Merge pull request #14013 from alalek:imgproc_dispatch_filter

modules/imgproc/CMakeLists.txt

@@ -1,6 +1,12 @@
 set(the_description "Image Processing")
 ocv_add_dispatched_file(accum SSE4_1 AVX AVX2)
+ocv_add_dispatched_file(bilateral_filter SSE2 AVX2)
+ocv_add_dispatched_file(box_filter SSE2 SSE4_1 AVX2)
+ocv_add_dispatched_file(filter SSE2 SSE4_1 AVX2)
 ocv_add_dispatched_file(color_hsv SSE2 SSE4_1 AVX2)
 ocv_add_dispatched_file(color_rgb SSE2 SSE4_1 AVX2)
 ocv_add_dispatched_file(color_yuv SSE2 SSE4_1 AVX2)
+ocv_add_dispatched_file(median_blur SSE2 SSE4_1 AVX2)
+ocv_add_dispatched_file(morph SSE2 SSE4_1 AVX2)
+ocv_add_dispatched_file(smooth SSE2 SSE4_1 AVX2)
 ocv_define_module(imgproc opencv_core WRAP java python js)

modules/imgproc/src/bilateral_filter.dispatch.cpp

@@ -0,0 +1,427 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, 2018, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Copyright (C) 2014-2015, Itseez Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#include "precomp.hpp"
+
+#include <vector>
+
+#include "opencv2/core/hal/intrin.hpp"
+#include "opencl_kernels_imgproc.hpp"
+
+#include "bilateral_filter.simd.hpp"
+#include "bilateral_filter.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
+
+/****************************************************************************************\
+                                   Bilateral Filtering
+\****************************************************************************************/
+
+namespace cv {
+
+#ifdef HAVE_OPENCL
+
+static bool ocl_bilateralFilter_8u(InputArray _src, OutputArray _dst, int d,
+                                   double sigma_color, double sigma_space,
+                                   int borderType)
+{
+    CV_INSTRUMENT_REGION();
+#ifdef __ANDROID__
+    if (ocl::Device::getDefault().isNVidia())
+        return false;
+#endif
+
+    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+    int i, j, maxk, radius;
+
+    if (depth != CV_8U || cn > 4)
+        return false;
+
+    if (sigma_color <= 0)
+        sigma_color = 1;
+    if (sigma_space <= 0)
+        sigma_space = 1;
+
+    double gauss_color_coeff = -0.5 / (sigma_color * sigma_color);
+    double gauss_space_coeff = -0.5 / (sigma_space * sigma_space);
+
+    if ( d <= 0 )
+        radius = cvRound(sigma_space * 1.5);
+    else
+        radius = d / 2;
+    radius = MAX(radius, 1);
+    d = radius * 2 + 1;
+
+    UMat src = _src.getUMat(), dst = _dst.getUMat(), temp;
+    if (src.u == dst.u)
+        return false;
+
+    copyMakeBorder(src, temp, radius, radius, radius, radius, borderType);
+    std::vector<float> _space_weight(d * d);
+    std::vector<int> _space_ofs(d * d);
+    float * const space_weight = &_space_weight[0];
+    int * const space_ofs = &_space_ofs[0];
+
+    // initialize space-related bilateral filter coefficients
+    for( i = -radius, maxk = 0; i <= radius; i++ )
+        for( j = -radius; j <= radius; j++ )
+        {
+            double r = std::sqrt((double)i * i + (double)j * j);
+            if ( r > radius )
+                continue;
+            space_weight[maxk] = (float)std::exp(r * r * gauss_space_coeff);
+            space_ofs[maxk++] = (int)(i * temp.step + j * cn);
+        }
+
+    char cvt[3][40];
+    String cnstr = cn > 1 ? format("%d", cn) : "";
+    String kernelName("bilateral");
+    size_t sizeDiv = 1;
+    if ((ocl::Device::getDefault().isIntel()) &&
+        (ocl::Device::getDefault().type() == ocl::Device::TYPE_GPU))
+    {
+            //Intel GPU
+            if (dst.cols % 4 == 0 && cn == 1) // For single channel x4 sized images.
+            {
+                kernelName = "bilateral_float4";
+                sizeDiv = 4;
+            }
+     }
+     ocl::Kernel k(kernelName.c_str(), ocl::imgproc::bilateral_oclsrc,
+            format("-D radius=%d -D maxk=%d -D cn=%d -D int_t=%s -D uint_t=uint%s -D convert_int_t=%s"
+            " -D uchar_t=%s -D float_t=%s -D convert_float_t=%s -D convert_uchar_t=%s -D gauss_color_coeff=(float)%f",
+            radius, maxk, cn, ocl::typeToStr(CV_32SC(cn)), cnstr.c_str(),
+            ocl::convertTypeStr(CV_8U, CV_32S, cn, cvt[0]),
+            ocl::typeToStr(type), ocl::typeToStr(CV_32FC(cn)),
+            ocl::convertTypeStr(CV_32S, CV_32F, cn, cvt[1]),
+            ocl::convertTypeStr(CV_32F, CV_8U, cn, cvt[2]), gauss_color_coeff));
+    if (k.empty())
+        return false;
+
+    Mat mspace_weight(1, d * d, CV_32FC1, space_weight);
+    Mat mspace_ofs(1, d * d, CV_32SC1, space_ofs);
+    UMat ucolor_weight, uspace_weight, uspace_ofs;
+
+    mspace_weight.copyTo(uspace_weight);
+    mspace_ofs.copyTo(uspace_ofs);
+
+    k.args(ocl::KernelArg::ReadOnlyNoSize(temp), ocl::KernelArg::WriteOnly(dst),
+           ocl::KernelArg::PtrReadOnly(uspace_weight),
+           ocl::KernelArg::PtrReadOnly(uspace_ofs));
+
+    size_t globalsize[2] = { (size_t)dst.cols / sizeDiv, (size_t)dst.rows };
+    return k.run(2, globalsize, NULL, false);
+}
+#endif
+
+
+static void
+bilateralFilter_8u( const Mat& src, Mat& dst, int d,
+    double sigma_color, double sigma_space,
+    int borderType )
+{
+    CV_INSTRUMENT_REGION();
+
+    int cn = src.channels();
+    int i, j, maxk, radius;
+
+    CV_Assert( (src.type() == CV_8UC1 || src.type() == CV_8UC3) && src.data != dst.data );
+
+    if( sigma_color <= 0 )
+        sigma_color = 1;
+    if( sigma_space <= 0 )
+        sigma_space = 1;
+
+    double gauss_color_coeff = -0.5/(sigma_color*sigma_color);
+    double gauss_space_coeff = -0.5/(sigma_space*sigma_space);
+
+    if( d <= 0 )
+        radius = cvRound(sigma_space*1.5);
+    else
+        radius = d/2;
+    radius = MAX(radius, 1);
+    d = radius*2 + 1;
+
+    Mat temp;
+    copyMakeBorder( src, temp, radius, radius, radius, radius, borderType );
+
+    std::vector<float> _color_weight(cn*256);
+    std::vector<float> _space_weight(d*d);
+    std::vector<int> _space_ofs(d*d);
+    float* color_weight = &_color_weight[0];
+    float* space_weight = &_space_weight[0];
+    int* space_ofs = &_space_ofs[0];
+
+    // initialize color-related bilateral filter coefficients
+
+    for( i = 0; i < 256*cn; i++ )
+        color_weight[i] = (float)std::exp(i*i*gauss_color_coeff);
+
+    // initialize space-related bilateral filter coefficients
+    for( i = -radius, maxk = 0; i <= radius; i++ )
+    {
+        j = -radius;
+
+        for( ; j <= radius; j++ )
+        {
+            double r = std::sqrt((double)i*i + (double)j*j);
+            if( r > radius )
+                continue;
+            space_weight[maxk] = (float)std::exp(r*r*gauss_space_coeff);
+            space_ofs[maxk++] = (int)(i*temp.step + j*cn);
+        }
+    }
+
+    CV_CPU_DISPATCH(bilateralFilterInvoker_8u, (dst, temp, radius, maxk, space_ofs, space_weight, color_weight),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+
+static void
+bilateralFilter_32f( const Mat& src, Mat& dst, int d,
+                     double sigma_color, double sigma_space,
+                     int borderType )
+{
+    CV_INSTRUMENT_REGION();
+
+    int cn = src.channels();
+    int i, j, maxk, radius;
+    double minValSrc=-1, maxValSrc=1;
+    const int kExpNumBinsPerChannel = 1 << 12;
+    int kExpNumBins = 0;
+    float lastExpVal = 1.f;
+    float len, scale_index;
+
+    CV_Assert( (src.type() == CV_32FC1 || src.type() == CV_32FC3) && src.data != dst.data );
+
+    if( sigma_color <= 0 )
+        sigma_color = 1;
+    if( sigma_space <= 0 )
+        sigma_space = 1;
+
+    double gauss_color_coeff = -0.5/(sigma_color*sigma_color);
+    double gauss_space_coeff = -0.5/(sigma_space*sigma_space);
+
+    if( d <= 0 )
+        radius = cvRound(sigma_space*1.5);
+    else
+        radius = d/2;
+    radius = MAX(radius, 1);
+    d = radius*2 + 1;
+    // compute the min/max range for the input image (even if multichannel)
+
+    minMaxLoc( src.reshape(1), &minValSrc, &maxValSrc );
+    if(std::abs(minValSrc - maxValSrc) < FLT_EPSILON)
+    {
+        src.copyTo(dst);
+        return;
+    }
+
+    // temporary copy of the image with borders for easy processing
+    Mat temp;
+    copyMakeBorder( src, temp, radius, radius, radius, radius, borderType );
+
+    // allocate lookup tables
+    std::vector<float> _space_weight(d*d);
+    std::vector<int> _space_ofs(d*d);
+    float* space_weight = &_space_weight[0];
+    int* space_ofs = &_space_ofs[0];
+
+    // assign a length which is slightly more than needed
+    len = (float)(maxValSrc - minValSrc) * cn;
+    kExpNumBins = kExpNumBinsPerChannel * cn;
+    std::vector<float> _expLUT(kExpNumBins+2);
+    float* expLUT = &_expLUT[0];
+
+    scale_index = kExpNumBins/len;
+
+    // initialize the exp LUT
+    for( i = 0; i < kExpNumBins+2; i++ )
+    {
+        if( lastExpVal > 0.f )
+        {
+            double val =  i / scale_index;
+            expLUT[i] = (float)std::exp(val * val * gauss_color_coeff);
+            lastExpVal = expLUT[i];
+        }
+        else
+            expLUT[i] = 0.f;
+    }
+
+    // initialize space-related bilateral filter coefficients
+    for( i = -radius, maxk = 0; i <= radius; i++ )
+        for( j = -radius; j <= radius; j++ )
+        {
+            double r = std::sqrt((double)i*i + (double)j*j);
+            if( r > radius || ( i == 0 && j == 0 ) )
+                continue;
+            space_weight[maxk] = (float)std::exp(r*r*gauss_space_coeff);
+            space_ofs[maxk++] = (int)(i*(temp.step/sizeof(float)) + j*cn);
+        }
+
+    // parallel_for usage
+    CV_CPU_DISPATCH(bilateralFilterInvoker_32f, (cn, radius, maxk, space_ofs, temp, dst, scale_index, space_weight, expLUT),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+#ifdef HAVE_IPP
+#define IPP_BILATERAL_PARALLEL 1
+
+#ifdef HAVE_IPP_IW
+class ipp_bilateralFilterParallel: public ParallelLoopBody
+{
+public:
+    ipp_bilateralFilterParallel(::ipp::IwiImage &_src, ::ipp::IwiImage &_dst, int _radius, Ipp32f _valSquareSigma, Ipp32f _posSquareSigma, ::ipp::IwiBorderType _borderType, bool *_ok):
+        src(_src), dst(_dst)
+    {
+        pOk = _ok;
+
+        radius          = _radius;
+        valSquareSigma  = _valSquareSigma;
+        posSquareSigma  = _posSquareSigma;
+        borderType      = _borderType;
+
+        *pOk = true;
+    }
+    ~ipp_bilateralFilterParallel() {}
+
+    virtual void operator() (const Range& range) const CV_OVERRIDE
+    {
+        if(*pOk == false)
+            return;
+
+        try
+        {
+            ::ipp::IwiTile tile = ::ipp::IwiRoi(0, range.start, dst.m_size.width, range.end - range.start);
+            CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterBilateral, src, dst, radius, valSquareSigma, posSquareSigma, ::ipp::IwDefault(), borderType, tile);
+        }
+        catch(const ::ipp::IwException &)
+        {
+            *pOk = false;
+            return;
+        }
+    }
+private:
+    ::ipp::IwiImage &src;
+    ::ipp::IwiImage &dst;
+
+    int                  radius;
+    Ipp32f               valSquareSigma;
+    Ipp32f               posSquareSigma;
+    ::ipp::IwiBorderType borderType;
+
+    bool  *pOk;
+    const ipp_bilateralFilterParallel& operator= (const ipp_bilateralFilterParallel&);
+};
+#endif
+
+static bool ipp_bilateralFilter(Mat &src, Mat &dst, int d, double sigmaColor, double sigmaSpace, int borderType)
+{
+#ifdef HAVE_IPP_IW
+    CV_INSTRUMENT_REGION_IPP();
+
+    int         radius         = IPP_MAX(((d <= 0)?cvRound(sigmaSpace*1.5):d/2), 1);
+    Ipp32f      valSquareSigma = (Ipp32f)((sigmaColor <= 0)?1:sigmaColor*sigmaColor);
+    Ipp32f      posSquareSigma = (Ipp32f)((sigmaSpace <= 0)?1:sigmaSpace*sigmaSpace);
+
+    // Acquire data and begin processing
+    try
+    {
+        ::ipp::IwiImage      iwSrc = ippiGetImage(src);
+        ::ipp::IwiImage      iwDst = ippiGetImage(dst);
+        ::ipp::IwiBorderSize borderSize(radius);
+        ::ipp::IwiBorderType ippBorder(ippiGetBorder(iwSrc, borderType, borderSize));
+        if(!ippBorder)
+            return false;
+
+        const int threads = ippiSuggestThreadsNum(iwDst, 2);
+        if(IPP_BILATERAL_PARALLEL && threads > 1) {
+            bool  ok      = true;
+            Range range(0, (int)iwDst.m_size.height);
+            ipp_bilateralFilterParallel invoker(iwSrc, iwDst, radius, valSquareSigma, posSquareSigma, ippBorder, &ok);
+            if(!ok)
+                return false;
+
+            parallel_for_(range, invoker, threads*4);
+
+            if(!ok)
+                return false;
+        } else {
+            CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterBilateral, iwSrc, iwDst, radius, valSquareSigma, posSquareSigma, ::ipp::IwDefault(), ippBorder);
+        }
+    }
+    catch (const ::ipp::IwException &)
+    {
+        return false;
+    }
+    return true;
+#else
+    CV_UNUSED(src); CV_UNUSED(dst); CV_UNUSED(d); CV_UNUSED(sigmaColor); CV_UNUSED(sigmaSpace); CV_UNUSED(borderType);
+    return false;
+#endif
+}
+#endif
+
+void bilateralFilter( InputArray _src, OutputArray _dst, int d,
+                      double sigmaColor, double sigmaSpace,
+                      int borderType )
+{
+    CV_INSTRUMENT_REGION();
+
+    _dst.create( _src.size(), _src.type() );
+
+    CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
+               ocl_bilateralFilter_8u(_src, _dst, d, sigmaColor, sigmaSpace, borderType))
+
+    Mat src = _src.getMat(), dst = _dst.getMat();
+
+    CV_IPP_RUN_FAST(ipp_bilateralFilter(src, dst, d, sigmaColor, sigmaSpace, borderType));
+
+    if( src.depth() == CV_8U )
+        bilateralFilter_8u( src, dst, d, sigmaColor, sigmaSpace, borderType );
+    else if( src.depth() == CV_32F )
+        bilateralFilter_32f( src, dst, d, sigmaColor, sigmaSpace, borderType );
+    else
+        CV_Error( CV_StsUnsupportedFormat,
+        "Bilateral filtering is only implemented for 8u and 32f images" );
+}
+
+} // namespace

modules/imgproc/src/bilateral_filter.simd.hpp

@@ -43,18 +43,25 @@
 
 #include "precomp.hpp"
 
-#include <vector>
-
 #include "opencv2/core/hal/intrin.hpp"
-#include "opencl_kernels_imgproc.hpp"
 
 /****************************************************************************************\
                                    Bilateral Filtering
 \****************************************************************************************/
 
-namespace cv
+namespace cv {
-{
+CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
+// forward declarations
+void bilateralFilterInvoker_8u(
+        Mat& dst, const Mat& temp, int radius, int maxk,
+        int* space_ofs, float *space_weight, float *color_weight);
+void bilateralFilterInvoker_32f(
+        int cn, int radius, int maxk, int *space_ofs,
+        const Mat& temp, Mat& dst, float scale_index, float *space_weight, float *expLUT);
 
+#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
+
+namespace {
 class BilateralFilter_8u_Invoker :
     public ParallelLoopBody
 {
@@ -68,6 +75,8 @@ public:
 
     virtual void operator() (const Range& range) const CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         int i, j, cn = dest->channels(), k;
         Size size = dest->size();
 
@@ -536,161 +545,20 @@ private:
     float *space_weight, *color_weight;
 };
 
-#ifdef HAVE_OPENCL
+}  // namespace anon
-
-static bool ocl_bilateralFilter_8u(InputArray _src, OutputArray _dst, int d,
-                                   double sigma_color, double sigma_space,
-                                   int borderType)
-{
-#ifdef __ANDROID__
-    if (ocl::Device::getDefault().isNVidia())
-        return false;
-#endif
-
-    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
-    int i, j, maxk, radius;
-
-    if (depth != CV_8U || cn > 4)
-        return false;
-
-    if (sigma_color <= 0)
-        sigma_color = 1;
-    if (sigma_space <= 0)
-        sigma_space = 1;
-
-    double gauss_color_coeff = -0.5 / (sigma_color * sigma_color);
-    double gauss_space_coeff = -0.5 / (sigma_space * sigma_space);
-
-    if ( d <= 0 )
-        radius = cvRound(sigma_space * 1.5);
-    else
-        radius = d / 2;
-    radius = MAX(radius, 1);
-    d = radius * 2 + 1;
-
-    UMat src = _src.getUMat(), dst = _dst.getUMat(), temp;
-    if (src.u == dst.u)
-        return false;
 
-    copyMakeBorder(src, temp, radius, radius, radius, radius, borderType);
+void bilateralFilterInvoker_8u(
-    std::vector<float> _space_weight(d * d);
+        Mat& dst, const Mat& temp, int radius, int maxk,
-    std::vector<int> _space_ofs(d * d);
+        int* space_ofs, float *space_weight, float *color_weight)
-    float * const space_weight = &_space_weight[0];
-    int * const space_ofs = &_space_ofs[0];
-
-    // initialize space-related bilateral filter coefficients
-    for( i = -radius, maxk = 0; i <= radius; i++ )
-        for( j = -radius; j <= radius; j++ )
-        {
-            double r = std::sqrt((double)i * i + (double)j * j);
-            if ( r > radius )
-                continue;
-            space_weight[maxk] = (float)std::exp(r * r * gauss_space_coeff);
-            space_ofs[maxk++] = (int)(i * temp.step + j * cn);
-        }
-
-    char cvt[3][40];
-    String cnstr = cn > 1 ? format("%d", cn) : "";
-    String kernelName("bilateral");
-    size_t sizeDiv = 1;
-    if ((ocl::Device::getDefault().isIntel()) &&
-        (ocl::Device::getDefault().type() == ocl::Device::TYPE_GPU))
-    {
-            //Intel GPU
-            if (dst.cols % 4 == 0 && cn == 1) // For single channel x4 sized images.
-            {
-                kernelName = "bilateral_float4";
-                sizeDiv = 4;
-            }
-     }
-     ocl::Kernel k(kernelName.c_str(), ocl::imgproc::bilateral_oclsrc,
-            format("-D radius=%d -D maxk=%d -D cn=%d -D int_t=%s -D uint_t=uint%s -D convert_int_t=%s"
-            " -D uchar_t=%s -D float_t=%s -D convert_float_t=%s -D convert_uchar_t=%s -D gauss_color_coeff=(float)%f",
-            radius, maxk, cn, ocl::typeToStr(CV_32SC(cn)), cnstr.c_str(),
-            ocl::convertTypeStr(CV_8U, CV_32S, cn, cvt[0]),
-            ocl::typeToStr(type), ocl::typeToStr(CV_32FC(cn)),
-            ocl::convertTypeStr(CV_32S, CV_32F, cn, cvt[1]),
-            ocl::convertTypeStr(CV_32F, CV_8U, cn, cvt[2]), gauss_color_coeff));
-    if (k.empty())
-        return false;
-
-    Mat mspace_weight(1, d * d, CV_32FC1, space_weight);
-    Mat mspace_ofs(1, d * d, CV_32SC1, space_ofs);
-    UMat ucolor_weight, uspace_weight, uspace_ofs;
-
-    mspace_weight.copyTo(uspace_weight);
-    mspace_ofs.copyTo(uspace_ofs);
-
-    k.args(ocl::KernelArg::ReadOnlyNoSize(temp), ocl::KernelArg::WriteOnly(dst),
-           ocl::KernelArg::PtrReadOnly(uspace_weight),
-           ocl::KernelArg::PtrReadOnly(uspace_ofs));
-
-    size_t globalsize[2] = { (size_t)dst.cols / sizeDiv, (size_t)dst.rows };
-    return k.run(2, globalsize, NULL, false);
-}
-
-#endif
-static void
-bilateralFilter_8u( const Mat& src, Mat& dst, int d,
-    double sigma_color, double sigma_space,
-    int borderType )
 {
-    int cn = src.channels();
+    CV_INSTRUMENT_REGION();
-    int i, j, maxk, radius;
-    Size size = src.size();
-
-    CV_Assert( (src.type() == CV_8UC1 || src.type() == CV_8UC3) && src.data != dst.data );
-
-    if( sigma_color <= 0 )
-        sigma_color = 1;
-    if( sigma_space <= 0 )
-        sigma_space = 1;
-
-    double gauss_color_coeff = -0.5/(sigma_color*sigma_color);
-    double gauss_space_coeff = -0.5/(sigma_space*sigma_space);
-
-    if( d <= 0 )
-        radius = cvRound(sigma_space*1.5);
-    else
-        radius = d/2;
-    radius = MAX(radius, 1);
-    d = radius*2 + 1;
-
-    Mat temp;
-    copyMakeBorder( src, temp, radius, radius, radius, radius, borderType );
-
-    std::vector<float> _color_weight(cn*256);
-    std::vector<float> _space_weight(d*d);
-    std::vector<int> _space_ofs(d*d);
-    float* color_weight = &_color_weight[0];
-    float* space_weight = &_space_weight[0];
-    int* space_ofs = &_space_ofs[0];
-
-    // initialize color-related bilateral filter coefficients
-
-    for( i = 0; i < 256*cn; i++ )
-        color_weight[i] = (float)std::exp(i*i*gauss_color_coeff);
-
-    // initialize space-related bilateral filter coefficients
-    for( i = -radius, maxk = 0; i <= radius; i++ )
-    {
-        j = -radius;
-
-        for( ; j <= radius; j++ )
-        {
-            double r = std::sqrt((double)i*i + (double)j*j);
-            if( r > radius )
-                continue;
-            space_weight[maxk] = (float)std::exp(r*r*gauss_space_coeff);
-            space_ofs[maxk++] = (int)(i*temp.step + j*cn);
-        }
-    }
-
     BilateralFilter_8u_Invoker body(dst, temp, radius, maxk, space_ofs, space_weight, color_weight);
-    parallel_for_(Range(0, size.height), body, dst.total()/(double)(1<<16));
+    parallel_for_(Range(0, dst.rows), body, dst.total()/(double)(1<<16));
 }
 
 
+namespace {
+
 class BilateralFilter_32f_Invoker :
     public ParallelLoopBody
 {
@@ -705,6 +573,8 @@ public:
 
     virtual void operator() (const Range& range) const CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         int i, j, k;
         Size size = dest->size();
 
@@ -1153,216 +1023,18 @@ private:
     float scale_index, *space_weight, *expLUT;
 };
 
+} // namespace anon
 
-static void
+void bilateralFilterInvoker_32f(
-bilateralFilter_32f( const Mat& src, Mat& dst, int d,
+        int cn, int radius, int maxk, int *space_ofs,
-                     double sigma_color, double sigma_space,
+        const Mat& temp, Mat& dst, float scale_index, float *space_weight, float *expLUT)
-                     int borderType )
 {
-    int cn = src.channels();
+    CV_INSTRUMENT_REGION();
-    int i, j, maxk, radius;
-    double minValSrc=-1, maxValSrc=1;
-    const int kExpNumBinsPerChannel = 1 << 12;
-    int kExpNumBins = 0;
-    float lastExpVal = 1.f;
-    float len, scale_index;
-    Size size = src.size();
-
-    CV_Assert( (src.type() == CV_32FC1 || src.type() == CV_32FC3) && src.data != dst.data );
-
-    if( sigma_color <= 0 )
-        sigma_color = 1;
-    if( sigma_space <= 0 )
-        sigma_space = 1;
-
-    double gauss_color_coeff = -0.5/(sigma_color*sigma_color);
-    double gauss_space_coeff = -0.5/(sigma_space*sigma_space);
-
-    if( d <= 0 )
-        radius = cvRound(sigma_space*1.5);
-    else
-        radius = d/2;
-    radius = MAX(radius, 1);
-    d = radius*2 + 1;
-    // compute the min/max range for the input image (even if multichannel)
-
-    minMaxLoc( src.reshape(1), &minValSrc, &maxValSrc );
-    if(std::abs(minValSrc - maxValSrc) < FLT_EPSILON)
-    {
-        src.copyTo(dst);
-        return;
-    }
-
-    // temporary copy of the image with borders for easy processing
-    Mat temp;
-    copyMakeBorder( src, temp, radius, radius, radius, radius, borderType );
-
-    // allocate lookup tables
-    std::vector<float> _space_weight(d*d);
-    std::vector<int> _space_ofs(d*d);
-    float* space_weight = &_space_weight[0];
-    int* space_ofs = &_space_ofs[0];
-
-    // assign a length which is slightly more than needed
-    len = (float)(maxValSrc - minValSrc) * cn;
-    kExpNumBins = kExpNumBinsPerChannel * cn;
-    std::vector<float> _expLUT(kExpNumBins+2);
-    float* expLUT = &_expLUT[0];
-
-    scale_index = kExpNumBins/len;
-
-    // initialize the exp LUT
-    for( i = 0; i < kExpNumBins+2; i++ )
-    {
-        if( lastExpVal > 0.f )
-        {
-            double val =  i / scale_index;
-            expLUT[i] = (float)std::exp(val * val * gauss_color_coeff);
-            lastExpVal = expLUT[i];
-        }
-        else
-            expLUT[i] = 0.f;
-    }
-
-    // initialize space-related bilateral filter coefficients
-    for( i = -radius, maxk = 0; i <= radius; i++ )
-        for( j = -radius; j <= radius; j++ )
-        {
-            double r = std::sqrt((double)i*i + (double)j*j);
-            if( r > radius || ( i == 0 && j == 0 ) )
-                continue;
-            space_weight[maxk] = (float)std::exp(r*r*gauss_space_coeff);
-            space_ofs[maxk++] = (int)(i*(temp.step/sizeof(float)) + j*cn);
-        }
-
-    // parallel_for usage
 
     BilateralFilter_32f_Invoker body(cn, radius, maxk, space_ofs, temp, dst, scale_index, space_weight, expLUT);
-    parallel_for_(Range(0, size.height), body, dst.total()/(double)(1<<16));
+    parallel_for_(Range(0, dst.rows), body, dst.total()/(double)(1<<16));
 }
 
