
			break;
		}
		if (*Handle == nullptr) {
			return -1;
		}
		else {
			// CUDA round-robin + VRAM check — NVIDIA only (see CreateANSODHandle).
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				const int assignedGPU = AssignNextGPU();
				modelConfig.gpuDeviceIndex = assignedGPU;
				CheckGPUVRAM(assignedGPU);
				(*Handle)->SetMaxSlotsPerGpu(GetPoolMaxSlotsPerGpu());
			}

			RegisterODHandle(*Handle);
			(*Handle)->SetLoadEngineOnCreation(_loadEngineOnCreation); //Set force to load the engine immediately
			bool result = (*Handle)->LoadModelFromFolder(licenseKey, modelConfig, modelName, className, modelFolder, labelMap);
			if (result) return 1;
			else return 0;
		}
	}
	catch (...) {
		return 0;
	}
}

ANSODENGINE_API int OptimizeModelStr(const char* modelFilePath, const char* modelFileZipPassword, int modelType, int modelDetectionType, int fp16, std::string& modelFolder) {
	try {
		bool optimizedResult = false;
		// NOTE: odMutex was removed here. OptimizeModelStr creates its own
		// temporary Engine<float> on the stack — no shared state with running
		// inference tasks. The global mutex was blocking ALL running tasks'
		// result delivery for the entire duration of TRT engine building.
		ANSCENTER::EngineType engineType = ANSCENTER::ANSLicenseHelper::CheckHardwareInformation();


		//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
			(modelType == 24))  // TensorRT Segmentation
		{
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU)modelType = 31; // RTYOLO
			else modelType = 30;// ONNXYOLO
		}
		else if ((modelType == 3) ||	// YoloV8/YoloV11 (Object Detection)
			(modelType == 17) ||	// YOLO V12
			(modelType == 20) ||	// ONNX Classification
			(modelType == 21) ||	// ONNX Pose
			(modelType == 23) ||	// ONNX Segmentation
			(modelType == 25))	    // OBB Segmentation
		{
			modelType = 30; // ONNXYOLO
		}
		else {
			// do nothing, use the modelType specified by user
		}

		
		if (modelType == 31) // If modelType is RTYOLO (31), handle separately.
		{
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				if (modelDetectionType == 0) {
					return ANSCENTER::ANSUtilityHelper::ModelOptimizer(modelFilePath, modelFileZipPassword, fp16, modelFolder, 224, 244) ? 1 : 0; // this is for classification models
				}
				else {
					return ANSCENTER::ANSUtilityHelper::ModelOptimizer(modelFilePath, modelFileZipPassword, fp16, modelFolder, 640, 640) ? 1 : 0; //  standard size for detection models, segmetation and others
				}
			}
		}
		// Create model handle dynamically
		std::unique_ptr<ANSCENTER::ANSODBase> Handle;
		ANSCENTER::ModelConfig modelConfig;
		bool _fp16 = (fp16 == 1);
		switch (modelType) {
		case 0:  Handle = std::make_unique<ANSCENTER::YOLOOD>(); modelConfig.modelType = ANSCENTER::ModelType::TENSORFLOW; break;
		case 1:  Handle = std::make_unique<ANSCENTER::YOLOOD>(); modelConfig.modelType = ANSCENTER::ModelType::YOLOV4; break;
		case 2:  Handle = std::make_unique<ANSCENTER::YOLOOD>(); modelConfig.modelType = ANSCENTER::ModelType::YOLOV5; break;
		case 5:  Handle = std::make_unique<ANSCENTER::OPENVINOOD>(); modelConfig.modelType = ANSCENTER::ModelType::OPENVINO; break;
		case 6:  Handle = std::make_unique<ANSCENTER::ANSFD>(); modelConfig.modelType = ANSCENTER::ModelType::FACEDETECT; break;
		case 10: Handle = std::make_unique<ANSCENTER::ANSANOMALIB>(); modelConfig.modelType = ANSCENTER::ModelType::ANOMALIB; break;
		case 11: Handle = std::make_unique<ANSCENTER::ANSPOSE>(); modelConfig.modelType = ANSCENTER::ModelType::POSE; break;
		case 12: Handle = std::make_unique<ANSCENTER::ANSSAM>(); modelConfig.modelType = ANSCENTER::ModelType::SAM; break;
		case 13: Handle = std::make_unique<ANSCENTER::ODHUBAPI>(); modelConfig.modelType = ANSCENTER::ModelType::ODHUBMODEL; break;
		case 15: Handle = std::make_unique<ANSCENTER::ANSOYOLOV10OVOD>(); modelConfig.modelType = ANSCENTER::ModelType::YOLOV10OVOD; break;
		case 16: Handle = std::make_unique<ANSCENTER::ANSCUSTOMDETECTOR>(); modelConfig.modelType = ANSCENTER::ModelType::CUSTOMDETECTOR; break;
		case 18: Handle = std::make_unique<ANSCENTER::ANSCUSTOMPY>(); modelConfig.modelType = ANSCENTER::ModelType::CUSTOMPY; break;
		case 19: Handle = std::make_unique<ANSCENTER::ANSMOTIONDETECTOR>(); modelConfig.modelType = ANSCENTER::ModelType::MOTIONDETECTOR; break;
		case 27: Handle = std::make_unique<ANSCENTER::ANSMOVIENET>(); modelConfig.modelType = ANSCENTER::ModelType::MOVIENET; break;
		case 28: Handle = std::make_unique<ANSCENTER::ANSONNXSAM3>(); modelConfig.modelType = ANSCENTER::ModelType::ONNXSAM3; break;
		case 29: {
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				Handle = std::make_unique<ANSCENTER::ANSSAM3>(); modelConfig.modelType = ANSCENTER::ModelType::RTSAM3;
			}
			else {
				Handle = std::make_unique<ANSCENTER::ANSONNXSAM3>(); modelConfig.modelType = ANSCENTER::ModelType::ONNXSAM3;
			}
			break;
		}
		case 30: Handle = std::make_unique<ANSCENTER::ANSONNXYOLO>(); modelConfig.modelType = ANSCENTER::ModelType::ONNXYOLO; break;
		case 31: {
			if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
				Handle = std::make_unique<ANSCENTER::ANSRTYOLO>(); modelConfig.modelType = ANSCENTER::ModelType::RTYOLO;
			}
			else {
				Handle = std::make_unique<ANSCENTER::ANSONNXYOLO>(); modelConfig.modelType = ANSCENTER::ModelType::ONNXYOLO;
			}
			break;
		}

		default: {
			if (modelDetectionType == 0) // classification
			{
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					Handle = std::make_unique<ANSCENTER::TENSORRTCL>();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					Handle = std::make_unique<ANSCENTER::ANSONNXCL>();
					modelConfig.modelType = ANSCENTER::ModelType::ONNXCL;
				}
				modelConfig.detectionType = ANSCENTER::DetectionType::CLASSIFICATION;

