4 * Copyright (C) 1991-1997, Thomas G. Lane.
5 * This file is part of the Independent JPEG Group's software.
6 * For conditions of distribution and use, see the accompanying README file.
8 * This file contains Huffman entropy encoding routines.
10 * Much of the complexity here has to do with supporting output suspension.
11 * If the data destination module demands suspension, we want to be able to
12 * back up to the start of the current MCU. To do this, we copy state
13 * variables into local working storage, and update them back to the
14 * permanent JPEG objects only upon successful completion of an MCU.
17 #define JPEG_INTERNALS
18 #include "jinclude12.h"
19 #include "jpeglib12.h"
20 #include "jchuff12.h" /* Declarations shared with jcphuff.c */
23 /* Expanded entropy encoder object for Huffman encoding.
25 * The savable_state subrecord contains fields that change within an MCU,
26 * but must not be updated permanently until we complete the MCU.
30 INT32 put_buffer; /* current bit-accumulation buffer */
31 int put_bits; /* # of bits now in it */
32 int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
35 /* This macro is to work around compilers with missing or broken
36 * structure assignment. You'll need to fix this code if you have
37 * such a compiler and you change MAX_COMPS_IN_SCAN.
40 #ifndef NO_STRUCT_ASSIGN
41 #define ASSIGN_STATE(dest,src) ((dest) = (src))
43 #if MAX_COMPS_IN_SCAN == 4
44 #define ASSIGN_STATE(dest,src) \
45 ((dest).put_buffer = (src).put_buffer, \
46 (dest).put_bits = (src).put_bits, \
47 (dest).last_dc_val[0] = (src).last_dc_val[0], \
48 (dest).last_dc_val[1] = (src).last_dc_val[1], \
49 (dest).last_dc_val[2] = (src).last_dc_val[2], \
50 (dest).last_dc_val[3] = (src).last_dc_val[3])
56 struct jpeg_entropy_encoder pub; /* public fields */
58 savable_state saved; /* Bit buffer & DC state at start of MCU */
60 /* These fields are NOT loaded into local working state. */
61 unsigned int restarts_to_go; /* MCUs left in this restart interval */
62 int next_restart_num; /* next restart number to write (0-7) */
64 /* Pointers to derived tables (these workspaces have image lifespan) */
65 c_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
66 c_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
68 #ifdef ENTROPY_OPT_SUPPORTED /* Statistics tables for optimization */
69 long * dc_count_ptrs[NUM_HUFF_TBLS];
70 long * ac_count_ptrs[NUM_HUFF_TBLS];
72 } huff_entropy_encoder;
74 typedef huff_entropy_encoder * huff_entropy_ptr;
76 /* Working state while writing an MCU.
77 * This struct contains all the fields that are needed by subroutines.
81 JOCTET * next_output_byte; /* => next byte to write in buffer */
82 size_t free_in_buffer; /* # of byte spaces remaining in buffer */
83 savable_state cur; /* Current bit buffer & DC state */
84 j_compress_ptr cinfo; /* dump_buffer needs access to this */
88 /* Forward declarations */
89 METHODDEF(boolean) encode_mcu_huff JPP((j_compress_ptr cinfo,
90 JBLOCKROW *MCU_data));
91 METHODDEF(void) finish_pass_huff JPP((j_compress_ptr cinfo));
92 #ifdef ENTROPY_OPT_SUPPORTED
93 METHODDEF(boolean) encode_mcu_gather JPP((j_compress_ptr cinfo,
94 JBLOCKROW *MCU_data));
95 METHODDEF(void) finish_pass_gather JPP((j_compress_ptr cinfo));
100 * Initialize for a Huffman-compressed scan.
101 * If gather_statistics is TRUE, we do not output anything during the scan,
102 * just count the Huffman symbols used and generate Huffman code tables.
