jchuff.c

一个国人自己实现图像库的程序（有参考价值）

    return FALSE;

  emit_byte(state, 0xFF, return FALSE);
  emit_byte(state, JPEG_RST0 + restart_num, return FALSE);

  /* Re-initialize DC predictions to 0 */
  for (ci = 0; ci < state->cinfo->comps_in_scan; ci++)
    state->cur.last_dc_val[ci] = 0;

  /* The restart counter is not updated until we successfully write the MCU. */

  return TRUE;
}


/*
 * Encode and output one MCU's worth of Huffman-compressed coefficients.
 */

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

  /* Load up working state */
  state.next_output_byte = cinfo->dest->next_output_byte;
  state.free_in_buffer = cinfo->dest->free_in_buffer;
  ASSIGN_STATE(state.cur, entropy->saved);
  state.cinfo = cinfo;

  /* Emit restart marker if needed */
  if (cinfo->restart_interval) {
    if (entropy->restarts_to_go == 0)
      if (! emit_restart(&state, entropy->next_restart_num))
	return FALSE;
  }

  /* Encode the MCU data blocks */
  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
    ci = cinfo->MCU_membership[blkn];
    compptr = cinfo->cur_comp_info[ci];
    if (! encode_one_block(&state,
			   MCU_data[blkn][0], state.cur.last_dc_val[ci],
			   entropy->dc_derived_tbls[compptr->dc_tbl_no],
			   entropy->ac_derived_tbls[compptr->ac_tbl_no]))
      return FALSE;
    /* Update last_dc_val */
    state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
  }

  /* Completed MCU, so update state */
  cinfo->dest->next_output_byte = state.next_output_byte;
  cinfo->dest->free_in_buffer = state.free_in_buffer;
  ASSIGN_STATE(entropy->saved, state.cur);

  /* Update restart-interval state too */
  if (cinfo->restart_interval) {
    if (entropy->restarts_to_go == 0) {
      entropy->restarts_to_go = cinfo->restart_interval;
      entropy->next_restart_num++;
      entropy->next_restart_num &= 7;
    }
    entropy->restarts_to_go--;
  }

  return TRUE;
}


/*
 * Finish up at the end of a Huffman-compressed scan.
 */

METHODDEF(void)
finish_pass_huff (j_compress_ptr cinfo)
{
  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
  working_state state;

  /* Load up working state ... flush_bits needs it */
  state.next_output_byte = cinfo->dest->next_output_byte;
  state.free_in_buffer = cinfo->dest->free_in_buffer;
  ASSIGN_STATE(state.cur, entropy->saved);
  state.cinfo = cinfo;

  /* Flush out the last data */
  if (! flush_bits(&state))
    ERREXIT(cinfo, JERR_CANT_SUSPEND);

  /* Update state */
  cinfo->dest->next_output_byte = state.next_output_byte;
  cinfo->dest->free_in_buffer = state.free_in_buffer;
  ASSIGN_STATE(entropy->saved, state.cur);
}


/*
 * Huffman coding optimization.
 *
 * We first scan the supplied data and count the number of uses of each symbol
 * that is to be Huffman-coded. (This process MUST agree with the code above.)
 * Then we build a Huffman coding tree for the observed counts.
 * Symbols which are not needed at all for the particular image are not
 * assigned any code, which saves space in the DHT marker as well as in
 * the compressed data.
 */

#ifdef ENTROPY_OPT_SUPPORTED


/* Process a single block's worth of coefficients */

LOCAL(void)
htest_one_block (j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val,
		 long dc_counts[], long ac_counts[])
{
  register int temp;
  register int nbits;
  register int k, r;
  
  /* Encode the DC coefficient difference per section F.1.2.1 */
  
  temp = block[0] - last_dc_val;
  if (temp < 0)
    temp = -temp;
  
  /* Find the number of bits needed for the magnitude of the coefficient */
  nbits = 0;
  while (temp) {
    nbits++;
    temp >>= 1;
  }
  /* Check for out-of-range coefficient values.
   * Since we're encoding a difference, the range limit is twice as much.
   */
  if (nbits > MAX_COEF_BITS+1)
    ERREXIT(cinfo, JERR_BAD_DCT_COEF);

  /* Count the Huffman symbol for the number of bits */
  dc_counts[nbits]++;
  
  /* Encode the AC coefficients per section F.1.2.2 */
  
  r = 0;			/* r = run length of zeros */
  
  for (k = 1; k < DCTSIZE2; k++) {
    if ((temp = block[jpeg_natural_order[k]]) == 0) {
      r++;
    } else {
      /* if run length > 15, must emit special run-length-16 codes (0xF0) */
      while (r > 15) {
	ac_counts[0xF0]++;
	r -= 16;
      }
      
      /* Find the number of bits needed for the magnitude of the coefficient */
      if (temp < 0)
	temp = -temp;
      
      /* Find the number of bits needed for the magnitude of the coefficient */
      nbits = 1;		/* there must be at least one 1 bit */
      while ((temp >>= 1))
	nbits++;
      /* Check for out-of-range coefficient values */
      if (nbits > MAX_COEF_BITS)
	ERREXIT(cinfo, JERR_BAD_DCT_COEF);
      
