#include <stdlib.h>
#include <stdio.h>
#include "cutil.h"

__global__ void kernel_launch(int size)
{
}


int main(int argc, char** argv)
{
        CUT_DEVICE_INIT(argc,argv);

        int size= 65536 * 512;//512 * 8192 * 8;


        unsigned int timer=0,timer1=0;
        CUT_SAFE_CALL( cutCreateTimer( &timer));
        CUT_SAFE_CALL( cutCreateTimer( &timer1));

        float avg_timer=0,avg_timer1=0;
        for(int i=0;i<100;i++ )
        {
                CUT_SAFE_CALL( cutStartTimer( timer));
                kernel_launch<<<size/512,512>>>(size);
                cudaThreadSynchronize();
                CUT_SAFE_CALL( cutStopTimer( timer));
                avg_timer += cutGetTimerValue( timer);
        }
        for(int i=0;i<200;i++ )
        {
                CUT_SAFE_CALL( cutStartTimer( timer1));
                kernel_launch<<<size/(512*2),512>>>(size);
                cudaThreadSynchronize();
                CUT_SAFE_CALL( cutStopTimer( timer1));
                avg_timer1 += cutGetTimerValue( timer1);
        }

        printf( "Time with num of blocks = %d kernel launch %f (ms)\n",size / 512, avg_timer / 100);
        printf( "Time with num of blocks = %d kernel launch %f (ms)\n",size / (512*2) , avg_timer1 / 100);

}