cudaDeviceSynchronize的优化

#include<iostream>
#include<cuda.h>
#include<cuda_runtime.h>
#include<time.h>


using namespace std;


__global__ void nullKernel()
{}




int main(void)
{
const int sum=1000000;
cout<<"Launching..."<<endl;
cudaEvent_t start,stop;
float time;
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start,0);
cout<<"Launching..."<<endl;
for(int i=0;i<10000;i++)
{
nullKernel<<<1,1>>>();
//cudaDeviceSynchronize();
}cudaDeviceSynchronize();
cout<<"Launching..."<<endl;
cudaEventRecord(stop,0);
cudaEventSynchronize(stop);
cout<<"Launching..."<<endl;
cudaEventElapsedTime(&time,start,stop);
cout<<time<<endl;
return 0;

}

图一是将cudaDeviceSynchronize放在for循环外面，而第二张图则是将cudaDeviceSynchronize放在了for循环里面，从而所耗时相差十倍。