4 * This file was part of the Independent JPEG Group's software:
5 * Copyright (C) 1991-1997, Thomas G. Lane.
6 * libjpeg-turbo Modifications:
7 * Copyright (C) 2009-2011, D. R. Commander.
8 * For conditions of distribution and use, see the accompanying README file.
10 * This file contains Huffman entropy encoding routines.
12 * Much of the complexity here has to do with supporting output suspension.
13 * If the data destination module demands suspension, we want to be able to
14 * back up to the start of the current MCU. To do this, we copy state
15 * variables into local working storage, and update them back to the
16 * permanent JPEG objects only upon successful completion of an MCU.
19 #define JPEG_INTERNALS
22 #include "jchuff.h" /* Declarations shared with jcphuff.c */
25 static unsigned char jpeg_nbits_table[65536];
26 static int jpeg_nbits_table_init = 0;
29 #define min(a,b) ((a)<(b)?(a):(b))
33 /* Expanded entropy encoder object for Huffman encoding.
35 * The savable_state subrecord contains fields that change within an MCU,
36 * but must not be updated permanently until we complete the MCU.
40 size_t put_buffer; /* current bit-accumulation buffer */
41 int put_bits; /* # of bits now in it */
42 int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
45 /* This macro is to work around compilers with missing or broken
46 * structure assignment. You'll need to fix this code if you have
47 * such a compiler and you change MAX_COMPS_IN_SCAN.
50 #ifndef NO_STRUCT_ASSIGN
51 #define ASSIGN_STATE(dest,src) ((dest) = (src))
53 #if MAX_COMPS_IN_SCAN == 4
54 #define ASSIGN_STATE(dest,src) \
55 ((dest).put_buffer = (src).put_buffer, \
56 (dest).put_bits = (src).put_bits, \
57 (dest).last_dc_val[0] = (src).last_dc_val[0], \
58 (dest).last_dc_val[1] = (src).last_dc_val[1], \
59 (dest).last_dc_val[2] = (src).last_dc_val[2], \
60 (dest).last_dc_val[3] = (src).last_dc_val[3])
66 struct jpeg_entropy_encoder pub; /* public fields */
68 savable_state saved; /* Bit buffer & DC state at start of MCU */
70 /* These fields are NOT loaded into local working state. */
71 unsigned int restarts_to_go; /* MCUs left in this restart interval */
72 int next_restart_num; /* next restart number to write (0-7) */
74 /* Pointers to derived tables (these workspaces have image lifespan) */
75 c_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
76 c_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
78 #ifdef ENTROPY_OPT_SUPPORTED /* Statistics tables for optimization */
79 long * dc_count_ptrs[NUM_HUFF_TBLS];
80 long * ac_count_ptrs[NUM_HUFF_TBLS];
82 } huff_entropy_encoder;
84 typedef huff_entropy_encoder * huff_entropy_ptr;
86 /* Working state while writing an MCU.
87 * This struct contains all the fields that are needed by subroutines.
91 JOCTET * next_output_byte; /* => next byte to write in buffer */
92 size_t free_in_buffer; /* # of byte spaces remaining in buffer */
93 savable_state cur; /* Current bit buffer & DC state */
94 j_compress_ptr cinfo; /* dump_buffer needs access to this */
98 /* Forward declarations */
99 METHODDEF(boolean) encode_mcu_huff JPP((j_compress_ptr cinfo,
100 JBLOCKROW *MCU_data));
101 METHODDEF(void) finish_pass_huff JPP((j_compress_ptr cinfo));
102 #ifdef ENTROPY_OPT_SUPPORTED
103 METHODDEF(boolean) encode_mcu_gather JPP((j_compress_ptr cinfo,
104 JBLOCKROW *MCU_data));
105 METHODDEF(void) finish_pass_gather JPP((j_compress_ptr cinfo));
110 * Initialize for a Huffman-compressed scan.
111 * If gather_statistics is TRUE, we do not output anything during the scan,
112 * just count the Huffman symbols used and generate Huffman code tables.