-#ifdef HAVE_IPP
-#define IPP_BILATERAL_PARALLEL 1
-
-#ifdef HAVE_IPP_IW
-class ipp_bilateralFilterParallel: public ParallelLoopBody
-{
-public:
-    ipp_bilateralFilterParallel(::ipp::IwiImage &_src, ::ipp::IwiImage &_dst, int _radius, Ipp32f _valSquareSigma, Ipp32f _posSquareSigma, ::ipp::IwiBorderType _borderType, bool *_ok):
-        src(_src), dst(_dst)
-    {
-        pOk = _ok;
-
-        radius          = _radius;
-        valSquareSigma  = _valSquareSigma;
-        posSquareSigma  = _posSquareSigma;
-        borderType      = _borderType;
-
-        *pOk = true;
-    }
-    ~ipp_bilateralFilterParallel() {}
-
-    virtual void operator() (const Range& range) const CV_OVERRIDE
-    {
-        if(*pOk == false)
-            return;
-
-        try
-        {
-            ::ipp::IwiTile tile = ::ipp::IwiRoi(0, range.start, dst.m_size.width, range.end - range.start);
-            CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterBilateral, src, dst, radius, valSquareSigma, posSquareSigma, ::ipp::IwDefault(), borderType, tile);
-        }
-        catch(const ::ipp::IwException &)
-        {
-            *pOk = false;
-            return;
-        }
-    }
-private:
-    ::ipp::IwiImage &src;
-    ::ipp::IwiImage &dst;
-
-    int                  radius;
-    Ipp32f               valSquareSigma;
-    Ipp32f               posSquareSigma;
-    ::ipp::IwiBorderType borderType;
-
-    bool  *pOk;
-    const ipp_bilateralFilterParallel& operator= (const ipp_bilateralFilterParallel&);
-};
 #endif
-
+CV_CPU_OPTIMIZATION_NAMESPACE_END
-static bool ipp_bilateralFilter(Mat &src, Mat &dst, int d, double sigmaColor, double sigmaSpace, int borderType)
+} // namespace
-{
-#ifdef HAVE_IPP_IW
-    CV_INSTRUMENT_REGION_IPP();
-
-    int         radius         = IPP_MAX(((d <= 0)?cvRound(sigmaSpace*1.5):d/2), 1);
-    Ipp32f      valSquareSigma = (Ipp32f)((sigmaColor <= 0)?1:sigmaColor*sigmaColor);
-    Ipp32f      posSquareSigma = (Ipp32f)((sigmaSpace <= 0)?1:sigmaSpace*sigmaSpace);
-
-    // Acquire data and begin processing
-    try
-    {
-        ::ipp::IwiImage      iwSrc = ippiGetImage(src);
-        ::ipp::IwiImage      iwDst = ippiGetImage(dst);
-        ::ipp::IwiBorderSize borderSize(radius);
-        ::ipp::IwiBorderType ippBorder(ippiGetBorder(iwSrc, borderType, borderSize));
-        if(!ippBorder)
-            return false;
-
-        const int threads = ippiSuggestThreadsNum(iwDst, 2);
-        if(IPP_BILATERAL_PARALLEL && threads > 1) {
-            bool  ok      = true;
-            Range range(0, (int)iwDst.m_size.height);
-            ipp_bilateralFilterParallel invoker(iwSrc, iwDst, radius, valSquareSigma, posSquareSigma, ippBorder, &ok);
-            if(!ok)
-                return false;
-
-            parallel_for_(range, invoker, threads*4);
-
-            if(!ok)
-                return false;
-        } else {
-            CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterBilateral, iwSrc, iwDst, radius, valSquareSigma, posSquareSigma, ::ipp::IwDefault(), ippBorder);
-        }
-    }
-    catch (const ::ipp::IwException &)
-    {
-        return false;
-    }
-    return true;
-#else
-    CV_UNUSED(src); CV_UNUSED(dst); CV_UNUSED(d); CV_UNUSED(sigmaColor); CV_UNUSED(sigmaSpace); CV_UNUSED(borderType);
-    return false;
-#endif
-}
-#endif
-
-}
-
-void cv::bilateralFilter( InputArray _src, OutputArray _dst, int d,
-                      double sigmaColor, double sigmaSpace,
-                      int borderType )
-{
-    CV_INSTRUMENT_REGION();
-
-    _dst.create( _src.size(), _src.type() );
-
-    CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
-               ocl_bilateralFilter_8u(_src, _dst, d, sigmaColor, sigmaSpace, borderType))
-
-    Mat src = _src.getMat(), dst = _dst.getMat();
-
-    CV_IPP_RUN_FAST(ipp_bilateralFilter(src, dst, d, sigmaColor, sigmaSpace, borderType));
-
-    if( src.depth() == CV_8U )
-        bilateralFilter_8u( src, dst, d, sigmaColor, sigmaSpace, borderType );
-    else if( src.depth() == CV_32F )
-        bilateralFilter_32f( src, dst, d, sigmaColor, sigmaSpace, borderType );
-    else
-        CV_Error( CV_StsUnsupportedFormat,
-        "Bilateral filtering is only implemented for 8u and 32f images" );
-}
-
-/* End of file. */

modules/imgproc/src/box_filter.dispatch.cpp

@@ -0,0 +1,557 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, 2018, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Copyright (C) 2014-2015, Itseez Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#include "precomp.hpp"
+
+#include <vector>
+
+#include "opencv2/core/hal/intrin.hpp"
+#include "opencl_kernels_imgproc.hpp"
+
+#include "opencv2/core/openvx/ovx_defs.hpp"
+
+#include "box_filter.simd.hpp"
+#include "box_filter.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
+
+
+namespace cv {
+
+#ifdef HAVE_OPENCL
+
+static bool ocl_boxFilter3x3_8UC1( InputArray _src, OutputArray _dst, int ddepth,
+                                   Size ksize, Point anchor, int borderType, bool normalize )
+{
+    const ocl::Device & dev = ocl::Device::getDefault();
+    int type = _src.type(), sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+
+    if (ddepth < 0)
+        ddepth = sdepth;
+
+    if (anchor.x < 0)
+        anchor.x = ksize.width / 2;
+    if (anchor.y < 0)
+        anchor.y = ksize.height / 2;
+
+    if ( !(dev.isIntel() && (type == CV_8UC1) &&
+         (_src.offset() == 0) && (_src.step() % 4 == 0) &&
+         (_src.cols() % 16 == 0) && (_src.rows() % 2 == 0) &&
+         (anchor.x == 1) && (anchor.y == 1) &&
+         (ksize.width == 3) && (ksize.height == 3)) )
+        return false;
+
+    float alpha = 1.0f / (ksize.height * ksize.width);
+    Size size = _src.size();
+    size_t globalsize[2] = { 0, 0 };
+    size_t localsize[2] = { 0, 0 };
+    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", 0, "BORDER_REFLECT_101" };
+
+    globalsize[0] = size.width / 16;
+    globalsize[1] = size.height / 2;
+
+    char build_opts[1024];
+    sprintf(build_opts, "-D %s %s", borderMap[borderType], normalize ? "-D NORMALIZE" : "");
+
+    ocl::Kernel kernel("boxFilter3x3_8UC1_cols16_rows2", cv::ocl::imgproc::boxFilter3x3_oclsrc, build_opts);
+    if (kernel.empty())
+        return false;
+
+    UMat src = _src.getUMat();
+    _dst.create(size, CV_MAKETYPE(ddepth, cn));
+    if (!(_dst.offset() == 0 && _dst.step() % 4 == 0))
+        return false;
+    UMat dst = _dst.getUMat();
+
+    int idxArg = kernel.set(0, ocl::KernelArg::PtrReadOnly(src));
+    idxArg = kernel.set(idxArg, (int)src.step);
+    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dst));
+    idxArg = kernel.set(idxArg, (int)dst.step);
+    idxArg = kernel.set(idxArg, (int)dst.rows);
+    idxArg = kernel.set(idxArg, (int)dst.cols);
+    if (normalize)
+        idxArg = kernel.set(idxArg, (float)alpha);
+
+    return kernel.run(2, globalsize, (localsize[0] == 0) ? NULL : localsize, false);
+}
+
+static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
+                           Size ksize, Point anchor, int borderType, bool normalize, bool sqr = false )
+{
+    const ocl::Device & dev = ocl::Device::getDefault();
+    int type = _src.type(), sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), esz = CV_ELEM_SIZE(type);
+    bool doubleSupport = dev.doubleFPConfig() > 0;
+
+    if (ddepth < 0)
+        ddepth = sdepth;
+
+    if (cn > 4 || (!doubleSupport && (sdepth == CV_64F || ddepth == CV_64F)) ||
+        _src.offset() % esz != 0 || _src.step() % esz != 0)
+        return false;
+
+    if (anchor.x < 0)
+        anchor.x = ksize.width / 2;
+    if (anchor.y < 0)
+        anchor.y = ksize.height / 2;
+
+    int computeUnits = ocl::Device::getDefault().maxComputeUnits();
+    float alpha = 1.0f / (ksize.height * ksize.width);
+    Size size = _src.size(), wholeSize;
+    bool isolated = (borderType & BORDER_ISOLATED) != 0;
+    borderType &= ~BORDER_ISOLATED;
+    int wdepth = std::max(CV_32F, std::max(ddepth, sdepth)),
+        wtype = CV_MAKE_TYPE(wdepth, cn), dtype = CV_MAKE_TYPE(ddepth, cn);
+
+    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", 0, "BORDER_REFLECT_101" };
+    size_t globalsize[2] = { (size_t)size.width, (size_t)size.height };
+    size_t localsize_general[2] = { 0, 1 }, * localsize = NULL;
+
+    UMat src = _src.getUMat();
+    if (!isolated)
+    {
+        Point ofs;
+        src.locateROI(wholeSize, ofs);
+    }
+
+    int h = isolated ? size.height : wholeSize.height;
+    int w = isolated ? size.width : wholeSize.width;
+
+    size_t maxWorkItemSizes[32];
+    ocl::Device::getDefault().maxWorkItemSizes(maxWorkItemSizes);
+    int tryWorkItems = (int)maxWorkItemSizes[0];
+
+    ocl::Kernel kernel;
+
+    if (dev.isIntel() && !(dev.type() & ocl::Device::TYPE_CPU) &&
+        ((ksize.width < 5 && ksize.height < 5 && esz <= 4) ||
+         (ksize.width == 5 && ksize.height == 5 && cn == 1)))
+    {
+        if (w < ksize.width || h < ksize.height)
+            return false;
+
+        // Figure out what vector size to use for loading the pixels.
+        int pxLoadNumPixels = cn != 1 || size.width % 4 ? 1 : 4;
+        int pxLoadVecSize = cn * pxLoadNumPixels;
+
+        // Figure out how many pixels per work item to compute in X and Y
+        // directions.  Too many and we run out of registers.
+        int pxPerWorkItemX = 1, pxPerWorkItemY = 1;
+        if (cn <= 2 && ksize.width <= 4 && ksize.height <= 4)
+        {
+            pxPerWorkItemX = size.width % 8 ? size.width % 4 ? size.width % 2 ? 1 : 2 : 4 : 8;
+            pxPerWorkItemY = size.height % 2 ? 1 : 2;
+        }
+        else if (cn < 4 || (ksize.width <= 4 && ksize.height <= 4))
+        {
+            pxPerWorkItemX = size.width % 2 ? 1 : 2;
+            pxPerWorkItemY = size.height % 2 ? 1 : 2;
+        }
+        globalsize[0] = size.width / pxPerWorkItemX;
+        globalsize[1] = size.height / pxPerWorkItemY;
+
+        // Need some padding in the private array for pixels
+        int privDataWidth = roundUp(pxPerWorkItemX + ksize.width - 1, pxLoadNumPixels);
+
+        // Make the global size a nice round number so the runtime can pick
+        // from reasonable choices for the workgroup size
+        const int wgRound = 256;
+        globalsize[0] = roundUp(globalsize[0], wgRound);
+
+        char build_options[1024], cvt[2][40];
+        sprintf(build_options, "-D cn=%d "
+                "-D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d "
+                "-D PX_LOAD_VEC_SIZE=%d -D PX_LOAD_NUM_PX=%d "
+                "-D PX_PER_WI_X=%d -D PX_PER_WI_Y=%d -D PRIV_DATA_WIDTH=%d -D %s -D %s "
+                "-D PX_LOAD_X_ITERATIONS=%d -D PX_LOAD_Y_ITERATIONS=%d "
+                "-D srcT=%s -D srcT1=%s -D dstT=%s -D dstT1=%s -D WT=%s -D WT1=%s "
+                "-D convertToWT=%s -D convertToDstT=%s%s%s -D PX_LOAD_FLOAT_VEC_CONV=convert_%s -D OP_BOX_FILTER",
+                cn, anchor.x, anchor.y, ksize.width, ksize.height,
+                pxLoadVecSize, pxLoadNumPixels,
+                pxPerWorkItemX, pxPerWorkItemY, privDataWidth, borderMap[borderType],
+                isolated ? "BORDER_ISOLATED" : "NO_BORDER_ISOLATED",
+                privDataWidth / pxLoadNumPixels, pxPerWorkItemY + ksize.height - 1,
+                ocl::typeToStr(type), ocl::typeToStr(sdepth), ocl::typeToStr(dtype),
+                ocl::typeToStr(ddepth), ocl::typeToStr(wtype), ocl::typeToStr(wdepth),
+                ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]),
+                ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]),
+                normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "",
+                ocl::typeToStr(CV_MAKE_TYPE(wdepth, pxLoadVecSize)) //PX_LOAD_FLOAT_VEC_CONV
+                );
+
+
+        if (!kernel.create("filterSmall", cv::ocl::imgproc::filterSmall_oclsrc, build_options))
+            return false;
+    }
+    else
+    {
+        localsize = localsize_general;
+        for ( ; ; )
+        {
+            int BLOCK_SIZE_X = tryWorkItems, BLOCK_SIZE_Y = std::min(ksize.height * 10, size.height);
+
+            while (BLOCK_SIZE_X > 32 && BLOCK_SIZE_X >= ksize.width * 2 && BLOCK_SIZE_X > size.width * 2)
+                BLOCK_SIZE_X /= 2;
+            while (BLOCK_SIZE_Y < BLOCK_SIZE_X / 8 && BLOCK_SIZE_Y * computeUnits * 32 < size.height)
+                BLOCK_SIZE_Y *= 2;
+
+            if (ksize.width > BLOCK_SIZE_X || w < ksize.width || h < ksize.height)
+                return false;
+
+            char cvt[2][50];
+            String opts = format("-D LOCAL_SIZE_X=%d -D BLOCK_SIZE_Y=%d -D ST=%s -D DT=%s -D WT=%s -D convertToDT=%s -D convertToWT=%s"
+                                 " -D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d -D %s%s%s%s%s"
+                                 " -D ST1=%s -D DT1=%s -D cn=%d",
+                                 BLOCK_SIZE_X, BLOCK_SIZE_Y, ocl::typeToStr(type), ocl::typeToStr(CV_MAKE_TYPE(ddepth, cn)),
+                                 ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
+                                 ocl::convertTypeStr(wdepth, ddepth, cn, cvt[0]),
+                                 ocl::convertTypeStr(sdepth, wdepth, cn, cvt[1]),
+                                 anchor.x, anchor.y, ksize.width, ksize.height, borderMap[borderType],
+                                 isolated ? " -D BORDER_ISOLATED" : "", doubleSupport ? " -D DOUBLE_SUPPORT" : "",
+                                 normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "",
+                                 ocl::typeToStr(sdepth), ocl::typeToStr(ddepth), cn);
+
+            localsize[0] = BLOCK_SIZE_X;
+            globalsize[0] = divUp(size.width, BLOCK_SIZE_X - (ksize.width - 1)) * BLOCK_SIZE_X;
+            globalsize[1] = divUp(size.height, BLOCK_SIZE_Y);
+
+            kernel.create("boxFilter", cv::ocl::imgproc::boxFilter_oclsrc, opts);
+            if (kernel.empty())
+                return false;
+
+            size_t kernelWorkGroupSize = kernel.workGroupSize();
+            if (localsize[0] <= kernelWorkGroupSize)
+                break;
+            if (BLOCK_SIZE_X < (int)kernelWorkGroupSize)
+                return false;
+
+            tryWorkItems = (int)kernelWorkGroupSize;
+        }
+    }
+
+    _dst.create(size, CV_MAKETYPE(ddepth, cn));
+    UMat dst = _dst.getUMat();
+
+    int idxArg = kernel.set(0, ocl::KernelArg::PtrReadOnly(src));
+    idxArg = kernel.set(idxArg, (int)src.step);
+    int srcOffsetX = (int)((src.offset % src.step) / src.elemSize());
+    int srcOffsetY = (int)(src.offset / src.step);
+    int srcEndX = isolated ? srcOffsetX + size.width : wholeSize.width;
+    int srcEndY = isolated ? srcOffsetY + size.height : wholeSize.height;
+    idxArg = kernel.set(idxArg, srcOffsetX);
+    idxArg = kernel.set(idxArg, srcOffsetY);
+    idxArg = kernel.set(idxArg, srcEndX);
+    idxArg = kernel.set(idxArg, srcEndY);
+    idxArg = kernel.set(idxArg, ocl::KernelArg::WriteOnly(dst));
+    if (normalize)
+        idxArg = kernel.set(idxArg, (float)alpha);
+
+    return kernel.run(2, globalsize, localsize, false);
+}
+
+#endif
+
+Ptr<BaseRowFilter> getRowSumFilter(int srcType, int sumType, int ksize, int anchor)
+{
+    CV_INSTRUMENT_REGION();
+
+    CV_CPU_DISPATCH(getRowSumFilter, (srcType, sumType, ksize, anchor),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+
+Ptr<BaseColumnFilter> getColumnSumFilter(int sumType, int dstType, int ksize, int anchor, double scale)
+{
+    CV_INSTRUMENT_REGION();
+
+    CV_CPU_DISPATCH(getColumnSumFilter, (sumType, dstType, ksize, anchor, scale),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+
+Ptr<FilterEngine> createBoxFilter(int srcType, int dstType, Size ksize,
+                                  Point anchor, bool normalize, int borderType)
+{
+    CV_INSTRUMENT_REGION();
+
+    CV_CPU_DISPATCH(createBoxFilter, (srcType, dstType, ksize, anchor, normalize, borderType),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+#ifdef HAVE_OPENVX
+    namespace ovx {
+        template <> inline bool skipSmallImages<VX_KERNEL_BOX_3x3>(int w, int h) { return w*h < 640 * 480; }
+    }
+    static bool openvx_boxfilter(InputArray _src, OutputArray _dst, int ddepth,
+                                 Size ksize, Point anchor,
+                                 bool normalize, int borderType)
+    {
+        if (ddepth < 0)
+            ddepth = CV_8UC1;
+        if (_src.type() != CV_8UC1 || ddepth != CV_8U || !normalize ||
+            _src.cols() < 3 || _src.rows() < 3 ||
+            ksize.width != 3 || ksize.height != 3 ||
+            (anchor.x >= 0 && anchor.x != 1) ||
+            (anchor.y >= 0 && anchor.y != 1) ||
+            ovx::skipSmallImages<VX_KERNEL_BOX_3x3>(_src.cols(), _src.rows()))
+            return false;
+
+        Mat src = _src.getMat();
+
+        if ((borderType & BORDER_ISOLATED) == 0 && src.isSubmatrix())
+            return false; //Process isolated borders only
+        vx_enum border;
+        switch (borderType & ~BORDER_ISOLATED)
+        {
+        case BORDER_CONSTANT:
+            border = VX_BORDER_CONSTANT;
+            break;
+        case BORDER_REPLICATE:
+            border = VX_BORDER_REPLICATE;
+            break;
+        default:
+            return false;
+        }
+
+        _dst.create(src.size(), CV_8UC1);
+        Mat dst = _dst.getMat();
+
+        try
+        {
+            ivx::Context ctx = ovx::getOpenVXContext();
+
+            Mat a;
+            if (dst.data != src.data)
+                a = src;
+            else
+                src.copyTo(a);
+
+            ivx::Image
+                ia = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
+                                                  ivx::Image::createAddressing(a.cols, a.rows, 1, (vx_int32)(a.step)), a.data),
+                ib = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
+                                                  ivx::Image::createAddressing(dst.cols, dst.rows, 1, (vx_int32)(dst.step)), dst.data);
+
+            //ATTENTION: VX_CONTEXT_IMMEDIATE_BORDER attribute change could lead to strange issues in multi-threaded environments
+            //since OpenVX standard says nothing about thread-safety for now
+            ivx::border_t prevBorder = ctx.immediateBorder();
+            ctx.setImmediateBorder(border, (vx_uint8)(0));
+            ivx::IVX_CHECK_STATUS(vxuBox3x3(ctx, ia, ib));
+            ctx.setImmediateBorder(prevBorder);
+        }
+        catch (const ivx::RuntimeError & e)
+        {
+            VX_DbgThrow(e.what());
+        }
+        catch (const ivx::WrapperError & e)
+        {
+            VX_DbgThrow(e.what());
+        }
+
+        return true;
+    }
+#endif
+
+#if defined(HAVE_IPP)
+static bool ipp_boxfilter(Mat &src, Mat &dst, Size ksize, Point anchor, bool normalize, int borderType)
+{
+#ifdef HAVE_IPP_IW
+    CV_INSTRUMENT_REGION_IPP();
+
+#if IPP_VERSION_X100 < 201801
+    // Problem with SSE42 optimization for 16s and some 8u modes
+    if(ipp::getIppTopFeatures() == ippCPUID_SSE42 && (((src.depth() == CV_16S || src.depth() == CV_16U) && (src.channels() == 3 || src.channels() == 4)) || (src.depth() == CV_8U && src.channels() == 3 && (ksize.width > 5 || ksize.height > 5))))
+        return false;
+
+    // Other optimizations has some degradations too
+    if((((src.depth() == CV_16S || src.depth() == CV_16U) && (src.channels() == 4)) || (src.depth() == CV_8U && src.channels() == 1 && (ksize.width > 5 || ksize.height > 5))))
+        return false;
+#endif
+
+    if(!normalize)
+        return false;
+
+    if(!ippiCheckAnchor(anchor, ksize))
+        return false;
+
+    try
+    {
+        ::ipp::IwiImage       iwSrc      = ippiGetImage(src);
+        ::ipp::IwiImage       iwDst      = ippiGetImage(dst);
+        ::ipp::IwiSize        iwKSize    = ippiGetSize(ksize);
+        ::ipp::IwiBorderSize  borderSize(iwKSize);
+        ::ipp::IwiBorderType  ippBorder(ippiGetBorder(iwSrc, borderType, borderSize));
+        if(!ippBorder)
+            return false;
+
+        CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterBox, iwSrc, iwDst, iwKSize, ::ipp::IwDefault(), ippBorder);
+    }
+    catch (const ::ipp::IwException &)
+    {
+        return false;
+    }
+
+    return true;
+#else
+    CV_UNUSED(src); CV_UNUSED(dst); CV_UNUSED(ksize); CV_UNUSED(anchor); CV_UNUSED(normalize); CV_UNUSED(borderType);
+    return false;
+#endif
+}
+#endif
+
+
+void boxFilter(InputArray _src, OutputArray _dst, int ddepth,
+               Size ksize, Point anchor,
+               bool normalize, int borderType)
+{
+    CV_INSTRUMENT_REGION();
+
+    CV_OCL_RUN(_dst.isUMat() &&
+               (borderType == BORDER_REPLICATE || borderType == BORDER_CONSTANT ||
+                borderType == BORDER_REFLECT || borderType == BORDER_REFLECT_101),
+               ocl_boxFilter3x3_8UC1(_src, _dst, ddepth, ksize, anchor, borderType, normalize))
+
+    CV_OCL_RUN(_dst.isUMat(), ocl_boxFilter(_src, _dst, ddepth, ksize, anchor, borderType, normalize))
+
+    Mat src = _src.getMat();
+    int stype = src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype);
+    if( ddepth < 0 )
+        ddepth = sdepth;
+    _dst.create( src.size(), CV_MAKETYPE(ddepth, cn) );
+    Mat dst = _dst.getMat();
+    if( borderType != BORDER_CONSTANT && normalize && (borderType & BORDER_ISOLATED) != 0 )
+    {
+        if( src.rows == 1 )
+            ksize.height = 1;
+        if( src.cols == 1 )
+            ksize.width = 1;
+    }
+
+    Point ofs;
+    Size wsz(src.cols, src.rows);
+    if(!(borderType&BORDER_ISOLATED))
+        src.locateROI( wsz, ofs );
+
+    CALL_HAL(boxFilter, cv_hal_boxFilter, src.ptr(), src.step, dst.ptr(), dst.step, src.cols, src.rows, sdepth, ddepth, cn,
+             ofs.x, ofs.y, wsz.width - src.cols - ofs.x, wsz.height - src.rows - ofs.y, ksize.width, ksize.height,
+             anchor.x, anchor.y, normalize, borderType&~BORDER_ISOLATED);
+
+    CV_OVX_RUN(true,
+               openvx_boxfilter(src, dst, ddepth, ksize, anchor, normalize, borderType))
+
+    CV_IPP_RUN_FAST(ipp_boxfilter(src, dst, ksize, anchor, normalize, borderType));
+
+    borderType = (borderType&~BORDER_ISOLATED);
+
+    Ptr<FilterEngine> f = createBoxFilter( src.type(), dst.type(),
+                        ksize, anchor, normalize, borderType );
+
+    f->apply( src, dst, wsz, ofs );
+}
+
+
+void blur(InputArray src, OutputArray dst,
+          Size ksize, Point anchor, int borderType)
+{
+    CV_INSTRUMENT_REGION();
+
+    boxFilter( src, dst, -1, ksize, anchor, true, borderType );
+}
+
+
+/****************************************************************************************\
+                                    Squared Box Filter
+\****************************************************************************************/
+
+static Ptr<BaseRowFilter> getSqrRowSumFilter(int srcType, int sumType, int ksize, int anchor)
+{
+    CV_INSTRUMENT_REGION();
+
+    CV_CPU_DISPATCH(getSqrRowSumFilter, (srcType, sumType, ksize, anchor),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+void sqrBoxFilter(InputArray _src, OutputArray _dst, int ddepth,
+                  Size ksize, Point anchor,
+                  bool normalize, int borderType)
+{
+    CV_INSTRUMENT_REGION();
+
+    int srcType = _src.type(), sdepth = CV_MAT_DEPTH(srcType), cn = CV_MAT_CN(srcType);
+    Size size = _src.size();
+
+    if( ddepth < 0 )
+        ddepth = sdepth < CV_32F ? CV_32F : CV_64F;
+
+    if( borderType != BORDER_CONSTANT && normalize )
+    {
+        if( size.height == 1 )
+            ksize.height = 1;
+        if( size.width == 1 )
+            ksize.width = 1;
+    }
+
+    CV_OCL_RUN(_dst.isUMat() && _src.dims() <= 2,
+               ocl_boxFilter(_src, _dst, ddepth, ksize, anchor, borderType, normalize, true))
+
+    int sumDepth = CV_64F;
+    if( sdepth == CV_8U )
+        sumDepth = CV_32S;
+    int sumType = CV_MAKETYPE( sumDepth, cn ), dstType = CV_MAKETYPE(ddepth, cn);
+
+    Mat src = _src.getMat();
+    _dst.create( size, dstType );
+    Mat dst = _dst.getMat();
+
+    Ptr<BaseRowFilter> rowFilter = getSqrRowSumFilter(srcType, sumType, ksize.width, anchor.x );
+    Ptr<BaseColumnFilter> columnFilter = getColumnSumFilter(sumType,
+                                                            dstType, ksize.height, anchor.y,
+                                                            normalize ? 1./(ksize.width*ksize.height) : 1);
+
+    Ptr<FilterEngine> f = makePtr<FilterEngine>(Ptr<BaseFilter>(), rowFilter, columnFilter,
+                                                srcType, dstType, sumType, borderType );
+    Point ofs;
+    Size wsz(src.cols, src.rows);
+    src.locateROI( wsz, ofs );
+
+    f->apply( src, dst, wsz, ofs );
+}
+
+} // namespace

modules/imgproc/src/box_filter.simd.hpp

@@ -42,21 +42,25 @@
 //M*/
 
 #include "precomp.hpp"
-
-#include <vector>
-
 #include "opencv2/core/hal/intrin.hpp"
-#include "opencl_kernels_imgproc.hpp"
 
-#include "opencv2/core/openvx/ovx_defs.hpp"
+namespace cv {
+CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
+// forward declarations
+Ptr<BaseRowFilter> getRowSumFilter(int srcType, int sumType, int ksize, int anchor);
+Ptr<BaseColumnFilter> getColumnSumFilter(int sumType, int dstType, int ksize, int anchor, double scale);
+Ptr<FilterEngine> createBoxFilter(int srcType, int dstType, Size ksize,
+                                  Point anchor, bool normalize, int borderType);
 
-namespace cv
+Ptr<BaseRowFilter> getSqrRowSumFilter(int srcType, int sumType, int ksize, int anchor);
-{
 
+
+#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
 /****************************************************************************************\
                                          Box Filter
 \****************************************************************************************/
 
+namespace {
 template<typename T, typename ST>
 struct RowSum :
         public BaseRowFilter
@@ -70,6 +74,8 @@ struct RowSum :
 
     virtual void operator()(const uchar* src, uchar* dst, int width, int cn) CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         const T* S = (const T*)src;
         ST* D = (ST*)dst;
         int i = 0, k, ksz_cn = ksize*cn;
@@ -183,6 +189,8 @@ struct ColumnSum :
 
     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         int i;
         ST* SUM;
         bool haveScale = scale != 1;
@@ -281,6 +289,8 @@ struct ColumnSum<int, uchar> :
 
     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         int* SUM;
         bool haveScale = scale != 1;
         double _scale = scale;
@@ -408,9 +418,6 @@ struct ColumnSum<int, uchar> :
             }
             dst += dststep;
         }
-#if CV_SIMD
-        vx_cleanup();
-#endif
     }
 
     double scale;
@@ -452,6 +459,8 @@ public BaseColumnFilter
 
     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         const int ds = divScale;
         const int dd = divDelta;
         ushort* SUM;
@@ -586,9 +595,6 @@ public BaseColumnFilter
             }
             dst += dststep;
         }
-#if CV_SIMD
-        vx_cleanup();
-#endif
     }
 
     double scale;
@@ -616,6 +622,8 @@ struct ColumnSum<int, short> :
 
     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         int i;
         int* SUM;
         bool haveScale = scale != 1;
@@ -739,9 +747,6 @@ struct ColumnSum<int, short> :
             }
             dst += dststep;
         }
-#if CV_SIMD
-        vx_cleanup();
-#endif
     }
 
     double scale;
@@ -767,6 +772,8 @@ struct ColumnSum<int, ushort> :
 
     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         int* SUM;
         bool haveScale = scale != 1;
         double _scale = scale;
@@ -888,9 +895,6 @@ struct ColumnSum<int, ushort> :
             }
             dst += dststep;
         }
-#if CV_SIMD
-        vx_cleanup();
-#endif
     }
 
     double scale;
@@ -915,6 +919,8 @@ struct ColumnSum<int, int> :
 
