/*
 *  This file is part of RawTherapee.
 *
 *  Copyright (c) 2004-2010 Gabor Horvath <hgabor@rawtherapee.com>
 *
 *  RawTherapee is free software: you can redistribute it and/or modify
 *  it under the terms of the GNU General Public License as published by
 *  the Free Software Foundation, either version 3 of the License, or
 *  (at your option) any later version.
 *
 *  RawTherapee is distributed in the hope that it will be useful,
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *  GNU General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License
 *  along with RawTherapee.  If not, see <https://www.gnu.org/licenses/>.
 */
#include <cmath>
#include <cstdlib>
#include <cstring>
#include <vector>

#include "gauss.h"

#include "boxblur.h"
#include "opthelper.h"
#include "rt_math.h"

namespace
{

void compute7x7kernel(float sigma, float kernel[7][7]) {
    const double temp = -2.f * rtengine::SQR(sigma);
    float sum = 0.f;
    for (int i = -3; i <= 3; ++i) {
        for (int j = -3; j <= 3; ++j) {
            if((rtengine::SQR(i) + rtengine::SQR(j)) <= rtengine::SQR(3.0 * 1.15)) {
                kernel[i + 3][j + 3] = std::exp((rtengine::SQR(i) + rtengine::SQR(j)) / temp);
                sum += kernel[i + 3][j + 3];
            } else {
                kernel[i + 3][j + 3] = 0.f;
            }
        }
    }

    for (int i = 0; i < 7; ++i) {
        for (int j = 0; j < 7; ++j) {
            kernel[i][j] /= sum;
        }
    }
}

void compute5x5kernel(float sigma, float kernel[5][5]) {
    const double temp = -2.f * rtengine::SQR(sigma);
    float sum = 0.f;
    for (int i = -2; i <= 2; ++i) {
        for (int j = -2; j <= 2; ++j) {
            if((rtengine::SQR(i) + rtengine::SQR(j)) <= rtengine::SQR(3.0 * 0.84)) {
                kernel[i + 2][j + 2] = std::exp((rtengine::SQR(i) + rtengine::SQR(j)) / temp);
                sum += kernel[i + 2][j + 2];
            } else {
                kernel[i + 2][j + 2] = 0.f;
            }
        }
    }

    for (int i = 0; i < 5; ++i) {
        for (int j = 0; j < 5; ++j) {
            kernel[i][j] /= sum;
        }
    }
}

template<class T> void calculateYvVFactors( const T sigma, T &b1, T &b2, T &b3, T &B, T M[3][3])
{
    // coefficient calculation
    T q;

    if (sigma < 2.5) {
        q = 3.97156 - 4.14554 * sqrt (1.0 - 0.26891 * sigma);
    } else {
        q = 0.98711 * sigma - 0.96330;
    }

    T b0 = 1.57825 + 2.44413 * q + 1.4281 * q * q + 0.422205 * q * q * q;
    b1 = 2.44413 * q + 2.85619 * q * q + 1.26661 * q * q * q;
    b2 = -1.4281 * q * q - 1.26661 * q * q * q;
    b3 = 0.422205 * q * q * q;
    B = 1.0 - (b1 + b2 + b3) / b0;

    b1 /= b0;
    b2 /= b0;
    b3 /= b0;

    // From: Bill Triggs, Michael Sdika: Boundary Conditions for Young-van Vliet Recursive Filtering
    M[0][0] = -b3 * b1 + 1.0 - b3 * b3 - b2;
    M[0][1] = (b3 + b1) * (b2 + b3 * b1);
    M[0][2] = b3 * (b1 + b3 * b2);
    M[1][0] = b1 + b3 * b2;
    M[1][1] = -(b2 - 1.0) * (b2 + b3 * b1);
    M[1][2] = -(b3 * b1 + b3 * b3 + b2 - 1.0) * b3;
    M[2][0] = b3 * b1 + b2 + b1 * b1 - b2 * b2;
    M[2][1] = b1 * b2 + b3 * b2 * b2 - b1 * b3 * b3 - b3 * b3 * b3 - b3 * b2 + b3;
    M[2][2] = b3 * (b1 + b3 * b2);

}

// classical filtering if the support window is small and src != dst
template<class T> void gauss3x3 (T** RESTRICT src, T** RESTRICT dst, const int W, const int H, const T c0, const T c1, const T c2, const T b0, const T b1)
{

    // first row
#ifdef _OPENMP
    #pragma omp single nowait
#endif
    {
        dst[0][0] = src[0][0];

        for (int j = 1; j < W - 1; j++)
        {
            dst[0][j] = b1 * (src[0][j - 1] + src[0][j + 1]) + b0 * src[0][j];
        }

        dst[0][W - 1] = src[0][W - 1];
    }

#ifdef _OPENMP
    #pragma omp for nowait
#endif

    for (int i = 1; i < H - 1; i++) {
        dst[i][0] = b1 * (src[i - 1][0] + src[i + 1][0]) + b0 * src[i][0];

        for (int j = 1; j < W - 1; j++) {
            dst[i][j] = c2 * (src[i - 1][j - 1] + src[i - 1][j + 1] + src[i + 1][j - 1] + src[i + 1][j + 1]) + c1 * (src[i - 1][j] + src[i][j - 1] + src[i][j + 1] + src[i + 1][j]) + c0 * src[i][j];
        }

        dst[i][W - 1] = b1 * (src[i - 1][W - 1] + src[i + 1][W - 1]) + b0 * src[i][W - 1];
    }

    // last row
#ifdef _OPENMP
    #pragma omp single
#endif
    {
        dst[H - 1][0] = src[H - 1][0];

        for (int j = 1; j < W - 1; j++) {
            dst[H - 1][j] = b1 * (src[H - 1][j - 1] + src[H - 1][j + 1]) + b0 * src[H - 1][j];
        }

        dst[H - 1][W - 1] = src[H - 1][W - 1];
    }
}

// classical filtering if the support window is small and src != dst
template<class T> void gauss3x3mult (T** RESTRICT src, T** RESTRICT dst, const int W, const int H, const T c0, const T c1, const T c2, const T b0, const T b1)
{

    // first row
#ifdef _OPENMP
    #pragma omp single nowait
#endif
    {
        dst[0][0] *= src[0][0];

        for (int j = 1; j < W - 1; j++)
        {
            dst[0][j] *= b1 * (src[0][j - 1] + src[0][j + 1]) + b0 * src[0][j];
        }

        dst[0][W - 1] *= src[0][W - 1];
    }

#ifdef _OPENMP
    #pragma omp for nowait
#endif

    for (int i = 1; i < H - 1; i++) {
        dst[i][0] *= b1 * (src[i - 1][0] + src[i + 1][0]) + b0 * src[i][0];

        for (int j = 1; j < W - 1; j++) {
            dst[i][j] *= c2 * (src[i - 1][j - 1] + src[i - 1][j + 1] + src[i + 1][j - 1] + src[i + 1][j + 1]) + c1 * (src[i - 1][j] + src[i][j - 1] + src[i][j + 1] + src[i + 1][j]) + c0 * src[i][j];
        }

        dst[i][W - 1] *= b1 * (src[i - 1][W - 1] + src[i + 1][W - 1]) + b0 * src[i][W - 1];
    }

    // last row
#ifdef _OPENMP
    #pragma omp single
#endif
    {
        dst[H - 1][0] *= src[H - 1][0];

        for (int j = 1; j < W - 1; j++) {
            dst[H - 1][j] *= b1 * (src[H - 1][j - 1] + src[H - 1][j + 1]) + b0 * src[H - 1][j];
        }

        dst[H - 1][W - 1] *= src[H - 1][W - 1];
    }
}

template<class T> void gauss3x3div (T** RESTRICT src, T** RESTRICT dst, T** RESTRICT divBuffer, const int W, const int H, const T c0, const T c1, const T c2, const T b0, const T b1)
{

    // first row
#ifdef _OPENMP
    #pragma omp single nowait
#endif
    {
        dst[0][0] = rtengine::max(divBuffer[0][0] / (src[0][0] > 0.f ? src[0][0] : 1.f), 0.f);

        for (int j = 1; j < W - 1; j++)
        {
            float tmp = (b1 * (src[0][j - 1] + src[0][j + 1]) + b0 * src[0][j]);
            dst[0][j] = rtengine::max(divBuffer[0][j] / (tmp > 0.f ? tmp : 1.f), 0.f);
        }

        dst[0][W - 1] = rtengine::max(divBuffer[0][W - 1] / (src[0][W - 1] > 0.f ? src[0][W - 1] : 1.f), 0.f);
    }

#ifdef _OPENMP
    #pragma omp for nowait
#endif

    for (int i = 1; i < H - 1; i++) {
        float tmp = (b1 * (src[i - 1][0] + src[i + 1][0]) + b0 * src[i][0]);
        dst[i][0] = rtengine::max(divBuffer[i][0] / (tmp > 0.f ? tmp : 1.f), 0.f);

        for (int j = 1; j < W - 1; j++) {
            tmp = (c2 * (src[i - 1][j - 1] + src[i - 1][j + 1] + src[i + 1][j - 1] + src[i + 1][j + 1]) + c1 * (src[i - 1][j] + src[i][j - 1] + src[i][j + 1] + src[i + 1][j]) + c0 * src[i][j]);
            dst[i][j] = rtengine::max(divBuffer[i][j] / (tmp > 0.f ? tmp : 1.f), 0.f);
        }

        tmp = (b1 * (src[i - 1][W - 1] + src[i + 1][W - 1]) + b0 * src[i][W - 1]);
        dst[i][W - 1] = rtengine::max(divBuffer[i][W - 1] / (tmp > 0.f ? tmp : 1.f), 0.f);
    }

    // last row
#ifdef _OPENMP
    #pragma omp single
#endif
    {
        dst[H - 1][0] = rtengine::max(divBuffer[H - 1][0] / (src[H - 1][0] > 0.f ? src[H - 1][0] : 1.f), 0.f);

        for (int j = 1; j < W - 1; j++) {
            float tmp = (b1 * (src[H - 1][j - 1] + src[H - 1][j + 1]) + b0 * src[H - 1][j]);
            dst[H - 1][j] = rtengine::max(divBuffer[H - 1][j] / (tmp > 0.f ? tmp : 1.f), 0.f);
        }

        dst[H - 1][W - 1] = rtengine::max(divBuffer[H - 1][W - 1] / (src[H - 1][W - 1] > 0.f ? src[H - 1][W - 1] : 1.f), 0.f);
    }
}

template<class T> void gauss7x7div (T** RESTRICT src, T** RESTRICT dst, T** RESTRICT divBuffer, const int W, const int H, float sigma)
{

    float kernel[7][7];
    compute7x7kernel(sigma, kernel);

    const float c31 = kernel[0][2];
    const float c30 = kernel[0][3];
    const float c22 = kernel[1][1];
    const float c21 = kernel[1][2];
    const float c20 = kernel[1][3];
    const float c11 = kernel[2][2];
    const float c10 = kernel[2][3];
    const float c00 = kernel[3][3];

