cabac.h

从FFMPEG转换而来的H264解码程序,VC下编译..

        "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx   \n\t"
        "movb %%cl, (%1)                        \n\t"
        "add  $1, %0                            \n\t"
        "test %%bx, %%bx                        \n\t"
        " jnz 2f                                \n\t"

        "mov  "FFBYTE     "(%2), %%"REG_c"        \n\t"
        "movzwl (%%"REG_c"), %%esi              \n\t"
        "bswap %%esi                            \n\t"
        "shrl $15, %%esi                        \n\t"
        "subl $0xFFFF, %%esi                    \n\t"
        "add  $2, %%"REG_c"                     \n\t"
        "mov  %%"REG_c", "FFBYTE    "(%2)         \n\t"

        "leal -1(%%ebx), %%ecx                  \n\t"
        "xorl %%ebx, %%ecx                      \n\t"
        "shrl $15, %%ecx                        \n\t"
        "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx   \n\t"
        "neg %%ecx                              \n\t"
        "add $7, %%ecx                          \n\t"

        "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)
        : "%"REG_c, "%ebx", "%edx", "%"REG_S, "memory"
    );
    bit&=1;
#else /* BRANCHLESS_CABAC_DECODER */


#if defined HAVE_FAST_CMOV
#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"
#else /* HAVE_FAST_CMOV */
#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"
#endif /* HAVE_FAST_CMOV */


#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"\
        "mov "FFBYTE"("cabac"), %%"REG_c"                                 \n\t"\
        "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
        "bswap  "tmp"                                                   \n\t"\
        "shr    $15         , "tmp"                                     \n\t"\
        "sub    $0xFFFF     , "tmp"                                     \n\t"\
        "add    $2          , %%"REG_c"                                 \n\t"\
        "mov    %%"REG_c"   , "FFBYTE    "("cabac")                       \n\t"\
        "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"

    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")
        "movl %%esi, "RANGE    "(%2)            \n\t"
        "movl %%ebx, "LOW      "(%2)            \n\t"

        :"=&a"(bit)
        :"r"(state), "r"(c)
        : "%"REG_c, "%ebx", "%edx", "%esi", "memory"
    );
    bit&=1;
#endif /* BRANCHLESS_CABAC_DECODER */
#else /* defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
    int s = *state;
    int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s];
    int bit, lps_mask av_unused;

    c->range -= RangeLPS;
#ifndef BRANCHLESS_CABAC_DECODER
    if(c->low < (c->range<<(CABAC_BITS+1))){
        bit= s&1;
        *state= ff_h264_mps_state[s];
        renorm_cabac_decoder_once(c);
    }else{
        bit= ff_h264_norm_shift[RangeLPS];
        c->low -= (c->range<<(CABAC_BITS+1));
        *state= ff_h264_lps_state[s];
        c->range = RangeLPS<<bit;
        c->low <<= bit;
        bit= (s&1)^1;

        if(!(c->low & CABAC_MASK)){
            refill2(c);
        }
    }
#else /* BRANCHLESS_CABAC_DECODER */
    lps_mask= ((c->range<<(CABAC_BITS+1)) - c->low)>>31;

    c->low -= (c->range<<(CABAC_BITS+1)) & lps_mask;
    c->range += (RangeLPS - c->range) & lps_mask;

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

    lps_mask= ff_h264_norm_shift[c->range];
    c->range<<= lps_mask;
    c->low  <<= lps_mask;
    if(!(c->low & CABAC_MASK))
        refill2(c);
#endif /* BRANCHLESS_CABAC_DECODER */
#endif /* defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
    return bit;
}

#ifdef __GNUC__
static int av_noinline get_cabac_noinline(CABACContext *c, uint8_t * const state){
#else
static int get_cabac_noinline(CABACContext *c, uint8_t * const state){
#endif
    return get_cabac_inline(c,state);
}

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

static int get_cabac_bypass(CABACContext *c){
#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"
        "movl "FFBYTE     "(%1), %%"REG_b"        \n\t"
        "subl $0xFFFF, %%eax                    \n\t"
        "movzwl (%%"REG_b"), %%ecx              \n\t"
        "bswap %%ecx                            \n\t"
        "shrl $15, %%ecx                        \n\t"
        "addl $2, %%"REG_b"                     \n\t"
        "addl %%ecx, %%eax                      \n\t"
        "movl %%"REG_b", "FFBYTE     "(%1)        \n\t"
        "1:                                     \n\t"
        "movl %%eax, "LOW      "(%1)            \n\t"

        :"=&d"(bit)
        :"r"(c)
        : "%eax", "%"REG_b, "%ecx", "memory"
    );
    return bit+1;
#else
    int range;
    c->low += c->low;

