OpenCV:reduce源碼解析

1.Demo

/**
 * @file core_reduce.cpp
 * @brief It demonstrates the usage of cv::reduce .
 *
 * It shows how to compute the row sum, column sum, row average,
 * column average, row minimum, column minimum, row maximum
 * and column maximum of a cv::Mat.
 *
 * @author KUANG Fangjun
 * @date August 2017
 */

#include <iostream>
#include <opencv2/core.hpp>

using namespace std;
using namespace cv;

int main()
{
    {
        //! [example]
        Mat m = (Mat_<uchar>(3,2) << 1,2,3,4,5,6);
        Mat col_sum, row_sum;
        reduce(m, col_sum, 0, REDUCE_SUM, CV_32F);
        reduce(m, row_sum, 1, REDUCE_SUM, CV_32F);
        /*
        m =
        [  1,   2;
           3,   4;
           5,   6]
        col_sum =
        [9, 12]
        row_sum =
        [3;
         7;
         11]
         */
        //! [example]

        Mat col_average, row_average, col_min, col_max, row_min, row_max;
        reduce(m, col_average, 0, REDUCE_AVG, CV_32F);
        cout << "col_average =\n" << col_average << endl;

        reduce(m, row_average, 1, REDUCE_AVG, CV_32F);
        cout << "row_average =\n" << row_average << endl;

        reduce(m, col_min, 0, REDUCE_MIN, CV_8U);
        cout << "col_min =\n" << col_min << endl;

        reduce(m, row_min, 1, REDUCE_MIN, CV_8U);
        cout << "row_min =\n" << row_min << endl;

        reduce(m, col_max, 0, REDUCE_MAX, CV_8U);
        cout << "col_max =\n" << col_max << endl;

        reduce(m, row_max, 1, REDUCE_MAX, CV_8U);
        cout << "row_max =\n" << row_max << endl;
    }

    {
        char d[] = {1,2,3,4,5,6};
        Mat m(3, 1, CV_8UC2, d);
        Mat col_sum_per_channel;
        reduce(m, col_sum_per_channel, 0, REDUCE_SUM, CV_32F);
    }

    return 0;
}

2.reduce源碼

matrix_operations.cpp文件

#define reduceSumR8u32s  reduceR_<uchar, int,   OpAdd<int> >
#define reduceSumR8u32f  reduceR_<uchar, float, OpAdd<int> >
#define reduceSumR8u64f  reduceR_<uchar, double,OpAdd<int> >
#define reduceSumR16u32f reduceR_<ushort,float, OpAdd<float> >
#define reduceSumR16u64f reduceR_<ushort,double,OpAdd<double> >
#define reduceSumR16s32f reduceR_<short, float, OpAdd<float> >
#define reduceSumR16s64f reduceR_<short, double,OpAdd<double> >
#define reduceSumR32f32f reduceR_<float, float, OpAdd<float> >
#define reduceSumR32f64f reduceR_<float, double,OpAdd<double> >
#define reduceSumR64f64f reduceR_<double,double,OpAdd<double> >

#define reduceSumC8u32s  reduceC_<uchar, int,   OpAdd<int> >
#define reduceSumC8u32f  reduceC_<uchar, float, OpAdd<int> >
#define reduceSumC16u32f reduceC_<ushort,float, OpAdd<float> >
#define reduceSumC16s32f reduceC_<short, float, OpAdd<float> >
#define reduceSumC32f32f reduceC_<float, float, OpAdd<float> >
#define reduceSumC64f64f reduceC_<double,double,OpAdd<double> >

//reduce rows
template<typename T, typename ST, class Op> static void
reduceR_( const Mat& srcmat, Mat& dstmat )
{
    typedef typename Op::rtype WT;
    Size size = srcmat.size();
    size.width *= srcmat.channels();
    AutoBuffer<WT> buffer(size.width);
    WT* buf = buffer.data();
    ST* dst = dstmat.ptr<ST>();
    const T* src = srcmat.ptr<T>();
    size_t srcstep = srcmat.step/sizeof(src[0]);
    int i;
    Op op;

    for( i = 0; i < size.width; i++ )
        buf[i] = src[i];

    for( ; --size.height; )
    {
        src += srcstep;
        i = 0;
        #if CV_ENABLE_UNROLLED
        for(; i <= size.width - 4; i += 4 )
        {
            WT s0, s1;
            s0 = op(buf[i], (WT)src[i]);
            s1 = op(buf[i+1], (WT)src[i+1]);
            buf[i] = s0; buf[i+1] = s1;

