#pragma once

#include <iostream>
#include <vector>
#include <chrono>
#include <numeric>
#include <future>

#include <pthread.h>
#include <semaphore.h>
#include <numa.h>

#include <dml/dml.hpp>

#include "util/dml-helper.hpp"
#include "util/task-data.hpp"

#define LOG_CODE_INFO "Location: " << __FILE__ << "@" << __LINE__ << "::" << __FUNCTION__ << std::endl
#define LOG_ERR { pthread_t t = pthread_self(); std::cerr << "--- BEGIN ERROR MSG ---" << std::endl << "Physical: [Node " << args->numa_node << " | Thread " << t << "]" << std::endl; } std::cerr << LOG_CODE_INFO
#define CHECK_STATUS(status,msg) { if (status != dml::status_code::ok) { LOG_ERR << "Status Code: " << StatusCodeToString(status) << std::endl << msg << std::endl; args->status = status; return nullptr; }}

std::shared_future<void> LAUNCH_;

template <typename path>
void* thread_function(void* argp) {
    TaskData* args = reinterpret_cast<TaskData*>(argp);

    // set numa node and core affinity of the current thread
    numa_run_on_node(args->numa_node);
    
    // allocate memory for the move operation on the requested numa nodes
    void* src = numa_alloc_onnode(args->size, args->nnode_src);
    void* dst = numa_alloc_onnode(args->size, args->nnode_dst);
    dml::data_view srcv = dml::make_view(reinterpret_cast<uint8_t*>(src), args->size);
    dml::data_view dstv = dml::make_view(reinterpret_cast<uint8_t*>(dst), args->size);

    std::memset(src, 0, args->size);
    std::memset(dst, 0, args->size);

    args->status = dml::status_code::ok;

    LAUNCH_.wait();

    if (args->batch_size > 1) {
        auto sequence = dml::sequence(args->batch_size, std::allocator<dml::byte_t>());

        for (uint32_t j = 0; j < args->batch_size; j++) {
            // block_on_fault() is required to submit the task in a way so that the
            // DSA engine can handle page faults itself together with the IOMMU which
            // requires the WQ to be configured to allow this too

            const auto status = sequence.add(dml::mem_copy.block_on_fault(), srcv, dstv);
            CHECK_STATUS(status, "Adding operation to batch failed!");
        }

        // we use the asynchronous submit-routine even though this is not required
        // here, however the project later on will only use async operation and
        // therefore this behaviour should be benchmarked

        auto handler = dml::submit<path>(dml::batch, sequence);

        auto result = handler.get();

        const dml::status_code status = result.status;
        CHECK_STATUS(status, "Batch completed with an Error!");
    }
    else {
        // we use the asynchronous submit-routine even though this is not required
        // here, however the project later on will only use async operation and
        // therefore this behaviour should be benchmarked
        // block_on_fault() is required to submit the task in a way so that the
        // DSA engine can handle page faults itself together with the IOMMU which
        // requires the WQ to be configured to allow this too
        auto handler = dml::submit<path>(dml::mem_copy.block_on_fault(), srcv, dstv);

        auto result = handler.get();

        const dml::status_code status = result.status;
        CHECK_STATUS(status, "Operation completed with an Error!");
    }

    // free the allocated memory regions on the selected nodes
    numa_free(src, args->size);
    numa_free(dst, args->size);

    return nullptr;
}

template <typename path>
std::vector<uint64_t> execute_dml_memcpy(std::vector<TaskData>& args, const uint64_t iterations) {
    std::vector<uint64_t> timing;

    // initialize numa library
    numa_available();

    // for each submitted task we link the semaphore
    // and create the thread, passing the argument

    for (uint64_t i = 0; i < iterations + 5; i++) {
        std::vector<std::thread> threads;
        std::promise<void> launch_promise;
        LAUNCH_ = launch_promise.get_future();

        for (auto& arg : args) {
            threads.emplace_back(thread_function<path>, &arg);
        }

        using namespace std::chrono_literals;
        std::this_thread::sleep_for(100ms);

        const auto time_start = std::chrono::steady_clock::now();

        launch_promise.set_value();

        for(std::thread& t : threads) { t.join(); }

        const auto time_end = std::chrono::steady_clock::now();

        if (i >= 5) timing.emplace_back(std::chrono::duration_cast<std::chrono::nanoseconds>(time_end - time_start).count());
    }

    return timing;
}