view release on metacpan or  search on metacpan

src/Source/LibJPEG/jchuff.c  view on Meta::CPAN

/*
 * Trial-encode one MCU's worth of Huffman-compressed coefficients.
 * No data is actually output, so no suspension return is possible.
 */

METHODDEF(boolean)
encode_mcu_gather (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
{
  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
  int blkn, ci;
  jpeg_component_info * compptr;

  /* Take care of restart intervals if needed */
  if (cinfo->restart_interval) {
    if (entropy->restarts_to_go == 0) {
      /* Re-initialize DC predictions to 0 */
      for (ci = 0; ci < cinfo->comps_in_scan; ci++)
	entropy->saved.last_dc_val[ci] = 0;
      /* Update restart state */
      entropy->restarts_to_go = cinfo->restart_interval;
    }
    entropy->restarts_to_go--;
  }

  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
    ci = cinfo->MCU_membership[blkn];
    compptr = cinfo->cur_comp_info[ci];
    htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci],
		    entropy->dc_count_ptrs[compptr->dc_tbl_no],
		    entropy->ac_count_ptrs[compptr->ac_tbl_no]);
    entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0];
  }

  return TRUE;
}


/*
 * Generate the best Huffman code table for the given counts, fill htbl.
 *
 * The JPEG standard requires that no symbol be assigned a codeword of all
 * one bits (so that padding bits added at the end of a compressed segment
 * can't look like a valid code).  Because of the canonical ordering of
 * codewords, this just means that there must be an unused slot in the
 * longest codeword length category.  Section K.2 of the JPEG spec suggests
 * reserving such a slot by pretending that symbol 256 is a valid symbol
 * with count 1.  In theory that's not optimal; giving it count zero but
 * including it in the symbol set anyway should give a better Huffman code.
 * But the theoretically better code actually seems to come out worse in
 * practice, because it produces more all-ones bytes (which incur stuffed
 * zero bytes in the final file).  In any case the difference is tiny.
 *
 * The JPEG standard requires Huffman codes to be no more than 16 bits long.
 * If some symbols have a very small but nonzero probability, the Huffman tree
 * must be adjusted to meet the code length restriction.  We currently use
 * the adjustment method suggested in JPEG section K.2.  This method is *not*
 * optimal; it may not choose the best possible limited-length code.  But
 * typically only very-low-frequency symbols will be given less-than-optimal
 * lengths, so the code is almost optimal.  Experimental comparisons against
 * an optimal limited-length-code algorithm indicate that the difference is
 * microscopic --- usually less than a hundredth of a percent of total size.
 * So the extra complexity of an optimal algorithm doesn't seem worthwhile.
 */

LOCAL(void)
jpeg_gen_optimal_table (j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[])
{
#define MAX_CLEN 32		/* assumed maximum initial code length */
  UINT8 bits[MAX_CLEN+1];	/* bits[k] = # of symbols with code length k */
  int codesize[257];		/* codesize[k] = code length of symbol k */
  int others[257];		/* next symbol in current branch of tree */
  int c1, c2;
  int p, i, j;
  long v;

  /* This algorithm is explained in section K.2 of the JPEG standard */

  MEMZERO(bits, SIZEOF(bits));
  MEMZERO(codesize, SIZEOF(codesize));
  for (i = 0; i < 257; i++)
    others[i] = -1;		/* init links to empty */
  
  freq[256] = 1;		/* make sure 256 has a nonzero count */
  /* Including the pseudo-symbol 256 in the Huffman procedure guarantees
   * that no real symbol is given code-value of all ones, because 256
   * will be placed last in the largest codeword category.
   */

  /* Huffman's basic algorithm to assign optimal code lengths to symbols */

  for (;;) {
    /* Find the smallest nonzero frequency, set c1 = its symbol */
    /* In case of ties, take the larger symbol number */
    c1 = -1;
    v = 1000000000L;
    for (i = 0; i <= 256; i++) {
      if (freq[i] && freq[i] <= v) {
	v = freq[i];
	c1 = i;
      }
    }

    /* Find the next smallest nonzero frequency, set c2 = its symbol */
    /* In case of ties, take the larger symbol number */
    c2 = -1;
    v = 1000000000L;
    for (i = 0; i <= 256; i++) {
      if (freq[i] && freq[i] <= v && i != c1) {
	v = freq[i];
	c2 = i;
      }
    }

    /* Done if we've merged everything into one frequency */
    if (c2 < 0)
      break;
    
    /* Else merge the two counts/trees */
    freq[c1] += freq[c2];
    freq[c2] = 0;