OPENCV：CUDA上下文初始化为不同的方法

我正在开发一个简单的c ++程序来评估一些Opencv GPU方法（cv :: cuda）的性能。
我在Unbuntu 15（使用CUDA 7.5）和GeForce 770上使用Opencv 3.1。

I'm working on a simple c++ program to evaluate the performance of some Opencv GPU methods (cv::cuda). I am using Opencv 3.1 on Unbuntu 15 (with CUDA 7.5) with a GeForce 770.

我以前读过，我们需要初始化CUDA环境，过程在第一次调用。所以，我用cv :: cuda :: getDevice（）和setDevice（）初始化我的程序。

I previously read that we need to initialize CUDA environnement to avoid slow process at first call. So, I initialize my program with a cv::cuda::getDevice() and setDevice().

然后，我测试2个方法：
- cv :: cuda :: resize（）（factor 0.5）
- 和cv :: cuda :: meanStdDev。

Then, I test 2 methods : - cv::cuda::resize() (factor 0.5) - and cv::cuda::meanStdDev.

初始化需要400ms。然后，调整大小需要2或3毫秒，没关系。
但是... meanStdDev需要476ms！
如果我运行两个连续的meanStdDev，第二个更快（3ms）。

Initialization takes 400ms. Then, resizing takes 2 or 3 ms, that's OK. But... meanStdDev takes 476ms !!! If I run 2 successive meanStdDev, the second one is much faster (3ms).

我真的不明白为什么初始化对调整大小（）但不是meanStdDev（）...

I really don't understand why the initialization has an effect on resize() but not on meanStdDev()...

我使用-DCUDA_ARCH_BIN = 3.0编译OPENCV。我尝试用-DCUDA_ARCH_PTX =但问题仍然是一样的。

I compile OPENCV with -DCUDA_ARCH_BIN=3.0. I try with -DCUDA_ARCH_PTX="" but the problem is still the same.

感谢您的帮助。

Pierre。

#include <opencv2/opencv.hpp>
#include <opencv2/cudaimgproc.hpp>
#include "opencv2/cudawarping.hpp"
#include "opencv2/cudaarithm.hpp"

using namespace std;

int main(int argc, char *argv[]) 
{

    double t_init_cuda = (double)cv::getTickCount();
    int CudaDevice;
    if(cv::cuda::getCudaEnabledDeviceCount()==0)
    {
        cerr<<endl<<"ERROR: NO CudaEnabledDevice"<<endl;
        exit(2);
    }
    else
    {
        CudaDevice = cv::cuda::getDevice();
        cv::cuda::setDevice(CudaDevice);
    }
    t_init_cuda = ((double)cv::getTickCount() - t_init_cuda)/cv::getTickFrequency() * 1000;
    cout<<endl<<"\t*T_INIT_CUDA="<<t_init_cuda<<"ms\n";;

    cv::Mat src = cv::imread(argv[1], 0);
    if (!src.data) exit(1);
    cv::cuda::GpuMat d_src(src);


    //CV::CUDA::RESIZE
    cv::cuda::GpuMat d_dst;
    double factor = 0.5;
    double t_gpu_resize = cv::getTickCount();
    cv::cuda::resize(d_src, d_dst, cv::Size( (int) ((float) (d_src.cols)*factor) , (int) ((float) (d_src.rows)*factor)), 0, 0, CV_INTER_AREA);
    t_gpu_resize = ((double)cv::getTickCount() - t_gpu_resize)/cv::getTickFrequency() * 1000;
    cout<<endl<<"D_SRC="<<d_src.rows<<"x"<<d_src.cols<<" => D_DST="<<d_dst.rows<<"x"<<d_dst.cols<<endl;
    cout<<endl<<"\t*T_GPU_RESIZE="<<t_gpu_resize<<"ms\n";;

    //CV::CUDA::MEANSTDDEV
    double t_meanstddev = (double)cv::getTickCount();
    cv::Scalar mean, stddev;
    std::vector<cv::cuda::GpuMat> d_src_split;
    cv::cuda::split(d_src, d_src_split);
    cv::cuda::meanStdDev (d_src_split[0], mean, stddev); 
    t_meanstddev = ((double)cv::getTickCount() - t_meanstddev)/cv::getTickFrequency() * 1000.0;
    cout<<endl<<"mean="<<mean.val[0]<<" | stddev="<<stddev.val[0]<<endl;    
    cout<<endl<<"\t*T_GPU_MEANSTDDEV="<<t_meanstddev<<"ms\n";

    return 0;
}