1 /*
2 * jchuff.c
3 *
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.
7 *
8 * This file contains Huffman entropy encoding routines.
9 *
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.
15 */
16
17 #define JPEG_INTERNALS
18 #include "jinclude.h"
19 #include "jpeglib.h"
20 #include "jchuff.h" /* Declarations shared with jcphuff.c */
21
22 #ifdef _FX_MANAGED_CODE_
23 #define savable_state savable_state_c
24 #endif
25
26 /* Expanded entropy encoder object for Huffman encoding.
27 *
28 * The savable_state subrecord contains fields that change within an MCU,
29 * but must not be updated permanently until we complete the MCU.
30 */
31
32 typedef struct {
33 INT32 put_buffer; /* current bit-accumulation buffer */
34 int put_bits; /* # of bits now in it */
35 int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
36 } savable_state;
37
38 /* This macro is to work around compilers with missing or broken
39 * structure assignment. You'll need to fix this code if you have
40 * such a compiler and you change MAX_COMPS_IN_SCAN.
41 */
42
43 #ifndef NO_STRUCT_ASSIGN
44 #define ASSIGN_STATE(dest,src) ((dest) = (src))
45 #else
46 #if MAX_COMPS_IN_SCAN == 4
47 #define ASSIGN_STATE(dest,src) \
48 ((dest).put_buffer = (src).put_buffer, \
49 (dest).put_bits = (src).put_bits, \
50 (dest).last_dc_val[0] = (src).last_dc_val[0], \
51 (dest).last_dc_val[1] = (src).last_dc_val[1], \
52 (dest).last_dc_val[2] = (src).last_dc_val[2], \
53 (dest).last_dc_val[3] = (src).last_dc_val[3])
54 #endif
55 #endif
56
57
58 typedef struct {
59 struct jpeg_entropy_encoder pub; /* public fields */
60
61 savable_state saved; /* Bit buffer & DC state at start of MCU */
62
63 /* These fields are NOT loaded into local working state. */
64 unsigned int restarts_to_go; /* MCUs left in this restart interval */
65 int next_restart_num; /* next restart number to write (0-7) */
66
67 /* Pointers to derived tables (these workspaces have image lifespan) */
68 c_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
69 c_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
70
71 #ifdef ENTROPY_OPT_SUPPORTED /* Statistics tables for optimization */
72 long * dc_count_ptrs[NUM_HUFF_TBLS];
73 long * ac_count_ptrs[NUM_HUFF_TBLS];
74 #endif
75 } huff_entropy_encoder;
76
77 typedef huff_entropy_encoder * huff_entropy_ptr;
78
79 /* Working state while writing an MCU.
80 * This struct contains all the fields that are needed by subroutines.
81 */
82
83 typedef struct {
84 JOCTET * next_output_byte; /* => next byte to write in buffer */
85 size_t free_in_buffer; /* # of byte spaces remaining in buffer */
86 savable_state cur; /* Current bit buffer & DC state */
87 j_compress_ptr cinfo; /* dump_buffer needs access to this */
88 } working_state;
89
90
91 /* Forward declarations */
92 METHODDEF(boolean) encode_mcu_huff JPP((j_compress_ptr cinfo,
93 JBLOCKROW *MCU_data));
94 METHODDEF(void) finish_pass_huff JPP((j_compress_ptr cinfo));
95 #ifdef ENTROPY_OPT_SUPPORTED
96 METHODDEF(boolean) encode_mcu_gather JPP((j_compress_ptr cinfo,
97 JBLOCKROW *MCU_data));
98 METHODDEF(void) finish_pass_gather JPP((j_compress_ptr cinfo));
99 #endif
100
101
102 /*
103 * Initialize for a Huffman-compressed scan.
104 * If gather_statistics is TRUE, we do not output anything during the scan,
105 * just count the Huffman symbols used and generate Huffman code tables.