#ifdef _OPENMP
    #pragma omp for schedule(dynamic, 16) nowait
#endif

    for (int i = 3; i < H - 3; ++i) {
        dst[i][0] = dst[i][1] = dst[i][2] = 1.f;
        // I tried hand written SSE code but gcc vectorizes better
        for (int j = 3; j < W - 3; ++j) {
            const float val = c31 * (src[i - 3][j - 1] + src[i - 3][j + 1] + src[i - 1][j - 3] + src[i - 1][j + 3] + src[i + 1][j - 3] + src[i + 1][j + 3] + src[i + 3][j - 1] + src[i + 3][j + 1]) +
                              c30 * (src[i - 3][j] + src[i][j - 3] + src[i][j + 3] + src[i + 3][j]) +
                              c22 * (src[i - 2][j - 2] + src[i - 2][j + 2] + src[i + 2][j - 2] + src[i + 2][j + 2]) +
                              c21 * (src[i - 2][j - 1] + src[i - 2][j + 1] * c21 + src[i - 1][j - 2] + src[i - 1][j + 2] + src[i + 1][j - 2] + src[i + 1][j + 2] + src[i + 2][j - 1] + src[i + 2][j + 1]) +
                              c20 * (src[i - 2][j] + src[i][j - 2] + src[i][j + 2] + src[i + 2][j]) +
                              c11 * (src[i - 1][j - 1] + src[i - 1][j + 1] + src[i + 1][j - 1] + src[i + 1][j + 1]) +
                              c10 * (src[i - 1][j] + src[i][j - 1] + src[i][j + 1] + src[i + 1][j]) +
                              c00 * src[i][j];

            dst[i][j] = divBuffer[i][j] / std::max(val, 0.00001f);
        }
        dst[i][W - 3] = dst[i][W - 2] = dst[i][W - 1] = 1.f;
    }

    // first and last rows
#ifdef _OPENMP
    #pragma omp single
#endif
    {
        for (int i = 0; i < 3; ++i) {
            for (int j = 0; j < W; ++j) {
                dst[i][j] = 1.f;
            }
        }
        for (int i = H - 3 ; i < H; ++i) {
            for (int j = 0; j < W; ++j) {
                dst[i][j] = 1.f;
            }
        }
    }
}

template<class T> void gauss5x5div (T** RESTRICT src, T** RESTRICT dst, T** RESTRICT divBuffer, const int W, const int H, float sigma)
{

    float kernel[5][5];
    compute5x5kernel(sigma, kernel);

    const float c21 = kernel[0][1];
    const float c20 = kernel[0][2];
    const float c11 = kernel[1][1];
    const float c10 = kernel[1][2];
    const float c00 = kernel[2][2];

#ifdef _OPENMP
    #pragma omp for schedule(dynamic, 16) nowait
#endif

    for (int i = 2; i < H - 2; ++i) {
        dst[i][0] = dst[i][1] = 1.f;
        // I tried hand written SSE code but gcc vectorizes better
        for (int j = 2; j < W - 2; ++j) {
            const float val = c21 * (src[i - 2][j - 1] + src[i - 2][j + 1] + src[i - 1][j - 2] + src[i - 1][j + 2] + src[i + 1][j - 2] + src[i + 1][j + 2] + src[i + 2][j - 1] + src[i + 2][j + 1]) +
                              c20 * (src[i - 2][j] + src[i][j - 2] + src[i][j + 2] + src[i + 2][j]) +
                              c11 * (src[i - 1][j - 1] + src[i - 1][j + 1] + src[i + 1][j - 1] + src[i + 1][j + 1]) +
                              c10 * (src[i - 1][j] + src[i][j - 1] + src[i][j + 1] + src[i + 1][j]) +
                              c00 * src[i][j];

            dst[i][j] = divBuffer[i][j] / std::max(val, 0.00001f);
        }
        dst[i][W - 2] = dst[i][W - 1] = 1.f;
    }

    // first and last rows
#ifdef _OPENMP
    #pragma omp single
#endif
    {
        for (int i = 0; i < 2; ++i) {
            for (int j = 0; j < W; ++j) {
                dst[i][j] = 1.f;
            }
        }
        for (int i = H - 2 ; i < H; ++i) {
            for (int j = 0; j < W; ++j) {
                dst[i][j] = 1.f;
            }
        }
    }
}

template<class T> void gauss7x7mult (T** RESTRICT src, T** RESTRICT dst, const int W, const int H, float sigma)
{

    float kernel[7][7];
    compute7x7kernel(sigma, kernel);
    const float c31 = kernel[0][2];
    const float c30 = kernel[0][3];
    const float c22 = kernel[1][1];
    const float c21 = kernel[1][2];
    const float c20 = kernel[1][3];
    const float c11 = kernel[2][2];
    const float c10 = kernel[2][3];
    const float c00 = kernel[3][3];

#ifdef _OPENMP
    #pragma omp for schedule(dynamic, 16)
#endif

    for (int i = 3; i < H - 3; ++i) {
        // I tried hand written SSE code but gcc vectorizes better
        for (int j = 3; j < W - 3; ++j) {
            const float val = c31 * (src[i - 3][j - 1] + src[i - 3][j + 1] + src[i - 1][j - 3] + src[i - 1][j + 3] + src[i + 1][j - 3] + src[i + 1][j + 3] + src[i + 3][j - 1] + src[i + 3][j + 1]) +
                              c30 * (src[i - 3][j] + src[i][j - 3] + src[i][j + 3] + src[i + 3][j]) +
                              c22 * (src[i - 2][j - 2] + src[i - 2][j + 2] + src[i + 2][j - 2] + src[i + 2][j + 2]) +
                              c21 * (src[i - 2][j - 1] + src[i - 2][j + 1] * c21 + src[i - 1][j - 2] + src[i - 1][j + 2] + src[i + 1][j - 2] + src[i + 1][j + 2] + src[i + 2][j - 1] + src[i + 2][j + 1]) +
                              c20 * (src[i - 2][j] + src[i][j - 2] + src[i][j + 2] + src[i + 2][j]) +
                              c11 * (src[i - 1][j - 1] + src[i - 1][j + 1] + src[i + 1][j - 1] + src[i + 1][j + 1]) +
                              c10 * (src[i - 1][j] + src[i][j - 1] + src[i][j + 1] + src[i + 1][j]) +
                              c00 * src[i][j];

            dst[i][j] *= val;
        }
    }
}

template<class T> void gauss5x5mult (T** RESTRICT src, T** RESTRICT dst, const int W, const int H, float sigma)
{

    float kernel[5][5];
    compute5x5kernel(sigma, kernel);

    const float c21 = kernel[0][1];
    const float c20 = kernel[0][2];
    const float c11 = kernel[1][1];
    const float c10 = kernel[1][2];
    const float c00 = kernel[2][2];

#ifdef _OPENMP
    #pragma omp for schedule(dynamic, 16)
#endif

    for (int i = 2; i < H - 2; ++i) {
        // I tried hand written SSE code but gcc vectorizes better
        for (int j = 2; j < W - 2; ++j) {
            const float val = c21 * (src[i - 2][j - 1] + src[i - 2][j + 1] + src[i - 1][j - 2] + src[i - 1][j + 2] + src[i + 1][j - 2] + src[i + 1][j + 2] + src[i + 2][j - 1] + src[i + 2][j + 1]) +
                              c20 * (src[i - 2][j] + src[i][j - 2] + src[i][j + 2] + src[i + 2][j]) +
                              c11 * (src[i - 1][j - 1] + src[i - 1][j + 1] + src[i + 1][j - 1] + src[i + 1][j + 1]) +
                              c10 * (src[i - 1][j] + src[i][j - 1] + src[i][j + 1] + src[i + 1][j]) +
                              c00 * src[i][j];

            dst[i][j] *= val;
        }
    }
}

// use separated filter if the support window is small and src == dst
template<class T> void gaussHorizontal3 (T** src, T** dst, int W, int H, const float c0, const float c1)
{
    T temp[W] ALIGNED16;
#ifdef _OPENMP
    #pragma omp for
#endif

    for (int i = 0; i < H; i++) {
        for (int j = 1; j < W - 1; j++) {
            temp[j] = (T)(c1 * (src[i][j - 1] + src[i][j + 1]) + c0 * src[i][j]);
        }

        dst[i][0] = src[i][0];
        memcpy (dst[i] + 1, temp + 1, (W - 2)*sizeof(T));

        dst[i][W - 1] = src[i][W - 1];
    }
}

#ifdef __SSE2__
template<class T> void gaussVertical3 (T** src, T** dst, int W, int H, const float c0, const float c1)
{
    vfloat Tv = F2V(0.f), Tm1v, Tp1v;
    vfloat Tv1 = F2V(0.f), Tm1v1, Tp1v1;
    vfloat c0v, c1v;
    c0v = F2V(c0);
    c1v = F2V(c1);

#ifdef _OPENMP
    #pragma omp for nowait
#endif

    // process 8 columns per iteration for better usage of cpu cache
    for (int i = 0; i < W - 7; i += 8) {
        Tm1v = LVFU( src[0][i] );
        Tm1v1 = LVFU( src[0][i + 4] );
        STVFU( dst[0][i], Tm1v);
        STVFU( dst[0][i + 4], Tm1v1);

        if (H > 1) {
            Tv = LVFU( src[1][i]);
            Tv1 = LVFU( src[1][i + 4]);
        }

        for (int j = 1; j < H - 1; j++) {
            Tp1v = LVFU( src[j + 1][i]);
            Tp1v1 = LVFU( src[j + 1][i + 4]);
            STVFU( dst[j][i], c1v * (Tp1v + Tm1v) + Tv * c0v);
            STVFU( dst[j][i + 4], c1v * (Tp1v1 + Tm1v1) + Tv1 * c0v);
            Tm1v = Tv;
            Tm1v1 = Tv1;
            Tv = Tp1v;
            Tv1 = Tp1v1;
        }

        STVFU( dst[H - 1][i], LVFU( src[H - 1][i]));
        STVFU( dst[H - 1][i + 4], LVFU( src[H - 1][i + 4]));
    }