     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         int* SUM;
         bool haveScale = scale != 1;
         double _scale = scale;
@@ -1022,9 +1028,6 @@ struct ColumnSum<int, int> :
             }
             dst += dststep;
         }
-#if CV_SIMD
-        vx_cleanup();
-#endif
     }
 
     double scale;
@@ -1050,6 +1053,8 @@ struct ColumnSum<int, float> :
 
     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         int* SUM;
         bool haveScale = scale != 1;
         double _scale = scale;
@@ -1154,9 +1159,6 @@ struct ColumnSum<int, float> :
             }
             dst += dststep;
         }
-#if CV_SIMD
-        vx_cleanup();
-#endif
     }
 
     double scale;
@@ -1164,243 +1166,13 @@ struct ColumnSum<int, float> :
     std::vector<int> sum;
 };
 
-#ifdef HAVE_OPENCL
+}  // namespace anon
 
-static bool ocl_boxFilter3x3_8UC1( InputArray _src, OutputArray _dst, int ddepth,
-                                   Size ksize, Point anchor, int borderType, bool normalize )
-{
-    const ocl::Device & dev = ocl::Device::getDefault();
-    int type = _src.type(), sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
-
-    if (ddepth < 0)
-        ddepth = sdepth;
-
-    if (anchor.x < 0)
-        anchor.x = ksize.width / 2;
-    if (anchor.y < 0)
-        anchor.y = ksize.height / 2;
-
-    if ( !(dev.isIntel() && (type == CV_8UC1) &&
-         (_src.offset() == 0) && (_src.step() % 4 == 0) &&
-         (_src.cols() % 16 == 0) && (_src.rows() % 2 == 0) &&
-         (anchor.x == 1) && (anchor.y == 1) &&
-         (ksize.width == 3) && (ksize.height == 3)) )
-        return false;
-
-    float alpha = 1.0f / (ksize.height * ksize.width);
-    Size size = _src.size();
-    size_t globalsize[2] = { 0, 0 };
-    size_t localsize[2] = { 0, 0 };
-    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", 0, "BORDER_REFLECT_101" };
-
-    globalsize[0] = size.width / 16;
-    globalsize[1] = size.height / 2;
-
-    char build_opts[1024];
-    sprintf(build_opts, "-D %s %s", borderMap[borderType], normalize ? "-D NORMALIZE" : "");
-
-    ocl::Kernel kernel("boxFilter3x3_8UC1_cols16_rows2", cv::ocl::imgproc::boxFilter3x3_oclsrc, build_opts);
-    if (kernel.empty())
-        return false;
-
-    UMat src = _src.getUMat();
-    _dst.create(size, CV_MAKETYPE(ddepth, cn));
-    if (!(_dst.offset() == 0 && _dst.step() % 4 == 0))
-        return false;
-    UMat dst = _dst.getUMat();
-
-    int idxArg = kernel.set(0, ocl::KernelArg::PtrReadOnly(src));
-    idxArg = kernel.set(idxArg, (int)src.step);
-    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dst));
-    idxArg = kernel.set(idxArg, (int)dst.step);
-    idxArg = kernel.set(idxArg, (int)dst.rows);
-    idxArg = kernel.set(idxArg, (int)dst.cols);
-    if (normalize)
-        idxArg = kernel.set(idxArg, (float)alpha);
-
-    return kernel.run(2, globalsize, (localsize[0] == 0) ? NULL : localsize, false);
-}
 
-static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
+Ptr<BaseRowFilter> getRowSumFilter(int srcType, int sumType, int ksize, int anchor)
-                           Size ksize, Point anchor, int borderType, bool normalize, bool sqr = false )
 {
-    const ocl::Device & dev = ocl::Device::getDefault();
+    CV_INSTRUMENT_REGION();
-    int type = _src.type(), sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), esz = CV_ELEM_SIZE(type);
-    bool doubleSupport = dev.doubleFPConfig() > 0;
-
-    if (ddepth < 0)
-        ddepth = sdepth;
-
-    if (cn > 4 || (!doubleSupport && (sdepth == CV_64F || ddepth == CV_64F)) ||
-        _src.offset() % esz != 0 || _src.step() % esz != 0)
-        return false;
-
-    if (anchor.x < 0)
-        anchor.x = ksize.width / 2;
-    if (anchor.y < 0)
-        anchor.y = ksize.height / 2;
-
-    int computeUnits = ocl::Device::getDefault().maxComputeUnits();
-    float alpha = 1.0f / (ksize.height * ksize.width);
-    Size size = _src.size(), wholeSize;
-    bool isolated = (borderType & BORDER_ISOLATED) != 0;
-    borderType &= ~BORDER_ISOLATED;
-    int wdepth = std::max(CV_32F, std::max(ddepth, sdepth)),
-        wtype = CV_MAKE_TYPE(wdepth, cn), dtype = CV_MAKE_TYPE(ddepth, cn);
-
-    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", 0, "BORDER_REFLECT_101" };
-    size_t globalsize[2] = { (size_t)size.width, (size_t)size.height };
-    size_t localsize_general[2] = { 0, 1 }, * localsize = NULL;
-
-    UMat src = _src.getUMat();
-    if (!isolated)
-    {
-        Point ofs;
-        src.locateROI(wholeSize, ofs);
-    }
-
-    int h = isolated ? size.height : wholeSize.height;
-    int w = isolated ? size.width : wholeSize.width;
-
-    size_t maxWorkItemSizes[32];
-    ocl::Device::getDefault().maxWorkItemSizes(maxWorkItemSizes);
-    int tryWorkItems = (int)maxWorkItemSizes[0];
-
-    ocl::Kernel kernel;
-
-    if (dev.isIntel() && !(dev.type() & ocl::Device::TYPE_CPU) &&
-        ((ksize.width < 5 && ksize.height < 5 && esz <= 4) ||
-         (ksize.width == 5 && ksize.height == 5 && cn == 1)))
-    {
-        if (w < ksize.width || h < ksize.height)
-            return false;
-
-        // Figure out what vector size to use for loading the pixels.
-        int pxLoadNumPixels = cn != 1 || size.width % 4 ? 1 : 4;
-        int pxLoadVecSize = cn * pxLoadNumPixels;
-
-        // Figure out how many pixels per work item to compute in X and Y
-        // directions.  Too many and we run out of registers.
-        int pxPerWorkItemX = 1, pxPerWorkItemY = 1;
-        if (cn <= 2 && ksize.width <= 4 && ksize.height <= 4)
-        {
-            pxPerWorkItemX = size.width % 8 ? size.width % 4 ? size.width % 2 ? 1 : 2 : 4 : 8;
-            pxPerWorkItemY = size.height % 2 ? 1 : 2;
-        }
-        else if (cn < 4 || (ksize.width <= 4 && ksize.height <= 4))
-        {
-            pxPerWorkItemX = size.width % 2 ? 1 : 2;
-            pxPerWorkItemY = size.height % 2 ? 1 : 2;
-        }
-        globalsize[0] = size.width / pxPerWorkItemX;
-        globalsize[1] = size.height / pxPerWorkItemY;
-
-        // Need some padding in the private array for pixels
-        int privDataWidth = roundUp(pxPerWorkItemX + ksize.width - 1, pxLoadNumPixels);
-
-        // Make the global size a nice round number so the runtime can pick
-        // from reasonable choices for the workgroup size
-        const int wgRound = 256;
-        globalsize[0] = roundUp(globalsize[0], wgRound);
-
-        char build_options[1024], cvt[2][40];
-        sprintf(build_options, "-D cn=%d "
-                "-D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d "
-                "-D PX_LOAD_VEC_SIZE=%d -D PX_LOAD_NUM_PX=%d "
-                "-D PX_PER_WI_X=%d -D PX_PER_WI_Y=%d -D PRIV_DATA_WIDTH=%d -D %s -D %s "
-                "-D PX_LOAD_X_ITERATIONS=%d -D PX_LOAD_Y_ITERATIONS=%d "
-                "-D srcT=%s -D srcT1=%s -D dstT=%s -D dstT1=%s -D WT=%s -D WT1=%s "
-                "-D convertToWT=%s -D convertToDstT=%s%s%s -D PX_LOAD_FLOAT_VEC_CONV=convert_%s -D OP_BOX_FILTER",
-                cn, anchor.x, anchor.y, ksize.width, ksize.height,
-                pxLoadVecSize, pxLoadNumPixels,
-                pxPerWorkItemX, pxPerWorkItemY, privDataWidth, borderMap[borderType],
-                isolated ? "BORDER_ISOLATED" : "NO_BORDER_ISOLATED",
-                privDataWidth / pxLoadNumPixels, pxPerWorkItemY + ksize.height - 1,
-                ocl::typeToStr(type), ocl::typeToStr(sdepth), ocl::typeToStr(dtype),
-                ocl::typeToStr(ddepth), ocl::typeToStr(wtype), ocl::typeToStr(wdepth),
-                ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]),
-                ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]),
-                normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "",
-                ocl::typeToStr(CV_MAKE_TYPE(wdepth, pxLoadVecSize)) //PX_LOAD_FLOAT_VEC_CONV
-                );
-
-
-        if (!kernel.create("filterSmall", cv::ocl::imgproc::filterSmall_oclsrc, build_options))
-            return false;
-    }
-    else
-    {
-        localsize = localsize_general;
-        for ( ; ; )
-        {
-            int BLOCK_SIZE_X = tryWorkItems, BLOCK_SIZE_Y = std::min(ksize.height * 10, size.height);
-
-            while (BLOCK_SIZE_X > 32 && BLOCK_SIZE_X >= ksize.width * 2 && BLOCK_SIZE_X > size.width * 2)
-                BLOCK_SIZE_X /= 2;
-            while (BLOCK_SIZE_Y < BLOCK_SIZE_X / 8 && BLOCK_SIZE_Y * computeUnits * 32 < size.height)
-                BLOCK_SIZE_Y *= 2;
-
-            if (ksize.width > BLOCK_SIZE_X || w < ksize.width || h < ksize.height)
-                return false;
-
-            char cvt[2][50];
-            String opts = format("-D LOCAL_SIZE_X=%d -D BLOCK_SIZE_Y=%d -D ST=%s -D DT=%s -D WT=%s -D convertToDT=%s -D convertToWT=%s"
-                                 " -D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d -D %s%s%s%s%s"
-                                 " -D ST1=%s -D DT1=%s -D cn=%d",
-                                 BLOCK_SIZE_X, BLOCK_SIZE_Y, ocl::typeToStr(type), ocl::typeToStr(CV_MAKE_TYPE(ddepth, cn)),
-                                 ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
-                                 ocl::convertTypeStr(wdepth, ddepth, cn, cvt[0]),
-                                 ocl::convertTypeStr(sdepth, wdepth, cn, cvt[1]),
-                                 anchor.x, anchor.y, ksize.width, ksize.height, borderMap[borderType],
-                                 isolated ? " -D BORDER_ISOLATED" : "", doubleSupport ? " -D DOUBLE_SUPPORT" : "",
-                                 normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "",
-                                 ocl::typeToStr(sdepth), ocl::typeToStr(ddepth), cn);
-
-            localsize[0] = BLOCK_SIZE_X;
-            globalsize[0] = divUp(size.width, BLOCK_SIZE_X - (ksize.width - 1)) * BLOCK_SIZE_X;
-            globalsize[1] = divUp(size.height, BLOCK_SIZE_Y);
-
-            kernel.create("boxFilter", cv::ocl::imgproc::boxFilter_oclsrc, opts);
-            if (kernel.empty())
-                return false;
-
-            size_t kernelWorkGroupSize = kernel.workGroupSize();
-            if (localsize[0] <= kernelWorkGroupSize)
-                break;
-            if (BLOCK_SIZE_X < (int)kernelWorkGroupSize)
-                return false;
-
-            tryWorkItems = (int)kernelWorkGroupSize;
-        }
-    }
-
-    _dst.create(size, CV_MAKETYPE(ddepth, cn));
-    UMat dst = _dst.getUMat();
-
-    int idxArg = kernel.set(0, ocl::KernelArg::PtrReadOnly(src));
-    idxArg = kernel.set(idxArg, (int)src.step);
-    int srcOffsetX = (int)((src.offset % src.step) / src.elemSize());
-    int srcOffsetY = (int)(src.offset / src.step);
-    int srcEndX = isolated ? srcOffsetX + size.width : wholeSize.width;
-    int srcEndY = isolated ? srcOffsetY + size.height : wholeSize.height;
-    idxArg = kernel.set(idxArg, srcOffsetX);
-    idxArg = kernel.set(idxArg, srcOffsetY);
-    idxArg = kernel.set(idxArg, srcEndX);
-    idxArg = kernel.set(idxArg, srcEndY);
-    idxArg = kernel.set(idxArg, ocl::KernelArg::WriteOnly(dst));
-    if (normalize)
-        idxArg = kernel.set(idxArg, (float)alpha);
-
-    return kernel.run(2, globalsize, localsize, false);
-}
-
-#endif
-
-}
-
 
-cv::Ptr<cv::BaseRowFilter> cv::getRowSumFilter(int srcType, int sumType, int ksize, int anchor)
-{
     int sdepth = CV_MAT_DEPTH(srcType), ddepth = CV_MAT_DEPTH(sumType);
     CV_Assert( CV_MAT_CN(sumType) == CV_MAT_CN(srcType) );
 
@@ -1434,9 +1206,10 @@ cv::Ptr<cv::BaseRowFilter> cv::getRowSumFilter(int srcType, int sumType, int ksi
 }
 
 
-cv::Ptr<cv::BaseColumnFilter> cv::getColumnSumFilter(int sumType, int dstType, int ksize,
+Ptr<BaseColumnFilter> getColumnSumFilter(int sumType, int dstType, int ksize, int anchor, double scale)
-                                                     int anchor, double scale)
 {
+    CV_INSTRUMENT_REGION();
+
     int sdepth = CV_MAT_DEPTH(sumType), ddepth = CV_MAT_DEPTH(dstType);
     CV_Assert( CV_MAT_CN(sumType) == CV_MAT_CN(dstType) );
 
@@ -1474,9 +1247,11 @@ cv::Ptr<cv::BaseColumnFilter> cv::getColumnSumFilter(int sumType, int dstType, i
 }
 
 
-cv::Ptr<cv::FilterEngine> cv::createBoxFilter( int srcType, int dstType, Size ksize,
+Ptr<FilterEngine> createBoxFilter(int srcType, int dstType, Size ksize,
-                    Point anchor, bool normalize, int borderType )
+                                  Point anchor, bool normalize, int borderType)
 {
+    CV_INSTRUMENT_REGION();
+
     int sdepth = CV_MAT_DEPTH(srcType);
     int cn = CV_MAT_CN(srcType), sumType = CV_64F;
     if( sdepth == CV_8U && CV_MAT_DEPTH(dstType) == CV_8U &&
@@ -1496,199 +1271,12 @@ cv::Ptr<cv::FilterEngine> cv::createBoxFilter( int srcType, int dstType, Size ks
            srcType, dstType, sumType, borderType );
 }
 
-#ifdef HAVE_OPENVX
-namespace cv
-{
-    namespace ovx {
-        template <> inline bool skipSmallImages<VX_KERNEL_BOX_3x3>(int w, int h) { return w*h < 640 * 480; }
-    }
-    static bool openvx_boxfilter(InputArray _src, OutputArray _dst, int ddepth,
-                                 Size ksize, Point anchor,
-                                 bool normalize, int borderType)
-    {
-        if (ddepth < 0)
-            ddepth = CV_8UC1;
-        if (_src.type() != CV_8UC1 || ddepth != CV_8U || !normalize ||
-            _src.cols() < 3 || _src.rows() < 3 ||
-            ksize.width != 3 || ksize.height != 3 ||
-            (anchor.x >= 0 && anchor.x != 1) ||
-            (anchor.y >= 0 && anchor.y != 1) ||
-            ovx::skipSmallImages<VX_KERNEL_BOX_3x3>(_src.cols(), _src.rows()))
-            return false;
-
-        Mat src = _src.getMat();
-
-        if ((borderType & BORDER_ISOLATED) == 0 && src.isSubmatrix())
-            return false; //Process isolated borders only
-        vx_enum border;
-        switch (borderType & ~BORDER_ISOLATED)
-        {
-        case BORDER_CONSTANT:
-            border = VX_BORDER_CONSTANT;
-            break;
-        case BORDER_REPLICATE:
-            border = VX_BORDER_REPLICATE;
-            break;
-        default:
-            return false;
-        }
-
-        _dst.create(src.size(), CV_8UC1);
-        Mat dst = _dst.getMat();
-
-        try
-        {
-            ivx::Context ctx = ovx::getOpenVXContext();
-
-            Mat a;
-            if (dst.data != src.data)
-                a = src;
-            else
-                src.copyTo(a);
-
-            ivx::Image
-                ia = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
-                                                  ivx::Image::createAddressing(a.cols, a.rows, 1, (vx_int32)(a.step)), a.data),
-                ib = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
-                                                  ivx::Image::createAddressing(dst.cols, dst.rows, 1, (vx_int32)(dst.step)), dst.data);
-
-            //ATTENTION: VX_CONTEXT_IMMEDIATE_BORDER attribute change could lead to strange issues in multi-threaded environments
-            //since OpenVX standard says nothing about thread-safety for now
-            ivx::border_t prevBorder = ctx.immediateBorder();
-            ctx.setImmediateBorder(border, (vx_uint8)(0));
-            ivx::IVX_CHECK_STATUS(vxuBox3x3(ctx, ia, ib));
-            ctx.setImmediateBorder(prevBorder);
-        }
-        catch (const ivx::RuntimeError & e)
-        {
-            VX_DbgThrow(e.what());
-        }
-        catch (const ivx::WrapperError & e)
-        {
-            VX_DbgThrow(e.what());
-        }
-
-        return true;
-    }
-}
-#endif
-
-#if defined(HAVE_IPP)
-namespace cv
-{
-static bool ipp_boxfilter(Mat &src, Mat &dst, Size ksize, Point anchor, bool normalize, int borderType)
-{
-#ifdef HAVE_IPP_IW
-    CV_INSTRUMENT_REGION_IPP();
-
-#if IPP_VERSION_X100 < 201801
-    // Problem with SSE42 optimization for 16s and some 8u modes
-    if(ipp::getIppTopFeatures() == ippCPUID_SSE42 && (((src.depth() == CV_16S || src.depth() == CV_16U) && (src.channels() == 3 || src.channels() == 4)) || (src.depth() == CV_8U && src.channels() == 3 && (ksize.width > 5 || ksize.height > 5))))
-        return false;
-
-    // Other optimizations has some degradations too
-    if((((src.depth() == CV_16S || src.depth() == CV_16U) && (src.channels() == 4)) || (src.depth() == CV_8U && src.channels() == 1 && (ksize.width > 5 || ksize.height > 5))))
-        return false;
-#endif
-
-    if(!normalize)
-        return false;
-
-    if(!ippiCheckAnchor(anchor, ksize))
-        return false;
-
-    try
-    {
-        ::ipp::IwiImage       iwSrc      = ippiGetImage(src);
-        ::ipp::IwiImage       iwDst      = ippiGetImage(dst);
-        ::ipp::IwiSize        iwKSize    = ippiGetSize(ksize);
-        ::ipp::IwiBorderSize  borderSize(iwKSize);
-        ::ipp::IwiBorderType  ippBorder(ippiGetBorder(iwSrc, borderType, borderSize));
-        if(!ippBorder)
-            return false;
-
-        CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterBox, iwSrc, iwDst, iwKSize, ::ipp::IwDefault(), ippBorder);
-    }
-    catch (const ::ipp::IwException &)
-    {
-        return false;
-    }
-
-    return true;
-#else
-    CV_UNUSED(src); CV_UNUSED(dst); CV_UNUSED(ksize); CV_UNUSED(anchor); CV_UNUSED(normalize); CV_UNUSED(borderType);
-    return false;
-#endif
-}
-}
-#endif
-
-
-void cv::boxFilter( InputArray _src, OutputArray _dst, int ddepth,
-                Size ksize, Point anchor,
-                bool normalize, int borderType )
-{
-    CV_INSTRUMENT_REGION();
-
-    CV_OCL_RUN(_dst.isUMat() &&
-               (borderType == BORDER_REPLICATE || borderType == BORDER_CONSTANT ||
-                borderType == BORDER_REFLECT || borderType == BORDER_REFLECT_101),
-               ocl_boxFilter3x3_8UC1(_src, _dst, ddepth, ksize, anchor, borderType, normalize))
-
-    CV_OCL_RUN(_dst.isUMat(), ocl_boxFilter(_src, _dst, ddepth, ksize, anchor, borderType, normalize))
-
-    Mat src = _src.getMat();
-    int stype = src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype);
-    if( ddepth < 0 )
-        ddepth = sdepth;
-    _dst.create( src.size(), CV_MAKETYPE(ddepth, cn) );
-    Mat dst = _dst.getMat();
-    if( borderType != BORDER_CONSTANT && normalize && (borderType & BORDER_ISOLATED) != 0 )
-    {
-        if( src.rows == 1 )
-            ksize.height = 1;
-        if( src.cols == 1 )
-            ksize.width = 1;
-    }
-
-    Point ofs;
-    Size wsz(src.cols, src.rows);
-    if(!(borderType&BORDER_ISOLATED))
-        src.locateROI( wsz, ofs );
-
-    CALL_HAL(boxFilter, cv_hal_boxFilter, src.ptr(), src.step, dst.ptr(), dst.step, src.cols, src.rows, sdepth, ddepth, cn,
-             ofs.x, ofs.y, wsz.width - src.cols - ofs.x, wsz.height - src.rows - ofs.y, ksize.width, ksize.height,
-             anchor.x, anchor.y, normalize, borderType&~BORDER_ISOLATED);
-
-    CV_OVX_RUN(true,
-               openvx_boxfilter(src, dst, ddepth, ksize, anchor, normalize, borderType))
-
-    CV_IPP_RUN_FAST(ipp_boxfilter(src, dst, ksize, anchor, normalize, borderType));
-
-    borderType = (borderType&~BORDER_ISOLATED);
-
-    Ptr<FilterEngine> f = createBoxFilter( src.type(), dst.type(),
-                        ksize, anchor, normalize, borderType );
-
-    f->apply( src, dst, wsz, ofs );
-}
-
-
-void cv::blur( InputArray src, OutputArray dst,
-           Size ksize, Point anchor, int borderType )
-{
-    CV_INSTRUMENT_REGION();
-
-    boxFilter( src, dst, -1, ksize, anchor, true, borderType );
-}
 
 
 /****************************************************************************************\
                                     Squared Box Filter
 \****************************************************************************************/
-
+namespace {
-namespace cv
-{
 
 template<typename T, typename ST>
 struct SqrRowSum :
@@ -1703,6 +1291,8 @@ struct SqrRowSum :
 
     virtual void operator()(const uchar* src, uchar* dst, int width, int cn) CV_OVERRIDE
     {
+        CV_INSTRUMENT_REGION();
+
         const T* S = (const T*)src;
         ST* D = (ST*)dst;
         int i = 0, k, ksz_cn = ksize*cn;
@@ -1727,7 +1317,9 @@ struct SqrRowSum :
     }
 };
 
-static Ptr<BaseRowFilter> getSqrRowSumFilter(int srcType, int sumType, int ksize, int anchor)
+} // namespace anon
+
+Ptr<BaseRowFilter> getSqrRowSumFilter(int srcType, int sumType, int ksize, int anchor)
 {
     int sdepth = CV_MAT_DEPTH(srcType), ddepth = CV_MAT_DEPTH(sumType);
     CV_Assert( CV_MAT_CN(sumType) == CV_MAT_CN(srcType) );
@@ -1753,52 +1345,6 @@ static Ptr<BaseRowFilter> getSqrRowSumFilter(int srcType, int sumType, int ksize
                srcType, sumType));
 }
 
-}
+#endif
-
+CV_CPU_OPTIMIZATION_NAMESPACE_END
-void cv::sqrBoxFilter( InputArray _src, OutputArray _dst, int ddepth,
+} // namespace
-                       Size ksize, Point anchor,
-                       bool normalize, int borderType )
-{
-    CV_INSTRUMENT_REGION();
-
-    int srcType = _src.type(), sdepth = CV_MAT_DEPTH(srcType), cn = CV_MAT_CN(srcType);
-    Size size = _src.size();
-
-    if( ddepth < 0 )
-        ddepth = sdepth < CV_32F ? CV_32F : CV_64F;
-
-    if( borderType != BORDER_CONSTANT && normalize )
-    {
-        if( size.height == 1 )
-            ksize.height = 1;
-        if( size.width == 1 )
-            ksize.width = 1;
-    }
-
-    CV_OCL_RUN(_dst.isUMat() && _src.dims() <= 2,
-               ocl_boxFilter(_src, _dst, ddepth, ksize, anchor, borderType, normalize, true))
-
-    int sumDepth = CV_64F;
-    if( sdepth == CV_8U )
-        sumDepth = CV_32S;
-    int sumType = CV_MAKETYPE( sumDepth, cn ), dstType = CV_MAKETYPE(ddepth, cn);
-
-    Mat src = _src.getMat();
-    _dst.create( size, dstType );
-    Mat dst = _dst.getMat();
-
-    Ptr<BaseRowFilter> rowFilter = getSqrRowSumFilter(srcType, sumType, ksize.width, anchor.x );
-    Ptr<BaseColumnFilter> columnFilter = getColumnSumFilter(sumType,
-                                                            dstType, ksize.height, anchor.y,
-                                                            normalize ? 1./(ksize.width*ksize.height) : 1);
-
-    Ptr<FilterEngine> f = makePtr<FilterEngine>(Ptr<BaseFilter>(), rowFilter, columnFilter,
-                                                srcType, dstType, sumType, borderType );
-    Point ofs;
-    Size wsz(src.cols, src.rows);
-    src.locateROI( wsz, ofs );
-
-    f->apply( src, dst, wsz, ofs );
-}
-
-/* End of file. */

