{
  int inrow, outrow;
  JDIMENSION outcol;
  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
  register JSAMPROW inptr0, inptr1, outptr;
  register int bias;

  /* Expand input data enough to let all the output samples be generated
   * by the standard loop.  Special-casing padded output would be more
   * efficient.
   */
  expand_right_edge(input_data, cinfo->max_v_samp_factor,
		    cinfo->image_width, output_cols * 2);

  inrow = 0;
  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
    outptr = output_data[outrow];
    inptr0 = input_data[inrow];
    inptr1 = input_data[inrow+1];
    bias = 1;			/* bias = 1,2,1,2,... for successive samples */
    for (outcol = 0; outcol < output_cols; outcol++) {
      *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
			      GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1])
			      + bias) >> 2);
      bias ^= 3;		/* 1=>2, 2=>1 */
      inptr0 += 2; inptr1 += 2;
    }
    inrow += 2;
  }
}


#ifdef INPUT_SMOOTHING_SUPPORTED

/*
 * Downsample pixel values of a single component.
 * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
 * with smoothing.  One row of context is required.
 */

METHODDEF(void)
h2v2_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
			JSAMPARRAY input_data, JSAMPARRAY output_data)
{
  int inrow, outrow;
  JDIMENSION colctr;
  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
  register JSAMPROW inptr0, inptr1, above_ptr, below_ptr, outptr;
  INT32 membersum, neighsum, memberscale, neighscale;

  /* Expand input data enough to let all the output samples be generated
   * by the standard loop.  Special-casing padded output would be more
   * efficient.
   */
  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
		    cinfo->image_width, output_cols * 2);

  /* We don't bother to form the individual "smoothed" input pixel values;
   * we can directly compute the output which is the average of the four
   * smoothed values.  Each of the four member pixels contributes a fraction
   * (1-8*SF) to its own smoothed image and a fraction SF to each of the three
   * other smoothed pixels, therefore a total fraction (1-5*SF)/4 to the final
   * output.  The four corner-adjacent neighbor pixels contribute a fraction
   * SF to just one smoothed pixel, or SF/4 to the final output; while the
   * eight edge-adjacent neighbors contribute SF to each of two smoothed
   * pixels, or SF/2 overall.  In order to use integer arithmetic, these
   * factors are scaled by 2^16 = 65536.
   * Also recall that SF = smoothing_factor / 1024.
   */

  memberscale = 16384 - cinfo->smoothing_factor * 80; /* scaled (1-5*SF)/4 */
  neighscale = cinfo->smoothing_factor * 16; /* scaled SF/4 */

  inrow = 0;
  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
    outptr = output_data[outrow];
    inptr0 = input_data[inrow];
    inptr1 = input_data[inrow+1];
    above_ptr = input_data[inrow-1];
    below_ptr = input_data[inrow+2];

    /* Special case for first column: pretend column -1 is same as column 0 */
    membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
    neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
	       GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[2]) +
	       GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[2]);
    neighsum += neighsum;
    neighsum += GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[2]) +
		GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[2]);
    membersum = membersum * memberscale + neighsum * neighscale;
    *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
    inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;

    for (colctr = output_cols - 2; colctr > 0; colctr--) {
      /* sum of pixels directly mapped to this output element */
      membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
		  GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
      /* sum of edge-neighbor pixels */
      neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
		 GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
		 GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[2]) +
		 GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[2]);
      /* The edge-neighbors count twice as much as corner-neighbors */
      neighsum += neighsum;
      /* Add in the corner-neighbors */
      neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[2]) +
		  GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[2]);
      /* form final output scaled up by 2^16 */
      membersum = membersum * memberscale + neighsum * neighscale;
      /* round, descale and output it */
      *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
      inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
    }

    /* Special case for last column */
    membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
    neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
	       GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[1]) +
	       GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[1]);
    neighsum += neighsum;
    neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[1]) +
		GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[1]);
    membersum = membersum * memberscale + neighsum * neighscale;
    *outptr = (JSAMPLE) ((membersum + 32768) >> 16);

    inrow += 2;
  }
}


/*
 * Downsample pixel values of a single component.
 * This version handles the special case of a full-size component,
 * with smoothing.  One row of context is required.
 */

METHODDEF(void)
fullsize_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr,
			    JSAMPARRAY input_data, JSAMPARRAY output_data)
{
  int outrow;
  JDIMENSION colctr;
  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
  register JSAMPROW inptr, above_ptr, below_ptr, outptr;
  INT32 membersum, neighsum, memberscale, neighscale;
  int colsum, lastcolsum, nextcolsum;