// Borders are done without SSE
    float temp[H] ALIGNED16;
#ifdef _OPENMP
    #pragma omp single
#endif

    for (int i = W - (W % 8); i < W; i++) {
        for (int j = 1; j < H - 1; j++) {
            temp[j] = c1 * (src[j - 1][i] + src[j + 1][i]) + c0 * src[j][i];
        }

        dst[0][i] = src[0][i];

        for (int j = 1; j < H - 1; j++) {
            dst[j][i] = temp[j];
        }

        dst[H - 1][i] = src[H - 1][i];
    }
}
#else
template<class T> void gaussVertical3 (T** src, T** dst, int W, int H, const float c0, const float c1)
{
    T temp[H] ALIGNED16;
#ifdef _OPENMP
    #pragma omp for
#endif

    for (int i = 0; i < W; i++) {
        for (int j = 1; j < H - 1; j++) {
            temp[j] = (T)(c1 * (src[j - 1][i] + src[j + 1][i]) + c0 * src[j][i]);
        }

        dst[0][i] = src[0][i];

        for (int j = 1; j < H - 1; j++) {
            dst[j][i] = temp[j];
        }

        dst[H - 1][i] = src[H - 1][i];
    }
}
#endif

#ifdef __SSE2__
// fast gaussian approximation if the support window is large
template<class T> void gaussHorizontalSse (T** src, T** dst, const int W, const int H, const float sigma)
{
    double b1, b2, b3, B, M[3][3];
    calculateYvVFactors<double>(sigma, b1, b2, b3, B, M);

    for (int i = 0; i < 3; i++)
        for (int j = 0; j < 3; j++) {
            M[i][j] *= (1.0 + b2 + (b1 - b3) * b3);
            M[i][j] /= (1.0 + b1 - b2 + b3) * (1.0 - b1 - b2 - b3);
        }

    vfloat Rv;
    vfloat Tv, Tm2v, Tm3v;
    vfloat Bv, b1v, b2v, b3v;
    vfloat temp2W, temp2Wp1;
    float tmp[W][4] ALIGNED16;
    Bv = F2V(B);
    b1v = F2V(b1);
    b2v = F2V(b2);
    b3v = F2V(b3);

#ifdef _OPENMP
    #pragma omp for nowait
#endif

    for (int i = 0; i < H - 3; i += 4) {
        Tv = _mm_set_ps(src[i][0], src[i + 1][0], src[i + 2][0], src[i + 3][0]);
        Tm3v = Tv * (Bv + b1v + b2v + b3v);
        STVF( tmp[0][0], Tm3v );

        Tm2v = _mm_set_ps(src[i][1], src[i + 1][1], src[i + 2][1], src[i + 3][1]) * Bv + Tm3v * b1v + Tv * (b2v + b3v);
        STVF( tmp[1][0], Tm2v );

        Rv = _mm_set_ps(src[i][2], src[i + 1][2], src[i + 2][2], src[i + 3][2]) * Bv + Tm2v * b1v + Tm3v * b2v + Tv * b3v;
        STVF( tmp[2][0], Rv );

        for (int j = 3; j < W; j++) {
            Tv = Rv;
            Rv = _mm_set_ps(src[i][j], src[i + 1][j], src[i + 2][j], src[i + 3][j]) * Bv + Tv * b1v + Tm2v * b2v + Tm3v * b3v;
            STVF( tmp[j][0], Rv );
            Tm3v = Tm2v;
            Tm2v = Tv;
        }

        Tv = _mm_set_ps(src[i][W - 1], src[i + 1][W - 1], src[i + 2][W - 1], src[i + 3][W - 1]);

        temp2Wp1 = Tv + F2V(M[2][0]) * (Rv - Tv) + F2V(M[2][1]) * ( Tm2v - Tv ) +  F2V(M[2][2]) * (Tm3v - Tv);
        temp2W = Tv + F2V(M[1][0]) * (Rv - Tv) + F2V(M[1][1]) * (Tm2v - Tv) + F2V(M[1][2]) * (Tm3v - Tv);

        Rv = Tv + F2V(M[0][0]) * (Rv - Tv) + F2V(M[0][1]) * (Tm2v - Tv) + F2V(M[0][2]) * (Tm3v - Tv);
        STVF(tmp[W - 1][0], Rv);

        Tm2v = Bv * Tm2v + b1v * Rv + b2v * temp2W + b3v * temp2Wp1;
        STVF(tmp[W - 2][0], Tm2v);

        Tm3v = Bv * Tm3v + b1v * Tm2v + b2v * Rv + b3v * temp2W;
        STVF(tmp[W - 3][0], Tm3v);

        Tv = Rv;
        Rv = Tm3v;
        Tm3v = Tv;

        for (int j = W - 4; j >= 0; j--) {
            Tv = Rv;
            Rv = LVF(tmp[j][0]) * Bv + Tv * b1v + Tm2v * b2v + Tm3v * b3v;
            STVF(tmp[j][0], Rv);
            Tm3v = Tm2v;
            Tm2v = Tv;
        }

        for (int j = 0; j < W; j++) {
            dst[i + 3][j] = tmp[j][0];
            dst[i + 2][j] = tmp[j][1];
            dst[i + 1][j] = tmp[j][2];
            dst[i + 0][j] = tmp[j][3];
        }


    }

// Borders are done without SSE
#ifdef _OPENMP
    #pragma omp single
#endif

    for (int i = H - (H % 4); i < H; i++) {
        tmp[0][0] = src[i][0] * (B + b1 + b2 + b3);
        tmp[1][0] = B * src[i][1] + b1 * tmp[0][0]  + src[i][0] * (b2 + b3);
        tmp[2][0] = B * src[i][2] + b1 * tmp[1][0]  + b2 * tmp[0][0]  + b3 * src[i][0];

        for (int j = 3; j < W; j++) {
            tmp[j][0] = B * src[i][j] + b1 * tmp[j - 1][0] + b2 * tmp[j - 2][0] + b3 * tmp[j - 3][0];
        }

        float temp2Wm1 = src[i][W - 1] + M[0][0] * (tmp[W - 1][0] - src[i][W - 1]) + M[0][1] * (tmp[W - 2][0] - src[i][W - 1]) + M[0][2] * (tmp[W - 3][0] - src[i][W - 1]);
        float temp2W   = src[i][W - 1] + M[1][0] * (tmp[W - 1][0] - src[i][W - 1]) + M[1][1] * (tmp[W - 2][0] - src[i][W - 1]) + M[1][2] * (tmp[W - 3][0] - src[i][W - 1]);
        float temp2Wp1 = src[i][W - 1] + M[2][0] * (tmp[W - 1][0] - src[i][W - 1]) + M[2][1] * (tmp[W - 2][0] - src[i][W - 1]) + M[2][2] * (tmp[W - 3][0] - src[i][W - 1]);

        tmp[W - 1][0] = temp2Wm1;
        tmp[W - 2][0] = B * tmp[W - 2][0] + b1 * tmp[W - 1][0] + b2 * temp2W + b3 * temp2Wp1;
        tmp[W - 3][0] = B * tmp[W - 3][0] + b1 * tmp[W - 2][0] + b2 * tmp[W - 1][0] + b3 * temp2W;

        for (int j = W - 4; j >= 0; j--) {
            tmp[j][0] = B * tmp[j][0] + b1 * tmp[j + 1][0] + b2 * tmp[j + 2][0] + b3 * tmp[j + 3][0];
        }

        for (int j = 0; j < W; j++) {
            dst[i][j] = tmp[j][0];
        }
    }
}
#endif

// fast gaussian approximation if the support window is large
template<class T> void gaussHorizontal (T** src, T** dst, const int W, const int H, const double sigma)
{
    double b1, b2, b3, B, M[3][3];
    calculateYvVFactors<double>(sigma, b1, b2, b3, B, M);

    for (int i = 0; i < 3; i++)
        for (int j = 0; j < 3; j++) {
            M[i][j] /= (1.0 + b1 - b2 + b3) * (1.0 + b2 + (b1 - b3) * b3);
        }

    double temp2[W] ALIGNED16;

#ifdef _OPENMP
    #pragma omp for
#endif

    for (int i = 0; i < H; i++) {

        temp2[0] = B * src[i][0] + b1 * src[i][0] + b2 * src[i][0] + b3 * src[i][0];
        temp2[1] = B * src[i][1] + b1 * temp2[0]  + b2 * src[i][0] + b3 * src[i][0];
        temp2[2] = B * src[i][2] + b1 * temp2[1]  + b2 * temp2[0]  + b3 * src[i][0];

        for (int j = 3; j < W; j++) {
            temp2[j] = B * src[i][j] + b1 * temp2[j - 1] + b2 * temp2[j - 2] + b3 * temp2[j - 3];
        }

        double temp2Wm1 = src[i][W - 1] + M[0][0] * (temp2[W - 1] - src[i][W - 1]) + M[0][1] * (temp2[W - 2] - src[i][W - 1]) + M[0][2] * (temp2[W - 3] - src[i][W - 1]);
        double temp2W   = src[i][W - 1] + M[1][0] * (temp2[W - 1] - src[i][W - 1]) + M[1][1] * (temp2[W - 2] - src[i][W - 1]) + M[1][2] * (temp2[W - 3] - src[i][W - 1]);
        double temp2Wp1 = src[i][W - 1] + M[2][0] * (temp2[W - 1] - src[i][W - 1]) + M[2][1] * (temp2[W - 2] - src[i][W - 1]) + M[2][2] * (temp2[W - 3] - src[i][W - 1]);

        temp2[W - 1] = temp2Wm1;
        temp2[W - 2] = B * temp2[W - 2] + b1 * temp2[W - 1] + b2 * temp2W + b3 * temp2Wp1;
        temp2[W - 3] = B * temp2[W - 3] + b1 * temp2[W - 2] + b2 * temp2[W - 1] + b3 * temp2W;

        for (int j = W - 4; j >= 0; j--) {
            temp2[j] = B * temp2[j] + b1 * temp2[j + 1] + b2 * temp2[j + 2] + b3 * temp2[j + 3];
        }

        for (int j = 0; j < W; j++) {
            dst[i][j] = (T)temp2[j];
        }

    }
}

#ifdef __SSE2__
template<class T> void gaussVerticalSse (T** src, T** dst, const int W, const int H, const float sigma)
{
    double b1, b2, b3, B, M[3][3];
    calculateYvVFactors<double>(sigma, b1, b2, b3, B, M);

    for (int i = 0; i < 3; i++)
        for (int j = 0; j < 3; j++) {
            M[i][j] *= (1.0 + b2 + (b1 - b3) * b3);
            M[i][j] /= (1.0 + b1 - b2 + b3) * (1.0 - b1 - b2 - b3);
        }

    float tmp[H][8] ALIGNED16;
    vfloat Rv;
    vfloat Tv, Tm2v, Tm3v;
    vfloat Rv1;
    vfloat Tv1, Tm2v1, Tm3v1;
    vfloat Bv, b1v, b2v, b3v;
    vfloat temp2W, temp2Wp1;
    vfloat temp2W1, temp2Wp11;
    Bv = F2V(B);
    b1v = F2V(b1);
    b2v = F2V(b2);
    b3v = F2V(b3);