modules/imgproc/src/filter.avx2.cpp

@@ -1,197 +0,0 @@
-/*M///////////////////////////////////////////////////////////////////////////////////////
-//
-//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
-//
-//  By downloading, copying, installing or using the software you agree to this license.
-//  If you do not agree to this license, do not download, install,
-//  copy or use the software.
-//
-//
-//                           License Agreement
-//                For Open Source Computer Vision Library
-//
-// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
-// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
-// Third party copyrights are property of their respective owners.
-//
-// Redistribution and use in source and binary forms, with or without modification,
-// are permitted provided that the following conditions are met:
-//
-//   * Redistribution's of source code must retain the above copyright notice,
-//     this list of conditions and the following disclaimer.
-//
-//   * Redistribution's in binary form must reproduce the above copyright notice,
-//     this list of conditions and the following disclaimer in the documentation
-//     and/or other materials provided with the distribution.
-//
-//   * The name of the copyright holders may not be used to endorse or promote products
-//     derived from this software without specific prior written permission.
-//
-// This software is provided by the copyright holders and contributors "as is" and
-// any express or implied warranties, including, but not limited to, the implied
-// warranties of merchantability and fitness for a particular purpose are disclaimed.
-// In no event shall the Intel Corporation or contributors be liable for any direct,
-// indirect, incidental, special, exemplary, or consequential damages
-// (including, but not limited to, procurement of substitute goods or services;
-// loss of use, data, or profits; or business interruption) however caused
-// and on any theory of liability, whether in contract, strict liability,
-// or tort (including negligence or otherwise) arising in any way out of
-// the use of this software, even if advised of the possibility of such damage.
-//
-//M*/
-
-#include "precomp.hpp"
-#include "filter.hpp"
-
-namespace cv
-{
-
-int RowVec_32f_AVX(const float* src0, const float* _kx, float* dst, int width, int cn, int _ksize)
-{
-    int i = 0, k;
-    for (; i <= width - 8; i += 8)
-    {
-        const float* src = src0 + i;
-        __m256 f, x0;
-        __m256 s0 = _mm256_set1_ps(0.0f);
-        for (k = 0; k < _ksize; k++, src += cn)
-        {
-            f = _mm256_set1_ps(_kx[k]);
-            x0 = _mm256_loadu_ps(src);
-#if CV_FMA3
-            s0 = _mm256_fmadd_ps(x0, f, s0);
-#else
-            s0 = _mm256_add_ps(s0, _mm256_mul_ps(x0, f));
-#endif
-        }
-        _mm256_storeu_ps(dst + i, s0);
-    }
-    _mm256_zeroupper();
-    return i;
-}
-
-int SymmColumnVec_32f_Symm_AVX(const float** src, const float* ky, float* dst, float delta, int width, int ksize2)
-{
-    int i = 0, k;
-    const float *S, *S2;
-    const __m128 d4 = _mm_set1_ps(delta);
-    const __m256 d8 = _mm256_set1_ps(delta);
-
-    for( ; i <= width - 16; i += 16 )
-    {
-        __m256 f = _mm256_set1_ps(ky[0]);
-        __m256 s0, s1;
-        __m256 x0;
-        S = src[0] + i;
-        s0 = _mm256_loadu_ps(S);
-#if CV_FMA3
-        s0 = _mm256_fmadd_ps(s0, f, d8);
-#else
-        s0 = _mm256_add_ps(_mm256_mul_ps(s0, f), d8);
-#endif
-        s1 = _mm256_loadu_ps(S+8);
-#if CV_FMA3
-        s1 = _mm256_fmadd_ps(s1, f, d8);
-#else
-        s1 = _mm256_add_ps(_mm256_mul_ps(s1, f), d8);
-#endif
-
-        for( k = 1; k <= ksize2; k++ )
-        {
-            S = src[k] + i;
-            S2 = src[-k] + i;
-            f = _mm256_set1_ps(ky[k]);
-            x0 = _mm256_add_ps(_mm256_loadu_ps(S), _mm256_loadu_ps(S2));
-#if CV_FMA3
-            s0 = _mm256_fmadd_ps(x0, f, s0);
-#else
-            s0 = _mm256_add_ps(s0, _mm256_mul_ps(x0, f));
-#endif
-            x0 = _mm256_add_ps(_mm256_loadu_ps(S+8), _mm256_loadu_ps(S2+8));
-#if CV_FMA3
-            s1 = _mm256_fmadd_ps(x0, f, s1);
-#else
-            s1 = _mm256_add_ps(s1, _mm256_mul_ps(x0, f));
-#endif
-        }
-
-        _mm256_storeu_ps(dst + i, s0);
-        _mm256_storeu_ps(dst + i + 8, s1);
-    }
-
-    for( ; i <= width - 4; i += 4 )
-    {
-        __m128 f = _mm_set1_ps(ky[0]);
-        __m128 x0, s0 = _mm_load_ps(src[0] + i);
-        s0 = _mm_add_ps(_mm_mul_ps(s0, f), d4);
-
-        for( k = 1; k <= ksize2; k++ )
-        {
-            f = _mm_set1_ps(ky[k]);
-            x0 = _mm_add_ps(_mm_load_ps(src[k]+i), _mm_load_ps(src[-k] + i));
-            s0 = _mm_add_ps(s0, _mm_mul_ps(x0, f));
-        }
-
-        _mm_storeu_ps(dst + i, s0);
-    }
-
-    _mm256_zeroupper();
-    return i;
-}
-
-int SymmColumnVec_32f_Unsymm_AVX(const float** src, const float* ky, float* dst, float delta, int width, int ksize2)
-{
-    int i = 0, k;
-    const float *S2;
-    const __m128 d4 = _mm_set1_ps(delta);
-    const __m256 d8 = _mm256_set1_ps(delta);
-
-    for (; i <= width - 16; i += 16)
-    {
-        __m256 f, s0 = d8, s1 = d8;
-        __m256 x0;
-
-        for (k = 1; k <= ksize2; k++)
-        {
-            const float *S = src[k] + i;
-            S2 = src[-k] + i;
-            f = _mm256_set1_ps(ky[k]);
-            x0 = _mm256_sub_ps(_mm256_loadu_ps(S), _mm256_loadu_ps(S2));
-#if CV_FMA3
-            s0 = _mm256_fmadd_ps(x0, f, s0);
-#else
-            s0 = _mm256_add_ps(s0, _mm256_mul_ps(x0, f));
-#endif
-            x0 = _mm256_sub_ps(_mm256_loadu_ps(S + 8), _mm256_loadu_ps(S2 + 8));
-#if CV_FMA3
-            s1 = _mm256_fmadd_ps(x0, f, s1);
-#else
-            s1 = _mm256_add_ps(s1, _mm256_mul_ps(x0, f));
-#endif
-        }
-
-        _mm256_storeu_ps(dst + i, s0);
-        _mm256_storeu_ps(dst + i + 8, s1);
-    }
-
-    for (; i <= width - 4; i += 4)
-    {
-        __m128 f, x0, s0 = d4;
-
-        for (k = 1; k <= ksize2; k++)
-        {
-            f = _mm_set1_ps(ky[k]);
-            x0 = _mm_sub_ps(_mm_load_ps(src[k] + i), _mm_load_ps(src[-k] + i));
-            s0 = _mm_add_ps(s0, _mm_mul_ps(x0, f));
-        }
-
-        _mm_storeu_ps(dst + i, s0);
-    }
-
-    _mm256_zeroupper();
-    return i;
-}
-
-}
-
-/* End of file. */

modules/imgproc/src/filter.hpp

@@ -45,17 +45,13 @@
 
 namespace cv
 {
-#if CV_TRY_AVX2
-    int RowVec_32f_AVX(const float* src0, const float* _kx, float* dst, int width, int cn, int _ksize);
-    int SymmColumnVec_32f_Symm_AVX(const float** src, const float* ky, float* dst, float delta, int width, int ksize2);
-    int SymmColumnVec_32f_Unsymm_AVX(const float** src, const float* ky, float* dst, float delta, int width, int ksize2);
-#endif
-
 #ifdef HAVE_OPENCL
     bool ocl_sepFilter2D( InputArray _src, OutputArray _dst, int ddepth,
                           InputArray _kernelX, InputArray _kernelY, Point anchor,
                           double delta, int borderType );
 #endif
+
+    void preprocess2DKernel(const Mat& kernel, std::vector<Point>& coords, std::vector<uchar>& coeffs);
 }
 
 #endif

modules/imgproc/src/fixedpoint.inl.hpp

@@ -9,10 +9,7 @@
 #ifndef _CV_FIXEDPOINT_HPP_
 #define _CV_FIXEDPOINT_HPP_
 
-#include "opencv2/core/softfloat.hpp"
+namespace {
-
-namespace
-{
 
 class fixedpoint64
 {

modules/imgproc/src/median_blur.dispatch.cpp

@@ -0,0 +1,317 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, 2018, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Copyright (C) 2014-2015, Itseez Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#include "precomp.hpp"
+
+#include <vector>
+
+#include "opencv2/core/hal/intrin.hpp"
+#include "opencl_kernels_imgproc.hpp"
+
+#include "opencv2/core/openvx/ovx_defs.hpp"
+
+#include "median_blur.simd.hpp"
+#include "median_blur.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
+
+namespace cv {
+
+#ifdef HAVE_OPENCL
+
+#define DIVUP(total, grain) ((total + grain - 1) / (grain))
+
+static bool ocl_medianFilter(InputArray _src, OutputArray _dst, int m)
+{
+    size_t localsize[2] = { 16, 16 };
+    size_t globalsize[2];
+    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+
+    if ( !((depth == CV_8U || depth == CV_16U || depth == CV_16S || depth == CV_32F) && cn <= 4 && (m == 3 || m == 5)) )
+        return false;
+
+    Size imgSize = _src.size();
+    bool useOptimized = (1 == cn) &&
+                        (size_t)imgSize.width >= localsize[0] * 8  &&
+                        (size_t)imgSize.height >= localsize[1] * 8 &&
+                        imgSize.width % 4 == 0 &&
+                        imgSize.height % 4 == 0 &&
+                        (ocl::Device::getDefault().isIntel());
+
+    cv::String kname = format( useOptimized ? "medianFilter%d_u" : "medianFilter%d", m) ;
+    cv::String kdefs = useOptimized ?
+                         format("-D T=%s -D T1=%s -D T4=%s%d -D cn=%d -D USE_4OPT", ocl::typeToStr(type),
+                         ocl::typeToStr(depth), ocl::typeToStr(depth), cn*4, cn)
+                         :
+                         format("-D T=%s -D T1=%s -D cn=%d", ocl::typeToStr(type), ocl::typeToStr(depth), cn) ;
+
+    ocl::Kernel k(kname.c_str(), ocl::imgproc::medianFilter_oclsrc, kdefs.c_str() );
+
+    if (k.empty())
+        return false;
+
+    UMat src = _src.getUMat();
+    _dst.create(src.size(), type);
+    UMat dst = _dst.getUMat();
+
+    k.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnly(dst));
+
+    if( useOptimized )
+    {
+        globalsize[0] = DIVUP(src.cols / 4, localsize[0]) * localsize[0];
+        globalsize[1] = DIVUP(src.rows / 4, localsize[1]) * localsize[1];
+    }
+    else
+    {
+        globalsize[0] = (src.cols + localsize[0] + 2) / localsize[0] * localsize[0];
+        globalsize[1] = (src.rows + localsize[1] - 1) / localsize[1] * localsize[1];
+    }
+
+    return k.run(2, globalsize, localsize, false);
+}
+
+#undef DIVUP
+
+#endif
+
+#ifdef HAVE_OPENVX
+namespace ovx {
+    template <> inline bool skipSmallImages<VX_KERNEL_MEDIAN_3x3>(int w, int h) { return w*h < 1280 * 720; }
+}
+static bool openvx_medianFilter(InputArray _src, OutputArray _dst, int ksize)
+{
+    if (_src.type() != CV_8UC1 || _dst.type() != CV_8U
+#ifndef VX_VERSION_1_1
+        || ksize != 3
+#endif
+        )
+        return false;
+
+    Mat src = _src.getMat();
+    Mat dst = _dst.getMat();
+
+    if (
+#ifdef VX_VERSION_1_1
+         ksize != 3 ? ovx::skipSmallImages<VX_KERNEL_NON_LINEAR_FILTER>(src.cols, src.rows) :
+#endif
+         ovx::skipSmallImages<VX_KERNEL_MEDIAN_3x3>(src.cols, src.rows)
+       )
+        return false;
+
+    try
+    {
+        ivx::Context ctx = ovx::getOpenVXContext();
+#ifdef VX_VERSION_1_1
+        if ((vx_size)ksize > ctx.nonlinearMaxDimension())
+            return false;
+#endif
+
+        Mat a;
+        if (dst.data != src.data)
+            a = src;
+        else
+            src.copyTo(a);
+
+        ivx::Image
+            ia = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
+                ivx::Image::createAddressing(a.cols, a.rows, 1, (vx_int32)(a.step)), a.data),
+            ib = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
+                ivx::Image::createAddressing(dst.cols, dst.rows, 1, (vx_int32)(dst.step)), dst.data);
+
+        //ATTENTION: VX_CONTEXT_IMMEDIATE_BORDER attribute change could lead to strange issues in multi-threaded environments
+        //since OpenVX standard says nothing about thread-safety for now
+        ivx::border_t prevBorder = ctx.immediateBorder();
+        ctx.setImmediateBorder(VX_BORDER_REPLICATE);
+#ifdef VX_VERSION_1_1
+        if (ksize == 3)
+#endif
+        {
+            ivx::IVX_CHECK_STATUS(vxuMedian3x3(ctx, ia, ib));
+        }
+#ifdef VX_VERSION_1_1
+        else
+        {
+            ivx::Matrix mtx;
+            if(ksize == 5)
+                mtx = ivx::Matrix::createFromPattern(ctx, VX_PATTERN_BOX, ksize, ksize);
+            else
+            {
+                vx_size supportedSize;
+                ivx::IVX_CHECK_STATUS(vxQueryContext(ctx, VX_CONTEXT_NONLINEAR_MAX_DIMENSION, &supportedSize, sizeof(supportedSize)));
+                if ((vx_size)ksize > supportedSize)
+                {
+                    ctx.setImmediateBorder(prevBorder);
+                    return false;
+                }
+                Mat mask(ksize, ksize, CV_8UC1, Scalar(255));
+                mtx = ivx::Matrix::create(ctx, VX_TYPE_UINT8, ksize, ksize);
+                mtx.copyFrom(mask);
+            }
+            ivx::IVX_CHECK_STATUS(vxuNonLinearFilter(ctx, VX_NONLINEAR_FILTER_MEDIAN, ia, mtx, ib));
+        }
+#endif
+        ctx.setImmediateBorder(prevBorder);
+    }
+    catch (const ivx::RuntimeError & e)
+    {
+        VX_DbgThrow(e.what());
+    }
+    catch (const ivx::WrapperError & e)
+    {
+        VX_DbgThrow(e.what());
+    }
+
+    return true;
+}
+#endif
+
+#ifdef HAVE_IPP
+static bool ipp_medianFilter(Mat &src0, Mat &dst, int ksize)
+{
+    CV_INSTRUMENT_REGION_IPP();
+
+#if IPP_VERSION_X100 < 201801
+    // Degradations for big kernel
+    if(ksize > 7)
+        return false;
+#endif
+
+    {
+        int         bufSize;
+        IppiSize    dstRoiSize = ippiSize(dst.cols, dst.rows), maskSize = ippiSize(ksize, ksize);
+        IppDataType ippType = ippiGetDataType(src0.type());
+        int         channels = src0.channels();
+        IppAutoBuffer<Ipp8u> buffer;
+
+        if(src0.isSubmatrix())
+            return false;
+
+        Mat src;
+        if(dst.data != src0.data)
+            src = src0;
+        else
+            src0.copyTo(src);
+
+        if(ippiFilterMedianBorderGetBufferSize(dstRoiSize, maskSize, ippType, channels, &bufSize) < 0)
+            return false;
+
+        buffer.allocate(bufSize);
+
+        switch(ippType)
+        {
+        case ipp8u:
+            if(channels == 1)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_8u_C1R, src.ptr<Ipp8u>(), (int)src.step, dst.ptr<Ipp8u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else if(channels == 3)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_8u_C3R, src.ptr<Ipp8u>(), (int)src.step, dst.ptr<Ipp8u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else if(channels == 4)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_8u_C4R, src.ptr<Ipp8u>(), (int)src.step, dst.ptr<Ipp8u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else
+                return false;
+        case ipp16u:
+            if(channels == 1)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16u_C1R, src.ptr<Ipp16u>(), (int)src.step, dst.ptr<Ipp16u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else if(channels == 3)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16u_C3R, src.ptr<Ipp16u>(), (int)src.step, dst.ptr<Ipp16u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else if(channels == 4)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16u_C4R, src.ptr<Ipp16u>(), (int)src.step, dst.ptr<Ipp16u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else
+                return false;
+        case ipp16s:
+            if(channels == 1)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16s_C1R, src.ptr<Ipp16s>(), (int)src.step, dst.ptr<Ipp16s>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else if(channels == 3)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16s_C3R, src.ptr<Ipp16s>(), (int)src.step, dst.ptr<Ipp16s>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else if(channels == 4)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16s_C4R, src.ptr<Ipp16s>(), (int)src.step, dst.ptr<Ipp16s>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else
+                return false;
+        case ipp32f:
+            if(channels == 1)
+                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_32f_C1R, src.ptr<Ipp32f>(), (int)src.step, dst.ptr<Ipp32f>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
+            else
+                return false;
+        default:
+            return false;
+        }
+    }
+}
+#endif
+
+void medianBlur( InputArray _src0, OutputArray _dst, int ksize )
+{
+    CV_INSTRUMENT_REGION();
+
+    CV_Assert( (ksize % 2 == 1) && (_src0.dims() <= 2 ));
+
+    if( ksize <= 1 || _src0.empty() )
+    {
+        _src0.copyTo(_dst);
+        return;
+    }
+
+    CV_OCL_RUN(_dst.isUMat(),
+               ocl_medianFilter(_src0,_dst, ksize))
+
+    Mat src0 = _src0.getMat();
+    _dst.create( src0.size(), src0.type() );
+    Mat dst = _dst.getMat();
+
+    CALL_HAL(medianBlur, cv_hal_medianBlur, src0.data, src0.step, dst.data, dst.step, src0.cols, src0.rows, src0.depth(),
+             src0.channels(), ksize);
+
+    CV_OVX_RUN(true,
+               openvx_medianFilter(_src0, _dst, ksize))
+
+    CV_IPP_RUN_FAST(ipp_medianFilter(src0, dst, ksize));
+
+#ifdef HAVE_TEGRA_OPTIMIZATION
+    if (tegra::useTegra() && tegra::medianBlur(src0, dst, ksize))
+        return;
+#endif
+
+    CV_CPU_DISPATCH(medianBlur, (src0, dst, ksize),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+}  // namespace
+
+/* End of file. */

modules/imgproc/src/median_blur.simd.hpp

@@ -46,9 +46,11 @@
 #include <vector>
 
 #include "opencv2/core/hal/intrin.hpp"
-#include "opencl_kernels_imgproc.hpp"
 
-#include "opencv2/core/openvx/ovx_defs.hpp"
+#ifdef _MSC_VER
+#pragma warning(disable: 4244)  // warning C4244: 'argument': conversion from 'int' to 'ushort', possible loss of data
+                                // triggered on intrinsic code from medianBlur_8u_O1()
+#endif
 
 /*
  * This file includes the code, contributed by Simon Perreault
@@ -71,12 +73,18 @@
                                       Median Filter
 \****************************************************************************************/
 
-namespace cv
+namespace cv {
-{
+CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
+// forward declarations
+void medianBlur(const Mat& src0, /*const*/ Mat& dst, int ksize);
+
+#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
 
 static void
 medianBlur_8u_O1( const Mat& _src, Mat& _dst, int ksize )
 {
+    CV_INSTRUMENT_REGION();
+
     typedef ushort HT;
 
     /**
@@ -330,9 +338,6 @@ medianBlur_8u_O1( const Mat& _src, Mat& _dst, int ksize )
                 }
             }
         }
-#if CV_SIMD
-        vx_cleanup();
-#endif
     }
 
 #undef HOP
@@ -342,6 +347,8 @@ medianBlur_8u_O1( const Mat& _src, Mat& _dst, int ksize )
 static void
 medianBlur_8u_Om( const Mat& _src, Mat& _dst, int m )
 {
+    CV_INSTRUMENT_REGION();
+
     #define N  16
     int     zone0[4][N];
     int     zone1[4][N*N];
@@ -671,6 +678,8 @@ template<class Op, class VecOp>
 static void
 medianBlur_SortNet( const Mat& _src, Mat& _dst, int m )
 {
+    CV_INSTRUMENT_REGION();
+
     typedef typename Op::value_type T;
     typedef typename Op::arg_type WT;
     typedef typename VecOp::arg_type VT;
@@ -770,9 +779,6 @@ medianBlur_SortNet( const Mat& _src, Mat& _dst, int m )
                 limit = size.width;
             }
         }
-#if CV_SIMD
-        vx_cleanup();
-#endif
     }
     else if( m == 5 )
     {
@@ -934,266 +940,15 @@ medianBlur_SortNet( const Mat& _src, Mat& _dst, int m )
                 limit = size.width;
             }
         }
-#if CV_SIMD
-        vx_cleanup();
-#endif
-    }
-}
-
-#ifdef HAVE_OPENCL
-
-#define DIVUP(total, grain) ((total + grain - 1) / (grain))
-
-static bool ocl_medianFilter(InputArray _src, OutputArray _dst, int m)
-{
-    size_t localsize[2] = { 16, 16 };
-    size_t globalsize[2];
-    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
-
-    if ( !((depth == CV_8U || depth == CV_16U || depth == CV_16S || depth == CV_32F) && cn <= 4 && (m == 3 || m == 5)) )
-        return false;
-
-    Size imgSize = _src.size();
-    bool useOptimized = (1 == cn) &&
-                        (size_t)imgSize.width >= localsize[0] * 8  &&
-                        (size_t)imgSize.height >= localsize[1] * 8 &&
-                        imgSize.width % 4 == 0 &&
-                        imgSize.height % 4 == 0 &&
-                        (ocl::Device::getDefault().isIntel());
-
-    cv::String kname = format( useOptimized ? "medianFilter%d_u" : "medianFilter%d", m) ;
-    cv::String kdefs = useOptimized ?
-                         format("-D T=%s -D T1=%s -D T4=%s%d -D cn=%d -D USE_4OPT", ocl::typeToStr(type),
-                         ocl::typeToStr(depth), ocl::typeToStr(depth), cn*4, cn)
-                         :
-                         format("-D T=%s -D T1=%s -D cn=%d", ocl::typeToStr(type), ocl::typeToStr(depth), cn) ;
-
-    ocl::Kernel k(kname.c_str(), ocl::imgproc::medianFilter_oclsrc, kdefs.c_str() );
-
-    if (k.empty())
-        return false;
-
-    UMat src = _src.getUMat();
-    _dst.create(src.size(), type);
-    UMat dst = _dst.getUMat();
-
-    k.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnly(dst));
-
-    if( useOptimized )
-    {
-        globalsize[0] = DIVUP(src.cols / 4, localsize[0]) * localsize[0];
-        globalsize[1] = DIVUP(src.rows / 4, localsize[1]) * localsize[1];
-    }
-    else
-    {
-        globalsize[0] = (src.cols + localsize[0] + 2) / localsize[0] * localsize[0];
-        globalsize[1] = (src.rows + localsize[1] - 1) / localsize[1] * localsize[1];
-    }
-
-    return k.run(2, globalsize, localsize, false);
-}
-
-#undef DIVUP
-
-#endif
-
-#ifdef HAVE_OPENVX
-namespace ovx {
-    template <> inline bool skipSmallImages<VX_KERNEL_MEDIAN_3x3>(int w, int h) { return w*h < 1280 * 720; }
-}
-static bool openvx_medianFilter(InputArray _src, OutputArray _dst, int ksize)
-{
-    if (_src.type() != CV_8UC1 || _dst.type() != CV_8U
-#ifndef VX_VERSION_1_1
-        || ksize != 3
-#endif
-        )
-        return false;
-
-    Mat src = _src.getMat();
-    Mat dst = _dst.getMat();
-
-    if (
-#ifdef VX_VERSION_1_1
-         ksize != 3 ? ovx::skipSmallImages<VX_KERNEL_NON_LINEAR_FILTER>(src.cols, src.rows) :
-#endif
-         ovx::skipSmallImages<VX_KERNEL_MEDIAN_3x3>(src.cols, src.rows)
-       )
-        return false;
-
-    try
-    {
-        ivx::Context ctx = ovx::getOpenVXContext();
-#ifdef VX_VERSION_1_1
-        if ((vx_size)ksize > ctx.nonlinearMaxDimension())
-            return false;
-#endif
-
-        Mat a;
-        if (dst.data != src.data)
-            a = src;
-        else
-            src.copyTo(a);
-
-        ivx::Image
-            ia = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
-                ivx::Image::createAddressing(a.cols, a.rows, 1, (vx_int32)(a.step)), a.data),
-            ib = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
-                ivx::Image::createAddressing(dst.cols, dst.rows, 1, (vx_int32)(dst.step)), dst.data);
-
-        //ATTENTION: VX_CONTEXT_IMMEDIATE_BORDER attribute change could lead to strange issues in multi-threaded environments
-        //since OpenVX standard says nothing about thread-safety for now
-        ivx::border_t prevBorder = ctx.immediateBorder();
-        ctx.setImmediateBorder(VX_BORDER_REPLICATE);
-#ifdef VX_VERSION_1_1
-        if (ksize == 3)
-#endif
-        {
-            ivx::IVX_CHECK_STATUS(vxuMedian3x3(ctx, ia, ib));
-        }
-#ifdef VX_VERSION_1_1
-        else
-        {
-            ivx::Matrix mtx;
-            if(ksize == 5)
-                mtx = ivx::Matrix::createFromPattern(ctx, VX_PATTERN_BOX, ksize, ksize);
-            else
-            {
-                vx_size supportedSize;
-                ivx::IVX_CHECK_STATUS(vxQueryContext(ctx, VX_CONTEXT_NONLINEAR_MAX_DIMENSION, &supportedSize, sizeof(supportedSize)));
-                if ((vx_size)ksize > supportedSize)
-                {
-                    ctx.setImmediateBorder(prevBorder);
-                    return false;
-                }
-                Mat mask(ksize, ksize, CV_8UC1, Scalar(255));
-                mtx = ivx::Matrix::create(ctx, VX_TYPE_UINT8, ksize, ksize);
-                mtx.copyFrom(mask);
-            }
-            ivx::IVX_CHECK_STATUS(vxuNonLinearFilter(ctx, VX_NONLINEAR_FILTER_MEDIAN, ia, mtx, ib));
-        }
-#endif
-        ctx.setImmediateBorder(prevBorder);
-    }
-    catch (const ivx::RuntimeError & e)
-    {
-        VX_DbgThrow(e.what());
     }
-    catch (const ivx::WrapperError & e)
-    {
-        VX_DbgThrow(e.what());
-    }
-
-    return true;
 }
-#endif
-
-#ifdef HAVE_IPP
-static bool ipp_medianFilter(Mat &src0, Mat &dst, int ksize)
-{
-    CV_INSTRUMENT_REGION_IPP();
-
-#if IPP_VERSION_X100 < 201801
-    // Degradations for big kernel
-    if(ksize > 7)
-        return false;
-#endif
-
-    {
-        int         bufSize;
-        IppiSize    dstRoiSize = ippiSize(dst.cols, dst.rows), maskSize = ippiSize(ksize, ksize);
-        IppDataType ippType = ippiGetDataType(src0.type());
-        int         channels = src0.channels();
-        IppAutoBuffer<Ipp8u> buffer;
-
-        if(src0.isSubmatrix())
-            return false;
-
-        Mat src;
-        if(dst.data != src0.data)
-            src = src0;
-        else
-            src0.copyTo(src);
-
-        if(ippiFilterMedianBorderGetBufferSize(dstRoiSize, maskSize, ippType, channels, &bufSize) < 0)
-            return false;
 
-        buffer.allocate(bufSize);
+} // namespace anon
 
-        switch(ippType)
+void medianBlur(const Mat& src0, /*const*/ Mat& dst, int ksize)
-        {
-        case ipp8u:
-            if(channels == 1)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_8u_C1R, src.ptr<Ipp8u>(), (int)src.step, dst.ptr<Ipp8u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else if(channels == 3)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_8u_C3R, src.ptr<Ipp8u>(), (int)src.step, dst.ptr<Ipp8u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else if(channels == 4)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_8u_C4R, src.ptr<Ipp8u>(), (int)src.step, dst.ptr<Ipp8u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else
-                return false;
-        case ipp16u:
-            if(channels == 1)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16u_C1R, src.ptr<Ipp16u>(), (int)src.step, dst.ptr<Ipp16u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else if(channels == 3)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16u_C3R, src.ptr<Ipp16u>(), (int)src.step, dst.ptr<Ipp16u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else if(channels == 4)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16u_C4R, src.ptr<Ipp16u>(), (int)src.step, dst.ptr<Ipp16u>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else
-                return false;
-        case ipp16s:
-            if(channels == 1)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16s_C1R, src.ptr<Ipp16s>(), (int)src.step, dst.ptr<Ipp16s>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else if(channels == 3)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16s_C3R, src.ptr<Ipp16s>(), (int)src.step, dst.ptr<Ipp16s>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else if(channels == 4)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_16s_C4R, src.ptr<Ipp16s>(), (int)src.step, dst.ptr<Ipp16s>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else
-                return false;
-        case ipp32f:
-            if(channels == 1)
-                return CV_INSTRUMENT_FUN_IPP(ippiFilterMedianBorder_32f_C1R, src.ptr<Ipp32f>(), (int)src.step, dst.ptr<Ipp32f>(), (int)dst.step, dstRoiSize, maskSize, ippBorderRepl, 0, buffer) >= 0;
-            else
-                return false;
-        default:
-            return false;
-        }
-    }
-}
-#endif
-}
-
-void medianBlur( InputArray _src0, OutputArray _dst, int ksize )
 {
     CV_INSTRUMENT_REGION();
 
-    CV_Assert( (ksize % 2 == 1) && (_src0.dims() <= 2 ));
-
-    if( ksize <= 1 || _src0.empty() )
-    {
-        _src0.copyTo(_dst);
-        return;
-    }
-
-    CV_OCL_RUN(_dst.isUMat(),
-               ocl_medianFilter(_src0,_dst, ksize))
-
-    Mat src0 = _src0.getMat();
-    _dst.create( src0.size(), src0.type() );
-    Mat dst = _dst.getMat();
-
-    CALL_HAL(medianBlur, cv_hal_medianBlur, src0.data, src0.step, dst.data, dst.step, src0.cols, src0.rows, src0.depth(),
-             src0.channels(), ksize);
-
-    CV_OVX_RUN(true,
-               openvx_medianFilter(_src0, _dst, ksize))
-
-    CV_IPP_RUN_FAST(ipp_medianFilter(src0, dst, ksize));
-
-#ifdef HAVE_TEGRA_OPTIMIZATION
-    if (tegra::useTegra() && tegra::medianBlur(src0, dst, ksize))
-        return;
-#endif
-
     bool useSortNet = ksize == 3 || (ksize == 5
 #if !(CV_SIMD)
             && ( src0.depth() > CV_8U || src0.channels() == 2 || src0.channels() > 4 )
@@ -1223,6 +978,7 @@ void medianBlur( InputArray _src0, OutputArray _dst, int ksize )
     }
     else
     {
+        // TODO AVX guard (external call)
         cv::copyMakeBorder( src0, src, 0, 0, ksize/2, ksize/2, BORDER_REPLICATE|BORDER_ISOLATED);
 
         int cn = src0.channels();
@@ -1237,6 +993,6 @@ void medianBlur( InputArray _src0, OutputArray _dst, int ksize )
     }
 }
 
-}
+#endif
-
+CV_CPU_OPTIMIZATION_NAMESPACE_END
-/* End of file. */
+} // namespace

modules/imgproc/src/morph.dispatch.cpp

@@ -48,779 +48,49 @@
 #include "opencv2/core/hal/intrin.hpp"
 #include <opencv2/core/utils/configuration.private.hpp>
 
+#include "morph.simd.hpp"
+#include "morph.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
+
+
 /****************************************************************************************\
                      Basic Morphological Operations: Erosion & Dilation
 \****************************************************************************************/
 
