libavcodec/common.h

   1 /**
   2  * @file common.h
   3  * common internal api header.
   4  */
   5
   6 #ifndef COMMON_H
   7 #define COMMON_H
   8
   9 #if defined(WIN32) && !defined(__MINGW32__) && !defined(__CYGWIN__)
  10 #    define CONFIG_WIN32
  11 #endif
  12
  13 //#define ALT_BITSTREAM_WRITER
  14 //#define ALIGNED_BITSTREAM_WRITER
  15
  16 #define ALT_BITSTREAM_READER
  17 //#define LIBMPEG2_BITSTREAM_READER
  18 //#define A32_BITSTREAM_READER
  19 #define LIBMPEG2_BITSTREAM_READER_HACK //add BERO
  20
  21 #ifdef HAVE_AV_CONFIG_H
  22 /* only include the following when compiling package */
  23 #    include "config.h"
  24
  25 #    include <stdlib.h>
  26 #    include <stdio.h>
  27 #    include <string.h>
  28 #    include <ctype.h>
  29 #    ifndef __BEOS__
  30 #        include <errno.h>
  31 #    else
  32 #        include "berrno.h"
  33 #    endif
  34 #    include <math.h>
  35
  36 #    ifndef ENODATA
  37 #        define ENODATA  61
  38 #    endif
  39
  40 #ifndef M_PI
  41 #define M_PI    3.14159265358979323846
  42 #endif
  43
  44 #include <stddef.h>
  45 #ifndef offsetof
  46 # define offsetof(T,F) ((unsigned int)((char *)&((T *)0)->F))
  47 #endif
  48
  49 #define AVOPTION_CODEC_BOOL(name, help, field) \
  50     { name, help, offsetof(AVCodecContext, field), FF_OPT_TYPE_BOOL }
  51 #define AVOPTION_CODEC_DOUBLE(name, help, field, minv, maxv, defval) \
  52     { name, help, offsetof(AVCodecContext, field), FF_OPT_TYPE_DOUBLE, minv, maxv, defval }
  53 #define AVOPTION_CODEC_FLAG(name, help, field, flag, defval) \
  54     { name, help, offsetof(AVCodecContext, field), FF_OPT_TYPE_FLAG, flag, 0, defval }
  55 #define AVOPTION_CODEC_INT(name, help, field, minv, maxv, defval) \
  56     { name, help, offsetof(AVCodecContext, field), FF_OPT_TYPE_INT, minv, maxv, defval }
  57 #define AVOPTION_CODEC_STRING(name, help, field, str, val) \
  58     { name, help, offsetof(AVCodecContext, field), FF_OPT_TYPE_STRING, .defval = val, .defstr = str }
  59 #define AVOPTION_CODEC_RCOVERRIDE(name, help, field) \
  60     { name, help, offsetof(AVCodecContext, field), FF_OPT_TYPE_RCOVERRIDE, .defval = 0, .defstr = NULL }
  61 #define AVOPTION_SUB(ptr) { .name = NULL, .help = (const char*)ptr }
  62 #define AVOPTION_END() AVOPTION_SUB(NULL)
  63
  64 struct AVOption;
  65 #ifdef HAVE_MMX
  66 extern const struct AVOption avoptions_common[3 + 5];
  67 #else
  68 extern const struct AVOption avoptions_common[3];
  69 #endif
  70 extern const struct AVOption avoptions_workaround_bug[11];
  71
  72 #endif /* HAVE_AV_CONFIG_H */
  73
  74 /* Suppress restrict if it was not defined in config.h.  */
  75 #ifndef restrict
  76 #    define restrict
  77 #endif
  78
  79 #if defined(__GNUC__) && (__GNUC__ > 3 || __GNUC__ == 3 && __GNUC_MINOR__ > 0)
  80 #    define always_inline __attribute__((always_inline)) inline
  81 #else
  82 #    define always_inline inline
  83 #endif
  84
  85 #ifdef EMULATE_FAST_INT
  86 /* note that we don't emulate 64bit ints */
  87 typedef signed char int_fast8_t;
  88 typedef signed int  int_fast16_t;
  89 typedef signed int  int_fast32_t;
  90 typedef unsigned char uint_fast8_t;
  91 typedef unsigned int  uint_fast16_t;
  92 typedef unsigned int  uint_fast32_t;
  93 #endif
  94
  95 #if defined(CONFIG_OS2) || defined(CONFIG_SUNOS)
  96 static inline float floorf(float f) {
  97     return floor(f);
  98 }
  99 #endif
 100
 101 #ifdef CONFIG_WIN32
 102
 103 /* windows */
 104
 105 typedef unsigned short uint16_t;
 106 typedef signed short int16_t;
 107 typedef unsigned char uint8_t;
