如何确保CUDA完成了她的计算

Question

我正在尝试使用CUDA为x，y的每个坐标实现KNN（k个最近邻居），如果我在最多7000个坐标上运行项目，之后它只是打印垃圾，是否有办法制作确保CUDA在调用memcpy函数之前完成计算？

主：

runCuda(cudaAllCordDevice, cudaKNearestCord, numOfJobsPerSlave, knn, size, max,  maxThreadNumber, numOfJobsForCuda);
        if (cudaDeviceSynchronize() == cudaSuccess){
            printf("cuda success");
        }else
        {
            printf("cuda fail");
        }
        cudaFinishedCalculation = (Coordinate*)malloc((knn+1)*sizeof(Coordinate)*numOfJobsForCuda);
        cudaMemcpy(cudaFinishedCalculation, cudaKNearestCord, numOfJobsForCuda*((knn+1)*sizeof(Coordinate)), cudaMemcpyDeviceToHost);

和我的内核（cu）：

__global__ void calcNCoordinates(Coordinate* cudaAllCoordArr, Coordinate* cudaKNearest,int startIndex, int knn, int size, Coordinate max, int cudaSizeToCalc)
{
    int i, j ,index;
    Coordinate* cudaTempKNearest = (Coordinate*)malloc((knn+1)*sizeof(Coordinate));
    int threadId = threadIdx.x; //get current thread Id
    index = threadId+startIndex;

    while(threadId < cudaSizeToCalc) //while treadId is smaller then the size to calc for cuda
    {
        calcKnnPerCoodinate(cudaAllCoordArr, cudaTempKNearest, knn, size, index, max);
        for(j = 0 ; j < knn+1 ; j++)
        {
            cudaKNearest[j+threadId*(knn+1)] = cudaTempKNearest[j];
        }
        threadId += blockDim.x; ////number of threads running in a block - 1024 in afeka pc
            index = threadId+startIndex; //update the current index
    }

    free(cudaTempKNearest);
}

void runCuda(Coordinate* cudaAllCoordArr, Coordinate* cudaKNearest,int startIndex, int knn, int size, Coordinate max, int maxNumberOfThreads, int numOfJobsForCuda)
{
    calcNCoordinates<<<1, maxNumberOfThreads>>>(cudaAllCoordArr, cudaKNearest, startIndex, knn, size, max, numOfJobsForCuda);
}

如果我尝试使用最多7k坐标运行它会打印成功，否则打印失败，我尝试cudaSyncronized()。我需要程序继续计算，直到它完成。