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