			}
			else if (modelDetectionType == 1) // detection
			{
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					Handle = std::make_unique<ANSCENTER::TENSORRTOD>();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					Handle = std::make_unique<ANSCENTER::YOLOOD>();
					modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
				}
				modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
			}
			else if (modelDetectionType == 2) // segmentation
			{
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU)
				{
					Handle = std::make_unique<ANSCENTER::TENSORRTSEG>();
					modelConfig.modelType = ANSCENTER::ModelType::RTSEG;
				}
				else {
					Handle = std::make_unique<ANSCENTER::ANSONNXSEG>();
					modelConfig.modelType = ANSCENTER::ModelType::ONNXSEG;

				}
				modelConfig.detectionType = ANSCENTER::DetectionType::SEGMENTATION;

			}
			else // default is detection
			{
				if (engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
					Handle = std::make_unique<ANSCENTER::TENSORRTOD>();
					modelConfig.modelType = ANSCENTER::ModelType::TENSORRT;
				}
				else {
					Handle = std::make_unique<ANSCENTER::YOLOOD>();
					modelConfig.modelType = ANSCENTER::ModelType::YOLOV8;
				}
				modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
			}
			break;
		}
		}
		// TensorRT-specific: bypass pool and cache for temporary optimizer engines
		if (Handle && engineType == ANSCENTER::EngineType::NVIDIA_GPU) {
			Handle->SetMaxSlotsPerGpu(0);
			Handle->SetSkipEngineCache(true);
			Handle->SetForceNoPool(true);
		}

		// RAII guard for TensorRT process-wide flags.
		// Without this, an exception in LoadModel or OptimizeModel permanently
		// leaves the flags set, breaking all subsequent engine creation.
		struct GlobalFlagGuard {
			bool active;
			GlobalFlagGuard(bool isNvidia) : active(isNvidia) {
				if (active) {
					g_forceNoPool = true;
					TRTEngineCache::globalBypass() = true;
				}
			}
			~GlobalFlagGuard() {
				if (active) {
					g_forceNoPool = false;
					TRTEngineCache::globalBypass() = false;
				}
			}
		} flagGuard(engineType == ANSCENTER::EngineType::NVIDIA_GPU);

		// Load and optimize model
		if (Handle && Handle->LoadModel(modelFilePath, modelFileZipPassword)) {
			optimizedResult = Handle->OptimizeModel(_fp16, modelFolder);
		}
		Handle.reset();  // Destroy engines BEFORE guard restores cache


		if (optimizedResult && !modelFolder.empty()) return 1;
		else return 0;
	}
	catch (const std::exception& e) {
		// GlobalFlagGuard destructor runs here — flags are always restored
		std::cerr << "OptimizeModelStr Exception: " << e.what() << std::endl;
		return -1;
	}
	catch (...) {
		// GlobalFlagGuard destructor runs here — flags are always restored
		std::cerr << "OptimizeModelStr: Unknown exception occurred." << std::endl;
		return -1;
	}
}
static int ReleaseANSODHandle_Impl(ANSCENTER::ANSODBase** Handle) {
	try {
		if (Handle == nullptr) return 0;
		if (*Handle == nullptr) return 0;

		// Only release if this handle was registered by us
		if (!UnregisterODHandle(*Handle)) {
			// Not in registry — already freed or not ours
			*Handle = nullptr;
			return 0;
		}
		(*Handle)->Destroy();
		delete *Handle;
		*Handle = nullptr;
		return 0;
	}
	catch (...) {
		*Handle = nullptr;
		return 1;
	}
}

extern "C" ANSODENGINE_API int ReleaseANSODHandle(ANSCENTER::ANSODBase** Handle) {
	__try {
		return ReleaseANSODHandle_Impl(Handle);
	}
	__except (EXCEPTION_EXECUTE_HANDLER) {
		return 1;
	}
}
ANSODENGINE_API std::string RunInference(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength) {
	try {
		if (Handle == nullptr || *Handle == nullptr) return "";
		if (jpeg_string == nullptr || bufferLength == 0) return "";
		cv::Mat frame = cv::imdecode(cv::Mat(1, bufferLength, CV_8UC1, jpeg_string), cv::IMREAD_COLOR);
		if (frame.empty()) {
			std::string result = "";
			result.clear();
			return result;
		}
		std::string detectionResult;
		(*Handle)->RunInference(frame, "Cam", detectionResult);
		frame.release();
		return detectionResult;
	}
	catch (...) {
		std::string result = "";
		result.clear();
		return result;
	}

}
ANSODENGINE_API std::string RunTiledInference(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength, int tiledWidth, int titledHeight, double overlap, const char* cameraId) {
	try {
		if (Handle == nullptr || *Handle == nullptr) return "";
		if (jpeg_string == nullptr || bufferLength == 0) return "";
		cv::Mat frame = cv::imdecode(cv::Mat(1, bufferLength, CV_8UC1, jpeg_string), cv::IMREAD_COLOR);
		if (frame.empty()) {
			std::string result = "";
			result.clear();
			return result;
		}
		std::vector<ANSCENTER::Object> outputs = (*Handle)->RunInferences(frame, tiledWidth, titledHeight, overlap, cameraId);
		frame.release();
		return ANSCENTER::ANSUtilityHelper::VectorDetectionToJsonString(outputs);
	}
	catch (...) {
		std::string result = "";
		result.clear();
		return result;
	}

}
ANSODENGINE_API std::string RunInferenceFromJpegString(ANSCENTER::ANSODBase** Handle, const char* jpeg_string, unsigned long jpeg_size, const char* cameraId) {
	try {
		if (Handle == nullptr || *Handle == nullptr) return "";
		std::vector<ANSCENTER::Object> outputs = (*Handle)->RunInferenceFromJpegString(jpeg_string, jpeg_size, cameraId);
		return ANSCENTER::ANSUtilityHelper::VectorDetectionToJsonString(outputs);
	}
	catch (...) {
		std::string result = "";
		result.clear();
		return result;

	}
}
ANSODENGINE_API std::string RunTiledInferenceFromJpegString(ANSCENTER::ANSODBase** Handle, const char* jpeg_string, unsigned long jpeg_size, int tiledWidth, int titledHeight, double overlap, const char* cameraId) {
	try {
		if (Handle == nullptr || *Handle == nullptr) return "";
		std::vector<ANSCENTER::Object> outputs = (*Handle)->RunTiledInferenceFromJpegString(jpeg_string, jpeg_size, tiledWidth, titledHeight, overlap, cameraId);
		return ANSCENTER::ANSUtilityHelper::VectorDetectionToJsonString(outputs);
	}
	catch (...) {
		std::string result = "";
		result.clear();
		return result;

