comparison cabac.h @ 3950:900d21b85dd6 libavcodec

renorm_cabac_decoder_once START/STOP_TIMER scores for athlon

3949:6020c2f3b16b

static inline void renorm_cabac_decoder_once(CABACContext *c){
#ifdef ARCH_X86
    int temp;
#if 0
    //P3:683
    asm(
        "lea -0x2000000(%0), %2     \n\t"
        "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
    //P3:680
    asm(
        "cmp $0x2000000, %0         \n\t"
        "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;
    //P3:665
    asm(
        "lea -0x2000000(%0), %%eax  \n\t"
        "cdq                        \n\t"
        "mov %0, %%eax              \n\t"
        "and %%edx, %0              \n\t"

        "add %%edx, %1              \n\t"
        : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
    );
#elif 0
    int temp2;
    //P3:673
    asm(
        "cmp $0x2000000, %0         \n\t"
        "sbb %%edx, %%edx           \n\t"
        "mov %0, %%eax              \n\t"
        "and %%edx, %0              \n\t"

        "add %%edx, %1              \n\t"
        : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
    );
#else
    int temp2;
    //P3:677
    asm(
        "cmp $0x2000000, %0         \n\t"
        "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
    //P3:675
    int shift= (uint32_t)(c->range - (0x200 << CABAC_BITS))>>31;
    c->range<<= shift;
    c->low  <<= shift;
#endif
    if(!(c->low & CABAC_MASK))

3950:900d21b85dd6

static inline void renorm_cabac_decoder_once(CABACContext *c){
#ifdef ARCH_X86
    int temp;
#if 0
    //P3:683    athlon:475
    asm(
        "lea -0x2000000(%0), %2     \n\t"
        "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
    //P3:680    athlon:474
    asm(
        "cmp $0x2000000, %0         \n\t"
        "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;
    //P3:665    athlon:517
    asm(
        "lea -0x2000000(%0), %%eax  \n\t"
        "cdq                        \n\t"
        "mov %0, %%eax              \n\t"
        "and %%edx, %0              \n\t"

        "add %%edx, %1              \n\t"
        : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
    );
#elif 0
    int temp2;
    //P3:673    athlon:509
    asm(
        "cmp $0x2000000, %0         \n\t"
        "sbb %%edx, %%edx           \n\t"
        "mov %0, %%eax              \n\t"
        "and %%edx, %0              \n\t"

        "add %%edx, %1              \n\t"
        : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
    );
#else
    int temp2;
    //P3:677    athlon:511
    asm(
        "cmp $0x2000000, %0         \n\t"
        "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
    //P3:675    athlon:476
    int shift= (uint32_t)(c->range - (0x200 << CABAC_BITS))>>31;
    c->range<<= shift;
    c->low  <<= shift;
#endif
    if(!(c->low & CABAC_MASK))