kdefx Library API Documentation
	Overview Class Hierarchy Classes Classes (annotated) Members Namespaces Source Files

kimageeffect.cpp

/* This file is part of the KDE libraries
    Copyright (C) 1998, 1999, 2001, 2002 Daniel M. Duley <mosfet@kde.org>
    (C) 1998, 1999 Christian Tibirna <ctibirna@total.net>
    (C) 1998, 1999 Dirk A. Mueller <mueller@kde.org>
    (C) 1999 Geert Jansen <g.t.jansen@stud.tue.nl>
    (C) 2000 Josef Weidendorfer <weidendo@in.tum.de>

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:

1. Redistributions of source code must retain the above copyright
   notice, this list of conditions and the following disclaimer.
2. Redistributions 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.

THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``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 AUTHOR 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.

*/

// $Id: kimageeffect.cpp,v 1.50.2.2 2004/06/28 14:39:15 lukas Exp $

#include <math.h>
#include <assert.h>

#include <qimage.h>
#include <stdlib.h>
#include <iostream>

#include "kimageeffect.h"
#include "kcpuinfo.h"

#include <config.h>

#if defined(__i386__) && ( defined(__GNUC__) || defined(__INTEL_COMPILER) )
#  if defined( HAVE_X86_MMX )
#    define USE_MMX_INLINE_ASM
#  endif
#  if defined( HAVE_X86_SSE2 )
#    define USE_SSE2_INLINE_ASM
#  endif
#endif

//======================================================================
//
// Utility stuff for effects ported from ImageMagick to QImage
//
//======================================================================
#define MaxRGB 255L
#define DegreesToRadians(x) ((x)*M_PI/180.0)
#define MagickSQ2PI 2.50662827463100024161235523934010416269302368164062
#define MagickEpsilon  1.0e-12
#define MagickPI  3.14159265358979323846264338327950288419716939937510

static inline unsigned int intensityValue(unsigned int color)
{
    return((unsigned int)((0.299*qRed(color) +
                           0.587*qGreen(color) +
                           0.1140000000000001*qBlue(color))));
}

static inline void liberateMemory(void **memory)
{
    assert(memory != (void **)NULL);
    if(*memory == (void *)NULL) return;
    free(*memory);
    *memory=(void *) NULL;
}

struct double_packet
{
    double red;
    double green;
    double blue;
    double alpha;
};

struct short_packet
{
    unsigned short int red;
    unsigned short int green;
    unsigned short int blue;
    unsigned short int alpha;
};


//======================================================================
//
// Gradient effects
//
//======================================================================

QImage KImageEffect::gradient(const QSize &size, const QColor &ca,
    const QColor &cb, GradientType eff, int ncols)
{
    int rDiff, gDiff, bDiff;
    int rca, gca, bca, rcb, gcb, bcb;

    QImage image(size, 32);

    if (size.width() == 0 || size.height() == 0) {
#ifndef NDEBUG
      std::cerr << "WARNING: KImageEffect::gradient: invalid image" << std::endl;
#endif
      return image;
    }

    register int x, y;

    rDiff = (rcb = cb.red())   - (rca = ca.red());
    gDiff = (gcb = cb.green()) - (gca = ca.green());
    bDiff = (bcb = cb.blue())  - (bca = ca.blue());

    if( eff == VerticalGradient || eff == HorizontalGradient ){

        uint *p;
        uint rgb;

        register int rl = rca << 16;
        register int gl = gca << 16;
        register int bl = bca << 16;

        if( eff == VerticalGradient ) {

            int rcdelta = ((1<<16) / size.height()) * rDiff;
            int gcdelta = ((1<<16) / size.height()) * gDiff;
            int bcdelta = ((1<<16) / size.height()) * bDiff;

            for ( y = 0; y < size.height(); y++ ) {
                p = (uint *) image.scanLine(y);

                rl += rcdelta;
                gl += gcdelta;
                bl += bcdelta;

                rgb = qRgb( (rl>>16), (gl>>16), (bl>>16) );

                for( x = 0; x < size.width(); x++ ) {
                    *p = rgb;
                    p++;
                }
            }

        }
        else {                  // must be HorizontalGradient

            unsigned int *o_src = (unsigned int *)image.scanLine(0);
            unsigned int *src = o_src;

            int rcdelta = ((1<<16) / size.width()) * rDiff;
            int gcdelta = ((1<<16) / size.width()) * gDiff;
            int bcdelta = ((1<<16) / size.width()) * bDiff;

            for( x = 0; x < size.width(); x++) {

                rl += rcdelta;
                gl += gcdelta;
                bl += bcdelta;

                *src++ = qRgb( (rl>>16), (gl>>16), (bl>>16));
            }

            src = o_src;

            // Believe it or not, manually copying in a for loop is faster
            // than calling memcpy for each scanline (on the order of ms...).
            // I think this is due to the function call overhead (mosfet).

            for (y = 1; y < size.height(); ++y) {

                p = (unsigned int *)image.scanLine(y);
                src = o_src;
                for(x=0; x < size.width(); ++x)
                    *p++ = *src++;
            }
        }
    }

    else {

        float rfd, gfd, bfd;
        float rd = rca, gd = gca, bd = bca;

        unsigned char *xtable[3];
        unsigned char *ytable[3];

        unsigned int w = size.width(), h = size.height();
        xtable[0] = new unsigned char[w];
        xtable[1] = new unsigned char[w];
        xtable[2] = new unsigned char[w];
        ytable[0] = new unsigned char[h];
        ytable[1] = new unsigned char[h];
        ytable[2] = new unsigned char[h];
        w*=2, h*=2;

        if ( eff == DiagonalGradient || eff == CrossDiagonalGradient) {
            // Diagonal dgradient code inspired by BlackBox (mosfet)
            // BlackBox dgradient is (C) Brad Hughes, <bhughes@tcac.net> and
            // Mike Cole <mike@mydot.com>.

            rfd = (float)rDiff/w;
            gfd = (float)gDiff/w;
            bfd = (float)bDiff/w;

            int dir;
            for (x = 0; x < size.width(); x++, rd+=rfd, gd+=gfd, bd+=bfd) {
                dir = eff == DiagonalGradient? x : size.width() - x - 1;
                xtable[0][dir] = (unsigned char) rd;
                xtable[1][dir] = (unsigned char) gd;
                xtable[2][dir] = (unsigned char) bd;
            }
            rfd = (float)rDiff/h;
            gfd = (float)gDiff/h;
            bfd = (float)bDiff/h;
            rd = gd = bd = 0;
            for (y = 0; y < size.height(); y++, rd+=rfd, gd+=gfd, bd+=bfd) {
                ytable[0][y] = (unsigned char) rd;
                ytable[1][y] = (unsigned char) gd;
                ytable[2][y] = (unsigned char) bd;
            }

            for (y = 0; y < size.height(); y++) {
                unsigned int *scanline = (unsigned int *)image.scanLine(y);
                for (x = 0; x < size.width(); x++) {
                    scanline[x] = qRgb(xtable[0][x] + ytable[0][y],
                                       xtable[1][x] + ytable[1][y],
                                       xtable[2][x] + ytable[2][y]);
                }
            }
        }