116 start_pass_huff (j_compress_ptr cinfo, boolean gather_statistics)
118 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
119 int ci, dctbl, actbl;
120 jpeg_component_info * compptr;
122 if (gather_statistics) {
123 #ifdef ENTROPY_OPT_SUPPORTED
124 entropy->pub.encode_mcu = encode_mcu_gather;
125 entropy->pub.finish_pass = finish_pass_gather;
127 ERREXIT(cinfo, JERR_NOT_COMPILED);
130 entropy->pub.encode_mcu = encode_mcu_huff;
131 entropy->pub.finish_pass = finish_pass_huff;
134 for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
135 compptr = cinfo->cur_comp_info[ci];
136 dctbl = compptr->dc_tbl_no;
137 actbl = compptr->ac_tbl_no;
138 if (gather_statistics) {
139 #ifdef ENTROPY_OPT_SUPPORTED
140 /* Check for invalid table indexes */
141 /* (make_c_derived_tbl does this in the other path) */
142 if (dctbl < 0 || dctbl >= NUM_HUFF_TBLS)
143 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, dctbl);
144 if (actbl < 0 || actbl >= NUM_HUFF_TBLS)
145 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, actbl);
146 /* Allocate and zero the statistics tables */
147 /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
148 if (entropy->dc_count_ptrs[dctbl] == NULL)
149 entropy->dc_count_ptrs[dctbl] = (long *)
150 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
152 MEMZERO(entropy->dc_count_ptrs[dctbl], 257 * SIZEOF(long));
153 if (entropy->ac_count_ptrs[actbl] == NULL)
154 entropy->ac_count_ptrs[actbl] = (long *)
155 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
157 MEMZERO(entropy->ac_count_ptrs[actbl], 257 * SIZEOF(long));
160 /* Compute derived values for Huffman tables */
161 /* We may do this more than once for a table, but it's not expensive */
162 jpeg_make_c_derived_tbl(cinfo, TRUE, dctbl,
163 & entropy->dc_derived_tbls[dctbl]);
164 jpeg_make_c_derived_tbl(cinfo, FALSE, actbl,
165 & entropy->ac_derived_tbls[actbl]);
167 /* Initialize DC predictions to 0 */
168 entropy->saved.last_dc_val[ci] = 0;
171 /* Initialize bit buffer to empty */
172 entropy->saved.put_buffer = 0;
173 entropy->saved.put_bits = 0;
175 /* Initialize restart stuff */
176 entropy->restarts_to_go = cinfo->restart_interval;
177 entropy->next_restart_num = 0;
182 * Compute the derived values for a Huffman table.
183 * This routine also performs some validation checks on the table.
185 * Note this is also used by jcphuff.c.
189 jpeg_make_c_derived_tbl (j_compress_ptr cinfo, boolean isDC, int tblno,
190 c_derived_tbl ** pdtbl)
194 int p, i, l, lastp, si, maxsymbol;
196 unsigned int huffcode[257];
199 /* Note that huffsize[] and huffcode[] are filled in code-length order,
200 * paralleling the order of the symbols themselves in htbl->huffval[].
203 /* Find the input Huffman table */
204 if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
205 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
207 isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
209 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
211 /* Allocate a workspace if we haven't already done so. */
213 *pdtbl = (c_derived_tbl *)
214 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
215 SIZEOF(c_derived_tbl));
218 /* Figure C.1: make table of Huffman code length for each symbol */
221 for (l = 1; l <= 16; l++) {
222 i = (int) htbl->bits[l];
223 if (i < 0 || p + i > 256) /* protect against table overrun */
224 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
226 huffsize[p++] = (char) l;
231 /* Figure C.2: generate the codes themselves */
232 /* We also validate that the counts represent a legal Huffman code tree. */
237 while (huffsize[p]) {
238 while (((int) huffsize[p]) == si) {
239 huffcode[p++] = code;
242 /* code is now 1 more than the last code used for codelength si; but
243 * it must still fit in si bits, since no code is allowed to be all ones.
245 if (((INT32) code) >= (((INT32) 1) << si))
246 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
251 /* Figure C.3: generate encoding tables */
252 /* These are code and size indexed by symbol value */
254 /* Set all codeless symbols to have code length 0;
255 * this lets us detect duplicate VAL entries here, and later
256 * allows emit_bits to detect any attempt to emit such symbols.
258 MEMZERO(dtbl->ehufsi, SIZEOF(dtbl->ehufsi));
260 /* This is also a convenient place to check for out-of-range
261 * and duplicated VAL entries. We allow 0..255 for AC symbols
262 * but only 0..15 for DC. (We could constrain them further
263 * based on data depth and mode, but this seems enough.)
