rosa/inc/alsk/executor/impl/dynamicpool.h

#ifndef ALSK_ALSK_EXECUTOR_IMPL_DYNAMICPOOL_H
#define ALSK_ALSK_EXECUTOR_IMPL_DYNAMICPOOL_H

#include <algorithm>
#include <future>
#include <vector>

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

#include "../utility/pool.h"

namespace alsk {
namespace exec {

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

public:
	std::size_t maxTaskCount = 1'000;

public:
	template<typename Impl>
	void config(Impl& impl) {
		impl.state.executor.config(cores);
	}

	template<typename Task, typename Impl, typename BTask, typename Parameters>
	void executeParallel(Impl& impl, BTask& task, Parameters const& parameters, std::size_t n) {
		std::size_t taskCount = std::min(maxTaskCount, n);

		if(cores > 1 && taskCount > 1) {
			Info info;
			std::vector<std::future<void>> futures(taskCount);
			std::size_t const step = n/taskCount;
			std::size_t const remain = n - step*(taskCount-1);

			typename Impl::State& state = impl.state;

			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 < taskCount-1; ++i)
				futures[i] = state.executor.run([&, b=i*step, k=step]{ run(b, k); });
			futures[taskCount-1] = state.executor.run([&, b=(taskCount-1)*step, k=remain]{ run(b, k); });

			state.executor.wait(futures);
		} else {
			Info info;
			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) {
		std::size_t taskCount = std::min(maxTaskCount, n);

		Value best{};

		if(cores > 1 && taskCount > 1) {
			Info info;
			std::vector<std::future<void>> futures(taskCount);
			std::size_t const step = n/taskCount;
			std::size_t const remainBase = n - step*taskCount;
			std::size_t remain = remainBase;

			typename Impl::State& state = impl.state;

			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(taskCount);
			for(std::size_t i = 0; i < taskCount-1; ++i) {
				std::size_t offset = !!remain;
				remain -= offset;
				futures[i] = state.executor.run([&, &best=bests[i], b=start, k=step+offset] { run(best, b, k); });
				start += step+offset;
			}

			futures[taskCount-1] = state.executor.run([&, &best=bests[taskCount-1], b=start, k=step] { run(best, b, k); });

			state.executor.wait(futures);

			if(taskCount) best = std::move(bests[0]);
			for(std::size_t i = 1; i < taskCount; ++i)
				best = Select::execute(impl, select, i, info, parameters, std::tuple<>{}, std::move(bests[i]), std::move(best));
		} else {
			Info info;
			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<DynamicPool<S>>: util::Pool {};

}
}

#endif