            s0 = op(buf[i+2], (WT)src[i+2]);
            s1 = op(buf[i+3], (WT)src[i+3]);
            buf[i+2] = s0; buf[i+3] = s1;
        }
        #endif
        for( ; i < size.width; i++ )
            buf[i] = op(buf[i], (WT)src[i]);
    }

    for( i = 0; i < size.width; i++ )
        dst[i] = (ST)buf[i];
}

//reduce cols
template<typename T, typename ST, class Op> static void
reduceC_( const Mat& srcmat, Mat& dstmat )
{
    typedef typename Op::rtype WT;
    Size size = srcmat.size();
    int cn = srcmat.channels();
    size.width *= cn;
    Op op;

    for( int y = 0; y < size.height; y++ )
    {
        const T* src = srcmat.ptr<T>(y);
        ST* dst = dstmat.ptr<ST>(y);
        if( size.width == cn )
            for( int k = 0; k < cn; k++ )
                dst[k] = src[k];
        else
        {
            for( int k = 0; k < cn; k++ )
            {
                WT a0 = src[k], a1 = src[k+cn];
                int i;
                for( i = 2*cn; i <= size.width - 4*cn; i += 4*cn )
                {
                    a0 = op(a0, (WT)src[i+k]);
                    a1 = op(a1, (WT)src[i+k+cn]);
                    a0 = op(a0, (WT)src[i+k+cn*2]);
                    a1 = op(a1, (WT)src[i+k+cn*3]);
                }

                for( ; i < size.width; i += cn )
                {
                    a0 = op(a0, (WT)src[i+k]);
                }
                a0 = op(a0, a1);
                dst[k] = (ST)a0;
            }
        }
    }
}

void cv::reduce(InputArray _src, OutputArray _dst, int dim, int op, int dtype)
{
    CV_INSTRUMENT_REGION();

    CV_Assert( _src.dims() <= 2 );
    int op0 = op;
    int stype = _src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype);
    if( dtype < 0 )
        dtype = _dst.fixedType() ? _dst.type() : stype;
    dtype = CV_MAKETYPE(dtype >= 0 ? dtype : stype, cn);
    int ddepth = CV_MAT_DEPTH(dtype);

    CV_Assert( cn == CV_MAT_CN(dtype) );
    CV_Assert( op == CV_REDUCE_SUM || op == CV_REDUCE_MAX ||
               op == CV_REDUCE_MIN || op == CV_REDUCE_AVG );

    CV_OCL_RUN(_dst.isUMat(),
               ocl_reduce(_src, _dst, dim, op, op0, stype, dtype))

    // Fake reference to source. Resolves issue 8693 in case of src == dst.
    UMat srcUMat;
    if (_src.isUMat())
        srcUMat = _src.getUMat();

    Mat src = _src.getMat();
    _dst.create(dim == 0 ? 1 : src.rows, dim == 0 ? src.cols : 1, dtype);
    Mat dst = _dst.getMat(), temp = dst;

    if( op == CV_REDUCE_AVG )
    {
        op = CV_REDUCE_SUM;
        if( sdepth < CV_32S && ddepth < CV_32S )
        {
            temp.create(dst.rows, dst.cols, CV_32SC(cn));
            ddepth = CV_32S;
        }
    }