  /* Expand input data enough to let all the output samples be generated
   * by the standard loop.  Special-casing padded output would be more
   * efficient.
   */
  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
		    cinfo->image_width, output_cols);

  /* Each of the eight neighbor pixels contributes a fraction SF to the
   * smoothed pixel, while the main pixel contributes (1-8*SF).  In order
   * to use integer arithmetic, these factors are multiplied by 2^16 = 65536.
   * Also recall that SF = smoothing_factor / 1024.
   */

  memberscale = 65536L - cinfo->smoothing_factor * 512L; /* scaled 1-8*SF */
  neighscale = cinfo->smoothing_factor * 64; /* scaled SF */

  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
    outptr = output_data[outrow];
    inptr = input_data[outrow];
    above_ptr = input_data[outrow-1];
    below_ptr = input_data[outrow+1];

    /* Special case for first column */
    colsum = GETJSAMPLE(*above_ptr++) + GETJSAMPLE(*below_ptr++) +
	     GETJSAMPLE(*inptr);
    membersum = GETJSAMPLE(*inptr++);
    nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
		 GETJSAMPLE(*inptr);
    neighsum = colsum + (colsum - membersum) + nextcolsum;
    membersum = membersum * memberscale + neighsum * neighscale;
    *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
    lastcolsum = colsum; colsum = nextcolsum;

    for (colctr = output_cols - 2; colctr > 0; colctr--) {
      membersum = GETJSAMPLE(*inptr++);
      above_ptr++; below_ptr++;
      nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
		   GETJSAMPLE(*inptr);
      neighsum = lastcolsum + (colsum - membersum) + nextcolsum;
      membersum = membersum * memberscale + neighsum * neighscale;
      *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
      lastcolsum = colsum; colsum = nextcolsum;
    }

    /* Special case for last column */
    membersum = GETJSAMPLE(*inptr);
    neighsum = lastcolsum + (colsum - membersum) + colsum;
    membersum = membersum * memberscale + neighsum * neighscale;
    *outptr = (JSAMPLE) ((membersum + 32768) >> 16);

  }
}

#endif /* INPUT_SMOOTHING_SUPPORTED */


/*
 * Module initialization routine for downsampling.
 * Note that we must select a routine for each component.
 */

GLOBAL(void)
jinit_downsampler (j_compress_ptr cinfo)
{
  my_downsample_ptr downsample;
  int ci;
  jpeg_component_info * compptr;
  boolean smoothok = TRUE;

  downsample = (my_downsample_ptr)
    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
				SIZEOF(my_downsampler));
  cinfo->downsample = (struct jpeg_downsampler *) downsample;
  downsample->pub.start_pass = start_pass_downsample;
  downsample->pub.downsample = sep_downsample;
  downsample->pub.need_context_rows = FALSE;

  if (cinfo->CCIR601_sampling)
    ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);

  /* Verify we can handle the sampling factors, and set up method pointers */
  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
       ci++, compptr++) {
    if (compptr->h_samp_factor == cinfo->max_h_samp_factor &&
	compptr->v_samp_factor == cinfo->max_v_samp_factor) {
#ifdef INPUT_SMOOTHING_SUPPORTED
      if (cinfo->smoothing_factor) {
	downsample->methods[ci] = fullsize_smooth_downsample;
	downsample->pub.need_context_rows = TRUE;
      } else
#endif
	downsample->methods[ci] = fullsize_downsample;
    } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
	       compptr->v_samp_factor == cinfo->max_v_samp_factor) {
      smoothok = FALSE;
      downsample->methods[ci] = h2v1_downsample;
    } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
	       compptr->v_samp_factor * 2 == cinfo->max_v_samp_factor) {
#ifdef INPUT_SMOOTHING_SUPPORTED
      if (cinfo->smoothing_factor) {
	downsample->methods[ci] = h2v2_smooth_downsample;
	downsample->pub.need_context_rows = TRUE;
      } else
#endif
	downsample->methods[ci] = h2v2_downsample;
    } else if ((cinfo->max_h_samp_factor % compptr->h_samp_factor) == 0 &&
	       (cinfo->max_v_samp_factor % compptr->v_samp_factor) == 0) {
      smoothok = FALSE;
      downsample->methods[ci] = int_downsample;
    } else
      ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
  }

#ifdef INPUT_SMOOTHING_SUPPORTED
  if (cinfo->smoothing_factor && !smoothok)
    TRACEMS(cinfo, 0, JTRC_SMOOTH_NOTIMPL);
#endif
}