#ifdef _OPENMP
    #pragma omp for nowait
#endif

    // process 8 columns per iteration for better usage of cpu cache
    for (int i = 0; i < W - 7; i += 8) {
        Tv = LVFU( src[0][i]);
        Tv1 = LVFU( src[0][i + 4]);
        Rv = Tv * (Bv + b1v + b2v + b3v);
        Rv1 = Tv1 * (Bv + b1v + b2v + b3v);
        Tm3v = Rv;
        Tm3v1 = Rv1;
        STVF( tmp[0][0], Rv );
        STVF( tmp[0][4], Rv1 );

        Rv = LVFU(src[1][i]) * Bv + Rv * b1v + Tv * (b2v + b3v);
        Rv1 = LVFU(src[1][i + 4]) * Bv + Rv1 * b1v + Tv1 * (b2v + b3v);
        Tm2v = Rv;
        Tm2v1 = Rv1;
        STVF( tmp[1][0], Rv );
        STVF( tmp[1][4], Rv1 );

        Rv = LVFU(src[2][i]) * Bv + Rv * b1v + Tm3v * b2v + Tv * b3v;
        Rv1 = LVFU(src[2][i + 4]) * Bv + Rv1 * b1v + Tm3v1 * b2v + Tv1 * b3v;
        STVF( tmp[2][0], Rv );
        STVF( tmp[2][4], Rv1 );

        for (int j = 3; j < H; j++) {
            Tv = Rv;
            Tv1 = Rv1;
            Rv = LVFU(src[j][i]) * Bv +  Tv * b1v + Tm2v * b2v + Tm3v * b3v;
            Rv1 = LVFU(src[j][i + 4]) * Bv +  Tv1 * b1v + Tm2v1 * b2v + Tm3v1 * b3v;
            STVF( tmp[j][0], Rv );
            STVF( tmp[j][4], Rv1 );
            Tm3v = Tm2v;
            Tm3v1 = Tm2v1;
            Tm2v = Tv;
            Tm2v1 = Tv1;
        }

        Tv = LVFU(src[H - 1][i]);
        Tv1 = LVFU(src[H - 1][i + 4]);

        temp2Wp1 = Tv + F2V(M[2][0]) * (Rv - Tv) + F2V(M[2][1]) * (Tm2v - Tv) + F2V(M[2][2]) * (Tm3v - Tv);
        temp2Wp11 = Tv1 + F2V(M[2][0]) * (Rv1 - Tv1) + F2V(M[2][1]) * (Tm2v1 - Tv1) + F2V(M[2][2]) * (Tm3v1 - Tv1);
        temp2W = Tv + F2V(M[1][0]) * (Rv - Tv) + F2V(M[1][1]) * (Tm2v - Tv) + F2V(M[1][2]) * (Tm3v - Tv);
        temp2W1 = Tv1 + F2V(M[1][0]) * (Rv1 - Tv1) + F2V(M[1][1]) * (Tm2v1 - Tv1) + F2V(M[1][2]) * (Tm3v1 - Tv1);

        Rv = Tv + F2V(M[0][0]) * (Rv - Tv) + F2V(M[0][1]) * (Tm2v - Tv) + F2V(M[0][2]) * (Tm3v - Tv);
        Rv1 = Tv1 + F2V(M[0][0]) * (Rv1 - Tv1) + F2V(M[0][1]) * (Tm2v1 - Tv1) + F2V(M[0][2]) * (Tm3v1 - Tv1);
        STVFU( dst[H - 1][i], Rv );
        STVFU( dst[H - 1][i + 4], Rv1 );

        Tm2v = Bv * Tm2v + b1v * Rv + b2v * temp2W + b3v * temp2Wp1;
        Tm2v1 = Bv * Tm2v1 + b1v * Rv1 + b2v * temp2W1 + b3v * temp2Wp11;
        STVFU( dst[H - 2][i], Tm2v );
        STVFU( dst[H - 2][i + 4], Tm2v1 );

        Tm3v = Bv * Tm3v + b1v * Tm2v + b2v * Rv + b3v * temp2W;
        Tm3v1 = Bv * Tm3v1 + b1v * Tm2v1 + b2v * Rv1 + b3v * temp2W1;
        STVFU( dst[H - 3][i], Tm3v );
        STVFU( dst[H - 3][i + 4], Tm3v1 );

        Tv = Rv;
        Tv1 = Rv1;
        Rv = Tm3v;
        Rv1 = Tm3v1;
        Tm3v = Tv;
        Tm3v1 = Tv1;

        for (int j = H - 4; j >= 0; j--) {
            Tv = Rv;
            Tv1 = Rv1;
            Rv = LVF(tmp[j][0]) * Bv +  Tv * b1v + Tm2v * b2v + Tm3v * b3v;
            Rv1 = LVF(tmp[j][4]) * Bv +  Tv1 * b1v + Tm2v1 * b2v + Tm3v1 * b3v;
            STVFU( dst[j][i], Rv );
            STVFU( dst[j][i + 4], Rv1 );
            Tm3v = Tm2v;
            Tm3v1 = Tm2v1;
            Tm2v = Tv;
            Tm2v1 = Tv1;
        }
    }

// Borders are done without SSE
#ifdef _OPENMP
    #pragma omp single
#endif

    for (int i = W - (W % 8); i < W; i++) {
        tmp[0][0] = src[0][i] * (B + b1 + b2 + b3);
        tmp[1][0] = B * src[1][i] + b1 * tmp[0][0] + src[0][i] * (b2 + b3);
        tmp[2][0] = B * src[2][i] + b1 * tmp[1][0] + b2 * tmp[0][0] + b3 * src[0][i];

        for (int j = 3; j < H; j++) {
            tmp[j][0] = B * src[j][i] + b1 * tmp[j - 1][0] + b2 * tmp[j - 2][0] + b3 * tmp[j - 3][0];
        }

        float temp2Hm1 = src[H - 1][i] + M[0][0] * (tmp[H - 1][0] - src[H - 1][i]) + M[0][1] * (tmp[H - 2][0] - src[H - 1][i]) + M[0][2] * (tmp[H - 3][0] - src[H - 1][i]);
        float temp2H   = src[H - 1][i] + M[1][0] * (tmp[H - 1][0] - src[H - 1][i]) + M[1][1] * (tmp[H - 2][0] - src[H - 1][i]) + M[1][2] * (tmp[H - 3][0] - src[H - 1][i]);
        float temp2Hp1 = src[H - 1][i] + M[2][0] * (tmp[H - 1][0] - src[H - 1][i]) + M[2][1] * (tmp[H - 2][0] - src[H - 1][i]) + M[2][2] * (tmp[H - 3][0] - src[H - 1][i]);

        tmp[H - 1][0] = temp2Hm1;
        tmp[H - 2][0] = B * tmp[H - 2][0] + b1 * tmp[H - 1][0] + b2 * temp2H + b3 * temp2Hp1;
        tmp[H - 3][0] = B * tmp[H - 3][0] + b1 * tmp[H - 2][0] + b2 * tmp[H - 1][0] + b3 * temp2H;

        for (int j = H - 4; j >= 0; j--) {
            tmp[j][0] = B * tmp[j][0] + b1 * tmp[j + 1][0] + b2 * tmp[j + 2][0] + b3 * tmp[j + 3][0];
        }

        for (int j = 0; j < H; j++) {
            dst[j][i] = tmp[j][0];
        }

    }
}
#endif

#ifdef __SSE2__
template<class T> void gaussVerticalSsemult (T** RESTRICT src, T** RESTRICT dst, const int W, const int H, const float sigma)
{
    double b1, b2, b3, B, M[3][3];
    calculateYvVFactors<double>(sigma, b1, b2, b3, B, M);

    for (int i = 0; i < 3; i++)
        for (int j = 0; j < 3; j++) {
            M[i][j] *= (1.0 + b2 + (b1 - b3) * b3);
            M[i][j] /= (1.0 + b1 - b2 + b3) * (1.0 - b1 - b2 - b3);
        }

    float tmp[H][8] ALIGNED16;
    vfloat Rv;
    vfloat Tv, Tm2v, Tm3v;
    vfloat Rv1;
    vfloat Tv1, Tm2v1, Tm3v1;
    vfloat Bv, b1v, b2v, b3v;
    vfloat temp2W, temp2Wp1;
    vfloat temp2W1, temp2Wp11;
    Bv = F2V(B);
    b1v = F2V(b1);
    b2v = F2V(b2);
    b3v = F2V(b3);

#ifdef _OPENMP
    #pragma omp for nowait
#endif

    // process 8 columns per iteration for better usage of cpu cache
    for (int i = 0; i < W - 7; i += 8) {
        Tv = LVFU( src[0][i]);
        Tv1 = LVFU( src[0][i + 4]);
        Rv = Tv * (Bv + b1v + b2v + b3v);
        Rv1 = Tv1 * (Bv + b1v + b2v + b3v);
        Tm3v = Rv;
        Tm3v1 = Rv1;
        STVF( tmp[0][0], Rv );
        STVF( tmp[0][4], Rv1 );

        Rv = LVFU(src[1][i]) * Bv + Rv * b1v + Tv * (b2v + b3v);
        Rv1 = LVFU(src[1][i + 4]) * Bv + Rv1 * b1v + Tv1 * (b2v + b3v);
        Tm2v = Rv;
        Tm2v1 = Rv1;
        STVF( tmp[1][0], Rv );
        STVF( tmp[1][4], Rv1 );

        Rv = LVFU(src[2][i]) * Bv + Rv * b1v + Tm3v * b2v + Tv * b3v;
        Rv1 = LVFU(src[2][i + 4]) * Bv + Rv1 * b1v + Tm3v1 * b2v + Tv1 * b3v;
        STVF( tmp[2][0], Rv );
        STVF( tmp[2][4], Rv1 );

        for (int j = 3; j < H; j++) {
            Tv = Rv;
            Tv1 = Rv1;
            Rv = LVFU(src[j][i]) * Bv +  Tv * b1v + Tm2v * b2v + Tm3v * b3v;
            Rv1 = LVFU(src[j][i + 4]) * Bv +  Tv1 * b1v + Tm2v1 * b2v + Tm3v1 * b3v;
            STVF( tmp[j][0], Rv );
            STVF( tmp[j][4], Rv1 );
            Tm3v = Tm2v;
            Tm3v1 = Tm2v1;
            Tm2v = Tv;
            Tm2v1 = Tv1;
        }

