bug 839193 - test for corresponding bug r=bent
[gecko.git] / media / libjpeg / jchuff.c
blob124d8845fa6389ed429812e1c55ab3b0003288c9
1 /*
2 * jchuff.c
4 * Copyright (C) 1991-1997, Thomas G. Lane.
5 * Copyright (C) 2009-2011, D. R. Commander.
6 * This file is part of the Independent JPEG Group's software.
7 * For conditions of distribution and use, see the accompanying README file.
9 * This file contains Huffman entropy encoding routines.
11 * Much of the complexity here has to do with supporting output suspension.
12 * If the data destination module demands suspension, we want to be able to
13 * back up to the start of the current MCU. To do this, we copy state
14 * variables into local working storage, and update them back to the
15 * permanent JPEG objects only upon successful completion of an MCU.
18 #define JPEG_INTERNALS
19 #include "jinclude.h"
20 #include "jpeglib.h"
21 #include "jchuff.h" /* Declarations shared with jcphuff.c */
22 #include <limits.h>
24 static const unsigned char jpeg_nbits_table[65536] = {
25 /* Number i needs jpeg_nbits_table[i] bits to be represented. */
26 #include "jpeg_nbits_table.h"
29 #ifndef min
30 #define min(a,b) ((a)<(b)?(a):(b))
31 #endif
34 /* Expanded entropy encoder object for Huffman encoding.
36 * The savable_state subrecord contains fields that change within an MCU,
37 * but must not be updated permanently until we complete the MCU.
40 typedef struct {
41 size_t put_buffer; /* current bit-accumulation buffer */
42 int put_bits; /* # of bits now in it */
43 int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
44 } savable_state;
46 /* This macro is to work around compilers with missing or broken
47 * structure assignment. You'll need to fix this code if you have
48 * such a compiler and you change MAX_COMPS_IN_SCAN.
51 #ifndef NO_STRUCT_ASSIGN
52 #define ASSIGN_STATE(dest,src) ((dest) = (src))
53 #else
54 #if MAX_COMPS_IN_SCAN == 4
55 #define ASSIGN_STATE(dest,src) \
56 ((dest).put_buffer = (src).put_buffer, \
57 (dest).put_bits = (src).put_bits, \
58 (dest).last_dc_val[0] = (src).last_dc_val[0], \
59 (dest).last_dc_val[1] = (src).last_dc_val[1], \
60 (dest).last_dc_val[2] = (src).last_dc_val[2], \
61 (dest).last_dc_val[3] = (src).last_dc_val[3])
62 #endif
63 #endif
66 typedef struct {
67 struct jpeg_entropy_encoder pub; /* public fields */
69 savable_state saved; /* Bit buffer & DC state at start of MCU */
71 /* These fields are NOT loaded into local working state. */
72 unsigned int restarts_to_go; /* MCUs left in this restart interval */
73 int next_restart_num; /* next restart number to write (0-7) */
75 /* Pointers to derived tables (these workspaces have image lifespan) */
76 c_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
77 c_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
79 #ifdef ENTROPY_OPT_SUPPORTED /* Statistics tables for optimization */
80 long * dc_count_ptrs[NUM_HUFF_TBLS];
81 long * ac_count_ptrs[NUM_HUFF_TBLS];
82 #endif
83 } huff_entropy_encoder;
85 typedef huff_entropy_encoder * huff_entropy_ptr;
87 /* Working state while writing an MCU.
88 * This struct contains all the fields that are needed by subroutines.
91 typedef struct {
92 JOCTET * next_output_byte; /* => next byte to write in buffer */
93 size_t free_in_buffer; /* # of byte spaces remaining in buffer */
94 savable_state cur; /* Current bit buffer & DC state */
95 j_compress_ptr cinfo; /* dump_buffer needs access to this */
96 } working_state;
99 /* Forward declarations */
100 METHODDEF(boolean) encode_mcu_huff JPP((j_compress_ptr cinfo,
101 JBLOCKROW *MCU_data));
102 METHODDEF(void) finish_pass_huff JPP((j_compress_ptr cinfo));
103 #ifdef ENTROPY_OPT_SUPPORTED
104 METHODDEF(boolean) encode_mcu_gather JPP((j_compress_ptr cinfo,
105 JBLOCKROW *MCU_data));
106 METHODDEF(void) finish_pass_gather JPP((j_compress_ptr cinfo));
107 #endif
111 * Initialize for a Huffman-compressed scan.