265 maxsymbol = isDC ? 15 : 255;
267 for (p = 0; p < lastp; p++) {
268 i = htbl->huffval[p];
269 if (i < 0 || i > maxsymbol || dtbl->ehufsi[i])
270 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
271 dtbl->ehufco[i] = huffcode[p];
272 dtbl->ehufsi[i] = huffsize[p];
275 if(!jpeg_nbits_table_init) {
276 for(i = 0; i < 65536; i++) {
277 int nbits = 0, temp = i;
278 while (temp) {temp >>= 1; nbits++;}
279 jpeg_nbits_table[i] = nbits;
281 jpeg_nbits_table_init = 1;
286 /* Outputting bytes to the file */
288 /* Emit a byte, taking 'action' if must suspend. */
289 #define emit_byte(state,val,action) \
290 { *(state)->next_output_byte++ = (JOCTET) (val); \
291 if (--(state)->free_in_buffer == 0) \
292 if (! dump_buffer(state)) \
297 dump_buffer (working_state * state)
298 /* Empty the output buffer; return TRUE if successful, FALSE if must suspend */
300 struct jpeg_destination_mgr * dest = state->cinfo->dest;
302 if (! (*dest->empty_output_buffer) (state->cinfo))
304 /* After a successful buffer dump, must reset buffer pointers */
305 state->next_output_byte = dest->next_output_byte;
306 state->free_in_buffer = dest->free_in_buffer;
311 /* Outputting bits to the file */
313 /* These macros perform the same task as the emit_bits() function in the
314 * original libjpeg code. In addition to reducing overhead by explicitly
315 * inlining the code, additional performance is achieved by taking into
316 * account the size of the bit buffer and waiting until it is almost full
317 * before emptying it. This mostly benefits 64-bit platforms, since 6
318 * bytes can be stored in a 64-bit bit buffer before it has to be emptied.
321 #define EMIT_BYTE() { \
324 c = (JOCTET)GETJOCTET(put_buffer >> put_bits); \
326 if (c == 0xFF) /* need to stuff a zero byte? */ \
330 #define PUT_BITS(code, size) { \
332 put_buffer = (put_buffer << size) | code; \
335 #define CHECKBUF15() { \
336 if (put_bits > 15) { \
342 #define CHECKBUF31() { \
343 if (put_bits > 31) { \
351 #define CHECKBUF47() { \
352 if (put_bits > 47) { \
362 #if __WORDSIZE==64 || defined(_WIN64)
364 #define EMIT_BITS(code, size) { \
366 PUT_BITS(code, size) \
369 #define EMIT_CODE(code, size) { \
370 temp2 &= (((INT32) 1)<<nbits) - 1; \
372 PUT_BITS(code, size) \
373 PUT_BITS(temp2, nbits) \
378 #define EMIT_BITS(code, size) { \
379 PUT_BITS(code, size) \
383 #define EMIT_CODE(code, size) { \
384 temp2 &= (((INT32) 1)<<nbits) - 1; \
385 PUT_BITS(code, size) \
387 PUT_BITS(temp2, nbits) \
394 #define BUFSIZE (DCTSIZE2 * 2)
396 #define LOAD_BUFFER() { \
397 if (state->free_in_buffer < BUFSIZE) { \
401 else buffer = state->next_output_byte; \
404 #define STORE_BUFFER() { \
406 bytes = buffer - _buffer; \
408 while (bytes > 0) { \
409 bytestocopy = min(bytes, state->free_in_buffer); \
410 MEMCOPY(state->next_output_byte, buffer, bytestocopy); \
411 state->next_output_byte += bytestocopy; \
412 buffer += bytestocopy; \
413 state->free_in_buffer -= bytestocopy; \
414 if (state->free_in_buffer == 0) \
415 if (! dump_buffer(state)) return FALSE; \
416 bytes -= bytestocopy; \
420 state->free_in_buffer -= (buffer - state->next_output_byte); \
421 state->next_output_byte = buffer; \
427 flush_bits (working_state * state)
429 JOCTET _buffer[BUFSIZE], *buffer;
430 size_t put_buffer; int put_bits;
431 size_t bytes, bytestocopy; int localbuf = 0;
433 put_buffer = state->cur.put_buffer;
434 put_bits = state->cur.put_bits;
437 /* fill any partial byte with ones */
439 while (put_bits >= 8) EMIT_BYTE()