	}

}
ANSODENGINE_API std::string RunInferenceFromCV(ANSCENTER::ANSODBase** Handle, cv::Mat image)
{
	try {
		if (Handle == nullptr || *Handle == nullptr) return "";
		std::string detectionResult;
		(*Handle)->RunInference(image, "Cam", detectionResult);
		return detectionResult;
	}
	catch (...) {
		std::string result = "";
		result.clear();
		return result;
	}
}
extern "C" ANSODENGINE_API void			RunDetectMovement(ANSCENTER::ANSODBase** Handle, cv::Mat image, const char* cameraId, std::vector<ANSCENTER::Object>& results) {
	try{
		if (Handle == nullptr || *Handle == nullptr) { results.clear(); return; }
		results = (*Handle)->DetectMovement(image, cameraId);
	}
	catch (...) {
		results.clear();
	}
}
extern "C" ANSODENGINE_API void RunTiledInferenceFromCV(ANSCENTER::ANSODBase** Handle, cv::Mat image, int tiledWidth, int titledHeight, double overlap, std::vector<ANSCENTER::Object>& results, const char* cameraId) {
	try {
		if (Handle == nullptr || *Handle == nullptr) { results.clear(); return; }
		results = (*Handle)->RunInferences(image, tiledWidth, titledHeight, overlap, cameraId);
	}
	catch (...) {
		results.clear();
	}
}
ANSODENGINE_API std::string RunInferenceInCroppedBBoxImages(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength, const char* cameraId, const char* strBboxes) {
	try {
		if (Handle == nullptr || *Handle == nullptr) return "";
		if (jpeg_string == nullptr || bufferLength == 0) return "";
		cv::Mat frame = cv::imdecode(cv::Mat(1, bufferLength, CV_8UC1, jpeg_string), cv::IMREAD_COLOR);
		if (frame.empty()) {
			std::string result = "";
			result.clear();
			return result;
		}
		std::vector<cv::Rect> bBoxes = ANSCENTER::ANSUtilityHelper::GetBoundingBoxesFromString(strBboxes);
		std::vector<ANSCENTER::Object> outputs = (*Handle)->RunInference(frame, bBoxes, cameraId);
		frame.release();
		return ANSCENTER::ANSUtilityHelper::VectorDetectionToJsonString(outputs);
	}
	catch (...) {
		std::string result = "";
		result.clear();
		return result;
	}

}
ANSODENGINE_API std::string RunInferenceInCroppedPolygonImages(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength, const char* cameraId, const char* strPolygon) {
	try {
		if (Handle == nullptr || *Handle == nullptr) return "";
		if (jpeg_string == nullptr || bufferLength == 0) return "";
		cv::Mat frame = cv::imdecode(cv::Mat(1, bufferLength, CV_8UC1, jpeg_string), cv::IMREAD_COLOR);
		if (frame.empty()) {
			std::string result = "";
			result.clear();
			return result;
		}
		std::vector<cv::Point> polygon = ANSCENTER::ANSUtilityHelper::StringToPolygon(strPolygon);
		std::vector<ANSCENTER::Object> outputs = (*Handle)->RunInference(frame, polygon, cameraId);
		frame.release();
		return ANSCENTER::ANSUtilityHelper::VectorDetectionToJsonString(outputs);
	}
	catch (...) {
		std::string result = "";
		result.clear();
		return result;
	}
}
ANSODENGINE_API std::string RunInferenceBinary(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_bytes, unsigned int width, unsigned int height) {
	try {
		if (Handle == nullptr || *Handle == nullptr) return "";
		if (jpeg_bytes == nullptr || width == 0 || height == 0) return "";
		cv::Mat frame = cv::Mat(height, width, CV_8UC3, jpeg_bytes).clone(); // make a copy
		if (frame.empty()) {
			std::string result = "";
			result.clear();
			return result;
		}
		std::string detectionResult;
		(*Handle)->RunInference(frame, "Cam", detectionResult);
		frame.release();
		return detectionResult;
	}
	catch (...) {
		std::string result = "";
		result.clear();
		return result;
	}
	
}
ANSODENGINE_API std::string RunInferenceImagePath(ANSCENTER::ANSODBase** Handle, const char* imageFilePath) {
	try {
		if (Handle == nullptr || *Handle == nullptr) return "";
		std::string stImageFileName(imageFilePath);
		cv::Mat frame = cv::imread(stImageFileName, cv::ImreadModes::IMREAD_COLOR);
		if (frame.empty()) {
			std::string result = "";
			result.clear();
			return result;
		}
		std::string detectionResult;
		(*Handle)->RunInference(frame, "Cam", detectionResult);
		frame.release();
		return detectionResult;
	}
	catch (...) {
		std::string result = "";
		result.clear();
		return result;
	}
	
}

extern "C" ANSODENGINE_API int			RunInference_LV(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength, LStrHandle detectionResult) {
	try {
		std::string st = RunInference(Handle, jpeg_string, bufferLength);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
}
extern "C" ANSODENGINE_API int			RunTiledInference_LV(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength, int tiledWidth, int titledHeight, double overlap, const char* cameraId, LStrHandle detectionResult) {
	try {
		std::string st = RunTiledInference(Handle, jpeg_string, bufferLength, tiledWidth, titledHeight, overlap, cameraId);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
}
extern "C" ANSODENGINE_API int			RunInferenceFromJpegString_LV(ANSCENTER::ANSODBase** Handle, const char* jpeg_string, unsigned long jpeg_size, const char* cameraId, LStrHandle detectionResult) {
	try {
		std::string st = RunInferenceFromJpegString(Handle, jpeg_string, jpeg_size, cameraId);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
	
}
extern "C" ANSODENGINE_API int			RunTiledInferenceFromJpegString_LV(ANSCENTER::ANSODBase** Handle, const char* jpeg_string, unsigned long jpeg_size, int tiledWidth, int titledHeight, double overlap, const char* cameraId, LStrHandle detectionResult) {
	try {
		std::string st = RunTiledInferenceFromJpegString(Handle, jpeg_string, jpeg_size, tiledWidth, titledHeight, overlap, cameraId);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
	
}
extern "C" ANSODENGINE_API int		    RunInferenceInCroppedBBoxImages_LV(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, int32 bufferLength, const char* cameraId, const char* strBboxes, LStrHandle detectionResult) {
	try {
		std::string st = RunInferenceInCroppedBBoxImages(Handle, jpeg_string, bufferLength, cameraId, strBboxes);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
}
extern "C" ANSODENGINE_API int		    RunInferenceInCroppedBBoxPolygonImages_LV(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, int32 bufferLength, const char* cameraId, const char* strPolygon, LStrHandle detectionResult) {
	try {
		std::string st = RunInferenceInCroppedPolygonImages(Handle, jpeg_string, bufferLength, cameraId, strPolygon);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
}
extern "C" ANSODENGINE_API int			RunInferenceBinary_LV(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_bytes, unsigned int width, unsigned int height, LStrHandle detectionResult) {
	try {
		std::string st = RunInferenceBinary(Handle, jpeg_bytes, width, height);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
	