        Tv = LVFU(src[H - 1][i]);
        Tv1 = LVFU(src[H - 1][i + 4]);

        temp2Wp1 = Tv + F2V(M[2][0]) * (Rv - Tv) + F2V(M[2][1]) * (Tm2v - Tv) + F2V(M[2][2]) * (Tm3v - Tv);
        temp2Wp11 = Tv1 + F2V(M[2][0]) * (Rv1 - Tv1) + F2V(M[2][1]) * (Tm2v1 - Tv1) + F2V(M[2][2]) * (Tm3v1 - Tv1);
        temp2W = Tv + F2V(M[1][0]) * (Rv - Tv) + F2V(M[1][1]) * (Tm2v - Tv) + F2V(M[1][2]) * (Tm3v - Tv);
        temp2W1 = Tv1 + F2V(M[1][0]) * (Rv1 - Tv1) + F2V(M[1][1]) * (Tm2v1 - Tv1) + F2V(M[1][2]) * (Tm3v1 - Tv1);

        Rv = Tv + F2V(M[0][0]) * (Rv - Tv) + F2V(M[0][1]) * (Tm2v - Tv) + F2V(M[0][2]) * (Tm3v - Tv);
        Rv1 = Tv1 + F2V(M[0][0]) * (Rv1 - Tv1) + F2V(M[0][1]) * (Tm2v1 - Tv1) + F2V(M[0][2]) * (Tm3v1 - Tv1);
        STVFU( dst[H - 1][i], LVFU(dst[H - 1][i]) * Rv );
        STVFU( dst[H - 1][i + 4], LVFU(dst[H - 1][i + 4]) * Rv1 );

        Tm2v = Bv * Tm2v + b1v * Rv + b2v * temp2W + b3v * temp2Wp1;
        Tm2v1 = Bv * Tm2v1 + b1v * Rv1 + b2v * temp2W1 + b3v * temp2Wp11;
        STVFU( dst[H - 2][i], LVFU(dst[H - 2][i]) * Tm2v );
        STVFU( dst[H - 2][i + 4], LVFU(dst[H - 2][i + 4]) * Tm2v1 );

        Tm3v = Bv * Tm3v + b1v * Tm2v + b2v * Rv + b3v * temp2W;
        Tm3v1 = Bv * Tm3v1 + b1v * Tm2v1 + b2v * Rv1 + b3v * temp2W1;
        STVFU( dst[H - 3][i], LVFU(dst[H - 3][i]) * Tm3v );
        STVFU( dst[H - 3][i + 4], LVFU(dst[H - 3][i + 4]) * Tm3v1 );

        Tv = Rv;
        Tv1 = Rv1;
        Rv = Tm3v;
        Rv1 = Tm3v1;
        Tm3v = Tv;
        Tm3v1 = Tv1;

        for (int j = H - 4; j >= 0; j--) {
            Tv = Rv;
            Tv1 = Rv1;
            Rv = LVF(tmp[j][0]) * Bv +  Tv * b1v + Tm2v * b2v + Tm3v * b3v;
            Rv1 = LVF(tmp[j][4]) * Bv +  Tv1 * b1v + Tm2v1 * b2v + Tm3v1 * b3v;
            STVFU( dst[j][i], LVFU(dst[j][i]) * Rv );
            STVFU( dst[j][i + 4], LVFU(dst[j][i + 4]) * Rv1 );
            Tm3v = Tm2v;
            Tm3v1 = Tm2v1;
            Tm2v = Tv;
            Tm2v1 = Tv1;
        }
    }

// Borders are done without SSE
#ifdef _OPENMP
    #pragma omp single
#endif

    for (int i = W - (W % 8); i < W; i++) {
        tmp[0][0] = src[0][i] * (B + b1 + b2 + b3);
        tmp[1][0] = B * src[1][i] + b1 * tmp[0][0] + src[0][i] * (b2 + b3);
        tmp[2][0] = B * src[2][i] + b1 * tmp[1][0] + b2 * tmp[0][0] + b3 * src[0][i];

        for (int j = 3; j < H; j++) {
            tmp[j][0] = B * src[j][i] + b1 * tmp[j - 1][0] + b2 * tmp[j - 2][0] + b3 * tmp[j - 3][0];
        }

        float temp2Hm1 = src[H - 1][i] + M[0][0] * (tmp[H - 1][0] - src[H - 1][i]) + M[0][1] * (tmp[H - 2][0] - src[H - 1][i]) + M[0][2] * (tmp[H - 3][0] - src[H - 1][i]);
        float temp2H   = src[H - 1][i] + M[1][0] * (tmp[H - 1][0] - src[H - 1][i]) + M[1][1] * (tmp[H - 2][0] - src[H - 1][i]) + M[1][2] * (tmp[H - 3][0] - src[H - 1][i]);
        float temp2Hp1 = src[H - 1][i] + M[2][0] * (tmp[H - 1][0] - src[H - 1][i]) + M[2][1] * (tmp[H - 2][0] - src[H - 1][i]) + M[2][2] * (tmp[H - 3][0] - src[H - 1][i]);

        tmp[H - 1][0] = temp2Hm1;
        tmp[H - 2][0] = B * tmp[H - 2][0] + b1 * tmp[H - 1][0] + b2 * temp2H + b3 * temp2Hp1;
        tmp[H - 3][0] = B * tmp[H - 3][0] + b1 * tmp[H - 2][0] + b2 * tmp[H - 1][0] + b3 * temp2H;

        for (int j = H - 4; j >= 0; j--) {
            tmp[j][0] = B * tmp[j][0] + b1 * tmp[j + 1][0] + b2 * tmp[j + 2][0] + b3 * tmp[j + 3][0];
        }

        for (int j = 0; j < H; j++) {
            dst[j][i] *= tmp[j][0];
        }

    }
}

template<class T> void gaussVerticalSsediv (T** src, T** dst, T** divBuffer, const int W, const int H, const float sigma)
{
    double b1, b2, b3, B, M[3][3];
    calculateYvVFactors<double>(sigma, b1, b2, b3, B, M);

    for (int i = 0; i < 3; i++)
        for (int j = 0; j < 3; j++) {
            M[i][j] *= (1.0 + b2 + (b1 - b3) * b3);
            M[i][j] /= (1.0 + b1 - b2 + b3) * (1.0 - b1 - b2 - b3);
        }

    float tmp[H][8] ALIGNED16;
    vfloat Rv;
    vfloat Tv, Tm2v, Tm3v;
    vfloat Rv1;
    vfloat Tv1, Tm2v1, Tm3v1;
    vfloat Bv, b1v, b2v, b3v;
    vfloat temp2W, temp2Wp1;
    vfloat temp2W1, temp2Wp11;
    vfloat onev = F2V(1.f);
    Bv = F2V(B);
    b1v = F2V(b1);
    b2v = F2V(b2);
    b3v = F2V(b3);

#ifdef _OPENMP
    #pragma omp for nowait
#endif

    // process 8 columns per iteration for better usage of cpu cache
    for (int i = 0; i < W - 7; i += 8) {
        Tv = LVFU( src[0][i]);
        Tv1 = LVFU( src[0][i + 4]);
        Rv = Tv * (Bv + b1v + b2v + b3v);
        Rv1 = Tv1 * (Bv + b1v + b2v + b3v);
        Tm3v = Rv;
        Tm3v1 = Rv1;
        STVF( tmp[0][0], Rv );
        STVF( tmp[0][4], Rv1 );

        Rv = LVFU(src[1][i]) * Bv + Rv * b1v + Tv * (b2v + b3v);
        Rv1 = LVFU(src[1][i + 4]) * Bv + Rv1 * b1v + Tv1 * (b2v + b3v);
        Tm2v = Rv;
        Tm2v1 = Rv1;
        STVF( tmp[1][0], Rv );
        STVF( tmp[1][4], Rv1 );

        Rv = LVFU(src[2][i]) * Bv + Rv * b1v + Tm3v * b2v + Tv * b3v;
        Rv1 = LVFU(src[2][i + 4]) * Bv + Rv1 * b1v + Tm3v1 * b2v + Tv1 * b3v;
        STVF( tmp[2][0], Rv );
        STVF( tmp[2][4], Rv1 );

        for (int j = 3; j < H; j++) {
            Tv = Rv;
            Tv1 = Rv1;
            Rv = LVFU(src[j][i]) * Bv +  Tv * b1v + Tm2v * b2v + Tm3v * b3v;
            Rv1 = LVFU(src[j][i + 4]) * Bv +  Tv1 * b1v + Tm2v1 * b2v + Tm3v1 * b3v;
            STVF( tmp[j][0], Rv );
            STVF( tmp[j][4], Rv1 );
            Tm3v = Tm2v;
            Tm3v1 = Tm2v1;
            Tm2v = Tv;
            Tm2v1 = Tv1;
        }

        Tv = LVFU(src[H - 1][i]);
        Tv1 = LVFU(src[H - 1][i + 4]);

        temp2Wp1 = Tv + F2V(M[2][0]) * (Rv - Tv) + F2V(M[2][1]) * (Tm2v - Tv) + F2V(M[2][2]) * (Tm3v - Tv);
        temp2Wp11 = Tv1 + F2V(M[2][0]) * (Rv1 - Tv1) + F2V(M[2][1]) * (Tm2v1 - Tv1) + F2V(M[2][2]) * (Tm3v1 - Tv1);
        temp2W = Tv + F2V(M[1][0]) * (Rv - Tv) + F2V(M[1][1]) * (Tm2v - Tv) + F2V(M[1][2]) * (Tm3v - Tv);
        temp2W1 = Tv1 + F2V(M[1][0]) * (Rv1 - Tv1) + F2V(M[1][1]) * (Tm2v1 - Tv1) + F2V(M[1][2]) * (Tm3v1 - Tv1);

        Rv = Tv + F2V(M[0][0]) * (Rv - Tv) + F2V(M[0][1]) * (Tm2v - Tv) + F2V(M[0][2]) * (Tm3v - Tv);
        Rv1 = Tv1 + F2V(M[0][0]) * (Rv1 - Tv1) + F2V(M[0][1]) * (Tm2v1 - Tv1) + F2V(M[0][2]) * (Tm3v1 - Tv1);

        STVFU( dst[H - 1][i], LVFU(divBuffer[H - 1][i]) / vself(vmaskf_gt(Rv, ZEROV), Rv, onev));
        STVFU( dst[H - 1][i + 4], LVFU(divBuffer[H - 1][i + 4]) / vself(vmaskf_gt(Rv1, ZEROV), Rv1, onev));

        Tm2v = Bv * Tm2v + b1v * Rv + b2v * temp2W + b3v * temp2Wp1;
        Tm2v1 = Bv * Tm2v1 + b1v * Rv1 + b2v * temp2W1 + b3v * temp2Wp11;
        STVFU( dst[H - 2][i], LVFU(divBuffer[H - 2][i]) / vself(vmaskf_gt(Tm2v, ZEROV), Tm2v, onev));
        STVFU( dst[H - 2][i + 4], LVFU(divBuffer[H - 2][i + 4]) / vself(vmaskf_gt(Tm2v1, ZEROV), Tm2v1, onev));

        Tm3v = Bv * Tm3v + b1v * Tm2v + b2v * Rv + b3v * temp2W;
        Tm3v1 = Bv * Tm3v1 + b1v * Tm2v1 + b2v * Rv1 + b3v * temp2W1;
        STVFU( dst[H - 3][i], LVFU(divBuffer[H - 3][i]) / vself(vmaskf_gt(Tm3v, ZEROV), Tm3v, onev));
        STVFU( dst[H - 3][i + 4], LVFU(divBuffer[H - 3][i + 4]) / vself(vmaskf_gt(Tm3v1, ZEROV), Tm3v1, onev));

        Tv = Rv;
        Tv1 = Rv1;
        Rv = Tm3v;
        Rv1 = Tm3v1;
        Tm3v = Tv;
        Tm3v1 = Tv1;

        for (int j = H - 4; j >= 0; j--) {
            Tv = Rv;
            Tv1 = Rv1;
            Rv = LVF(tmp[j][0]) * Bv +  Tv * b1v + Tm2v * b2v + Tm3v * b3v;
            Rv1 = LVF(tmp[j][4]) * Bv +  Tv1 * b1v + Tm2v1 * b2v + Tm3v1 * b3v;
            STVFU( dst[j][i], vmaxf(LVFU(divBuffer[j][i]) / vself(vmaskf_gt(Rv, ZEROV), Rv, onev), ZEROV));
            STVFU( dst[j][i + 4], vmaxf(LVFU(divBuffer[j][i + 4]) / vself(vmaskf_gt(Rv1, ZEROV), Rv1, onev), ZEROV));
            Tm3v = Tm2v;
            Tm3v1 = Tm2v1;
            Tm2v = Tv;
            Tm2v1 = Tv1;
        }
    }

// Borders are done without SSE
#ifdef _OPENMP
    #pragma omp single
#endif

    for (int i = W - (W % 8); i < W; i++) {
        tmp[0][0] = src[0][i] * (B + b1 + b2 + b3);
        tmp[1][0] = B * src[1][i] + b1 * tmp[0][0] + src[0][i] * (b2 + b3);
        tmp[2][0] = B * src[2][i] + b1 * tmp[1][0] + b2 * tmp[0][0] + b3 * src[0][i];

        for (int j = 3; j < H; j++) {
            tmp[j][0] = B * src[j][i] + b1 * tmp[j - 1][0] + b2 * tmp[j - 2][0] + b3 * tmp[j - 3][0];
        }

        float temp2Hm1 = src[H - 1][i] + M[0][0] * (tmp[H - 1][0] - src[H - 1][i]) + M[0][1] * (tmp[H - 2][0] - src[H - 1][i]) + M[0][2] * (tmp[H - 3][0] - src[H - 1][i]);
        float temp2H   = src[H - 1][i] + M[1][0] * (tmp[H - 1][0] - src[H - 1][i]) + M[1][1] * (tmp[H - 2][0] - src[H - 1][i]) + M[1][2] * (tmp[H - 3][0] - src[H - 1][i]);
        float temp2Hp1 = src[H - 1][i] + M[2][0] * (tmp[H - 1][0] - src[H - 1][i]) + M[2][1] * (tmp[H - 2][0] - src[H - 1][i]) + M[2][2] * (tmp[H - 3][0] - src[H - 1][i]);

        tmp[H - 1][0] = temp2Hm1;
        tmp[H - 2][0] = B * tmp[H - 2][0] + b1 * tmp[H - 1][0] + b2 * temp2H + b3 * temp2Hp1;
        tmp[H - 3][0] = B * tmp[H - 3][0] + b1 * tmp[H - 2][0] + b2 * tmp[H - 1][0] + b3 * temp2H;

        for (int j = H - 4; j >= 0; j--) {
            tmp[j][0] = B * tmp[j][0] + b1 * tmp[j + 1][0] + b2 * tmp[j + 2][0] + b3 * tmp[j + 3][0];
        }

        for (int j = 0; j < H; j++) {
            dst[j][i] = rtengine::max(divBuffer[j][i] / (tmp[j][0] > 0.f ? tmp[j][0] : 1.f), 0.f);
        }

    }
}

#endif

template<class T> void gaussVertical (T** src, T** dst, const int W, const int H, const double sigma)
{
    double b1, b2, b3, B, M[3][3];
    calculateYvVFactors<double>(sigma, b1, b2, b3, B, M);

    for (int i = 0; i < 3; i++)
        for (int j = 0; j < 3; j++) {
            M[i][j] /= (1.0 + b1 - b2 + b3) * (1.0 + b2 + (b1 - b3) * b3);
        }

    // process 'numcols' columns for better usage of L1 cpu cache (especially faster for large values of H)
    static const int numcols = 8;
    std::vector<double> temp2(H * numcols);
    double temp2Hm1[numcols], temp2H[numcols], temp2Hp1[numcols];

    auto coord =
    [](const int &x, const int &y) -> int {
        return x * numcols + y;
    };
#ifdef _OPENMP
    #pragma omp for nowait
#endif

    for (unsigned int i = 0; i < static_cast<unsigned>(std::max(0, W - numcols + 1)); i += numcols) {
        for (int k = 0; k < numcols; k++) {
            temp2[coord(0, k)] = B * src[0][i + k] + b1 * src[0][i + k] + b2 * src[0][i + k] + b3 * src[0][i + k];
            temp2[coord(1, k)] = B * src[1][i + k] + b1 * temp2[coord(0, k)] + b2 * src[0][i + k] + b3 * src[0][i + k];
            temp2[coord(2, k)] = B * src[2][i + k] + b1 * temp2[coord(1, k)] + b2 * temp2[coord(0, k)] + b3 * src[0][i + k];
        }

        for (int j = 3; j < H; j++) {
            for (int k = 0; k < numcols; k++) {
                temp2[coord(j, k)] = B * src[j][i + k] + b1 * temp2[coord(j - 1, k)] + b2 * temp2[coord(j - 2, k)] + b3 * temp2[coord(j - 3, k)];
            }
        }

        for (int k = 0; k < numcols; k++) {
            temp2Hm1[k] = src[H - 1][i + k] + M[0][0] * (temp2[coord(H - 1, k)] - src[H - 1][i + k]) + M[0][1] * (temp2[coord(H - 2, k)] - src[H - 1][i + k]) + M[0][2] * (temp2[coord(H - 3, k)] - src[H - 1][i + k]);
            temp2H[k]   = src[H - 1][i + k] + M[1][0] * (temp2[coord(H - 1, k)] - src[H - 1][i + k]) + M[1][1] * (temp2[coord(H - 2, k)] - src[H - 1][i + k]) + M[1][2] * (temp2[coord(H - 3, k)] - src[H - 1][i + k]);
            temp2Hp1[k] = src[H - 1][i + k] + M[2][0] * (temp2[coord(H - 1, k)] - src[H - 1][i + k]) + M[2][1] * (temp2[coord(H - 2, k)] - src[H - 1][i + k]) + M[2][2] * (temp2[coord(H - 3, k)] - src[H - 1][i + k]);
        }

        for (int k = 0; k < numcols; k++) {
            dst[H - 1][i + k] = temp2[coord(H - 1, k)] = temp2Hm1[k];
            dst[H - 2][i + k] = temp2[coord(H - 2, k)] = B * temp2[coord(H - 2, k)] + b1 * temp2[coord(H - 1, k)] + b2 * temp2H[k] + b3 * temp2Hp1[k];
            dst[H - 3][i + k] = temp2[coord(H - 3, k)] = B * temp2[coord(H - 3, k)] + b1 * temp2[coord(H - 2, k)] + b2 * temp2[coord(H - 1, k)] + b3 * temp2H[k];
        }

        for (int j = H - 4; j >= 0; j--) {
            for (int k = 0; k < numcols; k++) {
                dst[j][i + k] = temp2[coord(j, k)] = B * temp2[coord(j, k)] + b1 * temp2[coord(j + 1, k)] + b2 * temp2[coord(j + 2, k)] + b3 * temp2[coord(j + 3, k)];
            }
        }
    }

#ifdef _OPENMP
    #pragma omp single
#endif

    // process remaining columns
    for (int i = W - (W % numcols); i < W; i++) {
        temp2[coord(0, 0)] = B * src[0][i] + b1 * src[0][i] + b2 * src[0][i] + b3 * src[0][i];
        temp2[coord(1, 0)] = B * src[1][i] + b1 * temp2[coord(0, 0)]  + b2 * src[0][i] + b3 * src[0][i];
        temp2[coord(2, 0)] = B * src[2][i] + b1 * temp2[coord(1, 0)]  + b2 * temp2[coord(0, 0)]  + b3 * src[0][i];

        for (int j = 3; j < H; j++) {
            temp2[coord(j, 0)] = B * src[j][i] + b1 * temp2[coord(j - 1, 0)] + b2 * temp2[coord(j - 2, 0)] + b3 * temp2[coord(j - 3, 0)];
        }

        double temp2Hm1 = src[H - 1][i] + M[0][0] * (temp2[coord(H - 1, 0)] - src[H - 1][i]) + M[0][1] * (temp2[coord(H - 2, 0)] - src[H - 1][i]) + M[0][2] * (temp2[coord(H - 3, 0)] - src[H - 1][i]);
        double temp2H   = src[H - 1][i] + M[1][0] * (temp2[coord(H - 1, 0)] - src[H - 1][i]) + M[1][1] * (temp2[coord(H - 2, 0)] - src[H - 1][i]) + M[1][2] * (temp2[coord(H - 3, 0)] - src[H - 1][i]);
        double temp2Hp1 = src[H - 1][i] + M[2][0] * (temp2[coord(H - 1, 0)] - src[H - 1][i]) + M[2][1] * (temp2[coord(H - 2, 0)] - src[H - 1][i]) + M[2][2] * (temp2[coord(H - 3, 0)] - src[H - 1][i]);

        dst[H - 1][i] = temp2[coord(H - 1, 0)] = temp2Hm1;
        dst[H - 2][i] = temp2[coord(H - 2, 0)] = B * temp2[coord(H - 2, 0)] + b1 * temp2[coord(H - 1, 0)] + b2 * temp2H + b3 * temp2Hp1;
        dst[H - 3][i] = temp2[coord(H - 3, 0)] = B * temp2[coord(H - 3, 0)] + b1 * temp2[coord(H - 2, 0)] + b2 * temp2[coord(H - 1, 0)] + b3 * temp2H;

        for (int j = H - 4; j >= 0; j--) {
            dst[j][i] = temp2[coord(j, 0)] = B * temp2[coord(j, 0)] + b1 * temp2[coord(j + 1, 0)] + b2 * temp2[coord(j + 2, 0)] + b3 * temp2[coord(j + 3, 0)];
        }
    }
}

#ifndef __SSE2__
template<class T> void gaussVerticaldiv (T** src, T** dst, T** divBuffer, const int W, const int H, const double sigma)
{
    double b1, b2, b3, B, M[3][3];
    calculateYvVFactors<double>(sigma, b1, b2, b3, B, M);

    for (int i = 0; i < 3; i++)
        for (int j = 0; j < 3; j++) {
            M[i][j] /= (1.0 + b1 - b2 + b3) * (1.0 + b2 + (b1 - b3) * b3);
        }

    // process 'numcols' columns for better usage of L1 cpu cache (especially faster for large values of H)
    static const int numcols = 8;
    std::vector<double> temp2(H * numcols);
    double temp2Hm1[numcols], temp2H[numcols], temp2Hp1[numcols];

    auto coord =
    [](const int &x, const int &y) -> int {
        return x * numcols + y;
    };
#ifdef _OPENMP
    #pragma omp for nowait
#endif

    for (int i = 0; i < W - numcols + 1; i += numcols) {
        for (int k = 0; k < numcols; k++) {
            temp2[coord(0, k)] = B * src[0][i + k] + b1 * src[0][i + k] + b2 * src[0][i + k] + b3 * src[0][i + k];
            temp2[coord(1, k)] = B * src[1][i + k] + b1 * temp2[coord(0, k)] + b2 * src[0][i + k] + b3 * src[0][i + k];
            temp2[coord(2, k)] = B * src[2][i + k] + b1 * temp2[coord(1, k)] + b2 * temp2[coord(0, k)] + b3 * src[0][i + k];
        }

        for (int j = 3; j < H; j++) {
            for (int k = 0; k < numcols; k++) {
                temp2[coord(j, k)] = B * src[j][i + k] + b1 * temp2[coord(j - 1, k)] + b2 * temp2[coord(j - 2, k)] + b3 * temp2[coord(j - 3, k)];
            }
        }

        for (int k = 0; k < numcols; k++) {
            temp2Hm1[k] = src[H - 1][i + k] + M[0][0] * (temp2[coord(H - 1, k)] - src[H - 1][i + k]) + M[0][1] * (temp2[coord(H - 2, k)] - src[H - 1][i + k]) + M[0][2] * (temp2[coord(H - 3, k)] - src[H - 1][i + k]);
            temp2H[k]   = src[H - 1][i + k] + M[1][0] * (temp2[coord(H - 1, k)] - src[H - 1][i + k]) + M[1][1] * (temp2[coord(H - 2, k)] - src[H - 1][i + k]) + M[1][2] * (temp2[coord(H - 3, k)] - src[H - 1][i + k]);
            temp2Hp1[k] = src[H - 1][i + k] + M[2][0] * (temp2[coord(H - 1, k)] - src[H - 1][i + k]) + M[2][1] * (temp2[coord(H - 2, k)] - src[H - 1][i + k]) + M[2][2] * (temp2[coord(H - 3, k)] - src[H - 1][i + k]);
        }

        for (int k = 0; k < numcols; k++) {
            dst[H - 1][i + k] = rtengine::max(divBuffer[H - 1][i + k] / (temp2[coord(H - 1, k)] = temp2Hm1[k]), 0.0);
            dst[H - 2][i + k] = rtengine::max(divBuffer[H - 2][i + k] / (temp2[coord(H - 2, k)] = B * temp2[coord(H - 2, k)] + b1 * temp2[coord(H - 1, k)] + b2 * temp2H[k] + b3 * temp2Hp1[k]), 0.0);
            dst[H - 3][i + k] = rtengine::max(divBuffer[H - 3][i + k] / (temp2[coord(H - 3, k)] = B * temp2[coord(H - 3, k)] + b1 * temp2[coord(H - 2, k)] + b2 * temp2[coord(H - 1, k)] + b3 * temp2H[k]), 0.0);
        }

        for (int j = H - 4; j >= 0; j--) {
            for (int k = 0; k < numcols; k++) {
                dst[j][i + k] = rtengine::max(divBuffer[j][i + k] / (temp2[coord(j, k)] = B * temp2[coord(j, k)] + b1 * temp2[coord(j + 1, k)] + b2 * temp2[coord(j + 2, k)] + b3 * temp2[coord(j + 3, k)]), 0.0);
            }
        }
    }

#ifdef _OPENMP
    #pragma omp single
#endif

    // process remaining columns
    for (int i = W - (W % numcols); i < W; i++) {
        temp2[coord(0, 0)] = B * src[0][i] + b1 * src[0][i] + b2 * src[0][i] + b3 * src[0][i];
        temp2[coord(1, 0)] = B * src[1][i] + b1 * temp2[coord(0, 0)]  + b2 * src[0][i] + b3 * src[0][i];
        temp2[coord(2, 0)] = B * src[2][i] + b1 * temp2[coord(1, 0)]  + b2 * temp2[coord(0, 0)]  + b3 * src[0][i];

        for (int j = 3; j < H; j++) {
            temp2[coord(j, 0)] = B * src[j][i] + b1 * temp2[coord(j - 1, 0)] + b2 * temp2[coord(j - 2, 0)] + b3 * temp2[coord(j - 3, 0)];
        }

        double temp2Hm1 = src[H - 1][i] + M[0][0] * (temp2[coord(H - 1, 0)] - src[H - 1][i]) + M[0][1] * (temp2[coord(H - 2, 0)] - src[H - 1][i]) + M[0][2] * (temp2[coord(H - 3, 0)] - src[H - 1][i]);
        double temp2H   = src[H - 1][i] + M[1][0] * (temp2[coord(H - 1, 0)] - src[H - 1][i]) + M[1][1] * (temp2[coord(H - 2, 0)] - src[H - 1][i]) + M[1][2] * (temp2[coord(H - 3, 0)] - src[H - 1][i]);
        double temp2Hp1 = src[H - 1][i] + M[2][0] * (temp2[coord(H - 1, 0)] - src[H - 1][i]) + M[2][1] * (temp2[coord(H - 2, 0)] - src[H - 1][i]) + M[2][2] * (temp2[coord(H - 3, 0)] - src[H - 1][i]);

        dst[H - 1][i] = rtengine::max(divBuffer[H - 1][i] / (temp2[coord(H - 1, 0)] = temp2Hm1), 0.0);
        dst[H - 2][i] = rtengine::max(divBuffer[H - 2][i] / (temp2[coord(H - 2, 0)] = B * temp2[coord(H - 2, 0)] + b1 * temp2[coord(H - 1, 0)] + b2 * temp2H + b3 * temp2Hp1), 0.0);
        dst[H - 3][i] = rtengine::max(divBuffer[H - 3][i] / (temp2[coord(H - 3, 0)] = B * temp2[coord(H - 3, 0)] + b1 * temp2[coord(H - 2, 0)] + b2 * temp2[coord(H - 1, 0)] + b3 * temp2H), 0.0);

        for (int j = H - 4; j >= 0; j--) {
            dst[j][i] = rtengine::max(divBuffer[j][i] / (temp2[coord(j, 0)] = B * temp2[coord(j, 0)] + b1 * temp2[coord(j + 1, 0)] + b2 * temp2[coord(j + 2, 0)] + b3 * temp2[coord(j + 3, 0)]), 0.0);
        }
    }
}

template<class T> void gaussVerticalmult (T** src, T** dst, const int W, const int H, const double sigma)
{
    double b1, b2, b3, B, M[3][3];
    calculateYvVFactors<double>(sigma, b1, b2, b3, B, M);

    for (int i = 0; i < 3; i++)
        for (int j = 0; j < 3; j++) {
            M[i][j] /= (1.0 + b1 - b2 + b3) * (1.0 + b2 + (b1 - b3) * b3);
        }

    // process 'numcols' columns for better usage of L1 cpu cache (especially faster for large values of H)
    static const int numcols = 8;
    std::vector<double> temp2(H * numcols);
    double temp2Hm1[numcols], temp2H[numcols], temp2Hp1[numcols];

    auto coord =
    [](const int &x, const int &y) -> int {
        return x * numcols + y;
    };
#ifdef _OPENMP
    #pragma omp for nowait
#endif

    for (int i = 0; i < W - numcols + 1; i += numcols) {
        for (int k = 0; k < numcols; k++) {
            temp2[coord(0, k)] = B * src[0][i + k] + b1 * src[0][i + k] + b2 * src[0][i + k] + b3 * src[0][i + k];
            temp2[coord(1, k)] = B * src[1][i + k] + b1 * temp2[coord(0, k)] + b2 * src[0][i + k] + b3 * src[0][i + k];
            temp2[coord(2, k)] = B * src[2][i + k] + b1 * temp2[coord(1, k)] + b2 * temp2[coord(0, k)] + b3 * src[0][i + k];
        }

        for (int j = 3; j < H; j++) {
            for (int k = 0; k < numcols; k++) {
                temp2[coord(j, k)] = B * src[j][i + k] + b1 * temp2[coord(j - 1, k)] + b2 * temp2[coord(j - 2, k)] + b3 * temp2[coord(j - 3, k)];
            }
        }

        for (int k = 0; k < numcols; k++) {
            temp2Hm1[k] = src[H - 1][i + k] + M[0][0] * (temp2[coord(H - 1, k)] - src[H - 1][i + k]) + M[0][1] * (temp2[coord(H - 2, k)] - src[H - 1][i + k]) + M[0][2] * (temp2[coord(H - 3, k)] - src[H - 1][i + k]);
            temp2H[k]   = src[H - 1][i + k] + M[1][0] * (temp2[coord(H - 1, k)] - src[H - 1][i + k]) + M[1][1] * (temp2[coord(H - 2, k)] - src[H - 1][i + k]) + M[1][2] * (temp2[coord(H - 3, k)] - src[H - 1][i + k]);
            temp2Hp1[k] = src[H - 1][i + k] + M[2][0] * (temp2[coord(H - 1, k)] - src[H - 1][i + k]) + M[2][1] * (temp2[coord(H - 2, k)] - src[H - 1][i + k]) + M[2][2] * (temp2[coord(H - 3, k)] - src[H - 1][i + k]);
        }

        for (int k = 0; k < numcols; k++) {
            dst[H - 1][i + k] *= temp2[coord(H - 1, k)] = temp2Hm1[k];
            dst[H - 2][i + k] *= temp2[coord(H - 2, k)] = B * temp2[coord(H - 2, k)] + b1 * temp2[coord(H - 1, k)] + b2 * temp2H[k] + b3 * temp2Hp1[k];
            dst[H - 3][i + k] *= temp2[coord(H - 3, k)] = B * temp2[coord(H - 3, k)] + b1 * temp2[coord(H - 2, k)] + b2 * temp2[coord(H - 1, k)] + b3 * temp2H[k];
        }

        for (int j = H - 4; j >= 0; j--) {
            for (int k = 0; k < numcols; k++) {
                dst[j][i + k] *= (temp2[coord(j, k)] = B * temp2[coord(j, k)] + b1 * temp2[coord(j + 1, k)] + b2 * temp2[coord(j + 2, k)] + b3 * temp2[coord(j + 3, k)]);
            }
        }
    }

