cabac.h 28.4 KB
Newer Older
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
1 2 3 4
/*
 * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
 *
5 6 7
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
8 9
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
10
 * version 2.1 of the License, or (at your option) any later version.
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
11
 *
12
 * FFmpeg is distributed in the hope that it will be useful,
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
13 14 15 16 17
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
18
 * License along with FFmpeg; if not, write to the Free Software
19
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
20 21
 *
 */
22

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
23 24 25 26 27 28
/**
 * @file cabac.h
 * Context Adaptive Binary Arithmetic Coder.
 */


29
//#undef NDEBUG
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
30
#include <assert.h>
31 32 33
#ifdef ARCH_X86
#include "x86_cpu.h"
#endif
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
34

35
#define CABAC_BITS 16
Michael Niedermayer's avatar
Michael Niedermayer committed
36
#define CABAC_MASK ((1<<CABAC_BITS)-1)
37
#define BRANCHLESS_CABAC_DECODER 1
38
//#define ARCH_X86_DISABLED 1
Michael Niedermayer's avatar
Michael Niedermayer committed
39

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
40 41 42 43 44 45 46
typedef struct CABACContext{
    int low;
    int range;
    int outstanding_count;
#ifdef STRICT_LIMITS
    int symCount;
#endif
47 48
    const uint8_t *bytestream_start;
    const uint8_t *bytestream;
Michael Niedermayer's avatar
Michael Niedermayer committed
49
    const uint8_t *bytestream_end;
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
50 51 52
    PutBitContext pb;
}CABACContext;

53
extern uint8_t ff_h264_mlps_state[4*64];
54
extern uint8_t ff_h264_lps_range[4*2*64];  ///< rangeTabLPS
55 56
extern uint8_t ff_h264_mps_state[2*64];     ///< transIdxMPS
extern uint8_t ff_h264_lps_state[2*64];     ///< transIdxLPS
57
extern const uint8_t ff_h264_norm_shift[512];
Michael Niedermayer's avatar
Michael Niedermayer committed
58

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
59 60

void ff_init_cabac_encoder(CABACContext *c, uint8_t *buf, int buf_size);
61
void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size);
62
void ff_init_cabac_states(CABACContext *c);
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
63 64 65


static inline void put_cabac_bit(CABACContext *c, int b){
66 67
    put_bits(&c->pb, 1, b);
    for(;c->outstanding_count; c->outstanding_count--){
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83
        put_bits(&c->pb, 1, 1-b);
    }
}

static inline void renorm_cabac_encoder(CABACContext *c){
    while(c->range < 0x100){
        //FIXME optimize
        if(c->low<0x100){
            put_cabac_bit(c, 0);
        }else if(c->low<0x200){
            c->outstanding_count++;
            c->low -= 0x100;
        }else{
            put_cabac_bit(c, 1);
            c->low -= 0x200;
        }
84

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
85 86 87 88 89
        c->range+= c->range;
        c->low += c->low;
    }
}

90
static void put_cabac(CABACContext *c, uint8_t * const state, int bit){
91
    int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state];
92

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
93 94
    if(bit == ((*state)&1)){
        c->range -= RangeLPS;
95
        *state= ff_h264_mps_state[*state];
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
96 97 98
    }else{
        c->low += c->range - RangeLPS;
        c->range = RangeLPS;
99
        *state= ff_h264_lps_state[*state];
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
100
    }
101

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
102 103 104 105 106 107 108
    renorm_cabac_encoder(c);

#ifdef STRICT_LIMITS
    c->symCount++;
#endif
}

109
static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
    assert(c->range > RangeLPS);

    if(!bit){
        c->range -= RangeLPS;
    }else{
        c->low += c->range - RangeLPS;
        c->range = RangeLPS;
    }

    renorm_cabac_encoder(c);

#ifdef STRICT_LIMITS
    c->symCount++;
#endif
}

126 127 128
/**
 * @param bit 0 -> write zero bit, !=0 write one bit
 */
129
static void put_cabac_bypass(CABACContext *c, int bit){
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
130 131 132 133 134 135 136 137 138 139 140 141 142 143 144
    c->low += c->low;

    if(bit){
        c->low += c->range;
    }
//FIXME optimize
    if(c->low<0x200){
        put_cabac_bit(c, 0);
    }else if(c->low<0x400){
        c->outstanding_count++;
        c->low -= 0x200;
    }else{
        put_cabac_bit(c, 1);
        c->low -= 0x400;
    }
145