 108 typedef unsigned int uint32_t;
 109 typedef unsigned __int64 uint64_t;
 110 typedef signed char int8_t;
 111 typedef signed int int32_t;
 112 typedef signed __int64 int64_t;
 113
 114 #    ifndef __MINGW32__
 115 #        define int64_t_C(c)     (c ## i64)
 116 #        define uint64_t_C(c)    (c ## i64)
 117
 118 #    ifdef HAVE_AV_CONFIG_H
 119 #            define inline __inline
 120 #    endif
 121
 122 #    else
 123 #        define int64_t_C(c)     (c ## LL)
 124 #        define uint64_t_C(c)    (c ## ULL)
 125 #    endif /* __MINGW32__ */
 126
 127 #    ifdef HAVE_AV_CONFIG_H
 128 #        ifdef _DEBUG
 129 #            define DEBUG
 130 #        endif
 131
 132 #        define snprintf _snprintf
 133 #        define vsnprintf _vsnprintf
 134 #    endif
 135
 136 /* CONFIG_WIN32 end */
 137 #elif defined (CONFIG_OS2)
 138 /* OS/2 EMX */
 139
 140 #include <inttypes.h>
 141
 142 #ifndef int64_t_C
 143 #define int64_t_C(c)     (c ## LL)
 144 #define uint64_t_C(c)    (c ## ULL)
 145 #endif
 146
 147 #ifdef HAVE_AV_CONFIG_H
 148
 149 #ifdef USE_FASTMEMCPY
 150 #include "fastmemcpy.h"
 151 #endif
 152
 153 #include <float.h>
 154
 155 #endif /* HAVE_AV_CONFIG_H */
 156
 157 /* CONFIG_OS2 end */
 158 #else
 159
 160 /* unix */
 161
 162 #include <inttypes.h>
 163
 164 #ifndef int64_t_C
 165 #define int64_t_C(c)     (c ## LL)
 166 #define uint64_t_C(c)    (c ## ULL)
 167 #endif
 168
 169 #ifdef HAVE_AV_CONFIG_H
 170
 171 #        ifdef USE_FASTMEMCPY
 172 #            include "fastmemcpy.h"
 173 #        endif
 174 #    endif /* HAVE_AV_CONFIG_H */
 175
 176 #endif /* !CONFIG_WIN32 && !CONFIG_OS2 */
 177
 178 #ifdef HAVE_AV_CONFIG_H
 179
 180 #    include "bswap.h"
 181
 182 #    if defined(__MINGW32__) || defined(__CYGWIN__) || \
 183         defined(__OS2__) || (defined (__OpenBSD__) && !defined(__ELF__))
 184 #        define MANGLE(a) "_" #a
 185 #    else
 186 #        define MANGLE(a) #a
 187 #    endif
 188
 189 /* debug stuff */
 190
 191 #    ifndef DEBUG
 192 #        define NDEBUG
 193 #    endif
 194 #    include <assert.h>
 195
 196 /* dprintf macros */
 197 #    if defined(CONFIG_WIN32) && !defined(__MINGW32__)
 198
 199 inline void dprintf(const char* fmt,...) {}
 200
 201 #    else
 202
 203 #        ifdef DEBUG
 204 #            define dprintf(fmt,...) printf(fmt, __VA_ARGS__)
 205 #        else
 206 #            define dprintf(fmt,...)
 207 #        endif
 208
 209 #    endif /* !CONFIG_WIN32 */
 210
 211 #    define av_abort()      do { av_log(NULL, AV_LOG_ERROR, "Abort at %s:%d\n", __FILE__, __LINE__); abort(); } while (0)
 212
 213 //rounded divison & shift
 214 #define RSHIFT(a,b) ((a) > 0 ? ((a) + (1<<((b)-1)))>>(b) : ((a) + (1<<((b)-1))-1)>>(b))
 215 /* assume b>0 */
 216 #define ROUNDED_DIV(a,b) (((a)>0 ? (a) + ((b)>>1) : (a) - ((b)>>1))/(b))
 217 #define ABS(a) ((a) >= 0 ? (a) : (-(a)))
 218
 219 #define FFMAX(a,b) ((a) > (b) ? (a) : (b))
 220 #define FFMIN(a,b) ((a) > (b) ? (b) : (a))
 221
 222 extern const uint32_t inverse[256];
 223
 224 #ifdef ARCH_X86
 225 #    define FASTDIV(a,b) \
 226     ({\
 227         int ret,dmy;\
 228         asm volatile(\
 229             "mull %3"\
 230             :"=d"(ret),"=a"(dmy)\
 231             :"1"(a),"g"(inverse[b])\
 232             );\
 233         ret;\
 234     })
 235 #elif defined(CONFIG_FASTDIV)
 236 #    define FASTDIV(a,b)   ((uint32_t)((((uint64_t)a)*inverse[b])>>32))
 237 #else
 238 #    define FASTDIV(a,b)   ((a)/(b))
 239 #endif
 240
 241 #ifdef ARCH_X86
 242 // avoid +32 for shift optimization (gcc should do that ...)
