/proc/self/cwd/external/libjpeg_turbo/jchuff.c
Line | Count | Source (jump to first uncovered line) |
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, 2018-2022, 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 | | /* NOTE: Both GCC and Clang define __GNUC__ */ |
48 | | #if (defined(__GNUC__) && (defined(__arm__) || defined(__aarch64__))) || \ |
49 | | defined(_M_ARM) || defined(_M_ARM64) |
50 | | #if !defined(__thumb__) || defined(__thumb2__) |
51 | | #define USE_CLZ_INTRINSIC |
52 | | #endif |
53 | | #endif |
54 | | |
55 | | #ifdef USE_CLZ_INTRINSIC |
56 | | #if defined(_MSC_VER) && !defined(__clang__) |
57 | | #define JPEG_NBITS_NONZERO(x) (32 - _CountLeadingZeros(x)) |
58 | | #else |
59 | | #define JPEG_NBITS_NONZERO(x) (32 - __builtin_clz(x)) |
60 | | #endif |
61 | | #define JPEG_NBITS(x) (x ? JPEG_NBITS_NONZERO(x) : 0) |
62 | | #else |
63 | | #include "jpeg_nbits_table.h" |
64 | 0 | #define JPEG_NBITS(x) (jpeg_nbits_table[x]) |
65 | 0 | #define JPEG_NBITS_NONZERO(x) JPEG_NBITS(x) |
66 | | #endif |
67 | | |
68 | | |
69 | | /* Expanded entropy encoder object for Huffman encoding. |
70 | | * |
71 | | * The savable_state subrecord contains fields that change within an MCU, |
72 | | * but must not be updated permanently until we complete the MCU. |
73 | | */ |
74 | | |
75 | | #if defined(__x86_64__) && defined(__ILP32__) |
76 | | typedef unsigned long long bit_buf_type; |
77 | | #else |
78 | | typedef size_t bit_buf_type; |
79 | | #endif |
80 | | |
81 | | /* NOTE: The more optimal Huffman encoding algorithm is only used by the |
82 | | * intrinsics implementation of the Arm Neon SIMD extensions, which is why we |
83 | | * retain the old Huffman encoder behavior when using the GAS implementation. |
84 | | */ |
85 | | #if defined(WITH_SIMD) && !(defined(__arm__) || defined(__aarch64__) || \ |
86 | | defined(_M_ARM) || defined(_M_ARM64)) |
87 | | typedef unsigned long long simd_bit_buf_type; |
88 | | #else |
89 | | typedef bit_buf_type simd_bit_buf_type; |
90 | | #endif |
91 | | |
92 | | #if (defined(SIZEOF_SIZE_T) && SIZEOF_SIZE_T == 8) || defined(_WIN64) || \ |
93 | | (defined(__x86_64__) && defined(__ILP32__)) |
94 | 0 | #define BIT_BUF_SIZE 64 |
95 | | #elif (defined(SIZEOF_SIZE_T) && SIZEOF_SIZE_T == 4) || defined(_WIN32) |
96 | | #define BIT_BUF_SIZE 32 |
97 | | #else |
98 | | #error Cannot determine word size |
99 | | #endif |
100 | 2.79k | #define SIMD_BIT_BUF_SIZE (sizeof(simd_bit_buf_type) * 8) |
101 | | |
102 | | typedef struct { |
103 | | union { |
104 | | bit_buf_type c; |
105 | | simd_bit_buf_type simd; |
106 | | } put_buffer; /* current bit accumulation buffer */ |
107 | | int free_bits; /* # of bits available in it */ |
108 | | /* (Neon GAS: # of bits now in it) */ |
109 | | int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */ |
110 | | } savable_state; |
111 | | |
112 | | typedef struct { |
113 | | struct jpeg_entropy_encoder pub; /* public fields */ |
114 | | |
115 | | savable_state saved; /* Bit buffer & DC state at start of MCU */ |
116 | | |
117 | | /* These fields are NOT loaded into local working state. */ |
118 | | unsigned int restarts_to_go; /* MCUs left in this restart interval */ |
119 | | int next_restart_num; /* next restart number to write (0-7) */ |
120 | | |
121 | | /* Pointers to derived tables (these workspaces have image lifespan) */ |
122 | | c_derived_tbl *dc_derived_tbls[NUM_HUFF_TBLS]; |
123 | | c_derived_tbl *ac_derived_tbls[NUM_HUFF_TBLS]; |
124 | | |
125 | | #ifdef ENTROPY_OPT_SUPPORTED /* Statistics tables for optimization */ |
126 | | long *dc_count_ptrs[NUM_HUFF_TBLS]; |
127 | | long *ac_count_ptrs[NUM_HUFF_TBLS]; |
128 | | #endif |
129 | | |
130 | | int simd; |
131 | | } huff_entropy_encoder; |
132 | | |
133 | | typedef huff_entropy_encoder *huff_entropy_ptr; |
134 | | |
135 | | /* Working state while writing an MCU. |
136 | | * This struct contains all the fields that are needed by subroutines. |
137 | | */ |
138 | | |
139 | | typedef struct { |
140 | | JOCTET *next_output_byte; /* => next byte to write in buffer */ |
141 | | size_t free_in_buffer; /* # of byte spaces remaining in buffer */ |
142 | | savable_state cur; /* Current bit buffer & DC state */ |
143 | | j_compress_ptr cinfo; /* dump_buffer needs access to this */ |
144 | | int simd; |
145 | | } working_state; |
146 | | |
147 | | |
148 | | /* Forward declarations */ |
149 | | METHODDEF(boolean) encode_mcu_huff(j_compress_ptr cinfo, JBLOCKROW *MCU_data); |
150 | | METHODDEF(void) finish_pass_huff(j_compress_ptr cinfo); |
151 | | #ifdef ENTROPY_OPT_SUPPORTED |
152 | | METHODDEF(boolean) encode_mcu_gather(j_compress_ptr cinfo, |
153 | | JBLOCKROW *MCU_data); |
154 | | METHODDEF(void) finish_pass_gather(j_compress_ptr cinfo); |
155 | | #endif |
156 | | |
157 | | |
158 | | /* |
159 | | * Initialize for a Huffman-compressed scan. |
160 | | * If gather_statistics is TRUE, we do not output anything during the scan, |
161 | | * just count the Huffman symbols used and generate Huffman code tables. |
162 | | */ |
163 | | |
164 | | METHODDEF(void) |
165 | | start_pass_huff(j_compress_ptr cinfo, boolean gather_statistics) |
166 | 933 | { |
167 | 933 | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
168 | 933 | int ci, dctbl, actbl; |
169 | 933 | jpeg_component_info *compptr; |
170 | | |
171 | 933 | if (gather_statistics) { |
172 | 0 | #ifdef ENTROPY_OPT_SUPPORTED |
173 | 0 | entropy->pub.encode_mcu = encode_mcu_gather; |
174 | 0 | entropy->pub.finish_pass = finish_pass_gather; |
175 | | #else |