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
146 147 148 149 150
#ifdef STRICT_LIMITS
    c->symCount++;
#endif
}

151 152 153 154
/**
 *
 * @return the number of bytes written
 */
155
static int put_cabac_terminate(CABACContext *c, int bit){
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
156 157 158 159 160 161 162
    c->range -= 2;

    if(!bit){
        renorm_cabac_encoder(c);
    }else{
        c->low += c->range;
        c->range= 2;
163

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
164 165 166 167 168
        renorm_cabac_encoder(c);

        assert(c->low <= 0x1FF);
        put_cabac_bit(c, c->low>>9);
        put_bits(&c->pb, 2, ((c->low>>7)&3)|1);
169

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
170 171
        flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong
    }
172

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
173 174 175
#ifdef STRICT_LIMITS
    c->symCount++;
#endif
176

177
    return (put_bits_count(&c->pb)+7)>>3;
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
178 179
}

180 181 182
/**
 * put (truncated) unary binarization.
 */
183
static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){
184
    int i;
185

186
    assert(v <= max);
187

188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217
#if 1
    for(i=0; i<v; i++){
        put_cabac(c, state, 1);
        if(i < max_index) state++;
    }
    if(truncated==0 || v<max)
        put_cabac(c, state, 0);
#else
    if(v <= max_index){
        for(i=0; i<v; i++){
            put_cabac(c, state+i, 1);
        }
        if(truncated==0 || v<max)
            put_cabac(c, state+i, 0);
    }else{
        for(i=0; i<=max_index; i++){
            put_cabac(c, state+i, 1);
        }
        for(; i<v; i++){
            put_cabac(c, state+max_index, 1);
        }
        if(truncated==0 || v<max)
            put_cabac(c, state+max_index, 0);
    }
#endif
}

/**
 * put unary exp golomb k-th order binarization.
 */
218
static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){
219
    int i;
220

221 222 223
    if(v==0)
        put_cabac(c, state, 0);
    else{
Michael Niedermayer's avatar
cleanup  
Michael Niedermayer committed
224
        const int sign= v < 0;
225

Diego Biurrun's avatar
Diego Biurrun committed
226
        if(is_signed) v= FFABS(v);
227

228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259
        if(v<max){
            for(i=0; i<v; i++){
                put_cabac(c, state, 1);
                if(i < max_index) state++;
            }

            put_cabac(c, state, 0);
        }else{
            int m= 1<<k;

            for(i=0; i<max; i++){
                put_cabac(c, state, 1);
                if(i < max_index) state++;
            }

            v -= max;
            while(v >= m){ //FIXME optimize
                put_cabac_bypass(c, 1);
                v-= m;
                m+= m;
            }
            put_cabac_bypass(c, 0);
            while(m>>=1){
                put_cabac_bypass(c, v&m);
            }
        }

        if(is_signed)
            put_cabac_bypass(c, sign);
    }
}

Michael Niedermayer's avatar
Michael Niedermayer committed
260 261
static void refill(CABACContext *c){
#if CABAC_BITS == 16
Michael Niedermayer's avatar
() 10l  
Michael Niedermayer committed
262
        c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
Michael Niedermayer's avatar
Michael Niedermayer committed
263 264 265 266 267 268 269 270 271 272 273
#else
        c->low+= c->bytestream[0]<<1;
#endif
    c->low -= CABAC_MASK;
    c->bytestream+= CABAC_BITS/8;
}

static void refill2(CABACContext *c){
    int i, x;

    x= c->low ^ (c->low-1);
274
    i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
Michael Niedermayer's avatar
Michael Niedermayer committed
275 276

    x= -CABAC_MASK;
277

Michael Niedermayer's avatar
Michael Niedermayer committed
278 279 280 281 282
#if CABAC_BITS == 16
        x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
#else
        x+= c->bytestream[0]<<1;
#endif
283

Michael Niedermayer's avatar
Michael Niedermayer committed
284 285 286 287
    c->low += x<<i;
    c->bytestream+= CABAC_BITS/8;
}