}
extern "C" ANSODENGINE_API int			RunInferenceImagePath_LV(ANSCENTER::ANSODBase** Handle, const char* imageFilePath, LStrHandle detectionResult) {
	try {
		std::string st = RunInferenceImagePath(Handle, imageFilePath);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
	
}

extern "C" ANSODENGINE_API int OptimizeModel(const char* modelFilePath, const char* modelFileZipPassword,int modelType, int modelDetectionType, int fp16, LStrHandle optimizedModelFolder) {
	try {
		std::string st;
		int ret = OptimizeModelStr(modelFilePath, modelFileZipPassword, modelType, modelDetectionType, fp16, st);
		if (ret <= 0 || st.empty()) {
			return 0;
		}
		int size = static_cast<int>(st.length());
		if (size > 0 && optimizedModelFolder) {
			MgErr error = DSSetHandleSize(optimizedModelFolder, sizeof(int32) + size * sizeof(uChar));
			if (error == noErr) {
				(*optimizedModelFolder)->cnt = size;
				memcpy((*optimizedModelFolder)->str, st.c_str(), size);
				return 1;
			}
		}
		return 0;
	}
	catch (const std::exception& e) {
		return 0;
	}
	catch (...) {
		return 0;
	}
}
extern "C" __declspec(dllexport) const char* CreateANSODHandle_CS(ANSCENTER::ANSODBase** Handle, const char* licenseKey, const char* modelFilePath, const char* modelFileZipPassword, float modelThreshold, float modelConfThreshold, float modelNMSThreshold, int autoDetectEngine, int modelType, int detectionType, int loadEngineOnCreation) {
	try {
		static std::string result;
		result = CreateANSODHandle(Handle, licenseKey, modelFilePath, modelFileZipPassword, modelThreshold, modelConfThreshold, modelNMSThreshold, autoDetectEngine, modelType, detectionType, loadEngineOnCreation);
		return result.c_str();
	}
	catch (...) {
		return "";
	}

}
extern "C" __declspec(dllexport) const char* RunInferenceImagePath_CS(ANSCENTER::ANSODBase** Handle, const char* imageFilePath) {
	try {
		static std::string result;
		result = RunInferenceImagePath(Handle, imageFilePath);
		return result.c_str();
	}
	catch (...) {
		return "";
	}

}
extern "C" __declspec(dllexport) const char* RunInference_CS(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength) {
	try {
		static std::string result;
		result = RunInference(Handle, jpeg_string, bufferLength);
		return result.c_str();
	}
	catch (...) {
		return "";
	}

}
extern "C" __declspec(dllexport) const char* RunInferenceInCroppedBBoxImages_CS(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength, const char* cameraId, const char* strBboxes) {
	try {
		static std::string result;
		result = RunInferenceInCroppedBBoxImages(Handle, jpeg_string, bufferLength, cameraId, strBboxes);
		return result.c_str();
	}
	catch (...) {
		return "";
	}

}
extern "C" __declspec(dllexport) const char* RunInferenceInCroppedPolygonImages_CS(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength, const char* cameraId, const char* strPolygon) {
	try {
		static std::string result;
		result = RunInferenceInCroppedPolygonImages(Handle, jpeg_string, bufferLength, cameraId, strPolygon);
		return result.c_str();
	}
	catch (...) {
		return "";
	}
}
extern "C" __declspec(dllexport) const char* RunInferenceBinary_CS(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_bytes, unsigned int width, unsigned int height) {
	try {
		static std::string result;
		result = RunInferenceBinary(Handle, jpeg_bytes, width, height);
		return result.c_str();
	}
	catch (...) {
		return "";
	}

}
extern "C" __declspec(dllexport) const char* OptimizeModelStr_CS(const char* modelFilePath, const char* modelFileZipPassword, int modelType, int modelDetectionType, int fp16)
{
	try {
		static std::string result;
		result.clear();
		int ret = OptimizeModelStr(modelFilePath, modelFileZipPassword, modelType, modelDetectionType, fp16, result);
		return (ret > 0 && !result.empty()) ? result.c_str() : "";
	}
	catch (...) {
		return "";
	}

}

// with camera id
extern "C" ANSODENGINE_API int			RunDetectMovement_LV(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength, const char* cameraId, LStrHandle detectionResult) {
	try {
		cv::Mat frame = cv::imdecode(cv::Mat(1, bufferLength, CV_8UC1, jpeg_string), cv::IMREAD_COLOR);
		if (frame.empty()) {
			return 0;
		}
		std::vector<ANSCENTER::Object> outputs = (*Handle)->DetectMovement(frame, cameraId);
		frame.release();
		std::string st = ANSCENTER::ANSUtilityHelper::VectorDetectionToJsonString(outputs);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
	
}
extern "C" ANSODENGINE_API int			RunInferenceFromJpegStringWithCameraId_LV(ANSCENTER::ANSODBase** Handle, unsigned char* jpeg_string, unsigned int bufferLength, const char* cameraId, LStrHandle detectionResult) {
	try {
		cv::Mat frame = cv::imdecode(cv::Mat(1, bufferLength, CV_8UC1, jpeg_string), cv::IMREAD_COLOR);
		if (frame.empty()) {
			return 0;
		}
		std::string st;
		(*Handle)->RunInference(frame, cameraId, st);
		frame.release();

		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error;
		error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr)
		{
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) {
		return 0;
	}
}
static int RunInferenceComplete_LV_Impl(
	ANSCENTER::ANSODBase** Handle,
	cv::Mat** cvImage,
	const char* cameraId,
	int getJpegString,
	int jpegImageSize,
	const char* activeROIMode,
	LStrHandle detectionResult,
	LStrHandle imageStr)
{
	if (Handle == nullptr || *Handle == nullptr) {
		return -1;
	}
	if (!cvImage || !(*cvImage) || (*cvImage)->empty()) {
		return -2;
	}

	// RAII guard — prevents ReleaseANSODHandle from destroying the engine
	// while we are still using it. Destructor auto-releases refcount.
	ANSCENTER::ANSODBase* handleSnapshot = *Handle;  // snapshot the pointer value
	ODHandleGuard guard(AcquireODHandle(handleSnapshot));
	if (!guard) {
		return -3;  // Handle was already released
	}
	auto* engine = guard.get();

	try {
		auto _t0 = std::chrono::steady_clock::now();

		// Save/restore thread-local to support nested calls (custom model DLLs
		// calling back into ANSODEngine via ANSLIB.dll).
		GpuFrameData* savedFrame = tl_currentGpuFrame();

