Using opencv cuda library functions on TX1 is very slow at 1FPS

[packtpublishing]

Using opencv cuda library functions on TX1 is very slow at 1FPS

[robotchaoX]

This may be due to the long time it takes for the GPU to start up the first time. If the loop is executed 5 times, you will find that only the first execution takes longer. The code and test results are as follows:

#include <iostream>
#include "opencv2/opencv.hpp"
#include "opencv2/core/cuda.hpp"
//#include <cuda_runtime.h>

int main(int argc, char *argv[])
{
    //Read Two Images
    cv::Mat h_img1 = cv::imread("images/cameraman.tif");
    cv::Mat h_img2 = cv::imread("images/circles.png");
    cv::Mat h_result1;
    // execute 5 times
    for (int i = 0; i < 5; i++)
    {
        int64 work_begin = cv::getTickCount();
        //Create Memory for storing Images on device
        cv::cuda::GpuMat d_result1, d_img1, d_img2;
        //Upload Images to device
        d_img1.upload(h_img1);
        d_img2.upload(h_img2);
        cv::cuda::add(d_img1, d_img2, d_result1);
        //Download Result back to host
        d_result1.download(h_result1);
        int64 delta = cv::getTickCount() - work_begin;
        double freq = cv::getTickFrequency();
        double work_fps = freq / delta;
        std::cout << "Performance of Addition on Jetson TX1: " << std::endl;
        std::cout << "Time: " << (1 / work_fps) << std::endl;
        std::cout << "FPS: " << work_fps << std::endl;
    }
    cv::imshow("Image1 ", h_img1);
    cv::imshow("Image2 ", h_img2);
    cv::imshow("Result addition ", h_result1);
    cv::imwrite("result_add.png", h_result1);
    cv::waitKey();
    return 0;
}

Performance of Addition on Jetson TX1: 
Time: 0.789957
FPS: 1.26589
Performance of Addition on Jetson TX1: 
Time: 0.00194754
FPS: 513.469
Performance of Addition on Jetson TX1: 
Time: 0.00159426
FPS: 627.25
Performance of Addition on Jetson TX1: 
Time: 0.00177029
FPS: 564.879
Performance of Addition on Jetson TX1: 
Time: 0.0016342
FPS: 611.921

[gaurav-packt]

Thanks for the feedback.

With this closing the issue.