106 */
107
108 METHODDEF(void)
start_pass_huff(j_compress_ptr cinfo,boolean gather_statistics)109 start_pass_huff (j_compress_ptr cinfo, boolean gather_statistics)
110 {
111 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
112 int ci, dctbl, actbl;
113 jpeg_component_info * compptr;
114
115 if (gather_statistics) {
116 #ifdef ENTROPY_OPT_SUPPORTED
117 entropy->pub.encode_mcu = encode_mcu_gather;
118 entropy->pub.finish_pass = finish_pass_gather;
119 #else
120 ERREXIT(cinfo, JERR_NOT_COMPILED);
121 #endif
122 } else {
123 entropy->pub.encode_mcu = encode_mcu_huff;
124 entropy->pub.finish_pass = finish_pass_huff;
125 }
126
127 for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
128 compptr = cinfo->cur_comp_info[ci];
129 dctbl = compptr->dc_tbl_no;
130 actbl = compptr->ac_tbl_no;
131 if (gather_statistics) {
132 #ifdef ENTROPY_OPT_SUPPORTED
133 /* Check for invalid table indexes */
134 /* (make_c_derived_tbl does this in the other path) */
135 if (dctbl < 0 || dctbl >= NUM_HUFF_TBLS)
136 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, dctbl);
137 if (actbl < 0 || actbl >= NUM_HUFF_TBLS)
138 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, actbl);
139 /* Allocate and zero the statistics tables */
140 /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
141 if (entropy->dc_count_ptrs[dctbl] == NULL)
142 entropy->dc_count_ptrs[dctbl] = (long *)
143 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
144 257 * SIZEOF(long));
145 MEMZERO(entropy->dc_count_ptrs[dctbl], 257 * SIZEOF(long));
146 if (entropy->ac_count_ptrs[actbl] == NULL)
147 entropy->ac_count_ptrs[actbl] = (long *)
148 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
149 257 * SIZEOF(long));
150 MEMZERO(entropy->ac_count_ptrs[actbl], 257 * SIZEOF(long));
151 #endif
152 } else {
153 /* Compute derived values for Huffman tables */
154 /* We may do this more than once for a table, but it's not expensive */
155 jpeg_make_c_derived_tbl(cinfo, TRUE, dctbl,
156 & entropy->dc_derived_tbls[dctbl]);
157 jpeg_make_c_derived_tbl(cinfo, FALSE, actbl,
158 & entropy->ac_derived_tbls[actbl]);
159 }
160 /* Initialize DC predictions to 0 */
161 entropy->saved.last_dc_val[ci] = 0;
162 }
163
164 /* Initialize bit buffer to empty */
165 entropy->saved.put_buffer = 0;
166 entropy->saved.put_bits = 0;
167
168 /* Initialize restart stuff */
169 entropy->restarts_to_go = cinfo->restart_interval;
170 entropy->next_restart_num = 0;
171 }
172
173
174 /*
175 * Compute the derived values for a Huffman table.
176 * This routine also performs some validation checks on the table.
177 *
178 * Note this is also used by jcphuff.c.
179 */
180
181 GLOBAL(void)
jpeg_make_c_derived_tbl(j_compress_ptr cinfo,boolean isDC,int tblno,c_derived_tbl ** pdtbl)182 jpeg_make_c_derived_tbl (j_compress_ptr cinfo, boolean isDC, int tblno,
183 c_derived_tbl ** pdtbl)
184 {
185 JHUFF_TBL *htbl;
186 c_derived_tbl *dtbl;
187 int p, i, l, lastp, _si, maxsymbol;
188 char huffsize[257];
189 unsigned int huffcode[257];
190 unsigned int code;
191
192 /* Note that huffsize[] and huffcode[] are filled in code-length order,
193 * paralleling the order of the symbols themselves in htbl->huffval[].
194 */
195
196 /* Find the input Huffman table */
197 if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
198 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
199 htbl =
200 isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
201 if (htbl == NULL)
202 ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
203
204 /* Allocate a workspace if we haven't already done so. */
205 if (*pdtbl == NULL)
206 *pdtbl = (c_derived_tbl *)
207 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
208 SIZEOF(c_derived_tbl));
209 dtbl = *pdtbl;
210
211 /* Figure C.1: make table of Huffman code length for each symbol */
212
213 p = 0;
214 for (l = 1; l <= 16; l++) {
215 i = (int) htbl->bits[l];
216 if (i < 0 || p + i > 256) /* protect against table overrun */
217 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
218 while (i--)
219 huffsize[p++] = (char) l;
220 }
221 huffsize[p] = 0;
222 lastp = p;
223
224 /* Figure C.2: generate the codes themselves */
225 /* We also validate that the counts represent a legal Huffman code tree. */
226
227 code = 0;
228 _si = huffsize[0];
229 p = 0;
230 while (huffsize[p]) {
231 while (((int) huffsize[p]) == _si) {
232 huffcode[p++] = code;
233 code++;
234 }
235 /* code is now 1 more than the last code used for codelength si; but
236 * it must still fit in si bits, since no code is allowed to be all ones.