-using namespace std;
+namespace cv {
-
-namespace cv
-{
-
-template<typename T> struct MinOp
-{
-    typedef T type1;
-    typedef T type2;
-    typedef T rtype;
-    T operator ()(const T a, const T b) const { return std::min(a, b); }
-};
-
-template<typename T> struct MaxOp
-{
-    typedef T type1;
-    typedef T type2;
-    typedef T rtype;
-    T operator ()(const T a, const T b) const { return std::max(a, b); }
-};
-
-#undef CV_MIN_8U
-#undef CV_MAX_8U
-#define CV_MIN_8U(a,b)       ((a) - CV_FAST_CAST_8U((a) - (b)))
-#define CV_MAX_8U(a,b)       ((a) + CV_FAST_CAST_8U((b) - (a)))
-
-template<> inline uchar MinOp<uchar>::operator ()(const uchar a, const uchar b) const { return CV_MIN_8U(a, b); }
-template<> inline uchar MaxOp<uchar>::operator ()(const uchar a, const uchar b) const { return CV_MAX_8U(a, b); }
-
-struct MorphRowNoVec
-{
-    MorphRowNoVec(int, int) {}
-    int operator()(const uchar*, uchar*, int, int) const { return 0; }
-};
-
-struct MorphColumnNoVec
-{
-    MorphColumnNoVec(int, int) {}
-    int operator()(const uchar**, uchar*, int, int, int) const { return 0; }
-};
-
-struct MorphNoVec
-{
-    int operator()(uchar**, int, uchar*, int) const { return 0; }
-};
-
-#if CV_SIMD
-
-template<class VecUpdate> struct MorphRowVec
-{
-    typedef typename VecUpdate::vtype vtype;
-    typedef typename vtype::lane_type stype;
-    MorphRowVec(int _ksize, int _anchor) : ksize(_ksize), anchor(_anchor) {}
-    int operator()(const uchar* src, uchar* dst, int width, int cn) const
-    {
-        int i, k, _ksize = ksize*cn;
-        width *= cn;
-        VecUpdate updateOp;
-
-        for( i = 0; i <= width - 4*vtype::nlanes; i += 4*vtype::nlanes )
-        {
-            vtype s0 = vx_load((const stype*)src + i);
-            vtype s1 = vx_load((const stype*)src + i + vtype::nlanes);
-            vtype s2 = vx_load((const stype*)src + i + 2*vtype::nlanes);
-            vtype s3 = vx_load((const stype*)src + i + 3*vtype::nlanes);
-            for (k = cn; k < _ksize; k += cn)
-            {
-                s0 = updateOp(s0, vx_load((const stype*)src + i + k));
-                s1 = updateOp(s1, vx_load((const stype*)src + i + k + vtype::nlanes));
-                s2 = updateOp(s2, vx_load((const stype*)src + i + k + 2*vtype::nlanes));
-                s3 = updateOp(s3, vx_load((const stype*)src + i + k + 3*vtype::nlanes));
-            }
-            v_store((stype*)dst + i, s0);
-            v_store((stype*)dst + i + vtype::nlanes, s1);
-            v_store((stype*)dst + i + 2*vtype::nlanes, s2);
-            v_store((stype*)dst + i + 3*vtype::nlanes, s3);
-        }
-        if( i <= width - 2*vtype::nlanes )
-        {
-            vtype s0 = vx_load((const stype*)src + i);
-            vtype s1 = vx_load((const stype*)src + i + vtype::nlanes);
-            for( k = cn; k < _ksize; k += cn )
-            {
-                s0 = updateOp(s0, vx_load((const stype*)src + i + k));
-                s1 = updateOp(s1, vx_load((const stype*)src + i + k + vtype::nlanes));
-            }
-            v_store((stype*)dst + i, s0);
-            v_store((stype*)dst + i + vtype::nlanes, s1);
-            i += 2*vtype::nlanes;
-        }
-        if( i <= width - vtype::nlanes )
-        {
-            vtype s = vx_load((const stype*)src + i);
-            for( k = cn; k < _ksize; k += cn )
-                s = updateOp(s, vx_load((const stype*)src + i + k));
-            v_store((stype*)dst + i, s);
-            i += vtype::nlanes;
-        }
-        if( i <= width - vtype::nlanes/2 )
-        {
-            vtype s = vx_load_low((const stype*)src + i);
-            for( k = cn; k < _ksize; k += cn )
-                s = updateOp(s, vx_load_low((const stype*)src + i + k));
-            v_store_low((stype*)dst + i, s);
-            i += vtype::nlanes/2;
-        }
-
-        return i - i % cn;
-    }
-
-    int ksize, anchor;
-};
-
-
-template<class VecUpdate> struct MorphColumnVec
-{
-    typedef typename VecUpdate::vtype vtype;
-    typedef typename vtype::lane_type stype;
-    MorphColumnVec(int _ksize, int _anchor) : ksize(_ksize), anchor(_anchor) {}
-    int operator()(const uchar** _src, uchar* _dst, int dststep, int count, int width) const
-    {
-        int i = 0, k, _ksize = ksize;
-        VecUpdate updateOp;
-
-        for( i = 0; i < count + ksize - 1; i++ )
-            CV_Assert( ((size_t)_src[i] & (CV_SIMD_WIDTH-1)) == 0 );
-
-        const stype** src = (const stype**)_src;
-        stype* dst = (stype*)_dst;
-        dststep /= sizeof(dst[0]);
-
-        for( ; _ksize > 1 && count > 1; count -= 2, dst += dststep*2, src += 2 )
-        {
-            for( i = 0; i <= width - 4*vtype::nlanes; i += 4*vtype::nlanes)
-            {
-                const stype* sptr = src[1] + i;
-                vtype s0 = vx_load_aligned(sptr);
-                vtype s1 = vx_load_aligned(sptr + vtype::nlanes);
-                vtype s2 = vx_load_aligned(sptr + 2*vtype::nlanes);
-                vtype s3 = vx_load_aligned(sptr + 3*vtype::nlanes);
-
-                for( k = 2; k < _ksize; k++ )
-                {
-                    sptr = src[k] + i;
-                    s0 = updateOp(s0, vx_load_aligned(sptr));
-                    s1 = updateOp(s1, vx_load_aligned(sptr + vtype::nlanes));
-                    s2 = updateOp(s2, vx_load_aligned(sptr + 2*vtype::nlanes));
-                    s3 = updateOp(s3, vx_load_aligned(sptr + 3*vtype::nlanes));
-                }
-
-                sptr = src[0] + i;
-                v_store(dst + i, updateOp(s0, vx_load_aligned(sptr)));
-                v_store(dst + i + vtype::nlanes, updateOp(s1, vx_load_aligned(sptr + vtype::nlanes)));
-                v_store(dst + i + 2*vtype::nlanes, updateOp(s2, vx_load_aligned(sptr + 2*vtype::nlanes)));
-                v_store(dst + i + 3*vtype::nlanes, updateOp(s3, vx_load_aligned(sptr + 3*vtype::nlanes)));
-
-                sptr = src[k] + i;
-                v_store(dst + dststep + i, updateOp(s0, vx_load_aligned(sptr)));
-                v_store(dst + dststep + i + vtype::nlanes, updateOp(s1, vx_load_aligned(sptr + vtype::nlanes)));
-                v_store(dst + dststep + i + 2*vtype::nlanes, updateOp(s2, vx_load_aligned(sptr + 2*vtype::nlanes)));
-                v_store(dst + dststep + i + 3*vtype::nlanes, updateOp(s3, vx_load_aligned(sptr + 3*vtype::nlanes)));
-            }
-            if( i <= width - 2*vtype::nlanes )
-            {
-                const stype* sptr = src[1] + i;
-                vtype s0 = vx_load_aligned(sptr);
-                vtype s1 = vx_load_aligned(sptr + vtype::nlanes);
-
-                for( k = 2; k < _ksize; k++ )
-                {
-                    sptr = src[k] + i;
-                    s0 = updateOp(s0, vx_load_aligned(sptr));
-                    s1 = updateOp(s1, vx_load_aligned(sptr + vtype::nlanes));
-                }
-
-                sptr = src[0] + i;
-                v_store(dst + i, updateOp(s0, vx_load_aligned(sptr)));
-                v_store(dst + i + vtype::nlanes, updateOp(s1, vx_load_aligned(sptr + vtype::nlanes)));
-
-                sptr = src[k] + i;
-                v_store(dst + dststep + i, updateOp(s0, vx_load_aligned(sptr)));
-                v_store(dst + dststep + i + vtype::nlanes, updateOp(s1, vx_load_aligned(sptr + vtype::nlanes)));
-                i += 2*vtype::nlanes;
-            }
-            if( i <= width - vtype::nlanes )
-            {
-                vtype s0 = vx_load_aligned(src[1] + i);
-
-                for( k = 2; k < _ksize; k++ )
-                    s0 = updateOp(s0, vx_load_aligned(src[k] + i));
-
-                v_store(dst + i, updateOp(s0, vx_load_aligned(src[0] + i)));
-                v_store(dst + dststep + i, updateOp(s0, vx_load_aligned(src[k] + i)));
-                i += vtype::nlanes;
-            }
-            if( i <= width - vtype::nlanes/2 )
-            {
-                vtype s0 = vx_load_low(src[1] + i);
-
-                for( k = 2; k < _ksize; k++ )
-                    s0 = updateOp(s0, vx_load_low(src[k] + i));
-
-                v_store_low(dst + i, updateOp(s0, vx_load_low(src[0] + i)));
-                v_store_low(dst + dststep + i, updateOp(s0, vx_load_low(src[k] + i)));
-                i += vtype::nlanes/2;
-            }
-        }
-
-        for( ; count > 0; count--, dst += dststep, src++ )
-        {
-            for( i = 0; i <= width - 4*vtype::nlanes; i += 4*vtype::nlanes)
-            {
-                const stype* sptr = src[0] + i;
-                vtype s0 = vx_load_aligned(sptr);
-                vtype s1 = vx_load_aligned(sptr + vtype::nlanes);
-                vtype s2 = vx_load_aligned(sptr + 2*vtype::nlanes);
-                vtype s3 = vx_load_aligned(sptr + 3*vtype::nlanes);
-
-                for( k = 1; k < _ksize; k++ )
-                {
-                    sptr = src[k] + i;
-                    s0 = updateOp(s0, vx_load_aligned(sptr));
-                    s1 = updateOp(s1, vx_load_aligned(sptr + vtype::nlanes));
-                    s2 = updateOp(s2, vx_load_aligned(sptr + 2*vtype::nlanes));
-                    s3 = updateOp(s3, vx_load_aligned(sptr + 3*vtype::nlanes));
-                }
-                v_store(dst + i, s0);
-                v_store(dst + i + vtype::nlanes, s1);
-                v_store(dst + i + 2*vtype::nlanes, s2);
-                v_store(dst + i + 3*vtype::nlanes, s3);
-            }
-            if( i <= width - 2*vtype::nlanes )
-            {
-                const stype* sptr = src[0] + i;
-                vtype s0 = vx_load_aligned(sptr);
-                vtype s1 = vx_load_aligned(sptr + vtype::nlanes);
-
-                for( k = 1; k < _ksize; k++ )
-                {
-                    sptr = src[k] + i;
-                    s0 = updateOp(s0, vx_load_aligned(sptr));
-                    s1 = updateOp(s1, vx_load_aligned(sptr + vtype::nlanes));
-                }
-                v_store(dst + i, s0);
-                v_store(dst + i + vtype::nlanes, s1);
-                i += 2*vtype::nlanes;
-            }
-            if( i <= width - vtype::nlanes )
-            {
-                vtype s0 = vx_load_aligned(src[0] + i);
-
-                for( k = 1; k < _ksize; k++ )
-                    s0 = updateOp(s0, vx_load_aligned(src[k] + i));
-                v_store(dst + i, s0);
-                i += vtype::nlanes;
-            }
-            if( i <= width - vtype::nlanes/2 )
-            {
-                vtype s0 = vx_load_low(src[0] + i);
-
-                for( k = 1; k < _ksize; k++ )
-                    s0 = updateOp(s0, vx_load_low(src[k] + i));
-                v_store_low(dst + i, s0);
-                i += vtype::nlanes/2;
-            }
-        }
-
-        return i;
-    }
-
-    int ksize, anchor;
-};
-
-
-template<class VecUpdate> struct MorphVec
-{
-    typedef typename VecUpdate::vtype vtype;
-    typedef typename vtype::lane_type stype;
-    int operator()(uchar** _src, int nz, uchar* _dst, int width) const
-    {
-        const stype** src = (const stype**)_src;
-        stype* dst = (stype*)_dst;
-        int i, k;
-        VecUpdate updateOp;
-
-        for( i = 0; i <= width - 4*vtype::nlanes; i += 4*vtype::nlanes )
-        {
-            const stype* sptr = src[0] + i;
-            vtype s0 = vx_load(sptr);
-            vtype s1 = vx_load(sptr + vtype::nlanes);
-            vtype s2 = vx_load(sptr + 2*vtype::nlanes);
-            vtype s3 = vx_load(sptr + 3*vtype::nlanes);
-            for( k = 1; k < nz; k++ )
-            {
-                sptr = src[k] + i;
-                s0 = updateOp(s0, vx_load(sptr));
-                s1 = updateOp(s1, vx_load(sptr + vtype::nlanes));
-                s2 = updateOp(s2, vx_load(sptr + 2*vtype::nlanes));
-                s3 = updateOp(s3, vx_load(sptr + 3*vtype::nlanes));
-            }
-            v_store(dst + i, s0);
-            v_store(dst + i + vtype::nlanes, s1);
-            v_store(dst + i + 2*vtype::nlanes, s2);
-            v_store(dst + i + 3*vtype::nlanes, s3);
-        }
-        if( i <= width - 2*vtype::nlanes )
-        {
-            const stype* sptr = src[0] + i;
-            vtype s0 = vx_load(sptr);
-            vtype s1 = vx_load(sptr + vtype::nlanes);
-            for( k = 1; k < nz; k++ )
-            {
-                sptr = src[k] + i;
-                s0 = updateOp(s0, vx_load(sptr));
-                s1 = updateOp(s1, vx_load(sptr + vtype::nlanes));
-            }
-            v_store(dst + i, s0);
-            v_store(dst + i + vtype::nlanes, s1);
-            i += 2*vtype::nlanes;
-        }
-        if( i <= width - vtype::nlanes )
-        {
-            vtype s0 = vx_load(src[0] + i);
-            for( k = 1; k < nz; k++ )
-                s0 = updateOp(s0, vx_load(src[k] + i));
-            v_store(dst + i, s0);
-            i += vtype::nlanes;
-        }
-        if( i <= width - vtype::nlanes/2 )
-        {
-            vtype s0 = vx_load_low(src[0] + i);
-            for( k = 1; k < nz; k++ )
-                s0 = updateOp(s0, vx_load_low(src[k] + i));
-            v_store_low(dst + i, s0);
-            i += vtype::nlanes/2;
-        }
-        return i;
-    }
-};
-
-template <typename T> struct VMin
-{
-    typedef T vtype;
-    vtype operator()(const vtype& a, const vtype& b) const { return v_min(a,b); }
-};
-template <typename T> struct VMax
-{
-    typedef T vtype;
-    vtype operator()(const vtype& a, const vtype& b) const { return v_max(a,b); }
-};
-
-typedef MorphRowVec<VMin<v_uint8> > ErodeRowVec8u;
-typedef MorphRowVec<VMax<v_uint8> > DilateRowVec8u;
-typedef MorphRowVec<VMin<v_uint16> > ErodeRowVec16u;
-typedef MorphRowVec<VMax<v_uint16> > DilateRowVec16u;
-typedef MorphRowVec<VMin<v_int16> > ErodeRowVec16s;
-typedef MorphRowVec<VMax<v_int16> > DilateRowVec16s;
-typedef MorphRowVec<VMin<v_float32> > ErodeRowVec32f;
-typedef MorphRowVec<VMax<v_float32> > DilateRowVec32f;
-
-typedef MorphColumnVec<VMin<v_uint8> > ErodeColumnVec8u;
-typedef MorphColumnVec<VMax<v_uint8> > DilateColumnVec8u;
-typedef MorphColumnVec<VMin<v_uint16> > ErodeColumnVec16u;
-typedef MorphColumnVec<VMax<v_uint16> > DilateColumnVec16u;
-typedef MorphColumnVec<VMin<v_int16> > ErodeColumnVec16s;
-typedef MorphColumnVec<VMax<v_int16> > DilateColumnVec16s;
-typedef MorphColumnVec<VMin<v_float32> > ErodeColumnVec32f;
-typedef MorphColumnVec<VMax<v_float32> > DilateColumnVec32f;
-
-typedef MorphVec<VMin<v_uint8> > ErodeVec8u;
-typedef MorphVec<VMax<v_uint8> > DilateVec8u;
-typedef MorphVec<VMin<v_uint16> > ErodeVec16u;
-typedef MorphVec<VMax<v_uint16> > DilateVec16u;
-typedef MorphVec<VMin<v_int16> > ErodeVec16s;
-typedef MorphVec<VMax<v_int16> > DilateVec16s;
-typedef MorphVec<VMin<v_float32> > ErodeVec32f;
-typedef MorphVec<VMax<v_float32> > DilateVec32f;
-
-#else
-
-typedef MorphRowNoVec ErodeRowVec8u;
-typedef MorphRowNoVec DilateRowVec8u;
-
-typedef MorphColumnNoVec ErodeColumnVec8u;
-typedef MorphColumnNoVec DilateColumnVec8u;
-
-typedef MorphRowNoVec ErodeRowVec16u;
-typedef MorphRowNoVec DilateRowVec16u;
-typedef MorphRowNoVec ErodeRowVec16s;
-typedef MorphRowNoVec DilateRowVec16s;
-typedef MorphRowNoVec ErodeRowVec32f;
-typedef MorphRowNoVec DilateRowVec32f;
-
-typedef MorphColumnNoVec ErodeColumnVec16u;
-typedef MorphColumnNoVec DilateColumnVec16u;
-typedef MorphColumnNoVec ErodeColumnVec16s;
-typedef MorphColumnNoVec DilateColumnVec16s;
-typedef MorphColumnNoVec ErodeColumnVec32f;
-typedef MorphColumnNoVec DilateColumnVec32f;
-
-typedef MorphNoVec ErodeVec8u;
-typedef MorphNoVec DilateVec8u;
-typedef MorphNoVec ErodeVec16u;
-typedef MorphNoVec DilateVec16u;
-typedef MorphNoVec ErodeVec16s;
-typedef MorphNoVec DilateVec16s;
-typedef MorphNoVec ErodeVec32f;
-typedef MorphNoVec DilateVec32f;
-
-#endif
-
-typedef MorphRowNoVec ErodeRowVec64f;
-typedef MorphRowNoVec DilateRowVec64f;
-typedef MorphColumnNoVec ErodeColumnVec64f;
-typedef MorphColumnNoVec DilateColumnVec64f;
-typedef MorphNoVec ErodeVec64f;
-typedef MorphNoVec DilateVec64f;
-
-
-template<class Op, class VecOp> struct MorphRowFilter : public BaseRowFilter
-{
-    typedef typename Op::rtype T;
-
-    MorphRowFilter( int _ksize, int _anchor ) : vecOp(_ksize, _anchor)
-    {
-        ksize = _ksize;
-        anchor = _anchor;
-    }
-
-    void operator()(const uchar* src, uchar* dst, int width, int cn) CV_OVERRIDE
-    {
-        int i, j, k, _ksize = ksize*cn;
-        const T* S = (const T*)src;
-        Op op;
-        T* D = (T*)dst;
-
-        if( _ksize == cn )
-        {
-            for( i = 0; i < width*cn; i++ )
-                D[i] = S[i];
-            return;
-        }
-
-        int i0 = vecOp(src, dst, width, cn);
-        width *= cn;
-
-        for( k = 0; k < cn; k++, S++, D++ )
-        {
-            for( i = i0; i <= width - cn*2; i += cn*2 )
-            {
-                const T* s = S + i;
-                T m = s[cn];
-                for( j = cn*2; j < _ksize; j += cn )
-                    m = op(m, s[j]);
-                D[i] = op(m, s[0]);
-                D[i+cn] = op(m, s[j]);
-            }
-
-            for( ; i < width; i += cn )
-            {
-                const T* s = S + i;
-                T m = s[0];
-                for( j = cn; j < _ksize; j += cn )
-                    m = op(m, s[j]);
-                D[i] = m;
-            }
-        }
-    }
-
-    VecOp vecOp;
-};
-
-
-template<class Op, class VecOp> struct MorphColumnFilter : public BaseColumnFilter
-{
-    typedef typename Op::rtype T;
-
-    MorphColumnFilter( int _ksize, int _anchor ) : vecOp(_ksize, _anchor)
-    {
-        ksize = _ksize;
-        anchor = _anchor;
-    }
-
-    void operator()(const uchar** _src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE
-    {
-        int i, k, _ksize = ksize;
-        const T** src = (const T**)_src;
-        T* D = (T*)dst;
-        Op op;
-
-        int i0 = vecOp(_src, dst, dststep, count, width);
-        dststep /= sizeof(D[0]);
-
-        for( ; _ksize > 1 && count > 1; count -= 2, D += dststep*2, src += 2 )
-        {
-            i = i0;
-            #if CV_ENABLE_UNROLLED
-            for( ; i <= width - 4; i += 4 )
-            {
-                const T* sptr = src[1] + i;
-                T s0 = sptr[0], s1 = sptr[1], s2 = sptr[2], s3 = sptr[3];
-
-                for( k = 2; k < _ksize; k++ )
-                {
-                    sptr = src[k] + i;
-                    s0 = op(s0, sptr[0]); s1 = op(s1, sptr[1]);
-                    s2 = op(s2, sptr[2]); s3 = op(s3, sptr[3]);
-                }
-
-                sptr = src[0] + i;
-                D[i] = op(s0, sptr[0]);
-                D[i+1] = op(s1, sptr[1]);
-                D[i+2] = op(s2, sptr[2]);
-                D[i+3] = op(s3, sptr[3]);
-
-                sptr = src[k] + i;
-                D[i+dststep] = op(s0, sptr[0]);
-                D[i+dststep+1] = op(s1, sptr[1]);
-                D[i+dststep+2] = op(s2, sptr[2]);
-                D[i+dststep+3] = op(s3, sptr[3]);
-            }
-            #endif
-            for( ; i < width; i++ )
-            {
-                T s0 = src[1][i];
-
-                for( k = 2; k < _ksize; k++ )
-                    s0 = op(s0, src[k][i]);
-
-                D[i] = op(s0, src[0][i]);
-                D[i+dststep] = op(s0, src[k][i]);
-            }
-        }
-
-        for( ; count > 0; count--, D += dststep, src++ )
-        {
-            i = i0;
-            #if CV_ENABLE_UNROLLED
-            for( ; i <= width - 4; i += 4 )
-            {
-                const T* sptr = src[0] + i;
-                T s0 = sptr[0], s1 = sptr[1], s2 = sptr[2], s3 = sptr[3];
-
-                for( k = 1; k < _ksize; k++ )
-                {
-                    sptr = src[k] + i;
-                    s0 = op(s0, sptr[0]); s1 = op(s1, sptr[1]);
-                    s2 = op(s2, sptr[2]); s3 = op(s3, sptr[3]);
-                }
-
-                D[i] = s0; D[i+1] = s1;
-                D[i+2] = s2; D[i+3] = s3;
-            }
-            #endif
-            for( ; i < width; i++ )
-            {
-                T s0 = src[0][i];
-                for( k = 1; k < _ksize; k++ )
-                    s0 = op(s0, src[k][i]);
-                D[i] = s0;
-            }
-        }
-    }
-
-    VecOp vecOp;
-};
-
-
-template<class Op, class VecOp> struct MorphFilter : BaseFilter
-{
-    typedef typename Op::rtype T;
-
-    MorphFilter( const Mat& _kernel, Point _anchor )
-    {
-        anchor = _anchor;
-        ksize = _kernel.size();
-        CV_Assert( _kernel.type() == CV_8U );
-
-        std::vector<uchar> coeffs; // we do not really the values of non-zero
-        // kernel elements, just their locations
-        preprocess2DKernel( _kernel, coords, coeffs );
-        ptrs.resize( coords.size() );
-    }
-
-    void operator()(const uchar** src, uchar* dst, int dststep, int count, int width, int cn) CV_OVERRIDE
-    {
-        const Point* pt = &coords[0];
-        const T** kp = (const T**)&ptrs[0];
-        int i, k, nz = (int)coords.size();
-        Op op;
-
-        width *= cn;
-        for( ; count > 0; count--, dst += dststep, src++ )
-        {
-            T* D = (T*)dst;
-
-            for( k = 0; k < nz; k++ )
-                kp[k] = (const T*)src[pt[k].y] + pt[k].x*cn;
-
-            i = vecOp(&ptrs[0], nz, dst, width);
-            #if CV_ENABLE_UNROLLED
-            for( ; i <= width - 4; i += 4 )
-            {
-                const T* sptr = kp[0] + i;
-                T s0 = sptr[0], s1 = sptr[1], s2 = sptr[2], s3 = sptr[3];
-
-                for( k = 1; k < nz; k++ )
-                {
-                    sptr = kp[k] + i;
-                    s0 = op(s0, sptr[0]); s1 = op(s1, sptr[1]);
-                    s2 = op(s2, sptr[2]); s3 = op(s3, sptr[3]);
-                }
-
-                D[i] = s0; D[i+1] = s1;
-                D[i+2] = s2; D[i+3] = s3;
-            }
-            #endif
-            for( ; i < width; i++ )
-            {
-                T s0 = kp[0][i];
-                for( k = 1; k < nz; k++ )
-                    s0 = op(s0, kp[k][i]);
-                D[i] = s0;
-            }
-        }
-    }
-
-    std::vector<Point> coords;
-    std::vector<uchar*> ptrs;
-    VecOp vecOp;
-};
-
-}
 
 /////////////////////////////////// External Interface /////////////////////////////////////
 
-cv::Ptr<cv::BaseRowFilter> cv::getMorphologyRowFilter(int op, int type, int ksize, int anchor)
+Ptr<BaseRowFilter> getMorphologyRowFilter(int op, int type, int ksize, int anchor)
 {
-    int depth = CV_MAT_DEPTH(type);
+    CV_INSTRUMENT_REGION();
-    if( anchor < 0 )
-        anchor = ksize/2;
-    CV_Assert( op == MORPH_ERODE || op == MORPH_DILATE );
-    if( op == MORPH_ERODE )
-    {
-        if( depth == CV_8U )
-            return makePtr<MorphRowFilter<MinOp<uchar>,
-                                      ErodeRowVec8u> >(ksize, anchor);
-        if( depth == CV_16U )
-            return makePtr<MorphRowFilter<MinOp<ushort>,
-                                      ErodeRowVec16u> >(ksize, anchor);
-        if( depth == CV_16S )
-            return makePtr<MorphRowFilter<MinOp<short>,
-                                      ErodeRowVec16s> >(ksize, anchor);
-        if( depth == CV_32F )
-            return makePtr<MorphRowFilter<MinOp<float>,
-                                      ErodeRowVec32f> >(ksize, anchor);
-        if( depth == CV_64F )
-            return makePtr<MorphRowFilter<MinOp<double>,
-                                      ErodeRowVec64f> >(ksize, anchor);
-    }
-    else
-    {
-        if( depth == CV_8U )
-            return makePtr<MorphRowFilter<MaxOp<uchar>,
-                                      DilateRowVec8u> >(ksize, anchor);
-        if( depth == CV_16U )
-            return makePtr<MorphRowFilter<MaxOp<ushort>,
-                                      DilateRowVec16u> >(ksize, anchor);
-        if( depth == CV_16S )
-            return makePtr<MorphRowFilter<MaxOp<short>,
-                                      DilateRowVec16s> >(ksize, anchor);
-        if( depth == CV_32F )
-            return makePtr<MorphRowFilter<MaxOp<float>,
-                                      DilateRowVec32f> >(ksize, anchor);
-        if( depth == CV_64F )
-            return makePtr<MorphRowFilter<MaxOp<double>,
-                                      DilateRowVec64f> >(ksize, anchor);
-    }
 
-    CV_Error_( CV_StsNotImplemented, ("Unsupported data type (=%d)", type));
+    CV_CPU_DISPATCH(getMorphologyRowFilter, (op, type, ksize, anchor),
+        CV_CPU_DISPATCH_MODES_ALL);
 }
 
-cv::Ptr<cv::BaseColumnFilter> cv::getMorphologyColumnFilter(int op, int type, int ksize, int anchor)
+Ptr<BaseColumnFilter> getMorphologyColumnFilter(int op, int type, int ksize, int anchor)
 {
-    int depth = CV_MAT_DEPTH(type);
+    CV_INSTRUMENT_REGION();
-    if( anchor < 0 )
-        anchor = ksize/2;
-    CV_Assert( op == MORPH_ERODE || op == MORPH_DILATE );
-    if( op == MORPH_ERODE )
-    {
-        if( depth == CV_8U )
-            return makePtr<MorphColumnFilter<MinOp<uchar>,
-                                         ErodeColumnVec8u> >(ksize, anchor);
-        if( depth == CV_16U )
-            return makePtr<MorphColumnFilter<MinOp<ushort>,
-                                         ErodeColumnVec16u> >(ksize, anchor);
-        if( depth == CV_16S )
-            return makePtr<MorphColumnFilter<MinOp<short>,
-                                         ErodeColumnVec16s> >(ksize, anchor);
-        if( depth == CV_32F )
-            return makePtr<MorphColumnFilter<MinOp<float>,
-                                         ErodeColumnVec32f> >(ksize, anchor);
-        if( depth == CV_64F )
-            return makePtr<MorphColumnFilter<MinOp<double>,
-                                         ErodeColumnVec64f> >(ksize, anchor);
-    }
-    else
-    {
-        if( depth == CV_8U )
-            return makePtr<MorphColumnFilter<MaxOp<uchar>,
-                                         DilateColumnVec8u> >(ksize, anchor);
-        if( depth == CV_16U )
-            return makePtr<MorphColumnFilter<MaxOp<ushort>,
-                                         DilateColumnVec16u> >(ksize, anchor);
-        if( depth == CV_16S )
-            return makePtr<MorphColumnFilter<MaxOp<short>,
-                                         DilateColumnVec16s> >(ksize, anchor);
-        if( depth == CV_32F )
-            return makePtr<MorphColumnFilter<MaxOp<float>,
-                                         DilateColumnVec32f> >(ksize, anchor);
-        if( depth == CV_64F )
-            return makePtr<MorphColumnFilter<MaxOp<double>,
-                                         DilateColumnVec64f> >(ksize, anchor);
-    }
 