      /* Count Huffman symbol for run length / number of bits */
      ac_counts[(r << 4) + nbits]++;
      
      r = 0;
    }
  }

  /* If the last coef(s) were zero, emit an end-of-block code */
  if (r > 0)
    ac_counts[0]++;
}


/*
 * 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.
 * Note this is also used by jcphuff.c.
 *
 * 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.
 */

GLOBAL(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;

    /* Increment the codesize of everything in c1's tree branch */
    codesize[c1]++;
    while (others[c1] >= 0) {
      c1 = others[c1];
      codesize[c1]++;
    }
    
    others[c1] = c2;		/* chain c2 onto c1's tree branch */
    
    /* Increment the codesize of everything in c2's tree branch */
    codesize[c2]++;
    while (others[c2] >= 0) {
      c2 = others[c2];
      codesize[c2]++;
    }
  }

  /* Now count the number of symbols of each code length */
  for (i = 0; i <= 256; i++) {
    if (codesize[i]) {
      /* The JPEG standard seems to think that this can't happen, */
      /* but I'm paranoid... */
      if (codesize[i] > MAX_CLEN)
	ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW);

      bits[codesize[i]]++;
    }
  }

  /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure
   * Huffman procedure assigned any such lengths, we must adjust the coding.
   * Here is what the JPEG spec says about how this next bit works:
   * Since symbols are paired for the longest Huffman code, the symbols are
   * removed from this length category two at a time.  The prefix for the pair
   * (which is one bit shorter) is allocated to one of the pair; then,
   * skipping the BITS entry for that prefix length, a code word from the next
   * shortest nonzero BITS entry is converted into a prefix for two code words
   * one bit longer.
   */
  
  for (i = MAX_CLEN; i > 16; i--) {
    while (bits[i] > 0) {
      j = i - 2;		/* find length of new prefix to be used */
      while (bits[j] == 0)
	j--;
      
      bits[i] -= 2;		/* remove two symbols */
      bits[i-1]++;		/* one goes in this length */
      bits[j+1] += 2;		/* two new symbols in this length */
      bits[j]--;		/* symbol of this length is now a prefix */
    }
  }

  /* Remove the count for the pseudo-symbol 256 from the largest codelength */
  while (bits[i] == 0)		/* find largest codelength still in use */
    i--;
  bits[i]--;
  
  /* Return final symbol counts (only for lengths 0..16) */
  MEMCOPY(htbl->bits, bits, SIZEOF(htbl->bits));
  
  /* Return a list of the symbols sorted by code length */
  /* It's not real clear to me why we don't need to consider the codelength
   * changes made above, but the JPEG spec seems to think this works.
   */
  p = 0;
  for (i = 1; i <= MAX_CLEN; i++) {
    for (j = 0; j <= 255; j++) {
      if (codesize[j] == i) {
	htbl->huffval[p] = (UINT8) j;
	p++;
      }
    }
  }

  /* Set sent_table FALSE so updated table will be written to JPEG file. */
  htbl->sent_table = FALSE;
}


/*
 * Finish up a statistics-gathering pass and create the new Huffman tables.
 */

METHODDEF(void)
finish_pass_gather (j_compress_ptr cinfo)
{
  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
  int ci, dctbl, actbl;
  jpeg_component_info * compptr;
  JHUFF_TBL **htblptr;
  boolean did_dc[NUM_HUFF_TBLS];
  boolean did_ac[NUM_HUFF_TBLS];

  /* It's important not to apply jpeg_gen_optimal_table more than once
   * per table, because it clobbers the input frequency counts!
   */
  MEMZERO(did_dc, SIZEOF(did_dc));
  MEMZERO(did_ac, SIZEOF(did_ac));

  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
    compptr = cinfo->cur_comp_info[ci];
    dctbl = compptr->dc_tbl_no;
    actbl = compptr->ac_tbl_no;
    if (! did_dc[dctbl]) {
      htblptr = & cinfo->dc_huff_tbl_ptrs[dctbl];
      if (*htblptr == NULL)
	*htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]);
      did_dc[dctbl] = TRUE;
    }
    if (! did_ac[actbl]) {
      htblptr = & cinfo->ac_huff_tbl_ptrs[actbl];
      if (*htblptr == NULL)
	*htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]);
      did_ac[actbl] = TRUE;
    }
  }
}


#endif /* ENTROPY_OPT_SUPPORTED */


/*
 * Module initialization routine for Huffman entropy encoding.
 */

GLOBAL(void)
jinit_huff_encoder (j_compress_ptr cinfo)
{
  huff_entropy_ptr entropy;
  int i;

  entropy = (huff_entropy_ptr)
    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
				SIZEOF(huff_entropy_encoder));
  cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
  entropy->pub.start_pass = start_pass_huff;

  /* Mark tables unallocated */
  for (i = 0; i < NUM_HUFF_TBLS; i++) {
    entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
#ifdef ENTROPY_OPT_SUPPORTED
    entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL;
#endif
  }
}