106 start_pass_huff (j_compress_ptr cinfo, boolean gather_statistics)
108 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
109 int ci, dctbl, actbl;
110 jpeg_component_info * compptr;
112 fprintf (stderr,"=======================================================JPEG12\n");
114 if (gather_statistics) {
115 #ifdef ENTROPY_OPT_SUPPORTED
116 entropy->pub.encode_mcu = encode_mcu_gather;
117 entropy->pub.finish_pass = finish_pass_gather;
119 ERREXIT(cinfo, JERR_NOT_COMPILED);
122 entropy->pub.encode_mcu = encode_mcu_huff;
123 entropy->pub.finish_pass = finish_pass_huff;
126 for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
127 compptr = cinfo->cur_comp_info[ci];
128 dctbl = compptr->dc_tbl_no;
129 actbl = compptr->ac_tbl_no;
130 if (gather_statistics) {
131 #ifdef ENTROPY_OPT_SUPPORTED
132 /* Check for invalid table indexes */
133 /* (make_c_derived_tbl does this in the other path) */
134 if (dctbl < 0 || dctbl >= NUM_HUFF_TBLS)
135 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, dctbl);
136 if (actbl < 0 || actbl >= NUM_HUFF_TBLS)
137 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, actbl);
138 /* Allocate and zero the statistics tables */
139 /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
140 if (entropy->dc_count_ptrs[dctbl] == NULL)
141 entropy->dc_count_ptrs[dctbl] = (long *)
142 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
144 MEMZERO(entropy->dc_count_ptrs[dctbl], 257 * SIZEOF(long));
145 if (entropy->ac_count_ptrs[actbl] == NULL)
146 entropy->ac_count_ptrs[actbl] = (long *)
147 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
149 MEMZERO(entropy->ac_count_ptrs[actbl], 257 * SIZEOF(long));
152 /* Compute derived values for Huffman tables */
153 /* We may do this more than once for a table, but it's not expensive */
154 jpeg_make_c_derived_tbl(cinfo, TRUE, dctbl,
155 & entropy->dc_derived_tbls[dctbl]);
156 jpeg_make_c_derived_tbl(cinfo, FALSE, actbl,
157 & entropy->ac_derived_tbls[actbl]);
159 /* Initialize DC predictions to 0 */
160 entropy->saved.last_dc_val[ci] = 0;
163 /* Initialize bit buffer to empty */
164 entropy->saved.put_buffer = 0;
165 entropy->saved.put_bits = 0;
167 /* Initialize restart stuff */
168 entropy->restarts_to_go = cinfo->restart_interval;
169 entropy->next_restart_num = 0;
174 * Compute the derived values for a Huffman table.
175 * This routine also performs some validation checks on the table.
177 * Note this is also used by jcphuff.c.
181 jpeg_make_c_derived_tbl (j_compress_ptr cinfo, boolean isDC, int tblno,
182 c_derived_tbl ** pdtbl)
186 int p, i, l, lastp, si, maxsymbol;
188 unsigned int huffcode[257];
191 /* Note that huffsize[] and huffcode[] are filled in code-length order,
192 * paralleling the order of the symbols themselves in htbl->huffval[].
195 /* Find the input Huffman table */
196 if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
197 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
199 isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
201 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
203 /* Allocate a workspace if we haven't already done so. */
205 *pdtbl = (c_derived_tbl *)
206 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
207 SIZEOF(c_derived_tbl));
210 /* Figure C.1: make table of Huffman code length for each symbol */
213 for (l = 1; l <= 16; l++) {
214 i = (int) htbl->bits[l];
215 if (i < 0 || p + i > 256) { /* protect against table overrun */
216 printf ("JERR_BAD_HUFF_TABLE : protect against table overrun (i=%d p=%d)\n",i,p);
217 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
220 huffsize[p++] = (char) l;
225 /* Figure C.2: generate the codes themselves */
226 /* We also validate that the counts represent a legal Huffman code tree. */
231 while (huffsize[p]) {
232 while (((int) huffsize[p]) == si) {
233 huffcode[p++] = code;
236 /* code is now 1 more than the last code used for codelength si; but
237 * it must still fit in si bits, since no code is allowed to be all ones.