112 * If gather_statistics is TRUE, we do not output anything during the scan,
113 * just count the Huffman symbols used and generate Huffman code tables.
116 METHODDEF(void)
117 start_pass_huff (j_compress_ptr cinfo, boolean gather_statistics)
119 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
120 int ci, dctbl, actbl;
121 jpeg_component_info * compptr;
123 if (gather_statistics) {
124 #ifdef ENTROPY_OPT_SUPPORTED
125 entropy->pub.encode_mcu = encode_mcu_gather;
126 entropy->pub.finish_pass = finish_pass_gather;
127 #else
128 ERREXIT(cinfo, JERR_NOT_COMPILED);
129 #endif
130 } else {
131 entropy->pub.encode_mcu = encode_mcu_huff;
132 entropy->pub.finish_pass = finish_pass_huff;
135 for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
136 compptr = cinfo->cur_comp_info[ci];
137 dctbl = compptr->dc_tbl_no;
138 actbl = compptr->ac_tbl_no;
139 if (gather_statistics) {
140 #ifdef ENTROPY_OPT_SUPPORTED
141 /* Check for invalid table indexes */
142 /* (make_c_derived_tbl does this in the other path) */
143 if (dctbl < 0 || dctbl >= NUM_HUFF_TBLS)
144 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, dctbl);
145 if (actbl < 0 || actbl >= NUM_HUFF_TBLS)
146 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, actbl);
147 /* Allocate and zero the statistics tables */
148 /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
149 if (entropy->dc_count_ptrs[dctbl] == NULL)
150 entropy->dc_count_ptrs[dctbl] = (long *)
151 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
152 257 * SIZEOF(long));
153 MEMZERO(entropy->dc_count_ptrs[dctbl], 257 * SIZEOF(long));
154 if (entropy->ac_count_ptrs[actbl] == NULL)
155 entropy->ac_count_ptrs[actbl] = (long *)
156 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
157 257 * SIZEOF(long));
158 MEMZERO(entropy->ac_count_ptrs[actbl], 257 * SIZEOF(long));
159 #endif
160 } else {
161 /* Compute derived values for Huffman tables */
162 /* We may do this more than once for a table, but it's not expensive */
163 jpeg_make_c_derived_tbl(cinfo, TRUE, dctbl,
164 & entropy->dc_derived_tbls[dctbl]);
165 jpeg_make_c_derived_tbl(cinfo, FALSE, actbl,
166 & entropy->ac_derived_tbls[actbl]);
168 /* Initialize DC predictions to 0 */
169 entropy->saved.last_dc_val[ci] = 0;
172 /* Initialize bit buffer to empty */
173 entropy->saved.put_buffer = 0;
174 entropy->saved.put_bits = 0;
176 /* Initialize restart stuff */
177 entropy->restarts_to_go = cinfo->restart_interval;
178 entropy->next_restart_num = 0;
183 * Compute the derived values for a Huffman table.
184 * This routine also performs some validation checks on the table.
186 * Note this is also used by jcphuff.c.
189 GLOBAL(void)
190 jpeg_make_c_derived_tbl (j_compress_ptr cinfo, boolean isDC, int tblno,
191 c_derived_tbl ** pdtbl)
193 JHUFF_TBL *htbl;
194 c_derived_tbl *dtbl;
195 int p, i, l, lastp, si, maxsymbol;
196 char huffsize[257];
197 unsigned int huffcode[257];
198 unsigned int code;
200 /* Note that huffsize[] and huffcode[] are filled in code-length order,
201 * paralleling the order of the symbols themselves in htbl->huffval[].
204 /* Find the input Huffman table */
205 if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
206 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
207 htbl =
208 isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
209 if (htbl == NULL)
210 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
212 /* Allocate a workspace if we haven't already done so. */
213 if (*pdtbl == NULL)
214 *pdtbl = (c_derived_tbl *)
215 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
216 SIZEOF(c_derived_tbl));
217 dtbl = *pdtbl;
219 /* Figure C.1: make table of Huffman code length for each symbol */
221 p = 0;
222 for (l = 1; l <= 16; l++) {
223 i = (int) htbl->bits[l];
224 if (i < 0 || p + i > 256) /* protect against table overrun */
225 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
226 while (i--)
227 huffsize[p++] = (char) l;
229 huffsize[p] = 0;
230 lastp = p;
232 /* Figure C.2: generate the codes themselves */
233 /* We also validate that the counts represent a legal Huffman code tree. */
235 code = 0;
236 si = huffsize[0];
237 p = 0;
238 while (huffsize[p]) {
239 while (((int) huffsize[p]) == si) {
240 huffcode[p++] = code;
241 code++;
243 /* code is now 1 more than the last code used for codelength si; but
244 * it must still fit in si bits, since no code is allowed to be all ones.