        else if (eff == RectangleGradient ||
                 eff == PyramidGradient ||
                 eff == PipeCrossGradient ||
                 eff == EllipticGradient)
        {
            int rSign = rDiff>0? 1: -1;
            int gSign = gDiff>0? 1: -1;
            int bSign = bDiff>0? 1: -1;

            rfd = (float)rDiff / size.width();
            gfd = (float)gDiff / size.width();
            bfd = (float)bDiff / size.width();

            rd = (float)rDiff/2;
            gd = (float)gDiff/2;
            bd = (float)bDiff/2;

            for (x = 0; x < size.width(); x++, rd-=rfd, gd-=gfd, bd-=bfd)
            {
                xtable[0][x] = (unsigned char) abs((int)rd);
                xtable[1][x] = (unsigned char) abs((int)gd);
                xtable[2][x] = (unsigned char) abs((int)bd);
            }

            rfd = (float)rDiff/size.height();
            gfd = (float)gDiff/size.height();
            bfd = (float)bDiff/size.height();

            rd = (float)rDiff/2;
            gd = (float)gDiff/2;
            bd = (float)bDiff/2;

            for (y = 0; y < size.height(); y++, rd-=rfd, gd-=gfd, bd-=bfd)
            {
                ytable[0][y] = (unsigned char) abs((int)rd);
                ytable[1][y] = (unsigned char) abs((int)gd);
                ytable[2][y] = (unsigned char) abs((int)bd);
            }
            unsigned int rgb;
            int h = (size.height()+1)>>1;
            for (y = 0; y < h; y++) {
                unsigned int *sl1 = (unsigned int *)image.scanLine(y);
                unsigned int *sl2 = (unsigned int *)image.scanLine(QMAX(size.height()-y-1, y));

                int w = (size.width()+1)>>1;
                int x2 = size.width()-1;

                for (x = 0; x < w; x++, x2--) {
            rgb = 0;
                    if (eff == PyramidGradient) {
                        rgb = qRgb(rcb-rSign*(xtable[0][x]+ytable[0][y]),
                                   gcb-gSign*(xtable[1][x]+ytable[1][y]),
                                   bcb-bSign*(xtable[2][x]+ytable[2][y]));
                    }
                    if (eff == RectangleGradient) {
                        rgb = qRgb(rcb - rSign *
                                   QMAX(xtable[0][x], ytable[0][y]) * 2,
                                   gcb - gSign *
                                   QMAX(xtable[1][x], ytable[1][y]) * 2,
                                   bcb - bSign *
                                   QMAX(xtable[2][x], ytable[2][y]) * 2);
                    }
                    if (eff == PipeCrossGradient) {
                        rgb = qRgb(rcb - rSign *
                                   QMIN(xtable[0][x], ytable[0][y]) * 2,
                                   gcb - gSign *
                                   QMIN(xtable[1][x], ytable[1][y]) * 2,
                                   bcb - bSign *
                                   QMIN(xtable[2][x], ytable[2][y]) * 2);
                    }
                    if (eff == EllipticGradient) {
                        rgb = qRgb(rcb - rSign *
                                   (int)sqrt((xtable[0][x]*xtable[0][x] +
                                              ytable[0][y]*ytable[0][y])*2.0),
                                   gcb - gSign *
                                   (int)sqrt((xtable[1][x]*xtable[1][x] +
                                              ytable[1][y]*ytable[1][y])*2.0),
                                   bcb - bSign *
                                   (int)sqrt((xtable[2][x]*xtable[2][x] +
                                              ytable[2][y]*ytable[2][y])*2.0));
                    }

                    sl1[x] = sl2[x] = rgb;
                    sl1[x2] = sl2[x2] = rgb;
                }
            }
        }

        delete [] xtable[0];
        delete [] xtable[1];
        delete [] xtable[2];
        delete [] ytable[0];
        delete [] ytable[1];
        delete [] ytable[2];
    }

    // dither if necessary
    if (ncols && (QPixmap::defaultDepth() < 15 )) {
    if ( ncols < 2 || ncols > 256 )
        ncols = 3;
    QColor *dPal = new QColor[ncols];
    for (int i=0; i<ncols; i++) {
        dPal[i].setRgb ( rca + rDiff * i / ( ncols - 1 ),
                 gca + gDiff * i / ( ncols - 1 ),
                 bca + bDiff * i / ( ncols - 1 ) );
    }
    dither(image, dPal, ncols);
    delete [] dPal;
    }

    return image;
}


// -----------------------------------------------------------------------------

//CT this was (before Dirk A. Mueller's speedup changes)
//   merely the same code as in the above method, but it's supposedly
//   way less performant since it introduces a lot of supplementary tests
//   and simple math operations for the calculus of the balance.
//      (surprizingly, it isn't less performant, in the contrary :-)
//   Yes, I could have merged them, but then the excellent performance of
//   the balanced code would suffer with no other gain than a mere
//   source code and byte code size economy.

QImage KImageEffect::unbalancedGradient(const QSize &size, const QColor &ca,
    const QColor &cb, GradientType eff, int xfactor, int yfactor,
    int ncols)
{
    int dir; // general parameter used for direction switches

    bool _xanti = false , _yanti = false;

    if (xfactor < 0) _xanti = true; // negative on X direction
    if (yfactor < 0) _yanti = true; // negative on Y direction

    xfactor = abs(xfactor);
    yfactor = abs(yfactor);

    if (!xfactor) xfactor = 1;
    if (!yfactor) yfactor = 1;

    if (xfactor > 200 ) xfactor = 200;
    if (yfactor > 200 ) yfactor = 200;


    //    float xbal = xfactor/5000.;
    //    float ybal = yfactor/5000.;
    float xbal = xfactor/30./size.width();
    float ybal = yfactor/30./size.height();
    float rat;

    int rDiff, gDiff, bDiff;
    int rca, gca, bca, rcb, gcb, bcb;

    QImage image(size, 32);

    if (size.width() == 0 || size.height() == 0) {
#ifndef NDEBUG
      std::cerr << "WARNING: KImageEffect::unbalancedGradient : invalid image\n";
#endif
      return image;
    }

    register int x, y;
    unsigned int *scanline;

    rDiff = (rcb = cb.red())   - (rca = ca.red());
    gDiff = (gcb = cb.green()) - (gca = ca.green());
    bDiff = (bcb = cb.blue())  - (bca = ca.blue());

    if( eff == VerticalGradient || eff == HorizontalGradient){
        QColor cRow;

        uint *p;
        uint rgbRow;

    if( eff == VerticalGradient) {
      for ( y = 0; y < size.height(); y++ ) {
        dir = _yanti ? y : size.height() - 1 - y;
            p = (uint *) image.scanLine(dir);
            rat =  1 - exp( - (float)y  * ybal );

            cRow.setRgb( rcb - (int) ( rDiff * rat ),
                         gcb - (int) ( gDiff * rat ),
                         bcb - (int) ( bDiff * rat ) );

            rgbRow = cRow.rgb();

            for( x = 0; x < size.width(); x++ ) {
          *p = rgbRow;
          p++;
            }
      }
    }
    else {

      unsigned int *src = (unsigned int *)image.scanLine(0);
      for(x = 0; x < size.width(); x++ )
        {
          dir = _xanti ? x : size.width() - 1 - x;
          rat = 1 - exp( - (float)x  * xbal );

          src[dir] = qRgb(rcb - (int) ( rDiff * rat ),
                gcb - (int) ( gDiff * rat ),
                bcb - (int) ( bDiff * rat ));
        }