		// Lookup NV12 frame data BEFORE cloning (clone creates new cv::Mat*)
		GpuFrameData* gpuFrame = ANSGpuFrameRegistry::instance().lookup(*cvImage);

		cv::Mat localImage = (**cvImage).clone();

		int originalWidth = localImage.cols;
		int originalHeight = localImage.rows;

		ANS_DBG("LV_Inference", "START cam=%s %dx%d gpuFrame=%p nv12=%s",
				cameraId ? cameraId : "?", originalWidth, originalHeight,
				(void*)gpuFrame, gpuFrame ? "YES" : "NO");

		if (originalWidth == 0 || originalHeight == 0) {
			tl_currentGpuFrame() = savedFrame;
			return -2;
		}

		// Set thread-local so engines can access NV12 data without registry lookup.
		// Safe: *cvImage holds a refcount, keeping gpuFrame alive during inference.
		// Only use OWN gpuFrame — never inherit outer caller's frame (dimension mismatch on crops).
		tl_currentGpuFrame() = gpuFrame;
		auto _t1 = std::chrono::steady_clock::now();
		std::vector<ANSCENTER::Object> outputs = engine->RunInferenceWithOption(localImage, cameraId, activeROIMode);
		auto _t2 = std::chrono::steady_clock::now();
		tl_currentGpuFrame() = savedFrame;

		double prepMs = std::chrono::duration<double, std::milli>(_t1 - _t0).count();
		double infMs  = std::chrono::duration<double, std::milli>(_t2 - _t1).count();
		if (infMs > 500.0) {
			ANS_DBG("LV_Inference", "SLOW cam=%s prep=%.1fms inf=%.1fms results=%zu",
					cameraId ? cameraId : "?", prepMs, infMs, outputs.size());
		}
		bool getJpeg = (getJpegString == 1);
		std::string stImage;
		// NOTE: odMutex was removed here. All variables in this scope are local
		// (outputs, localImage, stImage, stDetectionResult) and the LStrHandle
		// parameters are per-call. The global mutex was blocking running tasks
		// for minutes when another task was optimizing a TRT engine.

		int maxImageSize = originalWidth;// std::max(originalWidth, originalHeight);
		bool resizeNeeded = (jpegImageSize > 0) && (jpegImageSize < maxImageSize);

		float ratio = 1.0f;
		int newWidth = originalWidth;
		int newHeight = originalHeight;

		if (resizeNeeded) {
			newWidth = jpegImageSize;
			newHeight = static_cast<int>(std::round(newWidth * static_cast<double>(originalHeight) / originalWidth));
			ratio = static_cast<float>(newWidth) / originalWidth;

			for (auto& obj : outputs) {
				obj.box.x = std::max(0, std::min(static_cast<int>(obj.box.x * ratio), newWidth - 1));
				obj.box.y = std::max(0, std::min(static_cast<int>(obj.box.y * ratio), newHeight - 1));
				obj.box.width = std::max(1, std::min(static_cast<int>(obj.box.width * ratio), newWidth - obj.box.x));
				obj.box.height = std::max(1, std::min(static_cast<int>(obj.box.height * ratio), newHeight - obj.box.y));
			}
		}
		else {
			for (auto& obj : outputs) {
				obj.box.x = std::max(0, std::min(static_cast<int>(obj.box.x), originalWidth - 1));
				obj.box.y = std::max(0, std::min(static_cast<int>(obj.box.y), originalHeight - 1));
				obj.box.width = std::max(1, std::min(static_cast<int>(obj.box.width), originalWidth - obj.box.x));
				obj.box.height = std::max(1, std::min(static_cast<int>(obj.box.height), originalHeight - obj.box.y));
			}
		}

		// Convert to JPEG if needed
		if (getJpeg) {
			cv::Mat processedImage = localImage;
			if (resizeNeeded) {
				cv::resize(localImage, processedImage, cv::Size(newWidth, newHeight), 0, 0, cv::INTER_AREA);
			}

			std::vector<uchar> buf;
			if (cv::imencode(".jpg", processedImage, buf, { cv::IMWRITE_JPEG_QUALITY, 50 })) {
				stImage.assign(buf.begin(), buf.end());
			}
			else {
				std::cerr << "Error: JPEG encoding failed!" << std::endl;
			}
		}

		std::string stDetectionResult = ANSCENTER::ANSUtilityHelper::VectorDetectionToJsonString(outputs);
		if (stDetectionResult.empty()) {
			return 0;
		}

		int size = static_cast<int>(stDetectionResult.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error != noErr) {
			return 0;
		}

		(*detectionResult)->cnt = size;
		memcpy((*detectionResult)->str, stDetectionResult.c_str(), size);

		if (getJpeg) {
			if (stImage.empty()) {
				return 0;
			}
			size = static_cast<int>(stImage.length());
			error = DSSetHandleSize(imageStr, sizeof(int32) + size * sizeof(uChar));
			if (error != noErr) {
				return 0;
			}

			(*imageStr)->cnt = size;
			memcpy((*imageStr)->str, stImage.c_str(), size);
		}

		return 1;
	}
	catch (const std::exception& ex) {
		return 0;
	}
	catch (...) {
		return 0;
	}
}

// SEH wrapper — no C++ objects allowed in this function
static int RunInferenceComplete_LV_SEH(
	ANSCENTER::ANSODBase** Handle,
	cv::Mat** cvImage,
	const char* cameraId,
	int getJpegString,
	int jpegImageSize,
	const char* activeROIMode,
	LStrHandle detectionResult,
	LStrHandle imageStr)
{
	__try {
		return RunInferenceComplete_LV_Impl(Handle, cvImage, cameraId, getJpegString, jpegImageSize, activeROIMode, detectionResult, imageStr);
	}
	__except (EXCEPTION_EXECUTE_HANDLER) {
		return -4;
	}
}

extern "C" ANSODENGINE_API int RunInferenceComplete_LV(
	ANSCENTER::ANSODBase** Handle,
	cv::Mat** cvImage,
	const char* cameraId,
	int getJpegString,
	int jpegImageSize,
	const char* activeROIMode,
	LStrHandle detectionResult,
	LStrHandle imageStr)
{
	return RunInferenceComplete_LV_SEH(Handle, cvImage, cameraId, getJpegString, jpegImageSize, activeROIMode, detectionResult, imageStr);
}

// V2: Accepts handle as uint64_t by value — eliminates pointer-to-pointer
// instability when LabVIEW calls concurrently from multiple tasks.
// LabVIEW CLFN: set Handle parameter to Numeric / Unsigned Pointer-sized Integer / Pass: Value
extern "C" ANSODENGINE_API int RunInferenceComplete_LV_V2(
	uint64_t handleVal,
	cv::Mat** cvImage,
	const char* cameraId,
	int getJpegString,
	int jpegImageSize,
	const char* activeROIMode,
	LStrHandle detectionResult,
	LStrHandle imageStr)
{
	// Cast the by-value integer back to the engine pointer — no double dereference
	ANSCENTER::ANSODBase* directHandle = reinterpret_cast<ANSCENTER::ANSODBase*>(handleVal);
	if (directHandle == nullptr) {
		return -1;
	}
	// Reuse a temporary pointer so we can call the existing Impl function
	ANSCENTER::ANSODBase* tempArr[1] = { directHandle };
	ANSCENTER::ANSODBase** pHandle = &tempArr[0];