246 if (((INT32) code) >= (((INT32) 1) << si))
247 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
248 code <<= 1;
249 si++;
252 /* Figure C.3: generate encoding tables */
253 /* These are code and size indexed by symbol value */
255 /* Set all codeless symbols to have code length 0;
256 * this lets us detect duplicate VAL entries here, and later
257 * allows emit_bits to detect any attempt to emit such symbols.
259 MEMZERO(dtbl->ehufsi, SIZEOF(dtbl->ehufsi));
261 /* This is also a convenient place to check for out-of-range
262 * and duplicated VAL entries. We allow 0..255 for AC symbols
263 * but only 0..15 for DC. (We could constrain them further
264 * based on data depth and mode, but this seems enough.)
266 maxsymbol = isDC ? 15 : 255;
268 for (p = 0; p < lastp; p++) {
269 i = htbl->huffval[p];
270 if (i < 0 || i > maxsymbol || dtbl->ehufsi[i])
271 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
272 dtbl->ehufco[i] = huffcode[p];
273 dtbl->ehufsi[i] = huffsize[p];
278 /* Outputting bytes to the file */
280 /* Emit a byte, taking 'action' if must suspend. */
281 #define emit_byte(state,val,action) \
282 { *(state)->next_output_byte++ = (JOCTET) (val); \
283 if (--(state)->free_in_buffer == 0) \
284 if (! dump_buffer(state)) \
285 { action; } }
288 LOCAL(boolean)
289 dump_buffer (working_state * state)
290 /* Empty the output buffer; return TRUE if successful, FALSE if must suspend */
292 struct jpeg_destination_mgr * dest = state->cinfo->dest;
294 dest->free_in_buffer = state->free_in_buffer;
296 if (! (*dest->empty_output_buffer) (state->cinfo))
297 return FALSE;
298 /* After a successful buffer dump, must reset buffer pointers */
299 state->next_output_byte = dest->next_output_byte;
300 state->free_in_buffer = dest->free_in_buffer;
301 return TRUE;
305 /* Outputting bits to the file */
307 /* These macros perform the same task as the emit_bits() function in the
308 * original libjpeg code. In addition to reducing overhead by explicitly
309 * inlining the code, additional performance is achieved by taking into
310 * account the size of the bit buffer and waiting until it is almost full
311 * before emptying it. This mostly benefits 64-bit platforms, since 6
312 * bytes can be stored in a 64-bit bit buffer before it has to be emptied.
315 #define EMIT_BYTE() { \
316 JOCTET c; \
317 put_bits -= 8; \
318 c = (JOCTET)GETJOCTET(put_buffer >> put_bits); \
319 *buffer++ = c; \
320 if (c == 0xFF) /* need to stuff a zero byte? */ \
321 *buffer++ = 0; \
324 #define PUT_BITS(code, size) { \
325 put_bits += size; \
326 put_buffer = (put_buffer << size) | code; \
329 #define CHECKBUF15() { \
330 if (put_bits > 15) { \
331 EMIT_BYTE() \
332 EMIT_BYTE() \
336 #define CHECKBUF31() { \
337 if (put_bits > 31) { \
338 EMIT_BYTE() \
339 EMIT_BYTE() \
340 EMIT_BYTE() \
341 EMIT_BYTE() \
345 #define CHECKBUF47() { \
346 if (put_bits > 47) { \
347 EMIT_BYTE() \
348 EMIT_BYTE() \
349 EMIT_BYTE() \
350 EMIT_BYTE() \
351 EMIT_BYTE() \
352 EMIT_BYTE() \
356 #if __WORDSIZE==64 || defined(_WIN64)
358 #define EMIT_BITS(code, size) { \
359 CHECKBUF47() \
360 PUT_BITS(code, size) \
363 #define EMIT_CODE(code, size) { \
364 temp2 &= (((INT32) 1)<<nbits) - 1; \
365 CHECKBUF31() \
366 PUT_BITS(code, size) \
367 PUT_BITS(temp2, nbits) \
370 #else
372 #define EMIT_BITS(code, size) { \
373 PUT_BITS(code, size) \
374 CHECKBUF15() \
377 #define EMIT_CODE(code, size) { \