441 state->cur.put_buffer = 0; /* and reset bit-buffer to empty */
442 state->cur.put_bits = 0;
449 /* Encode a single block's worth of coefficients */
452 encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
453 c_derived_tbl *dctbl, c_derived_tbl *actbl)
455 int temp, temp2, temp3;
458 JOCTET _buffer[BUFSIZE], *buffer;
459 size_t put_buffer; int put_bits;
460 int code_0xf0 = actbl->ehufco[0xf0], size_0xf0 = actbl->ehufsi[0xf0];
461 size_t bytes, bytestocopy; int localbuf = 0;
463 put_buffer = state->cur.put_buffer;
464 put_bits = state->cur.put_bits;
467 /* Encode the DC coefficient difference per section F.1.2.1 */
469 temp = temp2 = block[0] - last_dc_val;
471 /* This is a well-known technique for obtaining the absolute value without a
472 * branch. It is derived from an assembly language technique presented in
473 * "How to Optimize for the Pentium Processors", Copyright (c) 1996, 1997 by
476 temp3 = temp >> (CHAR_BIT * sizeof(int) - 1);
480 /* For a negative input, want temp2 = bitwise complement of abs(input) */
481 /* This code assumes we are on a two's complement machine */
484 /* Find the number of bits needed for the magnitude of the coefficient */
485 nbits = jpeg_nbits_table[temp];
487 /* Emit the Huffman-coded symbol for the number of bits */
488 code = dctbl->ehufco[nbits];
489 size = dctbl->ehufsi[nbits];
493 /* Mask off any extra bits in code */
494 temp2 &= (((INT32) 1)<<nbits) - 1;
496 /* Emit that number of bits of the value, if positive, */
497 /* or the complement of its magnitude, if negative. */
498 PUT_BITS(temp2, nbits)
501 /* Encode the AC coefficients per section F.1.2.2 */
503 r = 0; /* r = run length of zeros */
505 /* Manually unroll the k loop to eliminate the counter variable. This
506 * improves performance greatly on systems with a limited number of
507 * registers (such as x86.)
509 #define kloop(jpeg_natural_order_of_k) { \
510 if ((temp = block[jpeg_natural_order_of_k]) == 0) { \
514 /* Branch-less absolute value, bitwise complement, etc., same as above */ \
515 temp3 = temp >> (CHAR_BIT * sizeof(int) - 1); \
519 nbits = jpeg_nbits_table[temp]; \
520 /* if run length > 15, must emit special run-length-16 codes (0xF0) */ \
522 EMIT_BITS(code_0xf0, size_0xf0) \
525 /* Emit Huffman symbol for run length / number of bits */ \
526 temp3 = (r << 4) + nbits; \
527 code = actbl->ehufco[temp3]; \
528 size = actbl->ehufsi[temp3]; \
529 EMIT_CODE(code, size) \
534 /* One iteration for each value in jpeg_natural_order[] */
535 kloop(1); kloop(8); kloop(16); kloop(9); kloop(2); kloop(3);
536 kloop(10); kloop(17); kloop(24); kloop(32); kloop(25); kloop(18);
537 kloop(11); kloop(4); kloop(5); kloop(12); kloop(19); kloop(26);
538 kloop(33); kloop(40); kloop(48); kloop(41); kloop(34); kloop(27);
539 kloop(20); kloop(13); kloop(6); kloop(7); kloop(14); kloop(21);
540 kloop(28); kloop(35); kloop(42); kloop(49); kloop(56); kloop(57);
541 kloop(50); kloop(43); kloop(36); kloop(29); kloop(22); kloop(15);
542 kloop(23); kloop(30); kloop(37); kloop(44); kloop(51); kloop(58);
543 kloop(59); kloop(52); kloop(45); kloop(38); kloop(31); kloop(39);
544 kloop(46); kloop(53); kloop(60); kloop(61); kloop(54); kloop(47);
545 kloop(55); kloop(62); kloop(63);
547 /* If the last coef(s) were zero, emit an end-of-block code */
549 code = actbl->ehufco[0];
550 size = actbl->ehufsi[0];
551 EMIT_BITS(code, size)
554 state->cur.put_buffer = put_buffer;
555 state->cur.put_bits = put_bits;
563 * Emit a restart marker & resynchronize predictions.
