Split bitstream.h, put the bitstream writer stuff in the new file
[libav.git] / libavcodec / cabac.h
CommitLineData
d592f67f
MN
1/*
2 * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4 *
b78e7197
DB
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
d592f67f
MN
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
b78e7197 10 * version 2.1 of the License, or (at your option) any later version.
d592f67f 11 *
b78e7197 12 * FFmpeg is distributed in the hope that it will be useful,
d592f67f
MN
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
b78e7197 18 * License along with FFmpeg; if not, write to the Free Software
5509bffa 19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
d592f67f 20 */
115329f1 21
d592f67f 22/**
bad5537e 23 * @file libavcodec/cabac.h
d592f67f
MN
24 * Context Adaptive Binary Arithmetic Coder.
25 */
26
98790382
SS
27#ifndef AVCODEC_CABAC_H
28#define AVCODEC_CABAC_H
26b4fe82 29
b2755007 30#include "put_bits.h"
d592f67f 31
2848ce84 32//#undef NDEBUG
d592f67f 33#include <assert.h>
245976da 34#include "libavutil/x86_cpu.h"
d592f67f 35
5659b509 36#define CABAC_BITS 16
ec7eb896 37#define CABAC_MASK ((1<<CABAC_BITS)-1)
0bc2e7f0 38#define BRANCHLESS_CABAC_DECODER 1
a0f2c6ba 39//#define ARCH_X86_DISABLED 1
ec7eb896 40
d592f67f
MN
41typedef struct CABACContext{
42 int low;
43 int range;
44 int outstanding_count;
45#ifdef STRICT_LIMITS
46 int symCount;
47#endif
e96682e6
MN
48 const uint8_t *bytestream_start;
49 const uint8_t *bytestream;
bba83349 50 const uint8_t *bytestream_end;
d592f67f
MN
51 PutBitContext pb;
52}CABACContext;
53
68a205ed 54extern uint8_t ff_h264_mlps_state[4*64];
a0f2c6ba 55extern uint8_t ff_h264_lps_range[4*2*64]; ///< rangeTabLPS
d61c4e73
MN
56extern uint8_t ff_h264_mps_state[2*64]; ///< transIdxMPS
57extern uint8_t ff_h264_lps_state[2*64]; ///< transIdxLPS
f24a5159 58extern const uint8_t ff_h264_norm_shift[512];
ec7eb896 59
d592f67f
MN
60
61void ff_init_cabac_encoder(CABACContext *c, uint8_t *buf, int buf_size);
e96682e6 62void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size);
d61c4e73 63void ff_init_cabac_states(CABACContext *c);
d592f67f
MN
64
65
66static inline void put_cabac_bit(CABACContext *c, int b){
115329f1
DB
67 put_bits(&c->pb, 1, b);
68 for(;c->outstanding_count; c->outstanding_count--){
d592f67f
MN
69 put_bits(&c->pb, 1, 1-b);
70 }
71}
72
73static inline void renorm_cabac_encoder(CABACContext *c){
74 while(c->range < 0x100){
75 //FIXME optimize
76 if(c->low<0x100){
77 put_cabac_bit(c, 0);
78 }else if(c->low<0x200){
79 c->outstanding_count++;
80 c->low -= 0x100;
81 }else{
82 put_cabac_bit(c, 1);
83 c->low -= 0x200;
84 }
115329f1 85
d592f67f
MN
86 c->range+= c->range;
87 c->low += c->low;
88 }
89}
90
c46e2874 91#ifdef TEST
938dd846 92static void put_cabac(CABACContext *c, uint8_t * const state, int bit){