239 if (((INT32) code) >= (((INT32) 1) << si)) {
240 printf("JERR_BAD_HUFF_TABLE : (((INT32) code) >= (((INT32) 1) << si)) code %d si%d\v",code, si);
241 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
247 /* Figure C.3: generate encoding tables */
248 /* These are code and size indexed by symbol value */
250 /* Set all codeless symbols to have code length 0;
251 * this lets us detect duplicate VAL entries here, and later
252 * allows emit_bits to detect any attempt to emit such symbols.
254 MEMZERO(dtbl->ehufsi, SIZEOF(dtbl->ehufsi));
256 /* This is also a convenient place to check for out-of-range
257 * and duplicated VAL entries. We allow 0..255 for AC symbols
258 * but only 0..15 for DC. (We could constrain them further
259 * based on data depth and mode, but this seems enough.)
261 maxsymbol = isDC ? 15 : 255;
263 for (p = 0; p < lastp; p++) {
264 i = htbl->huffval[p];
265 if (i < 0 || i > maxsymbol || dtbl->ehufsi[i]) {
266 printf("JERR_BAD_HUFF_TABLE (i < 0 || i > maxsymbol || dtbl->ehufsi[i]) i %d maxsymbol %d dtbl->ehufsi[i] %d\n", i, maxsymbol, dtbl->ehufsi[i]);
267 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
269 dtbl->ehufco[i] = huffcode[p];
270 dtbl->ehufsi[i] = huffsize[p];
275 /* Outputting bytes to the file */
277 /* Emit a byte, taking 'action' if must suspend. */
278 #define emit_byte(state,val,action) \
279 { *(state)->next_output_byte++ = (JOCTET) (val); \
280 if (--(state)->free_in_buffer == 0) \
281 if (! dump_buffer(state)) \
286 dump_buffer (working_state * state)
287 /* Empty the output buffer; return TRUE if successful, FALSE if must suspend */
289 struct jpeg_destination_mgr * dest = state->cinfo->dest;
291 if (! (*dest->empty_output_buffer) (state->cinfo))
293 /* After a successful buffer dump, must reset buffer pointers */
294 state->next_output_byte = dest->next_output_byte;
295 state->free_in_buffer = dest->free_in_buffer;
300 /* Outputting bits to the file */
302 /* Only the right 24 bits of put_buffer are used; the valid bits are
303 * left-justified in this part. At most 16 bits can be passed to emit_bits
304 * in one call, and we never retain more than 7 bits in put_buffer
305 * between calls, so 24 bits are sufficient.
310 emit_bits (working_state * state, unsigned int code, int size)
311 /* Emit some bits; return TRUE if successful, FALSE if must suspend */
313 /* This routine is heavily used, so it's worth coding tightly. */
314 register INT32 put_buffer = (INT32) code;
315 register int put_bits = state->cur.put_bits;
317 /* if size is 0, caller used an invalid Huffman table entry */
319 ERREXIT(state->cinfo, JERR_HUFF_MISSING_CODE);
321 put_buffer &= (((INT32) 1)<<size) - 1; /* mask off any extra bits in code */
323 put_bits += size; /* new number of bits in buffer */
325 put_buffer <<= 24 - put_bits; /* align incoming bits */
327 put_buffer |= state->cur.put_buffer; /* and merge with old buffer contents */
329 while (put_bits >= 8) {
330 int c = (int) ((put_buffer >> 16) & 0xFF);
332 emit_byte(state, c, return FALSE);
333 if (c == 0xFF) { /* need to stuff a zero byte? */
334 emit_byte(state, 0, return FALSE);
340 state->cur.put_buffer = put_buffer; /* update state variables */
341 state->cur.put_bits = put_bits;
348 flush_bits (working_state * state)
350 if (! emit_bits(state, 0x7F, 7)) /* fill any partial byte with ones */
352 state->cur.put_buffer = 0; /* and reset bit-buffer to empty */
353 state->cur.put_bits = 0;
358 /* Encode a single block's worth of coefficients */
361 encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
362 c_derived_tbl *dctbl, c_derived_tbl *actbl)
364 register int temp, temp2;
366 register int k, r, i;
368 /* Encode the DC coefficient difference per section F.1.2.1 */
370 temp = temp2 = block[0] - last_dc_val;
373 temp = -temp; /* temp is abs value of input */
374 /* For a negative input, want temp2 = bitwise complement of abs(input) */
375 /* This code assumes we are on a two's complement machine */
379 /* Find the number of bits needed for the magnitude of the coefficient */
385 /* Check for out-of-range coefficient values.
