/* prime_counter_better_load.c
 * This is a threaded counter of prime numbers
 * It balances the load across threads by giving each thread a series of 100-number sequences, e.g. 
 * if there are two threads, then they will take every other sequence o 100 numbers (T0 gets 0-99, 200-299, etc. while
 * T1 gets 100-199, 300-399, etc.)
 *
 * gcc -o pcb -lpthread prime_counter_better_load.c my_timing.c
 */
 
#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#include "my_timing.h"

typedef struct {
    int id;
    int start;
    int N;
    int big_stride;
    int small_extent;
    int count;
} ThreadInfo;

// check whether a number is a prime
// return 1 if it is a prime, else 0
int is_prime (int n) {
    if (n < 2) return 0;
    if (n == 2) return 1;
    if (n%2 == 0) return 0;
    for (int i = 3; i < n; i += 2) {
        if (n%i == 0) return 0;
    }
    return 1;
}

// worker func: check the primes in
// the given range
void *thread_count_primes (void *thread_info) {
    double t1 = get_time_sec();
    ThreadInfo *ti = (ThreadInfo *)thread_info;
    // do the small_extent numbers every big stride
    for (int i = ti->start; i < ti->N; i = i + ti->big_stride) {
        for (int j = 0; j < ti->small_extent; j++) {
            if (is_prime(i+j))
                ti->count++;
        }
    }
    
    double t2 = get_time_sec();
    printf( "Thread %d took %f seconds\n", ti->id, t2-t1 );
    pthread_exit(NULL);
}

int main (int argc, char** argv) {
    int N = 1000000;
    int NUM_THREADS = 3;
    if (argc > 1) NUM_THREADS = atoi(argv[1]);
    double t1, t2;

    ThreadInfo srt[NUM_THREADS];
    pthread_t thread[NUM_THREADS];
    int small_extent = 100;
    int big_stride = NUM_THREADS * small_extent;
    for (int i = 0; i < NUM_THREADS; i++) {
        srt[i].id = i;
        srt[i].start = i*small_extent;
        srt[i].N = N;
        srt[i].big_stride = big_stride;
        srt[i].small_extent = small_extent;
        srt[i].count = 0;
    }

    t1 = get_time_sec();
    
    // create threads
    for (int i = 0; i < NUM_THREADS; i++) {
        pthread_create(&thread[i], NULL, thread_count_primes, &srt[i]);
    }
    
    // join threads
    for (int i = 0; i < NUM_THREADS; i++) {
        pthread_join(thread[i], NULL);
    }
    
    // sum counts
    int count = 0;
    for (int i = 0; i < NUM_THREADS; i++) {
        count += srt[i].count;
    }
    t2 = get_time_sec();
    
    printf("There are %d primes not larger than %d\n",count, N);
    printf("It took %f seconds to count the number with %d threads\n", t2 - t1, NUM_THREADS);
    /*
    ThreadInfo ti[3];
    pthread_t thread[3]; // use 2 of the threads to count N numbers
    
    ti[0].start = 0;
    ti[0].extent = N/2;
    ti[0].count = 0;
    ti[1].start = N/2;
    ti[1].extent = N/2;
    ti[1].count = 0;
    
    ti[2].start = 0;
    ti[2].extent = N;
    ti[2].count = 0;  
  
    t1 = get_time_sec();
    
    // create threads
    for (int i = 0; i < 2; i++) {
        pthread_create(&thread[i], NULL, thread_count_primes, &ti[i]);
    }
    
    // join threads
    for (int i = 0; i < 2; i++) {
        pthread_join(thread[i], NULL);
    }
    
    // sum counts
    int count = 0;
    for (int i = 0; i < 2; i++) {
        count += ti[i].count;
    }
    
    t2 = get_time_sec();
    printf("There are %d primes not larger than 100000\n",count);
    printf("It took %f seconds to count the number with 2 threads\n", t2 - t1);
    
    t1 = get_time_sec();
    
    // create thread
    pthread_create(&thread[2], NULL, thread_count_primes, &ti[2]);
    
    // join thread
    pthread_join(thread[2], NULL);
    
    // sum count
    count = ti[2].count;
    
    t2 = get_time_sec();
    printf("There are %d primes not larger than 100000\n",count);
    printf("It took %f seconds to count the number with 1 thread\n", t2 - t1);
 */   
    return EXIT_SUCCESS;
}