为什么OpenCV Gpu模块的性能比VisionWorks快？

Question

I have tried several functions of OpenCv gpu module and compared the same behavior with visionWorks immediate code. And surprisingly, it all circumstances the OpenCv Gpu Module is performing significantly faster than VisionWorks.

eg a Gaussian pyramid of level 4 implemented manually using opencv

#include <iostream>
#include <stdio.h>


#include <stdio.h>
#include <queue>
/* OPENCV RELATED */
#include <cv.h>
#include <highgui.h>
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/stitching/detail/util.hpp"
#include "opencv2/stitching/detail/warpers.hpp"
#include "opencv2/stitching/warpers.hpp"
#include <opencv2/gpu/gpu.hpp>  

#include "opencv2/opencv_modules.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/stitching/detail/autocalib.hpp"
#include "opencv2/stitching/detail/blenders.hpp"
#include "opencv2/stitching/detail/camera.hpp"
#include "opencv2/stitching/detail/exposure_compensate.hpp"
#include "opencv2/stitching/detail/matchers.hpp"
#include "opencv2/stitching/detail/motion_estimators.hpp"
#include "opencv2/stitching/detail/seam_finders.hpp"
#include "opencv2/stitching/detail/util.hpp"
#include "opencv2/stitching/detail/warpers.hpp"
#include "opencv2/stitching/warpers.hpp"
#include <opencv2/opencv.hpp>


using namespace std;
using namespace cv;

using namespace gpu;
using namespace cv::detail;


int main()
{
    Mat m = imread("br1.png");

    GpuMat d_m  = GpuMat (m);
    GpuMat d_m2;
    GpuMat l1,l2,l3,l4;
    int iter = 100;
    int64 e = getTickCount();
    float sum = 0;

    sum = 0;

    for(int i = 0 ; i < iter;  i++)
    {
        e = getTickCount();
        gpu::pyrDown(d_m,l1);
        gpu::pyrDown(l1,l2);
        gpu::pyrDown(l2,l3);
        gpu::pyrDown(l3,l4);
        sum+= (getTickCount() - e) / getTickFrequency(); 
    }

    cout <<"Time taken by Gussian Pyramid Level 4 \t\t\t"<<sum/iter<<" sec"<<endl;

    //imwrite("cv_res.jpg",res);
    return 0;
}

takes 2.5 ms on average for 100 iterations. Whereas, VisionWorks

    #include <VX/vx.h>
#include <VX/vxu.h>
#include <stdio.h>
#include <stdlib.h>
#include <iostream>
#include <stdio.h>


#include <stdio.h>
#include <queue>
/* OPENCV RELATED */
#include <cv.h>
#include <highgui.h>
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/stitching/detail/util.hpp"
#include "opencv2/stitching/detail/warpers.hpp"
#include "opencv2/stitching/warpers.hpp"
#include <opencv2/gpu/gpu.hpp>  

#include "opencv2/opencv_modules.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/stitching/detail/autocalib.hpp"
#include "opencv2/stitching/detail/blenders.hpp"
#include "opencv2/stitching/detail/camera.hpp"
#include "opencv2/stitching/detail/exposure_compensate.hpp"
#include "opencv2/stitching/detail/matchers.hpp"
#include "opencv2/stitching/detail/motion_estimators.hpp"
#include "opencv2/stitching/detail/seam_finders.hpp"
#include "opencv2/stitching/detail/util.hpp"
#include "opencv2/stitching/detail/warpers.hpp"
#include "opencv2/stitching/warpers.hpp"
#include <opencv2/opencv.hpp>


using namespace std;
using namespace cv;

using namespace gpu;
using namespace cv::detail;



vx_image createImageFromMat(vx_context& context, cv::Mat& mat);


vx_status createMatFromImage(vx_image& image, cv::Mat& mat);


/* Entry point. */
int main(int argc,char* argv[])
{

    Mat cv_src1 = imread("br1.png", IMREAD_GRAYSCALE);
  int width = 1280;
  int height = 720;

  int half_width = width/2;
  int half_height = height/2;
    Mat dstMat(cv_src1.size(), cv_src1.type());
  Mat half_dstMat(Size(width/16,height/16),cv_src1.type());

  /* Image data. */


    if (cv_src1.empty() )
    {
        std::cerr << "Can't load input images" << std::endl;
        return -1;
    }


  /* Create our context. */
  vx_context context = vxCreateContext();

  /* Image to process. */
  vx_image image = createImageFromMat(context, cv_src1);
   //NVXIO_CHECK_REFERENCE(image);