237 */
238 if (((INT32) code) >= (((INT32) 1) << _si))
239 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
240 code <<= 1;
241 _si++;
242 }
243
244 /* Figure C.3: generate encoding tables */
245 /* These are code and size indexed by symbol value */
246
247 /* Set all codeless symbols to have code length 0;
248 * this lets us detect duplicate VAL entries here, and later
249 * allows emit_bits to detect any attempt to emit such symbols.
250 */
251 MEMZERO(dtbl->ehufsi, SIZEOF(dtbl->ehufsi));
252
253 /* This is also a convenient place to check for out-of-range
254 * and duplicated VAL entries. We allow 0..255 for AC symbols
255 * but only 0..15 for DC. (We could constrain them further
256 * based on data depth and mode, but this seems enough.)
257 */
258 maxsymbol = isDC ? 15 : 255;
259
260 for (p = 0; p < lastp; p++) {
261 i = htbl->huffval[p];
262 if (i < 0 || i > maxsymbol || dtbl->ehufsi[i])
263 ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
264 dtbl->ehufco[i] = huffcode[p];
265 dtbl->ehufsi[i] = huffsize[p];
266 }
267 }
268
269
270 /* Outputting bytes to the file */
271
272 /* Emit a byte, taking 'action' if must suspend. */
273 #define emit_byte(state,val,action) \
274 { *(state)->next_output_byte++ = (JOCTET) (val); \
275 if (--(state)->free_in_buffer == 0) \
276 if (! dump_buffer(state)) \
277 { action; } }
278
279
280 LOCAL(boolean)
dump_buffer(working_state * state)281 dump_buffer (working_state * state)
282 /* Empty the output buffer; return TRUE if successful, FALSE if must suspend */
283 {
284 struct jpeg_destination_mgr * dest = state->cinfo->dest;
285
286 if (! (*dest->empty_output_buffer) (state->cinfo))
287 return FALSE;
288 /* After a successful buffer dump, must reset buffer pointers */
289 state->next_output_byte = dest->next_output_byte;
290 state->free_in_buffer = dest->free_in_buffer;
291 return TRUE;
292 }
293
294
295 /* Outputting bits to the file */
296
297 /* Only the right 24 bits of put_buffer are used; the valid bits are
298 * left-justified in this part. At most 16 bits can be passed to emit_bits
299 * in one call, and we never retain more than 7 bits in put_buffer
300 * between calls, so 24 bits are sufficient.
301 */
302
303 INLINE
LOCAL(boolean)304 LOCAL(boolean)
305 emit_bits (working_state * state, unsigned int code, int size)
306 /* Emit some bits; return TRUE if successful, FALSE if must suspend */
307 {
308 /* This routine is heavily used, so it's worth coding tightly. */
309 register INT32 put_buffer = (INT32) code;
310 register int put_bits = state->cur.put_bits;
311
312 /* if size is 0, caller used an invalid Huffman table entry */
313 if (size == 0)
314 ERREXIT(state->cinfo, JERR_HUFF_MISSING_CODE);
315
316 put_buffer &= (((INT32) 1)<<size) - 1; /* mask off any extra bits in code */
317
318 put_bits += size; /* new number of bits in buffer */
319
320 put_buffer <<= 24 - put_bits; /* align incoming bits */
321
322 put_buffer |= state->cur.put_buffer; /* and merge with old buffer contents */
323
324 while (put_bits >= 8) {
325 int c = (int) ((put_buffer >> 16) & 0xFF);
326
327 emit_byte(state, c, return FALSE);
328 if (c == 0xFF) { /* need to stuff a zero byte? */
329 emit_byte(state, 0, return FALSE);
330 }
331 put_buffer <<= 8;
332 put_bits -= 8;
333 }
334
335 state->cur.put_buffer = put_buffer; /* update state variables */
336 state->cur.put_bits = put_bits;
337
338 return TRUE;
339 }
340
341