386 * Since we're encoding a difference, the range limit is twice as much.
388 if (nbits > MAX_COEF_BITS+1)
389 ERREXIT(state->cinfo, JERR_BAD_DCT_COEF);
391 /* Emit the Huffman-coded symbol for the number of bits */
392 if (! emit_bits(state, dctbl->ehufco[nbits], dctbl->ehufsi[nbits]))
395 /* Emit that number of bits of the value, if positive, */
396 /* or the complement of its magnitude, if negative. */
397 if (nbits) /* emit_bits rejects calls with size 0 */
398 if (! emit_bits(state, (unsigned int) temp2, nbits))
401 /* Encode the AC coefficients per section F.1.2.2 */
403 r = 0; /* r = run length of zeros */
405 for (k = 1; k < DCTSIZE2; k++) {
406 if ((temp = block[jpeg_natural_order[k]]) == 0) {
409 /* if run length > 15, must emit special run-length-16 codes (0xF0) */
411 if (! emit_bits(state, actbl->ehufco[0xF0], actbl->ehufsi[0xF0]))
418 temp = -temp; /* temp is abs value of input */
419 /* This code assumes we are on a two's complement machine */
423 /* Find the number of bits needed for the magnitude of the coefficient */
424 nbits = 1; /* there must be at least one 1 bit */
427 /* Check for out-of-range coefficient values */
428 if (nbits > MAX_COEF_BITS)
429 ERREXIT(state->cinfo, JERR_BAD_DCT_COEF);
431 /* Emit Huffman symbol for run length / number of bits */
432 i = (r << 4) + nbits;
433 if (! emit_bits(state, actbl->ehufco[i], actbl->ehufsi[i]))
436 /* Emit that number of bits of the value, if positive, */
437 /* or the complement of its magnitude, if negative. */
438 if (! emit_bits(state, (unsigned int) temp2, nbits))
445 /* If the last coef(s) were zero, emit an end-of-block code */
447 if (! emit_bits(state, actbl->ehufco[0], actbl->ehufsi[0]))
455 * Emit a restart marker & resynchronize predictions.
459 emit_restart (working_state * state, int restart_num)
463 if (! flush_bits(state))
466 emit_byte(state, 0xFF, return FALSE);
467 emit_byte(state, JPEG_RST0 + restart_num, return FALSE);
469 /* Re-initialize DC predictions to 0 */
470 for (ci = 0; ci < state->cinfo->comps_in_scan; ci++)
471 state->cur.last_dc_val[ci] = 0;
473 /* The restart counter is not updated until we successfully write the MCU. */
480 * Encode and output one MCU's worth of Huffman-compressed coefficients.