 243 static inline  int32_t NEG_SSR32( int32_t a, int8_t s){
 244     asm ("sarl %1, %0\n\t"
 245          : "+r" (a)
 246          : "ic" ((uint8_t)(-s))
 247     );
 248     return a;
 249 }
 250 static inline uint32_t NEG_USR32(uint32_t a, int8_t s){
 251     asm ("shrl %1, %0\n\t"
 252          : "+r" (a)
 253          : "ic" ((uint8_t)(-s))
 254     );
 255     return a;
 256 }
 257 #else
 258 #    define NEG_SSR32(a,s) ((( int32_t)(a))>>(32-(s)))
 259 #    define NEG_USR32(a,s) (((uint32_t)(a))>>(32-(s)))
 260 #endif
 261
 262 /* bit output */
 263
 264 struct PutBitContext;
 265
 266 typedef void (*WriteDataFunc)(void *, uint8_t *, int);
 267
 268 typedef struct PutBitContext {
 269 #ifdef ALT_BITSTREAM_WRITER
 270     uint8_t *buf, *buf_end;
 271     int index;
 272 #else
 273     uint32_t bit_buf;
 274     int bit_left;
 275     uint8_t *buf, *buf_ptr, *buf_end;
 276 #endif
 277     int64_t data_out_size; /* in bytes */
 278 } PutBitContext;
 279
 280 void init_put_bits(PutBitContext *s, uint8_t *buffer, int buffer_size);
 281
 282 int64_t get_bit_count(PutBitContext *s); /* XXX: change function name */
 283 void align_put_bits(PutBitContext *s);
 284 void flush_put_bits(PutBitContext *s);
 285 void put_string(PutBitContext * pbc, char *s);
 286
 287 /* bit input */
 288
 289 typedef struct GetBitContext {
 290     const uint8_t *buffer, *buffer_end;
 291 #ifdef ALT_BITSTREAM_READER
 292     int index;
 293 #elif defined LIBMPEG2_BITSTREAM_READER
 294     uint8_t *buffer_ptr;
 295     uint32_t cache;
 296     int bit_count;
 297 #elif defined A32_BITSTREAM_READER
 298     uint32_t *buffer_ptr;
 299     uint32_t cache0;
 300     uint32_t cache1;
 301     int bit_count;
 302 #endif
 303     int size_in_bits;
 304 } GetBitContext;
 305
 306 static inline int get_bits_count(GetBitContext *s);
 307
 308 #define VLC_TYPE int16_t
 309
 310 typedef struct VLC {
 311     int bits;
 312     VLC_TYPE (*table)[2]; ///< code, bits
 313     int table_size, table_allocated;
 314 } VLC;
 315
 316 typedef struct RL_VLC_ELEM {
 317     int16_t level;
 318     int8_t len;
 319     uint8_t run;
 320 } RL_VLC_ELEM;
 321
 322 #ifdef ARCH_SPARC64
 323 #define UNALIGNED_STORES_ARE_BAD
 324 #endif
 325
 326 /* used to avoid missaligned exceptions on some archs (alpha, ...) */
 327 #ifdef ARCH_X86
 328 #    define unaligned32(a) (*(uint32_t*)(a))
 329 #else
 330 #    ifdef __GNUC__
 331 static inline uint32_t unaligned32(const void *v) {
 332     struct Unaligned {
 333         uint32_t i;
 334     } __attribute__((packed));
 335
 336     return ((const struct Unaligned *) v)->i;
 337 }
 338 #    elif defined(__DECC)
 339 static inline uint32_t unaligned32(const void *v) {
 340     return *(const __unaligned uint32_t *) v;
 341 }
 342 #    else
 343 static inline uint32_t unaligned32(const void *v) {
 344     return *(const uint32_t *) v;
 345 }
 346 #    endif
 347 #endif //!ARCH_X86
 348
 349 #ifndef ALT_BITSTREAM_WRITER
 350 static inline void put_bits(PutBitContext *s, int n, unsigned int value)
 351 {
 352     unsigned int bit_buf;
 353     int bit_left;
 354
 355 #ifdef STATS
 356     st_out_bit_counts[st_current_index] += n;
 357 #endif
 358     //    printf("put_bits=%d %x\n", n, value);
 359     assert(n == 32 || value < (1U << n));
 360
 361     bit_buf = s->bit_buf;
 362     bit_left = s->bit_left;
 363
 364     //    printf("n=%d value=%x cnt=%d buf=%x\n", n, value, bit_cnt, bit_buf);
 365     /* XXX: optimize */
 366     if (n < bit_left) {
 367         bit_buf = (bit_buf<<n) | value;
 368         bit_left-=n;
 369     } else {
 370         bit_buf<<=bit_left;
 371         bit_buf |= value >> (n - bit_left);
 372 #ifdef UNALIGNED_STORES_ARE_BAD
 373         if (3 & (int) s->buf_ptr) {
 374             s->buf_ptr[0] = bit_buf >> 24;
 375             s->buf_ptr[1] = bit_buf >> 16;
 376             s->buf_ptr[2] = bit_buf >>  8;
 377             s->buf_ptr[3] = bit_buf      ;
 378         } else
 379 #endif
 380         *(uint32_t *)s->buf_ptr = be2me_32(bit_buf);
 381         //printf("bitbuf = %08x\n", bit_buf);
 382         s->buf_ptr+=4;
 383         bit_left+=32 - n;
 384         bit_buf = value;
 385     }
 386
 387     s->bit_buf = bit_buf;
 388     s->bit_left = bit_left;
 389 }
 390 #endif
 391
 392
 393 #ifdef ALT_BITSTREAM_WRITER
 394 static inline void put_bits(PutBitContext *s, int n, unsigned int value)
 395 {
 396 #    ifdef ALIGNED_BITSTREAM_WRITER
 397 #        ifdef ARCH_X86
 398     asm volatile(
 399         "movl %0, %%ecx                 \n\t"
 400         "xorl %%eax, %%eax              \n\t"
 401         "shrdl %%cl, %1, %%eax          \n\t"