  /* Intermediate images. */
  vx_image dx = vxCreateImage(context, width, height, VX_DF_IMAGE_S16);
  vx_image dy = vxCreateImage(context, width, height, VX_DF_IMAGE_S16);
  vx_image mag = vxCreateImage(context, width, height, VX_DF_IMAGE_S16);
  vx_image half_image = vxCreateImage(context, half_width, half_height,  VX_DF_IMAGE_U8);
  vx_image half_image_2 = vxCreateImage(context, half_width/2, half_height/2,  VX_DF_IMAGE_U8);
  vx_image half_image_3 = vxCreateImage(context, half_width/4, half_height/4,  VX_DF_IMAGE_U8);
  vx_image half_image_4 = vxCreateImage(context, half_width/8, half_height/8,  VX_DF_IMAGE_U8);


  int64 e = getTickCount();
  int iter = 100;
  float sum = 0.0;



  e = getTickCount();
  iter = 100;
  for(int i = 0 ; i < iter; i ++)
  {
    /* RESIZEZ OPERATION */
    if(vxuHalfScaleGaussian(context,image,half_image,3) != VX_SUCCESS)
    {
      cout <<"ERROR :"<<"failed to perform scaling"<<endl;
    }

    if(vxuHalfScaleGaussian(context,half_image,half_image_2,3) != VX_SUCCESS)
    {
      cout <<"ERROR :"<<"failed to perform scaling"<<endl;
    }

    if(vxuHalfScaleGaussian(context,half_image_2,half_image_3,3) != VX_SUCCESS)
    {
      cout <<"ERROR :"<<"failed to perform scaling"<<endl;
    }

    if(vxuHalfScaleGaussian(context,half_image_3,half_image_4,3) != VX_SUCCESS)
    {
      cout <<"ERROR :"<<"failed to perform scaling"<<endl;
    }


    sum += (getTickCount() - e) / getTickFrequency();  
  }

  cout <<"Resize to half " <<sum/iter<<endl;

  createMatFromImage(half_image_4,half_dstMat);

  imwrite("RES.jpg",half_dstMat);
  /* Tidy up. */
  vxReleaseImage(&dx);
  vxReleaseImage(&dy);
  vxReleaseImage(&mag);
  vxReleaseContext(&context);
}



vx_image createImageFromMat(vx_context& context, cv::Mat& mat)
{
    vx_imagepatch_addressing_t src_addr = {
        mat.cols, mat.rows, sizeof(vx_uint8), mat.cols * sizeof(vx_uint8), VX_SCALE_UNITY, VX_SCALE_UNITY, 1, 1 };
    void* src_ptr = mat.data;

    vx_image image = vxCreateImageFromHandle(context, VX_DF_IMAGE_U8, &src_addr, &src_ptr, VX_IMPORT_TYPE_HOST);

    return image;
}


vx_status createMatFromImage(vx_image& image, cv::Mat& mat)
{
    vx_status status = VX_SUCCESS;
    vx_uint8 *ptr = NULL;

    cout <<"Creating image "<<mat.cols << " " <<mat.rows <<endl;
    vx_rectangle_t rect;
    vxGetValidRegionImage(image, &rect);
    vx_imagepatch_addressing_t addr = {
        mat.cols, mat.rows, sizeof(vx_uint8), mat.cols * sizeof(vx_uint8), VX_SCALE_UNITY, VX_SCALE_UNITY, 1, 1 };

    status = vxAccessImagePatch(image, &rect, 0, &addr, (void **)&ptr, VX_READ_ONLY);
    mat.data = ptr;

    return status;
}

takes 11.1 ms on single execution, and 96ms on average for 100 iterations.

If this is generally true, then what does visionWorks offer ?

I am running "cuda-repo-l4t-r21.3-6-5-local_6.5-50" version of L4T on Jetson TK1

Answer 1

You've made a mistake in VisionWorks code. You start timer only once e = getTickCount(); right before the loop, but you need to start it on each iteration.

iter = 100;
for(int i = 0 ; i < iter; i ++)
{
    // START TIMER
    e = getTickCount();

    /* RESIZEZ OPERATION */
    if(vxuHalfScaleGaussian(context,image,half_image,3) != VX_SUCCESS)
    {
        cout <<"ERROR :"<<"failed to perform scaling"<<endl;
    }

    if(vxuHalfScaleGaussian(context,half_image,half_image_2,3) != VX_SUCCESS)
    {
        cout <<"ERROR :"<<"failed to perform scaling"<<endl;
    }

    if(vxuHalfScaleGaussian(context,half_image_2,half_image_3,3) != VX_SUCCESS)
    {
        cout <<"ERROR :"<<"failed to perform scaling"<<endl;
    }

    if(vxuHalfScaleGaussian(context,half_image_3,half_image_4,3) != VX_SUCCESS)
    {
        cout <<"ERROR :"<<"failed to perform scaling"<<endl;
    }

    // STOP TIMER
    sum += (getTickCount() - e) / getTickFrequency();  
}

Answer 2

I think that the following code is mistake.

  Mat cv_src1 = imread("br1.png", IMREAD_GRAYSCALE);
  int width = 1280;
  int height = 720;

I think that you should be set as follows.

  Mat cv_src1 = imread("br1.png", IMREAD_GRAYSCALE);
  vx_uint32 width  = cv_src1.cols;
  vx_uint32 height = cv_src1.rows;

And, I made sample code to reproduce.
But, VisionWorks(about 0.3ms) faster than GpuMat(about 0.4ms) on my environment.

https://gist.github.com/atinfinity/9c8c067db739b190ba17f2bd8dbe75d6 https://gist.github.com/atinfinity/e8c2f2da6486be51881e3924c13a311c

My environment is as follows.

• GPU: NVIDIA GeForce GTX 680
• OS: Windows 10 Pro 64bit
• Compiler: Visual Studio 2013 Update5
• VisionWorks：NVIDIA VisionWorks v1.0.25
• OpenCV: OpenCV 3.1