f24a5159 93 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state];
115329f1 94
d592f67f
MN
95 if(bit == ((*state)&1)){
96 c->range -= RangeLPS;
d61c4e73 97 *state= ff_h264_mps_state[*state];
d592f67f
MN
98 }else{
99 c->low += c->range - RangeLPS;
100 c->range = RangeLPS;
d61c4e73 101 *state= ff_h264_lps_state[*state];
d592f67f 102 }
115329f1 103
d592f67f
MN
104 renorm_cabac_encoder(c);
105
106#ifdef STRICT_LIMITS
107 c->symCount++;
108#endif
109}
110
938dd846 111static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){
d592f67f
MN
112 assert(c->range > RangeLPS);
113
114 if(!bit){
115 c->range -= RangeLPS;
116 }else{
117 c->low += c->range - RangeLPS;
118 c->range = RangeLPS;
119 }
120
121 renorm_cabac_encoder(c);
122
123#ifdef STRICT_LIMITS
124 c->symCount++;
125#endif
126}
127
61ccfcc0
MN
128/**
129 * @param bit 0 -> write zero bit, !=0 write one bit
130 */
938dd846 131static void put_cabac_bypass(CABACContext *c, int bit){
d592f67f
MN
132 c->low += c->low;
133
134 if(bit){
135 c->low += c->range;
136 }
137//FIXME optimize
138 if(c->low<0x200){
139 put_cabac_bit(c, 0);
140 }else if(c->low<0x400){
141 c->outstanding_count++;
142 c->low -= 0x200;
143 }else{
144 put_cabac_bit(c, 1);
145 c->low -= 0x400;
146 }
115329f1 147
d592f67f
MN
148#ifdef STRICT_LIMITS
149 c->symCount++;
150#endif
151}
152
5e20f836
MN
153/**
154 *
155 * @return the number of bytes written
156 */
938dd846 157static int put_cabac_terminate(CABACContext *c, int bit){
d592f67f
MN
158 c->range -= 2;
159
160 if(!bit){
161 renorm_cabac_encoder(c);
162 }else{
163 c->low += c->range;
164 c->range= 2;
115329f1 165
d592f67f
MN
166 renorm_cabac_encoder(c);
167
168 assert(c->low <= 0x1FF);
169 put_cabac_bit(c, c->low>>9);
170 put_bits(&c->pb, 2, ((c->low>>7)&3)|1);
115329f1 171
d592f67f
MN
172 flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong
173 }
115329f1 174
d592f67f
MN
175#ifdef STRICT_LIMITS
176 c->symCount++;
177#endif
5e20f836 178
b46243ed 179 return (put_bits_count(&c->pb)+7)>>3;
d592f67f
MN
180}
181
61ccfcc0
MN
182/**
183 * put (truncated) unary binarization.
184 */
938dd846 185static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){
61ccfcc0 186 int i;
115329f1 187
61ccfcc0 188 assert(v <= max);
115329f1 189
61ccfcc0
MN
190#if 1
191 for(i=0; i<v; i++){
192 put_cabac(c, state, 1);
193 if(i < max_index) state++;
194 }
195 if(truncated==0 || v<max)
196 put_cabac(c, state, 0);
197#else
198 if(v <= max_index){
199 for(i=0; i<v; i++){
200 put_cabac(c, state+i, 1);
201 }
202 if(truncated==0 || v<max)
203 put_cabac(c, state+i, 0);
204 }else{
205 for(i=0; i<=max_index; i++){
206 put_cabac(c, state+i, 1);
207 }
208 for(; i<v; i++){
209 put_cabac(c, state+max_index, 1);
210 }
211 if(truncated==0 || v<max)
212 put_cabac(c, state+max_index, 0);
213 }
214#endif
215}
216
217/**
218 * put unary exp golomb k-th order binarization.
