#include <catch.hpp>
#include <tmp/debug.h>

#include "common.h"

std::atomic_ullong CallCounterV::count{};
std::atomic_ullong CallCounterII::count{};

namespace expar {

auto buildTestFunction() {
	auto callCounterv = alsk::edsl::makeOperand<int(), CallCounterV>();
	auto callCounterii = alsk::edsl::makeOperand<int(int, int), CallCounterII>();

	auto loop     = seq(3*callCounterv);
	auto serial   = link<R<0>()>(callCounterv, loop);
	auto itersel  = (callCounterv, &link<int(R<0>)>(5*link<R<0>(int)>(serial)) ->* callCounterii);
	auto farmsel  = link<int()>(7*link<R<1>()>(itersel, callCounterv)) ->* callCounterii;
	auto farm     = 11*farmsel;

	return farm;
}

constexpr auto countV() { return 11 * 7 * (1 + 5 * (1 + 3) + 1); }
constexpr auto countII() { return 11 * ((7 - 1) + 7 * 5); }

constexpr auto maxCores = 11*7 * 64;
constexpr std::array<decltype(maxCores), 8> coresList{1, 2, 3, 7, 11*7-1, 11*7, 11*7+1, maxCores};

}

namespace exacc {

auto buildTestFunction() {
	auto callCounterv = alsk::edsl::makeOperand<int(), CallCounterV>();
	auto callCounterii = alsk::edsl::makeOperand<int(int, int), CallCounterII>();

	auto loop     = seq(3*callCounterv);
	auto serial   = link<R<0>()>(callCounterv, loop);
	auto itersel  = (callCounterv, &link<int(R<0>)>(5*link<R<0>(int)>(serial)) ->* callCounterii);
	auto farm     = link<R<1>()>(7*itersel, callCounterv);
	auto farmsel  = link<int()>(11*farm) ->* callCounterii;

	return farmsel;
}

constexpr auto countV() { return 11 * (7 * (1 + 5 * (1 + 3)) + 1); }
constexpr auto countII() { return (11-1) + 11 * 7 * 5; }

constexpr auto maxCores = 11*7 * 64;
constexpr std::array<decltype(maxCores), 8> coresList{1, 2, 3, 7, 11*7-1, 11*7, 11*7+1, maxCores};

}

namespace repeatability {

using PRNG = std::mt19937;
int task(PRNG& rng, std::size_t) {
	std::uniform_int_distribution<int> dist(-100, 100);

	int a = dist(rng);
	std::this_thread::sleep_for(std::chrono::milliseconds(0));
	int b = dist(rng);

	return a - b;
}

int sel(int a, int b) { return a + b; }

auto buildTestFunction() {
	auto etask  = alsk::edsl::makeOperand<int(RNG, CtxId), FN(task)>();
	auto esel   = alsk::edsl::makeOperand<int(int, int), FN(sel)>();

	auto loop     = seq(3*etask);
	auto serial   = link<R<0>()>(etask, (2*loop));
	auto farmsel  = link<int()>(5*serial) ->* esel;

	return link<R<1>()>(farmsel, farmsel);
}

}

using namespace alsk;

using Executors = std::tuple<
		exec::DynamicPool<void>,
		exec::FirstLevelEqui<void>,
		exec::FirstLevelGreedy<void>,
		exec::FirstLevelNoOpti<void>,
		exec::Sequential<void>,
		exec::StaticPool<void>,
		exec::StaticPoolId<void>,
		exec::StaticThread<void>
>;
TEMPLATE_LIST_TEST_CASE("Executors", "[executors]", Executors) {
	CAPTURE(tmp::typeName<TestType>());

	SECTION("executeParallel") {
		using namespace expar;

		auto e = buildTestFunction();
		for(auto cores: coresList) {
			auto f = edsl::implement<GetExecutor<TestType>::template Template>(e);
			f.executor.cores = cores;

			CallCounterV::count = 0;
			CallCounterII::count = 0;

			f();

			REQUIRE(CallCounterV::count == countV());
			REQUIRE(CallCounterII::count == countII());
		}
	}

	SECTION("executeParallelAccumulate") {
		using namespace exacc;

		auto e = buildTestFunction();
		for(auto cores: coresList) {
			auto f = edsl::implement<GetExecutor<TestType>::template Template>(e);
			f.executor.cores = cores;

			CallCounterV::count = 0;
			CallCounterII::count = 0;

			f();

			REQUIRE(CallCounterV::count == countV());
			REQUIRE(CallCounterII::count == countII());
		}
	}
}

TEMPLATE_LIST_TEST_CASE("Repeatability", "[executors][repeatability]", Executors) {
	using namespace repeatability;

	CAPTURE(tmp::typeName<TestType>());

	auto e = buildTestFunction();
	for(std::size_t upTo: {2, 3, 4, 5, 6, 7, 8}) {
		CAPTURE(upTo);

		auto f = edsl::implement<GetExecutor<TestType>::template Template>(e);
		f.executor.repeatability.upTo(upTo);

		f.executor.cores = 1;

		auto expected = f();
		f.state.context.reset();

		for(std::size_t cores = 2; cores <= upTo; ++cores) {
			CAPTURE(cores);

			f.executor.cores = cores;

			REQUIRE(f() == expected);
			f.state.context.reset();
		}
	}
}