    ReduceFunc func = 0;
    if( dim == 0 )
    {
        if( op == CV_REDUCE_SUM )
        {
            if(sdepth == CV_8U && ddepth == CV_32S)
                func = GET_OPTIMIZED(reduceSumR8u32s);
            else if(sdepth == CV_8U && ddepth == CV_32F)
                func = reduceSumR8u32f;
            else if(sdepth == CV_8U && ddepth == CV_64F)
                func = reduceSumR8u64f;
            else if(sdepth == CV_16U && ddepth == CV_32F)
                func = reduceSumR16u32f;
            else if(sdepth == CV_16U && ddepth == CV_64F)
                func = reduceSumR16u64f;
            else if(sdepth == CV_16S && ddepth == CV_32F)
                func = reduceSumR16s32f;
            else if(sdepth == CV_16S && ddepth == CV_64F)
                func = reduceSumR16s64f;
            else if(sdepth == CV_32F && ddepth == CV_32F)
                func = GET_OPTIMIZED(reduceSumR32f32f);
            else if(sdepth == CV_32F && ddepth == CV_64F)
                func = reduceSumR32f64f;
            else if(sdepth == CV_64F && ddepth == CV_64F)
                func = reduceSumR64f64f;
        }
        else if(op == CV_REDUCE_MAX)
        {
            if(sdepth == CV_8U && ddepth == CV_8U)
                func = GET_OPTIMIZED(reduceMaxR8u);
            else if(sdepth == CV_16U && ddepth == CV_16U)
                func = reduceMaxR16u;
            else if(sdepth == CV_16S && ddepth == CV_16S)
                func = reduceMaxR16s;
            else if(sdepth == CV_32F && ddepth == CV_32F)
                func = GET_OPTIMIZED(reduceMaxR32f);
            else if(sdepth == CV_64F && ddepth == CV_64F)
                func = reduceMaxR64f;
        }
        else if(op == CV_REDUCE_MIN)
        {
            if(sdepth == CV_8U && ddepth == CV_8U)
                func = GET_OPTIMIZED(reduceMinR8u);
            else if(sdepth == CV_16U && ddepth == CV_16U)
                func = reduceMinR16u;
            else if(sdepth == CV_16S && ddepth == CV_16S)
                func = reduceMinR16s;
            else if(sdepth == CV_32F && ddepth == CV_32F)
                func = GET_OPTIMIZED(reduceMinR32f);
            else if(sdepth == CV_64F && ddepth == CV_64F)
                func = reduceMinR64f;
        }
    }
    else
    {
        if(op == CV_REDUCE_SUM)
        {
            if(sdepth == CV_8U && ddepth == CV_32S)
                func = GET_OPTIMIZED(reduceSumC8u32s);
            else if(sdepth == CV_8U && ddepth == CV_32F)
                func = reduceSumC8u32f;
            else if(sdepth == CV_8U && ddepth == CV_64F)
                func = reduceSumC8u64f;
            else if(sdepth == CV_16U && ddepth == CV_32F)
                func = reduceSumC16u32f;
            else if(sdepth == CV_16U && ddepth == CV_64F)
                func = reduceSumC16u64f;
            else if(sdepth == CV_16S && ddepth == CV_32F)
                func = reduceSumC16s32f;
            else if(sdepth == CV_16S && ddepth == CV_64F)
                func = reduceSumC16s64f;
            else if(sdepth == CV_32F && ddepth == CV_32F)
                func = GET_OPTIMIZED(reduceSumC32f32f);
            else if(sdepth == CV_32F && ddepth == CV_64F)
                func = reduceSumC32f64f;
            else if(sdepth == CV_64F && ddepth == CV_64F)
                func = reduceSumC64f64f;
        }
        else if(op == CV_REDUCE_MAX)
        {
            if(sdepth == CV_8U && ddepth == CV_8U)
                func = GET_OPTIMIZED(reduceMaxC8u);
            else if(sdepth == CV_16U && ddepth == CV_16U)
                func = reduceMaxC16u;
            else if(sdepth == CV_16S && ddepth == CV_16S)
                func = reduceMaxC16s;
            else if(sdepth == CV_32F && ddepth == CV_32F)
                func = GET_OPTIMIZED(reduceMaxC32f);
            else if(sdepth == CV_64F && ddepth == CV_64F)
                func = reduceMaxC64f;
        }
        else if(op == CV_REDUCE_MIN)
        {
            if(sdepth == CV_8U && ddepth == CV_8U)
                func = GET_OPTIMIZED(reduceMinC8u);
            else if(sdepth == CV_16U && ddepth == CV_16U)
                func = reduceMinC16u;
            else if(sdepth == CV_16S && ddepth == CV_16S)
                func = reduceMinC16s;
            else if(sdepth == CV_32F && ddepth == CV_32F)
                func = GET_OPTIMIZED(reduceMinC32f);
            else if(sdepth == CV_64F && ddepth == CV_64F)
                func = reduceMinC64f;
        }
    }

    if( !func )
        CV_Error( CV_StsUnsupportedFormat,
                  "Unsupported combination of input and output array formats" );

    func( src, temp );

    if( op0 == CV_REDUCE_AVG )
        temp.convertTo(dst, dst.type(), 1./(dim == 0 ? src.rows : src.cols));
}

3.參考文獻

https://github.com/opencv/opencv/blob/b39cd06249213220e802bb64260727711d9fc98c/samples/cpp/tutorial_code/snippets/core_reduce.cpp

https://github.com/opencv/opencv/blob/b39cd06249213220e802bb64260727711d9fc98c/modules/core/src/matrix_operations.cpp