176 | | ERREXIT(cinfo, JERR_NOT_COMPILED); |
177 | | #endif |
178 | 933 | } else { |
179 | 933 | entropy->pub.encode_mcu = encode_mcu_huff; |
180 | 933 | entropy->pub.finish_pass = finish_pass_huff; |
181 | 933 | } |
182 | | |
183 | 933 | entropy->simd = jsimd_can_huff_encode_one_block(); |
184 | | |
185 | 2.18k | for (ci = 0; ci < cinfo->comps_in_scan; ci++) { |
186 | 1.25k | compptr = cinfo->cur_comp_info[ci]; |
187 | 1.25k | dctbl = compptr->dc_tbl_no; |
188 | 1.25k | actbl = compptr->ac_tbl_no; |
189 | 1.25k | if (gather_statistics) { |
190 | 0 | #ifdef ENTROPY_OPT_SUPPORTED |
191 | | /* Check for invalid table indexes */ |
192 | | /* (make_c_derived_tbl does this in the other path) */ |
193 | 0 | if (dctbl < 0 || dctbl >= NUM_HUFF_TBLS) |
194 | 0 | ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, dctbl); |
195 | 0 | if (actbl < 0 || actbl >= NUM_HUFF_TBLS) |
196 | 0 | ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, actbl); |
197 | | /* Allocate and zero the statistics tables */ |
198 | | /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */ |
199 | 0 | if (entropy->dc_count_ptrs[dctbl] == NULL) |
200 | 0 | entropy->dc_count_ptrs[dctbl] = (long *) |
201 | 0 | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
202 | 0 | 257 * sizeof(long)); |
203 | 0 | memset(entropy->dc_count_ptrs[dctbl], 0, 257 * sizeof(long)); |
204 | 0 | if (entropy->ac_count_ptrs[actbl] == NULL) |
205 | 0 | entropy->ac_count_ptrs[actbl] = (long *) |
206 | 0 | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
207 | 0 | 257 * sizeof(long)); |
208 | 0 | memset(entropy->ac_count_ptrs[actbl], 0, 257 * sizeof(long)); |
209 | 0 | #endif |
210 | 1.25k | } else { |
211 | | /* Compute derived values for Huffman tables */ |
212 | | /* We may do this more than once for a table, but it's not expensive */ |
213 | 1.25k | jpeg_make_c_derived_tbl(cinfo, TRUE, dctbl, |
214 | 1.25k | &entropy->dc_derived_tbls[dctbl]); |
215 | 1.25k | jpeg_make_c_derived_tbl(cinfo, FALSE, actbl, |
216 | 1.25k | &entropy->ac_derived_tbls[actbl]); |
217 | 1.25k | } |
218 | | /* Initialize DC predictions to 0 */ |
219 | 1.25k | entropy->saved.last_dc_val[ci] = 0; |
220 | 1.25k | } |
221 | | |
222 | | /* Initialize bit buffer to empty */ |
223 | 933 | if (entropy->simd) { |
224 | 933 | entropy->saved.put_buffer.simd = 0; |
225 | | #if defined(__aarch64__) && !defined(NEON_INTRINSICS) |
226 | | entropy->saved.free_bits = 0; |
227 | | #else |
228 | 933 | entropy->saved.free_bits = SIMD_BIT_BUF_SIZE; |
229 | 933 | #endif |
230 | 933 | } else { |
231 | 0 | entropy->saved.put_buffer.c = 0; |
232 | 0 | entropy->saved.free_bits = BIT_BUF_SIZE; |
233 | 0 | } |
234 | | |
235 | | /* Initialize restart stuff */ |
236 | 933 | entropy->restarts_to_go = cinfo->restart_interval; |
237 | 933 | entropy->next_restart_num = 0; |
238 | 933 | } |
239 | | |
240 | | |
241 | | /* |
242 | | * Compute the derived values for a Huffman table. |
243 | | * This routine also performs some validation checks on the table. |
244 | | * |
245 | | * Note this is also used by jcphuff.c. |
246 | | */ |
247 | | |
248 | | GLOBAL(void) |
249 | | jpeg_make_c_derived_tbl(j_compress_ptr cinfo, boolean isDC, int tblno, |
250 | | c_derived_tbl **pdtbl) |
251 | 2.50k | { |
252 | 2.50k | JHUFF_TBL *htbl; |
253 | 2.50k | c_derived_tbl *dtbl; |
254 | 2.50k | int p, i, l, lastp, si, maxsymbol; |
255 | 2.50k | char huffsize[257]; |
256 | 2.50k | unsigned int huffcode[257]; |
257 | 2.50k | unsigned int code; |
258 | | |
259 | | /* Note that huffsize[] and huffcode[] are filled in code-length order, |
260 | | * paralleling the order of the symbols themselves in htbl->huffval[]. |
261 | | */ |
262 | | |
263 | | /* Find the input Huffman table */ |
264 | 2.50k | if (tblno < 0 || tblno >= NUM_HUFF_TBLS) |
265 | 0 | ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno); |
266 | 2.50k | htbl = |
267 | 2.50k | isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno]; |
268 | 2.50k | if (htbl == NULL) |
269 | 0 | ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno); |
270 | | |
271 | | /* Allocate a workspace if we haven't already done so. */ |
272 | 2.50k | if (*pdtbl == NULL) |
273 | 2.18k | *pdtbl = (c_derived_tbl *) |
274 | 2.18k | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
275 | 2.18k | sizeof(c_derived_tbl)); |
276 | 2.50k | dtbl = *pdtbl; |
277 | | |
278 | | /* Figure C.1: make table of Huffman code length for each symbol */ |
279 | | |
280 | 2.50k | p = 0; |
281 | 42.5k | for (l = 1; l <= 16; l++) { |
282 | 40.0k | i = (int)htbl->bits[l]; |
283 | 40.0k | if (i < 0 || p + i > 256) /* protect against table overrun */ |
284 | 0 | ERREXIT(cinfo, JERR_BAD_HUFF_TABLE); |
285 | 257k | while (i--) |
286 | 217k | huffsize[p++] = (char)l; |
287 | 40.0k | } |
288 | 2.50k | huffsize[p] = 0; |
289 | 2.50k | lastp = p; |
290 | | |
291 | | /* Figure C.2: generate the codes themselves */ |
292 | | /* We also validate that the counts represent a legal Huffman code tree. */ |
293 | | |
294 | 2.50k | code = 0; |
295 | 2.50k | si = huffsize[0]; |
296 | 2.50k | p = 0; |
297 | 31.9k | while (huffsize[p]) { |
298 | 247k | while (((int)huffsize[p]) == si) { |
299 | 217k | huffcode[p++] = code; |
300 | 217k | code++; |
301 | 217k | } |
302 | | /* code is now 1 more than the last code used for codelength si; but |
303 | | * it must still fit in si bits, since no code is allowed to be all ones. |
304 | | */ |
305 | 29.4k | if (((JLONG)code) >= (((JLONG)1) << si)) |
306 | 0 | ERREXIT(cinfo, JERR_BAD_HUFF_TABLE); |
307 | 29.4k | code <<= 1; |