567 emit_restart (working_state * state, int restart_num)
571 if (! flush_bits(state))
574 emit_byte(state, 0xFF, return FALSE);
575 emit_byte(state, JPEG_RST0 + restart_num, return FALSE);
577 /* Re-initialize DC predictions to 0 */
578 for (ci = 0; ci < state->cinfo->comps_in_scan; ci++)
579 state->cur.last_dc_val[ci] = 0;
581 /* The restart counter is not updated until we successfully write the MCU. */
588 * Encode and output one MCU's worth of Huffman-compressed coefficients.
592 encode_mcu_huff (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
594 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
597 jpeg_component_info * compptr;
599 /* Load up working state */
600 state.next_output_byte = cinfo->dest->next_output_byte;
601 state.free_in_buffer = cinfo->dest->free_in_buffer;
602 ASSIGN_STATE(state.cur, entropy->saved);
605 /* Emit restart marker if needed */
606 if (cinfo->restart_interval) {
607 if (entropy->restarts_to_go == 0)
608 if (! emit_restart(&state, entropy->next_restart_num))
612 /* Encode the MCU data blocks */
613 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
614 ci = cinfo->MCU_membership[blkn];
615 compptr = cinfo->cur_comp_info[ci];
616 if (! encode_one_block(&state,
617 MCU_data[blkn][0], state.cur.last_dc_val[ci],
618 entropy->dc_derived_tbls[compptr->dc_tbl_no],
619 entropy->ac_derived_tbls[compptr->ac_tbl_no]))
621 /* Update last_dc_val */
622 state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
625 /* Completed MCU, so update state */
626 cinfo->dest->next_output_byte = state.next_output_byte;
627 cinfo->dest->free_in_buffer = state.free_in_buffer;
628 ASSIGN_STATE(entropy->saved, state.cur);
630 /* Update restart-interval state too */
631 if (cinfo->restart_interval) {
632 if (entropy->restarts_to_go == 0) {
633 entropy->restarts_to_go = cinfo->restart_interval;
634 entropy->next_restart_num++;
635 entropy->next_restart_num &= 7;
637 entropy->restarts_to_go--;
645 * Finish up at the end of a Huffman-compressed scan.
649 finish_pass_huff (j_compress_ptr cinfo)
651 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
654 /* Load up working state ... flush_bits needs it */
655 state.next_output_byte = cinfo->dest->next_output_byte;
656 state.free_in_buffer = cinfo->dest->free_in_buffer;
657 ASSIGN_STATE(state.cur, entropy->saved);
660 /* Flush out the last data */
661 if (! flush_bits(&state))
662 ERREXIT(cinfo, JERR_CANT_SUSPEND);
665 cinfo->dest->next_output_byte = state.next_output_byte;
666 cinfo->dest->free_in_buffer = state.free_in_buffer;
667 ASSIGN_STATE(entropy->saved, state.cur);
672 * Huffman coding optimization.
674 * We first scan the supplied data and count the number of uses of each symbol
675 * that is to be Huffman-coded. (This process MUST agree with the code above.)
676 * Then we build a Huffman coding tree for the observed counts.
677 * Symbols which are not needed at all for the particular image are not
678 * assigned any code, which saves space in the DHT marker as well as in
679 * the compressed data.
682 #ifdef ENTROPY_OPT_SUPPORTED
685 /* Process a single block's worth of coefficients */
688 htest_one_block (j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val,
689 long dc_counts[], long ac_counts[])
695 /* Encode the DC coefficient difference per section F.1.2.1 */
697 temp = block[0] - last_dc_val;
701 /* Find the number of bits needed for the magnitude of the coefficient */
707 /* Check for out-of-range coefficient values.
708 * Since we're encoding a difference, the range limit is twice as much.
710 if (nbits > MAX_COEF_BITS+1)
711 ERREXIT(cinfo, JERR_BAD_DCT_COEF);
713 /* Count the Huffman symbol for the number of bits */
716 /* Encode the AC coefficients per section F.1.2.2 */
718 r = 0; /* r = run length of zeros */
720 for (k = 1; k < DCTSIZE2; k++) {
721 if ((temp = block[jpeg_natural_order[k]]) == 0) {
724 /* if run length > 15, must emit special run-length-16 codes (0xF0) */
730 /* Find the number of bits needed for the magnitude of the coefficient */
734 /* Find the number of bits needed for the magnitude of the coefficient */
735 nbits = 1; /* there must be at least one 1 bit */
738 /* Check for out-of-range coefficient values */
739 if (nbits > MAX_COEF_BITS)
740 ERREXIT(cinfo, JERR_BAD_DCT_COEF);
742 /* Count Huffman symbol for run length / number of bits */
743 ac_counts[(r << 4) + nbits]++;
749 /* If the last coef(s) were zero, emit an end-of-block code */
756 * Trial-encode one MCU's worth of Huffman-compressed coefficients.
