#ifndef ANSFACERECOGNISER_H
#define ANSFACERECOGNISER_H
#pragma once
#include "ANSFRCommon.h"
#include "hnswlib/hnswlib.h"
#include "cnn.hpp"
#include "face_reid.hpp"
#include "openvino/openvino.hpp"
#include <faiss/IndexFlat.h>
#include <faiss/IndexIDMap.h>
#include <faiss/gpu/GpuIndexFlat.h>
#include <faiss/gpu/StandardGpuResources.h>
#include <unordered_map>
#include "engine.h"
#include "engine/EnginePoolManager.h"
#include "ONNXEngine.h"
#define USE_ONNX_ENGINE
//#define CPU_MODE
//#define USE_CPU_BATCH_MODE
namespace ANSCENTER {
    class ANSFaceRecognizer : public ANSFRBase {
    public:
        virtual bool Initialize(std::string licenseKey,
            ModelConfig modelConfig,
            const std::string& modelZipFilePath,
            const std::string& modelZipPassword,
            std::string& labelMap) override;

        virtual bool LoadModel(const std::string& modelZipFilePath,
            const std::string& modelZipPassword) override;

        bool OptimizeModel(bool fp16, std::string& optimizedModelFolder);

        // Single face feature
        std::vector<float> Feature(const cv::Mat& image,const ANSCENTER::Object& bBox);

        // Full pipeline: embeddings -> FAISS search -> results
        std::vector<FaceResultObject> Match(const cv::Mat& input, const std::vector<ANSCENTER::Object>& bBox, const std::map<std::string, std::string>& userDict);

        cv::Mat GetCropFace(const cv::Mat& input, const ANSCENTER::Object& bBox);

        void Init();
        void AddEmbedding(const std::string& className, float embedding[]);
        void AddEmbedding(const std::string& className, const std::vector<float>& embedding);

        // Double-buffer support: atomically swap FAISS index + mapping
        void SwapIndex(std::shared_ptr<faiss::IndexIDMap> newIndex,
            std::unordered_map<faiss::idx_t, std::string>&& newFaceIdToUserId);

        int GetEmbeddingSize() const { return FACE_EMBEDDING_SIZE; }
        std::shared_ptr<faiss::gpu::StandardGpuResources> GetGpuResources() const { return m_gpuResources; }

        bool UpdateParamater(double knownPersonThreshold) {
            _modelConfig.unknownPersonThreshold = knownPersonThreshold;
            m_knownPersonThresh = _modelConfig.unknownPersonThreshold;
            return true;
        }

        ~ANSFaceRecognizer();
        bool Destroy();

        // L2-normalize a vector in-place (public — used by ANSFR::Reload)
        static void L2NormalizeInPlace(std::vector<float>& vec);

    private:
        bool LoadEngine(const std::string& xmlModelPath, bool engineOptimisation = true);

        // Batched forward: one embedding per Object.mask (caller must hold _mutex)
        std::vector<std::vector<float>> ForwardUnlocked(const cv::Mat& input,const std::vector<ANSCENTER::Object>& outputBbox);

        // FAISS search (caller must hold _mutex)
        std::tuple<std::vector<std::string>, std::vector<float>>
            SearchForFacesUnlocked(const std::vector<std::vector<float>>& detectedEmbeddings);

        std::string GetOpenVINODevice();

        // Single-face GPU inference
        std::vector<float> RunArcFace(const cv::Mat& input);

        // Batched GPU inference
        std::vector<std::vector<float>> RunArcFaceBatch(
            const std::vector<cv::Mat>& faceROIs,
            const std::vector<cv::cuda::GpuMat>& gpuFaceROIs = {});

    protected:
        const int GPU_FACE_WIDTH = 112;
        const int GPU_FACE_HEIGHT = 112;
        const int FACE_EMBEDDING_SIZE = 512;

        std::unordered_map<faiss::idx_t, std::string> _faceIdToUserId;
        faiss::idx_t _nextFaceId = 0; // Sequential ID counter for backward-compat AddEmbedding
        ModelConfig _modelConfig;
        std::string _modelFilePath;
        std::string _landmarkModelFilePath;

        ANSCENTER::Options m_options;
        const std::array<float, 3> SUB_VALS{ 0.5f, 0.5f, 0.5f };
        const std::array<float, 3> DIV_VALS{ 0.5f, 0.5f, 0.5f };
        const bool NORMALIZE = true;

        std::recursive_mutex _mutex;
        float m_knownPersonThresh = 0.35f;
        EngineType engineType;

#ifdef USE_ONNX_ENGINE
        std::unique_ptr<GlintArcFace> faceRecognizer = nullptr;
        const int CPU_FACE_WIDTH = 112;
        const int CPU_FACE_HEIGHT = 112;
#else
        std::unique_ptr<VectorCNN> faceRecognizer = nullptr; // OpenVINO
        const int CPU_FACE_WIDTH = 160;
        const int CPU_FACE_HEIGHT = 160;
#endif
        // Pooled GPU buffers to avoid per-frame allocation (Fix #8)
        cv::cuda::Stream m_gpuStream;
        cv::cuda::GpuMat m_gpuImg;
        cv::cuda::GpuMat m_gpuResized;
        cv::cuda::GpuMat m_gpuRgb;

        std::shared_ptr<Engine<float>> m_trtEngine = nullptr; // NVIDIA TensorRT
        EnginePoolManager<float>::PoolKey m_poolKey;
        bool m_usingSharedPool = false;
        int m_maxSlotsPerGpu{ -1 };  // -1 = elastic mode (on-demand slots, auto-cleanup)
        void SetMaxSlotsPerGpu(int n) override { m_maxSlotsPerGpu = n; }
        std::shared_ptr<faiss::IndexIDMap> faiss_index;
        std::shared_ptr<faiss::gpu::StandardGpuResources> m_gpuResources;
    };

}
#endif