308 | 29.4k | si++; |
309 | 29.4k | } |
310 | | |
311 | | /* Figure C.3: generate encoding tables */ |
312 | | /* These are code and size indexed by symbol value */ |
313 | | |
314 | | /* Set all codeless symbols to have code length 0; |
315 | | * this lets us detect duplicate VAL entries here, and later |
316 | | * allows emit_bits to detect any attempt to emit such symbols. |
317 | | */ |
318 | 2.50k | memset(dtbl->ehufco, 0, sizeof(dtbl->ehufco)); |
319 | 2.50k | memset(dtbl->ehufsi, 0, sizeof(dtbl->ehufsi)); |
320 | | |
321 | | /* This is also a convenient place to check for out-of-range |
322 | | * and duplicated VAL entries. We allow 0..255 for AC symbols |
323 | | * but only 0..15 for DC. (We could constrain them further |
324 | | * based on data depth and mode, but this seems enough.) |
325 | | */ |
326 | 2.50k | maxsymbol = isDC ? 15 : 255; |
327 | | |
328 | 220k | for (p = 0; p < lastp; p++) { |
329 | 217k | i = htbl->huffval[p]; |
330 | 217k | if (i < 0 || i > maxsymbol || dtbl->ehufsi[i]) |
331 | 0 | ERREXIT(cinfo, JERR_BAD_HUFF_TABLE); |
332 | 217k | dtbl->ehufco[i] = huffcode[p]; |
333 | 217k | dtbl->ehufsi[i] = huffsize[p]; |
334 | 217k | } |
335 | 2.50k | } |
336 | | |
337 | | |
338 | | /* Outputting bytes to the file */ |
339 | | |
340 | | /* Emit a byte, taking 'action' if must suspend. */ |
341 | 0 | #define emit_byte(state, val, action) { \ |
342 | 0 | *(state)->next_output_byte++ = (JOCTET)(val); \ |
343 | 0 | if (--(state)->free_in_buffer == 0) \ |
344 | 0 | if (!dump_buffer(state)) \ |
345 | 0 | { action; } \ |
346 | 0 | } |
347 | | |
348 | | |
349 | | LOCAL(boolean) |
350 | | dump_buffer(working_state *state) |
351 | | /* Empty the output buffer; return TRUE if successful, FALSE if must suspend */ |
352 | 1.43k | { |
353 | 1.43k | struct jpeg_destination_mgr *dest = state->cinfo->dest; |
354 | | |
355 | 1.43k | if (!(*dest->empty_output_buffer) (state->cinfo)) |
356 | 0 | return FALSE; |
357 | | /* After a successful buffer dump, must reset buffer pointers */ |
358 | 1.43k | state->next_output_byte = dest->next_output_byte; |
359 | 1.43k | state->free_in_buffer = dest->free_in_buffer; |
360 | 1.43k | return TRUE; |
361 | 1.43k | } |
362 | | |
363 | | |
364 | | /* Outputting bits to the file */ |
365 | | |
366 | | /* Output byte b and, speculatively, an additional 0 byte. 0xFF must be |
367 | | * encoded as 0xFF 0x00, so the output buffer pointer is advanced by 2 if the |
368 | | * byte is 0xFF. Otherwise, the output buffer pointer is advanced by 1, and |
369 | | * the speculative 0 byte will be overwritten by the next byte. |
370 | | */ |
371 | 4.13k | #define EMIT_BYTE(b) { \ |
372 | 4.13k | buffer[0] = (JOCTET)(b); \ |
373 | 4.13k | buffer[1] = 0; \ |
374 | 4.13k | buffer -= -2 + ((JOCTET)(b) < 0xFF); \ |
375 | 4.13k | } |
376 | | |
377 | | /* Output the entire bit buffer. If there are no 0xFF bytes in it, then write |
378 | | * directly to the output buffer. Otherwise, use the EMIT_BYTE() macro to |
379 | | * encode 0xFF as 0xFF 0x00. |
380 | | */ |
381 | | #if BIT_BUF_SIZE == 64 |
382 | | |
383 | 0 | #define FLUSH() { \ |
384 | 0 | if (put_buffer & 0x8080808080808080 & ~(put_buffer + 0x0101010101010101)) { \ |
385 | 0 | EMIT_BYTE(put_buffer >> 56) \ |
386 | 0 | EMIT_BYTE(put_buffer >> 48) \ |
387 | 0 | EMIT_BYTE(put_buffer >> 40) \ |
388 | 0 | EMIT_BYTE(put_buffer >> 32) \ |
389 | 0 | EMIT_BYTE(put_buffer >> 24) \ |
390 | 0 | EMIT_BYTE(put_buffer >> 16) \ |
391 | 0 | EMIT_BYTE(put_buffer >> 8) \ |
392 | 0 | EMIT_BYTE(put_buffer ) \ |
393 | 0 | } else { \ |
394 | 0 | buffer[0] = (JOCTET)(put_buffer >> 56); \ |
395 | 0 | buffer[1] = (JOCTET)(put_buffer >> 48); \ |
396 | 0 | buffer[2] = (JOCTET)(put_buffer >> 40); \ |
397 | 0 | buffer[3] = (JOCTET)(put_buffer >> 32); \ |
398 | 0 | buffer[4] = (JOCTET)(put_buffer >> 24); \ |
399 | 0 | buffer[5] = (JOCTET)(put_buffer >> 16); \ |
400 | 0 | buffer[6] = (JOCTET)(put_buffer >> 8); \ |
401 | 0 | buffer[7] = (JOCTET)(put_buffer); \ |
402 | 0 | buffer += 8; \ |
403 | 0 | } \ |
404 | 0 | } |
405 | | |
406 | | #else |
407 | | |
408 | | #define FLUSH() { \ |
409 | | if (put_buffer & 0x80808080 & ~(put_buffer + 0x01010101)) { \ |
410 | | EMIT_BYTE(put_buffer >> 24) \ |
411 | | EMIT_BYTE(put_buffer >> 16) \ |
412 | | EMIT_BYTE(put_buffer >> 8) \ |
413 | | EMIT_BYTE(put_buffer ) \ |
414 | | } else { \ |
415 | | buffer[0] = (JOCTET)(put_buffer >> 24); \ |
416 | | buffer[1] = (JOCTET)(put_buffer >> 16); \ |
417 | | buffer[2] = (JOCTET)(put_buffer >> 8); \ |
418 | | buffer[3] = (JOCTET)(put_buffer); \ |
419 | | buffer += 4; \ |
420 | | } \ |
421 | | } |
422 | | |
423 | | #endif |
424 | | |
425 | | /* Fill the bit buffer to capacity with the leading bits from code, then output |
426 | | * the bit buffer and put the remaining bits from code into the bit buffer. |
427 | | */ |
428 | 0 | #define PUT_AND_FLUSH(code, size) { \ |
429 | 0 | put_buffer = (put_buffer << (size + free_bits)) | (code >> -free_bits); \ |
430 | 0 | FLUSH() \ |
431 | 0 | free_bits += BIT_BUF_SIZE; \ |
432 | 0 | put_buffer = code; \ |
433 | 0 | } |
434 | | |
435 | | /* Insert code into the bit buffer and output the bit buffer if needed. |
436 | | * NOTE: We can't flush with free_bits == 0, since the left shift in |
437 | | * PUT_AND_FLUSH() would have undefined behavior. |
438 | | */ |
439 | 0 | #define PUT_BITS(code, size) { \ |
440 | 0 | free_bits -= size; \ |
441 | 0 | if (free_bits < 0) \ |
442 | 0 | PUT_AND_FLUSH(code, size) \ |
443 | 0 | else \ |