757 * No data is actually output, so no suspension return is possible.
761 encode_mcu_gather (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
763 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
765 jpeg_component_info * compptr;
767 /* Take care of restart intervals if needed */
768 if (cinfo->restart_interval) {
769 if (entropy->restarts_to_go == 0) {
770 /* Re-initialize DC predictions to 0 */
771 for (ci = 0; ci < cinfo->comps_in_scan; ci++)
772 entropy->saved.last_dc_val[ci] = 0;
773 /* Update restart state */
774 entropy->restarts_to_go = cinfo->restart_interval;
776 entropy->restarts_to_go--;
779 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
780 ci = cinfo->MCU_membership[blkn];
781 compptr = cinfo->cur_comp_info[ci];
782 htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci],
783 entropy->dc_count_ptrs[compptr->dc_tbl_no],
784 entropy->ac_count_ptrs[compptr->ac_tbl_no]);
785 entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0];
793 * Generate the best Huffman code table for the given counts, fill htbl.
794 * Note this is also used by jcphuff.c.
796 * The JPEG standard requires that no symbol be assigned a codeword of all
797 * one bits (so that padding bits added at the end of a compressed segment
798 * can't look like a valid code). Because of the canonical ordering of
799 * codewords, this just means that there must be an unused slot in the
800 * longest codeword length category. Section K.2 of the JPEG spec suggests
801 * reserving such a slot by pretending that symbol 256 is a valid symbol
802 * with count 1. In theory that's not optimal; giving it count zero but
803 * including it in the symbol set anyway should give a better Huffman code.
804 * But the theoretically better code actually seems to come out worse in
805 * practice, because it produces more all-ones bytes (which incur stuffed
806 * zero bytes in the final file). In any case the difference is tiny.
808 * The JPEG standard requires Huffman codes to be no more than 16 bits long.
809 * If some symbols have a very small but nonzero probability, the Huffman tree
810 * must be adjusted to meet the code length restriction. We currently use
811 * the adjustment method suggested in JPEG section K.2. This method is *not*
812 * optimal; it may not choose the best possible limited-length code. But
813 * typically only very-low-frequency symbols will be given less-than-optimal
814 * lengths, so the code is almost optimal. Experimental comparisons against
815 * an optimal limited-length-code algorithm indicate that the difference is
816 * microscopic --- usually less than a hundredth of a percent of total size.
817 * So the extra complexity of an optimal algorithm doesn't seem worthwhile.
821 jpeg_gen_optimal_table (j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[])
823 #define MAX_CLEN 32 /* assumed maximum initial code length */
824 UINT8 bits[MAX_CLEN+1]; /* bits[k] = # of symbols with code length k */
825 int codesize[257]; /* codesize[k] = code length of symbol k */
826 int others[257]; /* next symbol in current branch of tree */
831 /* This algorithm is explained in section K.2 of the JPEG standard */
833 MEMZERO(bits, SIZEOF(bits));
834 MEMZERO(codesize, SIZEOF(codesize));
835 for (i = 0; i < 257; i++)
836 others[i] = -1; /* init links to empty */
838 freq[256] = 1; /* make sure 256 has a nonzero count */
839 /* Including the pseudo-symbol 256 in the Huffman procedure guarantees
840 * that no real symbol is given code-value of all ones, because 256
841 * will be placed last in the largest codeword category.
