/* example of using threads to split up the work */
#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#include <sys/time.h>

typedef struct {
  double *array;
  int N;
  double sum;
} threadInfo;

void *fsum( void *ud ) {
  threadInfo *ti = (threadInfo *)ud;
  int i;

  for(i=0;i<ti->N;i++) {
    ti->sum += ti->array[i];
  }
}

int main(int argc, char *argv[]) {
  const int size = 1000000;
  int i, mux;
  double *bigarray = malloc( sizeof(double) * size );
  double sum;
  struct timeval start, end;
  double stime, ftime;
  threadInfo ti[4];
  pthread_t thread[4];

  for(i=0;i<size;i++) {
    bigarray[i] = drand48();
  }

  gettimeofday( &start, NULL );

  // now sum the array in a single loop and time it.
  for(mux=0;mux<8;mux++) {
    sum = 0.0;
    for(i=0;i<size;i++) {
      sum += bigarray[i];
    }
  }

  gettimeofday( &end, NULL );
  stime = start.tv_sec + start.tv_usec / 1e+6;
  ftime = end.tv_sec + end.tv_usec/1e+6;

  printf("single processor loop %.4f sum %.2f\n", ftime - stime, sum/size );

  gettimeofday( &start, NULL );

  for(i=0;i<4;i++) {
    ti[i].N = size/4;
    ti[i].array = &( bigarray[i*(size/4)] );
    ti[i].sum = 0.0;
  }

  for(i=0;i<4;i++) {
    pthread_create(&(thread[i]), NULL, fsum, &(ti[i]) );
  }

  for(i=0;i<4;i++) {
    pthread_join( thread[i], NULL );
  }

  sum = ti[0].sum + ti[1].sum + ti[2].sum + ti[3].sum;

  gettimeofday( &end, NULL);
  stime = start.tv_sec + start.tv_usec / 1e+6;
  ftime = end.tv_sec + end.tv_usec/1e+6;

  printf("multi-processor loop %.4f sum %.2f\n", ftime - stime, sum/size );


  return(0);
}