-    CV_Error_( CV_StsNotImplemented, ("Unsupported data type (=%d)", type));
+    CV_CPU_DISPATCH(getMorphologyColumnFilter, (op, type, ksize, anchor),
+        CV_CPU_DISPATCH_MODES_ALL);
 }
 
 
-cv::Ptr<cv::BaseFilter> cv::getMorphologyFilter(int op, int type, InputArray _kernel, Point anchor)
+Ptr<BaseFilter> getMorphologyFilter(int op, int type, InputArray _kernel, Point anchor)
 {
-    Mat kernel = _kernel.getMat();
+    CV_INSTRUMENT_REGION();
-    int depth = CV_MAT_DEPTH(type);
-    anchor = normalizeAnchor(anchor, kernel.size());
-    CV_Assert( op == MORPH_ERODE || op == MORPH_DILATE );
-    if( op == MORPH_ERODE )
-    {
-        if( depth == CV_8U )
-            return makePtr<MorphFilter<MinOp<uchar>, ErodeVec8u> >(kernel, anchor);
-        if( depth == CV_16U )
-            return makePtr<MorphFilter<MinOp<ushort>, ErodeVec16u> >(kernel, anchor);
-        if( depth == CV_16S )
-            return makePtr<MorphFilter<MinOp<short>, ErodeVec16s> >(kernel, anchor);
-        if( depth == CV_32F )
-            return makePtr<MorphFilter<MinOp<float>, ErodeVec32f> >(kernel, anchor);
-        if( depth == CV_64F )
-            return makePtr<MorphFilter<MinOp<double>, ErodeVec64f> >(kernel, anchor);
-    }
-    else
-    {
-        if( depth == CV_8U )
-            return makePtr<MorphFilter<MaxOp<uchar>, DilateVec8u> >(kernel, anchor);
-        if( depth == CV_16U )
-            return makePtr<MorphFilter<MaxOp<ushort>, DilateVec16u> >(kernel, anchor);
-        if( depth == CV_16S )
-            return makePtr<MorphFilter<MaxOp<short>, DilateVec16s> >(kernel, anchor);
-        if( depth == CV_32F )
-            return makePtr<MorphFilter<MaxOp<float>, DilateVec32f> >(kernel, anchor);
-        if( depth == CV_64F )
-            return makePtr<MorphFilter<MaxOp<double>, DilateVec64f> >(kernel, anchor);
-    }
 
-    CV_Error_( CV_StsNotImplemented, ("Unsupported data type (=%d)", type));
+    Mat kernel = _kernel.getMat();
+    CV_CPU_DISPATCH(getMorphologyFilter, (op, type, kernel, anchor),
+        CV_CPU_DISPATCH_MODES_ALL);
 }
 
 
-cv::Ptr<cv::FilterEngine> cv::createMorphologyFilter( int op, int type, InputArray _kernel,
+Ptr<FilterEngine> createMorphologyFilter(
-                                                      Point anchor, int _rowBorderType, int _columnBorderType,
+        int op, int type, InputArray _kernel,
-                                                      const Scalar& _borderValue )
+        Point anchor, int _rowBorderType, int _columnBorderType,
+        const Scalar& _borderValue)
 {
     Mat kernel = _kernel.getMat();
     anchor = normalizeAnchor(anchor, kernel.size());
@@ -862,7 +132,7 @@ cv::Ptr<cv::FilterEngine> cv::createMorphologyFilter( int op, int type, InputArr
 }
 
 
-cv::Mat cv::getStructuringElement(int shape, Size ksize, Point anchor)
+Mat getStructuringElement(int shape, Size ksize, Point anchor)
 {
     int i, j;
     int r = 0, c = 0;
@@ -915,9 +185,6 @@ cv::Mat cv::getStructuringElement(int shape, Size ksize, Point anchor)
     return elem;
 }
 
-namespace cv
-{
-
 // ===== 1. replacement implementation
 
 static bool halMorph(int op, int src_type, int dst_type,
@@ -1732,9 +999,7 @@ static void morphOp( int op, InputArray _src, OutputArray _dst,
                (src.isSubmatrix() && !isolated));
 }
 
-}
+void erode( InputArray src, OutputArray dst, InputArray kernel,
-
-void cv::erode( InputArray src, OutputArray dst, InputArray kernel,
                 Point anchor, int iterations,
                 int borderType, const Scalar& borderValue )
 {
@@ -1744,7 +1009,7 @@ void cv::erode( InputArray src, OutputArray dst, InputArray kernel,
 }
 
 
-void cv::dilate( InputArray src, OutputArray dst, InputArray kernel,
+void dilate( InputArray src, OutputArray dst, InputArray kernel,
                  Point anchor, int iterations,
                  int borderType, const Scalar& borderValue )
 {
@@ -1755,8 +1020,6 @@ void cv::dilate( InputArray src, OutputArray dst, InputArray kernel,
 
 #ifdef HAVE_OPENCL
 
-namespace cv {
-
 static bool ocl_morphologyEx(InputArray _src, OutputArray _dst, int op,
                              InputArray kernel, Point anchor, int iterations,
                              int borderType, const Scalar& borderValue)
@@ -1813,13 +1076,11 @@ static bool ocl_morphologyEx(InputArray _src, OutputArray _dst, int op,
     return true;
 }
 
-}
 #endif
 
 #define IPP_DISABLE_MORPH_ADV 1
 #ifdef HAVE_IPP
 #if !IPP_DISABLE_MORPH_ADV
-namespace cv {
 static bool ipp_morphologyEx(int op, InputArray _src, OutputArray _dst,
                      InputArray _kernel,
                      Point anchor, int iterations,
@@ -1884,11 +1145,10 @@ static bool ipp_morphologyEx(int op, InputArray _src, OutputArray _dst,
     return false;
 #endif
 }
-}
 #endif
 #endif
 
-void cv::morphologyEx( InputArray _src, OutputArray _dst, int op,
+void morphologyEx( InputArray _src, OutputArray _dst, int op,
                        InputArray _kernel, Point anchor, int iterations,
                        int borderType, const Scalar& borderValue )
 {
@@ -1985,6 +1245,8 @@ void cv::morphologyEx( InputArray _src, OutputArray _dst, int op,
     }
 }
 
+} // namespace cv
+
 CV_IMPL IplConvKernel *
 cvCreateStructuringElementEx( int cols, int rows,
                               int anchorX, int anchorY,

modules/imgproc/src/morph.simd.hpp

@@ -0,0 +1,846 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#include "precomp.hpp"
+#include <limits.h>
+#include "opencv2/core/hal/intrin.hpp"
+
+/****************************************************************************************\
+                     Basic Morphological Operations: Erosion & Dilation
+\****************************************************************************************/
+
+namespace cv {
+CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
+// forward declarations
+Ptr<BaseRowFilter> getMorphologyRowFilter(int op, int type, int ksize, int anchor);
+Ptr<BaseColumnFilter> getMorphologyColumnFilter(int op, int type, int ksize, int anchor);
+Ptr<BaseFilter> getMorphologyFilter(int op, int type, const Mat& kernel, Point anchor);
+
+#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
+
+namespace {
+template<typename T> struct MinOp
+{
+    typedef T type1;
+    typedef T type2;
+    typedef T rtype;
+    T operator ()(const T a, const T b) const { return std::min(a, b); }
+};
+
+template<typename T> struct MaxOp
+{
+    typedef T type1;
+    typedef T type2;
+    typedef T rtype;
+    T operator ()(const T a, const T b) const { return std::max(a, b); }
+};
+
+
+#if !defined(CV_SIMD)  // min/max operation are usually fast enough (without using of control flow 'if' statements)
+
+#undef CV_MIN_8U
+#undef CV_MAX_8U
+#define CV_MIN_8U(a,b)       ((a) - CV_FAST_CAST_8U((a) - (b)))
+#define CV_MAX_8U(a,b)       ((a) + CV_FAST_CAST_8U((b) - (a)))
+
+template<> inline uchar MinOp<uchar>::operator ()(const uchar a, const uchar b) const { return CV_MIN_8U(a, b); }
+template<> inline uchar MaxOp<uchar>::operator ()(const uchar a, const uchar b) const { return CV_MAX_8U(a, b); }
+
+#endif
+
+
+
+struct MorphRowNoVec
+{
+    MorphRowNoVec(int, int) {}
+    int operator()(const uchar*, uchar*, int, int) const { return 0; }
+};
+
+struct MorphColumnNoVec
+{
+    MorphColumnNoVec(int, int) {}
+    int operator()(const uchar**, uchar*, int, int, int) const { return 0; }
+};
+
+struct MorphNoVec
+{
+    int operator()(uchar**, int, uchar*, int) const { return 0; }
+};
+
+#if CV_SIMD
+
+template<class VecUpdate> struct MorphRowVec
+{
+    typedef typename VecUpdate::vtype vtype;
+    typedef typename vtype::lane_type stype;
+    MorphRowVec(int _ksize, int _anchor) : ksize(_ksize), anchor(_anchor) {}
+    int operator()(const uchar* src, uchar* dst, int width, int cn) const
+    {
+        CV_INSTRUMENT_REGION();
+
+        int i, k, _ksize = ksize*cn;
+        width *= cn;
+        VecUpdate updateOp;
+
+        for( i = 0; i <= width - 4*vtype::nlanes; i += 4*vtype::nlanes )
+        {
+            vtype s0 = vx_load((const stype*)src + i);
+            vtype s1 = vx_load((const stype*)src + i + vtype::nlanes);
+            vtype s2 = vx_load((const stype*)src + i + 2*vtype::nlanes);
+            vtype s3 = vx_load((const stype*)src + i + 3*vtype::nlanes);
+            for (k = cn; k < _ksize; k += cn)
+            {
+                s0 = updateOp(s0, vx_load((const stype*)src + i + k));
+                s1 = updateOp(s1, vx_load((const stype*)src + i + k + vtype::nlanes));
+                s2 = updateOp(s2, vx_load((const stype*)src + i + k + 2*vtype::nlanes));
+                s3 = updateOp(s3, vx_load((const stype*)src + i + k + 3*vtype::nlanes));
+            }
+            v_store((stype*)dst + i, s0);
+            v_store((stype*)dst + i + vtype::nlanes, s1);
+            v_store((stype*)dst + i + 2*vtype::nlanes, s2);
+            v_store((stype*)dst + i + 3*vtype::nlanes, s3);
+        }
+        if( i <= width - 2*vtype::nlanes )
+        {
+            vtype s0 = vx_load((const stype*)src + i);
+            vtype s1 = vx_load((const stype*)src + i + vtype::nlanes);
+            for( k = cn; k < _ksize; k += cn )
+            {
+                s0 = updateOp(s0, vx_load((const stype*)src + i + k));
+                s1 = updateOp(s1, vx_load((const stype*)src + i + k + vtype::nlanes));
+            }
+            v_store((stype*)dst + i, s0);
+            v_store((stype*)dst + i + vtype::nlanes, s1);
+            i += 2*vtype::nlanes;
+        }
+        if( i <= width - vtype::nlanes )
+        {
+            vtype s = vx_load((const stype*)src + i);
+            for( k = cn; k < _ksize; k += cn )
+                s = updateOp(s, vx_load((const stype*)src + i + k));
+            v_store((stype*)dst + i, s);
+            i += vtype::nlanes;
+        }
+        if( i <= width - vtype::nlanes/2 )
+        {
+            vtype s = vx_load_low((const stype*)src + i);
+            for( k = cn; k < _ksize; k += cn )
+                s = updateOp(s, vx_load_low((const stype*)src + i + k));
+            v_store_low((stype*)dst + i, s);
+            i += vtype::nlanes/2;
+        }
+
+        return i - i % cn;
+    }
+
+    int ksize, anchor;
+};
+
+
+template<class VecUpdate> struct MorphColumnVec
+{
+    typedef typename VecUpdate::vtype vtype;
+    typedef typename vtype::lane_type stype;
+    MorphColumnVec(int _ksize, int _anchor) : ksize(_ksize), anchor(_anchor) {}
+    int operator()(const uchar** _src, uchar* _dst, int dststep, int count, int width) const
+    {
+        CV_INSTRUMENT_REGION();
+
+        int i = 0, k, _ksize = ksize;
+        VecUpdate updateOp;
+
+        for( i = 0; i < count + ksize - 1; i++ )
+            CV_Assert( ((size_t)_src[i] & (CV_SIMD_WIDTH-1)) == 0 );
+
+        const stype** src = (const stype**)_src;
+        stype* dst = (stype*)_dst;
+        dststep /= sizeof(dst[0]);
+
+        for( ; _ksize > 1 && count > 1; count -= 2, dst += dststep*2, src += 2 )
+        {
+            for( i = 0; i <= width - 4*vtype::nlanes; i += 4*vtype::nlanes)
+            {
+                const stype* sptr = src[1] + i;
+                vtype s0 = vx_load_aligned(sptr);
+                vtype s1 = vx_load_aligned(sptr + vtype::nlanes);
+                vtype s2 = vx_load_aligned(sptr + 2*vtype::nlanes);
+                vtype s3 = vx_load_aligned(sptr + 3*vtype::nlanes);
+
+                for( k = 2; k < _ksize; k++ )
+                {
+                    sptr = src[k] + i;
+                    s0 = updateOp(s0, vx_load_aligned(sptr));
+                    s1 = updateOp(s1, vx_load_aligned(sptr + vtype::nlanes));
+                    s2 = updateOp(s2, vx_load_aligned(sptr + 2*vtype::nlanes));
+                    s3 = updateOp(s3, vx_load_aligned(sptr + 3*vtype::nlanes));
+                }
+
+                sptr = src[0] + i;
+                v_store(dst + i, updateOp(s0, vx_load_aligned(sptr)));
+                v_store(dst + i + vtype::nlanes, updateOp(s1, vx_load_aligned(sptr + vtype::nlanes)));
+                v_store(dst + i + 2*vtype::nlanes, updateOp(s2, vx_load_aligned(sptr + 2*vtype::nlanes)));
+                v_store(dst + i + 3*vtype::nlanes, updateOp(s3, vx_load_aligned(sptr + 3*vtype::nlanes)));
+
+                sptr = src[k] + i;
+                v_store(dst + dststep + i, updateOp(s0, vx_load_aligned(sptr)));
+                v_store(dst + dststep + i + vtype::nlanes, updateOp(s1, vx_load_aligned(sptr + vtype::nlanes)));
+                v_store(dst + dststep + i + 2*vtype::nlanes, updateOp(s2, vx_load_aligned(sptr + 2*vtype::nlanes)));
+                v_store(dst + dststep + i + 3*vtype::nlanes, updateOp(s3, vx_load_aligned(sptr + 3*vtype::nlanes)));
+            }
+            if( i <= width - 2*vtype::nlanes )
+            {
+                const stype* sptr = src[1] + i;
+                vtype s0 = vx_load_aligned(sptr);
+                vtype s1 = vx_load_aligned(sptr + vtype::nlanes);
+
+                for( k = 2; k < _ksize; k++ )
+                {
+                    sptr = src[k] + i;
+                    s0 = updateOp(s0, vx_load_aligned(sptr));
+                    s1 = updateOp(s1, vx_load_aligned(sptr + vtype::nlanes));
+                }
+
+                sptr = src[0] + i;
+                v_store(dst + i, updateOp(s0, vx_load_aligned(sptr)));
+                v_store(dst + i + vtype::nlanes, updateOp(s1, vx_load_aligned(sptr + vtype::nlanes)));
+
+                sptr = src[k] + i;
+                v_store(dst + dststep + i, updateOp(s0, vx_load_aligned(sptr)));
+                v_store(dst + dststep + i + vtype::nlanes, updateOp(s1, vx_load_aligned(sptr + vtype::nlanes)));
+                i += 2*vtype::nlanes;
+            }
+            if( i <= width - vtype::nlanes )
+            {
+                vtype s0 = vx_load_aligned(src[1] + i);
+
+                for( k = 2; k < _ksize; k++ )
+                    s0 = updateOp(s0, vx_load_aligned(src[k] + i));
+
+                v_store(dst + i, updateOp(s0, vx_load_aligned(src[0] + i)));
+                v_store(dst + dststep + i, updateOp(s0, vx_load_aligned(src[k] + i)));
+                i += vtype::nlanes;
+            }
+            if( i <= width - vtype::nlanes/2 )
+            {
+                vtype s0 = vx_load_low(src[1] + i);
+
+                for( k = 2; k < _ksize; k++ )
+                    s0 = updateOp(s0, vx_load_low(src[k] + i));
+
+                v_store_low(dst + i, updateOp(s0, vx_load_low(src[0] + i)));
+                v_store_low(dst + dststep + i, updateOp(s0, vx_load_low(src[k] + i)));
+                i += vtype::nlanes/2;
+            }
+        }
+
+        for( ; count > 0; count--, dst += dststep, src++ )
+        {
+            for( i = 0; i <= width - 4*vtype::nlanes; i += 4*vtype::nlanes)
+            {
+                const stype* sptr = src[0] + i;
+                vtype s0 = vx_load_aligned(sptr);
+                vtype s1 = vx_load_aligned(sptr + vtype::nlanes);
+                vtype s2 = vx_load_aligned(sptr + 2*vtype::nlanes);
+                vtype s3 = vx_load_aligned(sptr + 3*vtype::nlanes);
+
+                for( k = 1; k < _ksize; k++ )
+                {
+                    sptr = src[k] + i;
+                    s0 = updateOp(s0, vx_load_aligned(sptr));
+                    s1 = updateOp(s1, vx_load_aligned(sptr + vtype::nlanes));
+                    s2 = updateOp(s2, vx_load_aligned(sptr + 2*vtype::nlanes));
+                    s3 = updateOp(s3, vx_load_aligned(sptr + 3*vtype::nlanes));
+                }
+                v_store(dst + i, s0);
+                v_store(dst + i + vtype::nlanes, s1);
+                v_store(dst + i + 2*vtype::nlanes, s2);
+                v_store(dst + i + 3*vtype::nlanes, s3);
+            }
+            if( i <= width - 2*vtype::nlanes )
+            {
+                const stype* sptr = src[0] + i;
+                vtype s0 = vx_load_aligned(sptr);
+                vtype s1 = vx_load_aligned(sptr + vtype::nlanes);
+
+                for( k = 1; k < _ksize; k++ )
+                {
+                    sptr = src[k] + i;
+                    s0 = updateOp(s0, vx_load_aligned(sptr));
+                    s1 = updateOp(s1, vx_load_aligned(sptr + vtype::nlanes));
+                }
+                v_store(dst + i, s0);
+                v_store(dst + i + vtype::nlanes, s1);
+                i += 2*vtype::nlanes;
+            }
+            if( i <= width - vtype::nlanes )
+            {
+                vtype s0 = vx_load_aligned(src[0] + i);
+
+                for( k = 1; k < _ksize; k++ )
+                    s0 = updateOp(s0, vx_load_aligned(src[k] + i));
+                v_store(dst + i, s0);
+                i += vtype::nlanes;
+            }
+            if( i <= width - vtype::nlanes/2 )
+            {
+                vtype s0 = vx_load_low(src[0] + i);
+
+                for( k = 1; k < _ksize; k++ )
+                    s0 = updateOp(s0, vx_load_low(src[k] + i));
+                v_store_low(dst + i, s0);
+                i += vtype::nlanes/2;
+            }
+        }
+
+        return i;
+    }
+
+    int ksize, anchor;
+};
+
+
+template<class VecUpdate> struct MorphVec
+{
+    typedef typename VecUpdate::vtype vtype;
+    typedef typename vtype::lane_type stype;
+    int operator()(uchar** _src, int nz, uchar* _dst, int width) const
+    {
+        CV_INSTRUMENT_REGION();
+
+        const stype** src = (const stype**)_src;
+        stype* dst = (stype*)_dst;
+        int i, k;
+        VecUpdate updateOp;
+
+        for( i = 0; i <= width - 4*vtype::nlanes; i += 4*vtype::nlanes )
+        {
+            const stype* sptr = src[0] + i;
+            vtype s0 = vx_load(sptr);
+            vtype s1 = vx_load(sptr + vtype::nlanes);
+            vtype s2 = vx_load(sptr + 2*vtype::nlanes);
+            vtype s3 = vx_load(sptr + 3*vtype::nlanes);
+            for( k = 1; k < nz; k++ )
+            {
+                sptr = src[k] + i;
+                s0 = updateOp(s0, vx_load(sptr));
+                s1 = updateOp(s1, vx_load(sptr + vtype::nlanes));
+                s2 = updateOp(s2, vx_load(sptr + 2*vtype::nlanes));
+                s3 = updateOp(s3, vx_load(sptr + 3*vtype::nlanes));
+            }
+            v_store(dst + i, s0);
+            v_store(dst + i + vtype::nlanes, s1);
+            v_store(dst + i + 2*vtype::nlanes, s2);
+            v_store(dst + i + 3*vtype::nlanes, s3);
+        }
+        if( i <= width - 2*vtype::nlanes )
+        {
+            const stype* sptr = src[0] + i;
+            vtype s0 = vx_load(sptr);
+            vtype s1 = vx_load(sptr + vtype::nlanes);
+            for( k = 1; k < nz; k++ )
+            {
+                sptr = src[k] + i;
+                s0 = updateOp(s0, vx_load(sptr));
+                s1 = updateOp(s1, vx_load(sptr + vtype::nlanes));
+            }
+            v_store(dst + i, s0);
+            v_store(dst + i + vtype::nlanes, s1);
+            i += 2*vtype::nlanes;
+        }
+        if( i <= width - vtype::nlanes )
+        {
+            vtype s0 = vx_load(src[0] + i);
+            for( k = 1; k < nz; k++ )
+                s0 = updateOp(s0, vx_load(src[k] + i));
+            v_store(dst + i, s0);
+            i += vtype::nlanes;
+        }
+        if( i <= width - vtype::nlanes/2 )
+        {
+            vtype s0 = vx_load_low(src[0] + i);
+            for( k = 1; k < nz; k++ )
+                s0 = updateOp(s0, vx_load_low(src[k] + i));
+            v_store_low(dst + i, s0);
+            i += vtype::nlanes/2;
+        }
+        return i;
+    }
+};
+
+template <typename T> struct VMin
+{
+    typedef T vtype;
+    vtype operator()(const vtype& a, const vtype& b) const { return v_min(a,b); }
+};
+template <typename T> struct VMax
+{
+    typedef T vtype;
+    vtype operator()(const vtype& a, const vtype& b) const { return v_max(a,b); }
+};
+
+typedef MorphRowVec<VMin<v_uint8> > ErodeRowVec8u;
+typedef MorphRowVec<VMax<v_uint8> > DilateRowVec8u;
+typedef MorphRowVec<VMin<v_uint16> > ErodeRowVec16u;
+typedef MorphRowVec<VMax<v_uint16> > DilateRowVec16u;
+typedef MorphRowVec<VMin<v_int16> > ErodeRowVec16s;
+typedef MorphRowVec<VMax<v_int16> > DilateRowVec16s;
+typedef MorphRowVec<VMin<v_float32> > ErodeRowVec32f;
+typedef MorphRowVec<VMax<v_float32> > DilateRowVec32f;
+
+typedef MorphColumnVec<VMin<v_uint8> > ErodeColumnVec8u;
+typedef MorphColumnVec<VMax<v_uint8> > DilateColumnVec8u;
+typedef MorphColumnVec<VMin<v_uint16> > ErodeColumnVec16u;
+typedef MorphColumnVec<VMax<v_uint16> > DilateColumnVec16u;
+typedef MorphColumnVec<VMin<v_int16> > ErodeColumnVec16s;
+typedef MorphColumnVec<VMax<v_int16> > DilateColumnVec16s;
+typedef MorphColumnVec<VMin<v_float32> > ErodeColumnVec32f;
+typedef MorphColumnVec<VMax<v_float32> > DilateColumnVec32f;
+
+typedef MorphVec<VMin<v_uint8> > ErodeVec8u;
+typedef MorphVec<VMax<v_uint8> > DilateVec8u;
+typedef MorphVec<VMin<v_uint16> > ErodeVec16u;
+typedef MorphVec<VMax<v_uint16> > DilateVec16u;
+typedef MorphVec<VMin<v_int16> > ErodeVec16s;
+typedef MorphVec<VMax<v_int16> > DilateVec16s;
+typedef MorphVec<VMin<v_float32> > ErodeVec32f;
+typedef MorphVec<VMax<v_float32> > DilateVec32f;
+
+#else
+
+typedef MorphRowNoVec ErodeRowVec8u;
+typedef MorphRowNoVec DilateRowVec8u;
+
+typedef MorphColumnNoVec ErodeColumnVec8u;
+typedef MorphColumnNoVec DilateColumnVec8u;
+
+typedef MorphRowNoVec ErodeRowVec16u;
+typedef MorphRowNoVec DilateRowVec16u;
+typedef MorphRowNoVec ErodeRowVec16s;
+typedef MorphRowNoVec DilateRowVec16s;
+typedef MorphRowNoVec ErodeRowVec32f;
+typedef MorphRowNoVec DilateRowVec32f;
+
+typedef MorphColumnNoVec ErodeColumnVec16u;
+typedef MorphColumnNoVec DilateColumnVec16u;
+typedef MorphColumnNoVec ErodeColumnVec16s;
+typedef MorphColumnNoVec DilateColumnVec16s;
+typedef MorphColumnNoVec ErodeColumnVec32f;
+typedef MorphColumnNoVec DilateColumnVec32f;
+
+typedef MorphNoVec ErodeVec8u;
+typedef MorphNoVec DilateVec8u;
+typedef MorphNoVec ErodeVec16u;
+typedef MorphNoVec DilateVec16u;
+typedef MorphNoVec ErodeVec16s;
+typedef MorphNoVec DilateVec16s;
+typedef MorphNoVec ErodeVec32f;
+typedef MorphNoVec DilateVec32f;
+
+#endif
+
+typedef MorphRowNoVec ErodeRowVec64f;
+typedef MorphRowNoVec DilateRowVec64f;
+typedef MorphColumnNoVec ErodeColumnVec64f;
+typedef MorphColumnNoVec DilateColumnVec64f;
+typedef MorphNoVec ErodeVec64f;
+typedef MorphNoVec DilateVec64f;
+
+
+template<class Op, class VecOp> struct MorphRowFilter : public BaseRowFilter
+{
+    typedef typename Op::rtype T;
+
+    MorphRowFilter( int _ksize, int _anchor ) : vecOp(_ksize, _anchor)
+    {
+        ksize = _ksize;
+        anchor = _anchor;
+    }
+
+    void operator()(const uchar* src, uchar* dst, int width, int cn) CV_OVERRIDE
+    {
+        CV_INSTRUMENT_REGION();
+
+        int i, j, k, _ksize = ksize*cn;
+        const T* S = (const T*)src;
+        Op op;
+        T* D = (T*)dst;
+
+        if( _ksize == cn )
+        {
+            for( i = 0; i < width*cn; i++ )
+                D[i] = S[i];
+            return;
+        }
+
+        int i0 = vecOp(src, dst, width, cn);
+        width *= cn;
+
+        for( k = 0; k < cn; k++, S++, D++ )
+        {
+            for( i = i0; i <= width - cn*2; i += cn*2 )
+            {
+                const T* s = S + i;
+                T m = s[cn];
+                for( j = cn*2; j < _ksize; j += cn )
+                    m = op(m, s[j]);
+                D[i] = op(m, s[0]);
+                D[i+cn] = op(m, s[j]);
+            }
+
+            for( ; i < width; i += cn )
+            {
+                const T* s = S + i;
+                T m = s[0];
+                for( j = cn; j < _ksize; j += cn )
+                    m = op(m, s[j]);
+                D[i] = m;
+            }
+        }
+    }
+
+    VecOp vecOp;
+};
+
+
+template<class Op, class VecOp> struct MorphColumnFilter : public BaseColumnFilter
+{
+    typedef typename Op::rtype T;
+
+    MorphColumnFilter( int _ksize, int _anchor ) : vecOp(_ksize, _anchor)
+    {
+        ksize = _ksize;
+        anchor = _anchor;
+    }
+
+    void operator()(const uchar** _src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE
+    {
+        CV_INSTRUMENT_REGION();
+
+        int i, k, _ksize = ksize;
+        const T** src = (const T**)_src;
+        T* D = (T*)dst;
+        Op op;
+
+        int i0 = vecOp(_src, dst, dststep, count, width);
+        dststep /= sizeof(D[0]);
+
+        for( ; _ksize > 1 && count > 1; count -= 2, D += dststep*2, src += 2 )
+        {
+            i = i0;
+            #if CV_ENABLE_UNROLLED
+            for( ; i <= width - 4; i += 4 )
+            {
+                const T* sptr = src[1] + i;
+                T s0 = sptr[0], s1 = sptr[1], s2 = sptr[2], s3 = sptr[3];
+
+                for( k = 2; k < _ksize; k++ )
+                {
+                    sptr = src[k] + i;
+                    s0 = op(s0, sptr[0]); s1 = op(s1, sptr[1]);
+                    s2 = op(s2, sptr[2]); s3 = op(s3, sptr[3]);
+                }
+
+                sptr = src[0] + i;
+                D[i] = op(s0, sptr[0]);
+                D[i+1] = op(s1, sptr[1]);
+                D[i+2] = op(s2, sptr[2]);
+                D[i+3] = op(s3, sptr[3]);
+
+                sptr = src[k] + i;
+                D[i+dststep] = op(s0, sptr[0]);
+                D[i+dststep+1] = op(s1, sptr[1]);
+                D[i+dststep+2] = op(s2, sptr[2]);
+                D[i+dststep+3] = op(s3, sptr[3]);
+            }
+            #endif
+            for( ; i < width; i++ )
+            {
+                T s0 = src[1][i];
+
+                for( k = 2; k < _ksize; k++ )
+                    s0 = op(s0, src[k][i]);
+
+                D[i] = op(s0, src[0][i]);
+                D[i+dststep] = op(s0, src[k][i]);
+            }
+        }
+
+        for( ; count > 0; count--, D += dststep, src++ )
+        {
+            i = i0;
+            #if CV_ENABLE_UNROLLED
+            for( ; i <= width - 4; i += 4 )
+            {
+                const T* sptr = src[0] + i;
+                T s0 = sptr[0], s1 = sptr[1], s2 = sptr[2], s3 = sptr[3];
+
+                for( k = 1; k < _ksize; k++ )
+                {
+                    sptr = src[k] + i;
+                    s0 = op(s0, sptr[0]); s1 = op(s1, sptr[1]);
+                    s2 = op(s2, sptr[2]); s3 = op(s3, sptr[3]);
+                }
+
+                D[i] = s0; D[i+1] = s1;
+                D[i+2] = s2; D[i+3] = s3;
+            }
+            #endif
+            for( ; i < width; i++ )
+            {
+                T s0 = src[0][i];
+                for( k = 1; k < _ksize; k++ )
+                    s0 = op(s0, src[k][i]);
+                D[i] = s0;
+            }
+        }
+    }
+
+    VecOp vecOp;
+};
+
+
+template<class Op, class VecOp> struct MorphFilter : BaseFilter
+{
+    typedef typename Op::rtype T;
+
+    MorphFilter( const Mat& _kernel, Point _anchor )
+    {
+        anchor = _anchor;
+        ksize = _kernel.size();
+        CV_Assert( _kernel.type() == CV_8U );
+
+        std::vector<uchar> coeffs; // we do not really the values of non-zero
+        // kernel elements, just their locations
+        preprocess2DKernel( _kernel, coords, coeffs );
+        ptrs.resize( coords.size() );
+    }
+
+    void operator()(const uchar** src, uchar* dst, int dststep, int count, int width, int cn) CV_OVERRIDE
+    {
+        CV_INSTRUMENT_REGION();
+
+        const Point* pt = &coords[0];
+        const T** kp = (const T**)&ptrs[0];
+        int i, k, nz = (int)coords.size();
+        Op op;
+
+        width *= cn;
+        for( ; count > 0; count--, dst += dststep, src++ )
+        {
+            T* D = (T*)dst;
+
+            for( k = 0; k < nz; k++ )
+                kp[k] = (const T*)src[pt[k].y] + pt[k].x*cn;
+
+            i = vecOp(&ptrs[0], nz, dst, width);
+            #if CV_ENABLE_UNROLLED
+            for( ; i <= width - 4; i += 4 )
+            {
+                const T* sptr = kp[0] + i;
+                T s0 = sptr[0], s1 = sptr[1], s2 = sptr[2], s3 = sptr[3];
+
+                for( k = 1; k < nz; k++ )
+                {
+                    sptr = kp[k] + i;
+                    s0 = op(s0, sptr[0]); s1 = op(s1, sptr[1]);
+                    s2 = op(s2, sptr[2]); s3 = op(s3, sptr[3]);
+                }
+
+                D[i] = s0; D[i+1] = s1;
+                D[i+2] = s2; D[i+3] = s3;
+            }
+            #endif
+            for( ; i < width; i++ )
+            {
+                T s0 = kp[0][i];
+                for( k = 1; k < nz; k++ )
+                    s0 = op(s0, kp[k][i]);
+                D[i] = s0;
+            }
+        }
+    }
+
+    std::vector<Point> coords;
+    std::vector<uchar*> ptrs;
+    VecOp vecOp;
+};
+
+} // namespace anon
+
+/////////////////////////////////// External Interface /////////////////////////////////////
+
+Ptr<BaseRowFilter> getMorphologyRowFilter(int op, int type, int ksize, int anchor)
+{
+    CV_INSTRUMENT_REGION();
+
+    int depth = CV_MAT_DEPTH(type);
+    if( anchor < 0 )
+        anchor = ksize/2;
+    CV_Assert( op == MORPH_ERODE || op == MORPH_DILATE );
+    if( op == MORPH_ERODE )
+    {
+        if( depth == CV_8U )
+            return makePtr<MorphRowFilter<MinOp<uchar>,
+                                      ErodeRowVec8u> >(ksize, anchor);
+        if( depth == CV_16U )
+            return makePtr<MorphRowFilter<MinOp<ushort>,
+                                      ErodeRowVec16u> >(ksize, anchor);
+        if( depth == CV_16S )
+            return makePtr<MorphRowFilter<MinOp<short>,
+                                      ErodeRowVec16s> >(ksize, anchor);
+        if( depth == CV_32F )
+            return makePtr<MorphRowFilter<MinOp<float>,
+                                      ErodeRowVec32f> >(ksize, anchor);
+        if( depth == CV_64F )
+            return makePtr<MorphRowFilter<MinOp<double>,
+                                      ErodeRowVec64f> >(ksize, anchor);
+    }
+    else
+    {
+        if( depth == CV_8U )
+            return makePtr<MorphRowFilter<MaxOp<uchar>,
+                                      DilateRowVec8u> >(ksize, anchor);
+        if( depth == CV_16U )
+            return makePtr<MorphRowFilter<MaxOp<ushort>,
+                                      DilateRowVec16u> >(ksize, anchor);
+        if( depth == CV_16S )
+            return makePtr<MorphRowFilter<MaxOp<short>,
+                                      DilateRowVec16s> >(ksize, anchor);
+        if( depth == CV_32F )
+            return makePtr<MorphRowFilter<MaxOp<float>,
+                                      DilateRowVec32f> >(ksize, anchor);
+        if( depth == CV_64F )
+            return makePtr<MorphRowFilter<MaxOp<double>,
+                                      DilateRowVec64f> >(ksize, anchor);
+    }
+
+    CV_Error_( CV_StsNotImplemented, ("Unsupported data type (=%d)", type));
+}
+
+Ptr<BaseColumnFilter> getMorphologyColumnFilter(int op, int type, int ksize, int anchor)
+{
+    CV_INSTRUMENT_REGION();
+
+    int depth = CV_MAT_DEPTH(type);
+    if( anchor < 0 )
+        anchor = ksize/2;
+    CV_Assert( op == MORPH_ERODE || op == MORPH_DILATE );
+    if( op == MORPH_ERODE )
+    {
+        if( depth == CV_8U )
+            return makePtr<MorphColumnFilter<MinOp<uchar>,
+                                         ErodeColumnVec8u> >(ksize, anchor);
+        if( depth == CV_16U )
+            return makePtr<MorphColumnFilter<MinOp<ushort>,
+                                         ErodeColumnVec16u> >(ksize, anchor);
+        if( depth == CV_16S )
+            return makePtr<MorphColumnFilter<MinOp<short>,
+                                         ErodeColumnVec16s> >(ksize, anchor);
+        if( depth == CV_32F )
+            return makePtr<MorphColumnFilter<MinOp<float>,
+                                         ErodeColumnVec32f> >(ksize, anchor);
+        if( depth == CV_64F )
+            return makePtr<MorphColumnFilter<MinOp<double>,
+                                         ErodeColumnVec64f> >(ksize, anchor);
+    }
+    else
+    {
+        if( depth == CV_8U )
+            return makePtr<MorphColumnFilter<MaxOp<uchar>,
+                                         DilateColumnVec8u> >(ksize, anchor);
+        if( depth == CV_16U )
+            return makePtr<MorphColumnFilter<MaxOp<ushort>,
+                                         DilateColumnVec16u> >(ksize, anchor);
+        if( depth == CV_16S )
+            return makePtr<MorphColumnFilter<MaxOp<short>,
+                                         DilateColumnVec16s> >(ksize, anchor);
+        if( depth == CV_32F )
+            return makePtr<MorphColumnFilter<MaxOp<float>,
+                                         DilateColumnVec32f> >(ksize, anchor);
+        if( depth == CV_64F )
+            return makePtr<MorphColumnFilter<MaxOp<double>,
+                                         DilateColumnVec64f> >(ksize, anchor);
+    }
+
+    CV_Error_( CV_StsNotImplemented, ("Unsupported data type (=%d)", type));
+}
+
+Ptr<BaseFilter> getMorphologyFilter(int op, int type, const Mat& kernel, Point anchor)
+{
+    CV_INSTRUMENT_REGION();
+
+    int depth = CV_MAT_DEPTH(type);
+    anchor = normalizeAnchor(anchor, kernel.size());
+    CV_Assert( op == MORPH_ERODE || op == MORPH_DILATE );
+    if( op == MORPH_ERODE )
+    {
+        if( depth == CV_8U )
+            return makePtr<MorphFilter<MinOp<uchar>, ErodeVec8u> >(kernel, anchor);
+        if( depth == CV_16U )
+            return makePtr<MorphFilter<MinOp<ushort>, ErodeVec16u> >(kernel, anchor);
+        if( depth == CV_16S )
+            return makePtr<MorphFilter<MinOp<short>, ErodeVec16s> >(kernel, anchor);
+        if( depth == CV_32F )
+            return makePtr<MorphFilter<MinOp<float>, ErodeVec32f> >(kernel, anchor);
+        if( depth == CV_64F )
+            return makePtr<MorphFilter<MinOp<double>, ErodeVec64f> >(kernel, anchor);
+    }
+    else
+    {
+        if( depth == CV_8U )
+            return makePtr<MorphFilter<MaxOp<uchar>, DilateVec8u> >(kernel, anchor);
+        if( depth == CV_16U )
+            return makePtr<MorphFilter<MaxOp<ushort>, DilateVec16u> >(kernel, anchor);
+        if( depth == CV_16S )
+            return makePtr<MorphFilter<MaxOp<short>, DilateVec16s> >(kernel, anchor);
+        if( depth == CV_32F )
+            return makePtr<MorphFilter<MaxOp<float>, DilateVec32f> >(kernel, anchor);
+        if( depth == CV_64F )
+            return makePtr<MorphFilter<MaxOp<double>, DilateVec64f> >(kernel, anchor);
+    }
+
+    CV_Error_( CV_StsNotImplemented, ("Unsupported data type (=%d)", type));
+}
+
+#endif
+CV_CPU_OPTIMIZATION_NAMESPACE_END
+} // namespace

