#include <stdlib.h>
#include <stdio.h>
#include <errno.h>
#include <pthread.h>
#include "matrix_lib.h"
#define VERBOSE 0
typedef struct {
  int       id;
  int       Arow;
  int       Bcol;
} package_t; 

static matrix_t MA,MB,MC;
static int  size = ARRAY_SIZE;

void *mult_worker(void *arg){
  package_t p = *((package_t *)arg);
  if(VERBOSE)
     printf("thread  %d: processing A row %d, B col %d\n",
            p.id, p.Arow, p.Bcol );
  mult(size, p.Arow, p.Bcol, MA, MB, MC);
  if(VERBOSE)printf("thread  %d: complete\n", p.id);
  pthread_exit(NULL); 
}
int main(int argc, char **argv){
  int  row, column, num_threads, i;
  /* create thread id's          */
  pthread_t *threads = (pthread_t *)calloc(size*size, sizeof(pthread_t));
  /* create thread data          */
  package_t *thread_data = (package_t *)calloc(size*size, sizeof(package_t));
  /* check the memory allocation */
  if((threads == NULL) || (thread_data == NULL)){
    fprintf(stderr, "Calloc failed, sorry!\n");
     exit(EXIT_FAILURE);
  }
  /* initialize  MA              */
  matrix_init(size, MA, identity);
  /* initialize  MB              */
  matrix_init(size, MB, sum);
  /* display MA                  */
  matrix_printf("A", size, MA);
  /* display MB                  */
  matrix_printf("B", size, MB);
  /* make the data               */
  num_threads = 0;
  for(row = 0; row < size; row++) {
    for (column = 0; column < size; column++) {
      thread_data[num_threads].id = num_threads;
      thread_data[num_threads].Arow = row;
      thread_data[num_threads].Bcol = column;
      num_threads++; }}
  /* make the threads            */
  num_threads = 0;
  for(row = 0; row < size; row++) {
    for (column = 0; column < size; column++) {
      pthread_create(&threads[num_threads],
                     NULL, 
                     mult_worker, 
                     (void *)&thread_data[num_threads]);
      num_threads++;
      if(VERBOSE)
        printf("main thread : thread %d created\n", num_threads);
    }
  }
  /* wait for threads to finish  */
  for (i = 0; i < (size*size); i++) {
    pthread_join(threads[i], NULL);
    if(VERBOSE)
      printf("main thread : child %d has joined\n", i);
  }
  /* display result              */
  if(VERBOSE)
    printf("main thread : The result is;\n");
  matrix_printf("C", size, MC);
  free(threads);
  free(thread_data);
   exit(EXIT_SUCCESS);
}