      // Believe it or not, manually copying in a for loop is faster
      // than calling memcpy for each scanline (on the order of ms...).
      // I think this is due to the function call overhead (mosfet).

      for(y = 1; y < size.height(); ++y)
        {
          scanline = (unsigned int *)image.scanLine(y);
          for(x=0; x < size.width(); ++x)
        scanline[x] = src[x];
        }
    }
    }

    else {
      int w=size.width(), h=size.height();

      unsigned char *xtable[3];
      unsigned char *ytable[3];
      xtable[0] = new unsigned char[w];
      xtable[1] = new unsigned char[w];
      xtable[2] = new unsigned char[w];
      ytable[0] = new unsigned char[h];
      ytable[1] = new unsigned char[h];
      ytable[2] = new unsigned char[h];

      if ( eff == DiagonalGradient || eff == CrossDiagonalGradient)
    {
      for (x = 0; x < w; x++) {
              dir = _xanti ? x : w - 1 - x;
              rat = 1 - exp( - (float)x * xbal );

              xtable[0][dir] = (unsigned char) ( rDiff/2 * rat );
              xtable[1][dir] = (unsigned char) ( gDiff/2 * rat );
              xtable[2][dir] = (unsigned char) ( bDiff/2 * rat );
          }

      for (y = 0; y < h; y++) {
              dir = _yanti ? y : h - 1 - y;
              rat =  1 - exp( - (float)y  * ybal );

              ytable[0][dir] = (unsigned char) ( rDiff/2 * rat );
              ytable[1][dir] = (unsigned char) ( gDiff/2 * rat );
              ytable[2][dir] = (unsigned char) ( bDiff/2 * rat );
          }

      for (y = 0; y < h; y++) {
              unsigned int *scanline = (unsigned int *)image.scanLine(y);
              for (x = 0; x < w; x++) {
                  scanline[x] = qRgb(rcb - (xtable[0][x] + ytable[0][y]),
                                     gcb - (xtable[1][x] + ytable[1][y]),
                                     bcb - (xtable[2][x] + ytable[2][y]));
              }
          }
        }

      else if (eff == RectangleGradient ||
               eff == PyramidGradient ||
               eff == PipeCrossGradient ||
               eff == EllipticGradient)
      {
          int rSign = rDiff>0? 1: -1;
          int gSign = gDiff>0? 1: -1;
          int bSign = bDiff>0? 1: -1;

          for (x = 0; x < w; x++)
        {
                dir = _xanti ? x : w - 1 - x;
                rat =  1 - exp( - (float)x * xbal );

                xtable[0][dir] = (unsigned char) abs((int)(rDiff*(0.5-rat)));
                xtable[1][dir] = (unsigned char) abs((int)(gDiff*(0.5-rat)));
                xtable[2][dir] = (unsigned char) abs((int)(bDiff*(0.5-rat)));
            }

          for (y = 0; y < h; y++)
          {
              dir = _yanti ? y : h - 1 - y;

              rat =  1 - exp( - (float)y * ybal );

              ytable[0][dir] = (unsigned char) abs((int)(rDiff*(0.5-rat)));
              ytable[1][dir] = (unsigned char) abs((int)(gDiff*(0.5-rat)));
              ytable[2][dir] = (unsigned char) abs((int)(bDiff*(0.5-rat)));
          }

          for (y = 0; y < h; y++) {
              unsigned int *scanline = (unsigned int *)image.scanLine(y);
              for (x = 0; x < w; x++) {
                  if (eff == PyramidGradient)
                  {
                      scanline[x] = qRgb(rcb-rSign*(xtable[0][x]+ytable[0][y]),
                                         gcb-gSign*(xtable[1][x]+ytable[1][y]),
                                         bcb-bSign*(xtable[2][x]+ytable[2][y]));
                  }
                  if (eff == RectangleGradient)
                  {
                      scanline[x] = qRgb(rcb - rSign *
                                         QMAX(xtable[0][x], ytable[0][y]) * 2,
                                         gcb - gSign *
                                         QMAX(xtable[1][x], ytable[1][y]) * 2,
                                         bcb - bSign *
                                         QMAX(xtable[2][x], ytable[2][y]) * 2);
                  }
                  if (eff == PipeCrossGradient)
                  {
                      scanline[x] = qRgb(rcb - rSign *
                                         QMIN(xtable[0][x], ytable[0][y]) * 2,
                                         gcb - gSign *
                                         QMIN(xtable[1][x], ytable[1][y]) * 2,
                                         bcb - bSign *
                                         QMIN(xtable[2][x], ytable[2][y]) * 2);
                  }
                  if (eff == EllipticGradient)
                  {
                      scanline[x] = qRgb(rcb - rSign *
                                         (int)sqrt((xtable[0][x]*xtable[0][x] +
                                                    ytable[0][y]*ytable[0][y])*2.0),
                                         gcb - gSign *
                                         (int)sqrt((xtable[1][x]*xtable[1][x] +
                                                    ytable[1][y]*ytable[1][y])*2.0),
                                         bcb - bSign *
                                         (int)sqrt((xtable[2][x]*xtable[2][x] +
                                                    ytable[2][y]*ytable[2][y])*2.0));
                  }
              }
          }
      }

      if (ncols && (QPixmap::defaultDepth() < 15 )) {
          if ( ncols < 2 || ncols > 256 )
              ncols = 3;
          QColor *dPal = new QColor[ncols];
          for (int i=0; i<ncols; i++) {
              dPal[i].setRgb ( rca + rDiff * i / ( ncols - 1 ),
                               gca + gDiff * i / ( ncols - 1 ),
                               bca + bDiff * i / ( ncols - 1 ) );
          }
          dither(image, dPal, ncols);
          delete [] dPal;
      }

      delete [] xtable[0];
      delete [] xtable[1];
      delete [] xtable[2];
      delete [] ytable[0];
      delete [] ytable[1];
      delete [] ytable[2];

    }

    return image;
}

namespace {

struct KIE4Pack
{
    Q_UINT16 data[4];
};

struct KIE8Pack
{
    Q_UINT16 data[8];
};

}

//======================================================================
//
// Intensity effects
//
//======================================================================


/* This builds a 256 byte unsigned char lookup table with all
 * the possible percent values prior to applying the effect, then uses
 * integer math for the pixels. For any image larger than 9x9 this will be
 * less expensive than doing a float operation on the 3 color components of
 * each pixel. (mosfet)
 */
QImage& KImageEffect::intensity(QImage &image, float percent)
{
    if (image.width() == 0 || image.height() == 0) {
#ifndef NDEBUG
      std::cerr << "WARNING: KImageEffect::intensity : invalid image\n";
#endif
      return image;
    }

    int segColors = image.depth() > 8 ? 256 : image.numColors();
    int pixels = image.depth() > 8 ? image.width()*image.height() :
        image.numColors();
    unsigned int *data = image.depth() > 8 ? (unsigned int *)image.bits() :
        (unsigned int *)image.colorTable();

    bool brighten = (percent >= 0);
    if(percent < 0)
        percent = -percent;

