#include "ofxOnnxRuntime.h"

namespace ofxOnnxRuntime
{
#ifdef _MSC_VER
	static std::wstring to_wstring(const std::string &str)
	{
		unsigned len = str.size() * 2;
		setlocale(LC_CTYPE, "");
		wchar_t *p = new wchar_t[len];
		mbstowcs(p, str.c_str(), len);
		std::wstring wstr(p);
		delete[] p;
		return wstr;
	}
#endif
	void BaseHandler::setup(const std::string & onnx_path, const BaseSetting & base_setting, const int & batch_size, const bool debug, const bool timestamp)
	{
		// Store data types
		this->input_dtype = base_setting.input_dtype;
		this->output_dtype = base_setting.output_dtype;

		Ort::SessionOptions session_options;
		session_options.SetIntraOpNumThreads(1);
		session_options.SetIntraOpNumThreads(1);
		session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);

		if (base_setting.infer_type == INFER_CUDA) {
			OrtCUDAProviderOptions opts;
			opts.device_id = 0;
			opts.cudnn_conv_algo_search = OrtCudnnConvAlgoSearchExhaustive;
			opts.do_copy_in_default_stream = 0;
			opts.arena_extend_strategy = 0;
			session_options.AppendExecutionProvider_CUDA(opts);
		}

		this->timestamp = timestamp;
		this->debug = debug;
		this->batch_size = batch_size;
		this->setup2(onnx_path, session_options);
	}

	void BaseHandler::setup2(const std::string & onnx_path, const Ort::SessionOptions & session_options)
	{
		std::string path = ofToDataPath(onnx_path, true);

		std::wstring wpath(path.begin(), path.end());  // basic conversion

		ort_session = std::make_shared<Ort::Session>(ort_env, wpath.c_str(), session_options);

		setNames();
	}

	void BaseHandler::setNames()
	{
		Ort::AllocatorWithDefaultOptions allocator;

		// 1. Gets Input Name/s & Shape ([1, 3, 28, 28]) -- In most cases this is usually just one
		for (std::size_t i = 0; i < ort_session->GetInputCount(); i++) {
			input_node_names.emplace_back(ort_session->GetInputNameAllocated(i, allocator).get());
			input_node_dims = ort_session->GetInputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();

			// Some models might have negative shape values to indicate dynamic shape, e.g., for variable batch size. (?, 3, 28, 28) -> (1, 3, 28, 28)
			for (auto& s : input_node_dims) if (s < 0) s = batch_size;

			if (debug) std::cout << input_node_names.at(i) << " : " << PrintShape(input_node_dims) << std::endl;
		}

		// 2. Calculate the product of the dimensions
		for (auto& f : input_node_dims) {
			input_node_size *= f;
		}

		if (debug)  ofLog() << ofToString(input_node_size) + ", Batch Size:" + ofToString(input_node_dims[0]);

		// 2. Clear up output values
		output_node_dims.clear();
		output_values.clear();
		
		// 3. Gets Output name/s & Shapes
		for (std::size_t i = 0; i < ort_session->GetOutputCount(); i++) {
			output_node_names.emplace_back(ort_session->GetOutputNameAllocated(i, allocator).get());
			auto output_shapes = ort_session->GetOutputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
			
			output_values.emplace_back(nullptr);

			if (debug) std::cout << output_node_names.at(i) << " : " << PrintShape(output_shapes) << std::endl;
		}
	}

	float* BaseHandler::run()
	{
		
		auto start = std::chrono::high_resolution_clock::now(); // starting timestamp

		std::vector<Ort::Value> input_tensors;

		size_t num_images = input_imgs.size();

		if(input_imgs.size() != batch_size) {
			ofLog() << "Input images do not match batch size. Inference FAILED.";
			return dummy_output_tensor.front().GetTensorMutableData<float>();
		}

		// transform std::string -> const char*
		std::vector<const char*> input_names_char(input_node_names.size(), nullptr);
		std::transform(std::begin(input_node_names), std::end(input_node_names), std::begin(input_names_char),
			[&](const std::string& str) { return str.c_str(); });

		std::vector<const char*> output_names_char(output_node_names.size(), nullptr);
		std::transform(std::begin(output_node_names), std::end(output_node_names), std::begin(output_names_char),
			[&](const std::string& str) { return str.c_str(); });

		std::vector<float> batch_values_f;
		std::vector<int32_t> batch_values_int32;
		batch_values_f.reserve(input_node_size * batch_size); // Reserve space but don't initialize
		batch_values_int32.reserve(input_node_size * batch_size); // Reserve space but don't initialize


		if (input_dtype == ModelDataType::FLOAT32){
			// I have a list of imgs, these need to be converted from images into input for the model (int or float)
			for(size_t i = 0; i < batch_size; i++) {
				convertImageToMatFloat(input_imgs[i], batch_values_f, i);
			}

			// 2. Create tensor with batch values { input data, input size, model input dims, model input size}
			input_tensors.emplace_back(Ort::Value::CreateTensor<float>(
				memory_info_handler, batch_values_f.data(), input_node_size,
				input_node_dims.data(), input_node_dims.size()));
		} 
		else if (input_dtype == ModelDataType::INT32) {
			// I have a list of imgs, these need to be converted from images into input for the model (int or float)
			for(size_t i = 0; i < batch_size; i++) {
				convertImageToMatInt32(input_imgs[i], batch_values_int32, i);
			}

			// 2. Create tensor with batch values { input data, input size, model input dims, model input size}
			input_tensors.emplace_back(Ort::Value::CreateTensor<int32_t>(
			memory_info_handler, batch_values_int32.data(), input_node_size,
			input_node_dims.data(), input_node_dims.size()));
		}