219 */
938dd846 220static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){
61ccfcc0 221 int i;
115329f1 222
61ccfcc0
MN
223 if(v==0)
224 put_cabac(c, state, 0);
225 else{
8f8c0800 226 const int sign= v < 0;
115329f1 227
c26abfa5 228 if(is_signed) v= FFABS(v);
115329f1 229
61ccfcc0
MN
230 if(v<max){
231 for(i=0; i<v; i++){
232 put_cabac(c, state, 1);
233 if(i < max_index) state++;
234 }
235
236 put_cabac(c, state, 0);
237 }else{
238 int m= 1<<k;
239
240 for(i=0; i<max; i++){
241 put_cabac(c, state, 1);
242 if(i < max_index) state++;
243 }
244
245 v -= max;
246 while(v >= m){ //FIXME optimize
247 put_cabac_bypass(c, 1);
248 v-= m;
249 m+= m;
250 }
251 put_cabac_bypass(c, 0);
252 while(m>>=1){
253 put_cabac_bypass(c, v&m);
254 }
255 }
256
257 if(is_signed)
258 put_cabac_bypass(c, sign);
259 }
260}
c46e2874 261#endif /* TEST */
61ccfcc0 262
ec7eb896 263static void refill(CABACContext *c){
ec7eb896 264#if CABAC_BITS == 16
2ae7569d 265 c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
ec7eb896
MN
266#else
267 c->low+= c->bytestream[0]<<1;
268#endif
269 c->low -= CABAC_MASK;
270 c->bytestream+= CABAC_BITS/8;
271}
272
b250f9c6 273#if ! ( ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) )
ec7eb896
MN
274static void refill2(CABACContext *c){
275 int i, x;
276
277 x= c->low ^ (c->low-1);
f24a5159 278 i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
ec7eb896
MN
279
280 x= -CABAC_MASK;
115329f1 281
ec7eb896
MN
282#if CABAC_BITS == 16
283 x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
284#else
285 x+= c->bytestream[0]<<1;
286#endif
115329f1 287
ec7eb896
MN
288 c->low += x<<i;
289 c->bytestream+= CABAC_BITS/8;
290}
0f26eec2 291#endif
ec7eb896 292
d592f67f 293static inline void renorm_cabac_decoder(CABACContext *c){
f24a5159 294 while(c->range < 0x100){
d592f67f
MN
295 c->range+= c->range;
296 c->low+= c->low;
ec7eb896
MN
297 if(!(c->low & CABAC_MASK))
298 refill(c);
d592f67f
MN
299 }
300}
301
ec7eb896 302static inline void renorm_cabac_decoder_once(CABACContext *c){
400d0f8e 303#ifdef ARCH_X86_DISABLED
ec8f483a
MN
304 int temp;
305#if 0
4310580d 306 //P3:683 athlon:475
be449fca 307 __asm__(
f24a5159 308 "lea -0x100(%0), %2 \n\t"
ec8f483a
MN
309 "shr $31, %2 \n\t" //FIXME 31->63 for x86-64
310 "shl %%cl, %0 \n\t"
311 "shl %%cl, %1 \n\t"
312 : "+r"(c->range), "+r"(c->low), "+c"(temp)
313 );
314#elif 0
4310580d 315 //P3:680 athlon:474
be449fca 316 __asm__(
f24a5159 317 "cmp $0x100, %0 \n\t"
ec8f483a
MN
318 "setb %%cl \n\t" //FIXME 31->63 for x86-64
319 "shl %%cl, %0 \n\t"
320 "shl %%cl, %1 \n\t"
321 : "+r"(c->range), "+r"(c->low), "+c"(temp)
322 );
323#elif 1
324 int temp2;
4310580d 325 //P3:665 athlon:517
be449fca 326 __asm__(
f24a5159 327 "lea -0x100(%0), %%eax \n\t"
7e14b808 328 "cltd \n\t"
ec8f483a
MN
329 "mov %0, %%eax \n\t"
330 "and %%edx, %0 \n\t"
331 "and %1, %%edx \n\t"
332 "add %%eax, %0 \n\t"
333 "add %%edx, %1 \n\t"
334 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
335 );
336#elif 0
337 int temp2;
4310580d 338 //P3:673 athlon:509
be449fca 339 __asm__(
f24a5159 340 "cmp $0x100, %0 \n\t"
ec8f483a
MN
341 "sbb %%edx, %%edx \n\t"
342 "mov %0, %%eax \n\t"
343 "and %%edx, %0 \n\t"
344 "and %1, %%edx \n\t"
345 "add %%eax, %0 \n\t"
346 "add %%edx, %1 \n\t"
347 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
348 );
349#else
350 int temp2;
4310580d 351 //P3:677 athlon:511
be449fca 352 __asm__(
f24a5159 353 "cmp $0x100, %0 \n\t"
ec8f483a
MN
354 "lea (%0, %0), %%eax \n\t"
355 "lea (%1, %1), %%edx \n\t"
356 "cmovb %%eax, %0 \n\t"
357 "cmovb %%edx, %1 \n\t"
358 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
359 );
360#endif
361#else
4310580d 362 //P3:675 athlon:476
f24a5159 363 int shift= (uint32_t)(c->range - 0x100)>>31;
bfe328ca
LM
364 c->range<<= shift;
365 c->low <<= shift;
ec8f483a 366#endif
ec7eb896
MN
367 if(!(c->low & CABAC_MASK))
368 refill(c);
369}
370
5a6a6cc7 371static av_always_inline int get_cabac_inline(CABACContext *c, uint8_t * const state){
bfe328ca 372 //FIXME gcc generates duplicate load/stores for c->low and c->range
f7d0b683
MN
373#define LOW "0"
374#define RANGE "4"
b250f9c6 375#if ARCH_X86_64
755073fe
RD
376#define BYTESTART "16"
377#define BYTE "24"
378#define BYTEEND "32"
379#else
d61c4e73
MN
380#define BYTESTART "12"
381#define BYTE "16"
382#define BYTEEND "20"
755073fe 383#endif
b250f9c6 384#if ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS)
ba9fb5da
BR
385 int bit;
386
0bc2e7f0 387#ifndef BRANCHLESS_CABAC_DECODER
be449fca 388 __asm__ volatile(
4041a495 389 "movzbl (%1), %0 \n\t"
f7d0b683
MN
390 "movl "RANGE "(%2), %%ebx \n\t"
391 "movl "RANGE "(%2), %%edx \n\t"
f24a5159 392 "andl $0xC0, %%ebx \n\t"
4041a495 393 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t"
f7d0b683
MN
394 "movl "LOW "(%2), %%ebx \n\t"
395//eax:state ebx:low, edx:range, esi:RangeLPS
396 "subl %%esi, %%edx \n\t"
f24a5159
MN
397 "movl %%edx, %%ecx \n\t"
398 "shll $17, %%ecx \n\t"
399 "cmpl %%ecx, %%ebx \n\t"
f7d0b683 400 " ja 1f \n\t"
1f4d5e9f
MN
401
402#if 1
403 //athlon:4067 P3:4110
f24a5159 404 "lea -0x100(%%edx), %%ecx \n\t"
1f4d5e9f
MN
405 "shr $31, %%ecx \n\t"
406 "shl %%cl, %%edx \n\t"
407 "shl %%cl, %%ebx \n\t"
408#else
409 //athlon:4057 P3:4130
f24a5159 410 "cmp $0x100, %%edx \n\t" //FIXME avoidable
f7d0b683
MN
411 "setb %%cl \n\t"
412 "shl %%cl, %%edx \n\t"
413 "shl %%cl, %%ebx \n\t"
1f4d5e9f 414#endif
4041a495 415 "movzbl "MANGLE(ff_h264_mps_state)"(%0), %%ecx \n\t"
