MagickCore/accelerate-private.h

/*
  Copyright 1999-2014 ImageMagick Studio LLC, a non-profit organization
  dedicated to making software imaging solutions freely available.
  
  You may not use this file except in compliance with the License.
  obtain a copy of the License at
  
    http://www.imagemagick.org/script/license.php
  
  Unless required by applicable law or agreed to in writing, software
  distributed under the License is distributed on an "AS IS" BASIS,
  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  See the License for the specific language governing permissions and
  limitations under the License.

  MagickCore private methods for accelerated functions.
*/

#ifndef _MAGICKCORE_ACCELERATE_PRIVATE_H
#define _MAGICKCORE_ACCELERATE_PRIVATE_H

#if defined(__cplusplus) || defined(c_plusplus)
extern "C" {
#endif


#if defined(MAGICKCORE_OPENCL_SUPPORT)

#define OPENCL_DEFINE(VAR,...)	"\n #""define " #VAR " " #__VA_ARGS__ " \n"
#define OPENCL_ELIF(...)	"\n #""elif " #__VA_ARGS__ " \n"
#define OPENCL_ELSE()		"\n #""else " " \n"
#define OPENCL_ENDIF()		"\n #""endif " " \n"
#define OPENCL_IF(...)		"\n #""if " #__VA_ARGS__ " \n"
#define STRINGIFY(...) #__VA_ARGS__ "\n"

typedef struct _FloatPixelPacket
{
#ifdef MAGICK_PIXEL_RGBA  
  MagickRealType
    red,
    green,
    blue,
    opacity;
#endif
#ifdef MAGICK_PIXEL_BGRA 
  MagickRealType
    blue,
    green,
    red,
    opacity;
#endif
} FloatPixelPacket;

const char* accelerateKernels =
  STRINGIFY(
     typedef enum
     {
       UndefinedChannel,
       RedChannel = 0x0001,
       GrayChannel = 0x0001,
       CyanChannel = 0x0001,
       GreenChannel = 0x0002,
       MagentaChannel = 0x0002,
       BlueChannel = 0x0004,
       YellowChannel = 0x0004,
       AlphaChannel = 0x0008,
       OpacityChannel = 0x0008,
       MatteChannel = 0x0008,     /* deprecated */
       BlackChannel = 0x0020,
       IndexChannel = 0x0020,
       CompositeChannels = 0x002F,
       AllChannels = 0x7ffffff,
       /*
       Special purpose channel types.
       */
       TrueAlphaChannel = 0x0040, /* extract actual alpha channel from opacity */
       RGBChannels = 0x0080,      /* set alpha from  grayscale mask in RGB */
       GrayChannels = 0x0080,
       SyncChannels = 0x0100,     /* channels should be modified equally */
       DefaultChannels = ((AllChannels | SyncChannels) &~ OpacityChannel)
     } ChannelType;
  )

  OPENCL_IF((MAGICKCORE_QUANTUM_DEPTH == 8))

  STRINGIFY(
    inline CLQuantum ScaleCharToQuantum(const unsigned char value)
    {
      return((CLQuantum) value);
    }
  )

  OPENCL_ELIF((MAGICKCORE_QUANTUM_DEPTH == 16))

  STRINGIFY(
    inline CLQuantum ScaleCharToQuantum(const unsigned char value)
    {
      return((CLQuantum) (257.0f*value));
    }
  )

  OPENCL_ELIF((MAGICKCORE_QUANTUM_DEPTH == 32))

  STRINGIFY(
    inline CLQuantum ScaleCharToQuantum(const unsigned char value)
    {
      return((Quantum) (16843009.0*value));
    }
  )

  OPENCL_ENDIF()


  STRINGIFY(
    inline int ClampToCanvas(const int offset,const int range)
      {
        return clamp(offset, (int)0, range-1);
      }
  )

  STRINGIFY(
    inline int ClampToCanvasWithHalo(const int offset,const int range, const int edge, const int section)
      {
        return clamp(offset, section?(int)(0-edge):(int)0, section?(range-1):(range-1+edge));
      }
  )

  STRINGIFY(
    inline CLQuantum ClampToQuantum(const float value)
      {
        return (CLQuantum) (clamp(value, 0.0f, (float) QuantumRange) + 0.5f);
      }
  )

  STRINGIFY(
    inline uint ScaleQuantumToMap(CLQuantum value)
      {
        if (value >= (CLQuantum) MaxMap)
          return ((uint)MaxMap);
        else 
          return ((uint)value);
      }
  )

  STRINGIFY(
    inline float PerceptibleReciprocal(const float x)
    {
      float sign = x < (float) 0.0 ? (float) -1.0 : (float) 1.0;
      return((sign*x) >= MagickEpsilon ? (float) 1.0/x : sign*((float) 1.0/MagickEpsilon));
    }
  )

  OPENCL_DEFINE(GetPixelAlpha(pixel),(QuantumRange-(pixel).w))

  STRINGIFY(

  inline CLQuantum getBlue(CLPixelType p)		    { return p.x; }
  inline void setBlue(CLPixelType* p, CLQuantum value)	    { (*p).x = value; }
  inline float getBlueF4(float4 p)	                    { return p.x; }
  inline void setBlueF4(float4* p, float value)             { (*p).x = value; }

  inline CLQuantum getGreen(CLPixelType p)		    { return p.y; }
  inline void setGreen(CLPixelType* p, CLQuantum value)	    { (*p).y = value; }
  inline float getGreenF4(float4 p)			    { return p.y; }
  inline void setGreenF4(float4* p, float value)	    { (*p).y = value; }

  inline CLQuantum getRed(CLPixelType p)		    { return p.z; }
  inline void setRed(CLPixelType* p, CLQuantum value)	    { (*p).z = value; }
  inline float getRedF4(float4 p)			    { return p.z; }
  inline void setRedF4(float4* p, float value)	            { (*p).z = value; }

  inline CLQuantum getOpacity(CLPixelType p)		    { return p.w; }
  inline void setOpacity(CLPixelType* p, CLQuantum value)   { (*p).w = value; }
  inline float getOpacityF4(float4 p)			    { return p.w; }
  inline void setOpacityF4(float4* p, float value)          { (*p).w = value; }

  inline float GetPixelIntensity(int colorspace, CLPixelType p)
  {
    // this is for default intensity and sRGB (not RGB) color space
    float red = getRed(p);
    float green = getGreen(p);
    float blue = getBlue(p);

    if (colorspace == 0)
      return 0.212656*red+0.715158*green+0.072186*blue;
    else
    {
      // need encode gamma
    }
    return 0.0;
  }
  )

  STRINGIFY(
    __kernel 
    void ConvolveOptimized(const __global CLPixelType *input, __global CLPixelType *output,
    const unsigned int imageWidth, const unsigned int imageHeight,
    __constant float *filter, const unsigned int filterWidth, const unsigned int filterHeight,
    const uint matte, const ChannelType channel, __local CLPixelType *pixelLocalCache, __local float* filterCache) {

      int2 blockID;
      blockID.x = get_group_id(0);
      blockID.y = get_group_id(1);

      // image area processed by this workgroup
      int2 imageAreaOrg;
      imageAreaOrg.x = blockID.x * get_local_size(0);
      imageAreaOrg.y = blockID.y * get_local_size(1);

      int2 midFilterDimen;
      midFilterDimen.x = (filterWidth-1)/2;
      midFilterDimen.y = (filterHeight-1)/2;

      int2 cachedAreaOrg = imageAreaOrg - midFilterDimen;

      // dimension of the local cache
      int2 cachedAreaDimen;
      cachedAreaDimen.x = get_local_size(0) + filterWidth - 1;
      cachedAreaDimen.y = get_local_size(1) + filterHeight - 1;

      // cache the pixels accessed by this workgroup in local memory
      int localID = get_local_id(1)*get_local_size(0)+get_local_id(0);
      int cachedAreaNumPixels = cachedAreaDimen.x * cachedAreaDimen.y;
      int groupSize = get_local_size(0) * get_local_size(1);
      for (int i = localID; i < cachedAreaNumPixels; i+=groupSize) {

        int2 cachedAreaIndex;
        cachedAreaIndex.x = i % cachedAreaDimen.x;
        cachedAreaIndex.y = i / cachedAreaDimen.x;

        int2 imagePixelIndex;
        imagePixelIndex = cachedAreaOrg + cachedAreaIndex;

        // only support EdgeVirtualPixelMethod through ClampToCanvas
        // TODO: implement other virtual pixel method
        imagePixelIndex.x = ClampToCanvas(imagePixelIndex.x, imageWidth);
        imagePixelIndex.y = ClampToCanvas(imagePixelIndex.y, imageHeight);

        pixelLocalCache[i] = input[imagePixelIndex.y * imageWidth + imagePixelIndex.x];
      }

      // cache the filter
      for (int i = localID; i < filterHeight*filterWidth; i+=groupSize) {
        filterCache[i] = filter[i];
      }
      barrier(CLK_LOCAL_MEM_FENCE);


      int2 imageIndex;
      imageIndex.x = imageAreaOrg.x + get_local_id(0);
      imageIndex.y = imageAreaOrg.y + get_local_id(1);

      // if out-of-range, stops here and quit
      if (imageIndex.x >= imageWidth
        || imageIndex.y >= imageHeight) {
          return;
      }

      int filterIndex = 0;
      float4 sum = (float4)0.0f;
      float gamma = 0.0f;
      if (((channel & OpacityChannel) == 0) || (matte == 0)) {
        int cacheIndexY = get_local_id(1);
        for (int j = 0; j < filterHeight; j++) {
          int cacheIndexX = get_local_id(0);
          for (int i = 0; i < filterWidth; i++) {
            CLPixelType p = pixelLocalCache[cacheIndexY*cachedAreaDimen.x + cacheIndexX];
            float f = filterCache[filterIndex];

            sum.x += f * p.x;
            sum.y += f * p.y;
            sum.z += f * p.z; 
            sum.w += f * p.w;

            gamma += f;
            filterIndex++;
            cacheIndexX++;
          }
          cacheIndexY++;
        }
      }
      else {
        int cacheIndexY = get_local_id(1);
        for (int j = 0; j < filterHeight; j++) {
          int cacheIndexX = get_local_id(0);
          for (int i = 0; i < filterWidth; i++) {

            CLPixelType p = pixelLocalCache[cacheIndexY*cachedAreaDimen.x + cacheIndexX];
            float alpha = QuantumScale*(QuantumRange-p.w);
            float f = filterCache[filterIndex];
            float g = alpha * f;

            sum.x += g*p.x;
            sum.y += g*p.y;
            sum.z += g*p.z;
            sum.w += f*p.w;

            gamma += g;
            filterIndex++;
            cacheIndexX++;
          }
          cacheIndexY++;
        }
        gamma = PerceptibleReciprocal(gamma);
        sum.xyz = gamma*sum.xyz;
      }
      CLPixelType outputPixel;
      outputPixel.x = ClampToQuantum(sum.x);
      outputPixel.y = ClampToQuantum(sum.y);
      outputPixel.z = ClampToQuantum(sum.z);
      outputPixel.w = ((channel & OpacityChannel)!=0)?ClampToQuantum(sum.w):input[imageIndex.y * imageWidth + imageIndex.x].w;

      output[imageIndex.y * imageWidth + imageIndex.x] = outputPixel;
    }
  )

  STRINGIFY(
    __kernel 
    void Convolve(const __global CLPixelType *input, __global CLPixelType *output,
                  __constant float *filter, const unsigned int filterWidth, const unsigned int filterHeight,
                  const uint matte, const ChannelType channel) {

      int2 imageIndex;
      imageIndex.x = get_global_id(0);
      imageIndex.y = get_global_id(1);

      unsigned int imageWidth = get_global_size(0);
      unsigned int imageHeight = get_global_size(1);

      if (imageIndex.x >= imageWidth
          || imageIndex.y >= imageHeight)
          return;

      int2 midFilterDimen;
      midFilterDimen.x = (filterWidth-1)/2;
      midFilterDimen.y = (filterHeight-1)/2;

      int filterIndex = 0;
      float4 sum = (float4)0.0f;
      float gamma = 0.0f;
      if (((channel & OpacityChannel) == 0) || (matte == 0)) {
        for (int j = 0; j < filterHeight; j++) {
          int2 inputPixelIndex;
          inputPixelIndex.y = imageIndex.y - midFilterDimen.y + j;
          inputPixelIndex.y = ClampToCanvas(inputPixelIndex.y, imageHeight);
          for (int i = 0; i < filterWidth; i++) {
            inputPixelIndex.x = imageIndex.x - midFilterDimen.x + i;
            inputPixelIndex.x = ClampToCanvas(inputPixelIndex.x, imageWidth);
        
            CLPixelType p = input[inputPixelIndex.y * imageWidth + inputPixelIndex.x];
            float f = filter[filterIndex];

            sum.x += f * p.x;
            sum.y += f * p.y;
            sum.z += f * p.z; 
            sum.w += f * p.w;

            gamma += f;

            filterIndex++;
          }
        }
      }
      else {

        for (int j = 0; j < filterHeight; j++) {
          int2 inputPixelIndex;
          inputPixelIndex.y = imageIndex.y - midFilterDimen.y + j;
          inputPixelIndex.y = ClampToCanvas(inputPixelIndex.y, imageHeight);
          for (int i = 0; i < filterWidth; i++) {
            inputPixelIndex.x = imageIndex.x - midFilterDimen.x + i;
            inputPixelIndex.x = ClampToCanvas(inputPixelIndex.x, imageWidth);
        
            CLPixelType p = input[inputPixelIndex.y * imageWidth + inputPixelIndex.x];
            float alpha = QuantumScale*(QuantumRange-p.w);
            float f = filter[filterIndex];
            float g = alpha * f;

            sum.x += g*p.x;
            sum.y += g*p.y;
            sum.z += g*p.z;
            sum.w += f*p.w;

            gamma += g;


            filterIndex++;
          }
        }
        gamma = PerceptibleReciprocal(gamma);
        sum.xyz = gamma*sum.xyz;
      }

      CLPixelType outputPixel;
      outputPixel.x = ClampToQuantum(sum.x);
      outputPixel.y = ClampToQuantum(sum.y);
      outputPixel.z = ClampToQuantum(sum.z);
      outputPixel.w = ((channel & OpacityChannel)!=0)?ClampToQuantum(sum.w):input[imageIndex.y * imageWidth + imageIndex.x].w;

      output[imageIndex.y * imageWidth + imageIndex.x] = outputPixel;
    }
  )

  STRINGIFY(
     typedef enum
     {
       UndefinedFunction,
       PolynomialFunction,
       SinusoidFunction,
       ArcsinFunction,
       ArctanFunction
     } MagickFunction;
  )

  STRINGIFY(

    /*
    apply FunctionImageChannel(braightness-contrast)
    */
    CLPixelType ApplyFunction(CLPixelType pixel,const MagickFunction function,
        const unsigned int number_parameters,
        __constant float *parameters)
      {
        float4 result = (float4) 0.0f;
        switch (function)
        {
        case PolynomialFunction:
          {
            for (unsigned int i=0; i < number_parameters; i++)
              result = result*QuantumScale*convert_float4(pixel) + parameters[i];
            result *= QuantumRange;
            break;
          }
        case SinusoidFunction:
          {
            float  freq,phase,ampl,bias;
            freq  = ( number_parameters >= 1 ) ? parameters[0] : 1.0f;
            phase = ( number_parameters >= 2 ) ? parameters[1] : 0.0f;
            ampl  = ( number_parameters >= 3 ) ? parameters[2] : 0.5f;
            bias  = ( number_parameters >= 4 ) ? parameters[3] : 0.5f;
            result = QuantumRange*(ampl*sin(2.0f*MagickPI*
              (freq*QuantumScale*convert_float4(pixel) + phase/360.0f)) + bias);
            break;
          }
        case ArcsinFunction:
          {
            float  width,range,center,bias;
            width  = ( number_parameters >= 1 ) ? parameters[0] : 1.0f;
            center = ( number_parameters >= 2 ) ? parameters[1] : 0.5f;
            range  = ( number_parameters >= 3 ) ? parameters[2] : 1.0f;
            bias   = ( number_parameters >= 4 ) ? parameters[3] : 0.5f;
            result = 2.0f/width*(QuantumScale*convert_float4(pixel) - center);
            result = range/MagickPI*asin(result)+bias;
            result.x = ( result.x <= -1.0f ) ? bias - range/2.0f : result.x;
            result.x = ( result.x >= 1.0f ) ? bias + range/2.0f : result.x;
            result.y = ( result.y <= -1.0f ) ? bias - range/2.0f : result.y;
            result.y = ( result.y >= 1.0f ) ? bias + range/2.0f : result.y;
            result.z = ( result.z <= -1.0f ) ? bias - range/2.0f : result.x;
            result.z = ( result.z >= 1.0f ) ? bias + range/2.0f : result.x;
            result.w = ( result.w <= -1.0f ) ? bias - range/2.0f : result.w;
            result.w = ( result.w >= 1.0f ) ? bias + range/2.0f : result.w;
      
            result *= QuantumRange;
            break;
          }
        case ArctanFunction:
          {
            float slope,range,center,bias;
            slope  = ( number_parameters >= 1 ) ? parameters[0] : 1.0f;
            center = ( number_parameters >= 2 ) ? parameters[1] : 0.5f;
            range  = ( number_parameters >= 3 ) ? parameters[2] : 1.0f;
            bias   = ( number_parameters >= 4 ) ? parameters[3] : 0.5f;
            result = MagickPI*slope*(QuantumScale*convert_float4(pixel)-center);
            result = QuantumRange*(range/MagickPI*atan(result) + bias);
            break;
          }
        case UndefinedFunction:
          break;
        }
        return (CLPixelType) (ClampToQuantum(result.x), ClampToQuantum(result.y),
          ClampToQuantum(result.z), ClampToQuantum(result.w));
      }
    )

    STRINGIFY(
    /*
    Improve brightness / contrast of the image
    channel : define which channel is improved
    function : the function called to enchance the brightness contrast
    number_parameters : numbers of parameters 
    parameters : the parameter
    */
    __kernel void FunctionImage(__global CLPixelType *im,
                                        const ChannelType channel, const MagickFunction function,
                                        const unsigned int number_parameters, __constant float *parameters)
      {
        const int x = get_global_id(0);  
        const int y = get_global_id(1);  
        const int columns = get_global_size(0);  
        const int c = x + y * columns;
        im[c] = ApplyFunction(im[c], function, number_parameters, parameters); 
      }
    )

    STRINGIFY(
    /*
    */
    __kernel void Equalize(__global CLPixelType * restrict im,
      const ChannelType channel,  
      __global CLPixelType * restrict equalize_map,
      const float4 white, const float4 black)
      {
        const int x = get_global_id(0);  
        const int y = get_global_id(1);  
        const int columns = get_global_size(0);  
        const int c = x + y * columns;

        uint ePos;
        CLPixelType oValue, eValue;
        CLQuantum red, green, blue, opacity;

        //read from global
        oValue=im[c];

        if ((channel & SyncChannels) != 0)
        {
          if (getRedF4(white) != getRedF4(black))
          {
            ePos = ScaleQuantumToMap(getRed(oValue)); 
            eValue = equalize_map[ePos];
            red = getRed(eValue);
            ePos = ScaleQuantumToMap(getGreen(oValue)); 
            eValue = equalize_map[ePos];
            green = getRed(eValue);
            ePos = ScaleQuantumToMap(getBlue(oValue)); 
            eValue = equalize_map[ePos];
            blue = getRed(eValue);
            ePos = ScaleQuantumToMap(getOpacity(oValue)); 
            eValue = equalize_map[ePos];
            opacity = getRed(eValue);
 
            //write back
            im[c]=(CLPixelType)(blue, green, red, opacity);
          }

        }

        // for equalizing, we always need all channels?
        // otherwise something more

     }
    )

    STRINGIFY(
    /*
    */
    __kernel void Histogram(__global CLPixelType * restrict im,
      const ChannelType channel, const int colorspace,
      __global uint4 * restrict histogram)
      {
        const int x = get_global_id(0);  
        const int y = get_global_id(1);  
        const int columns = get_global_size(0);  
        const int c = x + y * columns;
        if ((channel & SyncChannels) != 0)
        {
          float intensity = GetPixelIntensity(colorspace,im[c]);
          uint pos = ScaleQuantumToMap(ClampToQuantum(intensity));
          atomic_inc((__global uint *)(&(histogram[pos]))+2); //red position
        }
        else
        {
          // for equalizing, we always need all channels?
          // otherwise something more
        }
      }
    )

    STRINGIFY(
      /*
      Reduce image noise and reduce detail levels by row
      im: input pixels filtered_in  filtered_im: output pixels
      filter : convolve kernel  width: convolve kernel size
      channel : define which channel is blured
      is_RGBA_BGRA : define the input is RGBA or BGRA
      */
      __kernel void BlurRow(__global CLPixelType *im, __global float4 *filtered_im,
                         const ChannelType channel, __constant float *filter,
                         const unsigned int width, 
                         const unsigned int imageColumns, const unsigned int imageRows,
                         __local CLPixelType *temp)
      {
        const int x = get_global_id(0);  
        const int y = get_global_id(1);  

        const int columns = imageColumns;  

        const unsigned int radius = (width-1)/2;
        const int wsize = get_local_size(0);  
        const unsigned int loadSize = wsize+width;

        //load chunk only for now
        //event_t e = async_work_group_copy(temp+radius, im+x+y*columns, wsize, 0);
        //wait_group_events(1,&e);

        //parallel load and clamp
        /*
        int count = 0;
        for (int i=0; i < loadSize; i=i+wsize)
        {
          int currentX = x + wsize*(count++);

          int localId = get_local_id(0);

          if ((localId+i) > loadSize)
            break;

          temp[localId+i] = im[y*columns+ClampToCanvas(currentX-radius, columns)];

          if (y==0 && get_group_id(0) == 0)
          {
            printf("(%d %d) temp %d load %d currentX %d\n", x, y, localId+i, ClampToCanvas(currentX-radius, columns), currentX);
          }
        }
        */

        //group coordinate
        const int groupX=get_local_size(0)*get_group_id(0);
        const int groupY=get_local_size(1)*get_group_id(1);

        //parallel load and clamp
        for (int i=get_local_id(0); i < loadSize; i=i+get_local_size(0))
        {
          //int cx = ClampToCanvas(groupX+i, columns);
          temp[i] = im[y * columns + ClampToCanvas(i+groupX-radius, columns)];

          if (0 && y==0 && get_group_id(1) == 0)
          {
            printf("(%d %d) temp %d load %d groupX %d\n", x, y, i, ClampToCanvas(groupX+i, columns), groupX);
          }
        }

        // barrier        
        barrier(CLK_LOCAL_MEM_FENCE);

        // only do the work if this is not a patched item
        if (get_global_id(0) < columns) 
        {
          // compute
          float4 result = (float4) 0;

          int i = 0;
          
          \n #ifndef UFACTOR   \n 
          \n #define UFACTOR 8 \n 
          \n #endif                  \n 

          for ( ; i+UFACTOR < width; ) 
          {
            \n #pragma unroll UFACTOR\n
            for (int j=0; j < UFACTOR; j++, i++)
            {
              result+=filter[i]*convert_float4(temp[i+get_local_id(0)]);
            }
          }

          for ( ; i < width; i++)
          {
            result+=filter[i]*convert_float4(temp[i+get_local_id(0)]);
          }

          result.x = ClampToQuantum(result.x);
          result.y = ClampToQuantum(result.y);
          result.z = ClampToQuantum(result.z);
          result.w = ClampToQuantum(result.w);

          // write back to global
          filtered_im[y*columns+x] = result;
        }
      }
    )

    STRINGIFY(
      /*
      Reduce image noise and reduce detail levels by row
      im: input pixels filtered_in  filtered_im: output pixels
      filter : convolve kernel  width: convolve kernel size
      channel : define which channel is blured
      is_RGBA_BGRA : define the input is RGBA or BGRA
      */
      __kernel void BlurRowSection(__global CLPixelType *im, __global float4 *filtered_im,
                         const ChannelType channel, __constant float *filter,
                         const unsigned int width, 
                         const unsigned int imageColumns, const unsigned int imageRows,
                         __local CLPixelType *temp, 
                         const unsigned int offsetRows, const unsigned int section)
      {
        const int x = get_global_id(0);  
        const int y = get_global_id(1);  

        const int columns = imageColumns;  

        const unsigned int radius = (width-1)/2;
        const int wsize = get_local_size(0);  
        const unsigned int loadSize = wsize+width;

        //group coordinate
        const int groupX=get_local_size(0)*get_group_id(0);
        const int groupY=get_local_size(1)*get_group_id(1);

        //offset the input data, assuming section is 0, 1 
        im += imageColumns * (offsetRows - radius * section);

        //parallel load and clamp
        for (int i=get_local_id(0); i < loadSize; i=i+get_local_size(0))
        {
          //int cx = ClampToCanvas(groupX+i, columns);
          temp[i] = im[y * columns + ClampToCanvas(i+groupX-radius, columns)];

          if (0 && y==0 && get_group_id(1) == 0)
          {
            printf("(%d %d) temp %d load %d groupX %d\n", x, y, i, ClampToCanvas(groupX+i, columns), groupX);
          }
        }

        // barrier        
        barrier(CLK_LOCAL_MEM_FENCE);

        // only do the work if this is not a patched item
        if (get_global_id(0) < columns) 
        {
          // compute
          float4 result = (float4) 0;

          int i = 0;
          
          \n #ifndef UFACTOR   \n 
          \n #define UFACTOR 8 \n 
          \n #endif                  \n 

          for ( ; i+UFACTOR < width; ) 
          {
            \n #pragma unroll UFACTOR\n
            for (int j=0; j < UFACTOR; j++, i++)
            {
              result+=filter[i]*convert_float4(temp[i+get_local_id(0)]);
            }
          }

          for ( ; i < width; i++)
          {
            result+=filter[i]*convert_float4(temp[i+get_local_id(0)]);
          }

          result.x = ClampToQuantum(result.x);
          result.y = ClampToQuantum(result.y);
          result.z = ClampToQuantum(result.z);
          result.w = ClampToQuantum(result.w);

          // write back to global
          filtered_im[y*columns+x] = result;
        }

      }
    )

    STRINGIFY(
      /*
      Reduce image noise and reduce detail levels by line
      im: input pixels filtered_in  filtered_im: output pixels
      filter : convolve kernel  width: convolve kernel size
      channel : define which channel is blured\
      is_RGBA_BGRA : define the input is RGBA or BGRA
      */
      __kernel void BlurColumn(const __global float4 *blurRowData, __global CLPixelType *filtered_im,
                                const ChannelType channel, __constant float *filter,
                                const unsigned int width, 
                                const unsigned int imageColumns, const unsigned int imageRows,
                                __local float4 *temp)
      {
        const int x = get_global_id(0);  
        const int y = get_global_id(1);

        //const int columns = get_global_size(0);
        //const int rows = get_global_size(1);  
        const int columns = imageColumns;  
        const int rows = imageRows;  

        unsigned int radius = (width-1)/2;
        const int wsize = get_local_size(1);  
        const unsigned int loadSize = wsize+width;

        //group coordinate
        const int groupX=get_local_size(0)*get_group_id(0);
        const int groupY=get_local_size(1)*get_group_id(1);
        //notice that get_local_size(0) is 1, so
        //groupX=get_group_id(0);
        
        //parallel load and clamp
        for (int i = get_local_id(1); i < loadSize; i=i+get_local_size(1))
        {
          temp[i] = blurRowData[ClampToCanvas(i+groupY-radius, rows) * columns + groupX];
        }
        
        // barrier        
        barrier(CLK_LOCAL_MEM_FENCE);

        // only do the work if this is not a patched item
        if (get_global_id(1) < rows)
        {
          // compute
          float4 result = (float4) 0;

          int i = 0;
          
          \n #ifndef UFACTOR   \n 
          \n #define UFACTOR 8 \n 
          \n #endif                  \n 
          
          for ( ; i+UFACTOR < width; ) 
          {
            \n #pragma unroll UFACTOR \n
            for (int j=0; j < UFACTOR; j++, i++)
            {
              result+=filter[i]*temp[i+get_local_id(1)];
            }
          }

          for ( ; i < width; i++)
          {
            result+=filter[i]*temp[i+get_local_id(1)];
          }

          result.x = ClampToQuantum(result.x);
          result.y = ClampToQuantum(result.y);
          result.z = ClampToQuantum(result.z);
          result.w = ClampToQuantum(result.w);

          // write back to global
          filtered_im[y*columns+x] = (CLPixelType) (result.x,result.y,result.z,result.w);
        }

      }
    )


    STRINGIFY(
      /*
      Reduce image noise and reduce detail levels by line
      im: input pixels filtered_in  filtered_im: output pixels
      filter : convolve kernel  width: convolve kernel size
      channel : define which channel is blured\
      is_RGBA_BGRA : define the input is RGBA or BGRA
      */
      __kernel void BlurColumnSection(const __global float4 *blurRowData, __global CLPixelType *filtered_im,
                                const ChannelType channel, __constant float *filter,
                                const unsigned int width, 
                                const unsigned int imageColumns, const unsigned int imageRows,
                                __local float4 *temp, 
                                const unsigned int offsetRows, const unsigned int section)
      {
        const int x = get_global_id(0);  
        const int y = get_global_id(1);

        //const int columns = get_global_size(0);
        //const int rows = get_global_size(1);  
        const int columns = imageColumns;  
        const int rows = imageRows;  

        unsigned int radius = (width-1)/2;
        const int wsize = get_local_size(1);  
        const unsigned int loadSize = wsize+width;

        //group coordinate
        const int groupX=get_local_size(0)*get_group_id(0);
        const int groupY=get_local_size(1)*get_group_id(1);
        //notice that get_local_size(0) is 1, so
        //groupX=get_group_id(0);
       
        // offset the input data
        blurRowData += imageColumns * radius * section;

        //parallel load and clamp
        for (int i = get_local_id(1); i < loadSize; i=i+get_local_size(1))
        {
          int pos = ClampToCanvasWithHalo(i+groupY-radius, rows, radius, section) * columns + groupX;
          temp[i] = *(blurRowData+pos);
        }
        
        // barrier        
        barrier(CLK_LOCAL_MEM_FENCE);

        // only do the work if this is not a patched item
        if (get_global_id(1) < rows)
        {
          // compute
          float4 result = (float4) 0;

          int i = 0;
          
          \n #ifndef UFACTOR   \n 
          \n #define UFACTOR 8 \n 
          \n #endif                  \n 
          
          for ( ; i+UFACTOR < width; ) 
          {
            \n #pragma unroll UFACTOR \n
            for (int j=0; j < UFACTOR; j++, i++)
            {
              result+=filter[i]*temp[i+get_local_id(1)];
            }
          }
          for ( ; i < width; i++)
          {
            result+=filter[i]*temp[i+get_local_id(1)];
          }

          result.x = ClampToQuantum(result.x);
          result.y = ClampToQuantum(result.y);
          result.z = ClampToQuantum(result.z);
          result.w = ClampToQuantum(result.w);

          // offset the output data
          filtered_im += imageColumns * offsetRows;

          // write back to global
          filtered_im[y*columns+x] = (CLPixelType) (result.x,result.y,result.z,result.w);
        }

      }
    )


    STRINGIFY(
    __kernel void UnsharpMaskBlurColumn(const __global CLPixelType* inputImage, 
          const __global float4 *blurRowData, __global CLPixelType *filtered_im,
          const unsigned int imageColumns, const unsigned int imageRows, 
          __local float4* cachedData, __local float* cachedFilter,
          const ChannelType channel, const __global float *filter, const unsigned int width, 
          const float gain, const float threshold)
    {
      const unsigned int radius = (width-1)/2;

      // cache the pixel shared by the workgroup
      const int groupX = get_group_id(0);
      const int groupStartY = get_group_id(1)*get_local_size(1) - radius;
      const int groupStopY = (get_group_id(1)+1)*get_local_size(1) + radius;

      if (groupStartY >= 0
          && groupStopY < imageRows) {
        event_t e = async_work_group_strided_copy(cachedData
                                                ,blurRowData+groupStartY*imageColumns+groupX
                                                ,groupStopY-groupStartY,imageColumns,0);
        wait_group_events(1,&e);
      }
      else {
        for (int i = get_local_id(1); i < (groupStopY - groupStartY); i+=get_local_size(1)) {
          cachedData[i] = blurRowData[ClampToCanvas(groupStartY+i,imageRows)*imageColumns+ groupX];
        }
        barrier(CLK_LOCAL_MEM_FENCE);
      }
      // cache the filter as well
      event_t e = async_work_group_copy(cachedFilter,filter,width,0);
      wait_group_events(1,&e);

      // only do the work if this is not a patched item
      //const int cy = get_group_id(1)*get_local_size(1)+get_local_id(1);
      const int cy = get_global_id(1);

      if (cy < imageRows) {
        float4 blurredPixel = (float4) 0.0f;

        int i = 0;

        \n #ifndef UFACTOR   \n 
          \n #define UFACTOR 8 \n 
          \n #endif                  \n 

          for ( ; i+UFACTOR < width; ) 
          {
            \n #pragma unroll UFACTOR \n
              for (int j=0; j < UFACTOR; j++, i++)
              {
                blurredPixel+=cachedFilter[i]*cachedData[i+get_local_id(1)];
              }
          }

        for ( ; i < width; i++)
        {
          blurredPixel+=cachedFilter[i]*cachedData[i+get_local_id(1)];
        }

        blurredPixel = floor((float4)(ClampToQuantum(blurredPixel.x), ClampToQuantum(blurredPixel.y)
                                      ,ClampToQuantum(blurredPixel.z), ClampToQuantum(blurredPixel.w)));

        float4 inputImagePixel = convert_float4(inputImage[cy*imageColumns+groupX]);
        float4 outputPixel = inputImagePixel - blurredPixel;

        float quantumThreshold = QuantumRange*threshold;

        int4 mask = isless(fabs(2.0f*outputPixel), (float4)quantumThreshold);
        outputPixel = select(inputImagePixel + outputPixel * gain, inputImagePixel, mask);

        //write back
        filtered_im[cy*imageColumns+groupX] = (CLPixelType) (ClampToQuantum(outputPixel.x), ClampToQuantum(outputPixel.y)
                                                            ,ClampToQuantum(outputPixel.z), ClampToQuantum(outputPixel.w));

      }
    }

    __kernel void UnsharpMaskBlurColumnSection(const __global CLPixelType* inputImage, 
          const __global float4 *blurRowData, __global CLPixelType *filtered_im,
          const unsigned int imageColumns, const unsigned int imageRows, 
          __local float4* cachedData, __local float* cachedFilter,
          const ChannelType channel, const __global float *filter, const unsigned int width, 
          const float gain, const float threshold, 
          const unsigned int offsetRows, const unsigned int section)
    {
      const unsigned int radius = (width-1)/2;

      // cache the pixel shared by the workgroup
      const int groupX = get_group_id(0);
      const int groupStartY = get_group_id(1)*get_local_size(1) - radius;
      const int groupStopY = (get_group_id(1)+1)*get_local_size(1) + radius;

      // offset the input data
      blurRowData += imageColumns * radius * section;

      if (groupStartY >= 0
          && groupStopY < imageRows) {
        event_t e = async_work_group_strided_copy(cachedData
                                                ,blurRowData+groupStartY*imageColumns+groupX
                                                ,groupStopY-groupStartY,imageColumns,0);
        wait_group_events(1,&e);
      }
      else {
        for (int i = get_local_id(1); i < (groupStopY - groupStartY); i+=get_local_size(1)) {
          int pos = ClampToCanvasWithHalo(groupStartY+i,imageRows, radius, section)*imageColumns+ groupX;
          cachedData[i] = *(blurRowData + pos);
        }
        barrier(CLK_LOCAL_MEM_FENCE);
      }
      // cache the filter as well
      event_t e = async_work_group_copy(cachedFilter,filter,width,0);
      wait_group_events(1,&e);

      // only do the work if this is not a patched item
      //const int cy = get_group_id(1)*get_local_size(1)+get_local_id(1);
      const int cy = get_global_id(1);

      if (cy < imageRows) {
        float4 blurredPixel = (float4) 0.0f;

        int i = 0;

        \n #ifndef UFACTOR   \n 
          \n #define UFACTOR 8 \n 
          \n #endif                  \n 

          for ( ; i+UFACTOR < width; ) 
          {
            \n #pragma unroll UFACTOR \n
              for (int j=0; j < UFACTOR; j++, i++)
              {
                blurredPixel+=cachedFilter[i]*cachedData[i+get_local_id(1)];
              }
          }

        for ( ; i < width; i++)
        {
          blurredPixel+=cachedFilter[i]*cachedData[i+get_local_id(1)];
        }

        blurredPixel = floor((float4)(ClampToQuantum(blurredPixel.x), ClampToQuantum(blurredPixel.y)
                                      ,ClampToQuantum(blurredPixel.z), ClampToQuantum(blurredPixel.w)));

        // offset the output data
        inputImage += imageColumns * offsetRows; 
        filtered_im += imageColumns * offsetRows;

        float4 inputImagePixel = convert_float4(inputImage[cy*imageColumns+groupX]);
        float4 outputPixel = inputImagePixel - blurredPixel;

        float quantumThreshold = QuantumRange*threshold;

        int4 mask = isless(fabs(2.0f*outputPixel), (float4)quantumThreshold);
        outputPixel = select(inputImagePixel + outputPixel * gain, inputImagePixel, mask);

        //write back
        filtered_im[cy*imageColumns+groupX] = (CLPixelType) (ClampToQuantum(outputPixel.x), ClampToQuantum(outputPixel.y)
                                                            ,ClampToQuantum(outputPixel.z), ClampToQuantum(outputPixel.w));

      }
     
    }
    )



  STRINGIFY(

  __kernel void HullPass1(const __global CLPixelType *inputImage, __global CLPixelType *outputImage
  , const unsigned int imageWidth, const unsigned int imageHeight
  , const int2 offset, const int polarity, const int matte) {

    int x = get_global_id(0);
    int y = get_global_id(1);

    CLPixelType v = inputImage[y*imageWidth+x];

    int2 neighbor;
    neighbor.y = y + offset.y;
    neighbor.x = x + offset.x;

    int2 clampedNeighbor;
    clampedNeighbor.x = ClampToCanvas(neighbor.x, imageWidth);
    clampedNeighbor.y = ClampToCanvas(neighbor.y, imageHeight);

    CLPixelType r = (clampedNeighbor.x == neighbor.x
                     && clampedNeighbor.y == neighbor.y)?inputImage[clampedNeighbor.y*imageWidth+clampedNeighbor.x]
    :(CLPixelType)0;

    int sv[4];
    sv[0] = (int)v.x;
    sv[1] = (int)v.y;
    sv[2] = (int)v.z;
    sv[3] = (int)v.w;

    int sr[4];
    sr[0] = (int)r.x;
    sr[1] = (int)r.y;
    sr[2] = (int)r.z;
    sr[3] = (int)r.w;

    if (polarity > 0) {
      \n #pragma unroll 4\n
      for (unsigned int i = 0; i < 4; i++) {
        sv[i] = (sr[i] >= (sv[i]+ScaleCharToQuantum(2)))?(sv[i]+ScaleCharToQuantum(1)):sv[i];
      }
    }
    else {
      \n #pragma unroll 4\n
      for (unsigned int i = 0; i < 4; i++) {
        sv[i] = (sr[i] <= (sv[i]-ScaleCharToQuantum(2)))?(sv[i]-ScaleCharToQuantum(1)):sv[i];
      }

    }

    v.x = (CLQuantum)sv[0];
    v.y = (CLQuantum)sv[1];
    v.z = (CLQuantum)sv[2];

    if (matte!=0)
      v.w = (CLQuantum)sv[3];

    outputImage[y*imageWidth+x] = v;

    }


  )



  STRINGIFY(

  __kernel void HullPass2(const __global CLPixelType *inputImage, __global CLPixelType *outputImage
  , const unsigned int imageWidth, const unsigned int imageHeight
  , const int2 offset, const int polarity, const int matte) {

    int x = get_global_id(0);
    int y = get_global_id(1);

    CLPixelType v = inputImage[y*imageWidth+x];

    int2 neighbor, clampedNeighbor;

    neighbor.y = y + offset.y;
    neighbor.x = x + offset.x;
    clampedNeighbor.x = ClampToCanvas(neighbor.x, imageWidth);
    clampedNeighbor.y = ClampToCanvas(neighbor.y, imageHeight);

    CLPixelType r = (clampedNeighbor.x == neighbor.x
      && clampedNeighbor.y == neighbor.y)?inputImage[clampedNeighbor.y*imageWidth+clampedNeighbor.x]
    :(CLPixelType)0;


    neighbor.y = y - offset.y;
    neighbor.x = x - offset.x;
    clampedNeighbor.x = ClampToCanvas(neighbor.x, imageWidth);
    clampedNeighbor.y = ClampToCanvas(neighbor.y, imageHeight);

    CLPixelType s = (clampedNeighbor.x == neighbor.x
      && clampedNeighbor.y == neighbor.y)?inputImage[clampedNeighbor.y*imageWidth+clampedNeighbor.x]
    :(CLPixelType)0;


    int sv[4];
    sv[0] = (int)v.x;
    sv[1] = (int)v.y;
    sv[2] = (int)v.z;
    sv[3] = (int)v.w;

    int sr[4];
    sr[0] = (int)r.x;
    sr[1] = (int)r.y;
    sr[2] = (int)r.z;
    sr[3] = (int)r.w;

    int ss[4];
    ss[0] = (int)s.x;
    ss[1] = (int)s.y;
    ss[2] = (int)s.z;
    ss[3] = (int)s.w;

    if (polarity > 0) {
      \n #pragma unroll 4\n
      for (unsigned int i = 0; i < 4; i++) {
        //sv[i] = (ss[i] >= (sv[i]+ScaleCharToQuantum(2)) && sr[i] > sv[i] )   ? (sv[i]+ScaleCharToQuantum(1)):sv[i];
        //
        //sv[i] =(!( (int)(ss[i] >= (sv[i]+ScaleCharToQuantum(2))) && (int) (sr[i] > sv[i] ) ))  ? sv[i]:(sv[i]+ScaleCharToQuantum(1));
        //sv[i] =(( (int)( ss[i] < (sv[i]+ScaleCharToQuantum(2))) || (int) ( sr[i] <= sv[i] ) ))  ? sv[i]:(sv[i]+ScaleCharToQuantum(1));
        sv[i] =(( (int)( ss[i] < (sv[i]+ScaleCharToQuantum(2))) + (int) ( sr[i] <= sv[i] ) ) !=0)  ? sv[i]:(sv[i]+ScaleCharToQuantum(1));
      }
    }
    else {
      \n #pragma unroll 4\n
      for (unsigned int i = 0; i < 4; i++) {
        //sv[i] = (ss[i] <= (sv[i]-ScaleCharToQuantum(2)) && sr[i] < sv[i] )   ? (sv[i]-ScaleCharToQuantum(1)):sv[i];
        //
        //sv[i] = ( (int)(ss[i] <= (sv[i]-ScaleCharToQuantum(2)) ) + (int)( sr[i] < sv[i] ) ==0)   ? sv[i]:(sv[i]-ScaleCharToQuantum(1));
        sv[i] = (( (int)(ss[i] > (sv[i]-ScaleCharToQuantum(2))) + (int)( sr[i] >= sv[i] )) !=0)   ? sv[i]:(sv[i]-ScaleCharToQuantum(1));
      }
    }

    v.x = (CLQuantum)sv[0];
    v.y = (CLQuantum)sv[1];
    v.z = (CLQuantum)sv[2];

    if (matte!=0)
      v.w = (CLQuantum)sv[3];

    outputImage[y*imageWidth+x] = v;

    }


  )

 
  STRINGIFY(
    __kernel void RadialBlur(const __global CLPixelType *im, __global CLPixelType *filtered_im,
                              const float4 bias,
                              const unsigned int channel, const unsigned int matte,
                              const float2 blurCenter,
                              __constant float *cos_theta, __constant float *sin_theta, 
                              const unsigned int cossin_theta_size)
      {
        const int x = get_global_id(0);  
        const int y = get_global_id(1);
        const int columns = get_global_size(0);
        const int rows = get_global_size(1);  
        unsigned int step = 1;
        float center_x = (float) x - blurCenter.x;
        float center_y = (float) y - blurCenter.y;
        float radius = hypot(center_x, center_y);
        
        //float blur_radius = hypot((float) columns/2.0f, (float) rows/2.0f);
        float blur_radius = hypot(blurCenter.x, blurCenter.y);

        if (radius > MagickEpsilon)
        {
          step = (unsigned int) (blur_radius / radius);
          if (step == 0)
            step = 1;
          if (step >= cossin_theta_size)
            step = cossin_theta_size-1;
        }

        float4 result;
        result.x = (float)bias.x;
        result.y = (float)bias.y;
        result.z = (float)bias.z;
        result.w = (float)bias.w;
        float normalize = 0.0f;

        if (((channel & OpacityChannel) == 0) || (matte == 0)) {
          for (unsigned int i=0; i<cossin_theta_size; i+=step)
          {
            result += convert_float4(im[
              ClampToCanvas(blurCenter.x+center_x*cos_theta[i]-center_y*sin_theta[i]+0.5f,columns)+ 
                ClampToCanvas(blurCenter.y+center_x*sin_theta[i]+center_y*cos_theta[i]+0.5f, rows)*columns]);
              normalize += 1.0f;
          }
          normalize = PerceptibleReciprocal(normalize);
          result = result * normalize;
        }
        else {
          float gamma = 0.0f;
          for (unsigned int i=0; i<cossin_theta_size; i+=step)
          {
            float4 p = convert_float4(im[
              ClampToCanvas(blurCenter.x+center_x*cos_theta[i]-center_y*sin_theta[i]+0.5f,columns)+ 
                ClampToCanvas(blurCenter.y+center_x*sin_theta[i]+center_y*cos_theta[i]+0.5f, rows)*columns]);
            
            float alpha = (float)(QuantumScale*(QuantumRange-p.w));
            result.x += alpha * p.x;
            result.y += alpha * p.y;
            result.z += alpha * p.z;
            result.w += p.w;
            gamma+=alpha;
            normalize += 1.0f;
          }
          gamma = PerceptibleReciprocal(gamma);
          normalize = PerceptibleReciprocal(normalize);
          result.x = gamma*result.x;
          result.y = gamma*result.y;
          result.z = gamma*result.z;
          result.w = normalize*result.w;
        }
        filtered_im[y * columns + x] = (CLPixelType) (ClampToQuantum(result.x), ClampToQuantum(result.y),
          ClampToQuantum(result.z), ClampToQuantum(result.w)); 
      }
  )
 
  STRINGIFY(
  typedef enum
  {
    UndefinedColorspace,
    RGBColorspace,            /* Linear RGB colorspace */
    GRAYColorspace,           /* greyscale (linear) image (faked 1 channel) */
    TransparentColorspace,
    OHTAColorspace,
    LabColorspace,
    XYZColorspace,
    YCbCrColorspace,
    YCCColorspace,
    YIQColorspace,
    YPbPrColorspace,
    YUVColorspace,
    CMYKColorspace,           /* negared linear RGB with black separated */
    sRGBColorspace,           /* Default: non-lienar sRGB colorspace */
    HSBColorspace,
    HSLColorspace,
    HWBColorspace,
    Rec601LumaColorspace,
    Rec601YCbCrColorspace,
    Rec709LumaColorspace,
    Rec709YCbCrColorspace,
    LogColorspace,
    CMYColorspace,            /* negated linear RGB colorspace */
    LuvColorspace,
    HCLColorspace,
    LCHColorspace,            /* alias for LCHuv */
    LMSColorspace,
    LCHabColorspace,          /* Cylindrical (Polar) Lab */
    LCHuvColorspace,          /* Cylindrical (Polar) Luv */
    scRGBColorspace,
    HSIColorspace,
    HSVColorspace,            /* alias for HSB */
    HCLpColorspace,
    YDbDrColorspace
  } ColorspaceType;
  )


  STRINGIFY(

  inline float3 ConvertRGBToHSB(CLPixelType pixel) {
    float3 HueSaturationBrightness;
    HueSaturationBrightness.x = 0.0f; // Hue
    HueSaturationBrightness.y = 0.0f; // Saturation
    HueSaturationBrightness.z = 0.0f; // Brightness

    float r=(float) getRed(pixel);
    float g=(float) getGreen(pixel);
    float b=(float) getBlue(pixel);

    float tmin=min(min(r,g),b);
    float tmax=max(max(r,g),b);

    if (tmax!=0.0f) {
      float delta=tmax-tmin;
      HueSaturationBrightness.y=delta/tmax;
      HueSaturationBrightness.z=QuantumScale*tmax;

      if (delta != 0.0f) {
	HueSaturationBrightness.x = ((r == tmax)?0.0f:((g == tmax)?2.0f:4.0f));
	HueSaturationBrightness.x += ((r == tmax)?(g-b):((g == tmax)?(b-r):(r-g)))/delta;
        HueSaturationBrightness.x/=6.0f;
        HueSaturationBrightness.x += (HueSaturationBrightness.x < 0.0f)?0.0f:1.0f;
      }
    }
    return HueSaturationBrightness;
  }

  inline CLPixelType ConvertHSBToRGB(float3 HueSaturationBrightness) {

    float hue = HueSaturationBrightness.x;
    float brightness = HueSaturationBrightness.z;
    float saturation = HueSaturationBrightness.y;
   
    CLPixelType rgb;

    if (saturation == 0.0f) {
      setRed(&rgb,ClampToQuantum(QuantumRange*brightness));
      setGreen(&rgb,getRed(rgb));
      setBlue(&rgb,getRed(rgb));
    }
    else {

      float h=6.0f*(hue-floor(hue));
      float f=h-floor(h);
      float p=brightness*(1.0f-saturation);
      float q=brightness*(1.0f-saturation*f);
      float t=brightness*(1.0f-(saturation*(1.0f-f)));
 
      float clampedBrightness = ClampToQuantum(QuantumRange*brightness);
      float clamped_t = ClampToQuantum(QuantumRange*t);
      float clamped_p = ClampToQuantum(QuantumRange*p);
      float clamped_q = ClampToQuantum(QuantumRange*q);     
      int ih = (int)h;
      setRed(&rgb, (ih == 1)?clamped_q:
	      (ih == 2 || ih == 3)?clamped_p:
	      (ih == 4)?clamped_t:
                 clampedBrightness);
 
      setGreen(&rgb, (ih == 1 || ih == 2)?clampedBrightness:
	      (ih == 3)?clamped_q:
	      (ih == 4 || ih == 5)?clamped_p:
                 clamped_t);

      setBlue(&rgb, (ih == 2)?clamped_t:
	      (ih == 3 || ih == 4)?clampedBrightness:
	      (ih == 5)?clamped_q:
                 clamped_p);
    }
    return rgb;
  }

  __kernel void Contrast(__global CLPixelType *im, const unsigned int sharpen)
  {

    const int sign = sharpen!=0?1:-1;
    const int x = get_global_id(0);  
    const int y = get_global_id(1);
    const int columns = get_global_size(0);
    const int c = x + y * columns;

    CLPixelType pixel = im[c];
    float3 HueSaturationBrightness = ConvertRGBToHSB(pixel);
    float brightness = HueSaturationBrightness.z;
    brightness+=0.5f*sign*(0.5f*(sinpi(brightness-0.5f)+1.0f)-brightness);
    brightness = clamp(brightness,0.0f,1.0f);
    HueSaturationBrightness.z = brightness;

    CLPixelType filteredPixel = ConvertHSBToRGB(HueSaturationBrightness);
    filteredPixel.w = pixel.w;
    im[c] = filteredPixel;
  }


  )

  STRINGIFY(

  inline void ConvertRGBToHSL(const CLQuantum red,const CLQuantum green, const CLQuantum blue,
    float *hue, float *saturation, float *lightness)
  {
  float
    c,
    tmax,
    tmin;

  /*
     Convert RGB to HSL colorspace.
     */
  tmax=max(QuantumScale*red,max(QuantumScale*green, QuantumScale*blue));
  tmin=min(QuantumScale*red,min(QuantumScale*green, QuantumScale*blue));

  c=tmax-tmin;

  *lightness=(tmax+tmin)/2.0;
  if (c <= 0.0)
  {
    *hue=0.0;
    *saturation=0.0;
    return;
  }

  if (tmax == (QuantumScale*red))
  {
    *hue=(QuantumScale*green-QuantumScale*blue)/c;
    if ((QuantumScale*green) < (QuantumScale*blue))
      *hue+=6.0;
  }
  else
    if (tmax == (QuantumScale*green))
      *hue=2.0+(QuantumScale*blue-QuantumScale*red)/c;
    else
      *hue=4.0+(QuantumScale*red-QuantumScale*green)/c;

  *hue*=60.0/360.0;
  if (*lightness <= 0.5)
    *saturation=c/(2.0*(*lightness));
  else
    *saturation=c/(2.0-2.0*(*lightness));
  }

  inline void ConvertHSLToRGB(const float hue,const float saturation, const float lightness,
      CLQuantum *red,CLQuantum *green,CLQuantum *blue)
  {
    float
      b,
      c,
      g,
      h,
      tmin,
      r,
      x;

    /*
       Convert HSL to RGB colorspace.
       */
    h=hue*360.0;
    if (lightness <= 0.5)
      c=2.0*lightness*saturation;
    else
      c=(2.0-2.0*lightness)*saturation;
    tmin=lightness-0.5*c;
    h-=360.0*floor(h/360.0);
    h/=60.0;
    x=c*(1.0-fabs(h-2.0*floor(h/2.0)-1.0));
    switch ((int) floor(h))
    {
      case 0:
        {
          r=tmin+c;
          g=tmin+x;
          b=tmin;
          break;
        }
      case 1:
        {
          r=tmin+x;
          g=tmin+c;
          b=tmin;
          break;
        }
      case 2:
        {
          r=tmin;
          g=tmin+c;
          b=tmin+x;
          break;
        }
      case 3:
        {
          r=tmin;
          g=tmin+x;
          b=tmin+c;
          break;
        }
      case 4:
        {
          r=tmin+x;
          g=tmin;
          b=tmin+c;
          break;
        }
      case 5:
        {
          r=tmin+c;
          g=tmin;
          b=tmin+x;
          break;
        }
      default:
        {
          r=0.0;
          g=0.0;
          b=0.0;
        }
    }
    *red=ClampToQuantum(QuantumRange*r);
    *green=ClampToQuantum(QuantumRange*g);
    *blue=ClampToQuantum(QuantumRange*b);
  }

  inline void ModulateHSL(const float percent_hue, const float percent_saturation,const float percent_lightness, 
    CLQuantum *red,CLQuantum *green,CLQuantum *blue)
  {
    float
      hue,
      lightness,
      saturation;

    /*
    Increase or decrease color lightness, saturation, or hue.
    */
    ConvertRGBToHSL(*red,*green,*blue,&hue,&saturation,&lightness);
    hue+=0.5*(0.01*percent_hue-1.0);
    while (hue < 0.0)
      hue+=1.0;
    while (hue >= 1.0)
      hue-=1.0;
    saturation*=0.01*percent_saturation;
    lightness*=0.01*percent_lightness;
    ConvertHSLToRGB(hue,saturation,lightness,red,green,blue);
  }

  __kernel void Modulate(__global CLPixelType *im, 
    const float percent_brightness, 
    const float percent_hue, 
    const float percent_saturation, 
    const int colorspace)
  {

    const int x = get_global_id(0);  
    const int y = get_global_id(1);
    const int columns = get_global_size(0);
    const int c = x + y * columns;

    CLPixelType pixel = im[c];

    CLQuantum
        blue,
        green,
        red;

    red=getRed(pixel);
    green=getGreen(pixel);
    blue=getBlue(pixel);

    switch (colorspace)
    {
      case HSLColorspace:
      default:
        {
          ModulateHSL(percent_hue, percent_saturation, percent_brightness, 
              &red, &green, &blue);
        }

    }

    CLPixelType filteredPixel;
   
    setRed(&filteredPixel, red);
    setGreen(&filteredPixel, green);
    setBlue(&filteredPixel, blue);
    filteredPixel.w = pixel.w;

    im[c] = filteredPixel;
  }
  )

  STRINGIFY(
  // Based on Box from resize.c
  float BoxResizeFilter(const float x)
  {
    return 1.0f;
  }
  )
    
  STRINGIFY(
  // Based on CubicBC from resize.c
  float CubicBC(const float x,const __global float* resizeFilterCoefficients)
  {
    /*
    Cubic Filters using B,C determined values:
    Mitchell-Netravali  B = 1/3 C = 1/3  "Balanced" cubic spline filter
    Catmull-Rom         B = 0   C = 1/2  Interpolatory and exact on linears
    Spline              B = 1   C = 0    B-Spline Gaussian approximation
    Hermite             B = 0   C = 0    B-Spline interpolator

    See paper by Mitchell and Netravali, Reconstruction Filters in Computer
    Graphics Computer Graphics, Volume 22, Number 4, August 1988
    http://www.cs.utexas.edu/users/fussell/courses/cs384g/lectures/mitchell/
    Mitchell.pdf.

    Coefficents are determined from B,C values:
    P0 = (  6 - 2*B       )/6 = coeff[0]
    P1 =         0
    P2 = (-18 +12*B + 6*C )/6 = coeff[1]
    P3 = ( 12 - 9*B - 6*C )/6 = coeff[2]
    Q0 = (      8*B +24*C )/6 = coeff[3]
    Q1 = (    -12*B -48*C )/6 = coeff[4]
    Q2 = (      6*B +30*C )/6 = coeff[5]
    Q3 = (    - 1*B - 6*C )/6 = coeff[6]

    which are used to define the filter:

    P0 + P1*x + P2*x^2 + P3*x^3      0 <= x < 1
    Q0 + Q1*x + Q2*x^2 + Q3*x^3      1 <= x < 2

    which ensures function is continuous in value and derivative (slope).
    */
    if (x < 1.0)
      return(resizeFilterCoefficients[0]+x*(x*
      (resizeFilterCoefficients[1]+x*resizeFilterCoefficients[2])));
    if (x < 2.0)
      return(resizeFilterCoefficients[3]+x*(resizeFilterCoefficients[4]+x*
      (resizeFilterCoefficients[5]+x*resizeFilterCoefficients[6])));
    return(0.0);
  }
  )

  STRINGIFY(
  float Sinc(const float x)
  {
    if (x != 0.0f)
    {
      const float alpha=(float) (MagickPI*x);
      return sinpi(x)/alpha;
    }
    return(1.0f);
  }
  )

  STRINGIFY(
  float Triangle(const float x)
  {
    /*
    1st order (linear) B-Spline, bilinear interpolation, Tent 1D filter, or
    a Bartlett 2D Cone filter.  Also used as a Bartlett Windowing function
    for Sinc().
    */
    return ((x<1.0f)?(1.0f-x):0.0f);
  }
  )


  STRINGIFY(
  float Hanning(const float x)
  {
    /*
    Cosine window function:
      0.5+0.5*cos(pi*x).
    */
    const float cosine=cos((MagickPI*x));
    return(0.5f+0.5f*cosine);
  }
  )

  STRINGIFY(
  float Hamming(const float x)
  {
    /*
      Offset cosine window function:
       .54 + .46 cos(pi x).
    */
    const float cosine=cos((MagickPI*x));
    return(0.54f+0.46f*cosine);
  }
  )

  STRINGIFY(
  float Blackman(const float x)
  {
    /*
      Blackman: 2nd order cosine windowing function:
        0.42 + 0.5 cos(pi x) + 0.08 cos(2pi x)

      Refactored by Chantal Racette and Nicolas Robidoux to one trig call and
      five flops.
    */
    const float cosine=cos((MagickPI*x));
    return(0.34f+cosine*(0.5f+cosine*0.16f));
  }
  )


  STRINGIFY(
  typedef enum {
    BoxWeightingFunction = 0,
    TriangleWeightingFunction,
    CubicBCWeightingFunction,
    HanningWeightingFunction,
    HammingWeightingFunction,
    BlackmanWeightingFunction,
    GaussianWeightingFunction,
    QuadraticWeightingFunction,
    JincWeightingFunction,
    SincWeightingFunction,
    SincFastWeightingFunction,
    KaiserWeightingFunction,
    WelshWeightingFunction,
    BohmanWeightingFunction,
    LagrangeWeightingFunction,
    CosineWeightingFunction,
  } ResizeWeightingFunctionType;
  )

  STRINGIFY(
  inline float applyResizeFilter(const float x, const ResizeWeightingFunctionType filterType, const __global float* filterCoefficients)
  {
    switch (filterType)
    {
    /* Call Sinc even for SincFast to get better precision on GPU 
       and to avoid thread divergence.  Sinc is pretty fast on GPU anyway...*/
    case SincWeightingFunction:
    case SincFastWeightingFunction:  
      return Sinc(x);
    case CubicBCWeightingFunction:
      return CubicBC(x,filterCoefficients);
    case BoxWeightingFunction:
      return BoxResizeFilter(x);
    case TriangleWeightingFunction:
      return Triangle(x);
    case HanningWeightingFunction:
      return Hanning(x);
    case HammingWeightingFunction:
      return Hamming(x);
    case BlackmanWeightingFunction:
      return Blackman(x);

    default:
      return 0.0f;
    }
  }
  )


  STRINGIFY(
  inline float getResizeFilterWeight(const __global float* resizeFilterCubicCoefficients, const ResizeWeightingFunctionType resizeFilterType
           , const ResizeWeightingFunctionType resizeWindowType
           , const float resizeFilterScale, const float resizeWindowSupport, const float resizeFilterBlur, const float x)
  {
    float scale;
    float xBlur = fabs(x/resizeFilterBlur);
    if (resizeWindowSupport < MagickEpsilon
        || resizeWindowType == BoxWeightingFunction)
    {
      scale = 1.0f;
    }
    else
    {
      scale = resizeFilterScale;
      scale = applyResizeFilter(xBlur*scale, resizeWindowType, resizeFilterCubicCoefficients);
    }
    float weight = scale * applyResizeFilter(xBlur, resizeFilterType, resizeFilterCubicCoefficients);
    return weight;
  }

  )

  ;
  const char* accelerateKernels2 =

  STRINGIFY(

  inline unsigned int getNumWorkItemsPerPixel(const unsigned int pixelPerWorkgroup, const unsigned int numWorkItems) {
    return (numWorkItems/pixelPerWorkgroup);
  }

  // returns the index of the pixel for the current workitem to compute.
  // returns -1 if this workitem doesn't need to participate in any computation
  inline int pixelToCompute(const unsigned itemID, const unsigned int pixelPerWorkgroup, const unsigned int numWorkItems) {
    const unsigned int numWorkItemsPerPixel = getNumWorkItemsPerPixel(pixelPerWorkgroup, numWorkItems);
    int pixelIndex = itemID/numWorkItemsPerPixel;
    pixelIndex = (pixelIndex<pixelPerWorkgroup)?pixelIndex:-1;
    return pixelIndex;
  }
 
  )

  STRINGIFY(
 __kernel __attribute__((reqd_work_group_size(256, 1, 1)))
 void ResizeHorizontalFilter(const __global CLPixelType* inputImage, const unsigned int inputColumns, const unsigned int inputRows, const unsigned int matte
  , const float xFactor, __global CLPixelType* filteredImage, const unsigned int filteredColumns, const unsigned int filteredRows
  , const int resizeFilterType, const int resizeWindowType
  , const __global float* resizeFilterCubicCoefficients
  , const float resizeFilterScale, const float resizeFilterSupport, const float resizeFilterWindowSupport, const float resizeFilterBlur
  , __local CLPixelType* inputImageCache, const int numCachedPixels, const unsigned int pixelPerWorkgroup, const unsigned int pixelChunkSize
  , __local float4* outputPixelCache, __local float* densityCache, __local float* gammaCache) {


    // calculate the range of resized image pixels computed by this workgroup
    const unsigned int startX = get_group_id(0)*pixelPerWorkgroup;
    const unsigned int stopX = min(startX + pixelPerWorkgroup,filteredColumns);
    const unsigned int actualNumPixelToCompute = stopX - startX;

    // calculate the range of input image pixels to cache
    float scale = max(1.0/xFactor+MagickEpsilon ,1.0f);
    const float support = max(scale*resizeFilterSupport,0.5f);
    scale = PerceptibleReciprocal(scale);

    const int cacheRangeStartX = max((int)((startX+0.5f)/xFactor+MagickEpsilon-support+0.5f),(int)(0));
    const int cacheRangeEndX = min((int)(cacheRangeStartX + numCachedPixels), (int)inputColumns);

    // cache the input pixels into local memory
    const unsigned int y = get_global_id(1);
    event_t e = async_work_group_copy(inputImageCache,inputImage+y*inputColumns+cacheRangeStartX,cacheRangeEndX-cacheRangeStartX,0);
    wait_group_events(1,&e);

    unsigned int totalNumChunks = (actualNumPixelToCompute+pixelChunkSize-1)/pixelChunkSize;
    for (unsigned int chunk = 0; chunk < totalNumChunks; chunk++)
    {

      const unsigned int chunkStartX = startX + chunk*pixelChunkSize;
      const unsigned int chunkStopX = min(chunkStartX + pixelChunkSize, stopX);
      const unsigned int actualNumPixelInThisChunk = chunkStopX - chunkStartX;

      // determine which resized pixel computed by this workitem
      const unsigned int itemID = get_local_id(0);
      const unsigned int numItems = getNumWorkItemsPerPixel(actualNumPixelInThisChunk, get_local_size(0));
      
      const int pixelIndex = pixelToCompute(itemID, actualNumPixelInThisChunk, get_local_size(0));

      float4 filteredPixel = (float4)0.0f;
      float density = 0.0f;
      float gamma = 0.0f;
      // -1 means this workitem doesn't participate in the computation
      if (pixelIndex != -1) {

        // x coordinated of the resized pixel computed by this workitem
        const int x = chunkStartX + pixelIndex;

        // calculate how many steps required for this pixel
        const float bisect = (x+0.5)/xFactor+MagickEpsilon;
        const unsigned int start = (unsigned int)max(bisect-support+0.5f,0.0f);
        const unsigned int stop  = (unsigned int)min(bisect+support+0.5f,(float)inputColumns);
        const unsigned int n = stop - start;

        // calculate how many steps this workitem will contribute
        unsigned int numStepsPerWorkItem = n / numItems;
        numStepsPerWorkItem += ((numItems*numStepsPerWorkItem)==n?0:1);

        const unsigned int startStep = (itemID%numItems)*numStepsPerWorkItem;
        if (startStep < n) {
          const unsigned int stopStep = min(startStep+numStepsPerWorkItem, n);

          unsigned int cacheIndex = start+startStep-cacheRangeStartX;
          if (matte == 0) {

            for (unsigned int i = startStep; i < stopStep; i++,cacheIndex++) {
              float4 cp = convert_float4(inputImageCache[cacheIndex]);

              float weight = getResizeFilterWeight(resizeFilterCubicCoefficients,(ResizeWeightingFunctionType)resizeFilterType
                , (ResizeWeightingFunctionType)resizeWindowType
                , resizeFilterScale, resizeFilterWindowSupport, resizeFilterBlur,scale*(start+i-bisect+0.5));

              filteredPixel += ((float4)weight)*cp;
              density+=weight;
            }


          }
          else {
            for (unsigned int i = startStep; i < stopStep; i++,cacheIndex++) {
              CLPixelType p = inputImageCache[cacheIndex];

              float weight = getResizeFilterWeight(resizeFilterCubicCoefficients,(ResizeWeightingFunctionType)resizeFilterType
                , (ResizeWeightingFunctionType)resizeWindowType
                , resizeFilterScale, resizeFilterWindowSupport, resizeFilterBlur,scale*(start+i-bisect+0.5));

              float alpha = weight * QuantumScale * GetPixelAlpha(p);
              float4 cp = convert_float4(p);

              filteredPixel.x += alpha * cp.x;
              filteredPixel.y += alpha * cp.y;
              filteredPixel.z += alpha * cp.z;
              filteredPixel.w += weight * cp.w;

              density+=weight;
              gamma+=alpha;
            }
         }
      }
    }

    // initialize the accumulators to zero
    if (itemID < actualNumPixelInThisChunk) {
      outputPixelCache[itemID] = (float4)0.0f;
      densityCache[itemID] = 0.0f;
      if (matte != 0)
        gammaCache[itemID] = 0.0f;
    }
    barrier(CLK_LOCAL_MEM_FENCE);

    // accumulatte the filtered pixel value and the density
    for (unsigned int i = 0; i < numItems; i++) {
      if (pixelIndex != -1) {
        if (itemID%numItems == i) {
          outputPixelCache[pixelIndex]+=filteredPixel;
          densityCache[pixelIndex]+=density;
          if (matte!=0) {
            gammaCache[pixelIndex]+=gamma;
          }
        }
      }
      barrier(CLK_LOCAL_MEM_FENCE);
    }

    if (itemID < actualNumPixelInThisChunk) {
      if (matte==0) {
        float density = densityCache[itemID];
        float4 filteredPixel = outputPixelCache[itemID];
        if (density!= 0.0f && density != 1.0)
        {
          density = PerceptibleReciprocal(density);
          filteredPixel *= (float4)density;
        }
        filteredImage[y*filteredColumns+chunkStartX+itemID] = (CLPixelType) (ClampToQuantum(filteredPixel.x)
                                                                       , ClampToQuantum(filteredPixel.y)
                                                                       , ClampToQuantum(filteredPixel.z)
                                                                       , ClampToQuantum(filteredPixel.w));
      }
      else {
        float density = densityCache[itemID];
        float gamma = gammaCache[itemID];
        float4 filteredPixel = outputPixelCache[itemID];

        if (density!= 0.0f && density != 1.0) {
          density = PerceptibleReciprocal(density);
          filteredPixel *= (float4)density;
          gamma *= density;
        }
        gamma = PerceptibleReciprocal(gamma);

        CLPixelType fp;
        fp = (CLPixelType) ( ClampToQuantum(gamma*filteredPixel.x)
          , ClampToQuantum(gamma*filteredPixel.y)
          , ClampToQuantum(gamma*filteredPixel.z)
          , ClampToQuantum(filteredPixel.w));

        filteredImage[y*filteredColumns+chunkStartX+itemID] = fp;

      }
    }

    } // end of chunking loop
  }
  )



  STRINGIFY(
 __kernel __attribute__((reqd_work_group_size(256, 1, 1)))
 void ResizeHorizontalFilterSinc(const __global CLPixelType* inputImage, const unsigned int inputColumns, const unsigned int inputRows, const unsigned int matte
  , const float xFactor, __global CLPixelType* filteredImage, const unsigned int filteredColumns, const unsigned int filteredRows
  , const int resizeFilterType, const int resizeWindowType
  , const __global float* resizeFilterCubicCoefficients
  , const float resizeFilterScale, const float resizeFilterSupport, const float resizeFilterWindowSupport, const float resizeFilterBlur
  , __local CLPixelType* inputImageCache, const int numCachedPixels, const unsigned int pixelPerWorkgroup, const unsigned int pixelChunkSize
  , __local float4* outputPixelCache, __local float* densityCache, __local float* gammaCache) {
    
    ResizeHorizontalFilter(inputImage,inputColumns,inputRows,matte
    ,xFactor, filteredImage, filteredColumns, filteredRows
    ,SincWeightingFunction, SincWeightingFunction
    ,resizeFilterCubicCoefficients
    ,resizeFilterScale, resizeFilterSupport, resizeFilterWindowSupport, resizeFilterBlur
    ,inputImageCache, numCachedPixels, pixelPerWorkgroup, pixelChunkSize
    ,outputPixelCache, densityCache, gammaCache);

  }
  )


  STRINGIFY(
 __kernel __attribute__((reqd_work_group_size(1, 256, 1)))
 void ResizeVerticalFilter(const __global CLPixelType* inputImage, const unsigned int inputColumns, const unsigned int inputRows, const unsigned int matte
  , const float yFactor, __global CLPixelType* filteredImage, const unsigned int filteredColumns, const unsigned int filteredRows
  , const int resizeFilterType, const int resizeWindowType
  , const __global float* resizeFilterCubicCoefficients
  , const float resizeFilterScale, const float resizeFilterSupport, const float resizeFilterWindowSupport, const float resizeFilterBlur
  , __local CLPixelType* inputImageCache, const int numCachedPixels, const unsigned int pixelPerWorkgroup, const unsigned int pixelChunkSize
  , __local float4* outputPixelCache, __local float* densityCache, __local float* gammaCache) {


    // calculate the range of resized image pixels computed by this workgroup
    const unsigned int startY = get_group_id(1)*pixelPerWorkgroup;
    const unsigned int stopY = min(startY + pixelPerWorkgroup,filteredRows);
    const unsigned int actualNumPixelToCompute = stopY - startY;

    // calculate the range of input image pixels to cache
    float scale = max(1.0/yFactor+MagickEpsilon ,1.0f);
    const float support = max(scale*resizeFilterSupport,0.5f);
    scale = PerceptibleReciprocal(scale);

    const int cacheRangeStartY = max((int)((startY+0.5f)/yFactor+MagickEpsilon-support+0.5f),(int)(0));
    const int cacheRangeEndY = min((int)(cacheRangeStartY + numCachedPixels), (int)inputRows);

    // cache the input pixels into local memory
    const unsigned int x = get_global_id(0);
    event_t e = async_work_group_strided_copy(inputImageCache, inputImage+cacheRangeStartY*inputColumns+x, cacheRangeEndY-cacheRangeStartY, inputColumns, 0);
    wait_group_events(1,&e);

    unsigned int totalNumChunks = (actualNumPixelToCompute+pixelChunkSize-1)/pixelChunkSize;
    for (unsigned int chunk = 0; chunk < totalNumChunks; chunk++)
    {

      const unsigned int chunkStartY = startY + chunk*pixelChunkSize;
      const unsigned int chunkStopY = min(chunkStartY + pixelChunkSize, stopY);
      const unsigned int actualNumPixelInThisChunk = chunkStopY - chunkStartY;

      // determine which resized pixel computed by this workitem
      const unsigned int itemID = get_local_id(1);
      const unsigned int numItems = getNumWorkItemsPerPixel(actualNumPixelInThisChunk, get_local_size(1));
      
      const int pixelIndex = pixelToCompute(itemID, actualNumPixelInThisChunk, get_local_size(1));

      float4 filteredPixel = (float4)0.0f;
      float density = 0.0f;
      float gamma = 0.0f;
      // -1 means this workitem doesn't participate in the computation
      if (pixelIndex != -1) {

        // x coordinated of the resized pixel computed by this workitem
        const int y = chunkStartY + pixelIndex;

        // calculate how many steps required for this pixel
        const float bisect = (y+0.5)/yFactor+MagickEpsilon;
        const unsigned int start = (unsigned int)max(bisect-support+0.5f,0.0f);
        const unsigned int stop  = (unsigned int)min(bisect+support+0.5f,(float)inputRows);
        const unsigned int n = stop - start;

        // calculate how many steps this workitem will contribute
        unsigned int numStepsPerWorkItem = n / numItems;
        numStepsPerWorkItem += ((numItems*numStepsPerWorkItem)==n?0:1);

        const unsigned int startStep = (itemID%numItems)*numStepsPerWorkItem;
        if (startStep < n) {
          const unsigned int stopStep = min(startStep+numStepsPerWorkItem, n);

          unsigned int cacheIndex = start+startStep-cacheRangeStartY;
          if (matte == 0) {

            for (unsigned int i = startStep; i < stopStep; i++,cacheIndex++) {
              float4 cp = convert_float4(inputImageCache[cacheIndex]);

              float weight = getResizeFilterWeight(resizeFilterCubicCoefficients,(ResizeWeightingFunctionType)resizeFilterType
                , (ResizeWeightingFunctionType)resizeWindowType
                , resizeFilterScale, resizeFilterWindowSupport, resizeFilterBlur,scale*(start+i-bisect+0.5));

              filteredPixel += ((float4)weight)*cp;
              density+=weight;
            }


          }
          else {
            for (unsigned int i = startStep; i < stopStep; i++,cacheIndex++) {
              CLPixelType p = inputImageCache[cacheIndex];

              float weight = getResizeFilterWeight(resizeFilterCubicCoefficients,(ResizeWeightingFunctionType)resizeFilterType
                , (ResizeWeightingFunctionType)resizeWindowType
                , resizeFilterScale, resizeFilterWindowSupport, resizeFilterBlur,scale*(start+i-bisect+0.5));

              float alpha = weight * QuantumScale * GetPixelAlpha(p);
              float4 cp = convert_float4(p);

              filteredPixel.x += alpha * cp.x;
              filteredPixel.y += alpha * cp.y;
              filteredPixel.z += alpha * cp.z;
              filteredPixel.w += weight * cp.w;

              density+=weight;
              gamma+=alpha;
            }
         }
      }
    }

    // initialize the accumulators to zero
    if (itemID < actualNumPixelInThisChunk) {
      outputPixelCache[itemID] = (float4)0.0f;
      densityCache[itemID] = 0.0f;
      if (matte != 0)
        gammaCache[itemID] = 0.0f;
    }
    barrier(CLK_LOCAL_MEM_FENCE);

    // accumulatte the filtered pixel value and the density
    for (unsigned int i = 0; i < numItems; i++) {
      if (pixelIndex != -1) {
        if (itemID%numItems == i) {
          outputPixelCache[pixelIndex]+=filteredPixel;
          densityCache[pixelIndex]+=density;
          if (matte!=0) {
            gammaCache[pixelIndex]+=gamma;
          }
        }
      }
      barrier(CLK_LOCAL_MEM_FENCE);
    }

    if (itemID < actualNumPixelInThisChunk) {
      if (matte==0) {
        float density = densityCache[itemID];
        float4 filteredPixel = outputPixelCache[itemID];
        if (density!= 0.0f && density != 1.0)
        {
          density = PerceptibleReciprocal(density);
          filteredPixel *= (float4)density;
        }
        filteredImage[(chunkStartY+itemID)*filteredColumns+x] = (CLPixelType) (ClampToQuantum(filteredPixel.x)
                                                                       , ClampToQuantum(filteredPixel.y)
                                                                       , ClampToQuantum(filteredPixel.z)
                                                                       , ClampToQuantum(filteredPixel.w));
      }
      else {
        float density = densityCache[itemID];
        float gamma = gammaCache[itemID];
        float4 filteredPixel = outputPixelCache[itemID];

        if (density!= 0.0f && density != 1.0) {
          density = PerceptibleReciprocal(density);
          filteredPixel *= (float4)density;
          gamma *= density;
        }
        gamma = PerceptibleReciprocal(gamma);

        CLPixelType fp;
        fp = (CLPixelType) ( ClampToQuantum(gamma*filteredPixel.x)
          , ClampToQuantum(gamma*filteredPixel.y)
          , ClampToQuantum(gamma*filteredPixel.z)
          , ClampToQuantum(filteredPixel.w));

        filteredImage[(chunkStartY+itemID)*filteredColumns+x] = fp;

      }
    }

    } // end of chunking loop
  }
  )



  STRINGIFY(
 __kernel __attribute__((reqd_work_group_size(1, 256, 1)))
 void ResizeVerticalFilterSinc(const __global CLPixelType* inputImage, const unsigned int inputColumns, const unsigned int inputRows, const unsigned int matte
  , const float yFactor, __global CLPixelType* filteredImage, const unsigned int filteredColumns, const unsigned int filteredRows
  , const int resizeFilterType, const int resizeWindowType
  , const __global float* resizeFilterCubicCoefficients
  , const float resizeFilterScale, const float resizeFilterSupport, const float resizeFilterWindowSupport, const float resizeFilterBlur
  , __local CLPixelType* inputImageCache, const int numCachedPixels, const unsigned int pixelPerWorkgroup, const unsigned int pixelChunkSize
  , __local float4* outputPixelCache, __local float* densityCache, __local float* gammaCache) {
    ResizeVerticalFilter(inputImage,inputColumns,inputRows,matte
      ,yFactor,filteredImage,filteredColumns,filteredRows
      ,SincWeightingFunction, SincWeightingFunction
      ,resizeFilterCubicCoefficients
      ,resizeFilterScale,resizeFilterSupport,resizeFilterWindowSupport,resizeFilterBlur
      ,inputImageCache,numCachedPixels,pixelPerWorkgroup,pixelChunkSize
      ,outputPixelCache,densityCache,gammaCache);
  }
  )

  STRINGIFY(


  __kernel void randomNumberGeneratorKernel(__global uint* seeds, const float normalizeRand
                                           , __global float* randomNumbers, const uint init
                                           ,const uint numRandomNumbers) {

    unsigned int id = get_global_id(0);
    unsigned int seed[4];

    if (init!=0) {
      seed[0] = seeds[id*4];
      seed[1] = 0x50a7f451;
      seed[2] = 0x5365417e;
      seed[3] = 0xc3a4171a;
    }
    else {
      seed[0] = seeds[id*4];
      seed[1] = seeds[id*4+1];
      seed[2] = seeds[id*4+2];
      seed[3] = seeds[id*4+3];
    }

    unsigned int numRandomNumbersPerItem = (numRandomNumbers+get_global_size(0)-1)/get_global_size(0);
    for (unsigned int i = 0; i < numRandomNumbersPerItem; i++) {
      do
      {
        unsigned int alpha=(unsigned int) (seed[1] ^ (seed[1] << 11));
        seed[1]=seed[2];
        seed[2]=seed[3];
        seed[3]=seed[0];
        seed[0]=(seed[0] ^ (seed[0] >> 19)) ^ (alpha ^ (alpha >> 8));
      } while (seed[0] == ~0UL);
      unsigned int pos = (get_group_id(0)*get_local_size(0)*numRandomNumbersPerItem) 
                          + get_local_size(0) * i + get_local_id(0);

      if (pos >= numRandomNumbers)
        break;
      randomNumbers[pos] = normalizeRand*seed[0];
    }

    /* save the seeds for the time*/
    seeds[id*4]   = seed[0];
    seeds[id*4+1] = seed[1];
    seeds[id*4+2] = seed[2];
    seeds[id*4+3] = seed[3];
  }

  )


  STRINGIFY(
  
  typedef enum
  {
    UndefinedNoise,
    UniformNoise,
    GaussianNoise,
    MultiplicativeGaussianNoise,
    ImpulseNoise,
    LaplacianNoise,
    PoissonNoise,
    RandomNoise
  } NoiseType;

  typedef struct {
    const global float* rns;
  } RandomNumbers;


  float GetPseudoRandomValue(RandomNumbers* r) {
    float v = *r->rns;
    r->rns++;
    return v;
  }
  )

  OPENCL_DEFINE(SigmaUniform, (attenuate*0.015625f))
  OPENCL_DEFINE(SigmaGaussian,(attenuate*0.015625f))
  OPENCL_DEFINE(SigmaImpulse,  (attenuate*0.1f))
  OPENCL_DEFINE(SigmaLaplacian, (attenuate*0.0390625f))
  OPENCL_DEFINE(SigmaMultiplicativeGaussian,  (attenuate*0.5f))
  OPENCL_DEFINE(SigmaPoisson,  (attenuate*12.5f))
  OPENCL_DEFINE(SigmaRandom,  (attenuate))
  OPENCL_DEFINE(TauGaussian,  (attenuate*0.078125f))

  STRINGIFY(
  float GenerateDifferentialNoise(RandomNumbers* r, CLQuantum pixel, NoiseType noise_type, float attenuate) {
 
    float 
      alpha,
      beta,
      noise,
      sigma;

    noise = 0.0f;
    alpha=GetPseudoRandomValue(r);
    switch(noise_type) {
    case UniformNoise:
    default:
      {
        noise=(pixel+QuantumRange*SigmaUniform*(alpha-0.5f));
        break;
      }
    case GaussianNoise:
      {
        float
          gamma,
          tau;

        if (alpha == 0.0f)
          alpha=1.0f;
        beta=GetPseudoRandomValue(r);
        gamma=sqrt(-2.0f*log(alpha));
        sigma=gamma*cospi((2.0f*beta));
        tau=gamma*sinpi((2.0f*beta));
        noise=(float)(pixel+sqrt((float) pixel)*SigmaGaussian*sigma+
                      QuantumRange*TauGaussian*tau);        
        break;
      }


    case ImpulseNoise:
    {
      if (alpha < (SigmaImpulse/2.0f))
        noise=0.0f;
      else
        if (alpha >= (1.0f-(SigmaImpulse/2.0f)))
          noise=(float)QuantumRange;
        else
          noise=(float)pixel;
      break;
    }
    case LaplacianNoise:
    {
      if (alpha <= 0.5f)
        {
          if (alpha <= MagickEpsilon)
            noise=(float) (pixel-QuantumRange);
          else
            noise=(float) (pixel+QuantumRange*SigmaLaplacian*log(2.0f*alpha)+
              0.5f);
          break;
        }
      beta=1.0f-alpha;
      if (beta <= (0.5f*MagickEpsilon))
        noise=(float) (pixel+QuantumRange);
      else
        noise=(float) (pixel-QuantumRange*SigmaLaplacian*log(2.0f*beta)+0.5f);
      break;
    }
    case MultiplicativeGaussianNoise:
    {
      sigma=1.0f;
      if (alpha > MagickEpsilon)
        sigma=sqrt(-2.0f*log(alpha));
      beta=GetPseudoRandomValue(r);
      noise=(float) (pixel+pixel*SigmaMultiplicativeGaussian*sigma*
        cospi((float) (2.0f*beta))/2.0f);
      break;
    }
    case PoissonNoise:
    {
      float 
        poisson;
      unsigned int i;
      poisson=exp(-SigmaPoisson*QuantumScale*pixel);
      for (i=0; alpha > poisson; i++)
      {
        beta=GetPseudoRandomValue(r);
        alpha*=beta;
      }
      noise=(float) (QuantumRange*i/SigmaPoisson);
      break;
    }
    case RandomNoise:
    {
      noise=(float) (QuantumRange*SigmaRandom*alpha);
      break;
    }

    };
    return noise;
  }

  __kernel
  void AddNoiseImage(const __global CLPixelType* inputImage, __global CLPixelType* filteredImage
                    ,const unsigned int inputColumns, const unsigned int inputRows
                    ,const ChannelType channel 
                    ,const NoiseType noise_type, const float attenuate
                    ,const __global float* randomNumbers, const unsigned int numRandomNumbersPerPixel
                    ,const unsigned int rowOffset) {

    unsigned int x = get_global_id(0);
    unsigned int y = get_global_id(1) + rowOffset;
    RandomNumbers r;
    r.rns = randomNumbers + (get_global_id(1) * inputColumns + get_global_id(0))*numRandomNumbersPerPixel;

    CLPixelType p = inputImage[y*inputColumns+x];
    CLPixelType q = filteredImage[y*inputColumns+x];

    if ((channel&RedChannel)!=0) {
      setRed(&q,ClampToQuantum(GenerateDifferentialNoise(&r,getRed(p),noise_type,attenuate)));
    }
    
    if ((channel&GreenChannel)!=0) {
      setGreen(&q,ClampToQuantum(GenerateDifferentialNoise(&r,getGreen(p),noise_type,attenuate)));
    }

    if ((channel&BlueChannel)!=0) {
      setBlue(&q,ClampToQuantum(GenerateDifferentialNoise(&r,getBlue(p),noise_type,attenuate)));
    }

    if ((channel & OpacityChannel) != 0) {
      setOpacity(&q,ClampToQuantum(GenerateDifferentialNoise(&r,getOpacity(p),noise_type,attenuate)));
    }

    filteredImage[y*inputColumns+x] = q;
  }

  )
  ;




#endif // MAGICKCORE_OPENCL_SUPPORT

#if defined(__cplusplus) || defined(c_plusplus)
}
#endif

#endif // _MAGICKCORE_ACCELERATE_PRIVATE_H