 402         "shrl %%cl, %1                  \n\t"
 403         "movl %0, %%ecx                 \n\t"
 404         "shrl $3, %%ecx                 \n\t"
 405         "andl $0xFFFFFFFC, %%ecx        \n\t"
 406         "bswapl %1                      \n\t"
 407         "orl %1, (%2, %%ecx)            \n\t"
 408         "bswapl %%eax                   \n\t"
 409         "addl %3, %0                    \n\t"
 410         "movl %%eax, 4(%2, %%ecx)       \n\t"
 411         : "=&r" (s->index), "=&r" (value)
 412         : "r" (s->buf), "r" (n), "0" (s->index), "1" (value<<(-n))
 413         : "%eax", "%ecx"
 414     );
 415 #        else
 416     int index= s->index;
 417     uint32_t *ptr= ((uint32_t *)s->buf)+(index>>5);
 418
 419     value<<= 32-n;
 420
 421     ptr[0] |= be2me_32(value>>(index&31));
 422     ptr[1]  = be2me_32(value<<(32-(index&31)));
 423 //if(n>24) printf("%d %d\n", n, value);
 424     index+= n;
 425     s->index= index;
 426 #        endif
 427 #    else //ALIGNED_BITSTREAM_WRITER
 428 #        ifdef ARCH_X86
 429     asm volatile(
 430         "movl $7, %%ecx                 \n\t"
 431         "andl %0, %%ecx                 \n\t"
 432         "addl %3, %%ecx                 \n\t"
 433         "negl %%ecx                     \n\t"
 434         "shll %%cl, %1                  \n\t"
 435         "bswapl %1                      \n\t"
 436         "movl %0, %%ecx                 \n\t"
 437         "shrl $3, %%ecx                 \n\t"
 438         "orl %1, (%%ecx, %2)            \n\t"
 439         "addl %3, %0                    \n\t"
 440         "movl $0, 4(%%ecx, %2)          \n\t"
 441         : "=&r" (s->index), "=&r" (value)
 442         : "r" (s->buf), "r" (n), "0" (s->index), "1" (value)
 443         : "%ecx"
 444     );
 445 #        else
 446     int index= s->index;
 447     uint32_t *ptr= (uint32_t*)(((uint8_t *)s->buf)+(index>>3));
 448
 449     ptr[0] |= be2me_32(value<<(32-n-(index&7) ));
 450     ptr[1] = 0;
 451 //if(n>24) printf("%d %d\n", n, value);
 452     index+= n;
 453     s->index= index;
 454 #        endif
 455 #    endif //!ALIGNED_BITSTREAM_WRITER
 456 }
 457 #endif
 458
 459
 460 static inline uint8_t* pbBufPtr(PutBitContext *s)
 461 {
 462 #ifdef ALT_BITSTREAM_WRITER
 463         return s->buf + (s->index>>3);
 464 #else
 465         return s->buf_ptr;
 466 #endif
 467 }
 468
 469 /* Bitstream reader API docs:
 470 name
 471     abritary name which is used as prefix for the internal variables
 472
 473 gb
 474     getbitcontext
 475
 476 OPEN_READER(name, gb)
 477     loads gb into local variables
 478
 479 CLOSE_READER(name, gb)
 480     stores local vars in gb
 481
 482 UPDATE_CACHE(name, gb)
 483     refills the internal cache from the bitstream
 484     after this call at least MIN_CACHE_BITS will be available,
 485
 486 GET_CACHE(name, gb)
 487     will output the contents of the internal cache, next bit is MSB of 32 or 64 bit (FIXME 64bit)
 488
 489 SHOW_UBITS(name, gb, num)
 490     will return the nest num bits
 491
 492 SHOW_SBITS(name, gb, num)
 493     will return the nest num bits and do sign extension
 494
 495 SKIP_BITS(name, gb, num)
 496     will skip over the next num bits
 497     note, this is equinvalent to SKIP_CACHE; SKIP_COUNTER
 498
 499 SKIP_CACHE(name, gb, num)
 500     will remove the next num bits from the cache (note SKIP_COUNTER MUST be called before UPDATE_CACHE / CLOSE_READER)
 501
 502 SKIP_COUNTER(name, gb, num)
 503     will increment the internal bit counter (see SKIP_CACHE & SKIP_BITS)
 504
 505 LAST_SKIP_CACHE(name, gb, num)
 506     will remove the next num bits from the cache if it is needed for UPDATE_CACHE otherwise it will do nothing
 507
 508 LAST_SKIP_BITS(name, gb, num)
 509     is equinvalent to SKIP_LAST_CACHE; SKIP_COUNTER
 510
 511 for examples see get_bits, show_bits, skip_bits, get_vlc
 512 */
 513
 514 static inline int unaligned32_be(const void *v)
 515 {
 516 #ifdef CONFIG_ALIGN
 517         const uint8_t *p=v;
 518         return (((p[0]<<8) | p[1])<<16) | (p[2]<<8) | (p[3]);
 519 #else
 520         return be2me_32( unaligned32(v)); //original
 521 #endif
 522 }
 523
 524 #ifdef ALT_BITSTREAM_READER
 525 #   define MIN_CACHE_BITS 25
 526
 527 #   define OPEN_READER(name, gb)\
 528         int name##_index= (gb)->index;\
 529         int name##_cache= 0;\
 530
 531 #   define CLOSE_READER(name, gb)\
 532         (gb)->index= name##_index;\
 533
 534 #   define UPDATE_CACHE(name, gb)\
 535         name##_cache= unaligned32_be( ((uint8_t *)(gb)->buffer)+(name##_index>>3) ) << (name##_index&0x07);\
 536
 537 #   define SKIP_CACHE(name, gb, num)\
 538         name##_cache <<= (num);\
 539
 540 // FIXME name?