484 encode_mcu_huff (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
486 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
489 jpeg_component_info * compptr;
491 /* Load up working state */
492 state.next_output_byte = cinfo->dest->next_output_byte;
493 state.free_in_buffer = cinfo->dest->free_in_buffer;
494 ASSIGN_STATE(state.cur, entropy->saved);
497 /* Emit restart marker if needed */
498 if (cinfo->restart_interval) {
499 if (entropy->restarts_to_go == 0)
500 if (! emit_restart(&state, entropy->next_restart_num))
504 /* Encode the MCU data blocks */
505 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
506 ci = cinfo->MCU_membership[blkn];
507 compptr = cinfo->cur_comp_info[ci];
508 if (! encode_one_block(&state,
509 MCU_data[blkn][0], state.cur.last_dc_val[ci],
510 entropy->dc_derived_tbls[compptr->dc_tbl_no],
511 entropy->ac_derived_tbls[compptr->ac_tbl_no]))
513 /* Update last_dc_val */
514 state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
517 /* Completed MCU, so update state */
518 cinfo->dest->next_output_byte = state.next_output_byte;
519 cinfo->dest->free_in_buffer = state.free_in_buffer;
520 ASSIGN_STATE(entropy->saved, state.cur);
522 /* Update restart-interval state too */
523 if (cinfo->restart_interval) {
524 if (entropy->restarts_to_go == 0) {
525 entropy->restarts_to_go = cinfo->restart_interval;
526 entropy->next_restart_num++;
527 entropy->next_restart_num &= 7;
529 entropy->restarts_to_go--;
537 * Finish up at the end of a Huffman-compressed scan.
541 finish_pass_huff (j_compress_ptr cinfo)
543 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
546 /* Load up working state ... flush_bits needs it */
547 state.next_output_byte = cinfo->dest->next_output_byte;
548 state.free_in_buffer = cinfo->dest->free_in_buffer;
549 ASSIGN_STATE(state.cur, entropy->saved);
552 /* Flush out the last data */
553 if (! flush_bits(&state))
554 ERREXIT(cinfo, JERR_CANT_SUSPEND);
557 cinfo->dest->next_output_byte = state.next_output_byte;
558 cinfo->dest->free_in_buffer = state.free_in_buffer;
559 ASSIGN_STATE(entropy->saved, state.cur);
564 * Huffman coding optimization.
566 * We first scan the supplied data and count the number of uses of each symbol
567 * that is to be Huffman-coded. (This process MUST agree with the code above.)
568 * Then we build a Huffman coding tree for the observed counts.
569 * Symbols which are not needed at all for the particular image are not
570 * assigned any code, which saves space in the DHT marker as well as in
571 * the compressed data.
574 #ifdef ENTROPY_OPT_SUPPORTED
577 /* Process a single block's worth of coefficients */
580 htest_one_block (j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val,
581 long dc_counts[], long ac_counts[])
587 /* Encode the DC coefficient difference per section F.1.2.1 */
589 temp = block[0] - last_dc_val;
593 /* Find the number of bits needed for the magnitude of the coefficient */
599 /* Check for out-of-range coefficient values.
600 * Since we're encoding a difference, the range limit is twice as much.
602 if (nbits > MAX_COEF_BITS+1)
603 ERREXIT(cinfo, JERR_BAD_DCT_COEF);
605 /* Count the Huffman symbol for the number of bits */
608 /* Encode the AC coefficients per section F.1.2.2 */
610 r = 0; /* r = run length of zeros */
612 for (k = 1; k < DCTSIZE2; k++) {
613 if ((temp = block[jpeg_natural_order[k]]) == 0) {
616 /* if run length > 15, must emit special run-length-16 codes (0xF0) */
622 /* Find the number of bits needed for the magnitude of the coefficient */
626 /* Find the number of bits needed for the magnitude of the coefficient */
627 nbits = 1; /* there must be at least one 1 bit */
630 /* Check for out-of-range coefficient values */
631 if (nbits > MAX_COEF_BITS)
632 ERREXIT(cinfo, JERR_BAD_DCT_COEF);
634 /* Count Huffman symbol for run length / number of bits */
635 ac_counts[(r << 4) + nbits]++;
641 /* If the last coef(s) were zero, emit an end-of-block code */
648 * Trial-encode one MCU's worth of Huffman-compressed coefficients.
649 * No data is actually output, so no suspension return is possible.