modules/imgproc/src/smooth.dispatch.cpp

@@ -0,0 +1,582 @@
+/*M///////////////////////////////////////////////////////////////////////////////////////
+//
+//  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
+//
+//  By downloading, copying, installing or using the software you agree to this license.
+//  If you do not agree to this license, do not download, install,
+//  copy or use the software.
+//
+//
+//                           License Agreement
+//                For Open Source Computer Vision Library
+//
+// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
+// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
+// Copyright (C) 2014-2015, Itseez Inc., all rights reserved.
+// Third party copyrights are property of their respective owners.
+//
+// Redistribution and use in source and binary forms, with or without modification,
+// are permitted provided that the following conditions are met:
+//
+//   * Redistribution's of source code must retain the above copyright notice,
+//     this list of conditions and the following disclaimer.
+//
+//   * Redistribution's in binary form must reproduce the above copyright notice,
+//     this list of conditions and the following disclaimer in the documentation
+//     and/or other materials provided with the distribution.
+//
+//   * The name of the copyright holders may not be used to endorse or promote products
+//     derived from this software without specific prior written permission.
+//
+// This software is provided by the copyright holders and contributors "as is" and
+// any express or implied warranties, including, but not limited to, the implied
+// warranties of merchantability and fitness for a particular purpose are disclaimed.
+// In no event shall the Intel Corporation or contributors be liable for any direct,
+// indirect, incidental, special, exemplary, or consequential damages
+// (including, but not limited to, procurement of substitute goods or services;
+// loss of use, data, or profits; or business interruption) however caused
+// and on any theory of liability, whether in contract, strict liability,
+// or tort (including negligence or otherwise) arising in any way out of
+// the use of this software, even if advised of the possibility of such damage.
+//
+//M*/
+
+#include "precomp.hpp"
+
+#include <vector>
+
+#include "opencv2/core/hal/intrin.hpp"
+#include "opencl_kernels_imgproc.hpp"
+
+#include "opencv2/core/openvx/ovx_defs.hpp"
+
+#include "filter.hpp"
+
+#include "opencv2/core/softfloat.hpp"
+
+namespace cv {
+#include "fixedpoint.inl.hpp"
+}
+
+#include "smooth.simd.hpp"
+#include "smooth.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
+
+namespace cv {
+
+/****************************************************************************************\
+                                     Gaussian Blur
+\****************************************************************************************/
+
+Mat getGaussianKernel(int n, double sigma, int ktype)
+{
+    CV_Assert(n > 0);
+    const int SMALL_GAUSSIAN_SIZE = 7;
+    static const float small_gaussian_tab[][SMALL_GAUSSIAN_SIZE] =
+    {
+        {1.f},
+        {0.25f, 0.5f, 0.25f},
+        {0.0625f, 0.25f, 0.375f, 0.25f, 0.0625f},
+        {0.03125f, 0.109375f, 0.21875f, 0.28125f, 0.21875f, 0.109375f, 0.03125f}
+    };
+
+    const float* fixed_kernel = n % 2 == 1 && n <= SMALL_GAUSSIAN_SIZE && sigma <= 0 ?
+        small_gaussian_tab[n>>1] : 0;
+
+    CV_Assert( ktype == CV_32F || ktype == CV_64F );
+    Mat kernel(n, 1, ktype);
+    float* cf = kernel.ptr<float>();
+    double* cd = kernel.ptr<double>();
+
+    double sigmaX = sigma > 0 ? sigma : ((n-1)*0.5 - 1)*0.3 + 0.8;
+    double scale2X = -0.5/(sigmaX*sigmaX);
+    double sum = 0;
+
+    int i;
+    for( i = 0; i < n; i++ )
+    {
+        double x = i - (n-1)*0.5;
+        double t = fixed_kernel ? (double)fixed_kernel[i] : std::exp(scale2X*x*x);
+        if( ktype == CV_32F )
+        {
+            cf[i] = (float)t;
+            sum += cf[i];
+        }
+        else
+        {
+            cd[i] = t;
+            sum += cd[i];
+        }
+    }
+
+    CV_DbgAssert(fabs(sum) > 0);
+    sum = 1./sum;
+    for( i = 0; i < n; i++ )
+    {
+        if( ktype == CV_32F )
+            cf[i] = (float)(cf[i]*sum);
+        else
+            cd[i] *= sum;
+    }
+
+    return kernel;
+}
+
+template <typename T>
+static std::vector<T> getFixedpointGaussianKernel( int n, double sigma )
+{
+    if (sigma <= 0)
+    {
+        if(n == 1)
+            return std::vector<T>(1, softdouble(1.0));
+        else if(n == 3)
+        {
+            T v3[] = { softdouble(0.25), softdouble(0.5), softdouble(0.25) };
+            return std::vector<T>(v3, v3 + 3);
+        }
+        else if(n == 5)
+        {
+            T v5[] = { softdouble(0.0625), softdouble(0.25), softdouble(0.375), softdouble(0.25), softdouble(0.0625) };
+            return std::vector<T>(v5, v5 + 5);
+        }
+        else if(n == 7)
+        {
+            T v7[] = { softdouble(0.03125), softdouble(0.109375), softdouble(0.21875), softdouble(0.28125), softdouble(0.21875), softdouble(0.109375), softdouble(0.03125) };
+            return std::vector<T>(v7, v7 + 7);
+        }
+    }
+
+
+    softdouble sigmaX = sigma > 0 ? softdouble(sigma) : mulAdd(softdouble(n),softdouble(0.15),softdouble(0.35));// softdouble(((n-1)*0.5 - 1)*0.3 + 0.8)
+    softdouble scale2X = softdouble(-0.5*0.25)/(sigmaX*sigmaX);
+    std::vector<softdouble> values(n);
+    softdouble sum(0.);
+    for(int i = 0, x = 1 - n; i < n; i++, x+=2 )
+    {
+        // x = i - (n - 1)*0.5
+        // t = std::exp(scale2X*x*x)
+        values[i] = exp(softdouble(x*x)*scale2X);
+        sum += values[i];
+    }
+    sum = softdouble::one()/sum;
+
+    std::vector<T> kernel(n);
+    for(int i = 0; i < n; i++ )
+    {
+        kernel[i] = values[i] * sum;
+    }
+
+    return kernel;
+};
+
+static void getGaussianKernel(int n, double sigma, int ktype, Mat& res) { res = getGaussianKernel(n, sigma, ktype); }
+template <typename T> static void getGaussianKernel(int n, double sigma, int, std::vector<T>& res) { res = getFixedpointGaussianKernel<T>(n, sigma); }
+
+template <typename T>
+static void createGaussianKernels( T & kx, T & ky, int type, Size &ksize,
+                                   double sigma1, double sigma2 )
+{
+    int depth = CV_MAT_DEPTH(type);
+    if( sigma2 <= 0 )
+        sigma2 = sigma1;
+
+    // automatic detection of kernel size from sigma
+    if( ksize.width <= 0 && sigma1 > 0 )
+        ksize.width = cvRound(sigma1*(depth == CV_8U ? 3 : 4)*2 + 1)|1;
+    if( ksize.height <= 0 && sigma2 > 0 )
+        ksize.height = cvRound(sigma2*(depth == CV_8U ? 3 : 4)*2 + 1)|1;
+
+    CV_Assert( ksize.width  > 0 && ksize.width  % 2 == 1 &&
+               ksize.height > 0 && ksize.height % 2 == 1 );
+
+    sigma1 = std::max( sigma1, 0. );
+    sigma2 = std::max( sigma2, 0. );
+
+    getGaussianKernel( ksize.width, sigma1, std::max(depth, CV_32F), kx );
+    if( ksize.height == ksize.width && std::abs(sigma1 - sigma2) < DBL_EPSILON )
+        ky = kx;
+    else
+        getGaussianKernel( ksize.height, sigma2, std::max(depth, CV_32F), ky );
+}
+
+Ptr<FilterEngine> createGaussianFilter( int type, Size ksize,
+                                        double sigma1, double sigma2,
+                                        int borderType )
+{
+    Mat kx, ky;
+    createGaussianKernels(kx, ky, type, ksize, sigma1, sigma2);
+
+    return createSeparableLinearFilter( type, type, kx, ky, Point(-1,-1), 0, borderType );
+}
+
+#ifdef HAVE_OPENCL
+
+static bool ocl_GaussianBlur_8UC1(InputArray _src, OutputArray _dst, Size ksize, int ddepth,
+                                  InputArray _kernelX, InputArray _kernelY, int borderType)
+{
+    const ocl::Device & dev = ocl::Device::getDefault();
+    int type = _src.type(), sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+
+    if ( !(dev.isIntel() && (type == CV_8UC1) &&
+         (_src.offset() == 0) && (_src.step() % 4 == 0) &&
+         ((ksize.width == 5 && (_src.cols() % 4 == 0)) ||
+         (ksize.width == 3 && (_src.cols() % 16 == 0) && (_src.rows() % 2 == 0)))) )
+        return false;
+
+    Mat kernelX = _kernelX.getMat().reshape(1, 1);
+    if (kernelX.cols % 2 != 1)
+        return false;
+    Mat kernelY = _kernelY.getMat().reshape(1, 1);
+    if (kernelY.cols % 2 != 1)
+        return false;
+
+    if (ddepth < 0)
+        ddepth = sdepth;
+
+    Size size = _src.size();
+    size_t globalsize[2] = { 0, 0 };
+    size_t localsize[2] = { 0, 0 };
+
+    if (ksize.width == 3)
+    {
+        globalsize[0] = size.width / 16;
+        globalsize[1] = size.height / 2;
+    }
+    else if (ksize.width == 5)
+    {
+        globalsize[0] = size.width / 4;
+        globalsize[1] = size.height / 1;
+    }
+
+    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", 0, "BORDER_REFLECT_101" };
+    char build_opts[1024];
+    sprintf(build_opts, "-D %s %s%s", borderMap[borderType & ~BORDER_ISOLATED],
+            ocl::kernelToStr(kernelX, CV_32F, "KERNEL_MATRIX_X").c_str(),
+            ocl::kernelToStr(kernelY, CV_32F, "KERNEL_MATRIX_Y").c_str());
+
+    ocl::Kernel kernel;
+
+    if (ksize.width == 3)
+        kernel.create("gaussianBlur3x3_8UC1_cols16_rows2", cv::ocl::imgproc::gaussianBlur3x3_oclsrc, build_opts);
+    else if (ksize.width == 5)
+        kernel.create("gaussianBlur5x5_8UC1_cols4", cv::ocl::imgproc::gaussianBlur5x5_oclsrc, build_opts);
+
+    if (kernel.empty())
+        return false;
+
+    UMat src = _src.getUMat();
+    _dst.create(size, CV_MAKETYPE(ddepth, cn));
+    if (!(_dst.offset() == 0 && _dst.step() % 4 == 0))
+        return false;
+    UMat dst = _dst.getUMat();
+
+    int idxArg = kernel.set(0, ocl::KernelArg::PtrReadOnly(src));
+    idxArg = kernel.set(idxArg, (int)src.step);
+    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dst));
+    idxArg = kernel.set(idxArg, (int)dst.step);
+    idxArg = kernel.set(idxArg, (int)dst.rows);
+    idxArg = kernel.set(idxArg, (int)dst.cols);
+
+    return kernel.run(2, globalsize, (localsize[0] == 0) ? NULL : localsize, false);
+}
+
+#endif
+
+#ifdef HAVE_OPENVX
+
+namespace ovx {
+    template <> inline bool skipSmallImages<VX_KERNEL_GAUSSIAN_3x3>(int w, int h) { return w*h < 320 * 240; }
+}
+static bool openvx_gaussianBlur(InputArray _src, OutputArray _dst, Size ksize,
+                                double sigma1, double sigma2, int borderType)
+{
+    if (sigma2 <= 0)
+        sigma2 = sigma1;
+    // automatic detection of kernel size from sigma
+    if (ksize.width <= 0 && sigma1 > 0)
+        ksize.width = cvRound(sigma1*6 + 1) | 1;
+    if (ksize.height <= 0 && sigma2 > 0)
+        ksize.height = cvRound(sigma2*6 + 1) | 1;
+
+    if (_src.type() != CV_8UC1 ||
+        _src.cols() < 3 || _src.rows() < 3 ||
+        ksize.width != 3 || ksize.height != 3)
+        return false;
+
+    sigma1 = std::max(sigma1, 0.);
+    sigma2 = std::max(sigma2, 0.);
+
+    if (!(sigma1 == 0.0 || (sigma1 - 0.8) < DBL_EPSILON) || !(sigma2 == 0.0 || (sigma2 - 0.8) < DBL_EPSILON) ||
+        ovx::skipSmallImages<VX_KERNEL_GAUSSIAN_3x3>(_src.cols(), _src.rows()))
+        return false;
+
+    Mat src = _src.getMat();
+    Mat dst = _dst.getMat();
+
+    if ((borderType & BORDER_ISOLATED) == 0 && src.isSubmatrix())
+        return false; //Process isolated borders only
+    vx_enum border;
+    switch (borderType & ~BORDER_ISOLATED)
+    {
+    case BORDER_CONSTANT:
+        border = VX_BORDER_CONSTANT;
+        break;
+    case BORDER_REPLICATE:
+        border = VX_BORDER_REPLICATE;
+        break;
+    default:
+        return false;
+    }
+
+    try
+    {
+        ivx::Context ctx = ovx::getOpenVXContext();
+
+        Mat a;
+        if (dst.data != src.data)
+            a = src;
+        else
+            src.copyTo(a);
+
+        ivx::Image
+            ia = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
+                ivx::Image::createAddressing(a.cols, a.rows, 1, (vx_int32)(a.step)), a.data),
+            ib = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
+                ivx::Image::createAddressing(dst.cols, dst.rows, 1, (vx_int32)(dst.step)), dst.data);
+
+        //ATTENTION: VX_CONTEXT_IMMEDIATE_BORDER attribute change could lead to strange issues in multi-threaded environments
+        //since OpenVX standard says nothing about thread-safety for now
+        ivx::border_t prevBorder = ctx.immediateBorder();
+        ctx.setImmediateBorder(border, (vx_uint8)(0));
+        ivx::IVX_CHECK_STATUS(vxuGaussian3x3(ctx, ia, ib));
+        ctx.setImmediateBorder(prevBorder);
+    }
+    catch (const ivx::RuntimeError & e)
+    {
+        VX_DbgThrow(e.what());
+    }
+    catch (const ivx::WrapperError & e)
+    {
+        VX_DbgThrow(e.what());
+    }
+    return true;
+}
+
+#endif
+
+#ifdef HAVE_IPP
+// IW 2017u2 has bug which doesn't allow use of partial inMem with tiling
+#if IPP_DISABLE_GAUSSIANBLUR_PARALLEL
+#define IPP_GAUSSIANBLUR_PARALLEL 0
+#else
+#define IPP_GAUSSIANBLUR_PARALLEL 1
+#endif
+
+#ifdef HAVE_IPP_IW
+
+class ipp_gaussianBlurParallel: public ParallelLoopBody
+{
+public:
+    ipp_gaussianBlurParallel(::ipp::IwiImage &src, ::ipp::IwiImage &dst, int kernelSize, float sigma, ::ipp::IwiBorderType &border, bool *pOk):
+        m_src(src), m_dst(dst), m_kernelSize(kernelSize), m_sigma(sigma), m_border(border), m_pOk(pOk) {
+        *m_pOk = true;
+    }
+    ~ipp_gaussianBlurParallel()
+    {
+    }
+
+    virtual void operator() (const Range& range) const CV_OVERRIDE
+    {
+        CV_INSTRUMENT_REGION_IPP();
+
+        if(!*m_pOk)
+            return;
+
+        try
+        {
+            ::ipp::IwiTile tile = ::ipp::IwiRoi(0, range.start, m_dst.m_size.width, range.end - range.start);
+            CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterGaussian, m_src, m_dst, m_kernelSize, m_sigma, ::ipp::IwDefault(), m_border, tile);
+        }
+        catch(const ::ipp::IwException &)
+        {
+            *m_pOk = false;
+            return;
+        }
+    }
+private:
+    ::ipp::IwiImage &m_src;
+    ::ipp::IwiImage &m_dst;
+
+    int m_kernelSize;
+    float m_sigma;
+    ::ipp::IwiBorderType &m_border;
+
+    volatile bool *m_pOk;
+    const ipp_gaussianBlurParallel& operator= (const ipp_gaussianBlurParallel&);
+};
+
+#endif
+
+static bool ipp_GaussianBlur(InputArray _src, OutputArray _dst, Size ksize,
+                   double sigma1, double sigma2, int borderType )
+{
+#ifdef HAVE_IPP_IW
+    CV_INSTRUMENT_REGION_IPP();
+
+#if IPP_VERSION_X100 < 201800 && ((defined _MSC_VER && defined _M_IX86) || (defined __GNUC__ && defined __i386__))
+    CV_UNUSED(_src); CV_UNUSED(_dst); CV_UNUSED(ksize); CV_UNUSED(sigma1); CV_UNUSED(sigma2); CV_UNUSED(borderType);
+    return false; // bug on ia32
+#else
+    if(sigma1 != sigma2)
+        return false;
+
+    if(sigma1 < FLT_EPSILON)
+        return false;
+
+    if(ksize.width != ksize.height)
+        return false;
+
+    // Acquire data and begin processing
+    try
+    {
+        Mat src = _src.getMat();
+        Mat dst = _dst.getMat();
+        ::ipp::IwiImage       iwSrc      = ippiGetImage(src);
+        ::ipp::IwiImage       iwDst      = ippiGetImage(dst);
+        ::ipp::IwiBorderSize  borderSize = ::ipp::iwiSizeToBorderSize(ippiGetSize(ksize));
+        ::ipp::IwiBorderType  ippBorder(ippiGetBorder(iwSrc, borderType, borderSize));
+        if(!ippBorder)
+            return false;
+
+        const int threads = ippiSuggestThreadsNum(iwDst, 2);
+        if(IPP_GAUSSIANBLUR_PARALLEL && threads > 1) {
+            bool ok;
+            ipp_gaussianBlurParallel invoker(iwSrc, iwDst, ksize.width, (float) sigma1, ippBorder, &ok);
+
+            if(!ok)
+                return false;
+            const Range range(0, (int) iwDst.m_size.height);
+            parallel_for_(range, invoker, threads*4);
+
+            if(!ok)
+                return false;
+        } else {
+            CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterGaussian, iwSrc, iwDst, ksize.width, sigma1, ::ipp::IwDefault(), ippBorder);
+        }
+    }
+    catch (const ::ipp::IwException &)
+    {
+        return false;
+    }
+
+    return true;
+#endif
+#else
+    CV_UNUSED(_src); CV_UNUSED(_dst); CV_UNUSED(ksize); CV_UNUSED(sigma1); CV_UNUSED(sigma2); CV_UNUSED(borderType);
+    return false;
+#endif
+}
+#endif
+
+void GaussianBlur(InputArray _src, OutputArray _dst, Size ksize,
+                  double sigma1, double sigma2,
+                  int borderType)
+{
+    CV_INSTRUMENT_REGION();
+
+    int type = _src.type();
+    Size size = _src.size();
+    _dst.create( size, type );
+
+    if( (borderType & ~BORDER_ISOLATED) != BORDER_CONSTANT &&
+        ((borderType & BORDER_ISOLATED) != 0 || !_src.getMat().isSubmatrix()) )
+    {
+        if( size.height == 1 )
+            ksize.height = 1;
+        if( size.width == 1 )
+            ksize.width = 1;
+    }
+
+    if( ksize.width == 1 && ksize.height == 1 )
+    {
+        _src.copyTo(_dst);
+        return;
+    }
+
+    bool useOpenCL = (ocl::isOpenCLActivated() && _dst.isUMat() && _src.dims() <= 2 &&
+               ((ksize.width == 3 && ksize.height == 3) ||
+               (ksize.width == 5 && ksize.height == 5)) &&
+               _src.rows() > ksize.height && _src.cols() > ksize.width);
+    CV_UNUSED(useOpenCL);
+
+    int sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+
+    Mat kx, ky;
+    createGaussianKernels(kx, ky, type, ksize, sigma1, sigma2);
+
+    CV_OCL_RUN(useOpenCL, ocl_GaussianBlur_8UC1(_src, _dst, ksize, CV_MAT_DEPTH(type), kx, ky, borderType));
+
+    CV_OCL_RUN(_dst.isUMat() && _src.dims() <= 2 && (size_t)_src.rows() > kx.total() && (size_t)_src.cols() > kx.total(),
+               ocl_sepFilter2D(_src, _dst, sdepth, kx, ky, Point(-1, -1), 0, borderType))
+
+    Mat src = _src.getMat();
+    Mat dst = _dst.getMat();
+
+    Point ofs;
+    Size wsz(src.cols, src.rows);
+    if(!(borderType & BORDER_ISOLATED))
+        src.locateROI( wsz, ofs );
+
+    CALL_HAL(gaussianBlur, cv_hal_gaussianBlur, src.ptr(), src.step, dst.ptr(), dst.step, src.cols, src.rows, sdepth, cn,
+             ofs.x, ofs.y, wsz.width - src.cols - ofs.x, wsz.height - src.rows - ofs.y, ksize.width, ksize.height,
+             sigma1, sigma2, borderType&~BORDER_ISOLATED);
+
+    CV_OVX_RUN(true,
+               openvx_gaussianBlur(src, dst, ksize, sigma1, sigma2, borderType))
+
+    CV_IPP_RUN_FAST(ipp_GaussianBlur(src, dst, ksize, sigma1, sigma2, borderType));
+
+    if(sdepth == CV_8U && ((borderType & BORDER_ISOLATED) || !_src.getMat().isSubmatrix()))
+    {
+        std::vector<ufixedpoint16> fkx, fky;
+        createGaussianKernels(fkx, fky, type, ksize, sigma1, sigma2);
+        if (src.data == dst.data)
+            src = src.clone();
+        CV_CPU_DISPATCH(GaussianBlurFixedPoint, (src, dst, (const uint16_t*)&fkx[0], (int)fkx.size(), (const uint16_t*)&fky[0], (int)fky.size(), borderType),
+            CV_CPU_DISPATCH_MODES_ALL);
+        return;
+    }
+
+    sepFilter2D(src, dst, sdepth, kx, ky, Point(-1, -1), 0, borderType);
+}
+
+} // namespace
+
+//////////////////////////////////////////////////////////////////////////////////////////
+
+CV_IMPL void
+cvSmooth( const void* srcarr, void* dstarr, int smooth_type,
+          int param1, int param2, double param3, double param4 )
+{
+    cv::Mat src = cv::cvarrToMat(srcarr), dst0 = cv::cvarrToMat(dstarr), dst = dst0;
+
+    CV_Assert( dst.size() == src.size() &&
+        (smooth_type == CV_BLUR_NO_SCALE || dst.type() == src.type()) );
+
+    if( param2 <= 0 )
+        param2 = param1;
+
+    if( smooth_type == CV_BLUR || smooth_type == CV_BLUR_NO_SCALE )
+        cv::boxFilter( src, dst, dst.depth(), cv::Size(param1, param2), cv::Point(-1,-1),
+            smooth_type == CV_BLUR, cv::BORDER_REPLICATE );
+    else if( smooth_type == CV_GAUSSIAN )
+        cv::GaussianBlur( src, dst, cv::Size(param1, param2), param3, param4, cv::BORDER_REPLICATE );
+    else if( smooth_type == CV_MEDIAN )
+        cv::medianBlur( src, dst, param1 );
+    else
+        cv::bilateralFilter( src, dst, param1, param3, param4, cv::BORDER_REPLICATE );
+
+    if( dst.data != dst0.data )
+        CV_Error( CV_StsUnmatchedFormats, "The destination image does not have the proper type" );
+}
+
+/* End of file. */

