Question

Well, this post is the reason for Why is memcpy() and memmove() faster than pointer increments?

I have a custom image structure with 4 channels and I like to extract 3 of them (RGB) into an openCV cv::MAt strcuture with 3 channels.

cv::Mat GetColorImgFromFrame(Image* pimg)
{
    int iHeight = pimg->m_imageInfo.m_height;
    int iWidth = pimg->m_imageInfo.m_width;

    cv::Mat cvmColorImg = cv::Mat::zeros(iHeight, iWidth, CV_8UC3);
    int iDestStep = cvmColorImg.channels(); // 3

    uchar* pucSrc = pimg->m_data;    
    uchar* pucDest = cvmColorImg.data;
    uchar* pucDestLimit = cvmColorImg.data + iHeight*iWidth*iDestStep;

    for (;pucDest < pucDestLimit;)
    {       
        *(pucDest++) = *(pucSrc++);
        *(pucDest++) = *(pucSrc++);
        *(pucDest++) = *(pucSrc++);
        pucSrc++;
    }

    return cvmColorImg;
}

It doesn’t seem much helpful to replace the three inner assignments with memcpy(pucDest, pucSrc, 3). Any suggestions?

Answer 1

Before going into “assembler like” mode by using SIMD intrinsic, you can try to tell your compiler that the source and destination pointer do not alias, using for example the __restrict__ keyword with gcc or __restrict with visual studio