444 | 0 | put_buffer = (put_buffer << size) | code; \ |
445 | 0 | } |
446 | | |
447 | 0 | #define PUT_CODE(code, size) { \ |
448 | 0 | temp &= (((JLONG)1) << nbits) - 1; \ |
449 | 0 | temp |= code << nbits; \ |
450 | 0 | nbits += size; \ |
451 | 0 | PUT_BITS(temp, nbits) \ |
452 | 0 | } |
453 | | |
454 | | |
455 | | /* Although it is exceedingly rare, it is possible for a Huffman-encoded |
456 | | * coefficient block to be larger than the 128-byte unencoded block. For each |
457 | | * of the 64 coefficients, PUT_BITS is invoked twice, and each invocation can |
458 | | * theoretically store 16 bits (for a maximum of 2048 bits or 256 bytes per |
459 | | * encoded block.) If, for instance, one artificially sets the AC |
460 | | * coefficients to alternating values of 32767 and -32768 (using the JPEG |
461 | | * scanning order-- 1, 8, 16, etc.), then this will produce an encoded block |
462 | | * larger than 200 bytes. |
463 | | */ |
464 | 1.79M | #define BUFSIZE (DCTSIZE2 * 8) |
465 | | |
466 | 1.79M | #define LOAD_BUFFER() { \ |
467 | 1.79M | if (state->free_in_buffer < BUFSIZE) { \ |
468 | 12.6k | localbuf = 1; \ |
469 | 12.6k | buffer = _buffer; \ |
470 | 12.6k | } else \ |
471 | 1.79M | buffer = state->next_output_byte; \ |
472 | 1.79M | } |
473 | | |
474 | 1.79M | #define STORE_BUFFER() { \ |
475 | 1.79M | if (localbuf) { \ |
476 | 12.6k | size_t bytes, bytestocopy; \ |
477 | 12.6k | bytes = buffer - _buffer; \ |
478 | 12.6k | buffer = _buffer; \ |
479 | 24.5k | while (bytes > 0) { \ |
480 | 11.9k | bytestocopy = MIN(bytes, state->free_in_buffer); \ |
481 | 11.9k | memcpy(state->next_output_byte, buffer, bytestocopy); \ |
482 | 11.9k | state->next_output_byte += bytestocopy; \ |
483 | 11.9k | buffer += bytestocopy; \ |
484 | 11.9k | state->free_in_buffer -= bytestocopy; \ |
485 | 11.9k | if (state->free_in_buffer == 0) \ |
486 | 11.9k | if (!dump_buffer(state)) return FALSE; \ |
487 | 11.9k | bytes -= bytestocopy; \ |
488 | 11.9k | } \ |
489 | 1.77M | } else { \ |
490 | 1.77M | state->free_in_buffer -= (buffer - state->next_output_byte); \ |
491 | 1.77M | state->next_output_byte = buffer; \ |
492 | 1.77M | } \ |
493 | 1.79M | } |
494 | | |
495 | | |
496 | | LOCAL(boolean) |
497 | | flush_bits(working_state *state) |
498 | 933 | { |
499 | 933 | JOCTET _buffer[BUFSIZE], *buffer, temp; |
500 | 933 | simd_bit_buf_type put_buffer; int put_bits; |
501 | 933 | int localbuf = 0; |
502 | | |
503 | 933 | if (state->simd) { |
504 | | #if defined(__aarch64__) && !defined(NEON_INTRINSICS) |
505 | | put_bits = state->cur.free_bits; |
506 | | #else |
507 | 933 | put_bits = SIMD_BIT_BUF_SIZE - state->cur.free_bits; |
508 | 933 | #endif |
509 | 933 | put_buffer = state->cur.put_buffer.simd; |
510 | 933 | } else { |
511 | 0 | put_bits = BIT_BUF_SIZE - state->cur.free_bits; |
512 | 0 | put_buffer = state->cur.put_buffer.c; |
513 | 0 | } |
514 | | |
515 | 933 | LOAD_BUFFER() |
516 | | |
517 | 4.26k | while (put_bits >= 8) { |
518 | 3.33k | put_bits -= 8; |
519 | 3.33k | temp = (JOCTET)(put_buffer >> put_bits); |
520 | 3.33k | EMIT_BYTE(temp) |
521 | 3.33k | } |
522 | 933 | if (put_bits) { |
523 | | /* fill partial byte with ones */ |
524 | 804 | temp = (JOCTET)((put_buffer << (8 - put_bits)) | (0xFF >> put_bits)); |
525 | 804 | EMIT_BYTE(temp) |
526 | 804 | } |
527 | | |
528 | 933 | if (state->simd) { /* and reset bit buffer to empty */ |
529 | 933 | state->cur.put_buffer.simd = 0; |
530 | | #if defined(__aarch64__) && !defined(NEON_INTRINSICS) |
531 | | state->cur.free_bits = 0; |
532 | | #else |
533 | 933 | state->cur.free_bits = SIMD_BIT_BUF_SIZE; |
534 | 933 | #endif |
535 | 933 | } else { |
536 | 0 | state->cur.put_buffer.c = 0; |
537 | 0 | state->cur.free_bits = BIT_BUF_SIZE; |
538 | 0 | } |
539 | 933 | STORE_BUFFER() |
540 | | |
541 | 933 | return TRUE; |
542 | 933 | } |
543 | | |
544 | | |
545 | | /* Encode a single block's worth of coefficients */ |
546 | | |
547 | | LOCAL(boolean) |
548 | | encode_one_block_simd(working_state *state, JCOEFPTR block, int last_dc_val, |
549 | | c_derived_tbl *dctbl, c_derived_tbl *actbl) |
550 | 1.79M | { |
551 | 1.79M | JOCTET _buffer[BUFSIZE], *buffer; |
552 | 1.79M | int localbuf = 0; |
553 | | |
554 | 1.79M | LOAD_BUFFER() |
555 | | |
556 | 1.79M | buffer = jsimd_huff_encode_one_block(state, buffer, block, last_dc_val, |
557 | 1.79M | dctbl, actbl); |
558 | | |
559 | 1.79M | STORE_BUFFER() |
560 | | |
561 | 1.79M | return TRUE; |
562 | 1.79M | } |
563 | | |
564 | | LOCAL(boolean) |
565 | | encode_one_block(working_state *state, JCOEFPTR block, int last_dc_val, |
566 | | c_derived_tbl *dctbl, c_derived_tbl *actbl) |
567 | 0 | { |
568 | 0 | int temp, nbits, free_bits; |
569 | 0 | bit_buf_type put_buffer; |
570 | 0 | JOCTET _buffer[BUFSIZE], *buffer; |
571 | 0 | int localbuf = 0; |
572 | |
|
573 | 0 | free_bits = state->cur.free_bits; |
574 | 0 | put_buffer = state->cur.put_buffer.c; |
575 | 0 | LOAD_BUFFER() |
576 | | |
577 | | /* Encode the DC coefficient difference per section F.1.2.1 */ |
578 | |
|
579 | 0 | temp = block[0] - last_dc_val; |
580 | | |
581 | | /* This is a well-known technique for obtaining the absolute value without a |
582 | | * branch. It is derived from an assembly language technique presented in |
583 | | * "How to Optimize for the Pentium Processors", Copyright (c) 1996, 1997 by |
584 | | * Agner Fog. This code assumes we are on a two's complement machine. |
585 | | */ |
586 | 0 | nbits = temp >> (CHAR_BIT * sizeof(int) - 1); |
587 | 0 | temp += nbits; |
588 | 0 | nbits ^= temp; |
589 | | |
590 | | /* Find the number of bits needed for the magnitude of the coefficient */ |
591 | 0 | nbits = JPEG_NBITS(nbits); |
592 | | |
593 | | /* Emit the Huffman-coded symbol for the number of bits. |
594 | | * Emit that number of bits of the value, if positive, |
595 | | * or the complement of its magnitude, if negative. |
596 | | */ |
597 | 0 | PUT_CODE(dctbl->ehufco[nbits], dctbl->ehufsi[nbits]) |
598 | | |
599 | | /* Encode the AC coefficients per section F.1.2.2 */ |
600 | |
|
601 | 0 | { |
602 | 0 | int r = 0; /* r = run length of zeros */ |
603 | | |
604 | | /* Manually unroll the k loop to eliminate the counter variable. This |
605 | | * improves performance greatly on systems with a limited number of |
606 | | * registers (such as x86.) |
607 | | */ |
608 | 0 | #define kloop(jpeg_natural_order_of_k) { \ |
609 | 0 | if ((temp = block[jpeg_natural_order_of_k]) == 0) { \ |
610 | 0 | r += 16; \ |
611 | 0 | } else { \ |
612 | | /* Branch-less absolute value, bitwise complement, etc., same as above */ \ |
613 | 0 | nbits = temp >> (CHAR_BIT * sizeof(int) - 1); \ |
614 | 0 | temp += nbits; \ |
615 | 0 | nbits ^= temp; \ |
616 | 0 | nbits = JPEG_NBITS_NONZERO(nbits); \ |
617 | | /* if run length > 15, must emit special run-length-16 codes (0xF0) */ \ |
618 | 0 | while (r >= 16 * 16) { \ |
619 | 0 | r -= 16 * 16; \ |
620 | 0 | PUT_BITS(actbl->ehufco[0xf0], actbl->ehufsi[0xf0]) \ |
621 | 0 | } \ |
622 | | /* Emit Huffman symbol for run length / number of bits */ \ |
623 | 0 | r += nbits; \ |
624 | 0 | PUT_CODE(actbl->ehufco[r], actbl->ehufsi[r]) \ |
625 | 0 | r = 0; \ |
626 | 0 | } \ |
627 | 0 | } |
628 | | |
629 | | /* One iteration for each value in jpeg_natural_order[] */ |
630 | 0 | kloop(1); kloop(8); kloop(16); kloop(9); kloop(2); kloop(3); |
631 | 0 | kloop(10); kloop(17); kloop(24); kloop(32); kloop(25); kloop(18); |
632 | 0 | kloop(11); kloop(4); kloop(5); kloop(12); kloop(19); kloop(26); |
633 | 0 | kloop(33); kloop(40); kloop(48); kloop(41); kloop(34); kloop(27); |
634 | 0 | kloop(20); kloop(13); kloop(6); kloop(7); kloop(14); kloop(21); |
635 | 0 | kloop(28); kloop(35); kloop(42); kloop(49); kloop(56); kloop(57); |
636 | 0 | kloop(50); kloop(43); kloop(36); kloop(29); kloop(22); kloop(15); |
637 | 0 | kloop(23); kloop(30); kloop(37); kloop(44); kloop(51); kloop(58); |
638 | 0 | kloop(59); kloop(52); kloop(45); kloop(38); kloop(31); kloop(39); |
639 | 0 | kloop(46); kloop(53); kloop(60); kloop(61); kloop(54); kloop(47); |
640 | 0 | kloop(55); kloop(62); kloop(63); |
641 | | |
642 | | /* If the last coef(s) were zero, emit an end-of-block code */ |
643 | 0 | if (r > 0) { |
644 | 0 | PUT_BITS(actbl->ehufco[0], actbl->ehufsi[0]) |
645 | 0 | } |
646 | 0 | } |
647 | |
|
648 | 0 | state->cur.put_buffer.c = put_buffer; |
649 | 0 | state->cur.free_bits = free_bits; |
650 | 0 | STORE_BUFFER() |
651 | | |
652 | 0 | return TRUE; |
653 | 0 | } |
654 | | |
655 | | |
656 | | /* |
657 | | * Emit a restart marker & resynchronize predictions. |
658 | | */ |
659 | | |
660 | | LOCAL(boolean) |
661 | | emit_restart(working_state *state, int restart_num) |
662 | 0 | { |
663 | 0 | int ci; |
664 | |
|
665 | 0 | if (!flush_bits(state)) |
666 | 0 | return FALSE; |
667 | | |
668 | 0 | emit_byte(state, 0xFF, return FALSE); |
669 | 0 | emit_byte(state, JPEG_RST0 + restart_num, return FALSE); |
670 | | |
671 | | /* Re-initialize DC predictions to 0 */ |
672 | 0 | for (ci = 0; ci < state->cinfo->comps_in_scan; ci++) |
673 | 0 | state->cur.last_dc_val[ci] = 0; |
674 | | |
675 | | /* The restart counter is not updated until we successfully write the MCU. */ |
676 | |
|
677 | 0 | return TRUE; |
678 | 0 | } |
679 | | |
680 | | |
681 | | /* |
682 | | * Encode and output one MCU's worth of Huffman-compressed coefficients. |
683 | | */ |
684 | | |
685 | | METHODDEF(boolean) |
686 | | encode_mcu_huff(j_compress_ptr cinfo, JBLOCKROW *MCU_data) |
687 | 1.67M | { |
688 | 1.67M | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
689 | 1.67M | working_state state; |
690 | 1.67M | int blkn, ci; |
691 | 1.67M | jpeg_component_info *compptr; |
692 | | |
693 | | /* Load up working state */ |
694 | 1.67M | state.next_output_byte = cinfo->dest->next_output_byte; |
695 | 1.67M | state.free_in_buffer = cinfo->dest->free_in_buffer; |
696 | 1.67M | state.cur = entropy->saved; |
697 | 1.67M | state.cinfo = cinfo; |
698 | 1.67M | state.simd = entropy->simd; |
699 | | |
700 | | /* Emit restart marker if needed */ |
701 | 1.67M | if (cinfo->restart_interval) { |
702 | 0 | if (entropy->restarts_to_go == 0) |
703 | 0 | if (!emit_restart(&state, entropy->next_restart_num)) |
704 | 0 | return FALSE; |
705 | 0 | } |
706 | | |
707 | | /* Encode the MCU data blocks */ |
708 | 1.67M | if (entropy->simd) { |
709 | 3.46M | for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) { |
710 | 1.79M | ci = cinfo->MCU_membership[blkn]; |
711 | 1.79M | compptr = cinfo->cur_comp_info[ci]; |
712 | 1.79M | if (!encode_one_block_simd(&state, |
713 | 1.79M | MCU_data[blkn][0], state.cur.last_dc_val[ci], |
714 | 1.79M | entropy->dc_derived_tbls[compptr->dc_tbl_no], |
715 | 1.79M | entropy->ac_derived_tbls[compptr->ac_tbl_no])) |
716 | 0 | return FALSE; |
717 | | /* Update last_dc_val */ |
718 | 1.79M | state.cur.last_dc_val[ci] = MCU_data[blkn][0][0]; |
719 | 1.79M | } |
720 | 1.67M | } else { |
721 | 0 | for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) { |
722 | 0 | ci = cinfo->MCU_membership[blkn]; |
723 | 0 | compptr = cinfo->cur_comp_info[ci]; |
724 | 0 | if (!encode_one_block(&state, |
725 | 0 | MCU_data[blkn][0], state.cur.last_dc_val[ci], |
726 | 0 | entropy->dc_derived_tbls[compptr->dc_tbl_no], |
727 | 0 | entropy->ac_derived_tbls[compptr->ac_tbl_no])) |
728 | 0 | return FALSE; |
729 | | /* Update last_dc_val */ |
730 | 0 | state.cur.last_dc_val[ci] = MCU_data[blkn][0][0]; |
731 | 0 | } |
732 | 0 | } |
733 | | |
734 | | /* Completed MCU, so update state */ |
735 | 1.67M | cinfo->dest->next_output_byte = state.next_output_byte; |
736 | 1.67M | cinfo->dest->free_in_buffer = state.free_in_buffer; |
737 | 1.67M | entropy->saved = state.cur; |
738 | | |
739 | | /* Update restart-interval state too */ |
740 | 1.67M | if (cinfo->restart_interval) { |
741 | 0 | if (entropy->restarts_to_go == 0) { |
742 | 0 | entropy->restarts_to_go = cinfo->restart_interval; |
743 | 0 | entropy->next_restart_num++; |
744 | 0 | entropy->next_restart_num &= 7; |
745 | 0 | } |
746 | 0 | entropy->restarts_to_go--; |
747 | 0 | } |
748 | | |
749 | 1.67M | return TRUE; |
750 | 1.67M | } |
751 | | |
752 | | |
753 | | /* |
754 | | * Finish up at the end of a Huffman-compressed scan. |
755 | | */ |
756 | | |
757 | | METHODDEF(void) |
758 | | finish_pass_huff(j_compress_ptr cinfo) |
759 | 933 | { |
760 | 933 | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
761 | 933 | working_state state; |
762 | | |
763 | | /* Load up working state ... flush_bits needs it */ |
764 | 933 | state.next_output_byte = cinfo->dest->next_output_byte; |
765 | 933 | state.free_in_buffer = cinfo->dest->free_in_buffer; |
766 | 933 | state.cur = entropy->saved; |
767 | 933 | state.cinfo = cinfo; |
768 | 933 | state.simd = entropy->simd; |
769 | | |
770 | | /* Flush out the last data */ |
771 | 933 | if (!flush_bits(&state)) |
772 | 0 | ERREXIT(cinfo, JERR_CANT_SUSPEND); |
773 | | |
774 | | /* Update state */ |
775 | 933 | cinfo->dest->next_output_byte = state.next_output_byte; |
776 | 933 | cinfo->dest->free_in_buffer = state.free_in_buffer; |
777 | 933 | entropy->saved = state.cur; |
778 | 933 | } |
779 | | |
780 | | |
781 | | /* |
782 | | * Huffman coding optimization. |
783 | | * |
784 | | * We first scan the supplied data and count the number of uses of each symbol |
785 | | * that is to be Huffman-coded. (This process MUST agree with the code above.) |
786 | | * Then we build a Huffman coding tree for the observed counts. |
787 | | * Symbols which are not needed at all for the particular image are not |
788 | | * assigned any code, which saves space in the DHT marker as well as in |
789 | | * the compressed data. |
790 | | */ |
791 | | |
792 | | #ifdef ENTROPY_OPT_SUPPORTED |
793 | | |
794 | | |
795 | | /* Process a single block's worth of coefficients */ |
796 | | |
797 | | LOCAL(void) |
798 | | htest_one_block(j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val, |
799 | | long dc_counts[], long ac_counts[]) |
800 | 0 | { |
801 | 0 | register int temp; |
802 | 0 | register int nbits; |
803 | 0 | register int k, r; |
804 | | |
805 | | /* Encode the DC coefficient difference per section F.1.2.1 */ |
806 | |
|
807 | 0 | temp = block[0] - last_dc_val; |
808 | 0 | if (temp < 0) |
809 | 0 | temp = -temp; |
810 | | |
811 | | /* Find the number of bits needed for the magnitude of the coefficient */ |
812 | 0 | nbits = 0; |
813 | 0 | while (temp) { |
814 | 0 | nbits++; |
815 | 0 | temp >>= 1; |
816 | 0 | } |
817 | | /* Check for out-of-range coefficient values. |
818 | | * Since we're encoding a difference, the range limit is twice as much. |
819 | | */ |
820 | 0 | if (nbits > MAX_COEF_BITS + 1) |
821 | 0 | ERREXIT(cinfo, JERR_BAD_DCT_COEF); |
822 | | |
823 | | /* Count the Huffman symbol for the number of bits */ |
824 | 0 | dc_counts[nbits]++; |
825 | | |
826 | | /* Encode the AC coefficients per section F.1.2.2 */ |
827 | |
|
828 | 0 | r = 0; /* r = run length of zeros */ |
829 | |
|
830 | 0 | for (k = 1; k < DCTSIZE2; k++) { |
831 | 0 | if ((temp = block[jpeg_natural_order[k]]) == 0) { |
832 | 0 | r++; |
833 | 0 | } else { |
834 | | /* if run length > 15, must emit special run-length-16 codes (0xF0) */ |
835 | 0 | while (r > 15) { |
836 | 0 | ac_counts[0xF0]++; |
837 | 0 | r -= 16; |
838 | 0 | } |
839 | | |
840 | | /* Find the number of bits needed for the magnitude of the coefficient */ |
841 | 0 | if (temp < 0) |
842 | 0 | temp = -temp; |
843 | | |
844 | | /* Find the number of bits needed for the magnitude of the coefficient */ |
845 | 0 | nbits = 1; /* there must be at least one 1 bit */ |
846 | 0 | while ((temp >>= 1)) |
847 | 0 | nbits++; |
848 | | /* Check for out-of-range coefficient values */ |
849 | 0 | if (nbits > MAX_COEF_BITS) |
850 | 0 | ERREXIT(cinfo, JERR_BAD_DCT_COEF); |
851 | | |
852 | | /* Count Huffman symbol for run length / number of bits */ |
853 | 0 | ac_counts[(r << 4) + nbits]++; |
854 | |
|
855 | 0 | r = 0; |
856 | 0 | } |
857 | 0 | } |
858 | | |
859 | | /* If the last coef(s) were zero, emit an end-of-block code */ |
860 | 0 | if (r > 0) |
861 | 0 | ac_counts[0]++; |
862 | 0 | } |
863 | | |
864 | | |
865 | | /* |
866 | | * Trial-encode one MCU's worth of Huffman-compressed coefficients. |
867 | | * No data is actually output, so no suspension return is possible. |
868 | | */ |
869 | | |
870 | | METHODDEF(boolean) |
871 | | encode_mcu_gather(j_compress_ptr cinfo, JBLOCKROW *MCU_data) |
872 | 0 | { |
873 | 0 | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
874 | 0 | int blkn, ci; |
875 | 0 | jpeg_component_info *compptr; |
876 | | |
877 | | /* Take care of restart intervals if needed */ |
878 | 0 | if (cinfo->restart_interval) { |
879 | 0 | if (entropy->restarts_to_go == 0) { |
880 | | /* Re-initialize DC predictions to 0 */ |
881 | 0 | for (ci = 0; ci < cinfo->comps_in_scan; ci++) |
882 | 0 | entropy->saved.last_dc_val[ci] = 0; |
883 | | /* Update restart state */ |
884 | 0 | entropy->restarts_to_go = cinfo->restart_interval; |
885 | 0 | } |
886 | 0 | entropy->restarts_to_go--; |
887 | 0 | } |
888 | |
|
889 | 0 | for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) { |
890 | 0 | ci = cinfo->MCU_membership[blkn]; |
891 | 0 | compptr = cinfo->cur_comp_info[ci]; |
892 | 0 | htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci], |
893 | 0 | entropy->dc_count_ptrs[compptr->dc_tbl_no], |
894 | 0 | entropy->ac_count_ptrs[compptr->ac_tbl_no]); |
895 | 0 | entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0]; |
896 | 0 | } |
897 | |
|
898 | 0 | return TRUE; |
899 | 0 | } |
900 | | |
901 | | |
902 | | /* |
903 | | * Generate the best Huffman code table for the given counts, fill htbl. |
904 | | * Note this is also used by jcphuff.c. |
905 | | * |
906 | | * The JPEG standard requires that no symbol be assigned a codeword of all |
907 | | * one bits (so that padding bits added at the end of a compressed segment |
908 | | * can't look like a valid code). Because of the canonical ordering of |
909 | | * codewords, this just means that there must be an unused slot in the |
910 | | * longest codeword length category. Annex K (Clause K.2) of |
911 | | * Rec. ITU-T T.81 (1992) | ISO/IEC 10918-1:1994 suggests reserving such a slot |
912 | | * by pretending that symbol 256 is a valid symbol with count 1. In theory |
913 | | * that's not optimal; giving it count zero but including it in the symbol set |
914 | | * anyway should give a better Huffman code. But the theoretically better code |
915 | | * actually seems to come out worse in practice, because it produces more |
916 | | * all-ones bytes (which incur stuffed zero bytes in the final file). In any |
917 | | * case the difference is tiny. |
918 | | * |
919 | | * The JPEG standard requires Huffman codes to be no more than 16 bits long. |
920 | | * If some symbols have a very small but nonzero probability, the Huffman tree |
921 | | * must be adjusted to meet the code length restriction. We currently use |
922 | | * the adjustment method suggested in JPEG section K.2. This method is *not* |
923 | | * optimal; it may not choose the best possible limited-length code. But |
924 | | * typically only very-low-frequency symbols will be given less-than-optimal |
925 | | * lengths, so the code is almost optimal. Experimental comparisons against |
926 | | * an optimal limited-length-code algorithm indicate that the difference is |
927 | | * microscopic --- usually less than a hundredth of a percent of total size. |
928 | | * So the extra complexity of an optimal algorithm doesn't seem worthwhile. |
929 | | */ |
930 | | |
931 | | GLOBAL(void) |
932 | | jpeg_gen_optimal_table(j_compress_ptr cinfo, JHUFF_TBL *htbl, long freq[]) |
933 | 0 | { |
934 | 0 | #define MAX_CLEN 32 /* assumed maximum initial code length */ |
935 | 0 | UINT8 bits[MAX_CLEN + 1]; /* bits[k] = # of symbols with code length k */ |
936 | 0 | int codesize[257]; /* codesize[k] = code length of symbol k */ |
937 | 0 | int others[257]; /* next symbol in current branch of tree */ |
938 | 0 | int c1, c2; |
939 | 0 | int p, i, j; |
940 | 0 | long v; |
941 | | |
942 | | /* This algorithm is explained in section K.2 of the JPEG standard */ |
943 | |
|
944 | 0 | memset(bits, 0, sizeof(bits)); |
945 | 0 | memset(codesize, 0, sizeof(codesize)); |
946 | 0 | for (i = 0; i < 257; i++) |
947 | 0 | others[i] = -1; /* init links to empty */ |
948 | |
|
949 | 0 | freq[256] = 1; /* make sure 256 has a nonzero count */ |
950 | | /* Including the pseudo-symbol 256 in the Huffman procedure guarantees |
951 | | * that no real symbol is given code-value of all ones, because 256 |
952 | | * will be placed last in the largest codeword category. |
953 | | */ |
954 | | |
955 | | /* Huffman's basic algorithm to assign optimal code lengths to symbols */ |
956 | |
|
957 | 0 | for (;;) { |
958 | | /* Find the smallest nonzero frequency, set c1 = its symbol */ |
959 | | /* In case of ties, take the larger symbol number */ |
960 | 0 | c1 = -1; |
961 | 0 | v = 1000000000L; |
962 | 0 | for (i = 0; i <= 256; i++) { |
963 | 0 | if (freq[i] && freq[i] <= v) { |
964 | 0 | v = freq[i]; |
965 | 0 | c1 = i; |
966 | 0 | } |
967 | 0 | } |
968 | | |
969 | | /* Find the next smallest nonzero frequency, set c2 = its symbol */ |
970 | | /* In case of ties, take the larger symbol number */ |
971 | 0 | c2 = -1; |
972 | 0 | v = 1000000000L; |
973 | 0 | for (i = 0; i <= 256; i++) { |
974 | 0 | if (freq[i] && freq[i] <= v && i != c1) { |
975 | 0 | v = freq[i]; |
976 | 0 | c2 = i; |
977 | 0 | } |
978 | 0 | } |
979 | | |
980 | | /* Done if we've merged everything into one frequency */ |
981 | 0 | if (c2 < 0) |
982 | 0 | break; |
983 | | |
984 | | /* Else merge the two counts/trees */ |
985 | 0 | freq[c1] += freq[c2]; |
986 | 0 | freq[c2] = 0; |
987 | | |
988 | | /* Increment the codesize of everything in c1's tree branch */ |
989 | 0 | codesize[c1]++; |
990 | 0 | while (others[c1] >= 0) { |
991 | 0 | c1 = others[c1]; |
992 | 0 | codesize[c1]++; |
993 | 0 | } |
994 | |
|
995 | 0 | others[c1] = c2; /* chain c2 onto c1's tree branch */ |
996 | | |
997 | | /* Increment the codesize of everything in c2's tree branch */ |
998 | 0 | codesize[c2]++; |
999 | 0 | while (others[c2] >= 0) { |
1000 | 0 | c2 = others[c2]; |
1001 | 0 | codesize[c2]++; |
1002 | 0 | } |
1003 | 0 | } |
1004 | | |
1005 | | /* Now count the number of symbols of each code length */ |
1006 | 0 | for (i = 0; i <= 256; i++) { |
1007 | 0 | if (codesize[i]) { |
1008 | | /* The JPEG standard seems to think that this can't happen, */ |
1009 | | /* but I'm paranoid... */ |
1010 | 0 | if (codesize[i] > MAX_CLEN) |
1011 | 0 | ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW); |
1012 | |
|
1013 | 0 | bits[codesize[i]]++; |
1014 | 0 | } |
1015 | 0 | } |
1016 | | |
1017 | | /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure |
1018 | | * Huffman procedure assigned any such lengths, we must adjust the coding. |
1019 | | * Here is what Rec. ITU-T T.81 | ISO/IEC 10918-1 says about how this next |
1020 | | * bit works: Since symbols are paired for the longest Huffman code, the |
1021 | | * symbols are removed from this length category two at a time. The prefix |
1022 | | * for the pair (which is one bit shorter) is allocated to one of the pair; |
1023 | | * then, skipping the BITS entry for that prefix length, a code word from the |
1024 | | * next shortest nonzero BITS entry is converted into a prefix for two code |
1025 | | * words one bit longer. |
1026 | | */ |
1027 | |
|
1028 | 0 | for (i = MAX_CLEN; i > 16; i--) { |
1029 | 0 | while (bits[i] > 0) { |
1030 | 0 | j = i - 2; /* find length of new prefix to be used */ |
1031 | 0 | while (bits[j] == 0) |
1032 | 0 | j--; |
1033 | |
|
1034 | 0 | bits[i] -= 2; /* remove two symbols */ |
1035 | 0 | bits[i - 1]++; /* one goes in this length */ |
1036 | 0 | bits[j + 1] += 2; /* two new symbols in this length */ |
1037 | 0 | bits[j]--; /* symbol of this length is now a prefix */ |
1038 | 0 | } |
1039 | 0 | } |
1040 | | |
1041 | | /* Remove the count for the pseudo-symbol 256 from the largest codelength */ |
1042 | 0 | while (bits[i] == 0) /* find largest codelength still in use */ |
1043 | 0 | i--; |
1044 | 0 | bits[i]--; |
1045 | | |
1046 | | /* Return final symbol counts (only for lengths 0..16) */ |
1047 | 0 | memcpy(htbl->bits, bits, sizeof(htbl->bits)); |
1048 | | |
1049 | | /* Return a list of the symbols sorted by code length */ |
1050 | | /* It's not real clear to me why we don't need to consider the codelength |
1051 | | * changes made above, but Rec. ITU-T T.81 | ISO/IEC 10918-1 seems to think |
1052 | | * this works. |
1053 | | */ |
1054 | 0 | p = 0; |
1055 | 0 | for (i = 1; i <= MAX_CLEN; i++) { |
1056 | 0 | for (j = 0; j <= 255; j++) { |
1057 | 0 | if (codesize[j] == i) { |
1058 | 0 | htbl->huffval[p] = (UINT8)j; |
1059 | 0 | p++; |
1060 | 0 | } |
1061 | 0 | } |
1062 | 0 | } |
1063 | | |
1064 | | /* Set sent_table FALSE so updated table will be written to JPEG file. */ |
1065 | 0 | htbl->sent_table = FALSE; |
1066 | 0 | } |
1067 | | |
1068 | | |
1069 | | /* |
1070 | | * Finish up a statistics-gathering pass and create the new Huffman tables. |
1071 | | */ |
1072 | | |
1073 | | METHODDEF(void) |
1074 | | finish_pass_gather(j_compress_ptr cinfo) |
1075 | 0 | { |
1076 | 0 | huff_entropy_ptr entropy = (huff_entropy_ptr)cinfo->entropy; |
1077 | 0 | int ci, dctbl, actbl; |
1078 | 0 | jpeg_component_info *compptr; |
1079 | 0 | JHUFF_TBL **htblptr; |
1080 | 0 | boolean did_dc[NUM_HUFF_TBLS]; |
1081 | 0 | boolean did_ac[NUM_HUFF_TBLS]; |
1082 | | |
1083 | | /* It's important not to apply jpeg_gen_optimal_table more than once |
1084 | | * per table, because it clobbers the input frequency counts! |
1085 | | */ |
1086 | 0 | memset(did_dc, 0, sizeof(did_dc)); |
1087 | 0 | memset(did_ac, 0, sizeof(did_ac)); |
1088 | |
|
1089 | 0 | for (ci = 0; ci < cinfo->comps_in_scan; ci++) { |
1090 | 0 | compptr = cinfo->cur_comp_info[ci]; |
1091 | 0 | dctbl = compptr->dc_tbl_no; |
1092 | 0 | actbl = compptr->ac_tbl_no; |
1093 | 0 | if (!did_dc[dctbl]) { |
1094 | 0 | htblptr = &cinfo->dc_huff_tbl_ptrs[dctbl]; |
1095 | 0 | if (*htblptr == NULL) |
1096 | 0 | *htblptr = jpeg_alloc_huff_table((j_common_ptr)cinfo); |
1097 | 0 | jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]); |
1098 | 0 | did_dc[dctbl] = TRUE; |
1099 | 0 | } |
1100 | 0 | if (!did_ac[actbl]) { |
1101 | 0 | htblptr = &cinfo->ac_huff_tbl_ptrs[actbl]; |
1102 | 0 | if (*htblptr == NULL) |
1103 | 0 | *htblptr = jpeg_alloc_huff_table((j_common_ptr)cinfo); |
1104 | 0 | jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]); |
1105 | 0 | did_ac[actbl] = TRUE; |
1106 | 0 | } |
1107 | 0 | } |
1108 | 0 | } |
1109 | | |
1110 | | |
1111 | | #endif /* ENTROPY_OPT_SUPPORTED */ |
1112 | | |
1113 | | |
1114 | | /* |
1115 | | * Module initialization routine for Huffman entropy encoding. |
1116 | | */ |
1117 | | |
1118 | | GLOBAL(void) |
1119 | | jinit_huff_encoder(j_compress_ptr cinfo) |
1120 | 933 | { |
1121 | 933 | huff_entropy_ptr entropy; |
1122 | 933 | int i; |
1123 | | |
1124 | 933 | entropy = (huff_entropy_ptr) |
1125 | 933 | (*cinfo->mem->alloc_small) ((j_common_ptr)cinfo, JPOOL_IMAGE, |
1126 | 933 | sizeof(huff_entropy_encoder)); |
1127 | 933 | cinfo->entropy = (struct jpeg_entropy_encoder *)entropy; |
1128 | 933 | entropy->pub.start_pass = start_pass_huff; |
1129 | | |
1130 | | /* Mark tables unallocated */ |
1131 | 4.66k | for (i = 0; i < NUM_HUFF_TBLS; i++) { |
1132 | 3.73k | entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL; |
1133 | 3.73k | #ifdef ENTROPY_OPT_SUPPORTED |
1134 | 3.73k | entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL; |
1135 | 3.73k | #endif |
1136 | 3.73k | } |
1137 | 933 | } |