		
		

		try {
			// 3. Run inference, { in names, input data, num of inputs, output names, num of outputs }
			ofLog() << "run";
			output_values = ort_session->Run(Ort::RunOptions{ nullptr }, 
				input_names_char.data(), input_tensors.data(),
				input_names_char.size(), output_names_char.data(), 
				output_names_char.size());
			ofLog() << "ran";
			
			if (debug) {
				// Gets the address of the first value
				auto& out = output_values.front();
				// Get tensor shape information
				Ort::TensorTypeAndShapeInfo info = out.GetTensorTypeAndShapeInfo();
				std::vector<int64_t> output_dims = info.GetShape();
				
				// Print the dimensions
				std::cout << "Output tensor dimensions: [";
				for (size_t i = 0; i < output_dims.size(); i++) {
					std::cout << output_dims[i];
					if (i < output_dims.size() - 1) {
						std::cout << ", ";
					}
				}
				std::cout << "]" << std::endl;
				
				// Optional: Print total number of elements
				size_t total_elements = 1;
				for (auto& dim : output_dims) {
					if (dim > 0) {  // Handle dynamic dimensions
						total_elements *= static_cast<size_t>(dim);
					}
				}
				std::cout << "Total elements: " << total_elements << std::endl;
			}
			
			// if (timestamp) {
			// 	auto end = std::chrono::high_resolution_clock::now();
			// 	std::chrono::duration<double, std::milli> elapsed = end - start;
			// 	std::cout << "Update loop took " << elapsed.count() << " ms" << std::endl;
			// }

			return output_values.front().GetTensorMutableData<float>();

		} catch (const Ort::Exception& ex) {
			std::cout << "ERROR running model inference: " << ex.what() << std::endl;
			return dummy_output_tensor.front().GetTensorMutableData<float>();
		}
		
	}

	/*
	*
	*	Utilties (｡･∀･)ﾉﾞ
	*
	*/

	// Add separate methods for float and int32 conversion
    void BaseHandler::convertImageToMatFloat(ofImage* img, std::vector<float>& values, size_t& idx) {
        // Your existing conversion code for float
        ofPixels& pix = img->getPixels();
        int width = img->getWidth();
        int height = img->getHeight();
        int channels = pix.getNumChannels();

        cv::Mat cvImage = cv::Mat(height, width, (channels == 3) ? CV_8UC3 : CV_8UC1, pix.getData());
        cv::InputArray inputArray(cvImage);
        image_array = cv::dnn::blobFromImage(inputArray, 1 / 255.0, cv::Size(input_node_dims[2], input_node_dims[3]), (0, 0, 0), false, false);

        std::memcpy(
            values.data() + idx * channels * width * height,
            image_array.data,
            channels * width * height * sizeof(float)
        );
    }

    void BaseHandler::convertImageToMatInt32(ofImage* img, std::vector<int32_t>& values, size_t& idx) {
        // New conversion code for int32
        ofPixels& pix = img->getPixels();
        int width = img->getWidth();
        int height = img->getHeight();
        int channels = pix.getNumChannels();
        
        cv::Mat cvImage = cv::Mat(height, width, (channels == 3) ? CV_8UC3 : CV_8UC1, pix.getData());
        cv::InputArray inputArray(cvImage);
        cv::Mat intMat = cv::dnn::blobFromImage(inputArray, 1 / 255.0, cv::Size(height, width), (0, 0, 0), false, false);

        intMat.convertTo(image_array, CV_32S);

        std::memcpy(
            values.data() + idx * channels * width * height,
            image_array.data,
            channels * width * height * sizeof(int32_t)
        );
    }

	void BaseHandler::setInputs(std::vector<ofImage*>& in) {
		this->input_imgs = in;
	}

	// Prints the shape of the given tensor (ex. input: (1, 1, 512, 512))
	std::string BaseHandler::PrintShape(const std::vector<int64_t>& v) {
		std::stringstream ss;
		for (std::size_t i = 0; i < v.size() - 1; i++) ss << v[i] << "x";
		ss << v[v.size() - 1];
		return ss.str();
	}

	Ort::Value BaseHandler::GenerateTensor(int batch_size) {
		// Random number generation setup
		std::random_device rd;
		std::mt19937 gen(rd());
		std::uniform_real_distribution<float> dis(0.0f, 255.0f); // Random values between 0 and 255

		// Calculate the total number of elements for a single tensor (without batch dimension) {?, 8} -> 8
		int tensor_size = CalculateProduct(input_node_dims);

		// Create a vector to hold all the values for the batch (8 * (4)batch_size) -> 32
		std::vector<float> batch_values(batch_size * tensor_size); 

		// Fill the batch with random values
		std::generate(batch_values.begin(), batch_values.end(), [&]() {
			return dis(gen);
		});

		// Fill the batch with random values
		std::generate(batch_values.begin(), batch_values.end(), [&]() {
			return dis(gen);
		});

		// Create the batched dimensions by inserting the batch size at the beginning of the original dimensions
		std::vector<int64_t> batched_dims = {  };  // Start with batch size
		batched_dims.insert(batched_dims.end(), input_node_dims.begin(), input_node_dims.end()); // Add the remaining dimensions
		batched_dims[0] = batch_size;

		return VectorToTensor(batch_values, batched_dims);
	}

	int BaseHandler::CalculateProduct(const std::vector<int64_t>& v) {
		int total = 1;
		for (auto& i : v) total *= i;
		return total;
	}

	Ort::Value BaseHandler::VectorToTensor(std::vector<float>& data, const std::vector<int64_t>& shape) {
		// Create a tensor from the provided data, shape, and memory info
		auto tensor = Ort::Value::CreateTensor<float>(memory_info_handler, data.data(), data.size(), shape.data(), shape.size());

		// Return the created tensor
		return tensor;
	}
}