342 LOCAL(boolean)
flush_bits(working_state * state)343 flush_bits (working_state * state)
344 {
345 if (! emit_bits(state, 0x7F, 7)) /* fill any partial byte with ones */
346 return FALSE;
347 state->cur.put_buffer = 0; /* and reset bit-buffer to empty */
348 state->cur.put_bits = 0;
349 return TRUE;
350 }
351
352
353 /* Encode a single block's worth of coefficients */
354
355 LOCAL(boolean)
encode_one_block(working_state * state,JCOEFPTR block,int last_dc_val,c_derived_tbl * dctbl,c_derived_tbl * actbl)356 encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
357 c_derived_tbl *dctbl, c_derived_tbl *actbl)
358 {
359 register int temp, temp2;
360 register int nbits;
361 register int k, r, i;
362
363 /* Encode the DC coefficient difference per section F.1.2.1 */
364
365 temp = temp2 = block[0] - last_dc_val;
366
367 if (temp < 0) {
368 temp = -temp; /* temp is abs value of input */
369 /* For a negative input, want temp2 = bitwise complement of abs(input) */
370 /* This code assumes we are on a two's complement machine */
371 temp2--;
372 }
373
374 /* Find the number of bits needed for the magnitude of the coefficient */
375 nbits = 0;
376 while (temp) {
377 nbits++;
378 temp >>= 1;
379 }
380 /* Check for out-of-range coefficient values.
381 * Since we're encoding a difference, the range limit is twice as much.
382 */
383 if (nbits > MAX_COEF_BITS+1)
384 ERREXIT(state->cinfo, JERR_BAD_DCT_COEF);
385
386 /* Emit the Huffman-coded symbol for the number of bits */
387 if (! emit_bits(state, dctbl->ehufco[nbits], dctbl->ehufsi[nbits]))
388 return FALSE;
389
390 /* Emit that number of bits of the value, if positive, */
391 /* or the complement of its magnitude, if negative. */
392 if (nbits) /* emit_bits rejects calls with size 0 */
393 if (! emit_bits(state, (unsigned int) temp2, nbits))
394 return FALSE;
395
396 /* Encode the AC coefficients per section F.1.2.2 */
397
398 r = 0; /* r = run length of zeros */
399
400 for (k = 1; k < DCTSIZE2; k++) {
401 if ((temp = block[jpeg_natural_order[k]]) == 0) {
402 r++;
403 } else {
404 /* if run length > 15, must emit special run-length-16 codes (0xF0) */
405 while (r > 15) {
406 if (! emit_bits(state, actbl->ehufco[0xF0], actbl->ehufsi[0xF0]))
407 return FALSE;
408 r -= 16;
409 }
410
411 temp2 = temp;
412 if (temp < 0) {
413 temp = -temp; /* temp is abs value of input */
414 /* This code assumes we are on a two's complement machine */
415 temp2--;
416 }
417
418 /* Find the number of bits needed for the magnitude of the coefficient */
419 nbits = 1; /* there must be at least one 1 bit */
420 while ((temp >>= 1))
421 nbits++;
422 /* Check for out-of-range coefficient values */
423 if (nbits > MAX_COEF_BITS)
424 ERREXIT(state->cinfo, JERR_BAD_DCT_COEF);
425
426 /* Emit Huffman symbol for run length / number of bits */
427 i = (r << 4) + nbits;
428 if (! emit_bits(state, actbl->ehufco[i], actbl->ehufsi[i]))
429 return FALSE;
430
431 /* Emit that number of bits of the value, if positive, */
432 /* or the complement of its magnitude, if negative. */
433 if (! emit_bits(state, (unsigned int) temp2, nbits))
434 return FALSE;
435
436 r = 0;
437 }
438 }
439
440 /* If the last coef(s) were zero, emit an end-of-block code */
441 if (r > 0)
442 if (! emit_bits(state, actbl->ehufco[0], actbl->ehufsi[0]))
443 return FALSE;
444
445 return TRUE;
446 }
447
448
449 /*
450 * Emit a restart marker & resynchronize predictions.
451 */
452
453 LOCAL(boolean)
emit_restart(working_state * state,int restart_num)454 emit_restart (working_state * state, int restart_num)
455 {
456 int ci;
457
458 if (! flush_bits(state))
459 return FALSE;
460
461 emit_byte(state, 0xFF, return FALSE);
462 emit_byte(state, JPEG_RST0 + restart_num, return FALSE);
463
464 /* Re-initialize DC predictions to 0 */
465 for (ci = 0; ci < state->cinfo->comps_in_scan; ci++)
466 state->cur.last_dc_val[ci] = 0;
467
468 /* The restart counter is not updated until we successfully write the MCU. */
469
470 return TRUE;
471 }
472
473
474 /*
475 * Encode and output one MCU's worth of Huffman-compressed coefficients.
476 */
477
478 METHODDEF(boolean)
encode_mcu_huff(j_compress_ptr cinfo,JBLOCKROW * MCU_data)479 encode_mcu_huff (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
480 {
481 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
482 working_state state;
483 int blkn, ci;
484 jpeg_component_info * compptr;
485
486 /* Load up working state */
487 state.next_output_byte = cinfo->dest->next_output_byte;
488 state.free_in_buffer = cinfo->dest->free_in_buffer;
489 ASSIGN_STATE(state.cur, entropy->saved);
490 state.cinfo = cinfo;
491
492 /* Emit restart marker if needed */
493 if (cinfo->restart_interval) {
494 if (entropy->restarts_to_go == 0)
495 if (! emit_restart(&state, entropy->next_restart_num))
496 return FALSE;
497 }
498
499 /* Encode the MCU data blocks */
500 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
501 ci = cinfo->MCU_membership[blkn];
502 compptr = cinfo->cur_comp_info[ci];
503 if (! encode_one_block(&state,
504 MCU_data[blkn][0], state.cur.last_dc_val[ci],
505 entropy->dc_derived_tbls[compptr->dc_tbl_no],
506 entropy->ac_derived_tbls[compptr->ac_tbl_no]))
507 return FALSE;
508 /* Update last_dc_val */
509 state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
510 }
511
512 /* Completed MCU, so update state */
513 cinfo->dest->next_output_byte = state.next_output_byte;
514 cinfo->dest->free_in_buffer = state.free_in_buffer;
515 ASSIGN_STATE(entropy->saved, state.cur);
516
517 /* Update restart-interval state too */
518 if (cinfo->restart_interval) {
519 if (entropy->restarts_to_go == 0) {
520 entropy->restarts_to_go = cinfo->restart_interval;
521 entropy->next_restart_num++;
522 entropy->next_restart_num &= 7;
523 }
524 entropy->restarts_to_go--;
525 }
526
527 return TRUE;
528 }
529
530
531 /*
532 * Finish up at the end of a Huffman-compressed scan.
533 */
534
535 METHODDEF(void)
finish_pass_huff(j_compress_ptr cinfo)536 finish_pass_huff (j_compress_ptr cinfo)
537 {
538 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
539 working_state state;
540
541 /* Load up working state ... flush_bits needs it */
542 state.next_output_byte = cinfo->dest->next_output_byte;
543 state.free_in_buffer = cinfo->dest->free_in_buffer;
544 ASSIGN_STATE(state.cur, entropy->saved);
545 state.cinfo = cinfo;
546
547 /* Flush out the last data */
548 if (! flush_bits(&state))
549 ERREXIT(cinfo, JERR_CANT_SUSPEND);
550
551 /* Update state */
552 cinfo->dest->next_output_byte = state.next_output_byte;
553 cinfo->dest->free_in_buffer = state.free_in_buffer;
554 ASSIGN_STATE(entropy->saved, state.cur);
555 }
556
557
558 /*
559 * Huffman coding optimization.
560 *
561 * We first scan the supplied data and count the number of uses of each symbol
562 * that is to be Huffman-coded. (This process MUST agree with the code above.)
563 * Then we build a Huffman coding tree for the observed counts.
564 * Symbols which are not needed at all for the particular image are not
565 * assigned any code, which saves space in the DHT marker as well as in
566 * the compressed data.
567 */
568
569 #ifdef ENTROPY_OPT_SUPPORTED
570
571
572 /* Process a single block's worth of coefficients */
573
574 LOCAL(void)
htest_one_block(j_compress_ptr cinfo,JCOEFPTR block,int last_dc_val,long dc_counts[],long ac_counts[])575 htest_one_block (j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val,
576 long dc_counts[], long ac_counts[])
577 {
578 register int temp;
579 register int nbits;
580 register int k, r;
581
582 /* Encode the DC coefficient difference per section F.1.2.1 */
583
584 temp = block[0] - last_dc_val;
585 if (temp < 0)
586 temp = -temp;
587
588 /* Find the number of bits needed for the magnitude of the coefficient */
589 nbits = 0;
590 while (temp) {
591 nbits++;
592 temp >>= 1;
593 }
594 /* Check for out-of-range coefficient values.
595 * Since we're encoding a difference, the range limit is twice as much.
596 */
597 if (nbits > MAX_COEF_BITS+1)
598 ERREXIT(cinfo, JERR_BAD_DCT_COEF);
599
600 /* Count the Huffman symbol for the number of bits */
601 dc_counts[nbits]++;
602
603 /* Encode the AC coefficients per section F.1.2.2 */
604
605 r = 0; /* r = run length of zeros */
606
607 for (k = 1; k < DCTSIZE2; k++) {
608 if ((temp = block[jpeg_natural_order[k]]) == 0) {
609 r++;
610 } else {
611 /* if run length > 15, must emit special run-length-16 codes (0xF0) */
612 while (r > 15) {
613 ac_counts[0xF0]++;
614 r -= 16;
615 }
616
617 /* Find the number of bits needed for the magnitude of the coefficient */
618 if (temp < 0)
619 temp = -temp;
620
621 /* Find the number of bits needed for the magnitude of the coefficient */
622 nbits = 1; /* there must be at least one 1 bit */
623 while ((temp >>= 1))
624 nbits++;
625 /* Check for out-of-range coefficient values */
626 if (nbits > MAX_COEF_BITS)
627 ERREXIT(cinfo, JERR_BAD_DCT_COEF);
628
629 /* Count Huffman symbol for run length / number of bits */
630 ac_counts[(r << 4) + nbits]++;
631
632 r = 0;
633 }
634 }
635
636 /* If the last coef(s) were zero, emit an end-of-block code */
637 if (r > 0)
638 ac_counts[0]++;
639 }
640
641
642 /*
643 * Trial-encode one MCU's worth of Huffman-compressed coefficients.
644 * No data is actually output, so no suspension return is possible.
645 */
646
647 METHODDEF(boolean)
encode_mcu_gather(j_compress_ptr cinfo,JBLOCKROW * MCU_data)648 encode_mcu_gather (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
649 {
650 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
651 int blkn, ci;
652 jpeg_component_info * compptr;
653
654 /* Take care of restart intervals if needed */
655 if (cinfo->restart_interval) {
656 if (entropy->restarts_to_go == 0) {
657 /* Re-initialize DC predictions to 0 */
658 for (ci = 0; ci < cinfo->comps_in_scan; ci++)
659 entropy->saved.last_dc_val[ci] = 0;
660 /* Update restart state */
661 entropy->restarts_to_go = cinfo->restart_interval;
662 }
663 entropy->restarts_to_go--;
664 }
665
666 for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
667 ci = cinfo->MCU_membership[blkn];
668 compptr = cinfo->cur_comp_info[ci];
669 htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci],
670 entropy->dc_count_ptrs[compptr->dc_tbl_no],
671 entropy->ac_count_ptrs[compptr->ac_tbl_no]);
672 entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0];
673 }
674
675 return TRUE;
676 }
677
678
679 /*
680 * Generate the best Huffman code table for the given counts, fill htbl.
681 * Note this is also used by jcphuff.c.
682 *
683 * The JPEG standard requires that no symbol be assigned a codeword of all
684 * one bits (so that padding bits added at the end of a compressed segment
685 * can't look like a valid code). Because of the canonical ordering of
686 * codewords, this just means that there must be an unused slot in the
687 * longest codeword length category. Section K.2 of the JPEG spec suggests
688 * reserving such a slot by pretending that symbol 256 is a valid symbol
689 * with count 1. In theory that's not optimal; giving it count zero but
690 * including it in the symbol set anyway should give a better Huffman code.
