为什么OpenCV Gpu模块的性能比VisionWorks快？

Question

我已经尝试了OpenCv gpu模块的几种功能，并将相同的行为与visionWorks立即代码进行了比较。 令人奇怪的是，这一切的情况下OpenCV的GPU模块正在执行比VisionWorks 显著快。

例如，使用opencv手动实现的4级高斯金字塔

#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;
}

100次迭代平均需要2.5毫秒。 而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;
}

一次执行需要11.1毫秒，而100次迭代平均需要96毫秒。

如果这通常是正确的，那么visionWorks提供什么？

我在Jetson TK1上运行L4T的“ cuda-repo-l4t-r21.3-6-5-local_6.5-50”版本

Answer 1

您在VisionWorks代码中犯了一个错误。 您只能启动一次计时器e = getTickCount(); 就在循环之前，但是您需要在每次迭代时启动它。

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

我认为以下代码是错误的。

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

我认为您应该设置如下。

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

而且，我制作了示例代码以进行重现。
但是，在我的环境中，VisionWorks（约0.3毫秒）比GpuMat（约0.4毫秒）快。

https://gist.github.com/atinfinity/9c8c067db739b190ba17f2bd8dbe75d6 https://gist.github.com/atinfinity/e8c2f2da2486648be51881e3924c13a311c

我的环境如下。

• GPU：NVIDIA GeForce GTX 680
• 操作系统：Windows 10 Pro 64bit
• 编译器：Visual Studio 2013 Update5
• VisionWorks：NVIDIA VisionWorks v1.0.25
• OpenCV：OpenCV 3.1