modules/imgproc/src/smooth.simd.hpp

@@ -46,120 +46,28 @@
 #include <vector>
 
 #include "opencv2/core/hal/intrin.hpp"
-#include "opencl_kernels_imgproc.hpp"
-
-#include "opencv2/core/openvx/ovx_defs.hpp"
 
 #include "filter.hpp"
 
-#include "fixedpoint.inl.hpp"
+#include "opencv2/core/softfloat.hpp"
-
-/****************************************************************************************\
-                                     Gaussian Blur
-\****************************************************************************************/
-
-cv::Mat cv::getGaussianKernel( int n, double sigma, int ktype )
-{
-    CV_Assert(n > 0);
-    const int SMALL_GAUSSIAN_SIZE = 7;
-    static const float small_gaussian_tab[][SMALL_GAUSSIAN_SIZE] =
-    {
-        {1.f},
-        {0.25f, 0.5f, 0.25f},
-        {0.0625f, 0.25f, 0.375f, 0.25f, 0.0625f},
-        {0.03125f, 0.109375f, 0.21875f, 0.28125f, 0.21875f, 0.109375f, 0.03125f}
-    };
-
-    const float* fixed_kernel = n % 2 == 1 && n <= SMALL_GAUSSIAN_SIZE && sigma <= 0 ?
-        small_gaussian_tab[n>>1] : 0;
-
-    CV_Assert( ktype == CV_32F || ktype == CV_64F );
-    Mat kernel(n, 1, ktype);
-    float* cf = kernel.ptr<float>();
-    double* cd = kernel.ptr<double>();
-
-    double sigmaX = sigma > 0 ? sigma : ((n-1)*0.5 - 1)*0.3 + 0.8;
-    double scale2X = -0.5/(sigmaX*sigmaX);
-    double sum = 0;
-
-    int i;
-    for( i = 0; i < n; i++ )
-    {
-        double x = i - (n-1)*0.5;
-        double t = fixed_kernel ? (double)fixed_kernel[i] : std::exp(scale2X*x*x);
-        if( ktype == CV_32F )
-        {
-            cf[i] = (float)t;
-            sum += cf[i];
-        }
-        else
-        {
-            cd[i] = t;
-            sum += cd[i];
-        }
-    }
-
-    CV_DbgAssert(fabs(sum) > 0);
-    sum = 1./sum;
-    for( i = 0; i < n; i++ )
-    {
-        if( ktype == CV_32F )
-            cf[i] = (float)(cf[i]*sum);
-        else
-            cd[i] *= sum;
-    }
-
-    return kernel;
-}
 
 namespace cv {
+CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
+// forward declarations
+void GaussianBlurFixedPoint(const Mat& src, /*const*/ Mat& dst,
+                            const uint16_t/*ufixedpoint16*/* fkx, int fkx_size,
+                            const uint16_t/*ufixedpoint16*/* fky, int fky_size,
+                            int borderType);
 
-template <typename T>
+#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
-static std::vector<T> getFixedpointGaussianKernel( int n, double sigma )
-{
-    if (sigma <= 0)
-    {
-        if(n == 1)
-            return std::vector<T>(1, softdouble(1.0));
-        else if(n == 3)
-        {
-            T v3[] = { softdouble(0.25), softdouble(0.5), softdouble(0.25) };
-            return std::vector<T>(v3, v3 + 3);
-        }
-        else if(n == 5)
-        {
-            T v5[] = { softdouble(0.0625), softdouble(0.25), softdouble(0.375), softdouble(0.25), softdouble(0.0625) };
-            return std::vector<T>(v5, v5 + 5);
-        }
-        else if(n == 7)
-        {
-            T v7[] = { softdouble(0.03125), softdouble(0.109375), softdouble(0.21875), softdouble(0.28125), softdouble(0.21875), softdouble(0.109375), softdouble(0.03125) };
-            return std::vector<T>(v7, v7 + 7);
-        }
-    }
-
-
-    softdouble sigmaX = sigma > 0 ? softdouble(sigma) : mulAdd(softdouble(n),softdouble(0.15),softdouble(0.35));// softdouble(((n-1)*0.5 - 1)*0.3 + 0.8)
-    softdouble scale2X = softdouble(-0.5*0.25)/(sigmaX*sigmaX);
-    std::vector<softdouble> values(n);
-    softdouble sum(0.);
-    for(int i = 0, x = 1 - n; i < n; i++, x+=2 )
-    {
-        // x = i - (n - 1)*0.5
-        // t = std::exp(scale2X*x*x)
-        values[i] = exp(softdouble(x*x)*scale2X);
-        sum += values[i];
-    }
-    sum = softdouble::one()/sum;
 
-    std::vector<T> kernel(n);
+#if defined(CV_CPU_BASELINE_MODE)
-    for(int i = 0; i < n; i++ )
+// included in dispatch.cpp
-    {
+#else
-        kernel[i] = values[i] * sum;
+#include "fixedpoint.inl.hpp"
-    }
+#endif
 
-    return kernel;
+namespace {
-};
 
 template <typename ET, typename FT>
 void hlineSmooth1N(const ET* src, int cn, const FT* m, int, FT* dst, int len, int)
@@ -2119,418 +2027,27 @@ private:
     fixedSmoothInvoker& operator=(const fixedSmoothInvoker&);
 };
 
-static void getGaussianKernel(int n, double sigma, int ktype, Mat& res) { res = getGaussianKernel(n, sigma, ktype); }
+}  // namespace anon
-template <typename T> static void getGaussianKernel(int n, double sigma, int, std::vector<T>& res) { res = getFixedpointGaussianKernel<T>(n, sigma); }
-
-template <typename T>
-static void createGaussianKernels( T & kx, T & ky, int type, Size &ksize,
-                                   double sigma1, double sigma2 )
-{
-    int depth = CV_MAT_DEPTH(type);
-    if( sigma2 <= 0 )
-        sigma2 = sigma1;
-
-    // automatic detection of kernel size from sigma
-    if( ksize.width <= 0 && sigma1 > 0 )
-        ksize.width = cvRound(sigma1*(depth == CV_8U ? 3 : 4)*2 + 1)|1;
-    if( ksize.height <= 0 && sigma2 > 0 )
-        ksize.height = cvRound(sigma2*(depth == CV_8U ? 3 : 4)*2 + 1)|1;
-
-    CV_Assert( ksize.width  > 0 && ksize.width  % 2 == 1 &&
-               ksize.height > 0 && ksize.height % 2 == 1 );
-
-    sigma1 = std::max( sigma1, 0. );
-    sigma2 = std::max( sigma2, 0. );
 
-    getGaussianKernel( ksize.width, sigma1, std::max(depth, CV_32F), kx );
+void GaussianBlurFixedPoint(const Mat& src, /*const*/ Mat& dst,
-    if( ksize.height == ksize.width && std::abs(sigma1 - sigma2) < DBL_EPSILON )
+                            const uint16_t/*ufixedpoint16*/* fkx, int fkx_size,
-        ky = kx;
+                            const uint16_t/*ufixedpoint16*/* fky, int fky_size,
-    else
+                            int borderType)
-        getGaussianKernel( ksize.height, sigma2, std::max(depth, CV_32F), ky );
-}
-
-}
-
-cv::Ptr<cv::FilterEngine> cv::createGaussianFilter( int type, Size ksize,
-                                        double sigma1, double sigma2,
-                                        int borderType )
-{
-    Mat kx, ky;
-    createGaussianKernels(kx, ky, type, ksize, sigma1, sigma2);
-
-    return createSeparableLinearFilter( type, type, kx, ky, Point(-1,-1), 0, borderType );
-}
-
-namespace cv
-{
-#ifdef HAVE_OPENCL
-
-static bool ocl_GaussianBlur_8UC1(InputArray _src, OutputArray _dst, Size ksize, int ddepth,
-                                  InputArray _kernelX, InputArray _kernelY, int borderType)
-{
-    const ocl::Device & dev = ocl::Device::getDefault();
-    int type = _src.type(), sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
-
-    if ( !(dev.isIntel() && (type == CV_8UC1) &&
-         (_src.offset() == 0) && (_src.step() % 4 == 0) &&
-         ((ksize.width == 5 && (_src.cols() % 4 == 0)) ||
-         (ksize.width == 3 && (_src.cols() % 16 == 0) && (_src.rows() % 2 == 0)))) )
-        return false;
-
-    Mat kernelX = _kernelX.getMat().reshape(1, 1);
-    if (kernelX.cols % 2 != 1)
-        return false;
-    Mat kernelY = _kernelY.getMat().reshape(1, 1);
-    if (kernelY.cols % 2 != 1)
-        return false;
-
-    if (ddepth < 0)
-        ddepth = sdepth;
-
-    Size size = _src.size();
-    size_t globalsize[2] = { 0, 0 };
-    size_t localsize[2] = { 0, 0 };
-
-    if (ksize.width == 3)
-    {
-        globalsize[0] = size.width / 16;
-        globalsize[1] = size.height / 2;
-    }
-    else if (ksize.width == 5)
-    {
-        globalsize[0] = size.width / 4;
-        globalsize[1] = size.height / 1;
-    }
-
-    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", 0, "BORDER_REFLECT_101" };
-    char build_opts[1024];
-    sprintf(build_opts, "-D %s %s%s", borderMap[borderType & ~BORDER_ISOLATED],
-            ocl::kernelToStr(kernelX, CV_32F, "KERNEL_MATRIX_X").c_str(),
-            ocl::kernelToStr(kernelY, CV_32F, "KERNEL_MATRIX_Y").c_str());
-
-    ocl::Kernel kernel;
-
-    if (ksize.width == 3)
-        kernel.create("gaussianBlur3x3_8UC1_cols16_rows2", cv::ocl::imgproc::gaussianBlur3x3_oclsrc, build_opts);
-    else if (ksize.width == 5)
-        kernel.create("gaussianBlur5x5_8UC1_cols4", cv::ocl::imgproc::gaussianBlur5x5_oclsrc, build_opts);
-
-    if (kernel.empty())
-        return false;
-
-    UMat src = _src.getUMat();
-    _dst.create(size, CV_MAKETYPE(ddepth, cn));
-    if (!(_dst.offset() == 0 && _dst.step() % 4 == 0))
-        return false;
-    UMat dst = _dst.getUMat();
-
-    int idxArg = kernel.set(0, ocl::KernelArg::PtrReadOnly(src));
-    idxArg = kernel.set(idxArg, (int)src.step);
-    idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dst));
-    idxArg = kernel.set(idxArg, (int)dst.step);
-    idxArg = kernel.set(idxArg, (int)dst.rows);
-    idxArg = kernel.set(idxArg, (int)dst.cols);
-
-    return kernel.run(2, globalsize, (localsize[0] == 0) ? NULL : localsize, false);
-}
-
-#endif
-
-#ifdef HAVE_OPENVX
-
-namespace ovx {
-    template <> inline bool skipSmallImages<VX_KERNEL_GAUSSIAN_3x3>(int w, int h) { return w*h < 320 * 240; }
-}
-static bool openvx_gaussianBlur(InputArray _src, OutputArray _dst, Size ksize,
-                                double sigma1, double sigma2, int borderType)
-{
-    if (sigma2 <= 0)
-        sigma2 = sigma1;
-    // automatic detection of kernel size from sigma
-    if (ksize.width <= 0 && sigma1 > 0)
-        ksize.width = cvRound(sigma1*6 + 1) | 1;
-    if (ksize.height <= 0 && sigma2 > 0)
-        ksize.height = cvRound(sigma2*6 + 1) | 1;
-
-    if (_src.type() != CV_8UC1 ||
-        _src.cols() < 3 || _src.rows() < 3 ||
-        ksize.width != 3 || ksize.height != 3)
-        return false;
-
-    sigma1 = std::max(sigma1, 0.);
-    sigma2 = std::max(sigma2, 0.);
-
-    if (!(sigma1 == 0.0 || (sigma1 - 0.8) < DBL_EPSILON) || !(sigma2 == 0.0 || (sigma2 - 0.8) < DBL_EPSILON) ||
-        ovx::skipSmallImages<VX_KERNEL_GAUSSIAN_3x3>(_src.cols(), _src.rows()))
-        return false;
-
-    Mat src = _src.getMat();
-    Mat dst = _dst.getMat();
-
-    if ((borderType & BORDER_ISOLATED) == 0 && src.isSubmatrix())
-        return false; //Process isolated borders only
-    vx_enum border;
-    switch (borderType & ~BORDER_ISOLATED)
-    {
-    case BORDER_CONSTANT:
-        border = VX_BORDER_CONSTANT;
-        break;
-    case BORDER_REPLICATE:
-        border = VX_BORDER_REPLICATE;
-        break;
-    default:
-        return false;
-    }
-
-    try
-    {
-        ivx::Context ctx = ovx::getOpenVXContext();
-
-        Mat a;
-        if (dst.data != src.data)
-            a = src;
-        else
-            src.copyTo(a);
-
-        ivx::Image
-            ia = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
-                ivx::Image::createAddressing(a.cols, a.rows, 1, (vx_int32)(a.step)), a.data),
-            ib = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,
-                ivx::Image::createAddressing(dst.cols, dst.rows, 1, (vx_int32)(dst.step)), dst.data);
-
-        //ATTENTION: VX_CONTEXT_IMMEDIATE_BORDER attribute change could lead to strange issues in multi-threaded environments
-        //since OpenVX standard says nothing about thread-safety for now
-        ivx::border_t prevBorder = ctx.immediateBorder();
-        ctx.setImmediateBorder(border, (vx_uint8)(0));
-        ivx::IVX_CHECK_STATUS(vxuGaussian3x3(ctx, ia, ib));
-        ctx.setImmediateBorder(prevBorder);
-    }
-    catch (const ivx::RuntimeError & e)
-    {
-        VX_DbgThrow(e.what());
-    }
-    catch (const ivx::WrapperError & e)
-    {
-        VX_DbgThrow(e.what());
-    }
-    return true;
-}
-
-#endif
-
-#ifdef HAVE_IPP
-// IW 2017u2 has bug which doesn't allow use of partial inMem with tiling
-#if IPP_DISABLE_GAUSSIANBLUR_PARALLEL
-#define IPP_GAUSSIANBLUR_PARALLEL 0
-#else
-#define IPP_GAUSSIANBLUR_PARALLEL 1
-#endif
-
-#ifdef HAVE_IPP_IW
-
-class ipp_gaussianBlurParallel: public ParallelLoopBody
-{
-public:
-    ipp_gaussianBlurParallel(::ipp::IwiImage &src, ::ipp::IwiImage &dst, int kernelSize, float sigma, ::ipp::IwiBorderType &border, bool *pOk):
-        m_src(src), m_dst(dst), m_kernelSize(kernelSize), m_sigma(sigma), m_border(border), m_pOk(pOk) {
-        *m_pOk = true;
-    }
-    ~ipp_gaussianBlurParallel()
-    {
-    }
-
-    virtual void operator() (const Range& range) const CV_OVERRIDE
-    {
-        CV_INSTRUMENT_REGION_IPP();
-
-        if(!*m_pOk)
-            return;
-
-        try
-        {
-            ::ipp::IwiTile tile = ::ipp::IwiRoi(0, range.start, m_dst.m_size.width, range.end - range.start);
-            CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterGaussian, m_src, m_dst, m_kernelSize, m_sigma, ::ipp::IwDefault(), m_border, tile);
-        }
-        catch(const ::ipp::IwException &)
-        {
-            *m_pOk = false;
-            return;
-        }
-    }
-private:
-    ::ipp::IwiImage &m_src;
-    ::ipp::IwiImage &m_dst;
-
-    int m_kernelSize;
-    float m_sigma;
-    ::ipp::IwiBorderType &m_border;
-
-    volatile bool *m_pOk;
-    const ipp_gaussianBlurParallel& operator= (const ipp_gaussianBlurParallel&);
-};
-
-#endif
-
-static bool ipp_GaussianBlur(InputArray _src, OutputArray _dst, Size ksize,
-                   double sigma1, double sigma2, int borderType )
-{
-#ifdef HAVE_IPP_IW
-    CV_INSTRUMENT_REGION_IPP();
-
-#if IPP_VERSION_X100 < 201800 && ((defined _MSC_VER && defined _M_IX86) || (defined __GNUC__ && defined __i386__))
-    CV_UNUSED(_src); CV_UNUSED(_dst); CV_UNUSED(ksize); CV_UNUSED(sigma1); CV_UNUSED(sigma2); CV_UNUSED(borderType);
-    return false; // bug on ia32
-#else
-    if(sigma1 != sigma2)
-        return false;
-
-    if(sigma1 < FLT_EPSILON)
-        return false;
-
-    if(ksize.width != ksize.height)
-        return false;
-
-    // Acquire data and begin processing
-    try
-    {
-        Mat src = _src.getMat();
-        Mat dst = _dst.getMat();
-        ::ipp::IwiImage       iwSrc      = ippiGetImage(src);
-        ::ipp::IwiImage       iwDst      = ippiGetImage(dst);
-        ::ipp::IwiBorderSize  borderSize = ::ipp::iwiSizeToBorderSize(ippiGetSize(ksize));
-        ::ipp::IwiBorderType  ippBorder(ippiGetBorder(iwSrc, borderType, borderSize));
-        if(!ippBorder)
-            return false;
-
-        const int threads = ippiSuggestThreadsNum(iwDst, 2);
-        if(IPP_GAUSSIANBLUR_PARALLEL && threads > 1) {
-            bool ok;
-            ipp_gaussianBlurParallel invoker(iwSrc, iwDst, ksize.width, (float) sigma1, ippBorder, &ok);
-
-            if(!ok)
-                return false;
-            const Range range(0, (int) iwDst.m_size.height);
-            parallel_for_(range, invoker, threads*4);
-
-            if(!ok)
-                return false;
-        } else {
-            CV_INSTRUMENT_FUN_IPP(::ipp::iwiFilterGaussian, iwSrc, iwDst, ksize.width, sigma1, ::ipp::IwDefault(), ippBorder);
-        }
-    }
-    catch (const ::ipp::IwException &)
-    {
-        return false;
-    }
-
-    return true;
-#endif
-#else
-    CV_UNUSED(_src); CV_UNUSED(_dst); CV_UNUSED(ksize); CV_UNUSED(sigma1); CV_UNUSED(sigma2); CV_UNUSED(borderType);
-    return false;
-#endif
-}
-#endif
-}
-
-void cv::GaussianBlur( InputArray _src, OutputArray _dst, Size ksize,
-                   double sigma1, double sigma2,
-                   int borderType )
 {
     CV_INSTRUMENT_REGION();
 
-    int type = _src.type();
+    CV_Assert(src.depth() == CV_8U && ((borderType & BORDER_ISOLATED) || !src.isSubmatrix()));
-    Size size = _src.size();
+    fixedSmoothInvoker<uint8_t, ufixedpoint16> invoker(
-    _dst.create( size, type );
+            src.ptr<uint8_t>(), src.step1(),
-
+            dst.ptr<uint8_t>(), dst.step1(), dst.cols, dst.rows, dst.channels(),
-    if( (borderType & ~BORDER_ISOLATED) != BORDER_CONSTANT &&
+            (const ufixedpoint16*)fkx, fkx_size, (const ufixedpoint16*)fky, fky_size,
-        ((borderType & BORDER_ISOLATED) != 0 || !_src.getMat().isSubmatrix()) )
+            borderType & ~BORDER_ISOLATED);
     {
-        if( size.height == 1 )
+        // TODO AVX guard (external call)
-            ksize.height = 1;
-        if( size.width == 1 )
-            ksize.width = 1;
-    }
-
-    if( ksize.width == 1 && ksize.height == 1 )
-    {
-        _src.copyTo(_dst);
-        return;
-    }
-
-    bool useOpenCL = (ocl::isOpenCLActivated() && _dst.isUMat() && _src.dims() <= 2 &&
-               ((ksize.width == 3 && ksize.height == 3) ||
-               (ksize.width == 5 && ksize.height == 5)) &&
-               _src.rows() > ksize.height && _src.cols() > ksize.width);
-    CV_UNUSED(useOpenCL);
-
-    int sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
-
-    Mat kx, ky;
-    createGaussianKernels(kx, ky, type, ksize, sigma1, sigma2);
-
-    CV_OCL_RUN(useOpenCL, ocl_GaussianBlur_8UC1(_src, _dst, ksize, CV_MAT_DEPTH(type), kx, ky, borderType));
-
-    CV_OCL_RUN(_dst.isUMat() && _src.dims() <= 2 && (size_t)_src.rows() > kx.total() && (size_t)_src.cols() > kx.total(),
-               ocl_sepFilter2D(_src, _dst, sdepth, kx, ky, Point(-1, -1), 0, borderType))
-
-    Mat src = _src.getMat();
-    Mat dst = _dst.getMat();
-
-    Point ofs;
-    Size wsz(src.cols, src.rows);
-    if(!(borderType & BORDER_ISOLATED))
-        src.locateROI( wsz, ofs );
-
-    CALL_HAL(gaussianBlur, cv_hal_gaussianBlur, src.ptr(), src.step, dst.ptr(), dst.step, src.cols, src.rows, sdepth, cn,
-             ofs.x, ofs.y, wsz.width - src.cols - ofs.x, wsz.height - src.rows - ofs.y, ksize.width, ksize.height,
-             sigma1, sigma2, borderType&~BORDER_ISOLATED);
-
-    CV_OVX_RUN(true,
-               openvx_gaussianBlur(src, dst, ksize, sigma1, sigma2, borderType))
-
-    CV_IPP_RUN_FAST(ipp_GaussianBlur(src, dst, ksize, sigma1, sigma2, borderType));
-
-    if(sdepth == CV_8U && ((borderType & BORDER_ISOLATED) || !_src.getMat().isSubmatrix()))
-    {
-        std::vector<ufixedpoint16> fkx, fky;
-        createGaussianKernels(fkx, fky, type, ksize, sigma1, sigma2);
-        if (src.data == dst.data)
-            src = src.clone();
-        fixedSmoothInvoker<uint8_t, ufixedpoint16> invoker(src.ptr<uint8_t>(), src.step1(), dst.ptr<uint8_t>(), dst.step1(), dst.cols, dst.rows, dst.channels(), &fkx[0], (int)fkx.size(), &fky[0], (int)fky.size(), borderType & ~BORDER_ISOLATED);
         parallel_for_(Range(0, dst.rows), invoker, std::max(1, std::min(getNumThreads(), getNumberOfCPUs())));
-        return;
     }
-
-    sepFilter2D(src, dst, sdepth, kx, ky, Point(-1, -1), 0, borderType);
 }
 
-//////////////////////////////////////////////////////////////////////////////////////////
+#endif
-
+CV_CPU_OPTIMIZATION_NAMESPACE_END
-CV_IMPL void
+} // namespace
-cvSmooth( const void* srcarr, void* dstarr, int smooth_type,
-          int param1, int param2, double param3, double param4 )
-{
-    cv::Mat src = cv::cvarrToMat(srcarr), dst0 = cv::cvarrToMat(dstarr), dst = dst0;
-
-    CV_Assert( dst.size() == src.size() &&
-        (smooth_type == CV_BLUR_NO_SCALE || dst.type() == src.type()) );
-
-    if( param2 <= 0 )
-        param2 = param1;
-
-    if( smooth_type == CV_BLUR || smooth_type == CV_BLUR_NO_SCALE )
-        cv::boxFilter( src, dst, dst.depth(), cv::Size(param1, param2), cv::Point(-1,-1),
-            smooth_type == CV_BLUR, cv::BORDER_REPLICATE );
-    else if( smooth_type == CV_GAUSSIAN )
-        cv::GaussianBlur( src, dst, cv::Size(param1, param2), param3, param4, cv::BORDER_REPLICATE );
-    else if( smooth_type == CV_MEDIAN )
-        cv::medianBlur( src, dst, param1 );
-    else
-        cv::bilateralFilter( src, dst, param1, param3, param4, cv::BORDER_REPLICATE );
-
-    if( dst.data != dst0.data )
-        CV_Error( CV_StsUnmatchedFormats, "The destination image does not have the proper type" );
-}
-
-/* End of file. */