	__try {
		return RunInferenceComplete_LV_Impl(pHandle, cvImage, cameraId, getJpegString, jpegImageSize, activeROIMode, detectionResult, imageStr);
	}
	__except (EXCEPTION_EXECUTE_HANDLER) {
		return -4;
	}
}

// ============================================================================
// V2 LabVIEW API — Accept handle as uint64_t by value.
// Eliminates Handle** pointer-to-pointer instability when LabVIEW calls
// concurrently from multiple tasks.
// LabVIEW CLFN: set Handle parameter to Numeric / Unsigned Pointer-sized Integer / Pass: Value
// ============================================================================

// Helper: cast uint64_t handle to ANSODBase** for delegation to existing functions
#define V2_HANDLE_SETUP(handleVal) \
	ANSCENTER::ANSODBase* _v2Direct = reinterpret_cast<ANSCENTER::ANSODBase*>(handleVal); \
	if (_v2Direct == nullptr) return 0; \
	ANSCENTER::ANSODBase* _v2Arr[1] = { _v2Direct }; \
	ANSCENTER::ANSODBase** Handle = &_v2Arr[0];

extern "C" ANSODENGINE_API int RunInference_LV_V2(uint64_t handleVal, unsigned char* jpeg_string, unsigned int bufferLength, LStrHandle detectionResult) {
	try {
		V2_HANDLE_SETUP(handleVal);
		std::string st = RunInference(Handle, jpeg_string, bufferLength);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int RunTiledInference_LV_V2(uint64_t handleVal, unsigned char* jpeg_string, unsigned int bufferLength, int tiledWidth, int titledHeight, double overlap, const char* cameraId, LStrHandle detectionResult) {
	try {
		V2_HANDLE_SETUP(handleVal);
		std::string st = RunTiledInference(Handle, jpeg_string, bufferLength, tiledWidth, titledHeight, overlap, cameraId);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int RunTiledInferenceFromJpegString_LV_V2(uint64_t handleVal, const char* jpeg_string, unsigned long jpeg_size, int tiledWidth, int titledHeight, double overlap, const char* cameraId, LStrHandle detectionResult) {
	try {
		V2_HANDLE_SETUP(handleVal);
		std::string st = RunTiledInferenceFromJpegString(Handle, jpeg_string, jpeg_size, tiledWidth, titledHeight, overlap, cameraId);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int RunInferenceFromJpegString_LV_V2(uint64_t handleVal, const char* jpeg_string, unsigned long jpeg_size, const char* cameraId, LStrHandle detectionResult) {
	try {
		V2_HANDLE_SETUP(handleVal);
		std::string st = RunInferenceFromJpegString(Handle, jpeg_string, jpeg_size, cameraId);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int RunInferenceBinary_LV_V2(uint64_t handleVal, unsigned char* jpeg_bytes, unsigned int width, unsigned int height, LStrHandle detectionResult) {
	try {
		V2_HANDLE_SETUP(handleVal);
		std::string st = RunInferenceBinary(Handle, jpeg_bytes, width, height);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int RunInferenceImagePath_LV_V2(uint64_t handleVal, const char* imageFilePath, LStrHandle detectionResult) {
	try {
		V2_HANDLE_SETUP(handleVal);
		std::string st = RunInferenceImagePath(Handle, imageFilePath);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int RunInferenceInCroppedBBoxImages_LV_V2(uint64_t handleVal, unsigned char* jpeg_string, int32 bufferLength, const char* cameraId, const char* strBboxes, LStrHandle detectionResult) {
	try {
		V2_HANDLE_SETUP(handleVal);
		std::string st = RunInferenceInCroppedBBoxImages(Handle, jpeg_string, bufferLength, cameraId, strBboxes);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int RunInferenceInCroppedBBoxPolygonImages_LV_V2(uint64_t handleVal, unsigned char* jpeg_string, int32 bufferLength, const char* cameraId, const char* strPolygon, LStrHandle detectionResult) {
	try {
		V2_HANDLE_SETUP(handleVal);
		std::string st = RunInferenceInCroppedPolygonImages(Handle, jpeg_string, bufferLength, cameraId, strPolygon);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int RunDetectMovement_LV_V2(uint64_t handleVal, unsigned char* jpeg_string, unsigned int bufferLength, const char* cameraId, LStrHandle detectionResult) {
	try {
		ANSCENTER::ANSODBase* directHandle = reinterpret_cast<ANSCENTER::ANSODBase*>(handleVal);
		if (directHandle == nullptr) return 0;
		cv::Mat frame = cv::imdecode(cv::Mat(1, bufferLength, CV_8UC1, jpeg_string), cv::IMREAD_COLOR);
		if (frame.empty()) return 0;
		std::vector<ANSCENTER::Object> outputs = directHandle->DetectMovement(frame, cameraId);
		frame.release();
		std::string st = ANSCENTER::ANSUtilityHelper::VectorDetectionToJsonString(outputs);
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int RunInferenceFromJpegStringWithCameraId_LV_V2(uint64_t handleVal, unsigned char* jpeg_string, unsigned int bufferLength, const char* cameraId, LStrHandle detectionResult) {
	try {
		ANSCENTER::ANSODBase* directHandle = reinterpret_cast<ANSCENTER::ANSODBase*>(handleVal);
		if (directHandle == nullptr) return 0;
		cv::Mat frame = cv::imdecode(cv::Mat(1, bufferLength, CV_8UC1, jpeg_string), cv::IMREAD_COLOR);
		if (frame.empty()) return 0;
		std::string st;
		directHandle->RunInference(frame, cameraId, st);
		frame.release();
		if (st.empty()) return 0;
		int size = static_cast<int>(st.length());
		MgErr error = DSSetHandleSize(detectionResult, sizeof(int32) + size * sizeof(uChar));
		if (error == noErr) {
			(*detectionResult)->cnt = size;
			memcpy((*detectionResult)->str, st.c_str(), size);
			return 1;
		}
		else return 0;
	}
	catch (...) { return 0; }
}

extern "C" ANSODENGINE_API int			RunInferenceComplete_CPP(ANSCENTER::ANSODBase** Handle, cv::Mat** cvImage, const char* cameraId, const char* activeROIMode, std::vector<ANSCENTER::Object> &detectionResult) {
	if (Handle == nullptr || *Handle == nullptr) {
		std::cerr << "Error: Handle is null in RunInferenceComplete_CPP!" << std::endl;
		return -1;
	}
	if (!cvImage || !(*cvImage) || (*cvImage)->empty()) {
		std::cerr << "Error: Invalid or empty input image in RunInferenceComplete_CPP!" << std::endl;
		return -2;
	}