f7d0b683
MN
416 "movb %%cl, (%1) \n\t"
417//eax:state ebx:low, edx:range, esi:RangeLPS
418 "test %%bx, %%bx \n\t"
419 " jnz 2f \n\t"
755073fe 420 "mov "BYTE "(%2), %%"REG_S" \n\t"
f7d0b683 421 "subl $0xFFFF, %%ebx \n\t"
755073fe 422 "movzwl (%%"REG_S"), %%ecx \n\t"
f7d0b683
MN
423 "bswap %%ecx \n\t"
424 "shrl $15, %%ecx \n\t"
755073fe 425 "add $2, %%"REG_S" \n\t"
f7d0b683 426 "addl %%ecx, %%ebx \n\t"
755073fe 427 "mov %%"REG_S", "BYTE "(%2) \n\t"
f7d0b683
MN
428 "jmp 2f \n\t"
429 "1: \n\t"
430//eax:state ebx:low, edx:range, esi:RangeLPS
f24a5159 431 "subl %%ecx, %%ebx \n\t"
f7d0b683 432 "movl %%esi, %%edx \n\t"
a6672acf 433 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \n\t"
f7d0b683
MN
434 "shll %%cl, %%ebx \n\t"
435 "shll %%cl, %%edx \n\t"
4041a495 436 "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx \n\t"
f24a5159 437 "movb %%cl, (%1) \n\t"
755073fe 438 "add $1, %0 \n\t"
f7d0b683
MN
439 "test %%bx, %%bx \n\t"
440 " jnz 2f \n\t"
441
755073fe
RD
442 "mov "BYTE "(%2), %%"REG_c" \n\t"
443 "movzwl (%%"REG_c"), %%esi \n\t"
f7d0b683
MN
444 "bswap %%esi \n\t"
445 "shrl $15, %%esi \n\t"
446 "subl $0xFFFF, %%esi \n\t"
755073fe
RD
447 "add $2, %%"REG_c" \n\t"
448 "mov %%"REG_c", "BYTE "(%2) \n\t"
f7d0b683
MN
449
450 "leal -1(%%ebx), %%ecx \n\t"
451 "xorl %%ebx, %%ecx \n\t"
f24a5159 452 "shrl $15, %%ecx \n\t"
a6672acf 453 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t"
d17faef0
MN
454 "neg %%ecx \n\t"
455 "add $7, %%ecx \n\t"
f7d0b683
MN
456
457 "shll %%cl , %%esi \n\t"
458 "addl %%esi, %%ebx \n\t"
459 "2: \n\t"
460 "movl %%edx, "RANGE "(%2) \n\t"
461 "movl %%ebx, "LOW "(%2) \n\t"
df3a80b5 462 :"=&a"(bit) //FIXME this is fragile gcc either runs out of registers or miscompiles it (for example if "+a"(bit) or "+m"(*state) is used
f7d0b683 463 :"r"(state), "r"(c)
755073fe 464 : "%"REG_c, "%ebx", "%edx", "%"REG_S, "memory"
f7d0b683 465 );
9ed92c65 466 bit&=1;
a0490b32 467#else /* BRANCHLESS_CABAC_DECODER */
13404b2e
MN
468
469
b250f9c6 470#if HAVE_FAST_CMOV
13404b2e
MN
471#define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
472 "mov "tmp" , %%ecx \n\t"\
473 "shl $17 , "tmp" \n\t"\
474 "cmp "low" , "tmp" \n\t"\
475 "cmova %%ecx , "range" \n\t"\
476 "sbb %%ecx , %%ecx \n\t"\
477 "and %%ecx , "tmp" \n\t"\
478 "sub "tmp" , "low" \n\t"\
479 "xor %%ecx , "ret" \n\t"
7073e9fc 480#else /* HAVE_FAST_CMOV */
13404b2e
MN
481#define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
482 "mov "tmp" , %%ecx \n\t"\
483 "shl $17 , "tmp" \n\t"\
484 "sub "low" , "tmp" \n\t"\
485 "sar $31 , "tmp" \n\t" /*lps_mask*/\
486 "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\
487 "and "tmp" , "range" \n\t" /*(RangeLPS - range)&lps_mask*/\
488 "add %%ecx , "range" \n\t" /*new range*/\
489 "shl $17 , %%ecx \n\t"\
490 "and "tmp" , %%ecx \n\t"\
491 "sub %%ecx , "low" \n\t"\
492 "xor "tmp" , "ret" \n\t"
7073e9fc 493#endif /* HAVE_FAST_CMOV */
ef0090a9 494
ef0090a9 495
13404b2e
MN
496#define BRANCHLESS_GET_CABAC(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
497 "movzbl "statep" , "ret" \n\t"\
498 "mov "range" , "tmp" \n\t"\
499 "and $0xC0 , "range" \n\t"\
500 "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
501 "sub "range" , "tmp" \n\t"\
502 BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
503 "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx \n\t"\
504 "shl %%cl , "range" \n\t"\
505 "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp" \n\t"\
506 "mov "tmpbyte" , "statep" \n\t"\
507 "shl %%cl , "low" \n\t"\
508 "test "lowword" , "lowword" \n\t"\
509 " jnz 1f \n\t"\
755073fe
RD
510 "mov "BYTE"("cabac"), %%"REG_c" \n\t"\
511 "movzwl (%%"REG_c") , "tmp" \n\t"\
13404b2e
MN
512 "bswap "tmp" \n\t"\
513 "shr $15 , "tmp" \n\t"\
514 "sub $0xFFFF , "tmp" \n\t"\
755073fe
RD
515 "add $2 , %%"REG_c" \n\t"\
516 "mov %%"REG_c" , "BYTE "("cabac") \n\t"\
13404b2e
MN
517 "lea -1("low") , %%ecx \n\t"\
518 "xor "low" , %%ecx \n\t"\
519 "shr $15 , %%ecx \n\t"\
520 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t"\
521 "neg %%ecx \n\t"\
522 "add $7 , %%ecx \n\t"\
523 "shl %%cl , "tmp" \n\t"\
524 "add "tmp" , "low" \n\t"\
525 "1: \n\t"
ef0090a9 526
be449fca 527 __asm__ volatile(
13404b2e
MN
528 "movl "RANGE "(%2), %%esi \n\t"
529 "movl "LOW "(%2), %%ebx \n\t"
530 BRANCHLESS_GET_CABAC("%0", "%2", "(%1)", "%%ebx", "%%bx", "%%esi", "%%edx", "%%dl")
f24a5159 531 "movl %%esi, "RANGE "(%2) \n\t"
b99f3cab 532 "movl %%ebx, "LOW "(%2) \n\t"
ef0090a9 533
ef0090a9
MN
534 :"=&a"(bit)
535 :"r"(state), "r"(c)
755073fe 536 : "%"REG_c, "%ebx", "%edx", "%esi", "memory"
ef0090a9 537 );
f1b37db4 538 bit&=1;
a0490b32 539#endif /* BRANCHLESS_CABAC_DECODER */
b250f9c6 540#else /* ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) */
bfe328ca 541 int s = *state;
a0f2c6ba 542 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s];
154e30f6 543 int bit, lps_mask av_unused;
115329f1 544
d592f67f 545 c->range -= RangeLPS;
0bc2e7f0 546#ifndef BRANCHLESS_CABAC_DECODER
3b6dc9ca 547 if(c->low < (c->range<<(CABAC_BITS+1))){
bfe328ca 548 bit= s&1;
d61c4e73 549 *state= ff_h264_mps_state[s];
ec7eb896 550 renorm_cabac_decoder_once(c);
d592f67f 551 }else{
f24a5159 552 bit= ff_h264_norm_shift[RangeLPS];
3b6dc9ca 553 c->low -= (c->range<<(CABAC_BITS+1));
d61c4e73 554 *state= ff_h264_lps_state[s];
260ceb63
MN
555 c->range = RangeLPS<<bit;
556 c->low <<= bit;
557 bit= (s&1)^1;
558
3b6dc9ca 559 if(!(c->low & CABAC_MASK)){
