modified_oneKernel_integralImage2D2D.c

    /********************************
    Author: Sravanthi Kota Venkata
    ********************************/

    #include <stdio.h>
    #include <stdlib.h>
    #include "disparity.h"
    #include "m5op.h"
    void integralImage2D2D(F2D* SAD, F2D* integralImg)
    {
        int nr, nc, i, j;

        nr = SAD->height;
        nc = SAD->width;
        int togo,size;
        for(i=0; i<nc; i++)
            subsref(integralImg,0,i) = subsref(SAD,0,i);

          /* original code
          for(i=1; i<nr; i++)
                for(j=0; j<nc; j++)
                {
                    subsref(integralImg,i,j) = subsref(integralImg, (i-1), j) + subsref(SAD,i,j);
                }
           */
          m5rh_start(101,1);
          int N=nr*nc;
          for(i=nr;i<N; i+=8)
          {
            togo = N - i;
            size = (togo > 8) ? 8 : togo;
            m5_rh_load_data(integralImg->data+i, size*4, 4, 1); //load data as many data elements as the number of size (each int/float size is 4 byte)
            m5_rh_load_data(SAD->data+i, size*4, 4, 1); //load data as many data elements as the number of size (each int/float size is 4 byte)
            m5_rh_load_data(0, 1, 5 , 0); //one addition

            for(j=0; j<size; j++)
            {
                asubsref(integralImg,i+j) = asubsref(integralImg, i-nc +j) + asubsref(SAD,i+j);
            }
            m5_rh_load_data(integramImg->data+i, size*4, 4, 1); //store results
            m5_rh_load_data(0, 0, 0, 0);
          }


         /*original code
          for(i=0; i<nr; i++)
            for(j=1; j<nc; j++)
                subsref(integralImg,i,j) = subsref(integralImg, i, (j-1)) + subsref(integralImg,i,j);
           */
          for(i=0; i<N; i+=8)
                {
                  togo = N - i;
                  size = (togo > 8) ? 8 : togo;
                  m5_rh_load_data(integralImg->data+i, size*4, 4, 1); //load data as many data elements as the number of size (each int/float size is 4 byte)
                  m5_rh_load_data(0, 1, 5 , 0); //one dependent addition
                  for(j=0; j<size; j++)
                  {
                        if (((i+j) % nc) != 0)
                           {
                              asubsref(integralImg,i+j) = asubsref(integralImg, i+j-1) + asubsref(integralImg,i+j);

                           }
                  }
                  m5_rh_load_data(integramImg->data+i, size*4, 4, 1); //store results
                  m5_rh_load_data(0, 0, 0, 0);

                }
         m5rh_start(101,0);

        return;

    }