#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#include <time.h>

const int m_rows = 40000;
const int m_cols = m_rows;

_Atomic int hv = 0;
int **matrix;

typedef struct{
  int start;
  int end;
} thread_args;

void* parallel_search_cols(void *arg){
  thread_args *args = (thread_args *)arg;
  for(int r = 0; r < m_rows; r++){
    for(int c = args->start; c < args->end; c++){
      if(matrix[r][c] > hv){
        hv = matrix[r][c];
      }
    }
  }
  free(args);
  return NULL;
}

int main(void){
  printf("1st --> Allocate memory\n");
  clock_t t_t1; t_t1 = clock();
  size_t rows_size = m_rows * sizeof(int*);
  matrix = malloc(rows_size);
  if(matrix == NULL){
    perror("malloc failled");
    return 1;
  }
  size_t data_size = (size_t)m_rows * m_cols * sizeof(int);
  int *data;
  if(posix_memalign((void**)&data, 64, data_size) != 0){
    perror("not able to alocate memory");
    free(matrix);
    return 1;
  }
  t_t1 = clock() - t_t1;
  double t1_ttaken = ((double)t_t1)/CLOCKS_PER_SEC;
  printf(" %f sec\n",t1_ttaken);

  printf("2nd --> Populate the matrix\n");
  clock_t t_t2; t_t2 = clock();
  for(int r = 0; r<m_rows;r++){
    matrix[r] = data + r * m_cols;
  }
  for(int r = 0; r < m_rows; r++){  
    for(int c = 0; c < m_cols; c++){
      matrix[r][c] = r*m_cols+c;
    }
  }
  t_t2 = clock() - t_t2;
  double t2_ttaken = ((double)t_t2)/CLOCKS_PER_SEC; 
  printf(" %f sec\n",t2_ttaken);  
  
  printf("3rd --> Search matrix\n");
  clock_t t_t3; t_t3 = clock();
  
  pthread_t t1;
  thread_args *args = malloc(sizeof(thread_args));
  args->start = 0;
  args->end = abs(m_cols/3);
  pthread_create(&t1,NULL,&parallel_search_cols,args);
  
  pthread_t t2;
  thread_args *args2 = malloc(sizeof(thread_args));
  args2->start = abs(m_cols/3);
  args2->end = m_cols-(m_cols/3);
  pthread_create(&t2,NULL,&parallel_search_cols,args2);  
  
  pthread_t t3;
  thread_args *args3 = malloc(sizeof(thread_args));
  args3->start = m_cols-abs((m_cols/3));
  args3->end = m_cols;
  pthread_create(&t3,NULL,&parallel_search_cols,args3);

  pthread_join(t1,NULL);
  pthread_join(t2,NULL);
  pthread_join(t3,NULL);

  t_t3 = clock() - t_t3;
  double t3_ttaken = ((double)t_t3)/CLOCKS_PER_SEC;
  printf(" %f sec\n",t3_ttaken);
 
  printf("Done\n The biggest value found is  %d", hv); 
  free(data);
  free(matrix);
  return 0;
}