	ODHandleGuard guard(AcquireODHandle(*Handle));
	if (!guard) {
		return -3;  // Handle was already released
	}
	auto* engine = guard.get();

	try {
		// Save/restore thread-local to support nested calls (e.g., custom model DLLs
		// calling back into ANSODEngine via ANSLIB.dll). Without this, the inner call
		// would overwrite the outer caller's valid GpuFrameData* with nullptr.
		GpuFrameData* savedFrame = tl_currentGpuFrame();

		GpuFrameData* gpuFrame = ANSGpuFrameRegistry::instance().lookup(*cvImage);

		cv::Mat localImage = (**cvImage).clone();
		int originalWidth = localImage.cols;
		int originalHeight = localImage.rows;

		if (originalWidth == 0 || originalHeight == 0) {
			tl_currentGpuFrame() = savedFrame;
			return -2;
		}
		// Only use gpuFrame if this cv::Mat* has NV12 data.
		// Do NOT propagate savedFrame to inner engines — the inner call's cv::Mat*
		// is a local copy (from ANSLIB.dll) that doesn't correspond to the outer
		// frame's NV12 geometry. Using savedFrame here would cause the inner engine
		// to read NV12 data that doesn't match its input image.
		tl_currentGpuFrame() = gpuFrame;
		detectionResult = engine->RunInferenceWithOption(localImage, cameraId, activeROIMode);
		tl_currentGpuFrame() = savedFrame;
		if (detectionResult.empty()) return 0;
		else return 1;
	}
	catch (const std::exception& ex) {
		std::cerr << "Exception in RunInferenceComplete_CPP: " << ex.what() << std::endl;
		return 0;
	}
	catch (...) {
		std::cerr << "Unknown exception in RunInferenceComplete_CPP!" << std::endl;
		return 0;
	}

}

extern "C" ANSODENGINE_API int			RunInference_CPP(ANSCENTER::ANSODBase** Handle, cv::Mat** cvImage, const char* cameraId, std::vector<ANSCENTER::Object>& detectionResult) {
	if (Handle == nullptr || *Handle == nullptr) {
		std::cerr << "Error: Handle is null in RunInference_CPP!" << std::endl;
		return -1;
	}
	if (!cvImage || !(*cvImage) || (*cvImage)->empty()) {
		std::cerr << "Error: Invalid or empty input image in RunInference_CPP!" << std::endl;
		return -2;
	}

	ODHandleGuard guard(AcquireODHandle(*Handle));
	if (!guard) {
		return -3;  // Handle was already released
	}
	auto* engine = guard.get();

	try {
		// Save/restore thread-local (same nested-call protection as RunInferenceComplete_CPP)
		GpuFrameData* savedFrame = tl_currentGpuFrame();

		GpuFrameData* gpuFrame = ANSGpuFrameRegistry::instance().lookup(*cvImage);

		cv::Mat localImage = (**cvImage).clone();
		int originalWidth = localImage.cols;
		int originalHeight = localImage.rows;

		if (originalWidth == 0 || originalHeight == 0) {
			tl_currentGpuFrame() = savedFrame;
			return -2;
		}
		// Only use gpuFrame if this cv::Mat* has NV12 data (see RunInferenceComplete_CPP comment)
		tl_currentGpuFrame() = gpuFrame;
		detectionResult = engine->RunInference(localImage, cameraId);
		tl_currentGpuFrame() = savedFrame;
		if (detectionResult.empty()) return 0;
		else return 1;
	}
	catch (const std::exception& ex) {
		std::cerr << "Exception in RunInference_CPP: " << ex.what() << std::endl;
		return 0;
	}
	catch (...) {
		std::cerr << "Unknown exception in RunInference_CPP!" << std::endl;
		return 0;
	}
}


// Utility function to convert a string to a vector of cv::Point
extern "C" __declspec(dllexport) int GetEngineType() {
	ANSCENTER::EngineType engineType = ANSCENTER::ANSLicenseHelper::CheckHardwareInformation();
	switch (engineType) {
	case ANSCENTER::EngineType::CPU:
		return 0; // CPU
	case ANSCENTER::EngineType::NVIDIA_GPU:
		return 1; // GPU
	case ANSCENTER::EngineType::OPENVINO_GPU:
		return 2; // NPU
	default:
		return 0; // Unknown
	}
}


extern "C" __declspec(dllexport) int GetActiveRect(ANSCENTER::ANSODBase** Handle, cv::Mat cvImage, cv::Rect& activeWindow) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}

		if (cvImage.empty()) {
			std::cerr << "Error: Input image is empty!" << std::endl;
			return -1;
		}

		activeWindow = (*Handle)->GetActiveWindow(cvImage);
		if (activeWindow.empty()) {
			std::cerr << "Error: Active window is empty!" << std::endl;
			return 0;
		}

		return 1; // Success
	}
	catch (...) {
		std::cerr << "Unknown exception in GetActiveWindow!" << std::endl;
		return -1;
	}

}
extern "C" __declspec(dllexport) int DetectMovement(ANSCENTER::ANSODBase** Handle, cv::Mat image, const char* cameraId, std::vector<ANSCENTER::Object>& results) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}

		if (image.empty()) {
			std::cerr << "Error: Input image is empty!" << std::endl;
			return -1;
		}

		results = (*Handle)->DetectMovement(image, cameraId);
		if (results.empty()) {
			std::cerr << "Error: No detection results!" << std::endl;
			return 0;
		}

		return 1; // Success
	}
	catch (...) {
		std::cerr << "Unknown exception in DetectMovement!" << std::endl;
		return -1;
	}
}

extern "C" __declspec(dllexport) int Optimize(ANSCENTER::ANSODBase** Handle, bool fp16) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}
		std::string modelFolder;
		bool result = (*Handle)->OptimizeModel(fp16, modelFolder);
		if (!result) {
			std::cerr << "Error: Model optimization failed!" << std::endl;
			return 0;
		}

		return 1; // Success
	}
	catch (...) {
		std::cerr << "Unknown exception in Optimize!" << std::endl;
		return -1;
	}
}

extern "C" __declspec(dllexport) int SetODParameters(ANSCENTER::ANSODBase** Handle, const char* parameters) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}

		if (parameters == nullptr) {
			std::cerr << "Error: Parameters string is null!" << std::endl;
			return -1;
		}
		std::string paramStr(parameters);
		ANSCENTER::Params param = ANSCENTER::ANSUtilityHelper::ParseCustomParameters(paramStr);
		bool result = (*Handle)->SetParameters(param);
		if (!result) {
			std::cerr << "Error: Setting parameters failed!" << std::endl;
			return 0;
		}

