#ifndef ALSK_ALSK_EXECUTOR_IMPL_FIRSTLEVEL_NOOPTI_H
#define ALSK_ALSK_EXECUTOR_IMPL_FIRSTLEVEL_NOOPTI_H

#include <thread>
#include <set>
#include <cmath>
#include <vector>

#include "../../executorbase.h"
#include "../../executorstate.h"
#include "../../../skeleton/traits.h"

namespace alsk {
namespace exec {

template<typename S>
struct FirstLevelNoOpti: ExecutorBase {
	using Tag = alsk::tag::Parallel;

public:
	struct Info {
		unsigned int parDepth;
	};

private:
	unsigned int threadLimit(unsigned int level) const { return level? 1 : cores; }

public:
	template<typename Impl>
	std::size_t contextIdCount(Impl&, std::size_t count) { return count; }

	template<typename Impl>
	std::size_t contextId(Impl&, std::size_t id) { return id; }

	template<typename Task, typename Impl, typename BTask, typename Parameters>
	void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
		auto const& parDepth = impl.executorInfo.parDepth;
		std::size_t const maxThreads = threadLimit(parDepth);

		std::size_t const nThreads = std::min(n, maxThreads);
		if(nThreads > 1) {
			Info info{parDepth+1};
			std::vector<std::thread> threads(nThreads-1);
			std::size_t const step = std::round(static_cast<double>(n)/nThreads);

			auto run = [&](std::size_t b, std::size_t k) {
				for(std::size_t i = 0; i < k; ++i)
					Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
			};

			for(std::size_t i = 0; i < nThreads-1; ++i)
				threads[i] = std::thread{run, i*step, step};

			run((nThreads-1)*step, n-(nThreads-1)*step);

			for(std::thread& thread: threads) thread.join();

		} else {
			Info info{parDepth};
			for(std::size_t i = 0; i < n; ++i)
				Task::execute(impl, task, i, info, parameters, std::tuple<>{});
		}
	}

	template<typename Value, typename Task, typename Select, typename Impl, typename BTask, typename BSelect, typename Parameters>
	Value executeParallelAccumulate(Impl& impl, BTask& task, BSelect& select, Parameters const& parameters, std::size_t n) {
		auto const& parDepth = impl.executorInfo.parDepth;
		std::size_t const maxThreads = threadLimit(parDepth); // TODO fix neighbours

		Value best{};

		std::size_t const nThreads = std::min(n, maxThreads);
		if(nThreads > 1) {
			Info info{parDepth+1};
			std::vector<std::thread> threads(nThreads-1);
			std::size_t const step = n/nThreads;
			std::size_t const remainBase = n - step*nThreads;
			std::size_t remain = remainBase;

			auto run = [&](Value& out, std::size_t b, std::size_t k) {
				Value best{};

				if(k)
					best = Task::execute(impl, task, b+0, info, parameters, std::tuple<>{});
				for(std::size_t i = 1; i < k; ++i) {
					Value current = Task::execute(impl, task, b+i, info, parameters, std::tuple<>{});
					best = Select::execute(impl, select, b+i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
				}

				out = std::move(best);
			};

			std::size_t start{};
			std::vector<Value> bests(nThreads);
			for(std::size_t i = 0; i < nThreads-1; ++i) {
				std::size_t offset = !!remain;
				remain -= offset;
				threads[i] = std::thread{run, std::ref(bests[i]), start, step+offset};
				start += step+offset;
			}

			run(bests[nThreads-1], start, step);

			for(std::thread& thread: threads) thread.join();

			if(nThreads)  best = std::move(bests[0]);
			for(std::size_t i = 1; i < nThreads; ++i)
				best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(bests[i]), std::move(best));
		} else {
			Info info{parDepth};
			if(n)
				best = Task::execute(impl, task, 0, info, parameters, std::tuple<>{});
			for(std::size_t i = 1; i < n; ++i) {
				Value current = Task::execute(impl, task, i, info, parameters, std::tuple<>{});
				best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(current), std::move(best));
			}
		}

		return best;
	}
};

template<typename S>
struct ExecutorState<FirstLevelNoOpti<S>> {};

}
}

#endif