#ifdef _OPENMP
    #pragma omp single
#endif

    // process remaining columns
    for (int i = W - (W % numcols); i < W; i++) {
        temp2[coord(0, 0)] = B * src[0][i] + b1 * src[0][i] + b2 * src[0][i] + b3 * src[0][i];
        temp2[coord(1, 0)] = B * src[1][i] + b1 * temp2[coord(0, 0)]  + b2 * src[0][i] + b3 * src[0][i];
        temp2[coord(2, 0)] = B * src[2][i] + b1 * temp2[coord(1, 0)]  + b2 * temp2[coord(0, 0)]  + b3 * src[0][i];

        for (int j = 3; j < H; j++) {
            temp2[coord(j, 0)] = B * src[j][i] + b1 * temp2[coord(j - 1, 0)] + b2 * temp2[coord(j - 2, 0)] + b3 * temp2[coord(j - 3, 0)];
        }

        double temp2Hm1 = src[H - 1][i] + M[0][0] * (temp2[coord(H - 1, 0)] - src[H - 1][i]) + M[0][1] * (temp2[coord(H - 2, 0)] - src[H - 1][i]) + M[0][2] * (temp2[coord(H - 3, 0)] - src[H - 1][i]);
        double temp2H   = src[H - 1][i] + M[1][0] * (temp2[coord(H - 1, 0)] - src[H - 1][i]) + M[1][1] * (temp2[coord(H - 2, 0)] - src[H - 1][i]) + M[1][2] * (temp2[coord(H - 3, 0)] - src[H - 1][i]);
        double temp2Hp1 = src[H - 1][i] + M[2][0] * (temp2[coord(H - 1, 0)] - src[H - 1][i]) + M[2][1] * (temp2[coord(H - 2, 0)] - src[H - 1][i]) + M[2][2] * (temp2[coord(H - 3, 0)] - src[H - 1][i]);

        dst[H - 1][i] *= temp2[coord(H - 1, 0)] = temp2Hm1;
        dst[H - 2][i] *= temp2[coord(H - 2, 0)] = B * temp2[coord(H - 2, 0)] + b1 * temp2[coord(H - 1, 0)] + b2 * temp2H + b3 * temp2Hp1;
        dst[H - 3][i] *= temp2[coord(H - 3, 0)] = B * temp2[coord(H - 3, 0)] + b1 * temp2[coord(H - 2, 0)] + b2 * temp2[coord(H - 1, 0)] + b3 * temp2H;

        for (int j = H - 4; j >= 0; j--) {
            dst[j][i] *= (temp2[coord(j, 0)] = B * temp2[coord(j, 0)] + b1 * temp2[coord(j + 1, 0)] + b2 * temp2[coord(j + 2, 0)] + b3 * temp2[coord(j + 3, 0)]);
        }
    }
}
#endif

template<class T> void gaussianBlurImpl(T** src, T** dst, const int W, const int H, const double sigma, bool useBoxBlur, eGaussType gausstype = GAUSS_STANDARD, T** buffer2 = nullptr)
{
    static constexpr auto GAUSS_3X3_LIMIT = 0.6;
    static constexpr auto GAUSS_5X5_LIMIT = 0.84;
    static constexpr auto GAUSS_7X7_LIMIT = 1.15;
    static constexpr auto GAUSS_DOUBLE = 25.0;

    if (useBoxBlur) {
        // special variant for very large sigma, currently only used by retinex algorithm
        // use iterated boxblur to approximate gaussian blur
        // Compute ideal averaging filter width and number of iterations
        int n = 1;
        double wIdeal = sqrt((12 * sigma * sigma) + 1);

        while(wIdeal > W || wIdeal > H) {
            n++;
            wIdeal = sqrt((12 * sigma * sigma / n) + 1);
        }

        if(n < 3) {
            n = 3;
            wIdeal = sqrt((12 * sigma * sigma / n) + 1);
        } else if(n > 6) {
            n = 6;
        }

        int wl = wIdeal;

        if(wl % 2 == 0) {
            wl--;
        }

        int wu = wl + 2;

        double mIdeal = (12 * sigma * sigma - n * wl * wl - 4 * n * wl - 3 * n) / (-4 * wl - 4);
        int m = round(mIdeal);

        int sizes[n];

        for(int i = 0; i < n; i++) {
            sizes[i] = ((i < m ? wl : wu) - 1) / 2;
        }

        rtengine::boxblur(src, dst, sizes[0], W, H, true);

        for(int i = 1; i < n; i++) {
            rtengine::boxblur(dst, dst, sizes[i], W, H, true);
        }
    } else {
        if (sigma < GAUSS_SKIP) {
            // don't perform filtering
#ifdef _OPENMP
#pragma omp single
#endif
            if (src != dst) {
                for(int i = 0; i < H; ++i) {
                    memcpy(dst[i], src[i], W * sizeof(T));
                }
            }
        } else if (sigma < GAUSS_3X3_LIMIT) {
            if(src != dst) {
                // If src != dst we can take the fast way
                // compute 3x3 kernel values
                double c0 = 1.0;
                double c1 = exp( -0.5 * (rtengine::SQR(1.0 / sigma)) );
                double c2 = exp( -rtengine::SQR(1.0 / sigma) );

                // normalize kernel values
                double sum = c0 + 4.0 * (c1 + c2);
                c0 /= sum;
                c1 /= sum;
                c2 /= sum;
                // compute kernel values for border pixels
                double b1 = exp (-1.0 / (2.0 * sigma * sigma));
                double bsum = 2.0 * b1 + 1.0;
                b1 /= bsum;
                double b0 = 1.0 / bsum;

                switch (gausstype) {
                case GAUSS_MULT     :
                    gauss3x3mult<T> (src, dst, W, H, c0, c1, c2, b0, b1);
                    break;

                case GAUSS_DIV      :
                    gauss3x3div<T> (src, dst, buffer2, W, H, c0, c1, c2, b0, b1);
                    break;

                case GAUSS_STANDARD :
                    gauss3x3<T> (src, dst, W, H, c0, c1, c2, b0, b1);
                    break;
                }
            } else {
                // compute kernel values for separated 3x3 gaussian blur
                double c1 = exp (-1.0 / (2.0 * sigma * sigma));
                double csum = 2.0 * c1 + 1.0;
                c1 /= csum;
                double c0 = 1.0 / csum;
                gaussHorizontal3<T> (src, dst, W, H, c0, c1);
                gaussVertical3<T>   (dst, dst, W, H, c0, c1);
            }
        } else {
#ifdef __SSE2__

            if (sigma < GAUSS_DOUBLE) {
                switch (gausstype) {
                case GAUSS_MULT : {
                    if (sigma <= GAUSS_5X5_LIMIT && src != dst) {
                        gauss5x5mult(src, dst, W, H, sigma);
                    } else if (sigma <= GAUSS_7X7_LIMIT && src != dst) {
                        gauss7x7mult(src, dst, W, H, sigma);
                    } else {
                        gaussHorizontalSse<T> (src, src, W, H, sigma);
                        gaussVerticalSsemult<T> (src, dst, W, H, sigma);
                    }
                    break;
                }

                case GAUSS_DIV : {
                    if (sigma <= GAUSS_5X5_LIMIT && src != dst) {
                        gauss5x5div (src, dst, buffer2, W, H, sigma);
                    } else if (sigma <= GAUSS_7X7_LIMIT && src != dst) {
                        gauss7x7div (src, dst, buffer2, W, H, sigma);
                    } else {
                        gaussHorizontalSse<T> (src, dst, W, H, sigma);
                        gaussVerticalSsediv<T> (dst, dst, buffer2, W, H, sigma);
                    }
                    break;
                }

                case GAUSS_STANDARD : {
                    gaussHorizontalSse<T> (src, dst, W, H, sigma);
                    gaussVerticalSse<T> (dst, dst, W, H, sigma);
                    break;
                }
                }
            } else { // large sigma only with double precision
                gaussHorizontal<T> (src, dst, W, H, sigma);
                gaussVertical<T>   (dst, dst, W, H, sigma);
            }

#else

            if (sigma < GAUSS_DOUBLE) {
                switch (gausstype) {
                case GAUSS_MULT : {
                    if (sigma <= GAUSS_5X5_LIMIT && src != dst) {
                        gauss5x5mult(src, dst, W, H, sigma);
                    } else if (sigma <= GAUSS_7X7_LIMIT && src != dst) {
                        gauss7x7mult(src, dst, W, H, sigma);
                    } else {
                        gaussHorizontal<T> (src, src, W, H, sigma);
                        gaussVerticalmult<T> (src, dst, W, H, sigma);
                    }
                    break;
                }

                case GAUSS_DIV : {
                    if (sigma <= GAUSS_5X5_LIMIT && src != dst) {
                        gauss5x5div (src, dst, buffer2, W, H, sigma);
                    } else if (sigma <= GAUSS_7X7_LIMIT && src != dst) {
                        gauss7x7div (src, dst, buffer2, W, H, sigma);
                    } else {
                        gaussHorizontal<T> (src, dst, W, H, sigma);
                        gaussVerticaldiv<T> (dst, dst, buffer2, W, H, sigma);
                    }
                    break;
                }

                case GAUSS_STANDARD : {
                    gaussHorizontal<T> (src, dst, W, H, sigma);
                    gaussVertical<T> (dst, dst, W, H, sigma);
                    break;
                }
                }
            } else { // large sigma only with double precision
                gaussHorizontal<T> (src, dst, W, H, sigma);
                gaussVertical<T>   (dst, dst, W, H, sigma);
            }

#endif
        }
    }
}
}

void gaussianBlur(float** src, float** dst, const int W, const int H, const double sigma, bool useBoxBlur, eGaussType gausstype, float** buffer2)
{
    gaussianBlurImpl<float>(src, dst, W, H, sigma, useBoxBlur, gausstype, buffer2);
}