ec7eb896 560 refill2(c);
260ceb63 561 }
d592f67f 562 }
a0490b32 563#else /* BRANCHLESS_CABAC_DECODER */
3b6dc9ca 564 lps_mask= ((c->range<<(CABAC_BITS+1)) - c->low)>>31;
115329f1 565
3b6dc9ca 566 c->low -= (c->range<<(CABAC_BITS+1)) & lps_mask;
ec7eb896 567 c->range += (RangeLPS - c->range) & lps_mask;
115329f1 568
2e1aee80 569 s^=lps_mask;
68a205ed 570 *state= (ff_h264_mlps_state+128)[s];
2e1aee80 571 bit= s&1;
115329f1 572
f24a5159 573 lps_mask= ff_h264_norm_shift[c->range];
ec7eb896
MN
574 c->range<<= lps_mask;
575 c->low <<= lps_mask;
576 if(!(c->low & CABAC_MASK))
577 refill2(c);
a0490b32 578#endif /* BRANCHLESS_CABAC_DECODER */
b250f9c6 579#endif /* ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) */
115329f1 580 return bit;
d592f67f
MN
581}
582
ab210908 583static int av_noinline av_unused get_cabac_noinline(CABACContext *c, uint8_t * const state){
851ded89
MN
584 return get_cabac_inline(c,state);
585}
586
ab210908 587static int av_unused get_cabac(CABACContext *c, uint8_t * const state){
851ded89
MN
588 return get_cabac_inline(c,state);
589}
590
ab210908 591static int av_unused get_cabac_bypass(CABACContext *c){
ebd624b6
MN
592#if 0 //not faster
593 int bit;
be449fca 594 __asm__ volatile(
ebd624b6
MN
595 "movl "RANGE "(%1), %%ebx \n\t"
596 "movl "LOW "(%1), %%eax \n\t"
597 "shl $17, %%ebx \n\t"
598 "add %%eax, %%eax \n\t"
599 "sub %%ebx, %%eax \n\t"
7e14b808 600 "cltd \n\t"
ebd624b6
MN
601 "and %%edx, %%ebx \n\t"
602 "add %%ebx, %%eax \n\t"
603 "test %%ax, %%ax \n\t"
604 " jnz 1f \n\t"
755073fe 605 "movl "BYTE "(%1), %%"REG_b" \n\t"
ebd624b6 606 "subl $0xFFFF, %%eax \n\t"
755073fe 607 "movzwl (%%"REG_b"), %%ecx \n\t"
ebd624b6
MN
608 "bswap %%ecx \n\t"
609 "shrl $15, %%ecx \n\t"
755073fe 610 "addl $2, %%"REG_b" \n\t"
ebd624b6 611 "addl %%ecx, %%eax \n\t"
755073fe 612 "movl %%"REG_b", "BYTE "(%1) \n\t"
ebd624b6
MN
613 "1: \n\t"
614 "movl %%eax, "LOW "(%1) \n\t"
615
616 :"=&d"(bit)
617 :"r"(c)
755073fe 618 : "%eax", "%"REG_b, "%ecx", "memory"
ebd624b6
MN
619 );
620 return bit+1;
621#else
f24a5159 622 int range;
d592f67f
MN
623 c->low += c->low;
624
ec7eb896
MN
625 if(!(c->low & CABAC_MASK))
626 refill(c);
115329f1 627
3b6dc9ca 628 range= c->range<<(CABAC_BITS+1);
f24a5159 629 if(c->low < range){
d592f67f
MN
630 return 0;
631 }else{
f24a5159 632 c->low -= range;
d592f67f
MN
633 return 1;
634 }
ebd624b6
MN
635#endif
636}
637
638
849f1035 639static av_always_inline int get_cabac_bypass_sign(CABACContext *c, int val){
b250f9c6 640#if ARCH_X86 && !(defined(PIC) && defined(__GNUC__))
be449fca 641 __asm__ volatile(
ebd624b6
MN
642 "movl "RANGE "(%1), %%ebx \n\t"
643 "movl "LOW "(%1), %%eax \n\t"
644 "shl $17, %%ebx \n\t"
645 "add %%eax, %%eax \n\t"
646 "sub %%ebx, %%eax \n\t"
