// nvcc -arch=sm_13 -o cpu_norm_const cpu_norm_const.cu

#include "../../cuda_by_example/common/book.h"


const int B_size = 1000000;
const int A_size = 15131;
//const int B_size = 100;
//const int A_size = 1000;

int main( void ) {
    unsigned long int   *a, *b; 
    int	c, i, j;

    cudaEvent_t     start, stop;

    // allocate memory on the cpu side
    a = (unsigned long int*)malloc( A_size*2*sizeof(unsigned long int) );
    b = (unsigned long int*)malloc( B_size*2*sizeof(unsigned long int) );

    FILE	*Ain, *Bin;
    Ain = fopen("A.txt", "r");
    Bin = fopen("B.txt", "r");

    unsigned long int	tmp1, tmp2;

    // fill in the host memory with data
    for (i=0; i<B_size; i++) {
       if(i < A_size) {
          fscanf(Ain, "%lu %lu\n", &tmp1, &tmp2);
	     a[i * 2 + 0] = tmp1;
	     a[i * 2 + 1] = tmp2;
       }

	  fscanf(Bin, "%lu %lu\n", &tmp1, &tmp2);
	  b[i * 2 + 0] = tmp1;
       b[i * 2 + 1] = tmp2;
       //b[i * 2 + 0] = 412316861472;
       //b[i * 2 + 1] = 2199040041218;
    }

    fclose(Ain); fclose(Bin);

    HANDLE_ERROR( cudaEventCreate( &start ) );
    HANDLE_ERROR( cudaEventCreate( &stop ) );
    HANDLE_ERROR( cudaEventRecord( start, 0 ) );

    // finish up on the CPU side
    c = 0;
    for(i=0; i<B_size; i++)
    {
       for(j=0; j<A_size; j++)
	  {
	     //printf("[%d, %d]: [%lu , %lu]\t[%lu , %lu]\n",i,j,b[i*2], a[j*2], b[i*2+1],a[j*2+1]);
	     if((b[i*2] == a[j*2]) && (b[i*2+1] == a[j*2+1]))
		{
             c++;
		   break;
		}
	  }
    }

    HANDLE_ERROR( cudaEventRecord( stop, 0 ) );
    HANDLE_ERROR( cudaEventSynchronize( stop ) );
    float   elapsedTime;
    HANDLE_ERROR( cudaEventElapsedTime( &elapsedTime,
    								start, stop ) );
    printf( "Time elapsed:  %3.1f ms\n", elapsedTime );

    HANDLE_ERROR( cudaEventDestroy( start ) );
    HANDLE_ERROR( cudaEventDestroy( stop ) );



    FILE *fout;
    fout = fopen("frequency.txt", "w");
    fprintf(fout, "%d\t%lf\n", c, (double)c / (double)B_size);
    fclose(fout);

    printf("Frequency = %d / %d = %lf\n\n", c, B_size, (double)c / B_size);


    // free memory on the cpu side
    free( a );
    free( b );
}