 541 #   define SKIP_COUNTER(name, gb, num)\
 542         name##_index += (num);\
 543
 544 #   define SKIP_BITS(name, gb, num)\
 545         {\
 546             SKIP_CACHE(name, gb, num)\
 547             SKIP_COUNTER(name, gb, num)\
 548         }\
 549
 550 #   define LAST_SKIP_BITS(name, gb, num) SKIP_COUNTER(name, gb, num)
 551 #   define LAST_SKIP_CACHE(name, gb, num) ;
 552
 553 #   define SHOW_UBITS(name, gb, num)\
 554         NEG_USR32(name##_cache, num)
 555
 556 #   define SHOW_SBITS(name, gb, num)\
 557         NEG_SSR32(name##_cache, num)
 558
 559 #   define GET_CACHE(name, gb)\
 560         ((uint32_t)name##_cache)
 561
 562 static inline int get_bits_count(GetBitContext *s){
 563     return s->index;
 564 }
 565 #elif defined LIBMPEG2_BITSTREAM_READER
 566 //libmpeg2 like reader
 567
 568 #   define MIN_CACHE_BITS 17
 569
 570 #   define OPEN_READER(name, gb)\
 571         int name##_bit_count=(gb)->bit_count;\
 572         int name##_cache= (gb)->cache;\
 573         uint8_t * name##_buffer_ptr=(gb)->buffer_ptr;\
 574
 575 #   define CLOSE_READER(name, gb)\
 576         (gb)->bit_count= name##_bit_count;\
 577         (gb)->cache= name##_cache;\
 578         (gb)->buffer_ptr= name##_buffer_ptr;\
 579
 580 #ifdef LIBMPEG2_BITSTREAM_READER_HACK
 581
 582 #   define UPDATE_CACHE(name, gb)\
 583     if(name##_bit_count >= 0){\
 584         name##_cache+= (int)be2me_16(*(uint16_t*)name##_buffer_ptr) << name##_bit_count;\
 585         ((uint16_t*)name##_buffer_ptr)++;\
 586         name##_bit_count-= 16;\
 587     }\
 588
 589 #else
 590
 591 #   define UPDATE_CACHE(name, gb)\
 592     if(name##_bit_count >= 0){\
 593         name##_cache+= ((name##_buffer_ptr[0]<<8) + name##_buffer_ptr[1]) << name##_bit_count;\
 594         name##_buffer_ptr+=2;\
 595         name##_bit_count-= 16;\
 596     }\
 597
 598 #endif
 599
 600 #   define SKIP_CACHE(name, gb, num)\
 601         name##_cache <<= (num);\
 602
 603 #   define SKIP_COUNTER(name, gb, num)\
 604         name##_bit_count += (num);\
 605
 606 #   define SKIP_BITS(name, gb, num)\
 607         {\
 608             SKIP_CACHE(name, gb, num)\
 609             SKIP_COUNTER(name, gb, num)\
 610         }\
 611
 612 #   define LAST_SKIP_BITS(name, gb, num) SKIP_BITS(name, gb, num)
 613 #   define LAST_SKIP_CACHE(name, gb, num) SKIP_CACHE(name, gb, num)
 614
 615 #   define SHOW_UBITS(name, gb, num)\
 616         NEG_USR32(name##_cache, num)
 617
 618 #   define SHOW_SBITS(name, gb, num)\
 619         NEG_SSR32(name##_cache, num)
 620
 621 #   define GET_CACHE(name, gb)\
 622         ((uint32_t)name##_cache)
 623
 624 static inline int get_bits_count(GetBitContext *s){
 625     return (s->buffer_ptr - s->buffer)*8 - 16 + s->bit_count;
 626 }
 627
 628 #elif defined A32_BITSTREAM_READER
 629
 630 #   define MIN_CACHE_BITS 32
 631
 632 #   define OPEN_READER(name, gb)\
 633         int name##_bit_count=(gb)->bit_count;\
 634         uint32_t name##_cache0= (gb)->cache0;\
 635         uint32_t name##_cache1= (gb)->cache1;\
 636         uint32_t * name##_buffer_ptr=(gb)->buffer_ptr;\
 637
 638 #   define CLOSE_READER(name, gb)\
 639         (gb)->bit_count= name##_bit_count;\
 640         (gb)->cache0= name##_cache0;\
 641         (gb)->cache1= name##_cache1;\
 642         (gb)->buffer_ptr= name##_buffer_ptr;\
 643
 644 #   define UPDATE_CACHE(name, gb)\
 645     if(name##_bit_count > 0){\
 646         const uint32_t next= be2me_32( *name##_buffer_ptr );\
 647         name##_cache0 |= NEG_USR32(next,name##_bit_count);\
 648         name##_cache1 |= next<<name##_bit_count;\
 649         name##_buffer_ptr++;\
 650         name##_bit_count-= 32;\
 651     }\
 652
 653 #ifdef ARCH_X86
 654 #   define SKIP_CACHE(name, gb, num)\
 655         asm(\
 656             "shldl %2, %1, %0           \n\t"\
 657             "shll %2, %1                \n\t"\
 658             : "+r" (name##_cache0), "+r" (name##_cache1)\
 659             : "Ic" ((uint8_t)num)\
 660            );
 661 #else
 662 #   define SKIP_CACHE(name, gb, num)\
 663         name##_cache0 <<= (num);\
 664         name##_cache0 |= NEG_USR32(name##_cache1,num);\
 665         name##_cache1 <<= (num);
 666 #endif
 667
 668 #   define SKIP_COUNTER(name, gb, num)\
 669         name##_bit_count += (num);\
 670
 671 #   define SKIP_BITS(name, gb, num)\
 672         {\
 673             SKIP_CACHE(name, gb, num)\
 674             SKIP_COUNTER(name, gb, num)\
 675         }\
 676
 677 #   define LAST_SKIP_BITS(name, gb, num) SKIP_BITS(name, gb, num)
 678 #   define LAST_SKIP_CACHE(name, gb, num) SKIP_CACHE(name, gb, num)
 679
 680 #   define SHOW_UBITS(name, gb, num)\
 681         NEG_USR32(name##_cache0, num)
 682
 683 #   define SHOW_SBITS(name, gb, num)\
 684         NEG_SSR32(name##_cache0, num)
 685
 686 #   define GET_CACHE(name, gb)\
 687         (name##_cache0)
 688
 689 static inline int get_bits_count(GetBitContext *s){
 690     return ((uint8_t*)s->buffer_ptr - s->buffer)*8 - 32 + s->bit_count;
 691 }
 692
 693 #endif
 694
 695 /**
 696  * read mpeg1 dc style vlc (sign bit + mantisse with no MSB).