#ifdef USE_MMX_INLINE_ASM
    bool haveMMX = KCPUInfo::haveExtension( KCPUInfo::IntelMMX );

    if(haveMMX)
    {
        Q_UINT16 p = Q_UINT16(256.0f*(percent));
        KIE4Pack mult = {{p,p,p,0}};

        __asm__ __volatile__(
        "pxor %%mm7, %%mm7\n\t"                // zero mm7 for unpacking
        "movq  (%0), %%mm6\n\t"                // copy intensity change to mm6
        : : "r"(&mult), "m"(mult));

        unsigned int rem = pixels % 4;
        pixels -= rem;
        Q_UINT32 *end = ( data + pixels );

        if (brighten)
        {
            while ( data != end ) {
                __asm__ __volatile__(
                "movq       (%0), %%mm0\n\t"
                "movq      8(%0), %%mm4\n\t"   // copy 4 pixels of data to mm0 and mm4
                "movq      %%mm0, %%mm1\n\t"
                "movq      %%mm0, %%mm3\n\t"
                "movq      %%mm4, %%mm5\n\t"   // copy to registers for unpacking
                "punpcklbw %%mm7, %%mm0\n\t"
                "punpckhbw %%mm7, %%mm1\n\t"   // unpack the two pixels from mm0
                "pmullw    %%mm6, %%mm0\n\t"
                "punpcklbw %%mm7, %%mm4\n\t"
                "pmullw    %%mm6, %%mm1\n\t"   // multiply by intensity*256
                "psrlw        $8, %%mm0\n\t"   // divide by 256
                "pmullw    %%mm6, %%mm4\n\t"
                "psrlw        $8, %%mm1\n\t"
                "psrlw        $8, %%mm4\n\t"
                "packuswb  %%mm1, %%mm0\n\t"   // pack solution into mm0. saturates at 255
                "movq      %%mm5, %%mm1\n\t"

                "punpckhbw %%mm7, %%mm1\n\t"   // unpack 4th pixel in mm1

                "pmullw    %%mm6, %%mm1\n\t"
                "paddusb   %%mm3, %%mm0\n\t"   // add intesity result to original of mm0
                "psrlw        $8, %%mm1\n\t"
                "packuswb  %%mm1, %%mm4\n\t"   // pack upper two pixels into mm4

                "movq      %%mm0, (%0)\n\t"    // rewrite to memory lower two pixels
                "paddusb   %%mm5, %%mm4\n\t"
                "movq      %%mm4, 8(%0)\n\t"   // rewrite upper two pixels
                : : "r"(data) );
                data += 4;
            }

            end += rem;
            while ( data != end ) {
                __asm__ __volatile__(
                "movd       (%0), %%mm0\n\t"   // repeat above but for
                "punpcklbw %%mm7, %%mm0\n\t"   // one pixel at a time
                "movq      %%mm0, %%mm3\n\t"
                "pmullw    %%mm6, %%mm0\n\t"
                "psrlw        $8, %%mm0\n\t"
                "paddw     %%mm3, %%mm0\n\t"
                "packuswb  %%mm0, %%mm0\n\t"
                "movd      %%mm0, (%0)\n\t"
                : : "r"(data) );
        data++;
            }
        }
        else
        {
            while ( data != end ) {
                __asm__ __volatile__(
                "movq       (%0), %%mm0\n\t"
                "movq      8(%0), %%mm4\n\t"
                "movq      %%mm0, %%mm1\n\t"
                "movq      %%mm0, %%mm3\n\t"

                "movq      %%mm4, %%mm5\n\t"

                "punpcklbw %%mm7, %%mm0\n\t"
                "punpckhbw %%mm7, %%mm1\n\t"
                "pmullw    %%mm6, %%mm0\n\t"
                "punpcklbw %%mm7, %%mm4\n\t"
                "pmullw    %%mm6, %%mm1\n\t"
                "psrlw        $8, %%mm0\n\t"
                "pmullw    %%mm6, %%mm4\n\t"
                "psrlw        $8, %%mm1\n\t"
                "psrlw        $8, %%mm4\n\t"
                "packuswb  %%mm1, %%mm0\n\t"
                "movq      %%mm5, %%mm1\n\t"

                "punpckhbw %%mm7, %%mm1\n\t"

                "pmullw    %%mm6, %%mm1\n\t"
                "psubusb   %%mm0, %%mm3\n\t"   // subtract darkening amount
                "psrlw        $8, %%mm1\n\t"
                "packuswb  %%mm1, %%mm4\n\t"

                "movq      %%mm3, (%0)\n\t"
                "psubusb   %%mm4, %%mm5\n\t"   // only change for this version is
                "movq      %%mm5, 8(%0)\n\t"   // subtraction here as we are darkening image
                : : "r"(data) );
                data += 4;
            }

            end += rem;
            while ( data != end ) {
                __asm__ __volatile__(
                "movd       (%0), %%mm0\n\t"
                "punpcklbw %%mm7, %%mm0\n\t"
                "movq      %%mm0, %%mm3\n\t"
                "pmullw    %%mm6, %%mm0\n\t"
                "psrlw        $8, %%mm0\n\t"
                "psubusw   %%mm0, %%mm3\n\t"
                "packuswb  %%mm3, %%mm3\n\t"
                "movd      %%mm3, (%0)\n\t"
                : : "r"(data) );
                data++;
            }
        }
        __asm__ __volatile__("emms");          // clear mmx state
    }
    else
#endif // USE_MMX_INLINE_ASM
    {
        unsigned char *segTbl = new unsigned char[segColors];
        int tmp;
        if(brighten){ // keep overflow check out of loops
            for(int i=0; i < segColors; ++i){
                tmp = (int)(i*percent);
                if(tmp > 255)
                    tmp = 255;
                segTbl[i] = tmp;
            }
        }
        else{
            for(int i=0; i < segColors; ++i){
                tmp = (int)(i*percent);
                if(tmp < 0)
                    tmp = 0;
                 segTbl[i] = tmp;
            }
        }

        if(brighten){ // same here
            for(int i=0; i < pixels; ++i){
                int r = qRed(data[i]);
                int g = qGreen(data[i]);
                int b = qBlue(data[i]);
                int a = qAlpha(data[i]);
                r = r + segTbl[r] > 255 ? 255 : r + segTbl[r];
                g = g + segTbl[g] > 255 ? 255 : g + segTbl[g];
                b = b + segTbl[b] > 255 ? 255 : b + segTbl[b];
                data[i] = qRgba(r, g, b,a);
            }
        }
        else{
            for(int i=0; i < pixels; ++i){
                int r = qRed(data[i]);
                int g = qGreen(data[i]);
                int b = qBlue(data[i]);
                int a = qAlpha(data[i]);
                r = r - segTbl[r] < 0 ? 0 : r - segTbl[r];
                g = g - segTbl[g] < 0 ? 0 : g - segTbl[g];
                b = b - segTbl[b] < 0 ? 0 : b - segTbl[b];
                data[i] = qRgba(r, g, b, a);
            }
        }
        delete [] segTbl;
    }

    return image;
}

QImage& KImageEffect::channelIntensity(QImage &image, float percent,
                                       RGBComponent channel)
{
    if (image.width() == 0 || image.height() == 0) {
#ifndef NDEBUG
      std::cerr << "WARNING: KImageEffect::channelIntensity : invalid image\n";
#endif
      return image;
    }

