cudaEvent_t startGPU, stopGPU, startCPU, stopCPU;
    float GPUtime, CPUtime;

        CUDA_CALL(cudaEventCreate(&startGPU));
        CUDA_CALL(cudaEventCreate(&stopGPU));
        CUDA_CALL(cudaEventRecord(startGPU, 0));
        for (int a = 0; a < 100; a++)
    {
        //do stuff
    }
        CUDA_CALL(cudaEventRecord(stopGPU, 0));
        CUDA_CALL(cudaEventSynchronize(stopGPU));
        cudaEventElapsedTime(&GPUtime, startGPU, stopGPU);


        CUDA_CALL(cudaEventCreate(&startCPU));
        CUDA_CALL(cudaEventCreate(&stopCPU));
        CUDA_CALL(cudaEventRecord(startCPU, 0));
        for (int a = 0; a < 100; a++)
    {
        //do stuff again
    }
        CUDA_CALL(cudaEventRecord(stopCPU, 0));
        CUDA_CALL(cudaEventSynchronize(stopCPU));
        cudaEventElapsedTime(&CPUtime, startCPU, stopCPU);

	printf("%f\n", GPUtime);
	printf("%f\n", CPUtime);

        CUDA_CALL(cudaEventDestroy(startGPU));
        CUDA_CALL(cudaEventDestroy(stopGPU));
        CUDA_CALL(cudaEventDestroy(startCPU));
        CUDA_CALL(cudaEventDestroy(stopCPU));
    	getchar();
        exit(EXIT_SUCCESS);
}