 697  * if MSB not set it is negative
 698  * @param n length in bits
 699  * @author BERO
 700  */
 701 static inline int get_xbits(GetBitContext *s, int n){
 702     register int tmp;
 703     register int32_t cache;
 704     OPEN_READER(re, s)
 705     UPDATE_CACHE(re, s)
 706     cache = GET_CACHE(re,s);
 707     if ((int32_t)cache<0) { //MSB=1
 708         tmp = NEG_USR32(cache,n);
 709     } else {
 710     //   tmp = (-1<<n) | NEG_USR32(cache,n) + 1; mpeg12.c algo
 711     //   tmp = - (NEG_USR32(cache,n) ^ ((1 << n) - 1)); h263.c algo
 712         tmp = - NEG_USR32(~cache,n);
 713     }
 714     LAST_SKIP_BITS(re, s, n)
 715     CLOSE_READER(re, s)
 716     return tmp;
 717 }
 718
 719 static inline int get_sbits(GetBitContext *s, int n){
 720     register int tmp;
 721     OPEN_READER(re, s)
 722     UPDATE_CACHE(re, s)
 723     tmp= SHOW_SBITS(re, s, n);
 724     LAST_SKIP_BITS(re, s, n)
 725     CLOSE_READER(re, s)
 726     return tmp;
 727 }
 728
 729 /**
 730  * reads 0-17 bits.
 731  * Note, the alt bitstream reader can read upto 25 bits, but the libmpeg2 reader cant
 732  */
 733 static inline unsigned int get_bits(GetBitContext *s, int n){
 734     register int tmp;
 735     OPEN_READER(re, s)
 736     UPDATE_CACHE(re, s)
 737     tmp= SHOW_UBITS(re, s, n);
 738     LAST_SKIP_BITS(re, s, n)
 739     CLOSE_READER(re, s)
 740     return tmp;
 741 }
 742
 743 unsigned int get_bits_long(GetBitContext *s, int n);
 744
 745 /**
 746  * shows 0-17 bits.
 747  * Note, the alt bitstream reader can read upto 25 bits, but the libmpeg2 reader cant
 748  */
 749 static inline unsigned int show_bits(GetBitContext *s, int n){
 750     register int tmp;
 751     OPEN_READER(re, s)
 752     UPDATE_CACHE(re, s)
 753     tmp= SHOW_UBITS(re, s, n);
 754 //    CLOSE_READER(re, s)
 755     return tmp;
 756 }
 757
 758 unsigned int show_bits_long(GetBitContext *s, int n);
 759
 760 static inline void skip_bits(GetBitContext *s, int n){
 761  //Note gcc seems to optimize this to s->index+=n for the ALT_READER :))
 762     OPEN_READER(re, s)
 763     UPDATE_CACHE(re, s)
 764     LAST_SKIP_BITS(re, s, n)
 765     CLOSE_READER(re, s)
 766 }
 767
 768 static inline unsigned int get_bits1(GetBitContext *s){
 769 #ifdef ALT_BITSTREAM_READER
 770     int index= s->index;
 771     uint8_t result= s->buffer[ index>>3 ];
 772     result<<= (index&0x07);
 773     result>>= 8 - 1;
 774     index++;
 775     s->index= index;
 776
 777     return result;
 778 #else
 779     return get_bits(s, 1);
 780 #endif
 781 }
 782
 783 static inline unsigned int show_bits1(GetBitContext *s){
 784     return show_bits(s, 1);
 785 }
 786
 787 static inline void skip_bits1(GetBitContext *s){
 788     skip_bits(s, 1);
 789 }
 790
 791 void init_get_bits(GetBitContext *s,
 792                    const uint8_t *buffer, int buffer_size);
 793
 794 int check_marker(GetBitContext *s, const char *msg);
 795 void align_get_bits(GetBitContext *s);
 796 int init_vlc(VLC *vlc, int nb_bits, int nb_codes,
 797              const void *bits, int bits_wrap, int bits_size,
 798              const void *codes, int codes_wrap, int codes_size);
 799 void free_vlc(VLC *vlc);
 800
 801 /**
 802  *
 803  * if the vlc code is invalid and max_depth=1 than no bits will be removed
 804  * if the vlc code is invalid and max_depth>1 than the number of bits removed
 805  * is undefined
 806  */
 807 #define GET_VLC(code, name, gb, table, bits, max_depth)\
 808 {\
 809     int n, index, nb_bits;\
 810 \
 811     index= SHOW_UBITS(name, gb, bits);\
 812     code = table[index][0];\
 813     n    = table[index][1];\
 814 \
 815     if(max_depth > 1 && n < 0){\
 816         LAST_SKIP_BITS(name, gb, bits)\
 817         UPDATE_CACHE(name, gb)\
 818 \
 819         nb_bits = -n;\
 820 \
 821         index= SHOW_UBITS(name, gb, nb_bits) + code;\
 822         code = table[index][0];\
 823         n    = table[index][1];\
 824         if(max_depth > 2 && n < 0){\
 825             LAST_SKIP_BITS(name, gb, nb_bits)\
 826             UPDATE_CACHE(name, gb)\
 827 \
 828             nb_bits = -n;\
 829 \
 830             index= SHOW_UBITS(name, gb, nb_bits) + code;\
 831             code = table[index][0];\
 832             n    = table[index][1];\
 833         }\
 834     }\
 835     SKIP_BITS(name, gb, n)\
 836 }
 837