    int segColors = image.depth() > 8 ? 256 : image.numColors();
    unsigned char *segTbl = new unsigned char[segColors];
    int pixels = image.depth() > 8 ? image.width()*image.height() :
        image.numColors();
    unsigned int *data = image.depth() > 8 ? (unsigned int *)image.bits() :
        (unsigned int *)image.colorTable();
    bool brighten = (percent >= 0);
    if(percent < 0)
        percent = -percent;

    if(brighten){ // keep overflow check out of loops
        for(int i=0; i < segColors; ++i){
            int tmp = (int)(i*percent);
            if(tmp > 255)
                tmp = 255;
            segTbl[i] = tmp;
        }
    }
    else{
        for(int i=0; i < segColors; ++i){
            int tmp = (int)(i*percent);
            if(tmp < 0)
                tmp = 0;
            segTbl[i] = tmp;
        }
    }

    if(brighten){ // same here
        if(channel == Red){ // and here ;-)
            for(int i=0; i < pixels; ++i){
                int c = qRed(data[i]);
                c = c + segTbl[c] > 255 ? 255 : c + segTbl[c];
                data[i] = qRgba(c, qGreen(data[i]), qBlue(data[i]), qAlpha(data[i]));
            }
        }
        if(channel == Green){
            for(int i=0; i < pixels; ++i){
                int c = qGreen(data[i]);
                c = c + segTbl[c] > 255 ? 255 : c + segTbl[c];
                data[i] = qRgba(qRed(data[i]), c, qBlue(data[i]), qAlpha(data[i]));
            }
        }
        else{
            for(int i=0; i < pixels; ++i){
                int c = qBlue(data[i]);
                c = c + segTbl[c] > 255 ? 255 : c + segTbl[c];
                data[i] = qRgba(qRed(data[i]), qGreen(data[i]), c, qAlpha(data[i]));
            }
        }

    }
    else{
        if(channel == Red){
            for(int i=0; i < pixels; ++i){
                int c = qRed(data[i]);
                c = c - segTbl[c] < 0 ? 0 : c - segTbl[c];
                data[i] = qRgba(c, qGreen(data[i]), qBlue(data[i]), qAlpha(data[i]));
            }
        }
        if(channel == Green){
            for(int i=0; i < pixels; ++i){
                int c = qGreen(data[i]);
                c = c - segTbl[c] < 0 ? 0 : c - segTbl[c];
                data[i] = qRgba(qRed(data[i]), c, qBlue(data[i]), qAlpha(data[i]));
            }
        }
        else{
            for(int i=0; i < pixels; ++i){
                int c = qBlue(data[i]);
                c = c - segTbl[c] < 0 ? 0 : c - segTbl[c];
                data[i] = qRgba(qRed(data[i]), qGreen(data[i]), c, qAlpha(data[i]));
            }
        }
    }
    delete [] segTbl;

    return image;
}

// Modulate an image with an RBG channel of another image
//
QImage& KImageEffect::modulate(QImage &image, QImage &modImage, bool reverse,
    ModulationType type, int factor, RGBComponent channel)
{
    if (image.width() == 0 || image.height() == 0 ||
        modImage.width() == 0 || modImage.height() == 0) {
#ifndef NDEBUG
      std::cerr << "WARNING: KImageEffect::modulate : invalid image\n";
#endif
      return image;
    }

    int r, g, b, h, s, v, a;
    QColor clr;
    int mod=0;
    unsigned int x1, x2, y1, y2;
    register int x, y;

    // for image, we handle only depth 32
    if (image.depth()<32) image = image.convertDepth(32);

    // for modImage, we handle depth 8 and 32
    if (modImage.depth()<8) modImage = modImage.convertDepth(8);

    unsigned int *colorTable2 = (modImage.depth()==8) ?
                 modImage.colorTable():0;
    unsigned int *data1, *data2;
    unsigned char *data2b;
    unsigned int color1, color2;

    x1 = image.width();    y1 = image.height();
    x2 = modImage.width(); y2 = modImage.height();

    for (y = 0; y < (int)y1; y++) {
        data1 =  (unsigned int *) image.scanLine(y);
    data2 =  (unsigned int *) modImage.scanLine( y%y2 );
    data2b = (unsigned char *) modImage.scanLine( y%y2 );

    x=0;
    while(x < (int)x1) {
      color2 = (colorTable2) ? colorTable2[*data2b] : *data2;
      if (reverse) {
          color1 = color2;
          color2 = *data1;
      }
      else
          color1 = *data1;

      if (type == Intensity || type == Contrast) {
              r = qRed(color1);
          g = qGreen(color1);
          b = qBlue(color1);
          if (channel != All) {
                mod = (channel == Red) ? qRed(color2) :
            (channel == Green) ? qGreen(color2) :
                (channel == Blue) ? qBlue(color2) :
            (channel == Gray) ? qGray(color2) : 0;
            mod = mod*factor/50;
          }

          if (type == Intensity) {
            if (channel == All) {
              r += r * factor/50 * qRed(color2)/256;
              g += g * factor/50 * qGreen(color2)/256;
              b += b * factor/50 * qBlue(color2)/256;
            }
            else {
              r += r * mod/256;
              g += g * mod/256;
              b += b * mod/256;
            }
          }
          else { // Contrast
            if (channel == All) {
          r += (r-128) * factor/50 * qRed(color2)/128;
              g += (g-128) * factor/50 * qGreen(color2)/128;
              b += (b-128) * factor/50 * qBlue(color2)/128;
            }
            else {
              r += (r-128) * mod/128;
              g += (g-128) * mod/128;
              b += (b-128) * mod/128;
            }
          }

          if (r<0) r=0; if (r>255) r=255;
          if (g<0) g=0; if (g>255) g=255;
          if (b<0) b=0; if (b>255) b=255;
          a = qAlpha(*data1);
          *data1 = qRgba(r, g, b, a);
      }
      else if (type == Saturation || type == HueShift) {
          clr.setRgb(color1);
          clr.hsv(&h, &s, &v);
              mod = (channel == Red) ? qRed(color2) :
            (channel == Green) ? qGreen(color2) :
                (channel == Blue) ? qBlue(color2) :
            (channel == Gray) ? qGray(color2) : 0;
          mod = mod*factor/50;

          if (type == Saturation) {
          s -= s * mod/256;
          if (s<0) s=0; if (s>255) s=255;
          }
          else { // HueShift
            h += mod;
        while(h<0) h+=360;
        h %= 360;
          }

          clr.setHsv(h, s, v);
          a = qAlpha(*data1);
          *data1 = clr.rgb() | ((uint)(a & 0xff) << 24);
      }
      data1++; data2++; data2b++; x++;
      if ( (x%x2) ==0) { data2 -= x2; data2b -= x2; }
        }
    }
    return image;
}



//======================================================================
//
// Blend effects
//
//======================================================================


// Nice and fast direct pixel manipulation
QImage& KImageEffect::blend(const QColor& clr, QImage& dst, float opacity)
{
    if (dst.width() <= 0 || dst.height() <= 0)
        return dst;

    if (opacity < 0.0 || opacity > 1.0) {
#ifndef NDEBUG
        std::cerr << "WARNING: KImageEffect::blend : invalid opacity. Range [0, 1]\n";
#endif
        return dst;
    }

    int depth = dst.depth();
    if (depth != 32)
        dst = dst.convertDepth(32);

    int pixels = dst.width() * dst.height();

#ifdef USE_SSE2_INLINE_ASM
    if ( KCPUInfo::haveExtension( KCPUInfo::IntelSSE2 ) && pixels > 16 ) {
        Q_UINT16 alpha = Q_UINT16( ( 1.0 - opacity ) * 256.0 );

        KIE8Pack packedalpha = { { alpha, alpha, alpha, 256,
                                      alpha, alpha, alpha, 256 } };

        Q_UINT16 red   = Q_UINT16( clr.red()   * 256 * opacity );
        Q_UINT16 green = Q_UINT16( clr.green() * 256 * opacity );
        Q_UINT16 blue  = Q_UINT16( clr.blue()  * 256 * opacity );

        KIE8Pack packedcolor = { { blue, green, red, 0,
                                      blue, green, red, 0 } };

        // Prepare the XMM5, XMM6 and XMM7 registers for unpacking and blending
        __asm__ __volatile__(
        "pxor        %%xmm7,  %%xmm7\n\t" // Zero out XMM7 for unpacking
        "movdqu        (%0),  %%xmm6\n\t" // Set up (1 - alpha) * 256 in XMM6
        "movdqu        (%1),  %%xmm5\n\t" // Set up color * alpha * 256 in XMM5
        : : "r"(&packedalpha), "r"(&packedcolor), "m"(packedcolor), "m"(packedalpha) );

        Q_UINT32 *data = reinterpret_cast<Q_UINT32*>( dst.bits() );

        // Check how many pixels we need to process to achieve 16 byte alignment
        int offset = (16 - (Q_UINT32( data ) & 0x0f)) / 4;

        // The main loop processes 8 pixels / iteration
        int remainder = (pixels - offset) % 8;
        pixels -= remainder;

        // Alignment loop
        for ( int i = 0; i < offset; i++ ) {
            __asm__ __volatile__(
            "movd         (%0,%1,4),      %%xmm0\n\t"  // Load one pixel to XMM1
            "punpcklbw       %%xmm7,      %%xmm0\n\t"  // Unpack the pixel
            "pmullw          %%xmm6,      %%xmm0\n\t"  // Multiply the pixel with (1 - alpha) * 256
            "paddw           %%xmm5,      %%xmm0\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%xmm0\n\t"  // Divide by 256
            "packuswb        %%xmm1,      %%xmm0\n\t"  // Pack the pixel to a dword
            "movd            %%xmm0,   (%0,%1,4)\n\t"  // Write the pixel to the image
            : : "r"(data), "r"(i) );
        }

        // Main loop
        for ( int i = offset; i < pixels; i += 8 ) {
            __asm__ __volatile(
            // Load 8 pixels to XMM registers 1 - 4
            "movq         (%0,%1,4),      %%xmm0\n\t"  // Load pixels 1 and 2 to XMM1
            "movq        8(%0,%1,4),      %%xmm1\n\t"  // Load pixels 3 and 4 to XMM2
            "movq       16(%0,%1,4),      %%xmm2\n\t"  // Load pixels 5 and 6 to XMM3
            "movq       24(%0,%1,4),      %%xmm3\n\t"  // Load pixels 7 and 8 to XMM4

            // Prefetch the pixels for next iteration
            "prefetchnta 32(%0,%1,4)            \n\t"

            // Blend pixels 1 and 2
            "punpcklbw       %%xmm7,      %%xmm0\n\t"  // Unpack the pixels
            "pmullw          %%xmm6,      %%xmm0\n\t"  // Multiply the pixels with (1 - alpha) * 256
            "paddw           %%xmm5,      %%xmm0\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%xmm0\n\t"  // Divide by 256

            // Blend pixels 3 and 4
            "punpcklbw       %%xmm7,      %%xmm1\n\t"  // Unpack the pixels
            "pmullw          %%xmm6,      %%xmm1\n\t"  // Multiply the pixels with (1 - alpha) * 256
            "paddw           %%xmm5,      %%xmm1\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%xmm1\n\t"  // Divide by 256

            // Blend pixels 5 and 6
            "punpcklbw       %%xmm7,      %%xmm2\n\t"  // Unpack the pixels
            "pmullw          %%xmm6,      %%xmm2\n\t"  // Multiply the pixels with (1 - alpha) * 256
            "paddw           %%xmm5,      %%xmm2\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%xmm2\n\t"  // Divide by 256

            // Blend pixels 7 and 8
            "punpcklbw       %%xmm7,      %%xmm3\n\t"  // Unpack the pixels
            "pmullw          %%xmm6,      %%xmm3\n\t"  // Multiply the pixels with (1 - alpha) * 256
            "paddw           %%xmm5,      %%xmm3\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%xmm3\n\t"  // Divide by 256

            // Pack the pixels into 2 double quadwords
            "packuswb        %%xmm1,      %%xmm0\n\t"  // Pack pixels 1 - 4 to a double qword
            "packuswb        %%xmm3,      %%xmm2\n\t"  // Pack pixles 5 - 8 to a double qword

            // Write the pixels back to the image
            "movdqa          %%xmm0,   (%0,%1,4)\n\t"  // Store pixels 1 - 4
            "movdqa          %%xmm2, 16(%0,%1,4)\n\t"  // Store pixels 5 - 8
            : : "r"(data), "r"(i) );
        }

        // Cleanup loop
        for ( int i = pixels; i < pixels + remainder; i++ ) {
            __asm__ __volatile__(
            "movd         (%0,%1,4),      %%xmm0\n\t"  // Load one pixel to XMM1
            "punpcklbw       %%xmm7,      %%xmm0\n\t"  // Unpack the pixel
            "pmullw          %%xmm6,      %%xmm0\n\t"  // Multiply the pixel with (1 - alpha) * 256
            "paddw           %%xmm5,      %%xmm0\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%xmm0\n\t"  // Divide by 256
            "packuswb        %%xmm1,      %%xmm0\n\t"  // Pack the pixel to a dword
            "movd            %%xmm0,   (%0,%1,4)\n\t"  // Write the pixel to the image
            : : "r"(data), "r"(i) );
        }
    } else
#endif

#ifdef USE_MMX_INLINE_ASM
    if ( KCPUInfo::haveExtension( KCPUInfo::IntelMMX ) && pixels > 1 ) {
        Q_UINT16 alpha = Q_UINT16( ( 1.0 - opacity ) * 256.0 );
        KIE4Pack packedalpha = { { alpha, alpha, alpha, 256 } };

        Q_UINT16 red   = Q_UINT16( clr.red()   * 256 * opacity );
        Q_UINT16 green = Q_UINT16( clr.green() * 256 * opacity );
        Q_UINT16 blue  = Q_UINT16( clr.blue()  * 256 * opacity );

        KIE4Pack packedcolor = { { blue, green, red, 0 } };

        __asm__ __volatile__(
        "pxor        %%mm7,    %%mm7\n\t"       // Zero out MM7 for unpacking
        "movq         (%0),    %%mm6\n\t"       // Set up (1 - alpha) * 256 in MM6
        "movq         (%1),    %%mm5\n\t"       // Set up color * alpha * 256 in MM5
        : : "r"(&packedalpha), "r"(&packedcolor), "m"(packedcolor), "m"(packedalpha) );

        Q_UINT32 *data = reinterpret_cast<Q_UINT32*>( dst.bits() );

        // The main loop processes 4 pixels / iteration
        int remainder = pixels % 4;
        pixels -= remainder;

        // Main loop
        for ( int i = 0; i < pixels; i += 4 ) {
            __asm__ __volatile__(
            // Load 4 pixels to MM registers 1 - 4
            "movd         (%0,%1,4),      %%mm0\n\t"  // Load the 1st pixel to MM0
            "movd        4(%0,%1,4),      %%mm1\n\t"  // Load the 2nd pixel to MM1
            "movd        8(%0,%1,4),      %%mm2\n\t"  // Load the 3rd pixel to MM2
            "movd       12(%0,%1,4),      %%mm3\n\t"  // Load the 4th pixel to MM3

            // Blend the first pixel
            "punpcklbw        %%mm7,      %%mm0\n\t"  // Unpack the pixel
            "pmullw           %%mm6,      %%mm0\n\t"  // Multiply the pixel with (1 - alpha) * 256
            "paddw            %%mm5,      %%mm0\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%mm0\n\t"  // Divide by 256

            // Blend the second pixel
            "punpcklbw        %%mm7,      %%mm1\n\t"  // Unpack the pixel
            "pmullw           %%mm6,      %%mm1\n\t"  // Multiply the pixel with (1 - alpha) * 256
            "paddw            %%mm5,      %%mm1\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%mm1\n\t"  // Divide by 256

            // Blend the third pixel
            "punpcklbw        %%mm7,      %%mm2\n\t"  // Unpack the pixel
            "pmullw           %%mm6,      %%mm2\n\t"  // Multiply the pixel with (1 - alpha) * 256
            "paddw            %%mm5,      %%mm2\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%mm2\n\t"  // Divide by 256

            // Blend the fourth pixel
            "punpcklbw        %%mm7,      %%mm3\n\t"  // Unpack the pixel
            "pmullw           %%mm6,      %%mm3\n\t"  // Multiply the pixel with (1 - alpha) * 256
            "paddw            %%mm5,      %%mm3\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%mm3\n\t"  // Divide by 256

            // Pack the pixels into 2 quadwords
            "packuswb         %%mm1,      %%mm0\n\t"  // Pack pixels 1 and 2 to a qword
            "packuswb         %%mm3,      %%mm2\n\t"  // Pack pixels 3 and 4 to a qword

            // Write the pixels back to the image
            "movq             %%mm0,  (%0,%1,4)\n\t"  // Store pixels 1 and 2
            "movq             %%mm2, 8(%0,%1,4)\n\t"  // Store pixels 3 and 4
            : : "r"(data), "r"(i) );
        }

        // Cleanup loop
        for ( int i = pixels; i < pixels + remainder; i++ ) {
            __asm__ __volatile__(
            "movd         (%0,%1,4),      %%mm0\n\t"  // Load one pixel to MM1
            "punpcklbw        %%mm7,      %%mm0\n\t"  // Unpack the pixel
            "pmullw           %%mm6,      %%mm0\n\t"  // Multiply the pixel with 1 - alpha * 256
            "paddw            %%mm5,      %%mm0\n\t"  // Add color * alpha * 256 to the result
            "psrlw               $8,      %%mm0\n\t"  // Divide by 256
            "packuswb         %%mm0,      %%mm0\n\t"  // Pack the pixel to a dword
            "movd             %%mm0,  (%0,%1,4)\n\t"  // Write the pixel to the image
            : : "r"(data), "r"(i) );
        }

        // Empty the MMX state
        __asm__ __volatile__("emms");
    } else
#endif // USE_MMX_INLINE_ASM

    {
        int rcol, gcol, bcol;
        clr.rgb(&rcol, &gcol, &bcol);

#ifdef WORDS_BIGENDIAN   // ARGB (skip alpha)
        register unsigned char *data = (unsigned char *)dst.bits() + 1;
#else                    // BGRA
        register unsigned char *data = (unsigned char *)dst.bits();
#endif

        for (register int i=0; i<pixels; i++)
        {
#ifdef WORDS_BIGENDIAN
            *data += (unsigned char)((rcol - *data) * opacity);
            data++;
            *data += (unsigned char)((gcol - *data) * opacity);
            data++;
            *data += (unsigned char)((bcol - *data) * opacity);
            data++;
#else
            *data += (unsigned char)((bcol - *data) * opacity);
            data++;
            *data += (unsigned char)((gcol - *data) * opacity);
            data++;
            *data += (unsigned char)((rcol - *data) * opacity);
            data++;
#endif
            data++; // skip alpha
        }
    }

    return dst;
}

// Nice and fast direct pixel manipulation
QImage& KImageEffect::blend(QImage& src, QImage& dst, float opacity)
{
    if (src.width() <= 0 || src.height() <= 0)
        return dst;
    if (dst.width() <= 0 || dst.height() <= 0)
        return dst;

    if (src.width() != dst.width() || src.height() != dst.height()) {
#ifndef NDEBUG
        std::cerr << "WARNING: KImageEffect::blend : src and destination images are not the same size\n";
#endif
        return dst;
    }

    if (opacity < 0.0 || opacity > 1.0) {
#ifndef NDEBUG
        std::cerr << "WARNING: KImageEffect::blend : invalid opacity. Range [0, 1]\n";
#endif
        return dst;
    }

    if (src.depth() != 32) src = src.convertDepth(32);
    if (dst.depth() != 32) dst = dst.convertDepth(32);

    int pixels = src.width() * src.height();

#ifdef USE_SSE2_INLINE_ASM
    if ( KCPUInfo::haveExtension( KCPUInfo::IntelSSE2 ) && pixels > 16 ) {
        Q_UINT16 alpha = Q_UINT16( opacity * 256.0 );
        KIE8Pack packedalpha = { { alpha, alpha, alpha, 0,
                                   alpha, alpha, alpha, 0 } };

        // Prepare the XMM6 and XMM7 registers for unpacking and blending
        __asm__ __volatile__(
        "pxor      %%xmm7, %%xmm7\n\t" // Zero out XMM7 for unpacking
        "movdqu      (%0), %%xmm6\n\t" // Set up alpha * 256 in XMM6
        : : "r"(&packedalpha), "m"(packedalpha) );

        Q_UINT32 *data1 = reinterpret_cast<Q_UINT32*>( src.bits() );
        Q_UINT32 *data2 = reinterpret_cast<Q_UINT32*>( dst.bits() );

        // Check how many pixels we need to process to achieve 16 byte alignment
        int offset = (16 - (Q_UINT32( data2 ) & 0x0f)) / 4;

        // The main loop processes 4 pixels / iteration
        int remainder = (pixels - offset) % 4;
        pixels -= remainder;

        // Alignment loop
        for ( int i = 0; i < offset; i++ ) {
            __asm__ __volatile__(
            "movd       (%1,%2,4),    %%xmm1\n\t"  // Load one dst pixel to XMM1
            "punpcklbw     %%xmm7,    %%xmm1\n\t"  // Unpack the pixel
            "movd       (%0,%2,4),    %%xmm0\n\t"  // Load one src pixel to XMM0
            "punpcklbw     %%xmm7,    %%xmm0\n\t"  // Unpack the pixel
            "psubw         %%xmm1,    %%xmm0\n\t"  // Subtract dst from src
            "pmullw        %%xmm6,    %%xmm0\n\t"  // Multiply the result with alpha * 256
            "psllw             $8,    %%xmm1\n\t"  // Multiply dst with 256
            "paddw         %%xmm1,    %%xmm0\n\t"  // Add dst to result
            "psrlw             $8,    %%xmm0\n\t"  // Divide by 256
            "packuswb      %%xmm1,    %%xmm0\n\t"  // Pack the pixel to a dword
            "movd          %%xmm0, (%1,%2,4)\n\t"  // Write the pixel to the image
            : : "r"(data1), "r"(data2), "r"(i) );
        }

        // Main loop
        for ( int i = offset; i < pixels; i += 4 ) {
            __asm__ __volatile__(
            // Load 4 src pixels to XMM0 and XMM2 and 4 dst pixels to XMM1 and XMM3
            "movq       (%0,%2,4),    %%xmm0\n\t"  // Load two src pixels to XMM0
            "movq       (%1,%2,4),    %%xmm1\n\t"  // Load two dst pixels to XMM1
            "movq      8(%0,%2,4),    %%xmm2\n\t"  // Load two src pixels to XMM2
            "movq      8(%1,%2,4),    %%xmm3\n\t"  // Load two dst pixels to XMM3

            // Prefetch the pixels for the iteration after the next one
            "prefetchnta 32(%0,%2,4)        \n\t"
            "prefetchnta 32(%1,%2,4)        \n\t"

            // Blend the first two pixels
            "punpcklbw     %%xmm7,    %%xmm1\n\t"  // Unpack the dst pixels
            "punpcklbw     %%xmm7,    %%xmm0\n\t"  // Unpack the src pixels
            "psubw         %%xmm1,    %%xmm0\n\t"  // Subtract dst from src
            "pmullw        %%xmm6,    %%xmm0\n\t"  // Multiply the result with alpha * 256
            "psllw             $8,    %%xmm1\n\t"  // Multiply dst with 256
            "paddw         %%xmm1,    %%xmm0\n\t"  // Add dst to the result
            "psrlw             $8,    %%xmm0\n\t"  // Divide by 256

            // Blend the next two pixels
            "punpcklbw     %%xmm7,    %%xmm3\n\t"  // Unpack the dst pixels
            "punpcklbw     %%xmm7,    %%xmm2\n\t"  // Unpack the src pixels
            "psubw         %%xmm3,    %%xmm2\n\t"  // Subtract dst from src
            "pmullw        %%xmm6,    %%xmm2\n\t"  // Multiply the result with alpha * 256
            "psllw             $8,    %%xmm3\n\t"  // Multiply dst with 256
            "paddw         %%xmm3,    %%xmm2\n\t"  // Add dst to the result
            "psrlw             $8,    %%xmm2\n\t"  // Divide by 256

            // Write the pixels back to the image
            "packuswb      %%xmm2,    %%xmm0\n\t"  // Pack the pixels to a double qword
            "movdqa        %%xmm0, (%1,%2,4)\n\t"  // Store the pixels
            : : "r"(data1), "r"(data2), "r"(i) );
        }

        // Cleanup loop
        for ( int i = pixels; i < pixels + remainder; i++ ) {
            __asm__ __volatile__(
            "movd       (%1,%2,4),    %%xmm1\n\t"  // Load one dst pixel to XMM1
            "punpcklbw     %%xmm7,    %%xmm1\n\t"  // Unpack the pixel
            "movd       (%0,%2,4),    %%xmm0\n\t"  // Load one src pixel to XMM0
            "punpcklbw     %%xmm7,    %%xmm0\n\t"  // Unpack the pixel
            "psubw         %%xmm1,    %%xmm0\n\t"  // Subtract dst from src
            "pmullw        %%xmm6,    %%xmm0\n\t"  // Multiply the result with alpha * 256
            "psllw             $8,    %%xmm1\n\t"  // Multiply dst with 256
            "paddw         %%xmm1,    %%xmm0\n\t"  // Add dst to result
            "psrlw             $8,    %%xmm0\n\t"  // Divide by 256
            "packuswb      %%xmm1,    %%xmm0\n\t"  // Pack the pixel to a dword
            "movd          %%xmm0, (%1,%2,4)\n\t"  // Write the pixel to the image
            : : "r"(data1), "r"(data2), "r"(i) );
        }
    } else
#endif // USE_SSE2_INLINE_ASM

#ifdef USE_MMX_INLINE_ASM
    if ( KCPUInfo::haveExtension( KCPUInfo::IntelMMX ) && pixels > 1 ) {
        Q_UINT16 alpha = Q_UINT16( opacity * 256.0 );
        KIE4Pack packedalpha = { { alpha, alpha, alpha, 0 } };

        // Prepare the MM6 and MM7 registers for blending and unpacking
        __asm__ __volatile__(
        "pxor       %%mm7,   %%mm7\n\t"      // Zero out MM7 for unpacking
        "movq        (%0),   %%mm6\n\t"      // Set up alpha * 256 in MM6
        : : "r"(&packedalpha), "m"(packedalpha) );

        Q_UINT32 *data1 = reinterpret_cast<Q_UINT32*>( src.bits() );
        Q_UINT32 *data2 = reinterpret_cast<Q_UINT32*>( dst.bits() );

        // The main loop processes 2 pixels / iteration
        int remainder = pixels % 2;
        pixels -= remainder;

        // Main loop
        for ( int i = 0; i < pixels; i += 2 ) {
            __asm__ __volatile__(
            // Load 2 src pixels to MM0 and MM2 and 2 dst pixels to MM1 and MM3
            "movd        (%0,%2,4),     %%mm0\n\t"  // Load the 1st src pixel to MM0
            "movd        (%1,%2,4),     %%mm1\n\t"  // Load the 1st dst pixel to MM1
            "movd       4(%0,%2,4),     %%mm2\n\t"  // Load the 2nd src pixel to MM2
            "movd       4(%1,%2,4),     %%mm3\n\t"  // Load the 2nd dst pixel to MM3

            // Blend the first pixel
            "punpcklbw       %%mm7,     %%mm0\n\t"  // Unpack the src pixel
            "punpcklbw       %%mm7,     %%mm1\n\t"  // Unpack the dst pixel
            "psubw           %%mm1,     %%mm0\n\t"  // Subtract dst from src
            "pmullw          %%mm6,     %%mm0\n\t"  // Multiply the result with alpha * 256
            "psllw              $8,     %%mm1\n\t"  // Multiply dst with 256
            "paddw           %%mm1,     %%mm0\n\t"  // Add dst to the result
            "psrlw              $8,     %%mm0\n\t"  // Divide by 256

            // Blend the second pixel
            "punpcklbw       %%mm7,     %%mm2\n\t"  // Unpack the src pixel
            "punpcklbw       %%mm7,     %%mm3\n\t"  // Unpack the dst pixel
            "psubw           %%mm3,     %%mm2\n\t"  // Subtract dst from src