#pragma once

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

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

#include <dml/dml.hpp>

#include "barrier.hpp"
#include "statuscode-tostring.hpp"
#include "task-data.hpp"

double avg(const std::vector<uint64_t>& v) {
    return static_cast<long double>(std::accumulate(v.begin(), v.end(), 0)) / static_cast<long double>(v.size());
}

double stdev(const std::vector<uint64_t>& v, const double mean) {
    std::vector<double> diff(v.size());
    std::transform(v.begin(), v.end(), diff.begin(), [mean](double x) { return x - mean; });
    const double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
    const double stdev = std::sqrt(sq_sum / static_cast<double>(v.size()));
    return stdev;
}

#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; }}

#define ADD_TIMING_MESSUREMENT { if (i >= 5) { submission_durations.emplace_back(std::chrono::duration_cast<std::chrono::microseconds>(se - st).count()); completion_durations.emplace_back(std::chrono::duration_cast<std::chrono::microseconds>(et - se).count()); combined_durations.emplace_back(std::chrono::duration_cast<std::chrono::microseconds>(et - st).count());}}

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

    std::vector<uint64_t> submission_durations;
    std::vector<uint64_t> completion_durations;
    std::vector<uint64_t> combined_durations;

    // 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);

    args->status = dml::status_code::ok;
    args->rep_completed = 0;

    // we add 5 as the first 5 iterations will not be meassured
    // to remove exceptional values encountered during warmup
    for (uint32_t i = 0; i < args->rep_count + 5; i++) {
        // synchronize the start of each iteration
        // using the barrier structure
        args->barrier_->wait();

        if (args->batch_submit) {
            const auto st = std::chrono::steady_clock::now();

            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);

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

            auto result = handler.get();

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

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

            ADD_TIMING_MESSUREMENT;
        }
        else if (args->batch_size > 1) {
            // implementation for non-batched batch submit follows here
            // this means we submit a bunch of work as single descriptors
            // but then dont wait for the completion immediately

            std::vector<dml::handler<dml::mem_copy_operation, std::allocator<uint8_t>>> handlers;

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

            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

                handlers.emplace_back(dml::submit<path>(dml::mem_copy.block_on_fault(), srcv, dstv));
            }

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

            for (auto& handler : handlers) {
                auto result = handler.get();
                const dml::status_code status = result.status;
                CHECK_STATUS(status, "Operation completed with an Error!");
            }

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

            ADD_TIMING_MESSUREMENT;
        }
        else {
            const auto st = std::chrono::steady_clock::now();

            // 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);

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

            auto result = handler.get();

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

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

            ADD_TIMING_MESSUREMENT;
        }

        // again: we do not count the first 5 repetitions
        if (i >= 5) args->rep_completed++;
    }

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

    args->combined_duration = avg(combined_durations);
    args->complete_duration = avg(completion_durations);
    args->submit_duration = avg(submission_durations);
    args->combined_duration_stdev = stdev(combined_durations, args->combined_duration);
    args->complete_duration_stdev = stdev(completion_durations, args->complete_duration);
    args->submit_duration_stdev = stdev(submission_durations, args->submit_duration);

    return nullptr;
}

template <typename path>
void execute_dml_memcpy(std::vector<TaskData>& args) {
    barrier task_barrier(args.size());
    std::vector<pthread_t> threads;

    // initialize numa library
    numa_available();

    // for each submitted task we link the semaphore
    // and create the thread, passing the argument
    for (auto& arg : args) {
        arg.barrier_ = &task_barrier;
        threads.emplace_back();

        if (pthread_create(&threads.back(), nullptr, thread_function<path>, &arg) != 0) {
            std::cerr << "Error creating thread" << std::endl;
            exit(1);
        }
    }

    for (pthread_t& t : threads) {
        pthread_join(t, nullptr);
    }
}