#pragma once
#include <cstddef>
#include <cstdint>
#include <type_traits>
#include <bit>
#include <stdexcept>
#include <string>
#include <array>
#include <memory>

namespace homcert::bfv {

constexpr std::size_t VECSIZE = 8192;
constexpr std::size_t PLAINMOD = 16760833; // 8192*2*1023+1 and prime

template<std::size_t N, typename = void>
struct smallest_uint;

template<std::size_t N>
struct smallest_uint<N, std::enable_if_t<(N <= UINT8_MAX)>> {
	using type = std::uint8_t;
};

template<std::size_t N>
struct smallest_uint<N, std::enable_if_t<(UINT8_MAX < N && N <= UINT16_MAX)>> {
	using type = std::uint16_t;
};

template<std::size_t N>
struct smallest_uint<N, std::enable_if_t<(UINT16_MAX < N && N <= UINT32_MAX)>> {
	using type = std::uint32_t;
};

template<std::size_t N>
struct smallest_uint<N, std::enable_if_t<(UINT32_MAX < N)>> {
	using type = std::uint64_t;
};

struct plaintext {
	using coefficient = typename smallest_uint<PLAINMOD>::type;

	std::array<coefficient, VECSIZE> coefficients;

	// Arithmetic...
};

using ciphertext = int;

// Exception for BFV errors in context

class bfv_exception : public std::exception {
public:
	enum class REASON {
		CTX_INVALID,
		CTX_NOT_IMPLEMENTED,
		CTX_NO_PRIVATE,
		CTX_NO_PUBLIC,
		CTX_ARITHMETIC,
		CTX_MEMORY,
		CTX_INVALID_ARGUMENT,
		CTX_COMPONENT,
		CTX_INTERNAL
	};
private:
	REASON m_reason;
	std::string m_message;
public:
	bfv_exception(REASON reason, std::string message);
	REASON reason() const noexcept;
	const char* what() const noexcept override;
};

// Inheriting from context class allows pluggability of implementations (e.g. SEAL vs. GPU)

struct context {
	static constexpr int PUBLIC_COMPONENT = 1;
	static constexpr int PRIVATE_COMPONENT = 2;

	context(const context& other) = delete;
	context(context&& other) = delete;
	context& operator=(const context& other) = delete;
	context& operator=(context&& other) = delete;

	context() = default;
	virtual ~context() = default;

	virtual void new_components() = 0; // context default constructs empty
	virtual void has_components(int& components) const = 0;
	virtual void clone_components(int components, std::shared_ptr<context>& ptr) const = 0;
	// virtual void dump_components(int components, void* buf, std::size_t& n) const = 0;
	// virtual void load_components(int components, const void* buf, std::size_t n) = 0;

	virtual void allocate(ciphertext& ct) = 0;
	virtual void free(ciphertext ct) = 0;
	virtual void serialize(ciphertext ct, void* buf, std::size_t& n) const = 0;
	virtual void deserialize(ciphertext ct, const void* buf, std::size_t n) const = 0;
	virtual void encrypt(const plaintext& pt, ciphertext ct) const = 0;
	virtual void decrypt(ciphertext ct, plaintext& pt) const = 0; // private
	virtual void add_cipher_plain(ciphertext a, const plaintext& b, ciphertext res) const = 0;
	virtual void add_cipher_cipher(ciphertext a, ciphertext b, ciphertext res) const = 0;
	virtual void sub_cipher_plain(ciphertext a, const plaintext& b, ciphertext res) const = 0;
	virtual void sub_cipher_cipher(ciphertext a, ciphertext b, ciphertext res) const = 0;
	virtual void mul_cipher_plain(ciphertext a, const plaintext& b, ciphertext res) const = 0;
	virtual void mul_cipher_cipher(ciphertext a, ciphertext b, ciphertext res) const = 0;
	virtual void rot_cipher_rows(ciphertext ct, int r, ciphertext res) const = 0;
	virtual void swap_cipher_rows(ciphertext ct, ciphertext res) const = 0;
	virtual void noise_budget(ciphertext ct, std::size_t& budget) const = 0; // private
};

/*
NOTE -> Implementations may defer operations or not wait for them to finish
For example, GPU implementations may dispatch multiplications when they arive
Additionally, a graph of running operations is maintained to handle data dependencies
Only when data is required to actually be present (e.g. decrypt), does the implementation wait
*/

// BFV implementation using Microsoft SEAL

struct seal_context : public context {
	seal_context();
	~seal_context();

	void new_components() override;
	void has_components(int& components) const override;
	void clone_components(int components, std::shared_ptr<context>& ptr) const override;
	void allocate(ciphertext& ct) override;
	void free(ciphertext ct) override;
	void serialize(ciphertext ct, void* buf, std::size_t& n) const override;
	void deserialize(ciphertext ct, const void* buf, std::size_t n) const override;
	void encrypt(const plaintext& pt, ciphertext ct) const override;
	void decrypt(ciphertext ct, plaintext& pt) const override;
	void add_cipher_plain(ciphertext a, const plaintext& b, ciphertext res) const override;
	void add_cipher_cipher(ciphertext a, ciphertext b, ciphertext res) const override;
	void sub_cipher_plain(ciphertext a, const plaintext& b, ciphertext res) const override;
	void sub_cipher_cipher(ciphertext a, ciphertext b, ciphertext res) const override;
	void mul_cipher_plain(ciphertext a, const plaintext& b, ciphertext res) const override;
	void mul_cipher_cipher(ciphertext a, ciphertext b, ciphertext res) const override;
	void rot_cipher_rows(ciphertext ct, int r, ciphertext res) const override;
	void swap_cipher_rows(ciphertext ct, ciphertext res) const override;
	void noise_budget(ciphertext ct, std::size_t& budget) const override;
private:
	int m_id;
};

/*
activate_context(std::shared_ptr<bfv::context> ctx) -> thread local pointer is set
Raw ciphertext and plaintext classes always have the full 8192 coefficients (defined in context as static constexpr)
bfv::vector<...>
	-> can be plaintext or ciphertext
	-> can be base (owns plain-/ciphertext) or component (view to part of base)
	-> can be local or remote
	-> can be a single vector or multiple vectors/components (variadic)
		-> arithmetic with component masks it out
		-> arithmetic with base does operation on all components
	-> tracks multiplicative depth
		-> warning/error if multiplicative depth exceeds limit
		-> use bootstrap member function to handle the warnings/errors
		-> callbacks to reach peer in context
			-> bootstrap_client (unchecked, just raw bootstrap, checks happen at an upper layer using other callbacks)
			-> bootstrap_server_await (waits for client to make request)
			-> bootstrap_server_serve (called immediately after request received with value to be returned)
	-> automatically does secure reveal when cipher is transformed to plain
	-> queues operations until used (cast to plaintext, communication with peer)
Programs are defined TWICE
	-> local stuff is executed
	-> remote stuff is hosted (e.g. bootstrapping server)
	-> defined once from each side (differ e.g. in the plaintext inputs etc.)
	-> program base class may be used to handle context setting
*/

}