static __inline __m128i_mm_unpackhi_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B);}static __inline __m128i_mm_unpackhi_epi64 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_unpacklo_epi8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_unpacklo_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_unpacklo_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B);}static __inline __m128i_mm_unpacklo_epi64 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_add_epi8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_paddb128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_add_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_paddw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_add_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_paddd128 ((__v4si)__A, (__v4si)__B);}static __inline __m128i_mm_add_epi64 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_paddq128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_adds_epi8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_adds_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_adds_epu8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_adds_epu16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_sub_epi8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psubb128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_sub_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psubw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_sub_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psubd128 ((__v4si)__A, (__v4si)__B);}static __inline __m128i_mm_sub_epi64 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psubq128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_subs_epi8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_subs_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_subs_epu8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_subs_epu16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_madd_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_mulhi_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_mullo_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pmullw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m64_mm_mul_su32 (__m64 __A, __m64 __B){  return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B);}static __inline __m128i_mm_mul_epu32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B);}static __inline __m128i_mm_sll_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psllw128 ((__v8hi)__A, (__v2di)__B);}static __inline __m128i_mm_sll_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pslld128 ((__v4si)__A, (__v2di)__B);}static __inline __m128i_mm_sll_epi64 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psllq128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_sra_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v2di)__B);}static __inline __m128i_mm_sra_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v2di)__B);}static __inline __m128i_mm_srl_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v2di)__B);}static __inline __m128i_mm_srl_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v2di)__B);}static __inline __m128i_mm_srl_epi64 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_slli_epi16 (__m128i __A, int __B){  return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B);}static __inline __m128i_mm_slli_epi32 (__m128i __A, int __B){  return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B);}static __inline __m128i_mm_slli_epi64 (__m128i __A, int __B){  return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B);}static __inline __m128i_mm_srai_epi16 (__m128i __A, int __B){  return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B);}static __inline __m128i_mm_srai_epi32 (__m128i __A, int __B){  return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B);}#if 0static __m128i __attribute__((__always_inline__))_mm_srli_si128 (__m128i __A, const int __B){  return ((__m128i)__builtin_ia32_psrldqi128 (__A, __B))}static __m128i __attribute__((__always_inline__))_mm_srli_si128 (__m128i __A, const int __B){  return ((__m128i)__builtin_ia32_pslldqi128 (__A, __B))}#endif#define _mm_srli_si128(__A, __B) ((__m128i)__builtin_ia32_psrldqi128 (__A, __B))#define _mm_slli_si128(__A, __B) ((__m128i)__builtin_ia32_pslldqi128 (__A, __B))static __inline __m128i_mm_srli_epi16 (__m128i __A, int __B){  return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B);}static __inline __m128i_mm_srli_epi32 (__m128i __A, int __B){  return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B);}static __inline __m128i_mm_srli_epi64 (__m128i __A, int __B){  return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B);}static __inline __m128i_mm_and_si128 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pand128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_andnot_si128 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_or_si128 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_por128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_xor_si128 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pxor128 ((__v2di)__A, (__v2di)__B);}static __inline __m128i_mm_cmpeq_epi8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pcmpeqb128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_cmpeq_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pcmpeqw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_cmpeq_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pcmpeqd128 ((__v4si)__A, (__v4si)__B);}static __inline __m128i_mm_cmplt_epi8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__B, (__v16qi)__A);}static __inline __m128i_mm_cmplt_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__B, (__v8hi)__A);}static __inline __m128i_mm_cmplt_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__B, (__v4si)__A);}static __inline __m128i_mm_cmpgt_epi8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_cmpgt_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_cmpgt_epi32 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__A, (__v4si)__B);}#define _mm_extract_epi16(__A, __B) __builtin_ia32_pextrw128 ((__v8hi)__A, __B)#define _mm_insert_epi16(__A, __B, __C) ((__m128i)__builtin_ia32_pinsrw128 ((__v8hi)__A, __B, __C))static __inline __m128i_mm_max_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_max_epu8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_min_epi16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_min_epu8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B);}static __inline int_mm_movemask_epi8 (__m128i __A){  return __builtin_ia32_pmovmskb128 ((__v16qi)__A);}static __inline __m128i_mm_mulhi_epu16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);}#define _mm_shufflehi_epi16(__A, __B) ((__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __B))#define _mm_shufflelo_epi16(__A, __B) ((__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __B))#define _mm_shuffle_epi32(__A, __B) ((__m128i)__builtin_ia32_pshufd ((__v4si)__A, __B))static __inline void_mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C){  __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C);}static __inline __m128i_mm_avg_epu8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B);}static __inline __m128i_mm_avg_epu16 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B);}static __inline __m128i_mm_sad_epu8 (__m128i __A, __m128i __B){  return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B);}static __inline void_mm_stream_si32 (int *__A, int __B){  __builtin_ia32_movnti (__A, __B);}static __inline void_mm_stream_si128 (__m128i *__A, __m128i __B){  __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B);}static __inline void_mm_stream_pd (double *__A, __m128d __B){  __builtin_ia32_movntpd (__A, (__v2df)__B);}static __inline __m128i_mm_movpi64_epi64 (__m64 __A){  return (__m128i)__builtin_ia32_movq2dq ((unsigned long long)__A);}static __inline void_mm_clflush (void const *__A){  return __builtin_ia32_clflush (__A);}static __inline void_mm_lfence (void){  __builtin_ia32_lfence ();}static __inline void_mm_mfence (void){  __builtin_ia32_mfence ();}static __inline __m128i_mm_cvtsi32_si128 (int __A){  return (__m128i) __builtin_ia32_loadd (&__A);}#ifdef __x86_64__static __inline __m128i_mm_cvtsi64x_si128 (long long __A){  return (__m128i) __builtin_ia32_movq2dq (__A);}#endifstatic __inline int_mm_cvtsi128_si32 (__m128i __A){  int __tmp;  __builtin_ia32_stored (&__tmp, (__v4si)__A);  return __tmp;}#ifdef __x86_64__static __inline long long_mm_cvtsi128_si64x (__m128i __A){  return __builtin_ia32_movdq2q ((__v2di)__A);}#endif#endif /* __SSE2__  */#endif /* _EMMINTRIN_H_INCLUDED */