653 encode_mcu_gather (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
655 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
657 jpeg_component_info * compptr;
659 /* Take care of restart intervals if needed */
660 if (cinfo->restart_interval) {
661 if (entropy->restarts_to_go == 0) {
662 /* Re-initialize DC predictions to 0 */
663 for (ci = 0; ci < cinfo->comps_in_scan; ci++)
664 entropy->saved.last_dc_val[ci] = 0;
665 /* Update restart state */
666 entropy->restarts_to_go = cinfo->restart_interval;
668 entropy->restarts_to_go--;
671 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
672 ci = cinfo->MCU_membership[blkn];
673 compptr = cinfo->cur_comp_info[ci];
674 htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci],
675 entropy->dc_count_ptrs[compptr->dc_tbl_no],
676 entropy->ac_count_ptrs[compptr->ac_tbl_no]);
677 entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0];
685 * Generate the best Huffman code table for the given counts, fill htbl.
686 * Note this is also used by jcphuff.c.
688 * The JPEG standard requires that no symbol be assigned a codeword of all
689 * one bits (so that padding bits added at the end of a compressed segment
690 * can't look like a valid code). Because of the canonical ordering of
691 * codewords, this just means that there must be an unused slot in the
692 * longest codeword length category. Section K.2 of the JPEG spec suggests
693 * reserving such a slot by pretending that symbol 256 is a valid symbol
694 * with count 1. In theory that's not optimal; giving it count zero but
695 * including it in the symbol set anyway should give a better Huffman code.
696 * But the theoretically better code actually seems to come out worse in
697 * practice, because it produces more all-ones bytes (which incur stuffed
698 * zero bytes in the final file). In any case the difference is tiny.
700 * The JPEG standard requires Huffman codes to be no more than 16 bits long.
701 * If some symbols have a very small but nonzero probability, the Huffman tree
702 * must be adjusted to meet the code length restriction. We currently use
703 * the adjustment method suggested in JPEG section K.2. This method is *not*
704 * optimal; it may not choose the best possible limited-length code. But
705 * typically only very-low-frequency symbols will be given less-than-optimal
706 * lengths, so the code is almost optimal. Experimental comparisons against
707 * an optimal limited-length-code algorithm indicate that the difference is
708 * microscopic --- usually less than a hundredth of a percent of total size.
709 * So the extra complexity of an optimal algorithm doesn't seem worthwhile.
713 jpeg_gen_optimal_table (j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[])
715 #define MAX_CLEN 32 /* assumed maximum initial code length */
716 UINT8 bits[MAX_CLEN+1]; /* bits[k] = # of symbols with code length k */
717 int codesize[257]; /* codesize[k] = code length of symbol k */
718 int others[257]; /* next symbol in current branch of tree */
723 /* This algorithm is explained in section K.2 of the JPEG standard */
725 MEMZERO(bits, SIZEOF(bits));
726 MEMZERO(codesize, SIZEOF(codesize));
727 for (i = 0; i < 257; i++)
728 others[i] = -1; /* init links to empty */
730 freq[256] = 1; /* make sure 256 has a nonzero count */
731 /* Including the pseudo-symbol 256 in the Huffman procedure guarantees
732 * that no real symbol is given code-value of all ones, because 256
733 * will be placed last in the largest codeword category.
736 /* Huffman's basic algorithm to assign optimal code lengths to symbols */
739 /* Find the smallest nonzero frequency, set c1 = its symbol */
740 /* In case of ties, take the larger symbol number */
743 for (i = 0; i <= 256; i++) {
744 if (freq[i] && freq[i] <= v) {
750 /* Find the next smallest nonzero frequency, set c2 = its symbol */
751 /* In case of ties, take the larger symbol number */
754 for (i = 0; i <= 256; i++) {
755 if (freq[i] && freq[i] <= v && i != c1) {
761 /* Done if we've merged everything into one frequency */
765 /* Else merge the two counts/trees */
766 freq[c1] += freq[c2];
769 /* Increment the codesize of everything in c1's tree branch */
771 while (others[c1] >= 0) {
776 others[c1] = c2; /* chain c2 onto c1's tree branch */
778 /* Increment the codesize of everything in c2's tree branch */
780 while (others[c2] >= 0) {
786 /* Now count the number of symbols of each code length */
787 for (i = 0; i <= 256; i++) {
789 /* The JPEG standard seems to think that this can't happen, */
790 /* but I'm paranoid... */
791 if (codesize[i] > MAX_CLEN)
792 ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW);
798 /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure
799 * Huffman procedure assigned any such lengths, we must adjust the coding.