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
288
static inline void renorm_cabac_decoder(CABACContext *c){
289
    while(c->range < 0x100){
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
290 291
        c->range+= c->range;
        c->low+= c->low;
Michael Niedermayer's avatar
Michael Niedermayer committed
292 293
        if(!(c->low & CABAC_MASK))
            refill(c);
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
294 295 296
    }
}

Michael Niedermayer's avatar
Michael Niedermayer committed
297
static inline void renorm_cabac_decoder_once(CABACContext *c){
298
#ifdef ARCH_X86_DISABLED
299 300
    int temp;
#if 0
301
    //P3:683    athlon:475
302
    asm(
303
        "lea -0x100(%0), %2         \n\t"
304 305 306 307 308 309
        "shr $31, %2                \n\t"  //FIXME 31->63 for x86-64
        "shl %%cl, %0               \n\t"
        "shl %%cl, %1               \n\t"
        : "+r"(c->range), "+r"(c->low), "+c"(temp)
    );
#elif 0
310
    //P3:680    athlon:474
311
    asm(
312
        "cmp $0x100, %0             \n\t"
313 314 315 316 317 318 319
        "setb %%cl                  \n\t"  //FIXME 31->63 for x86-64
        "shl %%cl, %0               \n\t"
        "shl %%cl, %1               \n\t"
        : "+r"(c->range), "+r"(c->low), "+c"(temp)
    );
#elif 1
    int temp2;
320
    //P3:665    athlon:517
321
    asm(
322
        "lea -0x100(%0), %%eax      \n\t"
323 324 325 326 327 328 329 330 331 332
        "cdq                        \n\t"
        "mov %0, %%eax              \n\t"
        "and %%edx, %0              \n\t"
        "and %1, %%edx              \n\t"
        "add %%eax, %0              \n\t"
        "add %%edx, %1              \n\t"
        : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
    );
#elif 0
    int temp2;
333
    //P3:673    athlon:509
334
    asm(
335
        "cmp $0x100, %0             \n\t"
336 337 338 339 340 341 342 343 344 345
        "sbb %%edx, %%edx           \n\t"
        "mov %0, %%eax              \n\t"
        "and %%edx, %0              \n\t"
        "and %1, %%edx              \n\t"
        "add %%eax, %0              \n\t"
        "add %%edx, %1              \n\t"
        : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
    );
#else
    int temp2;
346
    //P3:677    athlon:511
347
    asm(
348
        "cmp $0x100, %0             \n\t"
349 350 351 352 353 354 355 356
        "lea (%0, %0), %%eax        \n\t"
        "lea (%1, %1), %%edx        \n\t"
        "cmovb %%eax, %0            \n\t"
        "cmovb %%edx, %1            \n\t"
        : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
    );
#endif
#else
357
    //P3:675    athlon:476
358
    int shift= (uint32_t)(c->range - 0x100)>>31;
Loren Merritt's avatar
Loren Merritt committed
359 360
    c->range<<= shift;
    c->low  <<= shift;
361
#endif
Michael Niedermayer's avatar
Michael Niedermayer committed
362 363 364 365
    if(!(c->low & CABAC_MASK))
        refill(c);
}

366
static int av_always_inline get_cabac_inline(CABACContext *c, uint8_t * const state){
Loren Merritt's avatar
Loren Merritt committed
367
    //FIXME gcc generates duplicate load/stores for c->low and c->range
368 369
#define LOW          "0"
#define RANGE        "4"
370 371 372 373 374
#ifdef ARCH_X86_64
#define BYTESTART   "16"
#define BYTE        "24"
#define BYTEEND     "32"
#else
375 376 377
#define BYTESTART   "12"
#define BYTE        "16"
#define BYTEEND     "20"
378
#endif
379
#if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
380 381
    int bit;

382
#ifndef BRANCHLESS_CABAC_DECODER
383
    asm volatile(
Michael Niedermayer's avatar
Michael Niedermayer committed
384
        "movzbl (%1), %0                        \n\t"
385 386
        "movl "RANGE    "(%2), %%ebx            \n\t"
        "movl "RANGE    "(%2), %%edx            \n\t"
387
        "andl $0xC0, %%ebx                      \n\t"
Michael Niedermayer's avatar
Michael Niedermayer committed
388
        "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t"
389 390 391
        "movl "LOW      "(%2), %%ebx            \n\t"
//eax:state ebx:low, edx:range, esi:RangeLPS
        "subl %%esi, %%edx                      \n\t"
392 393 394
        "movl %%edx, %%ecx                      \n\t"
        "shll $17, %%ecx                        \n\t"
        "cmpl %%ecx, %%ebx                      \n\t"
395
        " ja 1f                                 \n\t"
396 397 398

#if 1
        //athlon:4067 P3:4110
399
        "lea -0x100(%%edx), %%ecx               \n\t"
400 401 402 403 404
        "shr $31, %%ecx                         \n\t"
        "shl %%cl, %%edx                        \n\t"
        "shl %%cl, %%ebx                        \n\t"
#else
        //athlon:4057 P3:4130
405
        "cmp $0x100, %%edx                      \n\t" //FIXME avoidable
406 407 408
        "setb %%cl                              \n\t"
        "shl %%cl, %%edx                        \n\t"
        "shl %%cl, %%ebx                        \n\t"
409
#endif
Michael Niedermayer's avatar
Michael Niedermayer committed
410
        "movzbl "MANGLE(ff_h264_mps_state)"(%0), %%ecx   \n\t"
411 412 413 414
        "movb %%cl, (%1)                        \n\t"
//eax:state ebx:low, edx:range, esi:RangeLPS
        "test %%bx, %%bx                        \n\t"
        " jnz 2f                                \n\t"
415
        "mov  "BYTE     "(%2), %%"REG_S"        \n\t"
416
        "subl $0xFFFF, %%ebx                    \n\t"
417
        "movzwl (%%"REG_S"), %%ecx              \n\t"
418 419
        "bswap %%ecx                            \n\t"
        "shrl $15, %%ecx                        \n\t"
420
        "add  $2, %%"REG_S"                     \n\t"
421
        "addl %%ecx, %%ebx                      \n\t"
422
        "mov  %%"REG_S", "BYTE    "(%2)         \n\t"
423 424 425
        "jmp 2f                                 \n\t"
        "1:                                     \n\t"
//eax:state ebx:low, edx:range, esi:RangeLPS
426
        "subl %%ecx, %%ebx                      \n\t"
427
        "movl %%esi, %%edx                      \n\t"
428
        "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx   \n\t"
429 430
        "shll %%cl, %%ebx                       \n\t"
        "shll %%cl, %%edx                       \n\t"
Michael Niedermayer's avatar
Michael Niedermayer committed
431
        "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx   \n\t"
432
        "movb %%cl, (%1)                        \n\t"
433
        "add  $1, %0                            \n\t"
434 435 436
        "test %%bx, %%bx                        \n\t"
        " jnz 2f                                \n\t"

437 438
        "mov  "BYTE     "(%2), %%"REG_c"        \n\t"
        "movzwl (%%"REG_c"), %%esi              \n\t"
439 440 441
        "bswap %%esi                            \n\t"
        "shrl $15, %%esi                        \n\t"
        "subl $0xFFFF, %%esi                    \n\t"
442 443
        "add  $2, %%"REG_c"                     \n\t"
        "mov  %%"REG_c", "BYTE    "(%2)         \n\t"
444 445 446

        "leal -1(%%ebx), %%ecx                  \n\t"
        "xorl %%ebx, %%ecx                      \n\t"
447
        "shrl $15, %%ecx                        \n\t"
448
        "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx   \n\t"
449 450
        "neg %%ecx                              \n\t"
        "add $7, %%ecx                          \n\t"
451 452 453 454 455 456 457 458

        "shll %%cl , %%esi                      \n\t"
        "addl %%esi, %%ebx                      \n\t"
        "2:                                     \n\t"
        "movl %%edx, "RANGE    "(%2)            \n\t"
        "movl %%ebx, "LOW      "(%2)            \n\t"
        :"=&a"(bit) //FIXME this is fragile gcc either runs out of registers or misscompiles it (for example if "+a"(bit) or "+m"(*state) is used
        :"r"(state), "r"(c)
459
        : "%"REG_c, "%ebx", "%edx", "%"REG_S, "memory"
460
    );
461
    bit&=1;
462
#else /* BRANCHLESS_CABAC_DECODER */
463 464


465
#if defined CMOV_IS_FAST
466 467 468 469 470 471 472 473 474
#define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
        "mov    "tmp"       , %%ecx                                     \n\t"\
        "shl    $17         , "tmp"                                     \n\t"\
        "cmp    "low"       , "tmp"                                     \n\t"\
        "cmova  %%ecx       , "range"                                   \n\t"\
        "sbb    %%ecx       , %%ecx                                     \n\t"\
        "and    %%ecx       , "tmp"                                     \n\t"\
        "sub    "tmp"       , "low"                                     \n\t"\
        "xor    %%ecx       , "ret"                                     \n\t"
475
#else /* CMOV_IS_FAST */
476 477 478 479 480 481 482 483 484 485 486 487
#define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
        "mov    "tmp"       , %%ecx                                     \n\t"\
        "shl    $17         , "tmp"                                     \n\t"\
        "sub    "low"       , "tmp"                                     \n\t"\
        "sar    $31         , "tmp"                                     \n\t" /*lps_mask*/\
        "sub    %%ecx       , "range"                                   \n\t" /*RangeLPS - range*/\
        "and    "tmp"       , "range"                                   \n\t" /*(RangeLPS - range)&lps_mask*/\
        "add    %%ecx       , "range"                                   \n\t" /*new range*/\
        "shl    $17         , %%ecx                                     \n\t"\
        "and    "tmp"       , %%ecx                                     \n\t"\
        "sub    %%ecx       , "low"                                     \n\t"\
        "xor    "tmp"       , "ret"                                     \n\t"
488
#endif /* CMOV_IS_FAST */
489 490


491 492 493 494 495 496 497 498 499 500 501 502 503 504
#define BRANCHLESS_GET_CABAC(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
        "movzbl "statep"    , "ret"                                     \n\t"\
        "mov    "range"     , "tmp"                                     \n\t"\
        "and    $0xC0       , "range"                                   \n\t"\
        "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
        "sub    "range"     , "tmp"                                     \n\t"\
        BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
        "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx          \n\t"\
        "shl    %%cl        , "range"                                   \n\t"\
        "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp"          \n\t"\
        "mov    "tmpbyte"   , "statep"                                  \n\t"\
        "shl    %%cl        , "low"                                     \n\t"\
        "test   "lowword"   , "lowword"                                 \n\t"\
        " jnz   1f                                                      \n\t"\
505 506
        "mov "BYTE"("cabac"), %%"REG_c"                                 \n\t"\
        "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
507 508 509
        "bswap  "tmp"                                                   \n\t"\
        "shr    $15         , "tmp"                                     \n\t"\
        "sub    $0xFFFF     , "tmp"                                     \n\t"\
510 511
        "add    $2          , %%"REG_c"                                 \n\t"\
        "mov    %%"REG_c"   , "BYTE    "("cabac")                       \n\t"\
512 513 514 515 516 517 518 519 520
        "lea    -1("low")   , %%ecx                                     \n\t"\
        "xor    "low"       , %%ecx                                     \n\t"\
        "shr    $15         , %%ecx                                     \n\t"\
        "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx            \n\t"\
        "neg    %%ecx                                                   \n\t"\
        "add    $7          , %%ecx                                     \n\t"\
        "shl    %%cl        , "tmp"                                     \n\t"\
        "add    "tmp"       , "low"                                     \n\t"\
        "1:                                                             \n\t"
521

522 523 524 525
    asm volatile(
        "movl "RANGE    "(%2), %%esi            \n\t"
        "movl "LOW      "(%2), %%ebx            \n\t"
        BRANCHLESS_GET_CABAC("%0", "%2", "(%1)", "%%ebx", "%%bx", "%%esi", "%%edx", "%%dl")
526
        "movl %%esi, "RANGE    "(%2)            \n\t"
527
        "movl %%ebx, "LOW      "(%2)            \n\t"
528 529 530

        :"=&a"(bit)
        :"r"(state), "r"(c)
531
        : "%"REG_c, "%ebx", "%edx", "%esi", "memory"
532
    );
533
    bit&=1;
534
#endif /* BRANCHLESS_CABAC_DECODER */
535
#else /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
Loren Merritt's avatar
Loren Merritt committed
536
    int s = *state;
537
    int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s];
538
    int bit, lps_mask attribute_unused;
539

Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
540
    c->range -= RangeLPS;
541
#ifndef BRANCHLESS_CABAC_DECODER
542
    if(c->low < (c->range<<(CABAC_BITS+1))){
Loren Merritt's avatar
Loren Merritt committed
543
        bit= s&1;
544
        *state= ff_h264_mps_state[s];
Michael Niedermayer's avatar
Michael Niedermayer committed
545
        renorm_cabac_decoder_once(c);
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
546
    }else{
547
        bit= ff_h264_norm_shift[RangeLPS];
548
        c->low -= (c->range<<(CABAC_BITS+1));
549
        *state= ff_h264_lps_state[s];
550 551 552 553
        c->range = RangeLPS<<bit;
        c->low <<= bit;
        bit= (s&1)^1;

554
        if(!(c->low & CABAC_MASK)){
Michael Niedermayer's avatar
Michael Niedermayer committed
555
            refill2(c);
556
        }
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
557
    }
558
#else /* BRANCHLESS_CABAC_DECODER */
559
    lps_mask= ((c->range<<(CABAC_BITS+1)) - c->low)>>31;
560

561
    c->low -= (c->range<<(CABAC_BITS+1)) & lps_mask;
Michael Niedermayer's avatar
Michael Niedermayer committed
562
    c->range += (RangeLPS - c->range) & lps_mask;
563

564
    s^=lps_mask;
565
    *state= (ff_h264_mlps_state+128)[s];
566
    bit= s&1;
567

568
    lps_mask= ff_h264_norm_shift[c->range];
Michael Niedermayer's avatar
Michael Niedermayer committed
569 570 571 572
    c->range<<= lps_mask;
    c->low  <<= lps_mask;
    if(!(c->low & CABAC_MASK))
        refill2(c);
573
#endif /* BRANCHLESS_CABAC_DECODER */
574
#endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
575
    return bit;
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
576 577
}

578 579 580 581 582 583 584 585
static int __attribute((noinline)) get_cabac_noinline(CABACContext *c, uint8_t * const state){
    return get_cabac_inline(c,state);
}

static int get_cabac(CABACContext *c, uint8_t * const state){
    return get_cabac_inline(c,state);
}

586
static int get_cabac_bypass(CABACContext *c){
587 588 589 590 591 592 593 594 595 596 597 598 599
#if 0 //not faster
    int bit;
    asm volatile(
        "movl "RANGE    "(%1), %%ebx            \n\t"
        "movl "LOW      "(%1), %%eax            \n\t"
        "shl $17, %%ebx                         \n\t"
        "add %%eax, %%eax                       \n\t"
        "sub %%ebx, %%eax                       \n\t"
        "cdq                                    \n\t"
        "and %%edx, %%ebx                       \n\t"
        "add %%ebx, %%eax                       \n\t"
        "test %%ax, %%ax                        \n\t"
        " jnz 1f                                \n\t"
600
        "movl "BYTE     "(%1), %%"REG_b"        \n\t"
601
        "subl $0xFFFF, %%eax                    \n\t"
602
        "movzwl (%%"REG_b"), %%ecx              \n\t"
603 604
        "bswap %%ecx                            \n\t"
        "shrl $15, %%ecx                        \n\t"
605
        "addl $2, %%"REG_b"                     \n\t"
606
        "addl %%ecx, %%eax                      \n\t"
607
        "movl %%"REG_b", "BYTE     "(%1)        \n\t"
608 609 610 611 612
        "1:                                     \n\t"
        "movl %%eax, "LOW      "(%1)            \n\t"

        :"=&d"(bit)
        :"r"(c)
613
        : "%eax", "%"REG_b, "%ecx", "memory"
614 615 616
    );
    return bit+1;
#else
617
    int range;
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
618 619
    c->low += c->low;

Michael Niedermayer's avatar
Michael Niedermayer committed
620 621
    if(!(c->low & CABAC_MASK))
        refill(c);
622

623
    range= c->range<<(CABAC_BITS+1);
624
    if(c->low < range){
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
625 626
        return 0;
    }else{
627
        c->low -= range;
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
628 629
        return 1;
    }
630 631 632 633
#endif
}


634
static av_always_inline int get_cabac_bypass_sign(CABACContext *c, int val){
Michael Niedermayer's avatar
PIC fix  
Michael Niedermayer committed
635
#if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
636 637 638 639 640 641 642 643 644 645 646 647 648
    asm volatile(
        "movl "RANGE    "(%1), %%ebx            \n\t"
        "movl "LOW      "(%1), %%eax            \n\t"
        "shl $17, %%ebx                         \n\t"
        "add %%eax, %%eax                       \n\t"
        "sub %%ebx, %%eax                       \n\t"
        "cdq                                    \n\t"
        "and %%edx, %%ebx                       \n\t"
        "add %%ebx, %%eax                       \n\t"
        "xor %%edx, %%ecx                       \n\t"
        "sub %%edx, %%ecx                       \n\t"
        "test %%ax, %%ax                        \n\t"
        " jnz 1f                                \n\t"
649
        "mov  "BYTE     "(%1), %%"REG_b"        \n\t"
650
        "subl $0xFFFF, %%eax                    \n\t"
651
        "movzwl (%%"REG_b"), %%edx              \n\t"
652 653
        "bswap %%edx                            \n\t"
        "shrl $15, %%edx                        \n\t"
654
        "add  $2, %%"REG_b"                     \n\t"
655
        "addl %%edx, %%eax                      \n\t"
656
        "mov  %%"REG_b", "BYTE     "(%1)        \n\t"
657 658 659 660 661
        "1:                                     \n\t"
        "movl %%eax, "LOW      "(%1)            \n\t"

        :"+c"(val)
        :"r"(c)
662
        : "%eax", "%"REG_b, "%edx", "memory"
663 664 665 666 667 668 669 670 671
    );
    return val;
#else
    int range, mask;
    c->low += c->low;

    if(!(c->low & CABAC_MASK))
        refill(c);

672
    range= c->range<<(CABAC_BITS+1);
673 674 675 676 677 678
    c->low -= range;
    mask= c->low >> 31;
    range &= mask;
    c->low += range;
    return (val^mask)-mask;
#endif
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
679
}
680

681 682
//FIXME the x86 code from this file should be moved into i386/h264 or cabac something.c/h (note ill kill you if you move my code away from under my fingers before iam finished with it!)
//FIXME use some macros to avoid duplicatin get_cabac (cant be done yet as that would make optimization work hard)
683
#if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
684 685 686
static int decode_significance_x86(CABACContext *c, int max_coeff, uint8_t *significant_coeff_ctx_base, int *index){
    void *end= significant_coeff_ctx_base + max_coeff - 1;
    int minusstart= -(int)significant_coeff_ctx_base;
687
    int minusindex= 4-(int)index;
688 689 690 691 692 693 694
    int coeff_count;
    asm volatile(
        "movl "RANGE    "(%3), %%esi            \n\t"
        "movl "LOW      "(%3), %%ebx            \n\t"

        "2:                                     \n\t"

695
        BRANCHLESS_GET_CABAC("%%edx", "%3", "(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
696

697
        "test $1, %%edx                         \n\t"
698 699
        " jz 3f                                 \n\t"

700 701
        BRANCHLESS_GET_CABAC("%%edx", "%3", "61(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")

702
        "mov  %2, %%"REG_a"                     \n\t"
703
        "movl %4, %%ecx                         \n\t"
704 705
        "add  %1, %%"REG_c"                     \n\t"
        "movl %%ecx, (%%"REG_a")                \n\t"
706

707
        "test $1, %%edx                         \n\t"
708 709
        " jnz 4f                                \n\t"

710 711
        "add  $4, %%"REG_a"                     \n\t"
        "mov  %%"REG_a", %2                     \n\t"
Michael Niedermayer's avatar
Michael Niedermayer committed
712

713
        "3:                                     \n\t"
714 715
        "add  $1, %1                            \n\t"
        "cmp  %5, %1                            \n\t"
716
        " jb 2b                                 \n\t"
717
        "mov  %2, %%"REG_a"                     \n\t"
718
        "movl %4, %%ecx                         \n\t"
719 720
        "add  %1, %%"REG_c"                     \n\t"
        "movl %%ecx, (%%"REG_a")                \n\t"
721
        "4:                                     \n\t"
722
        "add  %6, %%eax                         \n\t"
723 724 725 726 727 728
        "shr $2, %%eax                          \n\t"

        "movl %%esi, "RANGE    "(%3)            \n\t"
        "movl %%ebx, "LOW      "(%3)            \n\t"
        :"=&a"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index)\
        :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex)\
729
        : "%"REG_c, "%ebx", "%edx", "%esi", "memory"\
730 731 732
    );
    return coeff_count;
}
733 734 735 736

static int decode_significance_8x8_x86(CABACContext *c, uint8_t *significant_coeff_ctx_base, int *index, uint8_t *sig_off){
    int minusindex= 4-(int)index;
    int coeff_count;
737
    long last=0;
738 739 740 741
    asm volatile(
        "movl "RANGE    "(%3), %%esi            \n\t"
        "movl "LOW      "(%3), %%ebx            \n\t"

742
        "mov %1, %%"REG_D"                      \n\t"
743 744
        "2:                                     \n\t"

745 746 747
        "mov %6, %%"REG_a"                      \n\t"
        "movzbl (%%"REG_a", %%"REG_D"), %%edi   \n\t"
        "add %5, %%"REG_D"                      \n\t"
748

749
        BRANCHLESS_GET_CABAC("%%edx", "%3", "(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
750 751 752 753 754 755

        "mov %1, %%edi                          \n\t"
        "test $1, %%edx                         \n\t"
        " jz 3f                                 \n\t"

        "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t"
756
        "add %5, %%"REG_D"                      \n\t"
757

758
        BRANCHLESS_GET_CABAC("%%edx", "%3", "15(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
759

760
        "mov %2, %%"REG_a"                      \n\t"
761
        "mov %1, %%edi                          \n\t"
762
        "movl %%edi, (%%"REG_a")                \n\t"
763 764 765 766

        "test $1, %%edx                         \n\t"
        " jnz 4f                                \n\t"

767 768
        "add $4, %%"REG_a"                      \n\t"
        "mov %%"REG_a", %2                      \n\t"
769 770 771 772 773 774

        "3:                                     \n\t"
        "addl $1, %%edi                         \n\t"
        "mov %%edi, %1                          \n\t"
        "cmpl $63, %%edi                        \n\t"
        " jb 2b                                 \n\t"
775 776
        "mov %2, %%"REG_a"                      \n\t"
        "movl %%edi, (%%"REG_a")                \n\t"
777 778 779 780 781 782 783 784
        "4:                                     \n\t"
        "addl %4, %%eax                         \n\t"
        "shr $2, %%eax                          \n\t"

        "movl %%esi, "RANGE    "(%3)            \n\t"
        "movl %%ebx, "LOW      "(%3)            \n\t"
        :"=&a"(coeff_count),"+m"(last), "+m"(index)\
        :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off)\
785
        : "%"REG_c, "%ebx", "%edx", "%esi", "%"REG_D, "memory"\
786 787 788
    );
    return coeff_count;
}
789
#endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
790

791 792 793 794
/**
 *
 * @return the number of bytes read or 0 if no end
 */
795
static int get_cabac_terminate(CABACContext *c){
796
    c->range -= 2;
797
    if(c->low < c->range<<(CABAC_BITS+1)){
Michael Niedermayer's avatar
Michael Niedermayer committed
798
        renorm_cabac_decoder_once(c);
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
799 800
        return 0;
    }else{
801
        return c->bytestream - c->bytestream_start;
802
    }
Michael Niedermayer's avatar
CABAC  
Michael Niedermayer committed
803 804
}

805 806 807
/**
 * get (truncated) unnary binarization.
 */
808
static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){
809
    int i;
810 811

    for(i=0; i<max; i++){
812 813
        if(get_cabac(c, state)==0)
            return i;
814

815 816 817 818 819 820 821 822 823
        if(i< max_index) state++;
    }

    return truncated ? max : -1;
}

/**
 * get unary exp golomb k-th order binarization.
 */
824
static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){
825 826
    int i, v;
    int m= 1<<k;
827 828

    if(get_cabac(c, state)==0)
829
        return 0;
830

831
    if(0 < max_index) state++;
832 833

    for(i=1; i<max; i++){
834 835 836 837 838 839 840 841 842
        if(get_cabac(c, state)==0){
            if(is_signed && get_cabac_bypass(c)){
                return -i;
            }else
                return i;
        }

        if(i < max_index) state++;
    }
843

844 845 846 847
    while(get_cabac_bypass(c)){
        i+= m;
        m+= m;
    }
848

849 850 851 852 853 854 855 856 857 858 859
    v=0;
    while(m>>=1){
        v+= v + get_cabac_bypass(c);
    }
    i += v;

    if(is_signed && get_cabac_bypass(c)){
        return -i;
    }else
        return i;
}