#include <stdio.h>
#include <string.h>
#include <stdlib.h>
#include <unistd.h>
#include <math.h>

#include "../../../context.h"

int cpu_num = 1;
int length = 1024;
int split  = 8;
int gpu_num = 0;
int CPU_ANY = -1;
int CPU_CUDA = -1;

void *start_taskManager(struct Context *context) {
    goto initDataGears(context, Gearef(context, LoopCounter), Gearef(context, TaskManager));
    return 0;
}

#ifdef USE_CUDAWorker
#ifdef USE_CUDA_MAIN_THREAD
extern volatile int cuda_initialized;
#endif
#endif

__code initDataGears(struct LoopCounter* loopCounter, struct TaskManager* taskManager) {
    // loopCounter->tree = createRedBlackTree(context);
    loopCounter->i = 0;
    taskManager->taskManager = (union Data*)createTaskManagerImpl(context, cpu_num, gpu_num, 0);
#ifdef USE_CUDAWorker
#ifdef USE_CUDA_MAIN_THREAD
    while(! cuda_initialized) {};
#endif
#endif
    goto meta(context, C_code1);
}

__code initDataGears_stub(struct Context* context) {
    struct TaskManager* taskManager =  Gearef(context, TaskManager);
    taskManager->taskManager = 0;
#if (! defined(USE_CUDAWorker) || ! defined(USE_CUDA_MAIN_THREAD))
    struct LoopCounter* loopCounter = Gearef(context, LoopCounter);
    goto initDataGears(context, loopCounter, taskManager);
#else
    cuda_initialized = 0;
    pthread_t thread;
    pthread_create(&thread, NULL, (void*)&start_taskManager, context);
    while (taskManager->taskManager == 0);
    TaskManager *t = (TaskManager*)taskManager->taskManager;
    TaskManagerImpl *im = (TaskManagerImpl*)t->taskManager;
    struct Queue *q = (Queue *)im->workers[0];
    createCUDAWorker(context,0,q, im);
    pthread_join(thread,0);
    exit(0);
#endif
}

__code code1(struct Time* time) {
    printf("cpus:\t\t%d\n", cpu_num);
    printf("gpus:\t\t%d\n", gpu_num);
    printf("length:\t\t%d\n", length);
    printf("length/task:\t%d\n", length/split);
    /* puts("queue"); */
    /* print_queue(context->data[ActiveQueue]->queue.first); */
    /* puts("tree"); */
    /* print_tree(context->data[Tree]->tree.root); */
    /* puts("result"); */

    goto meta(context, C_createTask1);
}

__code createTask1(struct LoopCounter* loopCounter, struct TaskManager* taskManager) {
    int logN = (int)log2((float)length);
    struct Context** tasks = (struct Context**)ALLOC_ARRAY(context, Context, logN*(1+logN)/2 + 2);
    int taskCount = 0;

    struct SortArray* outputSortArray = &ALLOCATE_DATA_GEAR(context, SortArray)->SortArray;
    struct SortArray* inputSortArray = outputSortArray;
    struct Time* time = createTimeImpl(context);

    // par goto makeArray(outputSortArray, time, _exit);
    // par goto makeArray(outputSortArray, time, time->start); ?
    struct Context* task = NEW(struct Context);
    initContext(task);
    task->next = C_makeArray;
    task->idgCount = 0;
    task->idg = task->dataNum;
    task->maxIdg = task->idg;
    task->odg = task->maxIdg;
    task->data[task->odg] = (union Data*)outputSortArray;
    task->data[task->odg+1] = (union Data*)time;
    task->maxOdg = task->odg + 2;
    tasks[taskCount] = task;

    taskCount++;
    for (int i=2; i <= length; i=2*i) {
        int first = 1;
        for (int j=i>>1; j > 0; j=j>>1) {
            outputSortArray = &ALLOCATE_DATA_GEAR(context, SortArray)->SortArray;
            struct Context* task = NEW(struct Context);
            initContext(task);
            struct Integer* integer1 = &ALLOCATE_DATA_GEAR(context, Integer)->Integer;
            struct Integer* integer2 = &ALLOCATE_DATA_GEAR(context, Integer)->Integer;
            integer1->value = j;
            integer2->value = first;
            inputSortArray->prefix = length/2/split;
            task->next = C_bitonicSwap;
            task->iterator = createOneDimIterator(context, split);
            task->idgCount = 1;
            task->idg = task->dataNum;
            task->data[task->idg] = (union Data*)inputSortArray;
            task->data[task->idg+1] = (union Data*)integer1;
            task->data[task->idg+2] = (union Data*)integer2;
            task->maxIdg = task->idg + 3;
            task->odg = task->maxIdg;
            task->data[task->odg] = (union Data*)outputSortArray;
            task->maxOdg = task->odg + 1;
            tasks[taskCount] = task;
            taskCount++;
            first = 0;
            inputSortArray = outputSortArray;
        }
    }

    // par goto printArray(inputSortArray, time, __exit)
    // par goto printArray(inputSortArray, time, time->exit)?
    task = NEW(struct Context);
    initContext(task);
    task->next = C_printArray;
    task->idgCount = 2;
    task->idg = task->dataNum;
    task->data[task->idg] = (union Data*)inputSortArray;
    task->data[task->idg+1] = (union Data*)time;
    task->maxIdg = task->idg + 2;
    task->odg = task->maxIdg;
    task->maxOdg = task->odg;
    tasks[taskCount] = task;

    taskManager->contexts = tasks;
    // goto code2();
    taskManager->next1 = C_code2;
    goto meta(context, taskManager->taskManager->TaskManager.spawnTasks);
    //goto meta(context, taskManager->taskManager->TaskManager.shutdown);
}

__code code2(struct LoopCounter* loopCounter, struct TaskManager* taskManager, struct Time* time) {
    sleep(40);
    taskManager->next = C_exit_code;
    goto meta(context, taskManager->taskManager->TaskManager.shutdown);
}

void init(int argc, char** argv) {
    for (int i = 1; argv[i]; ++i) {
        if (strcmp(argv[i], "-cpu") == 0)
            cpu_num = (int)atoi(argv[i+1]);
        else if (strcmp(argv[i], "-l") == 0)
            length = (int)atoi(argv[i+1]);
        else if (strcmp(argv[i], "-s") == 0)
            split = (int)atoi(argv[i+1]);
        else if (strcmp(argv[i], "-cuda") == 0) {
            gpu_num = 1;
            CPU_CUDA = 0;
        }
    }
}

int main(int argc, char** argv) {
    init(argc, argv);
    struct Context* main_context = NEW(struct Context);
    initContext(main_context);
    main_context->next = C_initDataGears;

    goto start_code(main_context);
}

__code makeArray(struct Time* timeInterface, __code next(struct SortArray* output, struct Time* output1, ...)){
    struct SortArray* output = *O_output;
    struct Time* output1 = *O_output1;
    if (output->loopCounter == 0){
        output->array = (Integer**)ALLOC_ARRAY(context, Integer, length);
        srand((unsigned) time(NULL));
    }
    if (output->loopCounter == GET_SIZE(output->array)){
        printf("created Array\n");
        output->loopCounter = 0;
        timeInterface->time = (union Data*)output1;
        timeInterface->next = context->next;
        *O_output = output;
        *O_output1 = output1;
        goto meta(context, output1->start);
    }
    struct Integer* integer = new Integer();
    integer->value = rand() % 1000;
    output->array[output->loopCounter] = integer;
    //printf("%d\n", output->array[output->loopCounter]->value);
    output->loopCounter++;
    *O_output = output;
    *O_output1 = output1;
    goto meta(context, C_makeArray);
}

__code makeArray_stub(struct Context* context) {
    SortArray** O_output = (struct SortArray**)&context->data[context->odg];
    Time** O_output1 = (struct Time**)&context->data[context->odg+1];
    goto makeArray(context,
                   Gearef(context, Time),
                   context->next,
                   O_output,
                   O_output1);
}

__code printArray(struct Time* timeInterface, struct SortArray* inputArray, struct Time* inputTime, __code next(...)){
    timeInterface->time = (union Data*)inputTime;
    timeInterface->next = C_printArray1;
    goto meta(context, inputTime->end);
}

__code printArray_stub(struct Context* context) {
    goto printArray(context,
                   Gearef(context, Time),
                   &context->data[context->idg]->SortArray,
                   &context->data[context->idg+1]->Time,
                   context->next);
}

__code printArray1(struct SortArray* inputArray, __code next(...)){
    //printf("%d\n", inputArray->array[inputArray->loopCounter]->value);
    inputArray->loopCounter++;
    if (inputArray->loopCounter == GET_SIZE(inputArray->array)){
        inputArray->loopCounter = 0;
        goto meta(context, next);
    }
    if (inputArray->array[inputArray->loopCounter-1]->value > inputArray->array[inputArray->loopCounter]->value) {
        printf("wrong result\n");
        goto meta(context, next);
    }
    goto meta(context, C_printArray1);
}

__code printArray1_stub(struct Context* context) {
    goto printArray1(context,
                   &context->data[context->idg]->SortArray,
                   context->next);
}