691 * But the theoretically better code actually seems to come out worse in
692 * practice, because it produces more all-ones bytes (which incur stuffed
693 * zero bytes in the final file). In any case the difference is tiny.
694 *
695 * The JPEG standard requires Huffman codes to be no more than 16 bits long.
696 * If some symbols have a very small but nonzero probability, the Huffman tree
697 * must be adjusted to meet the code length restriction. We currently use
698 * the adjustment method suggested in JPEG section K.2. This method is *not*
699 * optimal; it may not choose the best possible limited-length code. But
700 * typically only very-low-frequency symbols will be given less-than-optimal
701 * lengths, so the code is almost optimal. Experimental comparisons against
702 * an optimal limited-length-code algorithm indicate that the difference is
703 * microscopic --- usually less than a hundredth of a percent of total size.
704 * So the extra complexity of an optimal algorithm doesn't seem worthwhile.
705 */
706
707 GLOBAL(void)
jpeg_gen_optimal_table(j_compress_ptr cinfo,JHUFF_TBL * htbl,long freq[])708 jpeg_gen_optimal_table (j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[])
709 {
710 #define MAX_CLEN 32 /* assumed maximum initial code length */
711 UINT8 bits[MAX_CLEN+1]; /* bits[k] = # of symbols with code length k */
712 int codesize[257]; /* codesize[k] = code length of symbol k */
713 int others[257]; /* next symbol in current branch of tree */
714 int c1, c2;
715 int p, i, j;
716 long v;
717
718 /* This algorithm is explained in section K.2 of the JPEG standard */
719
720 MEMZERO(bits, SIZEOF(bits));
721 MEMZERO(codesize, SIZEOF(codesize));
722 for (i = 0; i < 257; i++)
723 others[i] = -1; /* init links to empty */
724
725 freq[256] = 1; /* make sure 256 has a nonzero count */
726 /* Including the pseudo-symbol 256 in the Huffman procedure guarantees
727 * that no real symbol is given code-value of all ones, because 256
728 * will be placed last in the largest codeword category.
729 */
730
731 /* Huffman's basic algorithm to assign optimal code lengths to symbols */
732
733 for (;;) {
734 /* Find the smallest nonzero frequency, set c1 = its symbol */
735 /* In case of ties, take the larger symbol number */
736 c1 = -1;
737 v = 1000000000L;
738 for (i = 0; i <= 256; i++) {
739 if (freq[i] && freq[i] <= v) {
740 v = freq[i];
741 c1 = i;
742 }
743 }
744
745 /* Find the next smallest nonzero frequency, set c2 = its symbol */
746 /* In case of ties, take the larger symbol number */
747 c2 = -1;
748 v = 1000000000L;
749 for (i = 0; i <= 256; i++) {
750 if (freq[i] && freq[i] <= v && i != c1) {
751 v = freq[i];
752 c2 = i;
753 }
754 }
755
756 /* Done if we've merged everything into one frequency */
757 if (c2 < 0)
758 break;
759
760 /* Else merge the two counts/trees */
761 freq[c1] += freq[c2];
762 freq[c2] = 0;
763
764 /* Increment the codesize of everything in c1's tree branch */
765 codesize[c1]++;
766 while (others[c1] >= 0) {
767 c1 = others[c1];
768 codesize[c1]++;
769 }
770
771 others[c1] = c2; /* chain c2 onto c1's tree branch */
772
773 /* Increment the codesize of everything in c2's tree branch */
774 codesize[c2]++;
775 while (others[c2] >= 0) {
776 c2 = others[c2];
777 codesize[c2]++;
778 }
779 }
780
781 /* Now count the number of symbols of each code length */
782 for (i = 0; i <= 256; i++) {
783 if (codesize[i]) {
784 /* The JPEG standard seems to think that this can't happen, */
785 /* but I'm paranoid... */
786 if (codesize[i] > MAX_CLEN)
787 ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW);
788
789 bits[codesize[i]]++;
790 }
791 }
792
793 /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure
794 * Huffman procedure assigned any such lengths, we must adjust the coding.
