1 | /* |
2 | * jchuff.c |
3 | * |
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, 2014-2016, 2018-2021, D. R. Commander. |
8 | * Copyright (C) 2015, Matthieu Darbois. |
9 | * Copyright (C) 2018, Matthias Räncker. |
10 | * Copyright (C) 2020, Arm Limited. |
11 | * For conditions of distribution and use, see the accompanying README.ijg |
12 | * file. |
13 | * |
14 | * This file contains Huffman entropy encoding routines. |
15 | * |
16 | * Much of the complexity here has to do with supporting output suspension. |
17 | * If the data destination module demands suspension, we want to be able to |
18 | * back up to the start of the current MCU. To do this, we copy state |
19 | * variables into local working storage, and update them back to the |
20 | * permanent JPEG objects only upon successful completion of an MCU. |
21 | * |
22 | * NOTE: All referenced figures are from |
23 | * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994. |
24 | */ |
25 | |
26 | #define JPEG_INTERNALS |
27 | #include "jinclude.h" |
28 | #include "jpeglib.h" |
29 | #include "jsimd.h" |
30 | #include "jconfigint.h" |
31 | #include <limits.h> |
32 | |
33 | /* |
34 | * NOTE: If USE_CLZ_INTRINSIC is defined, then clz/bsr instructions will be |
35 | * used for bit counting rather than the lookup table. This will reduce the |
36 | * memory footprint by 64k, which is important for some mobile applications |
37 | * that create many isolated instances of libjpeg-turbo (web browsers, for |
38 | * instance.) This may improve performance on some mobile platforms as well. |
39 | * This feature is enabled by default only on Arm processors, because some x86 |
40 | * chips have a slow implementation of bsr, and the use of clz/bsr cannot be |
41 | * shown to have a significant performance impact even on the x86 chips that |
42 | * have a fast implementation of it. When building for Armv6, you can |
43 | * explicitly disable the use of clz/bsr by adding -mthumb to the compiler |
44 | * flags (this defines __thumb__). |
45 | */ |
46 | |
47 | #if defined(__arm__) || defined(__aarch64__) || defined(_M_ARM) || \ |
48 | defined(_M_ARM64) |
49 | #if !defined(__thumb__) || defined(__thumb2__) |
50 | #define USE_CLZ_INTRINSIC |
51 | #endif |
52 | #endif |
53 | |
54 | #ifdef USE_CLZ_INTRINSIC |
55 | #if defined(_MSC_VER) && !defined(__clang__) |
56 | #define JPEG_NBITS_NONZERO(x) (32 - _CountLeadingZeros(x)) |
57 | #else |
58 | #define JPEG_NBITS_NONZERO(x) (32 - __builtin_clz(x)) |
59 | #endif |
60 | #define JPEG_NBITS(x) (x ? JPEG_NBITS_NONZERO(x) : 0) |
61 | #else |
62 | #include "jpeg_nbits_table.h" |
63 | #define JPEG_NBITS(x) (jpeg_nbits_table[x]) |
64 | #define JPEG_NBITS_NONZERO(x) JPEG_NBITS(x) |
65 | #endif |
66 | |
67 | |
68 | /* Expanded entropy encoder object for Huffman encoding. |
69 | * |
70 | * The savable_state subrecord contains fields that change within an MCU, |
71 | * but must not be updated permanently until we complete the MCU. |
72 | */ |
73 | |
74 | #if defined(__x86_64__) && defined(__ILP32__) |
75 | typedef unsigned long long bit_buf_type; |
76 | #else |
77 | typedef size_t bit_buf_type; |
78 | #endif |
79 | |
80 | /* NOTE: The more optimal Huffman encoding algorithm is only used by the |
81 | * intrinsics implementation of the Arm Neon SIMD extensions, which is why we |
82 | * retain the old Huffman encoder behavior when using the GAS implementation. |
83 | */ |
84 | #if defined(WITH_SIMD) && !(defined(__arm__) || defined(__aarch64__) || \ |
85 | defined(_M_ARM) || defined(_M_ARM64)) |
86 | typedef unsigned long long simd_bit_buf_type; |
87 | #else |
88 | typedef bit_buf_type simd_bit_buf_type; |
89 | #endif |
90 | |
91 | #if (defined(SIZEOF_SIZE_T) && SIZEOF_SIZE_T == 8) || defined(_WIN64) || \ |
92 | (defined(__x86_64__) && defined(__ILP32__)) |
93 | #define BIT_BUF_SIZE 64 |
94 | #elif (defined(SIZEOF_SIZE_T) && SIZEOF_SIZE_T == 4) || defined(_WIN32) |
95 | #define BIT_BUF_SIZE 32 |
96 | #else |
97 | #error Cannot determine word size |
98 | #endif |
99 | #define SIMD_BIT_BUF_SIZE (sizeof(simd_bit_buf_type) * 8) |
100 | |
101 | typedef struct { |
102 | union { |
103 | bit_buf_type c; |
104 | simd_bit_buf_type simd; |
105 | } put_buffer; /* current bit accumulation buffer */ |
106 | int free_bits; /* # of bits available in it */ |
107 | /* (Neon GAS: # of bits now in it) */ |
108 | int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */ |
109 | } savable_state; |
110 | |
111 | typedef struct { |
112 | struct jpeg_entropy_encoder pub; /* public fields */ |
113 | |
114 | savable_state saved; /* Bit buffer & DC state at start of MCU */ |
115 | |
116 | /* These fields are NOT loaded into local working state. */ |
117 | unsigned int restarts_to_go; /* MCUs left in this restart interval */ |
118 | int next_restart_num; /* next restart number to write (0-7) */ |
119 | |
120 | /* Pointers to derived tables (these workspaces have image lifespan) */ |
121 | c_derived_tbl *dc_derived_tbls[NUM_HUFF_TBLS]; |
122 | c_derived_tbl *ac_derived_tbls[NUM_HUFF_TBLS]; |
123 | |
124 | #ifdef ENTROPY_OPT_SUPPORTED /* Statistics tables for optimization */ |
125 | long *dc_count_ptrs[NUM_HUFF_TBLS]; |
126 | long *ac_count_ptrs[NUM_HUFF_TBLS]; |
127 | #endif |
128 | |
129 | int simd; |
130 | } huff_entropy_encoder; |
131 | |
132 | typedef huff_entropy_encoder *huff_entropy_ptr; |
133 | |
134 | /* Working state while writing an MCU. |
135 | * This struct contains all the fields that are needed by subroutines. |
136 | */ |
137 | |
138 | typedef struct { |
139 | JOCTET *next_output_byte; /* => next byte to write in buffer */ |
140 | size_t free_in_buffer; /* # of byte spaces remaining in buffer */ |
141 | savable_state cur; /* Current bit buffer & DC state */ |
142 | j_compress_ptr cinfo; /* dump_buffer needs access to this */ |
143 | int simd; |
144 | } working_state; |
145 | |
146 | |
147 | /* Forward declarations */ |
148 | METHODDEF(boolean) encode_mcu_huff(j_compress_ptr cinfo, JBLOCKROW *MCU_data); |
149 | METHODDEF(void) finish_pass_huff(j_compress_ptr cinfo); |
150 | #ifdef ENTROPY_OPT_SUPPORTED |
151 | METHODDEF(boolean) encode_mcu_gather(j_compress_ptr cinfo, |
152 | JBLOCKROW *MCU_data); |
153 | METHODDEF(void) finish_pass_gather(j_compress_ptr cinfo); |
154 | #endif |
155 | |
156 | |
157 | /* |
158 | * Initialize for a Huffman-compressed scan. |
159 | * If gather_statistics is TRUE, we do not output anything during the scan, |
160 | * just count the Huffman symbols used and generate Huffman code tables. |
161 | */ |
162 | |
163 | METHODDEF(void) |
164 | start_pass_huff(j_compress_ptr cinfo, boolean gather_statistics) |
165 | { |
166 | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
167 | int ci, dctbl, actbl; |
168 | jpeg_component_info *compptr; |
169 | |
170 | if (gather_statistics) { |
171 | #ifdef ENTROPY_OPT_SUPPORTED |
172 | entropy->pub.encode_mcu = encode_mcu_gather; |
173 | entropy->pub.finish_pass = finish_pass_gather; |
174 | #else |
175 | ERREXIT(cinfo, JERR_NOT_COMPILED); |
176 | #endif |
177 | } else { |
178 | entropy->pub.encode_mcu = encode_mcu_huff; |
179 | entropy->pub.finish_pass = finish_pass_huff; |
180 | } |
181 | |
182 | entropy->simd = jsimd_can_huff_encode_one_block(); |
183 | |
184 | for (ci = 0; ci < cinfo->comps_in_scan; ci++) { |
185 | compptr = cinfo->cur_comp_info[ci]; |
186 | dctbl = compptr->dc_tbl_no; |
187 | actbl = compptr->ac_tbl_no; |
188 | if (gather_statistics) { |
189 | #ifdef ENTROPY_OPT_SUPPORTED |
190 | /* Check for invalid table indexes */ |
191 | /* (make_c_derived_tbl does this in the other path) */ |
192 | if (dctbl < 0 || dctbl >= NUM_HUFF_TBLS) |
193 | ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, dctbl); |
194 | if (actbl < 0 || actbl >= NUM_HUFF_TBLS) |
195 | ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, actbl); |
196 | /* Allocate and zero the statistics tables */ |
197 | /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */ |
198 | if (entropy->dc_count_ptrs[dctbl] == NULL) |
199 | entropy->dc_count_ptrs[dctbl] = (long *) |
200 | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
201 | 257 * sizeof(long)); |
202 | MEMZERO(entropy->dc_count_ptrs[dctbl], 257 * sizeof(long)); |
203 | if (entropy->ac_count_ptrs[actbl] == NULL) |
204 | entropy->ac_count_ptrs[actbl] = (long *) |
205 | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
206 | 257 * sizeof(long)); |
207 | MEMZERO(entropy->ac_count_ptrs[actbl], 257 * sizeof(long)); |
208 | #endif |
209 | } else { |
210 | /* Compute derived values for Huffman tables */ |
211 | /* We may do this more than once for a table, but it's not expensive */ |
212 | jpeg_make_c_derived_tbl(cinfo, TRUE, dctbl, |
213 | &entropy->dc_derived_tbls[dctbl]); |
214 | jpeg_make_c_derived_tbl(cinfo, FALSE, actbl, |
215 | &entropy->ac_derived_tbls[actbl]); |
216 | } |
217 | /* Initialize DC predictions to 0 */ |
218 | entropy->saved.last_dc_val[ci] = 0; |
219 | } |
220 | |
221 | /* Initialize bit buffer to empty */ |
222 | if (entropy->simd) { |
223 | entropy->saved.put_buffer.simd = 0; |
224 | #if defined(__aarch64__) && !defined(NEON_INTRINSICS) |
225 | entropy->saved.free_bits = 0; |
226 | #else |
227 | entropy->saved.free_bits = SIMD_BIT_BUF_SIZE; |
228 | #endif |
229 | } else { |
230 | entropy->saved.put_buffer.c = 0; |
231 | entropy->saved.free_bits = BIT_BUF_SIZE; |
232 | } |
233 | |
234 | /* Initialize restart stuff */ |
235 | entropy->restarts_to_go = cinfo->restart_interval; |
236 | entropy->next_restart_num = 0; |
237 | } |
238 | |
239 | |
240 | /* |
241 | * Compute the derived values for a Huffman table. |
242 | * This routine also performs some validation checks on the table. |
243 | * |
244 | * Note this is also used by jcphuff.c. |
245 | */ |
246 | |
247 | GLOBAL(void) |
248 | jpeg_make_c_derived_tbl(j_compress_ptr cinfo, boolean isDC, int tblno, |
249 | c_derived_tbl **pdtbl) |
250 | { |
251 | JHUFF_TBL *htbl; |
252 | c_derived_tbl *dtbl; |
253 | int p, i, l, lastp, si, maxsymbol; |
254 | char huffsize[257]; |
255 | unsigned int huffcode[257]; |
256 | unsigned int code; |
257 | |
258 | /* Note that huffsize[] and huffcode[] are filled in code-length order, |
259 | * paralleling the order of the symbols themselves in htbl->huffval[]. |
260 | */ |
261 | |
262 | /* Find the input Huffman table */ |
263 | if (tblno < 0 || tblno >= NUM_HUFF_TBLS) |
264 | ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno); |
265 | htbl = |
266 | isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno]; |
267 | if (htbl == NULL) |
268 | ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno); |
269 | |
270 | /* Allocate a workspace if we haven't already done so. */ |
271 | if (*pdtbl == NULL) |
272 | *pdtbl = (c_derived_tbl *) |
273 | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
274 | sizeof(c_derived_tbl)); |
275 | dtbl = *pdtbl; |
276 | |
277 | /* Figure C.1: make table of Huffman code length for each symbol */ |
278 | |
279 | p = 0; |
280 | for (l = 1; l <= 16; l++) { |
281 | i = (int)htbl->bits[l]; |
282 | if (i < 0 || p + i > 256) /* protect against table overrun */ |
283 | ERREXIT(cinfo, JERR_BAD_HUFF_TABLE); |
284 | while (i--) |
285 | huffsize[p++] = (char)l; |
286 | } |
287 | huffsize[p] = 0; |
288 | lastp = p; |
289 | |
290 | /* Figure C.2: generate the codes themselves */ |
291 | /* We also validate that the counts represent a legal Huffman code tree. */ |
292 | |
293 | code = 0; |
294 | si = huffsize[0]; |
295 | p = 0; |
296 | while (huffsize[p]) { |
297 | while (((int)huffsize[p]) == si) { |
298 | huffcode[p++] = code; |
299 | code++; |
300 | } |
301 | /* code is now 1 more than the last code used for codelength si; but |
302 | * it must still fit in si bits, since no code is allowed to be all ones. |
303 | */ |
304 | if (((JLONG)code) >= (((JLONG)1) << si)) |
305 | ERREXIT(cinfo, JERR_BAD_HUFF_TABLE); |
306 | code <<= 1; |
307 | si++; |
308 | } |
309 | |
310 | /* Figure C.3: generate encoding tables */ |
311 | /* These are code and size indexed by symbol value */ |
312 | |
313 | /* Set all codeless symbols to have code length 0; |
314 | * this lets us detect duplicate VAL entries here, and later |
315 | * allows emit_bits to detect any attempt to emit such symbols. |
316 | */ |
317 | MEMZERO(dtbl->ehufco, sizeof(dtbl->ehufco)); |
318 | MEMZERO(dtbl->ehufsi, sizeof(dtbl->ehufsi)); |
319 | |
320 | /* This is also a convenient place to check for out-of-range |
321 | * and duplicated VAL entries. We allow 0..255 for AC symbols |
322 | * but only 0..15 for DC. (We could constrain them further |
323 | * based on data depth and mode, but this seems enough.) |
324 | */ |
325 | maxsymbol = isDC ? 15 : 255; |
326 | |
327 | for (p = 0; p < lastp; p++) { |
328 | i = htbl->huffval[p]; |
329 | if (i < 0 || i > maxsymbol || dtbl->ehufsi[i]) |
330 | ERREXIT(cinfo, JERR_BAD_HUFF_TABLE); |
331 | dtbl->ehufco[i] = huffcode[p]; |
332 | dtbl->ehufsi[i] = huffsize[p]; |
333 | } |
334 | } |
335 | |
336 | |
337 | /* Outputting bytes to the file */ |
338 | |
339 | /* Emit a byte, taking 'action' if must suspend. */ |
340 | #define emit_byte(state, val, action) { \ |
341 | *(state)->next_output_byte++ = (JOCTET)(val); \ |
342 | if (--(state)->free_in_buffer == 0) \ |
343 | if (!dump_buffer(state)) \ |
344 | { action; } \ |
345 | } |
346 | |
347 | |
348 | LOCAL(boolean) |
349 | dump_buffer(working_state *state) |
350 | /* Empty the output buffer; return TRUE if successful, FALSE if must suspend */ |
351 | { |
352 | struct jpeg_destination_mgr *dest = state->cinfo->dest; |
353 | |
354 | if (!(*dest->empty_output_buffer) (state->cinfo)) |
355 | return FALSE; |
356 | /* After a successful buffer dump, must reset buffer pointers */ |
357 | state->next_output_byte = dest->next_output_byte; |
358 | state->free_in_buffer = dest->free_in_buffer; |
359 | return TRUE; |
360 | } |
361 | |
362 | |
363 | /* Outputting bits to the file */ |
364 | |
365 | /* Output byte b and, speculatively, an additional 0 byte. 0xFF must be |
366 | * encoded as 0xFF 0x00, so the output buffer pointer is advanced by 2 if the |
367 | * byte is 0xFF. Otherwise, the output buffer pointer is advanced by 1, and |
368 | * the speculative 0 byte will be overwritten by the next byte. |
369 | */ |
370 | #define EMIT_BYTE(b) { \ |
371 | buffer[0] = (JOCTET)(b); \ |
372 | buffer[1] = 0; \ |
373 | buffer -= -2 + ((JOCTET)(b) < 0xFF); \ |
374 | } |
375 | |
376 | /* Output the entire bit buffer. If there are no 0xFF bytes in it, then write |
377 | * directly to the output buffer. Otherwise, use the EMIT_BYTE() macro to |
378 | * encode 0xFF as 0xFF 0x00. |
379 | */ |
380 | #if BIT_BUF_SIZE == 64 |
381 | |
382 | #define FLUSH() { \ |
383 | if (put_buffer & 0x8080808080808080 & ~(put_buffer + 0x0101010101010101)) { \ |
384 | EMIT_BYTE(put_buffer >> 56) \ |
385 | EMIT_BYTE(put_buffer >> 48) \ |
386 | EMIT_BYTE(put_buffer >> 40) \ |
387 | EMIT_BYTE(put_buffer >> 32) \ |
388 | EMIT_BYTE(put_buffer >> 24) \ |
389 | EMIT_BYTE(put_buffer >> 16) \ |
390 | EMIT_BYTE(put_buffer >> 8) \ |
391 | EMIT_BYTE(put_buffer ) \ |
392 | } else { \ |
393 | buffer[0] = (JOCTET)(put_buffer >> 56); \ |
394 | buffer[1] = (JOCTET)(put_buffer >> 48); \ |
395 | buffer[2] = (JOCTET)(put_buffer >> 40); \ |
396 | buffer[3] = (JOCTET)(put_buffer >> 32); \ |
397 | buffer[4] = (JOCTET)(put_buffer >> 24); \ |
398 | buffer[5] = (JOCTET)(put_buffer >> 16); \ |
399 | buffer[6] = (JOCTET)(put_buffer >> 8); \ |
400 | buffer[7] = (JOCTET)(put_buffer); \ |
401 | buffer += 8; \ |
402 | } \ |
403 | } |
404 | |
405 | #else |
406 | |
407 | #define FLUSH() { \ |
408 | if (put_buffer & 0x80808080 & ~(put_buffer + 0x01010101)) { \ |
409 | EMIT_BYTE(put_buffer >> 24) \ |
410 | EMIT_BYTE(put_buffer >> 16) \ |
411 | EMIT_BYTE(put_buffer >> 8) \ |
412 | EMIT_BYTE(put_buffer ) \ |
413 | } else { \ |
414 | buffer[0] = (JOCTET)(put_buffer >> 24); \ |
415 | buffer[1] = (JOCTET)(put_buffer >> 16); \ |
416 | buffer[2] = (JOCTET)(put_buffer >> 8); \ |
417 | buffer[3] = (JOCTET)(put_buffer); \ |
418 | buffer += 4; \ |
419 | } \ |
420 | } |
421 | |
422 | #endif |
423 | |
424 | /* Fill the bit buffer to capacity with the leading bits from code, then output |
425 | * the bit buffer and put the remaining bits from code into the bit buffer. |
426 | */ |
427 | #define PUT_AND_FLUSH(code, size) { \ |
428 | put_buffer = (put_buffer << (size + free_bits)) | (code >> -free_bits); \ |
429 | FLUSH() \ |
430 | free_bits += BIT_BUF_SIZE; \ |
431 | put_buffer = code; \ |
432 | } |
433 | |
434 | /* Insert code into the bit buffer and output the bit buffer if needed. |
435 | * NOTE: We can't flush with free_bits == 0, since the left shift in |
436 | * PUT_AND_FLUSH() would have undefined behavior. |
437 | */ |
438 | #define PUT_BITS(code, size) { \ |
439 | free_bits -= size; \ |
440 | if (free_bits < 0) \ |
441 | PUT_AND_FLUSH(code, size) \ |
442 | else \ |
443 | put_buffer = (put_buffer << size) | code; \ |
444 | } |
445 | |
446 | #define PUT_CODE(code, size) { \ |
447 | temp &= (((JLONG)1) << nbits) - 1; \ |
448 | temp |= code << nbits; \ |
449 | nbits += size; \ |
450 | PUT_BITS(temp, nbits) \ |
451 | } |
452 | |
453 | |
454 | /* Although it is exceedingly rare, it is possible for a Huffman-encoded |
455 | * coefficient block to be larger than the 128-byte unencoded block. For each |
456 | * of the 64 coefficients, PUT_BITS is invoked twice, and each invocation can |
457 | * theoretically store 16 bits (for a maximum of 2048 bits or 256 bytes per |
458 | * encoded block.) If, for instance, one artificially sets the AC |
459 | * coefficients to alternating values of 32767 and -32768 (using the JPEG |
460 | * scanning order-- 1, 8, 16, etc.), then this will produce an encoded block |
461 | * larger than 200 bytes. |
462 | */ |
463 | #define BUFSIZE (DCTSIZE2 * 8) |
464 | |
465 | #define LOAD_BUFFER() { \ |
466 | if (state->free_in_buffer < BUFSIZE) { \ |
467 | localbuf = 1; \ |
468 | buffer = _buffer; \ |
469 | } else \ |
470 | buffer = state->next_output_byte; \ |
471 | } |
472 | |
473 | #define STORE_BUFFER() { \ |
474 | if (localbuf) { \ |
475 | size_t bytes, bytestocopy; \ |
476 | bytes = buffer - _buffer; \ |
477 | buffer = _buffer; \ |
478 | while (bytes > 0) { \ |
479 | bytestocopy = MIN(bytes, state->free_in_buffer); \ |
480 | MEMCOPY(state->next_output_byte, buffer, bytestocopy); \ |
481 | state->next_output_byte += bytestocopy; \ |
482 | buffer += bytestocopy; \ |
483 | state->free_in_buffer -= bytestocopy; \ |
484 | if (state->free_in_buffer == 0) \ |
485 | if (!dump_buffer(state)) return FALSE; \ |
486 | bytes -= bytestocopy; \ |
487 | } \ |
488 | } else { \ |
489 | state->free_in_buffer -= (buffer - state->next_output_byte); \ |
490 | state->next_output_byte = buffer; \ |
491 | } \ |
492 | } |
493 | |
494 | |
495 | LOCAL(boolean) |
496 | flush_bits(working_state *state) |
497 | { |
498 | JOCTET _buffer[BUFSIZE], *buffer, temp; |
499 | simd_bit_buf_type put_buffer; int put_bits; |
500 | int localbuf = 0; |
501 | |
502 | if (state->simd) { |
503 | #if defined(__aarch64__) && !defined(NEON_INTRINSICS) |
504 | put_bits = state->cur.free_bits; |
505 | #else |
506 | put_bits = SIMD_BIT_BUF_SIZE - state->cur.free_bits; |
507 | #endif |
508 | put_buffer = state->cur.put_buffer.simd; |
509 | } else { |
510 | put_bits = BIT_BUF_SIZE - state->cur.free_bits; |
511 | put_buffer = state->cur.put_buffer.c; |
512 | } |
513 | |
514 | LOAD_BUFFER() |
515 | |
516 | while (put_bits >= 8) { |
517 | put_bits -= 8; |
518 | temp = (JOCTET)(put_buffer >> put_bits); |
519 | EMIT_BYTE(temp) |
520 | } |
521 | if (put_bits) { |
522 | /* fill partial byte with ones */ |
523 | temp = (JOCTET)((put_buffer << (8 - put_bits)) | (0xFF >> put_bits)); |
524 | EMIT_BYTE(temp) |
525 | } |
526 | |
527 | if (state->simd) { /* and reset bit buffer to empty */ |
528 | state->cur.put_buffer.simd = 0; |
529 | #if defined(__aarch64__) && !defined(NEON_INTRINSICS) |
530 | state->cur.free_bits = 0; |
531 | #else |
532 | state->cur.free_bits = SIMD_BIT_BUF_SIZE; |
533 | #endif |
534 | } else { |
535 | state->cur.put_buffer.c = 0; |
536 | state->cur.free_bits = BIT_BUF_SIZE; |
537 | } |
538 | STORE_BUFFER() |
539 | |
540 | return TRUE; |
541 | } |
542 | |
543 | |
544 | /* Encode a single block's worth of coefficients */ |
545 | |
546 | LOCAL(boolean) |
547 | encode_one_block_simd(working_state *state, JCOEFPTR block, int last_dc_val, |
548 | c_derived_tbl *dctbl, c_derived_tbl *actbl) |
549 | { |
550 | JOCTET _buffer[BUFSIZE], *buffer; |
551 | int localbuf = 0; |
552 | |
553 | LOAD_BUFFER() |
554 | |
555 | buffer = jsimd_huff_encode_one_block(state, buffer, block, last_dc_val, |
556 | dctbl, actbl); |
557 | |
558 | STORE_BUFFER() |
559 | |
560 | return TRUE; |
561 | } |
562 | |
563 | LOCAL(boolean) |
564 | encode_one_block(working_state *state, JCOEFPTR block, int last_dc_val, |
565 | c_derived_tbl *dctbl, c_derived_tbl *actbl) |
566 | { |
567 | int temp, nbits, free_bits; |
568 | bit_buf_type put_buffer; |
569 | JOCTET _buffer[BUFSIZE], *buffer; |
570 | int localbuf = 0; |
571 | |
572 | free_bits = state->cur.free_bits; |
573 | put_buffer = state->cur.put_buffer.c; |
574 | LOAD_BUFFER() |
575 | |
576 | /* Encode the DC coefficient difference per section F.1.2.1 */ |
577 | |
578 | temp = block[0] - last_dc_val; |
579 | |
580 | /* This is a well-known technique for obtaining the absolute value without a |
581 | * branch. It is derived from an assembly language technique presented in |
582 | * "How to Optimize for the Pentium Processors", Copyright (c) 1996, 1997 by |
583 | * Agner Fog. This code assumes we are on a two's complement machine. |
584 | */ |
585 | nbits = temp >> (CHAR_BIT * sizeof(int) - 1); |
586 | temp += nbits; |
587 | nbits ^= temp; |
588 | |
589 | /* Find the number of bits needed for the magnitude of the coefficient */ |
590 | nbits = JPEG_NBITS(nbits); |
591 | |
592 | /* Emit the Huffman-coded symbol for the number of bits. |
593 | * Emit that number of bits of the value, if positive, |
594 | * or the complement of its magnitude, if negative. |
595 | */ |
596 | PUT_CODE(dctbl->ehufco[nbits], dctbl->ehufsi[nbits]) |
597 | |
598 | /* Encode the AC coefficients per section F.1.2.2 */ |
599 | |
600 | { |
601 | int r = 0; /* r = run length of zeros */ |
602 | |
603 | /* Manually unroll the k loop to eliminate the counter variable. This |
604 | * improves performance greatly on systems with a limited number of |
605 | * registers (such as x86.) |
606 | */ |
607 | #define kloop(jpeg_natural_order_of_k) { \ |
608 | if ((temp = block[jpeg_natural_order_of_k]) == 0) { \ |
609 | r += 16; \ |
610 | } else { \ |
611 | /* Branch-less absolute value, bitwise complement, etc., same as above */ \ |
612 | nbits = temp >> (CHAR_BIT * sizeof(int) - 1); \ |
613 | temp += nbits; \ |
614 | nbits ^= temp; \ |
615 | nbits = JPEG_NBITS_NONZERO(nbits); \ |
616 | /* if run length > 15, must emit special run-length-16 codes (0xF0) */ \ |
617 | while (r >= 16 * 16) { \ |
618 | r -= 16 * 16; \ |
619 | PUT_BITS(actbl->ehufco[0xf0], actbl->ehufsi[0xf0]) \ |
620 | } \ |
621 | /* Emit Huffman symbol for run length / number of bits */ \ |
622 | r += nbits; \ |
623 | PUT_CODE(actbl->ehufco[r], actbl->ehufsi[r]) \ |
624 | r = 0; \ |
625 | } \ |
626 | } |
627 | |
628 | /* One iteration for each value in jpeg_natural_order[] */ |
629 | kloop(1); kloop(8); kloop(16); kloop(9); kloop(2); kloop(3); |
630 | kloop(10); kloop(17); kloop(24); kloop(32); kloop(25); kloop(18); |
631 | kloop(11); kloop(4); kloop(5); kloop(12); kloop(19); kloop(26); |
632 | kloop(33); kloop(40); kloop(48); kloop(41); kloop(34); kloop(27); |
633 | kloop(20); kloop(13); kloop(6); kloop(7); kloop(14); kloop(21); |
634 | kloop(28); kloop(35); kloop(42); kloop(49); kloop(56); kloop(57); |
635 | kloop(50); kloop(43); kloop(36); kloop(29); kloop(22); kloop(15); |
636 | kloop(23); kloop(30); kloop(37); kloop(44); kloop(51); kloop(58); |
637 | kloop(59); kloop(52); kloop(45); kloop(38); kloop(31); kloop(39); |
638 | kloop(46); kloop(53); kloop(60); kloop(61); kloop(54); kloop(47); |
639 | kloop(55); kloop(62); kloop(63); |
640 | |
641 | /* If the last coef(s) were zero, emit an end-of-block code */ |
642 | if (r > 0) { |
643 | PUT_BITS(actbl->ehufco[0], actbl->ehufsi[0]) |
644 | } |
645 | } |
646 | |
647 | state->cur.put_buffer.c = put_buffer; |
648 | state->cur.free_bits = free_bits; |
649 | STORE_BUFFER() |
650 | |
651 | return TRUE; |
652 | } |
653 | |
654 | |
655 | /* |
656 | * Emit a restart marker & resynchronize predictions. |
657 | */ |
658 | |
659 | LOCAL(boolean) |
660 | emit_restart(working_state *state, int restart_num) |
661 | { |
662 | int ci; |
663 | |
664 | if (!flush_bits(state)) |
665 | return FALSE; |
666 | |
667 | emit_byte(state, 0xFF, return FALSE); |
668 | emit_byte(state, JPEG_RST0 + restart_num, return FALSE); |
669 | |
670 | /* Re-initialize DC predictions to 0 */ |
671 | for (ci = 0; ci < state->cinfo->comps_in_scan; ci++) |
672 | state->cur.last_dc_val[ci] = 0; |
673 | |
674 | /* The restart counter is not updated until we successfully write the MCU. */ |
675 | |
676 | return TRUE; |
677 | } |
678 | |
679 | |
680 | /* |
681 | * Encode and output one MCU's worth of Huffman-compressed coefficients. |
682 | */ |
683 | |
684 | METHODDEF(boolean) |
685 | encode_mcu_huff(j_compress_ptr cinfo, JBLOCKROW *MCU_data) |
686 | { |
687 | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
688 | working_state state; |
689 | int blkn, ci; |
690 | jpeg_component_info *compptr; |
691 | |
692 | /* Load up working state */ |
693 | state.next_output_byte = cinfo->dest->next_output_byte; |
694 | state.free_in_buffer = cinfo->dest->free_in_buffer; |
695 | state.cur = entropy->saved; |
696 | state.cinfo = cinfo; |
697 | state.simd = entropy->simd; |
698 | |
699 | /* Emit restart marker if needed */ |
700 | if (cinfo->restart_interval) { |
701 | if (entropy->restarts_to_go == 0) |
702 | if (!emit_restart(&state, entropy->next_restart_num)) |
703 | return FALSE; |
704 | } |
705 | |
706 | /* Encode the MCU data blocks */ |
707 | if (entropy->simd) { |
708 | for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) { |
709 | ci = cinfo->MCU_membership[blkn]; |
710 | compptr = cinfo->cur_comp_info[ci]; |
711 | if (!encode_one_block_simd(&state, |
712 | MCU_data[blkn][0], state.cur.last_dc_val[ci], |
713 | entropy->dc_derived_tbls[compptr->dc_tbl_no], |
714 | entropy->ac_derived_tbls[compptr->ac_tbl_no])) |
715 | return FALSE; |
716 | /* Update last_dc_val */ |
717 | state.cur.last_dc_val[ci] = MCU_data[blkn][0][0]; |
718 | } |
719 | } else { |
720 | for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) { |
721 | ci = cinfo->MCU_membership[blkn]; |
722 | compptr = cinfo->cur_comp_info[ci]; |
723 | if (!encode_one_block(&state, |
724 | MCU_data[blkn][0], state.cur.last_dc_val[ci], |
725 | entropy->dc_derived_tbls[compptr->dc_tbl_no], |
726 | entropy->ac_derived_tbls[compptr->ac_tbl_no])) |
727 | return FALSE; |
728 | /* Update last_dc_val */ |
729 | state.cur.last_dc_val[ci] = MCU_data[blkn][0][0]; |
730 | } |
731 | } |
732 | |
733 | /* Completed MCU, so update state */ |
734 | cinfo->dest->next_output_byte = state.next_output_byte; |
735 | cinfo->dest->free_in_buffer = state.free_in_buffer; |
736 | entropy->saved = state.cur; |
737 | |
738 | /* Update restart-interval state too */ |
739 | if (cinfo->restart_interval) { |
740 | if (entropy->restarts_to_go == 0) { |
741 | entropy->restarts_to_go = cinfo->restart_interval; |
742 | entropy->next_restart_num++; |
743 | entropy->next_restart_num &= 7; |
744 | } |
745 | entropy->restarts_to_go--; |
746 | } |
747 | |
748 | return TRUE; |
749 | } |
750 | |
751 | |
752 | /* |
753 | * Finish up at the end of a Huffman-compressed scan. |
754 | */ |
755 | |
756 | METHODDEF(void) |
757 | finish_pass_huff(j_compress_ptr cinfo) |
758 | { |
759 | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
760 | working_state state; |
761 | |
762 | /* Load up working state ... flush_bits needs it */ |
763 | state.next_output_byte = cinfo->dest->next_output_byte; |
764 | state.free_in_buffer = cinfo->dest->free_in_buffer; |
765 | state.cur = entropy->saved; |
766 | state.cinfo = cinfo; |
767 | state.simd = entropy->simd; |
768 | |
769 | /* Flush out the last data */ |
770 | if (!flush_bits(&state)) |
771 | ERREXIT(cinfo, JERR_CANT_SUSPEND); |
772 | |
773 | /* Update state */ |
774 | cinfo->dest->next_output_byte = state.next_output_byte; |
775 | cinfo->dest->free_in_buffer = state.free_in_buffer; |
776 | entropy->saved = state.cur; |
777 | } |
778 | |
779 | |
780 | /* |
781 | * Huffman coding optimization. |
782 | * |
783 | * We first scan the supplied data and count the number of uses of each symbol |
784 | * that is to be Huffman-coded. (This process MUST agree with the code above.) |
785 | * Then we build a Huffman coding tree for the observed counts. |
786 | * Symbols which are not needed at all for the particular image are not |
787 | * assigned any code, which saves space in the DHT marker as well as in |
788 | * the compressed data. |
789 | */ |
790 | |
791 | #ifdef ENTROPY_OPT_SUPPORTED |
792 | |
793 | |
794 | /* Process a single block's worth of coefficients */ |
795 | |
796 | LOCAL(void) |
797 | htest_one_block(j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val, |
798 | long dc_counts[], long ac_counts[]) |
799 | { |
800 | register int temp; |
801 | register int nbits; |
802 | register int k, r; |
803 | |
804 | /* Encode the DC coefficient difference per section F.1.2.1 */ |
805 | |
806 | temp = block[0] - last_dc_val; |
807 | if (temp < 0) |
808 | temp = -temp; |
809 | |
810 | /* Find the number of bits needed for the magnitude of the coefficient */ |
811 | nbits = 0; |
812 | while (temp) { |
813 | nbits++; |
814 | temp >>= 1; |
815 | } |
816 | /* Check for out-of-range coefficient values. |
817 | * Since we're encoding a difference, the range limit is twice as much. |
818 | */ |
819 | if (nbits > MAX_COEF_BITS + 1) |
820 | ERREXIT(cinfo, JERR_BAD_DCT_COEF); |
821 | |
822 | /* Count the Huffman symbol for the number of bits */ |
823 | dc_counts[nbits]++; |
824 | |
825 | /* Encode the AC coefficients per section F.1.2.2 */ |
826 | |
827 | r = 0; /* r = run length of zeros */ |
828 | |
829 | for (k = 1; k < DCTSIZE2; k++) { |
830 | if ((temp = block[jpeg_natural_order[k]]) == 0) { |
831 | r++; |
832 | } else { |
833 | /* if run length > 15, must emit special run-length-16 codes (0xF0) */ |
834 | while (r > 15) { |
835 | ac_counts[0xF0]++; |
836 | r -= 16; |
837 | } |
838 | |
839 | /* Find the number of bits needed for the magnitude of the coefficient */ |
840 | if (temp < 0) |
841 | temp = -temp; |
842 | |
843 | /* Find the number of bits needed for the magnitude of the coefficient */ |
844 | nbits = 1; /* there must be at least one 1 bit */ |
845 | while ((temp >>= 1)) |
846 | nbits++; |
847 | /* Check for out-of-range coefficient values */ |
848 | if (nbits > MAX_COEF_BITS) |
849 | ERREXIT(cinfo, JERR_BAD_DCT_COEF); |
850 | |
851 | /* Count Huffman symbol for run length / number of bits */ |
852 | ac_counts[(r << 4) + nbits]++; |
853 | |
854 | r = 0; |
855 | } |
856 | } |
857 | |
858 | /* If the last coef(s) were zero, emit an end-of-block code */ |
859 | if (r > 0) |
860 | ac_counts[0]++; |
861 | } |
862 | |
863 | |
864 | /* |
865 | * Trial-encode one MCU's worth of Huffman-compressed coefficients. |
866 | * No data is actually output, so no suspension return is possible. |
867 | */ |
868 | |
869 | METHODDEF(boolean) |
870 | encode_mcu_gather(j_compress_ptr cinfo, JBLOCKROW *MCU_data) |
871 | { |
872 | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
873 | int blkn, ci; |
874 | jpeg_component_info *compptr; |
875 | |
876 | /* Take care of restart intervals if needed */ |
877 | if (cinfo->restart_interval) { |
878 | if (entropy->restarts_to_go == 0) { |
879 | /* Re-initialize DC predictions to 0 */ |
880 | for (ci = 0; ci < cinfo->comps_in_scan; ci++) |
881 | entropy->saved.last_dc_val[ci] = 0; |
882 | /* Update restart state */ |
883 | entropy->restarts_to_go = cinfo->restart_interval; |
884 | } |
885 | entropy->restarts_to_go--; |
886 | } |
887 | |
888 | for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) { |
889 | ci = cinfo->MCU_membership[blkn]; |
890 | compptr = cinfo->cur_comp_info[ci]; |
891 | htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci], |
892 | entropy->dc_count_ptrs[compptr->dc_tbl_no], |
893 | entropy->ac_count_ptrs[compptr->ac_tbl_no]); |
894 | entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0]; |
895 | } |
896 | |
897 | return TRUE; |
898 | } |
899 | |
900 | |
901 | /* |
902 | * Generate the best Huffman code table for the given counts, fill htbl. |
903 | * Note this is also used by jcphuff.c. |
904 | * |
905 | * The JPEG standard requires that no symbol be assigned a codeword of all |
906 | * one bits (so that padding bits added at the end of a compressed segment |
907 | * can't look like a valid code). Because of the canonical ordering of |
908 | * codewords, this just means that there must be an unused slot in the |
909 | * longest codeword length category. Annex K (Clause K.2) of |
910 | * Rec. ITU-T T.81 (1992) | ISO/IEC 10918-1:1994 suggests reserving such a slot |
911 | * by pretending that symbol 256 is a valid symbol with count 1. In theory |
912 | * that's not optimal; giving it count zero but including it in the symbol set |
913 | * anyway should give a better Huffman code. But the theoretically better code |
914 | * actually seems to come out worse in practice, because it produces more |
915 | * all-ones bytes (which incur stuffed zero bytes in the final file). In any |
916 | * case the difference is tiny. |
917 | * |
918 | * The JPEG standard requires Huffman codes to be no more than 16 bits long. |
919 | * If some symbols have a very small but nonzero probability, the Huffman tree |
920 | * must be adjusted to meet the code length restriction. We currently use |
921 | * the adjustment method suggested in JPEG section K.2. This method is *not* |
922 | * optimal; it may not choose the best possible limited-length code. But |
923 | * typically only very-low-frequency symbols will be given less-than-optimal |
924 | * lengths, so the code is almost optimal. Experimental comparisons against |
925 | * an optimal limited-length-code algorithm indicate that the difference is |
926 | * microscopic --- usually less than a hundredth of a percent of total size. |
927 | * So the extra complexity of an optimal algorithm doesn't seem worthwhile. |
928 | */ |
929 | |
930 | GLOBAL(void) |
931 | jpeg_gen_optimal_table(j_compress_ptr cinfo, JHUFF_TBL *htbl, long freq[]) |
932 | { |
933 | #define MAX_CLEN 32 /* assumed maximum initial code length */ |
934 | UINT8 bits[MAX_CLEN + 1]; /* bits[k] = # of symbols with code length k */ |
935 | int codesize[257]; /* codesize[k] = code length of symbol k */ |
936 | int others[257]; /* next symbol in current branch of tree */ |
937 | int c1, c2; |
938 | int p, i, j; |
939 | long v; |
940 | |
941 | /* This algorithm is explained in section K.2 of the JPEG standard */ |
942 | |
943 | MEMZERO(bits, sizeof(bits)); |
944 | MEMZERO(codesize, sizeof(codesize)); |
945 | for (i = 0; i < 257; i++) |
946 | others[i] = -1; /* init links to empty */ |
947 | |
948 | freq[256] = 1; /* make sure 256 has a nonzero count */ |
949 | /* Including the pseudo-symbol 256 in the Huffman procedure guarantees |
950 | * that no real symbol is given code-value of all ones, because 256 |
951 | * will be placed last in the largest codeword category. |
952 | */ |
953 | |
954 | /* Huffman's basic algorithm to assign optimal code lengths to symbols */ |
955 | |
956 | for (;;) { |
957 | /* Find the smallest nonzero frequency, set c1 = its symbol */ |
958 | /* In case of ties, take the larger symbol number */ |
959 | c1 = -1; |
960 | v = 1000000000L; |
961 | for (i = 0; i <= 256; i++) { |
962 | if (freq[i] && freq[i] <= v) { |
963 | v = freq[i]; |
964 | c1 = i; |
965 | } |
966 | } |
967 | |
968 | /* Find the next smallest nonzero frequency, set c2 = its symbol */ |
969 | /* In case of ties, take the larger symbol number */ |
970 | c2 = -1; |
971 | v = 1000000000L; |
972 | for (i = 0; i <= 256; i++) { |
973 | if (freq[i] && freq[i] <= v && i != c1) { |
974 | v = freq[i]; |
975 | c2 = i; |
976 | } |
977 | } |
978 | |
979 | /* Done if we've merged everything into one frequency */ |
980 | if (c2 < 0) |
981 | break; |
982 | |
983 | /* Else merge the two counts/trees */ |
984 | freq[c1] += freq[c2]; |
985 | freq[c2] = 0; |
986 | |
987 | /* Increment the codesize of everything in c1's tree branch */ |
988 | codesize[c1]++; |
989 | while (others[c1] >= 0) { |
990 | c1 = others[c1]; |
991 | codesize[c1]++; |
992 | } |
993 | |
994 | others[c1] = c2; /* chain c2 onto c1's tree branch */ |
995 | |
996 | /* Increment the codesize of everything in c2's tree branch */ |
997 | codesize[c2]++; |
998 | while (others[c2] >= 0) { |
999 | c2 = others[c2]; |
1000 | codesize[c2]++; |
1001 | } |
1002 | } |
1003 | |
1004 | /* Now count the number of symbols of each code length */ |
1005 | for (i = 0; i <= 256; i++) { |
1006 | if (codesize[i]) { |
1007 | /* The JPEG standard seems to think that this can't happen, */ |
1008 | /* but I'm paranoid... */ |
1009 | if (codesize[i] > MAX_CLEN) |
1010 | ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW); |
1011 | |
1012 | bits[codesize[i]]++; |
1013 | } |
1014 | } |
1015 | |
1016 | /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure |
1017 | * Huffman procedure assigned any such lengths, we must adjust the coding. |
1018 | * Here is what Rec. ITU-T T.81 | ISO/IEC 10918-1 says about how this next |
1019 | * bit works: Since symbols are paired for the longest Huffman code, the |
1020 | * symbols are removed from this length category two at a time. The prefix |
1021 | * for the pair (which is one bit shorter) is allocated to one of the pair; |
1022 | * then, skipping the BITS entry for that prefix length, a code word from the |
1023 | * next shortest nonzero BITS entry is converted into a prefix for two code |
1024 | * words one bit longer. |
1025 | */ |
1026 | |
1027 | for (i = MAX_CLEN; i > 16; i--) { |
1028 | while (bits[i] > 0) { |
1029 | j = i - 2; /* find length of new prefix to be used */ |
1030 | while (bits[j] == 0) |
1031 | j--; |
1032 | |
1033 | bits[i] -= 2; /* remove two symbols */ |
1034 | bits[i - 1]++; /* one goes in this length */ |
1035 | bits[j + 1] += 2; /* two new symbols in this length */ |
1036 | bits[j]--; /* symbol of this length is now a prefix */ |
1037 | } |
1038 | } |
1039 | |
1040 | /* Remove the count for the pseudo-symbol 256 from the largest codelength */ |
1041 | while (bits[i] == 0) /* find largest codelength still in use */ |
1042 | i--; |
1043 | bits[i]--; |
1044 | |
1045 | /* Return final symbol counts (only for lengths 0..16) */ |
1046 | MEMCOPY(htbl->bits, bits, sizeof(htbl->bits)); |
1047 | |
1048 | /* Return a list of the symbols sorted by code length */ |
1049 | /* It's not real clear to me why we don't need to consider the codelength |
1050 | * changes made above, but Rec. ITU-T T.81 | ISO/IEC 10918-1 seems to think |
1051 | * this works. |
1052 | */ |
1053 | p = 0; |
1054 | for (i = 1; i <= MAX_CLEN; i++) { |
1055 | for (j = 0; j <= 255; j++) { |
1056 | if (codesize[j] == i) { |
1057 | htbl->huffval[p] = (UINT8)j; |
1058 | p++; |
1059 | } |
1060 | } |
1061 | } |
1062 | |
1063 | /* Set sent_table FALSE so updated table will be written to JPEG file. */ |
1064 | htbl->sent_table = FALSE; |
1065 | } |
1066 | |
1067 | |
1068 | /* |
1069 | * Finish up a statistics-gathering pass and create the new Huffman tables. |
1070 | */ |
1071 | |
1072 | METHODDEF(void) |
1073 | finish_pass_gather(j_compress_ptr cinfo) |
1074 | { |
1075 | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
1076 | int ci, dctbl, actbl; |
1077 | jpeg_component_info *compptr; |
1078 | JHUFF_TBL **htblptr; |
1079 | boolean did_dc[NUM_HUFF_TBLS]; |
1080 | boolean did_ac[NUM_HUFF_TBLS]; |
1081 | |
1082 | /* It's important not to apply jpeg_gen_optimal_table more than once |
1083 | * per table, because it clobbers the input frequency counts! |
1084 | */ |
1085 | MEMZERO(did_dc, sizeof(did_dc)); |
1086 | MEMZERO(did_ac, sizeof(did_ac)); |
1087 | |
1088 | for (ci = 0; ci < cinfo->comps_in_scan; ci++) { |
1089 | compptr = cinfo->cur_comp_info[ci]; |
1090 | dctbl = compptr->dc_tbl_no; |
1091 | actbl = compptr->ac_tbl_no; |
1092 | if (!did_dc[dctbl]) { |
1093 | htblptr = &cinfo->dc_huff_tbl_ptrs[dctbl]; |
1094 | if (*htblptr == NULL) |
1095 | *htblptr = jpeg_alloc_huff_table((j_common_ptr)cinfo); |
1096 | jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]); |
1097 | did_dc[dctbl] = TRUE; |
1098 | } |
1099 | if (!did_ac[actbl]) { |
1100 | htblptr = &cinfo->ac_huff_tbl_ptrs[actbl]; |
1101 | if (*htblptr == NULL) |
1102 | *htblptr = jpeg_alloc_huff_table((j_common_ptr)cinfo); |
1103 | jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]); |
1104 | did_ac[actbl] = TRUE; |
1105 | } |
1106 | } |
1107 | } |
1108 | |
1109 | |
1110 | #endif /* ENTROPY_OPT_SUPPORTED */ |
1111 | |
1112 | |
1113 | /* |
1114 | * Module initialization routine for Huffman entropy encoding. |
1115 | */ |
1116 | |
1117 | GLOBAL(void) |
1118 | jinit_huff_encoder(j_compress_ptr cinfo) |
1119 | { |
1120 | huff_entropy_ptr entropy; |
1121 | int i; |
1122 | |
1123 | entropy = (huff_entropy_ptr) |
1124 | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
1125 | sizeof(huff_entropy_encoder)); |
1126 | cinfo->entropy = (struct jpeg_entropy_encoder *)entropy; |
1127 | entropy->pub.start_pass = start_pass_huff; |
1128 | |
1129 | /* Mark tables unallocated */ |
1130 | for (i = 0; i < NUM_HUFF_TBLS; i++) { |
1131 | entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL; |
1132 | #ifdef ENTROPY_OPT_SUPPORTED |
1133 | entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL; |
1134 | #endif |
1135 | } |
1136 | } |
1137 | |