7e14b808 647 "cltd \n\t"
ebd624b6
MN
648 "and %%edx, %%ebx \n\t"
649 "add %%ebx, %%eax \n\t"
650 "xor %%edx, %%ecx \n\t"
651 "sub %%edx, %%ecx \n\t"
652 "test %%ax, %%ax \n\t"
653 " jnz 1f \n\t"
755073fe 654 "mov "BYTE "(%1), %%"REG_b" \n\t"
ebd624b6 655 "subl $0xFFFF, %%eax \n\t"
755073fe 656 "movzwl (%%"REG_b"), %%edx \n\t"
ebd624b6
MN
657 "bswap %%edx \n\t"
658 "shrl $15, %%edx \n\t"
755073fe 659 "add $2, %%"REG_b" \n\t"
ebd624b6 660 "addl %%edx, %%eax \n\t"
755073fe 661 "mov %%"REG_b", "BYTE "(%1) \n\t"
ebd624b6
MN
662 "1: \n\t"
663 "movl %%eax, "LOW "(%1) \n\t"
664
665 :"+c"(val)
666 :"r"(c)
755073fe 667 : "%eax", "%"REG_b, "%edx", "memory"
ebd624b6
MN
668 );
669 return val;
670#else
671 int range, mask;
672 c->low += c->low;
673
674 if(!(c->low & CABAC_MASK))
675 refill(c);
676
3b6dc9ca 677 range= c->range<<(CABAC_BITS+1);
ebd624b6
MN
678 c->low -= range;
679 mask= c->low >> 31;
680 range &= mask;
681 c->low += range;
682 return (val^mask)-mask;
683#endif
d592f67f 684}
ebd624b6 685
5e20f836
MN
686/**
687 *
688 * @return the number of bytes read or 0 if no end
689 */
ab210908 690static int av_unused get_cabac_terminate(CABACContext *c){
f24a5159 691 c->range -= 2;
3b6dc9ca 692 if(c->low < c->range<<(CABAC_BITS+1)){
ec7eb896 693 renorm_cabac_decoder_once(c);
d592f67f
MN
694 return 0;
695 }else{
5e20f836 696 return c->bytestream - c->bytestream_start;
115329f1 697 }
d592f67f
MN
698}
699
0f26eec2 700#if 0
61ccfcc0 701/**
1fe2d0fd 702 * Get (truncated) unary binarization.
61ccfcc0 703 */
938dd846 704static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){
61ccfcc0 705 int i;
115329f1
DB
706
707 for(i=0; i<max; i++){
61ccfcc0
MN
708 if(get_cabac(c, state)==0)
709 return i;
115329f1 710
61ccfcc0
MN
711 if(i< max_index) state++;
712 }
713
714 return truncated ? max : -1;
715}
716
717/**
718 * get unary exp golomb k-th order binarization.
719 */
938dd846 720static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){
61ccfcc0
MN
721 int i, v;
722 int m= 1<<k;
115329f1
DB
723
724 if(get_cabac(c, state)==0)
61ccfcc0 725 return 0;
115329f1 726
61ccfcc0 727 if(0 < max_index) state++;
115329f1
DB
728
729 for(i=1; i<max; i++){
61ccfcc0
MN
730 if(get_cabac(c, state)==0){
731 if(is_signed && get_cabac_bypass(c)){
732 return -i;
733 }else
734 return i;
735 }
736
737 if(i < max_index) state++;
738 }
115329f1 739
61ccfcc0
MN
740 while(get_cabac_bypass(c)){
741 i+= m;
742 m+= m;
743 }
115329f1 744
61ccfcc0
MN
745 v=0;
746 while(m>>=1){
747 v+= v + get_cabac_bypass(c);
748 }
749 i += v;
750
751 if(is_signed && get_cabac_bypass(c)){
752 return -i;
753 }else
754 return i;
755}
0f26eec2 756#endif /* 0 */
26b4fe82 757
98790382 758#endif /* AVCODEC_CABAC_H */