795 * Here is what the JPEG spec says about how this next bit works:
796 * Since symbols are paired for the longest Huffman code, the symbols are
797 * removed from this length category two at a time. The prefix for the pair
798 * (which is one bit shorter) is allocated to one of the pair; then,
799 * skipping the BITS entry for that prefix length, a code word from the next
800 * shortest nonzero BITS entry is converted into a prefix for two code words
801 * one bit longer.
802 */
803
804 for (i = MAX_CLEN; i > 16; i--) {
805 while (bits[i] > 0) {
806 j = i - 2; /* find length of new prefix to be used */
807 while (bits[j] == 0)
808 j--;
809
810 bits[i] -= 2; /* remove two symbols */
811 bits[i-1]++; /* one goes in this length */
812 bits[j+1] += 2; /* two new symbols in this length */
813 bits[j]--; /* symbol of this length is now a prefix */
814 }
815 }
816
817 /* Remove the count for the pseudo-symbol 256 from the largest codelength */
818 while (bits[i] == 0) /* find largest codelength still in use */
819 i--;
820 bits[i]--;
821
822 /* Return final symbol counts (only for lengths 0..16) */
823 MEMCOPY(htbl->bits, bits, SIZEOF(htbl->bits));
824
825 /* Return a list of the symbols sorted by code length */
826 /* It's not real clear to me why we don't need to consider the codelength
827 * changes made above, but the JPEG spec seems to think this works.
828 */
829 p = 0;
830 for (i = 1; i <= MAX_CLEN; i++) {
831 for (j = 0; j <= 255; j++) {
832 if (codesize[j] == i) {
833 htbl->huffval[p] = (UINT8) j;
834 p++;
835 }
836 }
837 }
838
839 /* Set sent_table FALSE so updated table will be written to JPEG file. */
840 htbl->sent_table = FALSE;
841 }
842
843
844 /*
845 * Finish up a statistics-gathering pass and create the new Huffman tables.
846 */
847
848 METHODDEF(void)
finish_pass_gather(j_compress_ptr cinfo)849 finish_pass_gather (j_compress_ptr cinfo)
850 {
851 huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
852 int ci, dctbl, actbl;
853 jpeg_component_info * compptr;
854 JHUFF_TBL **htblptr;
855 boolean did_dc[NUM_HUFF_TBLS];
856 boolean did_ac[NUM_HUFF_TBLS];
857
858 /* It's important not to apply jpeg_gen_optimal_table more than once
859 * per table, because it clobbers the input frequency counts!
860 */
861 MEMZERO(did_dc, SIZEOF(did_dc));
862 MEMZERO(did_ac, SIZEOF(did_ac));
863
864 for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
865 compptr = cinfo->cur_comp_info[ci];
866 dctbl = compptr->dc_tbl_no;
867 actbl = compptr->ac_tbl_no;
868 if (! did_dc[dctbl]) {
869 htblptr = & cinfo->dc_huff_tbl_ptrs[dctbl];
870 if (*htblptr == NULL)
871 *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
872 jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]);
873 did_dc[dctbl] = TRUE;
874 }
875 if (! did_ac[actbl]) {
876 htblptr = & cinfo->ac_huff_tbl_ptrs[actbl];
877 if (*htblptr == NULL)
878 *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
879 jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]);
880 did_ac[actbl] = TRUE;
881 }
882 }
883 }
884
885
886 #endif /* ENTROPY_OPT_SUPPORTED */
887
888
889 /*
890 * Module initialization routine for Huffman entropy encoding.
891 */
892
893 GLOBAL(void)
jinit_huff_encoder(j_compress_ptr cinfo)894 jinit_huff_encoder (j_compress_ptr cinfo)
895 {
896 huff_entropy_ptr entropy;
897 int i;
898
899 entropy = (huff_entropy_ptr)
900 (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
901 SIZEOF(huff_entropy_encoder));
902 cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
903 entropy->pub.start_pass = start_pass_huff;
904
905 /* Mark tables unallocated */
906 for (i = 0; i < NUM_HUFF_TBLS; i++) {
907 entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
908 #ifdef ENTROPY_OPT_SUPPORTED
909 entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL;
910 #endif
911 }
912 }
913