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

    range= c->range<<(CABAC_BITS+1);
    if(c->low < range){
        return 0;
    }else{
        c->low -= range;
        return 1;
    }
#endif
}


static av_always_inline int get_cabac_bypass_sign(CABACContext *c, int val){
#if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
    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"
        "mov  "FFBYTE     "(%1), %%"REG_b"        \n\t"
        "subl $0xFFFF, %%eax                    \n\t"
        "movzwl (%%"REG_b"), %%edx              \n\t"
        "bswap %%edx                            \n\t"
        "shrl $15, %%edx                        \n\t"
        "add  $2, %%"REG_b"                     \n\t"
        "addl %%edx, %%eax                      \n\t"
        "mov  %%"REG_b", "FFBYTE     "(%1)        \n\t"
        "1:                                     \n\t"
        "movl %%eax, "LOW      "(%1)            \n\t"

        :"+c"(val)
        :"r"(c)
        : "%eax", "%"REG_b, "%edx", "memory"
    );
    return val;
#else
    int range, mask;
    c->low += c->low;

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

    range= c->range<<(CABAC_BITS+1);
    c->low -= range;
    mask= c->low >> 31;
    range &= mask;
    c->low += range;
    return (val^mask)-mask;
#endif
}

//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)
#if defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS)
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;
    int minusindex= 4-(int)index;
    int coeff_count;
    asm volatile(
        "movl "RANGE    "(%3), %%esi            \n\t"
        "movl "LOW      "(%3), %%ebx            \n\t"

        "2:                                     \n\t"

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

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

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

        "mov  %2, %%"REG_a"                     \n\t"
        "movl %4, %%ecx                         \n\t"
        "add  %1, %%"REG_c"                     \n\t"
        "movl %%ecx, (%%"REG_a")                \n\t"

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

        "add  $4, %%"REG_a"                     \n\t"
        "mov  %%"REG_a", %2                     \n\t"

        "3:                                     \n\t"
        "add  $1, %1                            \n\t"
        "cmp  %5, %1                            \n\t"
        " jb 2b                                 \n\t"
        "mov  %2, %%"REG_a"                     \n\t"
        "movl %4, %%ecx                         \n\t"
        "add  %1, %%"REG_c"                     \n\t"
        "movl %%ecx, (%%"REG_a")                \n\t"
        "4:                                     \n\t"
        "add  %6, %%eax                         \n\t"
        "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)\
        : "%"REG_c, "%ebx", "%edx", "%esi", "memory"\
    );
    return coeff_count;
}

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;
    long last=0;
    asm volatile(
        "movl "RANGE    "(%3), %%esi            \n\t"
        "movl "LOW      "(%3), %%ebx            \n\t"

        "mov %1, %%"REG_D"                      \n\t"
        "2:                                     \n\t"

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

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

        "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"
        "add %5, %%"REG_D"                      \n\t"

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

        "mov %2, %%"REG_a"                      \n\t"
        "mov %1, %%edi                          \n\t"
        "movl %%edi, (%%"REG_a")                \n\t"

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

        "add $4, %%"REG_a"                      \n\t"
        "mov %%"REG_a", %2                      \n\t"

        "3:                                     \n\t"
        "addl $1, %%edi                         \n\t"
        "mov %%edi, %1                          \n\t"
        "cmpl $63, %%edi                        \n\t"
        " jb 2b                                 \n\t"
        "mov %2, %%"REG_a"                      \n\t"
        "movl %%edi, (%%"REG_a")                \n\t"
        "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)\
        : "%"REG_c, "%ebx", "%edx", "%esi", "%"REG_D, "memory"\
    );
    return coeff_count;
}
#endif /* defined(ARCH_X86) && && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */

/**
 *
 * @return the number of bytes read or 0 if no end
 */
static int get_cabac_terminate(CABACContext *c){
    c->range -= 2;
    if(c->low < c->range<<(CABAC_BITS+1)){
        renorm_cabac_decoder_once(c);
        return 0;
    }else{
        return c->bytestream - c->bytestream_start;
    }
}

/**
 * get (truncated) unnary binarization.
 */
static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){
    int i;

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

        if(i< max_index) state++;
    }

    return truncated ? max : -1;
}

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

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

    if(0 < max_index) state++;

    for(i=1; i<max; i++){
        if(get_cabac(c, state)==0){
            if(is_signed && get_cabac_bypass(c)){
                return -i;
            }else
                return i;
        }

        if(i < max_index) state++;
    }

    while(get_cabac_bypass(c)){
        i+= m;
        m+= m;
    }

    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;
}

#endif /* CABAC_H */