378 temp2 &= (((INT32) 1)<<nbits) - 1; \
379 PUT_BITS(code, size) \
380 CHECKBUF15() \
381 PUT_BITS(temp2, nbits) \
382 CHECKBUF15() \
385 #endif
388 #define BUFSIZE (DCTSIZE2 * 2)
390 #define LOAD_BUFFER() { \
391 if (state->free_in_buffer < BUFSIZE) { \
392 localbuf = 1; \
393 buffer = _buffer; \
395 else buffer = state->next_output_byte; \
398 #define STORE_BUFFER() { \
399 if (localbuf) { \
400 bytes = buffer - _buffer; \
401 buffer = _buffer; \
402 while (bytes > 0) { \
403 bytestocopy = min(bytes, state->free_in_buffer); \
404 MEMCOPY(state->next_output_byte, buffer, bytestocopy); \
405 state->next_output_byte += bytestocopy; \
406 buffer += bytestocopy; \
407 state->free_in_buffer -= bytestocopy; \
408 if (state->free_in_buffer == 0) \
409 if (! dump_buffer(state)) return FALSE; \
410 bytes -= bytestocopy; \
413 else { \
414 state->free_in_buffer -= (buffer - state->next_output_byte); \
415 state->next_output_byte = buffer; \
420 LOCAL(boolean)
421 flush_bits (working_state * state)
423 JOCTET _buffer[BUFSIZE], *buffer;
424 size_t put_buffer; int put_bits;
425 size_t bytes, bytestocopy; int localbuf = 0;
427 put_buffer = state->cur.put_buffer;
428 put_bits = state->cur.put_bits;
429 LOAD_BUFFER()
431 /* fill any partial byte with ones */
432 PUT_BITS(0x7F, 7)
433 while (put_bits >= 8) EMIT_BYTE()
435 state->cur.put_buffer = 0; /* and reset bit-buffer to empty */
436 state->cur.put_bits = 0;
437 STORE_BUFFER()
439 return TRUE;
443 /* Encode a single block's worth of coefficients */
445 LOCAL(boolean)
446 encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
447 c_derived_tbl *dctbl, c_derived_tbl *actbl)
449 int temp, temp2, temp3;
450 int nbits;
451 int r, code, size;
452 JOCTET _buffer[BUFSIZE], *buffer;
453 size_t put_buffer; int put_bits;
454 int code_0xf0 = actbl->ehufco[0xf0], size_0xf0 = actbl->ehufsi[0xf0];
455 size_t bytes, bytestocopy; int localbuf = 0;
457 put_buffer = state->cur.put_buffer;
458 put_bits = state->cur.put_bits;
459 LOAD_BUFFER()
461 /* Encode the DC coefficient difference per section F.1.2.1 */
463 temp = temp2 = block[0] - last_dc_val;
465 /* This is a well-known technique for obtaining the absolute value without a
466 * branch. It is derived from an assembly language technique presented in
467 * "How to Optimize for the Pentium Processors", Copyright (c) 1996, 1997 by
468 * Agner Fog.
470 temp3 = temp >> (CHAR_BIT * sizeof(int) - 1);
471 temp ^= temp3;
472 temp -= temp3;
474 /* For a negative input, want temp2 = bitwise complement of abs(input) */
475 /* This code assumes we are on a two's complement machine */
476 temp2 += temp3;
478 /* Find the number of bits needed for the magnitude of the coefficient */
479 nbits = jpeg_nbits_table[temp];
481 /* Emit the Huffman-coded symbol for the number of bits */
482 code = dctbl->ehufco[nbits];
483 size = dctbl->ehufsi[nbits];
484 PUT_BITS(code, size)
485 CHECKBUF15()
487 /* Mask off any extra bits in code */
488 temp2 &= (((INT32) 1)<<nbits) - 1;
490 /* Emit that number of bits of the value, if positive, */
491 /* or the complement of its magnitude, if negative. */
492 PUT_BITS(temp2, nbits)
493 CHECKBUF15()
495 /* Encode the AC coefficients per section F.1.2.2 */
497 r = 0; /* r = run length of zeros */
499 /* Manually unroll the k loop to eliminate the counter variable. This
500 * improves performance greatly on systems with a limited number of
501 * registers (such as x86.)
503 #define kloop(jpeg_natural_order_of_k) { \
504 if ((temp = block[jpeg_natural_order_of_k]) == 0) { \
505 r++; \
506 } else { \
507 temp2 = temp; \
508 /* Branch-less absolute value, bitwise complement, etc., same as above */ \
509 temp3 = temp >> (CHAR_BIT * sizeof(int) - 1); \
510 temp ^= temp3; \
511 temp -= temp3; \
512 temp2 += temp3; \
513 nbits = jpeg_nbits_table[temp]; \
514 /* if run length > 15, must emit special run-length-16 codes (0xF0) */ \
515 while (r > 15) { \
516 EMIT_BITS(code_0xf0, size_0xf0) \
517 r -= 16; \
519 /* Emit Huffman symbol for run length / number of bits */ \
520 temp3 = (r << 4) + nbits; \
521 code = actbl->ehufco[temp3]; \
522 size = actbl->ehufsi[temp3]; \
523 EMIT_CODE(code, size) \
524 r = 0; \
528 /* One iteration for each value in jpeg_natural_order[] */
529 kloop(1); kloop(8); kloop(16); kloop(9); kloop(2); kloop(3);
530 kloop(10); kloop(17); kloop(24); kloop(32); kloop(25); kloop(18);
531 kloop(11); kloop(4); kloop(5); kloop(12); kloop(19); kloop(26);
532 kloop(33); kloop(40); kloop(48); kloop(41); kloop(34); kloop(27);
533 kloop(20); kloop(13); kloop(6); kloop(7); kloop(14); kloop(21);
534 kloop(28); kloop(35); kloop(42); kloop(49); kloop(56); kloop(57);
535 kloop(50); kloop(43); kloop(36); kloop(29); kloop(22); kloop(15);
536 kloop(23); kloop(30); kloop(37); kloop(44); kloop(51); kloop(58);
537 kloop(59); kloop(52); kloop(45); kloop(38); kloop(31); kloop(39);
538 kloop(46); kloop(53); kloop(60); kloop(61); kloop(54); kloop(47);
539 kloop(55); kloop(62); kloop(63);
541 /* If the last coef(s) were zero, emit an end-of-block code */
542 if (r > 0) {
543 code = actbl->ehufco[0];
544 size = actbl->ehufsi[0];
545 EMIT_BITS(code, size)
548 state->cur.put_buffer = put_buffer;
549 state->cur.put_bits = put_bits;
550 STORE_BUFFER()
552 return TRUE;
557 * Emit a restart marker & resynchronize predictions.
560 LOCAL(boolean)
561 emit_restart (working_state * state, int restart_num)
563 int ci;
565 if (! flush_bits(state))
566 return FALSE;
568 emit_byte(state, 0xFF, return FALSE);
569 emit_byte(state, JPEG_RST0 + restart_num, return FALSE);
571 /* Re-initialize DC predictions to 0 */
572 for (ci = 0; ci < state->cinfo->comps_in_scan; ci++)
573 state->cur.last_dc_val[ci] = 0;
575 /* The restart counter is not updated until we successfully write the MCU. */
577 return TRUE;
582 * Encode and output one MCU's worth of Huffman-compressed coefficients.
585 METHODDEF(boolean)
586 encode_mcu_huff (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
588 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
589 working_state state;
590 int blkn, ci;
591 jpeg_component_info * compptr;
593 /* Load up working state */
594 state.next_output_byte = cinfo->dest->next_output_byte;
595 state.free_in_buffer = cinfo->dest->free_in_buffer;
596 ASSIGN_STATE(state.cur, entropy->saved);
597 state.cinfo = cinfo;
599 /* Emit restart marker if needed */
600 if (cinfo->restart_interval) {
601 if (entropy->restarts_to_go == 0)
602 if (! emit_restart(&state, entropy->next_restart_num))
603 return FALSE;
606 /* Encode the MCU data blocks */
607 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
608 ci = cinfo->MCU_membership[blkn];
609 compptr = cinfo->cur_comp_info[ci];
610 if (! encode_one_block(&state,
611 MCU_data[blkn][0], state.cur.last_dc_val[ci],
612 entropy->dc_derived_tbls[compptr->dc_tbl_no],
613 entropy->ac_derived_tbls[compptr->ac_tbl_no]))
614 return FALSE;
615 /* Update last_dc_val */
616 state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
619 /* Completed MCU, so update state */
620 cinfo->dest->next_output_byte = state.next_output_byte;
621 cinfo->dest->free_in_buffer = state.free_in_buffer;
622 ASSIGN_STATE(entropy->saved, state.cur);
624 /* Update restart-interval state too */
625 if (cinfo->restart_interval) {
626 if (entropy->restarts_to_go == 0) {
627 entropy->restarts_to_go = cinfo->restart_interval;
628 entropy->next_restart_num++;
629 entropy->next_restart_num &= 7;
631 entropy->restarts_to_go--;
634 return TRUE;
639 * Finish up at the end of a Huffman-compressed scan.
642 METHODDEF(void)
643 finish_pass_huff (j_compress_ptr cinfo)
645 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
646 working_state state;
648 /* Load up working state ... flush_bits needs it */
649 state.next_output_byte = cinfo->dest->next_output_byte;
650 state.free_in_buffer = cinfo->dest->free_in_buffer;
651 ASSIGN_STATE(state.cur, entropy->saved);
652 state.cinfo = cinfo;
654 /* Flush out the last data */
655 if (! flush_bits(&state))
656 ERREXIT(cinfo, JERR_CANT_SUSPEND);
658 /* Update state */
659 cinfo->dest->next_output_byte = state.next_output_byte;
660 cinfo->dest->free_in_buffer = state.free_in_buffer;
661 ASSIGN_STATE(entropy->saved, state.cur);
666 * Huffman coding optimization.
668 * We first scan the supplied data and count the number of uses of each symbol
669 * that is to be Huffman-coded. (This process MUST agree with the code above.)
670 * Then we build a Huffman coding tree for the observed counts.
671 * Symbols which are not needed at all for the particular image are not
672 * assigned any code, which saves space in the DHT marker as well as in
673 * the compressed data.
676 #ifdef ENTROPY_OPT_SUPPORTED
679 /* Process a single block's worth of coefficients */
681 LOCAL(void)
682 htest_one_block (j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val,
683 long dc_counts[], long ac_counts[])
685 register int temp;
686 register int nbits;
687 register int k, r;
689 /* Encode the DC coefficient difference per section F.1.2.1 */
691 temp = block[0] - last_dc_val;
692 if (temp < 0)
693 temp = -temp;
695 /* Find the number of bits needed for the magnitude of the coefficient */
696 nbits = 0;
697 while (temp) {
698 nbits++;
699 temp >>= 1;
701 /* Check for out-of-range coefficient values.
702 * Since we're encoding a difference, the range limit is twice as much.
704 if (nbits > MAX_COEF_BITS+1)
705 ERREXIT(cinfo, JERR_BAD_DCT_COEF);
707 /* Count the Huffman symbol for the number of bits */
708 dc_counts[nbits]++;
710 /* Encode the AC coefficients per section F.1.2.2 */
712 r = 0; /* r = run length of zeros */
714 for (k = 1; k < DCTSIZE2; k++) {
715 if ((temp = block[jpeg_natural_order[k]]) == 0) {
716 r++;
717 } else {
718 /* if run length > 15, must emit special run-length-16 codes (0xF0) */
719 while (r > 15) {
720 ac_counts[0xF0]++;
721 r -= 16;
724 /* Find the number of bits needed for the magnitude of the coefficient */
725 if (temp < 0)
726 temp = -temp;
728 /* Find the number of bits needed for the magnitude of the coefficient */
729 nbits = 1; /* there must be at least one 1 bit */
730 while ((temp >>= 1))
731 nbits++;
732 /* Check for out-of-range coefficient values */
733 if (nbits > MAX_COEF_BITS)
734 ERREXIT(cinfo, JERR_BAD_DCT_COEF);
736 /* Count Huffman symbol for run length / number of bits */
737 ac_counts[(r << 4) + nbits]++;
739 r = 0;
743 /* If the last coef(s) were zero, emit an end-of-block code */
744 if (r > 0)
745 ac_counts[0]++;
750 * Trial-encode one MCU's worth of Huffman-compressed coefficients.
751 * No data is actually output, so no suspension return is possible.
754 METHODDEF(boolean)
755 encode_mcu_gather (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
757 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
758 int blkn, ci;
759 jpeg_component_info * compptr;
761 /* Take care of restart intervals if needed */
762 if (cinfo->restart_interval) {
763 if (entropy->restarts_to_go == 0) {
764 /* Re-initialize DC predictions to 0 */
765 for (ci = 0; ci < cinfo->comps_in_scan; ci++)
766 entropy->saved.last_dc_val[ci] = 0;
767 /* Update restart state */
768 entropy->restarts_to_go = cinfo->restart_interval;
770 entropy->restarts_to_go--;
773 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
774 ci = cinfo->MCU_membership[blkn];
775 compptr = cinfo->cur_comp_info[ci];
776 htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci],
777 entropy->dc_count_ptrs[compptr->dc_tbl_no],
778 entropy->ac_count_ptrs[compptr->ac_tbl_no]);
779 entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0];
782 return TRUE;
787 * Generate the best Huffman code table for the given counts, fill htbl.
788 * Note this is also used by jcphuff.c.
790 * The JPEG standard requires that no symbol be assigned a codeword of all
791 * one bits (so that padding bits added at the end of a compressed segment
792 * can't look like a valid code). Because of the canonical ordering of
793 * codewords, this just means that there must be an unused slot in the
794 * longest codeword length category. Section K.2 of the JPEG spec suggests
795 * reserving such a slot by pretending that symbol 256 is a valid symbol
796 * with count 1. In theory that's not optimal; giving it count zero but
797 * including it in the symbol set anyway should give a better Huffman code.
798 * But the theoretically better code actually seems to come out worse in
799 * practice, because it produces more all-ones bytes (which incur stuffed
800 * zero bytes in the final file). In any case the difference is tiny.
802 * The JPEG standard requires Huffman codes to be no more than 16 bits long.
803 * If some symbols have a very small but nonzero probability, the Huffman tree
804 * must be adjusted to meet the code length restriction. We currently use
805 * the adjustment method suggested in JPEG section K.2. This method is *not*
806 * optimal; it may not choose the best possible limited-length code. But
807 * typically only very-low-frequency symbols will be given less-than-optimal
808 * lengths, so the code is almost optimal. Experimental comparisons against
809 * an optimal limited-length-code algorithm indicate that the difference is
810 * microscopic --- usually less than a hundredth of a percent of total size.
811 * So the extra complexity of an optimal algorithm doesn't seem worthwhile.
814 GLOBAL(void)
815 jpeg_gen_optimal_table (j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[])
817 #define MAX_CLEN 32 /* assumed maximum initial code length */
818 UINT8 bits[MAX_CLEN+1]; /* bits[k] = # of symbols with code length k */
819 int codesize[257]; /* codesize[k] = code length of symbol k */
820 int others[257]; /* next symbol in current branch of tree */
821 int c1, c2;
822 int p, i, j;
823 long v;
825 /* This algorithm is explained in section K.2 of the JPEG standard */
827 MEMZERO(bits, SIZEOF(bits));
828 MEMZERO(codesize, SIZEOF(codesize));
829 for (i = 0; i < 257; i++)
830 others[i] = -1; /* init links to empty */
832 freq[256] = 1; /* make sure 256 has a nonzero count */
833 /* Including the pseudo-symbol 256 in the Huffman procedure guarantees
834 * that no real symbol is given code-value of all ones, because 256
835 * will be placed last in the largest codeword category.
838 /* Huffman's basic algorithm to assign optimal code lengths to symbols */
840 for (;;) {
841 /* Find the smallest nonzero frequency, set c1 = its symbol */
842 /* In case of ties, take the larger symbol number */
843 c1 = -1;
844 v = 1000000000L;
845 for (i = 0; i <= 256; i++) {
846 if (freq[i] && freq[i] <= v) {
847 v = freq[i];
848 c1 = i;
852 /* Find the next smallest nonzero frequency, set c2 = its symbol */
853 /* In case of ties, take the larger symbol number */
854 c2 = -1;
855 v = 1000000000L;
856 for (i = 0; i <= 256; i++) {
857 if (freq[i] && freq[i] <= v && i != c1) {
858 v = freq[i];
859 c2 = i;
863 /* Done if we've merged everything into one frequency */
864 if (c2 < 0)
865 break;
867 /* Else merge the two counts/trees */
868 freq[c1] += freq[c2];
869 freq[c2] = 0;
871 /* Increment the codesize of everything in c1's tree branch */
872 codesize[c1]++;
873 while (others[c1] >= 0) {
874 c1 = others[c1];
875 codesize[c1]++;
878 others[c1] = c2; /* chain c2 onto c1's tree branch */
880 /* Increment the codesize of everything in c2's tree branch */
881 codesize[c2]++;
882 while (others[c2] >= 0) {
883 c2 = others[c2];
884 codesize[c2]++;
888 /* Now count the number of symbols of each code length */
889 for (i = 0; i <= 256; i++) {
890 if (codesize[i]) {
891 /* The JPEG standard seems to think that this can't happen, */
892 /* but I'm paranoid... */
893 if (codesize[i] > MAX_CLEN)
894 ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW);
896 bits[codesize[i]]++;
900 /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure
901 * Huffman procedure assigned any such lengths, we must adjust the coding.
902 * Here is what the JPEG spec says about how this next bit works:
903 * Since symbols are paired for the longest Huffman code, the symbols are
904 * removed from this length category two at a time. The prefix for the pair
905 * (which is one bit shorter) is allocated to one of the pair; then,
906 * skipping the BITS entry for that prefix length, a code word from the next
907 * shortest nonzero BITS entry is converted into a prefix for two code words
908 * one bit longer.
911 for (i = MAX_CLEN; i > 16; i--) {
912 while (bits[i] > 0) {
913 j = i - 2; /* find length of new prefix to be used */
914 while (bits[j] == 0)
915 j--;
917 bits[i] -= 2; /* remove two symbols */
918 bits[i-1]++; /* one goes in this length */
919 bits[j+1] += 2; /* two new symbols in this length */
920 bits[j]--; /* symbol of this length is now a prefix */
924 /* Remove the count for the pseudo-symbol 256 from the largest codelength */
925 while (bits[i] == 0) /* find largest codelength still in use */
926 i--;
927 bits[i]--;
929 /* Return final symbol counts (only for lengths 0..16) */
930 MEMCOPY(htbl->bits, bits, SIZEOF(htbl->bits));
932 /* Return a list of the symbols sorted by code length */
933 /* It's not real clear to me why we don't need to consider the codelength
934 * changes made above, but the JPEG spec seems to think this works.
936 p = 0;
937 for (i = 1; i <= MAX_CLEN; i++) {
938 for (j = 0; j <= 255; j++) {
939 if (codesize[j] == i) {
940 htbl->huffval[p] = (UINT8) j;
941 p++;
946 /* Set sent_table FALSE so updated table will be written to JPEG file. */
947 htbl->sent_table = FALSE;
952 * Finish up a statistics-gathering pass and create the new Huffman tables.
955 METHODDEF(void)
956 finish_pass_gather (j_compress_ptr cinfo)
958 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
959 int ci, dctbl, actbl;
960 jpeg_component_info * compptr;
961 JHUFF_TBL **htblptr;
962 boolean did_dc[NUM_HUFF_TBLS];
963 boolean did_ac[NUM_HUFF_TBLS];
965 /* It's important not to apply jpeg_gen_optimal_table more than once
966 * per table, because it clobbers the input frequency counts!
968 MEMZERO(did_dc, SIZEOF(did_dc));
969 MEMZERO(did_ac, SIZEOF(did_ac));
971 for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
972 compptr = cinfo->cur_comp_info[ci];
973 dctbl = compptr->dc_tbl_no;
974 actbl = compptr->ac_tbl_no;
975 if (! did_dc[dctbl]) {
976 htblptr = & cinfo->dc_huff_tbl_ptrs[dctbl];
977 if (*htblptr == NULL)
978 *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
979 jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]);
980 did_dc[dctbl] = TRUE;
982 if (! did_ac[actbl]) {
983 htblptr = & cinfo->ac_huff_tbl_ptrs[actbl];
984 if (*htblptr == NULL)
985 *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
986 jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]);
987 did_ac[actbl] = TRUE;
993 #endif /* ENTROPY_OPT_SUPPORTED */
997 * Module initialization routine for Huffman entropy encoding.
1000 GLOBAL(void)
1001 jinit_huff_encoder (j_compress_ptr cinfo)
1003 huff_entropy_ptr entropy;
1004 int i;
1006 entropy = (huff_entropy_ptr)
1007 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
1008 SIZEOF(huff_entropy_encoder));
1009 cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
1010 entropy->pub.start_pass = start_pass_huff;
1012 /* Mark tables unallocated */
1013 for (i = 0; i < NUM_HUFF_TBLS; i++) {
1014 entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
1015 #ifdef ENTROPY_OPT_SUPPORTED
1016 entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL;
1017 #endif