import jdk.incubator.vector.*;

final sclass BWIntegralImage_vectorized implements MakesBufferedImage, IBWIntegralImage {
  int w, h;
  int vecShift; // vector size used as dual log
  int[] data; // 1 entry per pixel
  ifdef BWIntegralImage_CountAccesses
    long accesses;
  endifdef
  
  *() {}
  *(File f) { this(loadImage2(f)); }
  
  *(MakesBufferedImage img) { this(toBufferedImage(img)); }
  
  //*(BufferedImage img) { this(BWImage(img)); }
  
  *(BufferedImage image) ctex {
    alloc(image.getWidth(), image.getHeight());
    // Just grab directly into data array
    PixelGrabber pixelGrabber = new PixelGrabber(image, 0, 0, w, h, data, 0, w);
    if (!pixelGrabber.grabPixels())
      fail("Could not grab pixels");
      
    VectorSpecies<Int> species = IntVector.SPECIES_PREFERRED;
    vecShift = dualLog(species.length());
    int upperBound = species.loopBound(w);

    // for brightness of pixels,
    // for now we cheat by just using one of the channels

    // first row
    int i = 0, sum = 0;
    for (; i < upperBound; i += species.length()) {
      IntVector v = IntVector.fromArray(species, data, i);
      v = v.and(255).add(sum);
      v.intoArray(data, i);
      sum += v.reduceLanes(VectorOperators.ADD);
    }

    for (; i < w; i++)
      data[i] = (data[i] & 255) + sum;
      
    // subsequent rows
    int i = w;
    for (int y = 1; y < h; y++) {
      sum = 0;
      int x = 0;
      for (; x < upperBound; x += species.length(), i += species.length()) {
        IntVector v = IntVector.fromArray(species, data, i);
        IntVector v_last = IntVector.fromArray(species, data, i-w);
        v = v.and(255).add(v_last).add(sum);
        v.intoArray(data, i);
        sum += v.reduceLanes(VectorOperators.ADD);
      }

      for (; x < w; x++, i++)
        data[i] = (data[i] & 255) + sum;
    }
  }

  private void alloc(int w, int h) {
    this.w = w;
    this.h = h;
    if (w*h > 8*1024*1024)
      fail("Image too large (more than 8 MP): " + w + "*" + h);
    data = new int[w*h];
  }
  
  // pixels outside of image are considered black
  int get(int x, int y) {
    ifdef BWIntegralImage_CountAccesses
      ++accesses;
    endifdef
    ret x < 0 || y < 0 || x >= w || y >= h ? 0
      : data[min(y, h-1)*w+min(x, w-1)];
  }
  
  public double getPixelAverage(int x1, int y1, int x2, int y2) {
    int area = (x2-x1)*(y2-y1);
    ret doubleRatio(bwIntegralImage_sumRect(this, x1, y1, x2, y2), area);
  }
  
  int getPixel(int x, int y) {
    ret bwIntegralImage_sumRect(this, x, y, x+1, y+1);
  }
  
  int getPixel(Pt p) { ret getPixel(p.x, p.y); }
  
  public int getWidth() { ret w; }
  public int getHeight() { ret h; }
  
  // unoptimized
  public BufferedImage getBufferedImage() {
    ret scaleDownUsingIntegralImageBW(this, w).getBufferedImage();
  }
}

Began life as a copy of #1019595