waifu2x-caffe/common/waifu2x.cpp

#include "waifu2x.h"
#include "stImage.h"
#include "cNet.h"
#include <caffe/caffe.hpp>
#include <cudnn.h>
#include <mutex>
#include <opencv2/core.hpp>
#include <opencv2/imgproc.hpp>
#include <tclap/CmdLine.h>
#include <boost/filesystem.hpp>
#include <boost/algorithm/string.hpp>
#include <chrono>
#include <unordered_map>
#include <cuda_runtime.h>

#include <boost/iostreams/stream.hpp>
#include <boost/iostreams/device/file_descriptor.hpp>
#include <msgpack.hpp>

#include <fcntl.h>
#include <zlib.h>
#ifdef _MSC_VER
#include <io.h>
#endif

//#if defined(WIN32) || defined(WIN64)
//#include <Windows.h>
//#endif

#define CV_VERSION_STR CVAUX_STR(CV_MAJOR_VERSION) CVAUX_STR(CV_MINOR_VERSION) CVAUX_STR(CV_SUBMINOR_VERSION)

// <20>r<EFBFBD><72><EFBFBD>h<EFBFBD><68><EFBFBD>[<5B>h
#ifdef _DEBUG
#define CV_EXT_STR "d.lib"
#else
#define CV_EXT_STR ".lib"
#endif

#ifdef _MSC_VER

#pragma comment(lib, "opencv_core" CV_VERSION_STR CV_EXT_STR)
#pragma comment(lib, "opencv_imgcodecs" CV_VERSION_STR CV_EXT_STR)
#pragma comment(lib, "opencv_imgproc" CV_VERSION_STR CV_EXT_STR)
#pragma comment(lib, "IlmImf" CV_EXT_STR)
#pragma comment(lib, "ippicvmt.lib")
#pragma comment(lib, "libjasper" CV_EXT_STR)
#pragma comment(lib, "libjpeg-turbo" CV_EXT_STR)
#pragma comment(lib, "libpng" CV_EXT_STR)
#pragma comment(lib, "libtiff" CV_EXT_STR)
#pragma comment(lib, "libwebp" CV_EXT_STR)
#pragma comment(lib, "zlib" CV_EXT_STR)

#pragma comment(lib, "libopenblas.dll.a")
#pragma comment(lib, "cudart.lib")
#pragma comment(lib, "curand.lib")
#pragma comment(lib, "cublas.lib")
#pragma comment(lib, "cudnn.lib")

#ifdef _DEBUG
#pragma comment(lib, "caffe-d.lib")
#pragma comment(lib, "caffeproto-d.lib")
#pragma comment(lib, "libprotobufd.lib")
#pragma comment(lib, "glogd.lib")
#pragma comment(lib, "gflagsd.lib")
#pragma comment(lib, "libboost_system-vc140-mt-gd-1_61.lib")
#pragma comment(lib, "boost_thread-vc140-mt-gd-1_61.lib")
#pragma comment(lib, "boost_filesystem-vc140-mt-gd-1_61.lib")
#pragma comment(lib, "boost_iostreams-vc140-mt-gd-1_61.lib")
//#pragma comment(lib, "zlibstaticd.lib")


#else
#pragma comment(lib, "caffe.lib")
#pragma comment(lib, "caffeproto.lib")
#pragma comment(lib, "libprotobuf.lib")
#pragma comment(lib, "glog.lib")
#pragma comment(lib, "gflags.lib")
#pragma comment(lib, "libboost_system-vc140-mt-1_61.lib")
#pragma comment(lib, "boost_thread-vc140-mt-1_61.lib")
#pragma comment(lib, "boost_filesystem-vc140-mt-1_61.lib")
#pragma comment(lib, "boost_iostreams-vc140-mt-1_61.lib")
#endif
#endif

const int ScaleBase = 2; // TODO: <20><><EFBFBD>f<EFBFBD><66><EFBFBD>̊g<CC8A>嗦<EFBFBD>ɂ<EFBFBD><C982><EFBFBD><EFBFBD>ĉ•ςł<CF82><C582><EFBFBD><EFBFBD>悤<EFBFBD>ɂ<EFBFBD><C982><EFBFBD>

// <20><><EFBFBD>͉摜<CD89>ɒlj<C992><C789><EFBFBD><EFBFBD><EFBFBD><EFBFBD>p<EFBFBD>f<EFBFBD>B<EFBFBD><42><EFBFBD>O
const int OuterPadding = 0;

// <20>Œ<EFBFBD><C592><EFBFBD><EFBFBD>K<EFBFBD>v<EFBFBD><76>CUDA<44>h<EFBFBD><68><EFBFBD>C<EFBFBD>o<EFBFBD>[<5B>̃o<CC83>[<5B>W<EFBFBD><57><EFBFBD><EFBFBD>
const int MinCudaDriverVersion = 7050;

static std::once_flag waifu2x_once_flag;
static std::once_flag waifu2x_cudnn_once_flag;
static std::once_flag waifu2x_cuda_once_flag;

std::string Waifu2x::ExeDir;


#ifndef CUDA_CHECK_WAIFU2X
#define CUDA_CHECK_WAIFU2X(condition) \
 do { \
    cudaError_t error = condition; \
    if(error != cudaSuccess) throw error; \
 } while (0)
#endif

#define CUDA_HOST_SAFE_FREE(ptr) \
	do { \
		if (ptr) { \
			cudaFreeHost(ptr); \
			ptr = nullptr; \
		} \
	} while (0)

#define SAFE_DELETE_WAIFU2X(ptr) \
	do { \
		if (ptr) { \
			delete [] ptr; \
			ptr = nullptr; \
		} \
	} while (0)

namespace
{
	class IgnoreErrorCV
	{
	private:
		static int handleError(int status, const char* func_name,
			const char* err_msg, const char* file_name,
			int line, void* userdata)
		{
			return 0;
		}

	public:
		IgnoreErrorCV()
		{
			cv::redirectError(handleError);
		}
	};

	IgnoreErrorCV g_IgnoreErrorCV;

	class CudaDeviceSet
	{
	private:
		int orgDevice;
		bool mIsSet;

	public:
		CudaDeviceSet(const std::string &process, const int devno) : orgDevice(0), mIsSet(false)
		{
			if (process == "gpu" || process == "cudnn")
			{
				int count = 0;
				if (cudaGetDeviceCount(&count) != CUDA_SUCCESS)
					return;

				if (devno >= count || count < 0)
					return;

				if (cudaGetDevice(&orgDevice) != CUDA_SUCCESS)
					return;

				if (cudaSetDevice(devno) != CUDA_SUCCESS)
					return;

				mIsSet = true;
			}
		}

		~CudaDeviceSet()
		{
			if (mIsSet)
				cudaSetDevice(orgDevice);
		}
	};
}

class CcuDNNAlgorithmElement
{
private:
	typedef std::unordered_map<uint64_t, uint8_t> AlgoMap;

	AlgoMap mAlgo;
	bool mIsModefy;

	uint8_t kernel_w;
	uint8_t kernel_h;
	uint8_t pad_w;
	uint8_t pad_h;
	uint8_t stride_w;
	uint8_t stride_h;
	uint16_t batch_size;

private:
	static uint64_t InfoToKey(uint16_t num_input, uint16_t num_output, uint16_t width, uint16_t height)
	{
		return (uint64_t)num_input << 8 * 3 | (uint64_t)num_output << 8 * 2 | (uint64_t)width << 8 * 1 | (uint64_t)height << 8 * 0;
	}

public:
	CcuDNNAlgorithmElement() : mIsModefy(false)
	{}
	~CcuDNNAlgorithmElement()
	{}

	void SetLayerData(uint8_t kernel_w, uint8_t kernel_h, uint8_t pad_w, uint8_t pad_h, uint8_t stride_w, uint8_t stride_h, uint16_t batch_size)
	{
		this->kernel_w = kernel_w;
		this->kernel_h = kernel_h;
		this->pad_w = pad_w;
		this->pad_h = pad_h;
		this->stride_w = stride_w;
		this->stride_h = stride_h;
		this->batch_size = batch_size;
	}

	void GetLayerData(uint8_t &kernel_w, uint8_t &kernel_h, uint8_t &pad_w, uint8_t &pad_h, uint8_t &stride_w, uint8_t &stride_h, uint16_t &batch_size)
	{
		kernel_w = this->kernel_w;
		kernel_h = this->kernel_h;
		pad_w = this->pad_w;
		pad_h = this->pad_h;
		stride_w = this->stride_w;
		stride_h = this->stride_h;
		batch_size = this->batch_size;
	}

	int GetAlgorithm(uint16_t num_input, uint16_t num_output, uint16_t width, uint16_t height) const
	{
		const uint64_t key = InfoToKey(num_input, num_output, width, height);
		const auto it = mAlgo.find(key);
		if (it != mAlgo.end())
			return it->second;

		return -1;
	}

	void SetAlgorithm(uint8_t algo, uint16_t num_input, uint16_t num_output, uint16_t width, uint16_t height)
	{
		const uint64_t key = InfoToKey(num_input, num_output, width, height);
		auto it = mAlgo.find(key);
		if (it == mAlgo.end() || it->second != algo)
			mIsModefy = true;

		mAlgo[key] = algo;
	}

	bool IsModefy() const
	{
		return mIsModefy;
	}

	void Saved()
	{
		mIsModefy = false;
	}

	MSGPACK_DEFINE(mAlgo, kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h, batch_size);
};

class CcuDNNAlgorithm
{
private:
	typedef std::unordered_map<uint64_t, CcuDNNAlgorithmElement> AlgoEmlMap;

	AlgoEmlMap mAlgoEmlMap;
	std::string mDataPath;

private:
	static uint64_t InfoToKey(uint8_t kernel_w, uint8_t kernel_h, uint8_t pad_w, uint8_t pad_h, uint8_t stride_w, uint8_t stride_h, uint16_t batch_size)
	{
		return
			(uint64_t)kernel_w << 8 * 7 | (uint64_t)kernel_h << 8 * 6 |
			(uint64_t)pad_w << 8 * 5 | (uint64_t)pad_h << 8 * 4 |
			(uint64_t)stride_w << 8 * 3 | (uint64_t)stride_h << 8 * 2 |
			(uint64_t)batch_size;
	}

	std::string GetDataPath(uint8_t kernel_w, uint8_t kernel_h, uint8_t pad_w, uint8_t pad_h, uint8_t stride_w, uint8_t stride_h, uint16_t batch_size) const
	{
		std::string SavePath = mDataPath;
		SavePath +=
			std::to_string(kernel_w) + "x" + std::to_string(kernel_h) + " " +
			std::to_string(pad_w) + "x" + std::to_string(pad_w) + " " +
			std::to_string(stride_w) + "x" + std::to_string(stride_w) + " " +
			std::to_string(batch_size);
		SavePath += ".dat";

		return SavePath;
	}

	bool Load(uint8_t kernel_w, uint8_t kernel_h, uint8_t pad_w, uint8_t pad_h, uint8_t stride_w, uint8_t stride_h, uint16_t batch_size)
	{
		const std::string SavePath = GetDataPath(kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h, batch_size);

		std::vector<char> sbuf;

		FILE *fp = fopen(SavePath.c_str(), "rb");
		if (!fp)
			return false;

		fseek(fp, 0, SEEK_END);
		const auto size = ftell(fp);
		fseek(fp, 0, SEEK_SET);

		sbuf.resize(size);

		if (fread(sbuf.data(), 1, sbuf.size(), fp) != sbuf.size())
		{
			fclose(fp);
			return false;
		}

		fclose(fp);

		try
		{
			CcuDNNAlgorithmElement elm;
			msgpack::unpack(sbuf.data(), sbuf.size()).get().convert(elm);
			sbuf.clear();

			const uint64_t key = InfoToKey(kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h, batch_size);
			mAlgoEmlMap[key] = std::move(elm);
		}
		catch (...)
		{
			boost::filesystem::remove(SavePath);
		}

		return true;
	}

public:
	CcuDNNAlgorithm()
	{}

	~CcuDNNAlgorithm()
	{
		Save();
	}

	int GetAlgorithm(uint16_t num_input, uint16_t num_output, uint16_t batch_size,
		uint16_t width, uint16_t height, uint16_t kernel_w, uint16_t kernel_h, uint16_t pad_w, uint16_t pad_h, uint16_t stride_w, uint16_t stride_h)
	{
		const uint64_t key = InfoToKey(kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h, batch_size);
		const auto it = mAlgoEmlMap.find(key);
		if (it != mAlgoEmlMap.end())
		{
			const auto &elm = it->second;
			return elm.GetAlgorithm(num_input, num_output, width, height);
		}

		if (Load(kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h, batch_size))
			return mAlgoEmlMap[key].GetAlgorithm(num_input, num_output, width, height);

		return -1;
	}

	void SetAlgorithm(int algo, uint16_t num_input, uint16_t num_output, uint16_t batch_size,
		uint16_t width, uint16_t height, uint16_t kernel_w, uint16_t kernel_h, uint16_t pad_w, uint16_t pad_h, uint16_t stride_w, uint16_t stride_h)
	{
		if (algo < 0 || algo > 255)
			return;

		const uint64_t key = InfoToKey(kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h, batch_size);
		auto &eml = mAlgoEmlMap[key];
		eml.SetAlgorithm(algo, num_input, num_output, width, height);
		eml.SetLayerData(kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h, batch_size);
	}

	void Save()
	{
		for (auto &p : mAlgoEmlMap)
		{
			auto &eml = p.second;
			if (eml.IsModefy())
			{
				try
				{
					msgpack::sbuffer sbuf;
					msgpack::pack(sbuf, eml);

					uint8_t kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h;
					uint16_t batch_size;
					eml.GetLayerData(kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h, batch_size);

					const std::string SavePath = GetDataPath(kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h, batch_size);
					FILE *fp = fopen(SavePath.c_str(), "wb");
					if (fp)
					{
						fwrite(sbuf.data(), 1, sbuf.size(), fp);
						fclose(fp);

						eml.Saved();
					}
				}
				catch(...)
				{}
			}
		}
	}

	void SetDataPath(std::string path)
	{
		mDataPath = path;
	}
};

CcuDNNAlgorithm g_ConvCcuDNNAlgorithm;
CcuDNNAlgorithm g_DeconvCcuDNNAlgorithm;


// CUDA<44><41><EFBFBD>g<EFBFBD><67><EFBFBD>邩<EFBFBD>`<60>F<EFBFBD>b<EFBFBD>N
Waifu2x::eWaifu2xCudaError Waifu2x::can_use_CUDA()
{
	static eWaifu2xCudaError CudaFlag = eWaifu2xCudaError_NotFind;
	std::call_once(waifu2x_cuda_once_flag, [&]()
	{
		int driverVersion = 0;
		if (cudaDriverGetVersion(&driverVersion) == cudaSuccess)
		{
			if (driverVersion > 0)
			{
				int runtimeVersion;
				if (cudaRuntimeGetVersion(&runtimeVersion) == cudaSuccess)
				{
					if (runtimeVersion >= MinCudaDriverVersion && driverVersion >= runtimeVersion)
					{
						cudaDeviceProp prop;
						cudaGetDeviceProperties(&prop, 0);
						if (prop.major >= 3 && prop.minor >= 5 || prop.major >= 4)
							CudaFlag = eWaifu2xCudaError_OK;
						else
							CudaFlag = eWaifu2xCudaError_OldDevice;
					}
					else
						CudaFlag = eWaifu2xCudaError_OldVersion;
				}
				else
					CudaFlag = eWaifu2xCudaError_NotFind;
			}
			else
				CudaFlag = eWaifu2xCudaError_NotFind;
		}
		else
			CudaFlag = eWaifu2xCudaError_NotFind;
	});

	return CudaFlag;
}

// cuDNN<4E><4E><EFBFBD>g<EFBFBD><67><EFBFBD>邩<EFBFBD>`<60>F<EFBFBD>b<EFBFBD>N<EFBFBD>B<EFBFBD><42><EFBFBD><EFBFBD>Windows<77>̂<EFBFBD>
Waifu2x::eWaifu2xcuDNNError Waifu2x::can_use_cuDNN()
{
	static eWaifu2xcuDNNError cuDNNFlag = eWaifu2xcuDNNError_NotFind;
	std::call_once(waifu2x_cudnn_once_flag, [&]()
	{
#if defined(WIN32) || defined(WIN64)
		HMODULE hModule = LoadLibrary(TEXT(CUDNN_DLL_NAME));
		if (hModule != NULL)
		{
			typedef cudnnStatus_t(CUDNNWINAPI* cudnnCreateType)(cudnnHandle_t *);
			typedef cudnnStatus_t(CUDNNWINAPI* cudnnDestroyType)(cudnnHandle_t);
			typedef size_t(CUDNNWINAPI* cudnnGetVersionType)();

			cudnnCreateType cudnnCreateFunc = (cudnnCreateType)GetProcAddress(hModule, "cudnnCreate");
			cudnnDestroyType cudnnDestroyFunc = (cudnnDestroyType)GetProcAddress(hModule, "cudnnDestroy");
			cudnnGetVersionType cudnnGetVersionFunc = (cudnnGetVersionType)GetProcAddress(hModule, "cudnnGetVersion");
			if (cudnnCreateFunc != nullptr && cudnnDestroyFunc != nullptr && cudnnGetVersionFunc != nullptr)
			{
				if (cudnnGetVersionFunc() >= CUDNN_REQUIRE_VERION)
				{
					cudnnHandle_t h;
					if (cudnnCreateFunc(&h) == CUDNN_STATUS_SUCCESS)
					{
						if (cudnnDestroyFunc(h) == CUDNN_STATUS_SUCCESS)
							cuDNNFlag = eWaifu2xcuDNNError_OK;
						else
							cuDNNFlag = eWaifu2xcuDNNError_CannotCreate;
					}
					else
						cuDNNFlag = eWaifu2xcuDNNError_CannotCreate;
				}
				else
					cuDNNFlag = eWaifu2xcuDNNError_OldVersion;
			}
			else
				cuDNNFlag = eWaifu2xcuDNNError_NotFind;

			FreeLibrary(hModule);
		}
#endif
	});

	return cuDNNFlag;
}

void Waifu2x::init_liblary(int argc, char** argv)
{
	if (argc > 0)
		ExeDir = argv[0];

	std::call_once(waifu2x_once_flag, [argc, argv]()
	{
		assert(argc >= 1);

		int tmpargc = 1;
		char* tmpargvv[] = {argv[0]};
		char** tmpargv = tmpargvv;
		// glog<6F><67><EFBFBD>̏<EFBFBD><CC8F><EFBFBD><EFBFBD><EFBFBD>
		caffe::GlobalInit(&tmpargc, &tmpargv);
	});
}

void Waifu2x::quit_liblary()
{
	g_ConvCcuDNNAlgorithm.Save();
	g_DeconvCcuDNNAlgorithm.Save();

	//caffe::GlobalFinalize();
}

void Waifu2x::quit_thread_liblary()
{
	//caffe::ThreadFinalize();
}

Waifu2x::Waifu2x() : mIsInited(false), mNoiseLevel(0), mIsCuda(false), mOutputBlock(nullptr), mOutputBlockSize(0), mGPUNo(0)
{}

Waifu2x::~Waifu2x()
{
	Destroy();
}

Waifu2x::eWaifu2xError Waifu2x::Init(const eWaifu2xModelType mode, const int noise_level,
	const boost::filesystem::path &model_dir, const std::string &process, const int GPUNo)
{
	Waifu2x::eWaifu2xError ret;

	if (mIsInited)
		return Waifu2x::eWaifu2xError_OK;

	try
	{
		std::string Process = process;
		const auto cuDNNCheckStartTime = std::chrono::system_clock::now();

		if (Process == "gpu")
		{
			if (can_use_CUDA() != eWaifu2xCudaError_OK)
				return Waifu2x::eWaifu2xError_FailedCudaCheck;
			// cuDNN<4E><4E><EFBFBD>g<EFBFBD><67><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>Ȃ<EFBFBD>cuDNN<4E><4E><EFBFBD>g<EFBFBD><67>
			else if (can_use_cuDNN() == eWaifu2xcuDNNError_OK)
				Process = "cudnn";
		}

		mMode = mode;
		mNoiseLevel = noise_level;
		mProcess = Process;
		mGPUNo = GPUNo;

		const auto cuDNNCheckEndTime = std::chrono::system_clock::now();

		if (Process == "cudnn")
		{
			// exe<78>̃f<CC83>B<EFBFBD><42><EFBFBD>N<EFBFBD>g<EFBFBD><67><EFBFBD><EFBFBD>cuDNN<4E>̃A<CC83><41><EFBFBD>S<EFBFBD><53><EFBFBD>Y<EFBFBD><59><EFBFBD>f<EFBFBD>[<5B>^<5E>ۑ<EFBFBD>
			boost::filesystem::path cudnn_data_base_dir_path(ExeDir);
			if (cudnn_data_base_dir_path.is_relative())
				cudnn_data_base_dir_path = boost::filesystem::system_complete(cudnn_data_base_dir_path);

			if (!boost::filesystem::is_directory(cudnn_data_base_dir_path))
				cudnn_data_base_dir_path = cudnn_data_base_dir_path.branch_path();

			if (!boost::filesystem::exists(cudnn_data_base_dir_path))
			{
				// exe<78>̃f<CC83>B<EFBFBD><42><EFBFBD>N<EFBFBD>g<EFBFBD><67><EFBFBD><EFBFBD><EFBFBD>擾<EFBFBD>ł<EFBFBD><C582>Ȃ<EFBFBD><C882><EFBFBD><EFBFBD>΃J<CE83><4A><EFBFBD><EFBFBD><EFBFBD>g<EFBFBD>f<EFBFBD>B<EFBFBD><42><EFBFBD>N<EFBFBD>g<EFBFBD><67><EFBFBD>ɕۑ<C995>

				cudnn_data_base_dir_path = boost::filesystem::current_path();

				if (cudnn_data_base_dir_path.is_relative())
					cudnn_data_base_dir_path = boost::filesystem::system_complete(cudnn_data_base_dir_path);

				if (!boost::filesystem::exists(cudnn_data_base_dir_path))
					cudnn_data_base_dir_path = "./";
			}

			if (boost::filesystem::exists(cudnn_data_base_dir_path))
			{
				const boost::filesystem::path cudnn_data_dir_path(cudnn_data_base_dir_path / "cudnn_data");

				bool isOK = false;
				if (boost::filesystem::exists(cudnn_data_dir_path))
					isOK = true;

				if (!isOK)
				{
					boost::system::error_code error;
					const bool result = boost::filesystem::create_directory(cudnn_data_dir_path, error);
					if (result && !error)
						isOK = true;
				}

				if (isOK)
				{
					cudaDeviceProp prop;
					if (cudaGetDeviceProperties(&prop, mGPUNo) == cudaSuccess)
					{
						std::string conv_filename(prop.name);
						conv_filename += " conv ";

						std::string deconv_filename(prop.name);
						deconv_filename += " deconv ";

						const boost::filesystem::path conv_data_path = cudnn_data_dir_path / conv_filename;
						const boost::filesystem::path deconv_data_path = cudnn_data_dir_path / deconv_filename;

						g_ConvCcuDNNAlgorithm.SetDataPath(conv_data_path.string());
						g_DeconvCcuDNNAlgorithm.SetDataPath(deconv_data_path.string());
					}
				}
			}
		}

		const boost::filesystem::path mode_dir_path(GetModeDirPath(model_dir));
		if (!boost::filesystem::exists(mode_dir_path))
			return Waifu2x::eWaifu2xError_FailedOpenModelFile;

		CudaDeviceSet devset(process, mGPUNo);

		if (mProcess == "cpu")
		{
			caffe::Caffe::set_mode(caffe::Caffe::CPU);
			mIsCuda = false;
		}
		else
		{
			caffe::Caffe::set_mode(caffe::Caffe::GPU);
			mIsCuda = true;
		}

		caffe::Caffe::SetGetcuDNNAlgorithmFunc(GetcuDNNAlgorithm);
		caffe::Caffe::SetSetcuDNNAlgorithmFunc(SetcuDNNAlgorithm);

		mInputPlane = 0;
		mMaxNetOffset = 0;

		const boost::filesystem::path info_path = GetInfoPath(mode_dir_path);

		stInfo info;
		ret = cNet::GetInfo(info_path, info);
		if (ret != Waifu2x::eWaifu2xError_OK)
			return ret;

		mHasNoiseScaleOnly = info.has_noise_scale;
		mInputPlane = info.channels;

		if (mode == eWaifu2xModelTypeNoise && info.has_noise_only) // <20>m<EFBFBD>C<EFBFBD>Y<EFBFBD><59><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>ƒm<C283>C<EFBFBD>Y<EFBFBD><59><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>f<EFBFBD><66><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>݂<EFBFBD><DD82><EFBFBD><EFBFBD>̂ł<CC82><C582><EFBFBD><EFBFBD>΃m<CE83>C<EFBFBD>Y<EFBFBD><59><EFBFBD><EFBFBD><EFBFBD>X<EFBFBD>P<EFBFBD>[<5B><><EFBFBD><EFBFBD><EFBFBD>f<EFBFBD><66><EFBFBD>͎g<CD8E><67><EFBFBD>Ȃ<EFBFBD><C882>悤<EFBFBD>ɂ<EFBFBD><C982><EFBFBD>
			mHasNoiseScaleOnly = false;

		if (mode == eWaifu2xModelTypeNoise || mode == eWaifu2xModelTypeNoiseScale || mode == eWaifu2xModelTypeAutoScale)
		{
			std::string base_name;

			mNoiseNet.reset(new cNet);

			eWaifu2xModelType Mode = mode;
			if (mHasNoiseScaleOnly) // <20>m<EFBFBD>C<EFBFBD>Y<EFBFBD><59><EFBFBD><EFBFBD><EFBFBD>Ɗg<C68A><67><EFBFBD>𓯎<EFBFBD><F093AF8E>ɍs<C98D><73>
			{
				// <20>m<EFBFBD>C<EFBFBD>Y<EFBFBD><59><EFBFBD><EFBFBD><EFBFBD>g<EFBFBD><67><EFBFBD>l<EFBFBD>b<EFBFBD>g<EFBFBD>̍\<5C>z<EFBFBD><7A>eWaifu2xModelTypeNoiseScale<6C><65><EFBFBD>w<EFBFBD>肷<EFBFBD><E882B7><EFBFBD>K<EFBFBD>v<EFBFBD><76><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
				Mode = eWaifu2xModelTypeNoiseScale;
				base_name = "noise" + std::to_string(noise_level) + "_scale2.0x_model";
			}
			else // <20>m<EFBFBD>C<EFBFBD>Y<EFBFBD><59><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
			{
				Mode = eWaifu2xModelTypeNoise;
				base_name = "noise" + std::to_string(noise_level) + "_model";
			}

			const boost::filesystem::path model_path = mode_dir_path / (base_name + ".prototxt");
			const boost::filesystem::path param_path = mode_dir_path / (base_name + ".json");

			ret = mNoiseNet->ConstractNet(Mode, model_path, param_path, info, mProcess);
			if (ret != Waifu2x::eWaifu2xError_OK)
				return ret;

			mMaxNetOffset = mNoiseNet->GetNetOffset();
		}

		// <20>g<EFBFBD>傪<EFBFBD>K<EFBFBD>v<EFBFBD>ȏꍇ<C88F>̓<EFBFBD><CD83>`<60><><EFBFBD><EFBFBD><EFBFBD>l<EFBFBD><6C><EFBFBD>̊g<CC8A><67><EFBFBD>̂<EFBFBD><CC82>߂<EFBFBD>mScaleNet<65><74><EFBFBD>\<5C>z<EFBFBD><7A><EFBFBD><EFBFBD><EFBFBD>K<EFBFBD>v<EFBFBD><76><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
		if (mode == eWaifu2xModelTypeScale || mode == eWaifu2xModelTypeNoiseScale || mode == eWaifu2xModelTypeAutoScale)
		{
			const std::string base_name = "scale2.0x_model";

			const boost::filesystem::path model_path = mode_dir_path / (base_name + ".prototxt");
			const boost::filesystem::path param_path = mode_dir_path / (base_name + ".json");

			mScaleNet.reset(new cNet);

			ret = mScaleNet->ConstractNet(eWaifu2xModelTypeScale, model_path, param_path, info, mProcess);
			if (ret != Waifu2x::eWaifu2xError_OK)
				return ret;

			assert(mInputPlane == 0 || mInputPlane == mScaleNet->GetInputPlane());

			mMaxNetOffset = std::max(mScaleNet->GetNetOffset(), mMaxNetOffset);
		}
		else
		{
			
		}

		mIsInited = true;
	}
	catch (...)
	{
		return Waifu2x::eWaifu2xError_InvalidParameter;
	}

	return Waifu2x::eWaifu2xError_OK;
}

boost::filesystem::path Waifu2x::GetModeDirPath(const boost::filesystem::path &model_dir)
{
	boost::filesystem::path mode_dir_path(model_dir);
	if (!mode_dir_path.is_absolute()) // model_dir<69><72><EFBFBD><EFBFBD><EFBFBD>΃p<CE83>X<EFBFBD>Ȃ<EFBFBD><C882><EFBFBD><EFBFBD>΃p<CE83>X<EFBFBD>ɒ<EFBFBD><C992><EFBFBD>
	{
		// <20>܂<EFBFBD><DC82>̓J<CD83><4A><EFBFBD><EFBFBD><EFBFBD>g<EFBFBD>f<EFBFBD>B<EFBFBD><42><EFBFBD>N<EFBFBD>g<EFBFBD><67><EFBFBD><EFBFBD><EFBFBD>ɂ<EFBFBD><C982>邩<EFBFBD>T<EFBFBD><54>
		mode_dir_path = boost::filesystem::absolute(model_dir);
		if (!boost::filesystem::exists(mode_dir_path) && !ExeDir.empty()) // <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>argv[0]<5D><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>s<EFBFBD>t<EFBFBD>@<40>C<EFBFBD><43><EFBFBD>̂<EFBFBD><CC82><EFBFBD><EFBFBD>t<EFBFBD>H<EFBFBD><48><EFBFBD>_<EFBFBD>𐄒肵<F0908492>A<EFBFBD><41><EFBFBD>̃t<CC83>H<EFBFBD><48><EFBFBD>_<EFBFBD><5F><EFBFBD>ɂ<EFBFBD><C982>邩<EFBFBD>T<EFBFBD><54>
		{
			boost::filesystem::path a0(ExeDir);
			if (a0.is_absolute())
				mode_dir_path = a0.branch_path() / model_dir;
		}
	}

	return mode_dir_path;
}

boost::filesystem::path Waifu2x::GetInfoPath(const boost::filesystem::path &mode_dir_path)
{
	const boost::filesystem::path info_path = mode_dir_path / "info.json";

	return info_path;
}

Waifu2x::eWaifu2xError Waifu2x::waifu2x(const boost::filesystem::path &input_file, const boost::filesystem::path &output_file,
	const boost::optional<double> scale_ratio, const boost::optional<int> scale_width, const boost::optional<int> scale_height, 
	const waifu2xCancelFunc cancel_func, const int crop_w, const int crop_h,
	const boost::optional<int> output_quality, const int output_depth, const bool use_tta,
	const int batch_size)
{
	Waifu2x::eWaifu2xError ret;

	if (!mIsInited)
		return Waifu2x::eWaifu2xError_NotInitialized;

	stImage image;
	ret = image.Load(input_file);
	if (ret != Waifu2x::eWaifu2xError_OK)
		return ret;

	image.Preprocess(mInputPlane, mMaxNetOffset);

	const bool isReconstructNoise = mMode == eWaifu2xModelTypeNoise || mMode == eWaifu2xModelTypeNoiseScale || (mMode == eWaifu2xModelTypeAutoScale && image.RequestDenoise());
	const bool isReconstructScale = mMode == eWaifu2xModelTypeScale || mMode == eWaifu2xModelTypeNoiseScale || mMode == eWaifu2xModelTypeAutoScale;

	auto factor = CalcScaleRatio(scale_ratio, scale_width, scale_height, image);

	if (!isReconstructScale)
		factor = Factor(1.0, 1.0);

	cv::Mat reconstruct_image;
	ret = ReconstructImage(factor, crop_w, crop_h, use_tta, batch_size, isReconstructNoise, isReconstructScale, cancel_func, image);
	if (ret != Waifu2x::eWaifu2xError_OK)
		return ret;

	if(!scale_width || !scale_height)
		image.Postprocess(mInputPlane, factor, output_depth);
	else
		image.Postprocess(mInputPlane, *scale_width, *scale_height, output_depth);

	ret = image.Save(output_file, output_quality);
	if (ret != Waifu2x::eWaifu2xError_OK)
		return ret;

	return Waifu2x::eWaifu2xError_OK;
}

Waifu2x::eWaifu2xError Waifu2x::waifu2x(const double factor, const void* source, void* dest, const int width, const int height,
	const int in_channel, const int in_stride, const int out_channel, const int out_stride,
	const int crop_w, const int crop_h, const bool use_tta, const int batch_size)
{
	Waifu2x::eWaifu2xError ret;

	if (!mIsInited)
		return Waifu2x::eWaifu2xError_NotInitialized;

	int cvrSetting = -1;
	if (in_channel == 3 && out_channel == 3)
		cvrSetting = CV_BGR2RGB;
	else if (in_channel == 4 && out_channel == 4)
		cvrSetting = CV_BGRA2RGBA;
	else if (in_channel == 3 && out_channel == 4)
		cvrSetting = CV_BGR2RGBA;
	else if (in_channel == 4 && out_channel == 3)
		cvrSetting = CV_BGRA2RGB;
	else if (!(in_channel == 1 && out_channel == 1))
		return Waifu2x::eWaifu2xError_InvalidParameter;

	stImage image;
	ret = image.Load(source, width, height, in_channel, in_stride);
	if (ret != Waifu2x::eWaifu2xError_OK)
		return ret;

	image.Preprocess(mInputPlane, mMaxNetOffset);

	const bool isReconstructNoise = mMode == eWaifu2xModelTypeNoise || mMode == eWaifu2xModelTypeNoiseScale;
	const bool isReconstructScale = mMode == eWaifu2xModelTypeScale || mMode == eWaifu2xModelTypeNoiseScale || mMode == eWaifu2xModelTypeAutoScale;

	Factor nowFactor = Factor(factor, 1.0);
	if (!isReconstructScale)
		nowFactor = Factor(1.0, 1.0);

	cv::Mat reconstruct_image;
	ret = ReconstructImage(nowFactor, crop_w, crop_h, use_tta, batch_size, isReconstructNoise, isReconstructScale, nullptr, image);
	if (ret != Waifu2x::eWaifu2xError_OK)
		return ret;

	image.Postprocess(mInputPlane, nowFactor, 8);

	cv::Mat out_bgr_image = image.GetEndImage();
	image.Clear();

	cv::Mat out_image;
	if (cvrSetting >= 0)
		cv::cvtColor(out_bgr_image, out_image, cvrSetting); // BGR<47><52><EFBFBD><EFBFBD>RGB<47>ɖ߂<C996>
	else
		out_image = out_bgr_image;
	out_bgr_image.release();

	// <20>o<EFBFBD>͔z<CD94><7A><EFBFBD>֏<EFBFBD><D68F><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
	{
		const auto width = out_image.size().width;
		const auto stride = out_image.step1();
		for (int i = 0; i < out_image.size().height; i++)
			memcpy((uint8_t *)dest + out_stride * i, out_image.data + stride * i, out_stride);
	}

	return Waifu2x::eWaifu2xError_OK;
}

Factor Waifu2x::CalcScaleRatio(const boost::optional<double> scale_ratio, const boost::optional<int> scale_width, const boost::optional<int> scale_height,
	const stImage &image)
{
	if (scale_ratio)
		return Factor(*scale_ratio, 1.0);

	if (scale_width && scale_height)
	{
		const auto d1 = image.GetScaleFromWidth(*scale_width);
		const auto d2 = image.GetScaleFromWidth(*scale_height);

		return d1.toDouble() >= d2.toDouble() ? d1 : d2;
	}

	if (scale_width)
		return image.GetScaleFromWidth(*scale_width);

	if(scale_height)
		return image.GetScaleFromHeight(*scale_height);

	return Factor(1.0, 1.0);
}

int Waifu2x::GetcuDNNAlgorithm(const char * layer_name, int num_input, int num_output, int batch_size,
	int width, int height, int kernel_w, int kernel_h, int pad_w, int pad_h, int stride_w, int stride_h)
{
	// g_ConvCcuDNNAlgorithm<68><6D>g_DeconvCcuDNNAlgorithm<68><6D><EFBFBD>t<EFBFBD>ɂȂ<C982><C882>Ă<EFBFBD><C482>܂<EFBFBD><DC82>Ă<EFBFBD><C482>邪<EFBFBD>A<EFBFBD>t<EFBFBD>@<40>C<EFBFBD><43><EFBFBD><EFBFBD><EFBFBD>ɂ<EFBFBD><C982><EFBFBD><EFBFBD>e<EFBFBD><65><EFBFBD><EFBFBD><EFBFBD>Ȃ<EFBFBD><C882>̂ƌ݊<C68C><DD8A><EFBFBD><EFBFBD><EFBFBD><EFBFBD>Ȃ<EFBFBD><C882>Ȃ<EFBFBD><C882>̂ł<CC82><C582>̂܂܎d<DC8E>l<EFBFBD>Ƃ<EFBFBD><C682><EFBFBD>
	if (strcmp(layer_name, "Deconvolution") == 0)
		return g_ConvCcuDNNAlgorithm.GetAlgorithm(num_input, num_output, batch_size, width, height, kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h);
	else if (strcmp(layer_name, "Convolution") == 0)
		return g_DeconvCcuDNNAlgorithm.GetAlgorithm(num_input, num_output, batch_size, width, height, kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h);

	return -1;
}

void Waifu2x::SetcuDNNAlgorithm(int algo, const char * layer_name, int num_input, int num_output, int batch_size,
	int width, int height, int kernel_w, int kernel_h, int pad_w, int pad_h, int stride_w, int stride_h)
{
	// g_ConvCcuDNNAlgorithm<68><6D>g_DeconvCcuDNNAlgorithm<68><6D><EFBFBD>t<EFBFBD>ɂȂ<C982><C882>Ă<EFBFBD><C482>܂<EFBFBD><DC82>Ă<EFBFBD><C482>邪<EFBFBD>A<EFBFBD>t<EFBFBD>@<40>C<EFBFBD><43><EFBFBD><EFBFBD><EFBFBD>ɂ<EFBFBD><C982><EFBFBD><EFBFBD>e<EFBFBD><65><EFBFBD><EFBFBD><EFBFBD>Ȃ<EFBFBD><C882>̂ƌ݊<C68C><DD8A><EFBFBD><EFBFBD><EFBFBD><EFBFBD>Ȃ<EFBFBD><C882>Ȃ<EFBFBD><C882>̂ł<CC82><C582>̂܂܎d<DC8E>l<EFBFBD>Ƃ<EFBFBD><C682><EFBFBD>
	if (strcmp(layer_name, "Deconvolution") == 0)
		return g_ConvCcuDNNAlgorithm.SetAlgorithm(algo, num_input, num_output, batch_size, width, height, kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h);
	else if (strcmp(layer_name, "Convolution") == 0)
		return g_DeconvCcuDNNAlgorithm.SetAlgorithm(algo, num_input, num_output, batch_size, width, height, kernel_w, kernel_h, pad_w, pad_h, stride_w, stride_h);
}

Waifu2x::eWaifu2xError Waifu2x::ReconstructImage(const Factor factor, const int crop_w, const int crop_h, const bool use_tta, const int batch_size,
	const bool isReconstructNoise, const bool isReconstructScale, const Waifu2x::waifu2xCancelFunc cancel_func, stImage &image)
{
	Waifu2x::eWaifu2xError ret;

	Factor nowFactor = factor;

	if (isReconstructNoise)
	{
		if (!mHasNoiseScaleOnly) // <20>m<EFBFBD>C<EFBFBD>Y<EFBFBD><59><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
		{
			cv::Mat im;
			cv::Size_<int> size;
			image.GetScalePaddingedRGB(im, size, mNoiseNet->GetNetOffset(), OuterPadding, crop_w, crop_h, 1);

			ret = ReconstructByNet(mNoiseNet, crop_w, crop_h, use_tta, batch_size, cancel_func, im);
			if (ret != Waifu2x::eWaifu2xError_OK)
				return ret;

			image.SetReconstructedRGB(im, size, 1);
		}
		else // <20>m<EFBFBD>C<EFBFBD>Y<EFBFBD><59><EFBFBD><EFBFBD><EFBFBD>Ɗg<C68A><67>
		{
			ret = ReconstructNoiseScale(crop_w, crop_h, use_tta, batch_size, cancel_func, image);
			if (ret != Waifu2x::eWaifu2xError_OK)
				return ret;

			//nowFactor /= mNoiseNet->GetInnerScale();
			nowFactor = nowFactor.MultiDenominator(mNoiseNet->GetInnerScale());
		}
	}

	if (cancel_func && cancel_func())
		return Waifu2x::eWaifu2xError_Cancel;

	const int scaleNum = ceil(log(nowFactor.toDouble()) / log(ScaleBase));

	if (isReconstructScale)
	{
		for (int i = 0; i < scaleNum; i++)
		{
			ret = ReconstructScale(crop_w, crop_h, use_tta, batch_size, cancel_func, image);
			if (ret != Waifu2x::eWaifu2xError_OK)
				return ret;
		}
	}

	return Waifu2x::eWaifu2xError_OK;
}

Waifu2x::eWaifu2xError Waifu2x::ReconstructScale(const int crop_w, const int crop_h, const bool use_tta, const int batch_size,
	const Waifu2x::waifu2xCancelFunc cancel_func, stImage &image)
{
	Waifu2x::eWaifu2xError ret;

	if (image.HasAlpha())
	{
		cv::Mat im;
		cv::Size_<int> size;
		image.GetScalePaddingedA(im, size, mScaleNet->GetNetOffset(), OuterPadding, crop_w, crop_h, mScaleNet->GetScale() / mScaleNet->GetInnerScale());

		ret = ReconstructByNet(mScaleNet, crop_w, crop_h, use_tta, batch_size, cancel_func, im);
		if (ret != Waifu2x::eWaifu2xError_OK)
			return ret;

		image.SetReconstructedA(im, size, mScaleNet->GetInnerScale());
	}

	cv::Mat im;
	cv::Size_<int> size;
	image.GetScalePaddingedRGB(im, size, mScaleNet->GetNetOffset(), OuterPadding, crop_w, crop_h, mScaleNet->GetScale() / mScaleNet->GetInnerScale());

	ret = ReconstructByNet(mScaleNet, crop_w, crop_h, use_tta, batch_size, cancel_func, im);
	if (ret != Waifu2x::eWaifu2xError_OK)
		return ret;

	image.SetReconstructedRGB(im, size, mScaleNet->GetInnerScale());

	return Waifu2x::eWaifu2xError_OK;
}

Waifu2x::eWaifu2xError Waifu2x::ReconstructNoiseScale(const int crop_w, const int crop_h, const bool use_tta, const int batch_size,
	const Waifu2x::waifu2xCancelFunc cancel_func, stImage &image)
{
	Waifu2x::eWaifu2xError ret;

	if (image.HasAlpha())
	{
		// <20><><EFBFBD>`<60><><EFBFBD><EFBFBD><EFBFBD>l<EFBFBD><6C><EFBFBD>ɂ̓m<CD83>C<EFBFBD>Y<EFBFBD><59><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>s<EFBFBD><73><EFBFBD>Ȃ<EFBFBD>

		cv::Mat im;
		cv::Size_<int> size;
		image.GetScalePaddingedA(im, size, mScaleNet->GetNetOffset(), OuterPadding, crop_w, crop_h, mScaleNet->GetScale() / mScaleNet->GetInnerScale());

		ret = ReconstructByNet(mScaleNet, crop_w, crop_h, use_tta, batch_size, cancel_func, im);
		if (ret != Waifu2x::eWaifu2xError_OK)
			return ret;

		image.SetReconstructedA(im, size, mScaleNet->GetInnerScale());
	}

	cv::Mat im;
	cv::Size_<int> size;
	image.GetScalePaddingedRGB(im, size, mNoiseNet->GetNetOffset(), OuterPadding, crop_w, crop_h, mNoiseNet->GetScale() / mNoiseNet->GetInnerScale());

	ret = ReconstructByNet(mNoiseNet, crop_w, crop_h, use_tta, batch_size, cancel_func, im);
	if (ret != Waifu2x::eWaifu2xError_OK)
		return ret;

	image.SetReconstructedRGB(im, size, mNoiseNet->GetInnerScale());

	return Waifu2x::eWaifu2xError_OK;
}


Waifu2x::eWaifu2xError Waifu2x::ReconstructByNet(std::shared_ptr<cNet> net, const int crop_w, const int crop_h, const bool use_tta, const int batch_size,
	const Waifu2x::waifu2xCancelFunc cancel_func, cv::Mat &im)
{
	Waifu2x::eWaifu2xError ret;

	if (!use_tta) // <20><><EFBFBD>ʂɏ<CA82><C98F><EFBFBD>
	{
		ret = ProcessNet(net, crop_w, crop_h, use_tta, batch_size, im);
		if (ret != Waifu2x::eWaifu2xError_OK)
			return ret;
	}
	else // Test-Time Augmentation Mode
	{
		const auto RotateClockwise90 = [](cv::Mat &mat)
		{
			cv::transpose(mat, mat);
			cv::flip(mat, mat, 1);
		};

		const auto RotateClockwise90N = [RotateClockwise90](cv::Mat &mat, const int rotateNum)
		{
			for (int i = 0; i < rotateNum; i++)
				RotateClockwise90(mat);
		};

		const auto RotateCounterclockwise90 = [](cv::Mat &mat)
		{
			cv::transpose(mat, mat);
			cv::flip(mat, mat, 0);
		};

		const auto RotateCounterclockwise90N = [RotateCounterclockwise90](cv::Mat &mat, const int rotateNum)
		{
			for (int i = 0; i < rotateNum; i++)
				RotateCounterclockwise90(mat);
		};

		cv::Mat reconstruct_image;
		for (int i = 0; i < 8; i++)
		{
			cv::Mat in(im.clone());

			const int rotateNum = i % 4;
			RotateClockwise90N(in, rotateNum);

			if (i >= 4)
				cv::flip(in, in, 1); // <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>]

			const int cw = (rotateNum % 2 == 0) ? crop_w : crop_h;
			const int ch = (rotateNum % 2 == 0) ? crop_h : crop_w;

			ret = ProcessNet(net, cw, ch, use_tta, batch_size, in);
			if (ret != Waifu2x::eWaifu2xError_OK)
				return ret;

			if (i >= 4)
				cv::flip(in, in, 1); // <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>]

			RotateCounterclockwise90N(in, rotateNum);

			if (i == 0)
				reconstruct_image = in;
			else
				reconstruct_image += in;
		}

		reconstruct_image /= 8.0;

		im = reconstruct_image;
	}

	return Waifu2x::eWaifu2xError_OK;
}

Waifu2x::eWaifu2xError Waifu2x::ProcessNet(std::shared_ptr<cNet> net, const int crop_w, const int crop_h, const bool use_tta, const int batch_size, cv::Mat &im)
{
	Waifu2x::eWaifu2xError ret;

	CudaDeviceSet devset(mProcess, mGPUNo);

	const auto OutputMemorySize = net->GetOutputMemorySize(crop_w, crop_h, OuterPadding, batch_size);
	if (OutputMemorySize > mOutputBlockSize)
	{
		if (mIsCuda)
		{
			CUDA_HOST_SAFE_FREE(mOutputBlock);
			CUDA_CHECK_WAIFU2X(cudaHostAlloc(&mOutputBlock, OutputMemorySize, cudaHostAllocDefault));
		}
		else
		{
			SAFE_DELETE_WAIFU2X(mOutputBlock);
			mOutputBlock = new float[OutputMemorySize];
		}

		mOutputBlockSize = OutputMemorySize;
	}

	ret = net->ReconstructImage(use_tta, crop_w, crop_h, OuterPadding, batch_size, mOutputBlock, im, im);
	if (ret != Waifu2x::eWaifu2xError_OK)
		return ret;

	return Waifu2x::eWaifu2xError_OK;
}

void Waifu2x::Destroy()
{
	CudaDeviceSet devset(mProcess, mGPUNo);

	mNoiseNet.reset();
	mScaleNet.reset();

	if (mIsCuda)
	{
		CUDA_HOST_SAFE_FREE(mOutputBlock);
	}
	else
	{
		SAFE_DELETE_WAIFU2X(mOutputBlock);
	}

	mIsInited = false;
}

const std::string& Waifu2x::used_process() const
{
	return mProcess;
}

std::string Waifu2x::GetModelName(const boost::filesystem::path & model_dir)
{
	const boost::filesystem::path mode_dir_path(GetModeDirPath(model_dir));
	if (!boost::filesystem::exists(mode_dir_path))
		return std::string();

	const boost::filesystem::path info_path = mode_dir_path / "info.json";

	return cNet::GetModelName(info_path);
}

bool Waifu2x::GetInfo(const boost::filesystem::path &model_dir, stInfo &info)
{
	const boost::filesystem::path mode_dir_path(GetModeDirPath(model_dir));
	if (!boost::filesystem::exists(mode_dir_path))
		return false;

	const boost::filesystem::path info_path = mode_dir_path / "info.json";

	return cNet::GetInfo(info_path, info) == Waifu2x::eWaifu2xError_OK;
}