 838 #define GET_RL_VLC(level, run, name, gb, table, bits, max_depth)\
 839 {\
 840     int n, index, nb_bits;\
 841 \
 842     index= SHOW_UBITS(name, gb, bits);\
 843     level = table[index].level;\
 844     n     = table[index].len;\
 845 \
 846     if(max_depth > 1 && n < 0){\
 847         LAST_SKIP_BITS(name, gb, bits)\
 848         UPDATE_CACHE(name, gb)\
 849 \
 850         nb_bits = -n;\
 851 \
 852         index= SHOW_UBITS(name, gb, nb_bits) + level;\
 853         level = table[index].level;\
 854         n     = table[index].len;\
 855     }\
 856     run= table[index].run;\
 857     SKIP_BITS(name, gb, n)\
 858 }
 859
 860 // deprecated, dont use get_vlc for new code, use get_vlc2 instead or use GET_VLC directly
 861 static inline int get_vlc(GetBitContext *s, VLC *vlc)
 862 {
 863     int code;
 864     VLC_TYPE (*table)[2]= vlc->table;
 865
 866     OPEN_READER(re, s)
 867     UPDATE_CACHE(re, s)
 868
 869     GET_VLC(code, re, s, table, vlc->bits, 3)
 870
 871     CLOSE_READER(re, s)
 872     return code;
 873 }
 874
 875 /**
 876  * parses a vlc code, faster then get_vlc()
 877  * @param bits is the number of bits which will be read at once, must be
 878  *             identical to nb_bits in init_vlc()
 879  * @param max_depth is the number of times bits bits must be readed to completly
 880  *                  read the longest vlc code
 881  *                  = (max_vlc_length + bits - 1) / bits
 882  */
 883 static always_inline int get_vlc2(GetBitContext *s, VLC_TYPE (*table)[2],
 884                                   int bits, int max_depth)
 885 {
 886     int code;
 887
 888     OPEN_READER(re, s)
 889     UPDATE_CACHE(re, s)
 890
 891     GET_VLC(code, re, s, table, bits, max_depth)
 892
 893     CLOSE_READER(re, s)
 894     return code;
 895 }
 896
 897 //#define TRACE
 898
 899 #ifdef TRACE
 900
 901 static inline void print_bin(int bits, int n){
 902     int i;
 903
 904     for(i=n-1; i>=0; i--){
 905         printf("%d", (bits>>i)&1);
 906     }
 907     for(i=n; i<24; i++)
 908         printf(" ");
 909 }
 910
 911 static inline int get_bits_trace(GetBitContext *s, int n, char *file, char *func, int line){
 912     int r= get_bits(s, n);
 913
 914     print_bin(r, n);
 915     printf("%5d %2d %3d bit @%5d in %s %s:%d\n", r, n, r, get_bits_count(s)-n, file, func, line);
 916     return r;
 917 }
 918 static inline int get_vlc_trace(GetBitContext *s, VLC_TYPE (*table)[2], int bits, int max_depth, char *file, char *func, int line){
 919     int show= show_bits(s, 24);
 920     int pos= get_bits_count(s);
 921     int r= get_vlc2(s, table, bits, max_depth);
 922     int len= get_bits_count(s) - pos;
 923     int bits2= show>>(24-len);
 924
 925     print_bin(bits2, len);
 926
 927     printf("%5d %2d %3d vlc @%5d in %s %s:%d\n", bits2, len, r, pos, file, func, line);
 928     return r;
 929 }
 930 static inline int get_xbits_trace(GetBitContext *s, int n, char *file, char *func, int line){
 931     int show= show_bits(s, n);
 932     int r= get_xbits(s, n);
 933
 934     print_bin(show, n);
 935     printf("%5d %2d %3d xbt @%5d in %s %s:%d\n", show, n, r, get_bits_count(s)-n, file, func, line);
 936     return r;
 937 }
 938
 939 #define get_bits(s, n)  get_bits_trace(s, n, __FILE__, __PRETTY_FUNCTION__, __LINE__)
 940 #define get_bits1(s)    get_bits_trace(s, 1, __FILE__, __PRETTY_FUNCTION__, __LINE__)
 941 #define get_xbits(s, n) get_xbits_trace(s, n, __FILE__, __PRETTY_FUNCTION__, __LINE__)
 942 #define get_vlc(s, vlc)            get_vlc_trace(s, (vlc)->table, (vlc)->bits, 3, __FILE__, __PRETTY_FUNCTION__, __LINE__)
 943 #define get_vlc2(s, tab, bits, max) get_vlc_trace(s, tab, bits, max, __FILE__, __PRETTY_FUNCTION__, __LINE__)
 944
 945 #define tprintf printf
 946
 947 #else //TRACE
 948 #define tprintf(_arg...) {}
 949 #endif
 950
 951 /* define it to include statistics code (useful only for optimizing
 952    codec efficiency */
 953 //#define STATS
 954
 955 #ifdef STATS
 956
 957 enum {
 958     ST_UNKNOWN,
 959     ST_DC,
 960     ST_INTRA_AC,
 961     ST_INTER_AC,
 962     ST_INTRA_MB,
 963     ST_INTER_MB,
 964     ST_MV,
 965     ST_NB,
 966 };
 967
 968 extern int st_current_index;
 969 extern unsigned int st_bit_counts[ST_NB];
 970 extern unsigned int st_out_bit_counts[ST_NB];
 971
 972 void print_stats(void);
 973 #endif
 974
 975 /* misc math functions */
 976 extern const uint8_t ff_log2_tab[256];
 977
 978 static inline int av_log2(unsigned int v)
 979 {
 980     int n;
 981
 982     n = 0;
 983     if (v & 0xffff0000) {
 984         v >>= 16;
 985         n += 16;
 986     }
 987     if (v & 0xff00) {
 988         v >>= 8;
 989         n += 8;
 990     }
 991     n += ff_log2_tab[v];
 992
 993     return n;
 994 }
 995
 996 static inline int av_log2_16bit(unsigned int v)
 997 {
 998     int n;
 999
1000     n = 0;
1001     if (v & 0xff00) {
1002         v >>= 8;
1003         n += 8;
1004     }
1005     n += ff_log2_tab[v];
1006
1007     return n;
1008 }
1009
1010
1011 /* median of 3 */
1012 static inline int mid_pred(int a, int b, int c)
1013 {
1014     int vmin, vmax;
1015     vmax = vmin = a;
1016     if (b < vmin)
1017         vmin = b;
1018     else
1019         vmax = b;
1020
1021     if (c < vmin)
1022         vmin = c;
1023     else if (c > vmax)
1024         vmax = c;
1025
1026     return a + b + c - vmin - vmax;
1027 }
1028
1029 static inline int clip(int a, int amin, int amax)
1030 {
1031     if (a < amin)
1032         return amin;
1033     else if (a > amax)
1034         return amax;
1035     else
1036         return a;
1037 }
1038
1039 /* math */
1040 extern const uint8_t ff_sqrt_tab[128];
1041
1042 int64_t ff_gcd(int64_t a, int64_t b);
1043
1044 static inline int ff_sqrt(int a)
1045 {
1046     int ret=0;
1047     int s;
1048     int ret_sq=0;
1049
1050     if(a<128) return ff_sqrt_tab[a];
1051
1052     for(s=15; s>=0; s--){
1053         int b= ret_sq + (1<<(s*2)) + (ret<<s)*2;
1054         if(b<=a){
1055             ret_sq=b;
1056             ret+= 1<<s;
1057         }
1058     }
1059     return ret;
1060 }
1061
1062 /**
1063  * converts fourcc string to int
1064  */
1065 static inline int ff_get_fourcc(const char *s){
1066     assert( strlen(s)==4 );
1067
1068     return (s[0]) + (s[1]<<8) + (s[2]<<16) + (s[3]<<24);
1069 }
1070
1071 #define MKTAG(a,b,c,d) (a | (b << 8) | (c << 16) | (d << 24))
1072 #define MKBETAG(a,b,c,d) (d | (c << 8) | (b << 16) | (a << 24))
1073
1074
1075 #ifdef ARCH_X86
1076 #define MASK_ABS(mask, level)\
1077             asm volatile(\
1078                 "cdq                    \n\t"\
1079                 "xorl %1, %0            \n\t"\
1080                 "subl %1, %0            \n\t"\
1081                 : "+a" (level), "=&d" (mask)\
1082             );
1083 #else
1084 #define MASK_ABS(mask, level)\
1085             mask= level>>31;\
1086             level= (level^mask)-mask;
1087 #endif
1088
1089
1090 #if __CPU__ >= 686 && !defined(RUNTIME_CPUDETECT)
1091 #define COPY3_IF_LT(x,y,a,b,c,d)\
1092 asm volatile (\
1093     "cmpl %0, %3        \n\t"\
1094     "cmovl %3, %0       \n\t"\
1095     "cmovl %4, %1       \n\t"\
1096     "cmovl %5, %2       \n\t"\
1097     : "+r" (x), "+r" (a), "+r" (c)\
1098     : "r" (y), "r" (b), "r" (d)\
1099 );
1100 #else
1101 #define COPY3_IF_LT(x,y,a,b,c,d)\
1102 if((y)<(x)){\
1103      (x)=(y);\
1104      (a)=(b);\
1105      (c)=(d);\
1106 }
1107 #endif
1108
1109 #ifdef ARCH_X86
1110 static inline long long rdtsc()
1111 {
1112         long long l;
1113         asm volatile(   "rdtsc\n\t"
1114                 : "=A" (l)
1115         );
1116         return l;
1117 }
1118
1119 #define START_TIMER \
1120 static uint64_t tsum=0;\
1121 static int tcount=0;\
1122 static int tskip_count=0;\
1123 uint64_t tend;\
1124 uint64_t tstart= rdtsc();\
1125
1126 #define STOP_TIMER(id) \
1127 tend= rdtsc();\
1128 if(tcount<2 || tend - tstart < 4*tsum/tcount){\
1129     tsum+= tend - tstart;\
1130     tcount++;\
1131 }else\
1132     tskip_count++;\
1133 if(256*256*256*64%(tcount+tskip_count)==0){\
1134     fprintf(stderr, "%Ld dezicycles in %s, %d runs, %d skips\n", tsum*10/tcount, id, tcount, tskip_count);\
1135 }
1136 #endif
1137
1138 #define CLAMP_TO_8BIT(d) ((d > 0xff) ? 0xff : (d < 0) ? 0 : d)
1139
1140 /* avoid usage of various functions */
1141 #define malloc please_use_av_malloc
1142 #define free please_use_av_free
1143 #define realloc please_use_av_realloc
1144
1145 #define CHECKED_ALLOCZ(p, size)\
1146 {\
1147     p= av_mallocz(size);\
1148     if(p==NULL && (size)!=0){\
1149         perror("malloc");\
1150         goto fail;\
1151     }\
1152 }
1153
1154 #endif /* HAVE_AV_CONFIG_H */
1155
1156 #endif /* COMMON_H */