		return 1; // Success
	}
	catch (...) {
		std::cerr << "Unknown exception in SetParameters!" << std::endl;
		return -1;
	}
}

extern "C" __declspec(dllexport) int GetODParameters(ANSCENTER::ANSODBase** Handle, ANSCENTER::Params& param) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}
		param = (*Handle)->GetParameters();
		return 1; // Success
	}
	catch (...) {
		std::cerr << "Unknown exception in GetParameters!" << std::endl;
		return -1;
	}
}
extern "C" __declspec(dllexport) int GetConfiguredParameters(ANSCENTER::ANSODBase** Handle, LStrHandle stParam) {
	__try {
		return [&]() -> int {
			try {
				if (Handle == nullptr || *Handle == nullptr) {
					std::cerr << "Error: Handle is null!" << std::endl;
					return -1;
				}
				if (stParam == nullptr) {
					std::cerr << "Error: stParam LStrHandle is null!" << std::endl;
					return -1;
				}
				ANSCENTER::Params param;
				bool result = (*Handle)->ConfigureParameters(param);
				if (!result) {
					std::cerr << "Error: Getting parameters failed!" << std::endl;
					return 0;
				}
				std::string st = ANSCENTER::ANSUtilityHelper::SerializeCustomParamters(param);
				if (st.empty()) return 0;
				int size = static_cast<int>(st.length());
				MgErr error;
				error = DSSetHandleSize(stParam, sizeof(int32) + size * sizeof(uChar));
				if (error == noErr)
				{
					(*stParam)->cnt = size;
					memcpy((*stParam)->str, st.c_str(), size);
					return 1;
				}
				else return 0;
			}
			catch (const std::exception& e) {
				std::cerr << "GetConfiguredParameters exception: " << e.what() << std::endl;
				return 0;
			}
			catch (...) {
				std::cerr << "GetConfiguredParameters unknown exception" << std::endl;
				return 0;
			}
		}();
	}
	__except (EXCEPTION_EXECUTE_HANDLER) {
		std::cerr << "GetConfiguredParameters SEH exception (code: " << GetExceptionCode() << ")" << std::endl;
		return -2;
	}
}
extern "C" __declspec(dllexport) int GetConfiguredParameters_CPP(ANSCENTER::ANSODBase** Handle, std::string& stParam) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}
		ANSCENTER::Params param;
		bool result = (*Handle)->ConfigureParameters(param);
		if (!result) {
			std::cerr << "Error: Getting parameters failed!" << std::endl;
			return 0;
		}
		stParam = ANSCENTER::ANSUtilityHelper::SerializeCustomParamters(param);
		if (stParam.empty()) return 0;
		else return 1;
	}
	catch (...) {
		return 0;
	}

}

extern "C" ANSODENGINE_API int			ShutdownPythonEngine() {
	ANSCENTER::ANSCUSTOMPY::SafeShutdownAll(false);
	return 1;
}
extern "C" ANSODENGINE_API int			ShutdownPythonEngine_CPP() {
	ANSCENTER::ANSCUSTOMPY::SafeShutdownAll(true);
	return 1;
}
extern "C" __declspec(dllexport) int UpdateDetectionMinScore(ANSCENTER::ANSODBase** Handle, float detectionScore) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}

		bool result = (*Handle)->UpdateDetectionThreshold(detectionScore);
		if (!result) {
			std::cerr << "Error: Updating detection threshold failed!" << std::endl;
			return 0;
		}

		return 1; // Success
	}
	catch (...) {
		std::cerr << "Unknown exception in UpdateDetectionMinScore!" << std::endl;
		return -1;
	}
}
extern "C" __declspec(dllexport) int SetPrompt(ANSCENTER::ANSODBase** Handle, const char* textPrompt) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}
		if (textPrompt == nullptr) {
			std::cerr << "Error: Text prompt is null!" << std::endl;
			return -1;
		}
		std::string promptStr(textPrompt);
		bool result = (*Handle)->SetPrompt(promptStr);
		if (!result) {
			std::cerr << "Error: Setting text prompt failed!" << std::endl;
			return 0;
		}
		return 1; // Success
	
	}
	catch (...) {
		std::cerr << "Unknown exception in SetPrompt!" << std::endl;
		return -1;
	}
}
extern "C" __declspec(dllexport) int SetTracker(ANSCENTER::ANSODBase** Handle, int trackerType, int enableTracker) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}
		ANSCENTER::TrackerType ansTrackerType;
		switch (trackerType) {
			case 0:
				ansTrackerType = ANSCENTER::TrackerType::BYTETRACK;
				break;
			case 1:
				ansTrackerType = ANSCENTER::TrackerType::UCMC;
				break;
			case 2:
				ansTrackerType = ANSCENTER::TrackerType::OCSORT;
				break;
			default:
				ansTrackerType = ANSCENTER::TrackerType::BYTETRACK;
				break;
		}
		bool enable = false;
		if (enableTracker == 1)enable = true;
		bool result = (*Handle)->SetTracker(ansTrackerType, enable);
		if (!result) {
			std::cerr << "Error: Setting tracker failed!" << std::endl;
			return 0;
		}
		return 1; // Success
		
	}
	catch (...) {
		std::cerr << "Unknown exception in SetTracker!" << std::endl;
		return -1;
	}
}
extern "C" __declspec(dllexport) int SetTrackerParameters(ANSCENTER::ANSODBase** Handle, const char* trackerParams) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			std::cerr << "Error: Handle is null!" << std::endl;
			return -1;
		}
		if (trackerParams == nullptr) {
			std::cerr << "Error: Tracker parameters string is null!" << std::endl;
			return -1;
		}
		std::string paramStr(trackerParams);
		bool result = (*Handle)->SetTrackerParameters(paramStr);
		if (!result) {
			std::cerr << "Error: Setting tracker parameters failed!" << std::endl;
			return 0;
		}
		return 1; // Success
	}
	catch (...) {
		std::cerr << "Unknown exception in SetTrackerParameters!" << std::endl;
		return -1;
	}
}

// Set stabilization parameters (JSON string).
// All keys are optional — omit to keep current defaults.
// Example: {"ema_alpha":0.4, "class_consistency_frames":8, "hysteresis_enter":0.5, "hysteresis_keep":0.3}
extern "C" __declspec(dllexport) int SetStabilizationParameters(ANSCENTER::ANSODBase** Handle, const char* stabParams) {
	try {
		if (Handle == nullptr || *Handle == nullptr) {
			return -1;
		}
		if (stabParams == nullptr) {
			return -1;
		}
		std::string paramStr(stabParams);
		bool result = (*Handle)->SetStabilizationParameters(paramStr);
		return result ? 1 : 0;
	}
	catch (...) {
		return -1;
	}
}