844 /* Huffman's basic algorithm to assign optimal code lengths to symbols */
847 /* Find the smallest nonzero frequency, set c1 = its symbol */
848 /* In case of ties, take the larger symbol number */
851 for (i = 0; i <= 256; i++) {
852 if (freq[i] && freq[i] <= v) {
858 /* Find the next smallest nonzero frequency, set c2 = its symbol */
859 /* In case of ties, take the larger symbol number */
862 for (i = 0; i <= 256; i++) {
863 if (freq[i] && freq[i] <= v && i != c1) {
869 /* Done if we've merged everything into one frequency */
873 /* Else merge the two counts/trees */
874 freq[c1] += freq[c2];
877 /* Increment the codesize of everything in c1's tree branch */
879 while (others[c1] >= 0) {
884 others[c1] = c2; /* chain c2 onto c1's tree branch */
886 /* Increment the codesize of everything in c2's tree branch */
888 while (others[c2] >= 0) {
894 /* Now count the number of symbols of each code length */
895 for (i = 0; i <= 256; i++) {
897 /* The JPEG standard seems to think that this can't happen, */
898 /* but I'm paranoid... */
899 if (codesize[i] > MAX_CLEN)
900 ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW);
906 /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure
907 * Huffman procedure assigned any such lengths, we must adjust the coding.
908 * Here is what the JPEG spec says about how this next bit works:
909 * Since symbols are paired for the longest Huffman code, the symbols are
910 * removed from this length category two at a time. The prefix for the pair
911 * (which is one bit shorter) is allocated to one of the pair; then,
912 * skipping the BITS entry for that prefix length, a code word from the next
913 * shortest nonzero BITS entry is converted into a prefix for two code words
917 for (i = MAX_CLEN; i > 16; i--) {
918 while (bits[i] > 0) {
919 j = i - 2; /* find length of new prefix to be used */
923 bits[i] -= 2; /* remove two symbols */
924 bits[i-1]++; /* one goes in this length */
925 bits[j+1] += 2; /* two new symbols in this length */
926 bits[j]--; /* symbol of this length is now a prefix */
930 /* Remove the count for the pseudo-symbol 256 from the largest codelength */
931 while (bits[i] == 0) /* find largest codelength still in use */
935 /* Return final symbol counts (only for lengths 0..16) */
936 MEMCOPY(htbl->bits, bits, SIZEOF(htbl->bits));
938 /* Return a list of the symbols sorted by code length */
939 /* It's not real clear to me why we don't need to consider the codelength
940 * changes made above, but the JPEG spec seems to think this works.
943 for (i = 1; i <= MAX_CLEN; i++) {
944 for (j = 0; j <= 255; j++) {
945 if (codesize[j] == i) {
946 htbl->huffval[p] = (UINT8) j;
952 /* Set sent_table FALSE so updated table will be written to JPEG file. */
953 htbl->sent_table = FALSE;
958 * Finish up a statistics-gathering pass and create the new Huffman tables.
962 finish_pass_gather (j_compress_ptr cinfo)
964 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
965 int ci, dctbl, actbl;
966 jpeg_component_info * compptr;
968 boolean did_dc[NUM_HUFF_TBLS];
969 boolean did_ac[NUM_HUFF_TBLS];
971 /* It's important not to apply jpeg_gen_optimal_table more than once
972 * per table, because it clobbers the input frequency counts!
974 MEMZERO(did_dc, SIZEOF(did_dc));
975 MEMZERO(did_ac, SIZEOF(did_ac));
977 for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
978 compptr = cinfo->cur_comp_info[ci];
979 dctbl = compptr->dc_tbl_no;
980 actbl = compptr->ac_tbl_no;
981 if (! did_dc[dctbl]) {
982 htblptr = & cinfo->dc_huff_tbl_ptrs[dctbl];
983 if (*htblptr == NULL)
984 *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
985 jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]);
986 did_dc[dctbl] = TRUE;
988 if (! did_ac[actbl]) {
989 htblptr = & cinfo->ac_huff_tbl_ptrs[actbl];
990 if (*htblptr == NULL)
991 *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
992 jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]);
993 did_ac[actbl] = TRUE;
999 #endif /* ENTROPY_OPT_SUPPORTED */
1003 * Module initialization routine for Huffman entropy encoding.
1007 jinit_huff_encoder (j_compress_ptr cinfo)
1009 huff_entropy_ptr entropy;
1012 entropy = (huff_entropy_ptr)
1013 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
1014 SIZEOF(huff_entropy_encoder));
1015 cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
1016 entropy->pub.start_pass = start_pass_huff;
1018 /* Mark tables unallocated */
1019 for (i = 0; i < NUM_HUFF_TBLS; i++) {
1020 entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
1021 #ifdef ENTROPY_OPT_SUPPORTED
1022 entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL;