__kernel
  void GenerateNoiseImage(const __global CLPixelType* inputImage, __global CLPixelType* filteredImage
                    ,const unsigned int inputPixelCount, const unsigned int pixelsPerWorkItem
                    ,const ChannelType channel 
                    ,const NoiseType noise_type, const float attenuate
                    ,const unsigned int seed0, const unsigned int seed1
                                        ,const unsigned int numRandomNumbersPerPixel) {

        mwc64x_state_t rng;
        rng.x = seed0;
        rng.c = seed1;

        uint span = pixelsPerWorkItem * numRandomNumbersPerPixel;       // length of RNG substream each workitem will use
        uint offset = span * get_local_size(0) * get_group_id(0);       // offset of this workgroup's RNG substream (in master stream);

        MWC64X_SeedStreams(&rng, offset, span);                                         // Seed the RNG streams

        uint pos = get_local_size(0) * get_group_id(0) * pixelsPerWorkItem + get_local_id(0);   // pixel to process

        uint count = pixelsPerWorkItem;

        while (count > 0) {
                if (pos < inputPixelCount) {
                        CLPixelType p = inputImage[pos];

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

                        if ((channel&BlueChannel)!=0) {
                          setBlue(&p,ClampToQuantum(mwcGenerateDifferentialNoise(&rng,getBlue(p),noise_type,attenuate)));
                        }

                        if ((channel & OpacityChannel) != 0) {
                          setOpacity(&p,ClampToQuantum(mwcGenerateDifferentialNoise(&rng,getOpacity(p),noise_type,attenuate)));
                        }

                        filteredImage[pos] = p;
                        //filteredImage[pos] = (CLPixelType)(MWC64X_NextUint(&rng) % 256, MWC64X_NextUint(&rng) % 256, MWC64X_NextUint(&rng) % 256, 255);
                }
                pos += get_local_size(0);
                --count;
        }
  }