800 * Here is what the JPEG spec says about how this next bit works:
801 * Since symbols are paired for the longest Huffman code, the symbols are
802 * removed from this length category two at a time. The prefix for the pair
803 * (which is one bit shorter) is allocated to one of the pair; then,
804 * skipping the BITS entry for that prefix length, a code word from the next
805 * shortest nonzero BITS entry is converted into a prefix for two code words
809 for (i = MAX_CLEN; i > 16; i--) {
810 while (bits[i] > 0) {
811 j = i - 2; /* find length of new prefix to be used */
815 bits[i] -= 2; /* remove two symbols */
816 bits[i-1]++; /* one goes in this length */
817 bits[j+1] += 2; /* two new symbols in this length */
818 bits[j]--; /* symbol of this length is now a prefix */
822 /* Remove the count for the pseudo-symbol 256 from the largest codelength */
823 while (bits[i] == 0) /* find largest codelength still in use */
827 /* Return final symbol counts (only for lengths 0..16) */
828 MEMCOPY(htbl->bits, bits, SIZEOF(htbl->bits));
830 /* Return a list of the symbols sorted by code length */
831 /* It's not real clear to me why we don't need to consider the codelength
832 * changes made above, but the JPEG spec seems to think this works.
835 for (i = 1; i <= MAX_CLEN; i++) {
836 for (j = 0; j <= 255; j++) {
837 if (codesize[j] == i) {
838 htbl->huffval[p] = (UINT8) j;
844 /* Set sent_table FALSE so updated table will be written to JPEG file. */
845 htbl->sent_table = FALSE;
850 * Finish up a statistics-gathering pass and create the new Huffman tables.
854 finish_pass_gather (j_compress_ptr cinfo)
856 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
857 int ci, dctbl, actbl;
858 jpeg_component_info * compptr;
860 boolean did_dc[NUM_HUFF_TBLS];
861 boolean did_ac[NUM_HUFF_TBLS];
863 /* It's important not to apply jpeg_gen_optimal_table more than once
864 * per table, because it clobbers the input frequency counts!
866 MEMZERO(did_dc, SIZEOF(did_dc));
867 MEMZERO(did_ac, SIZEOF(did_ac));
869 for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
870 compptr = cinfo->cur_comp_info[ci];
871 dctbl = compptr->dc_tbl_no;
872 actbl = compptr->ac_tbl_no;
873 if (! did_dc[dctbl]) {
874 htblptr = & cinfo->dc_huff_tbl_ptrs[dctbl];
875 if (*htblptr == NULL)
876 *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
877 jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]);
878 did_dc[dctbl] = TRUE;
880 if (! did_ac[actbl]) {
881 htblptr = & cinfo->ac_huff_tbl_ptrs[actbl];
882 if (*htblptr == NULL)
883 *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
884 jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]);
885 did_ac[actbl] = TRUE;
891 #endif /* ENTROPY_OPT_SUPPORTED */
895 * Module initialization routine for Huffman entropy encoding.
899 jinit_huff_encoder (j_compress_ptr cinfo)
901 huff_entropy_ptr entropy;
904 entropy = (huff_entropy_ptr)
905 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
906 SIZEOF(huff_entropy_encoder));
907 cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
908 entropy->pub.start_pass = start_pass_huff;
910 /* Mark tables unallocated */
911 for (i = 0; i < NUM_HUFF_TBLS; i++) {
912 entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
913 #ifdef ENTROPY_OPT_SUPPORTED
914 entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL;