#include <stdio.h>
#include <stdlib.h>
#include <semaphore.h>
#include <time.h>
#include <pthread.h>
#include <stdint.h>
#include <immintrin.h>		/* for SSE/AVX intrinsics */

#define NB_TH 4
#define NB_JOBS 500000000
#define NB_EVENTS (NB_JOBS/100)

struct event {
  int result;
};

struct trace {
  int nb_events;
  struct event *event_array;
  char padding[128];
};

struct trace t_th[NB_TH];
sem_t thread_ready;

#define TIME_DIFF(t1, t2) (((t2).tv_sec-(t1).tv_sec)*1e9+((t2).tv_nsec-(t1).tv_nsec))

void analyze_events_avx(int thread_id) {
  int i;
  int result = 0;
  int result_array[8];

  __m256i result_256 = _mm256_loadu_si256((__m256i*)&t_th[thread_id].event_array[0].result);
  __m256i* vtab = (__m256i*)&t_th[thread_id].event_array[8].result;
  for(i=8; i<NB_EVENTS; i+=8) {
    result_256 = _mm256_add_epi32(*vtab, result_256);
    vtab++;
  }

  _mm256_storeu_si256((__m256i*)result_array, result_256);
  for(i=0; i<8; i++)
    result += result_array[i];

  int average_result = result/NB_EVENTS;
  printf("average result for thread %d: %d\n", thread_id, average_result);
}


void analyze_events(int thread_id) {
  analyze_events_avx(thread_id);
  return;

  int i;
  int result = 0;

  for(i=0; i<NB_EVENTS; i++) {
    result += t_th[thread_id].event_array[i].result;
  }

  int average_result = result/NB_EVENTS;
  printf("average result for thread %d: %d\n", thread_id, average_result);
}

void* thread_function(void* arg) {
  int my_id= *(int*)arg;
  int i;
  t_th[my_id].nb_events = 0;
  t_th[my_id].event_array = malloc(sizeof(struct event)*NB_EVENTS);
  sem_post(&thread_ready);
  printf("Thread %d starts processing jobs\n", my_id);

  for(i=0; i<NB_JOBS; i++) {
    int nb_events = t_th[my_id].nb_events++;
    t_th[my_id].event_array[nb_events%NB_EVENTS].result = i;
  }
  printf("Thread %d finished processing jobs\n", my_id);

  struct timespec t1, t2;
  clock_gettime(CLOCK_THREAD_CPUTIME_ID, &t1);
  analyze_events(my_id);
  clock_gettime(CLOCK_THREAD_CPUTIME_ID, &t2);
  double duration = TIME_DIFF(t1, t2)/1e6;
  printf("analysis took %lf ms\n", duration);
  return NULL;
}

int main() {
  int i;
  sem_init(&thread_ready, 0, 0);
  pthread_t tids[NB_TH];

  for(i=0; i<NB_TH; i++) {
    pthread_create(&tids[i], NULL, thread_function, &i);
    sem_wait(&thread_ready);
  }

  for(i=0; i<NB_TH; i++) {
    pthread_join(tids[i], NULL);
  }
  return 0;
}