x64 (amd64) Intrinsics List

Article
2012-07-10

This document shows the intrinsics that are supported in Visual C++ in Visual Studio 2012 when x64 (also referred to as amd64) is targeted.

For more information, see these resources:

Intrinsic name	Function prototype
__addgsbyte	void __addgsbyte(unsigned long, unsigned char);
__addgsdword	void __addgsdword(unsigned long, unsigned long);
__addgsqword	void __addgsqword(unsigned long, unsigned __int64);
__addgsword	void __addgsword(unsigned long, unsigned short);
__code_seg	void __code_seg(const char *);
__cpuid	void __cpuid(int[4], int);
__cpuidex	void __cpuidex(int[4], int, int);
__debugbreak	void __cdecl __debugbreak(void);
__emul	__int64 __emul(int, int);
__emulu	unsigned __int64 __emulu(unsigned int, unsigned int);
__fastfail	void __fastfail(unsigned int);
__faststorefence	void __faststorefence(void);
__getcallerseflags	unsigned int __getcallerseflags(void);
__halt	void __halt(void);
__inbyte	unsigned char __inbyte(unsigned short);
__inbytestring	void __inbytestring(unsigned short, unsigned char *, unsigned long);
__incgsbyte	void __incgsbyte(unsigned long);
__incgsdword	void __incgsdword(unsigned long);
__incgsqword	void __incgsqword(unsigned long);
__incgsword	void __incgsword(unsigned long);
__indword	unsigned long __indword(unsigned short);
__indwordstring	void __indwordstring(unsigned short, unsigned long *, unsigned long);
__int2c	void __int2c(void);
__invlpg	void __invlpg(void *);
__inword	unsigned short __inword(unsigned short);
__inwordstring	void __inwordstring(unsigned short, unsigned short *, unsigned long);
__lidt	void __lidt(void *);
__ll_lshift	unsigned __int64 __ll_lshift(unsigned __int64, int);
__ll_rshift	__int64 __ll_rshift(__int64, int);
__llwpcb	void __llwpcb(void *);
__lwpins32	unsigned char __lwpins32(unsigned int, unsigned int, unsigned int);
__lwpins64	unsigned char __lwpins64(unsigned __int64, unsigned int, unsigned int);
__lwpval32	void __lwpval32(unsigned int, unsigned int, unsigned int);
__lwpval64	void __lwpval64(unsigned __int64, unsigned int, unsigned int);
__lzcnt	unsigned int __lzcnt(unsigned int);
__lzcnt16	unsigned short __lzcnt16(unsigned short);
__lzcnt64	unsigned __int64 __lzcnt64(unsigned __int64);
__movsb	void __movsb(unsigned char , unsigned char const , size_t);
__movsd	void __movsd(unsigned long , unsigned long const , size_t);
__movsq	void __movsq(unsigned long long , unsigned long long const , size_t);
__movsw	void __movsw(unsigned short , unsigned short const , size_t);
__mulh	__int64 __mulh(__int64, __int64);
__nop	void __nop(void);
__nvreg_restore_fence	void __nvreg_restore_fence(void);
__nvreg_save_fence	void __nvreg_save_fence(void);
__outbyte	void __outbyte(unsigned short, unsigned char);
__outbytestring	void __outbytestring(unsigned short, unsigned char *, unsigned long);
__outdword	void __outdword(unsigned short, unsigned long);
__outdwordstring	void __outdwordstring(unsigned short, unsigned long *, unsigned long);
__outword	void __outword(unsigned short, unsigned short);
__outwordstring	void __outwordstring(unsigned short, unsigned short *, unsigned long);
__popcnt	unsigned int __popcnt(unsigned int);
__popcnt16	unsigned short __popcnt16(unsigned short);
__popcnt64	unsigned __int64 __popcnt64(unsigned __int64);
__rdtsc	unsigned __int64 __rdtsc(void);
__rdtscp	unsigned __int64 __rdtscp(unsigned int *);
__readcr0	unsigned __int64 __readcr0(void);
__readcr2	unsigned __int64 __readcr2(void);
__readcr3	unsigned __int64 __readcr3(void);
__readcr4	unsigned __int64 __readcr4(void);
__readcr8	unsigned __int64 __readcr8(void);
__readdr	unsigned __int64 __readdr(unsigned int);
__readeflags	unsigned __int64 __readeflags(void);
__readgsbyte	unsigned char __readgsbyte(unsigned long);
__readgsdword	unsigned long __readgsdword(unsigned long);
__readgsqword	unsigned __int64 __readgsqword(unsigned long);
__readgsword	unsigned short __readgsword(unsigned long);
__readmsr	unsigned __int64 __readmsr(unsigned long);
__readpmc	unsigned __int64 __readpmc(unsigned long);
__segmentlimit	unsigned long __segmentlimit(unsigned long);
__shiftleft128	unsigned __int64 __shiftleft128(unsigned __int64 _LowPart, unsigned __int64 _HighPart, unsigned char _Shift);
__shiftright128	unsigned __int64 __shiftright128(unsigned __int64 _LowPart, unsigned __int64 _HighPart, unsigned char _Shift);
__sidt	void __sidt(void *);
__slwpcb	void *__slwpcb();
__stosb	void __stosb(unsigned char *, unsigned char, size_t);
__stosd	void __stosd(unsigned long *, unsigned long, size_t);
__stosq	void __stosq(unsigned __int64 *, unsigned __int64, size_t);
__stosw	void __stosw(unsigned short *, unsigned short, size_t);
__svm_clgi	void __svm_clgi(void);
__svm_invlpga	void __svm_invlpga(void *, int);
__svm_skinit	void __svm_skinit(int);
__svm_stgi	void __svm_stgi(void);
__svm_vmload	void __svm_vmload(size_t);
__svm_vmrun	void __svm_vmrun(size_t);
__svm_vmsave	void __svm_vmsave(size_t);
__ud2	void __ud2(void);
__ull_rshift	unsigned __int64 __ull_rshift(unsigned __int64, int);
__umulh	unsigned __int64 __umulh(unsigned __int64, unsigned __int64);
__vmx_off	void __vmx_off(void);
__vmx_on	unsigned char __vmx_on(unsigned __int64 *);
__vmx_vmclear	unsigned char __vmx_vmclear(unsigned __int64 *);
__vmx_vmlaunch	unsigned char __vmx_vmlaunch(void);
__vmx_vmptrld	unsigned char __vmx_vmptrld(unsigned __int64 *);
__vmx_vmptrst	void __vmx_vmptrst(unsigned __int64 *);
__vmx_vmread	unsigned char __vmx_vmread(size_t, size_t *);
__vmx_vmresume	unsigned char __vmx_vmresume(void);
__vmx_vmwrite	unsigned char __vmx_vmwrite(size_t, size_t);
__wbinvd	void __wbinvd(void);
__writecr0	void __writecr0(unsigned __int64);
__writecr3	void __writecr3(unsigned __int64);
__writecr4	void __writecr4(unsigned __int64);
__writecr8	void __writecr8(unsigned __int64);
__writedr	void __writedr(unsigned int, unsigned __int64);
__writeeflags	void __writeeflags(unsigned __int64);
__writegsbyte	void __writegsbyte(unsigned long, unsigned char);
__writegsdword	void __writegsdword(unsigned long, unsigned long);
__writegsqword	void __writegsqword(unsigned long, unsigned __int64);
__writegsword	void __writegsword(unsigned long, unsigned short);
__writemsr	void __writemsr(unsigned long, unsigned __int64);
_AddressOfReturnAddress	void * _AddressOfReturnAddress(void);
_andn_u32	unsigned int _andn_u32(unsigned int, unsigned int);
_andn_u64	unsigned __int64 _andn_u64(unsigned __int64, unsigned __int64);
_bextr_u32	unsigned int _bextr_u32(unsigned int, unsigned int, unsigned int );
_bextr_u32	unsigned int _bextr_u32(unsigned int, unsigned int, unsigned int);
_bextr_u64	unsigned __int64 _bextr_u64(unsigned __int64, unsigned int, unsigned int );
_bextr_u64	unsigned __int64 _bextr_u64(unsigned __int64, unsigned int, unsigned int);
_bextri_u32	unsigned int _bextri_u32(unsigned int, unsigned int);
_bextri_u64	unsigned __int64 _bextri_u64(unsigned __int64, unsigned int);
_BitScanForward	unsigned char _BitScanForward(unsigned long * _Index, unsigned long _Mask);
_BitScanForward64	unsigned char _BitScanForward64(unsigned long * _Index, unsigned __int64 _Mask);
_BitScanReverse	unsigned char _BitScanReverse(unsigned long * _Index, unsigned long _Mask);
_BitScanReverse64	unsigned char _BitScanReverse64(unsigned long * _Index, unsigned __int64 _Mask);
_bittest	unsigned char _bittest(long const *, long);
_bittest64	unsigned char _bittest64(__int64 const *, __int64);
_bittestandcomplement	unsigned char _bittestandcomplement(long *, long);
_bittestandcomplement64	unsigned char _bittestandcomplement64(__int64 *, __int64);
_bittestandreset	unsigned char _bittestandreset(long *, long);
_bittestandreset64	unsigned char _bittestandreset64(__int64 *, __int64);
_bittestandset	unsigned char _bittestandset(long *, long);
_bittestandset64	unsigned char _bittestandset64(__int64 *, __int64);
_blcfill_u32	unsigned int _blcfill_u32(unsigned int);
_blcfill_u64	unsigned __int64 _blcfill_u64(unsigned __int64);
_blci_u32	unsigned int _blci_u32(unsigned int);
_blci_u64	unsigned __int64 _blci_u64(unsigned __int64);
_blcic_u32	unsigned int _blcic_u32(unsigned int);
_blcic_u64	unsigned __int64 _blcic_u64(unsigned __int64);
_blcmsk_u32	unsigned int _blcmsk_u32(unsigned int);
_blcmsk_u64	unsigned __int64 _blcmsk_u64(unsigned __int64);
_blcs_u32	unsigned int _blcs_u32(unsigned int);
_blcs_u64	unsigned __int64 _blcs_u64(unsigned __int64);
_blsfill_u32	unsigned int _blsfill_u32(unsigned int);
_blsfill_u64	unsigned __int64 _blsfill_u64(unsigned __int64);
_blsi_u32	unsigned int _blsi_u32(unsigned int);
_blsi_u32	unsigned int _blsi_u32(unsigned int);
_blsi_u64	unsigned __int64 _blsi_u64(unsigned __int64);
_blsi_u64	unsigned __int64 _blsi_u64(unsigned __int64);
_blsic_u32	unsigned int _blsic_u32(unsigned int);
_blsic_u64	unsigned __int64 _blsic_u64(unsigned __int64);
_blsmsk_u32	unsigned int _blsmsk_u32(unsigned int);
_blsmsk_u32	unsigned int _blsmsk_u32(unsigned int);
_blsmsk_u64	unsigned __int64 _blsmsk_u64(unsigned __int64);
_blsmsk_u64	unsigned __int64 _blsmsk_u64(unsigned __int64);
_blsr_u32	unsigned int _blsr_u32(unsigned int);
_blsr_u32	unsigned int _blsr_u32(unsigned int);
_blsr_u64	unsigned __int64 _blsr_u64(unsigned __int64);
_blsr_u64	unsigned __int64 _blsr_u64(unsigned __int64);
_byteswap_uint64	unsigned __int64 __cdecl _byteswap_uint64(unsigned __int64);
_byteswap_ulong	unsigned long __cdecl _byteswap_ulong(unsigned long);
_byteswap_ushort	unsigned short __cdecl _byteswap_ushort(unsigned short);
_bzhi_u32	unsigned int _bzhi_u32(unsigned int, unsigned int );
_bzhi_u64	unsigned __int64 _bzhi_u64(unsigned __int64, unsigned int );
_disable	void __cdecl _disable(void);
_enable	void __cdecl _enable(void);
_fxrstor	void __cdecl _fxrstor(void const *);
_fxrstor64	void __cdecl _fxrstor64(void const *);
_fxsave	void __cdecl _fxsave(void *);
_fxsave64	void __cdecl _fxsave64(void *);
_InterlockedAnd	long _InterlockedAnd(long volatile * _Value, long _Mask);
_InterlockedAnd_np	long _InterlockedAnd_np(long volatile * _Value, long _Mask);
_InterlockedAnd16	short _InterlockedAnd16(short volatile * _Value, short _Mask);
_InterlockedAnd16_np	short _InterlockedAnd16_np(short volatile * _Value, short _Mask);
_InterlockedAnd64	__int64 _InterlockedAnd64(__int64 volatile * _Value, __int64 _Mask);
_InterlockedAnd64_np	__int64 _InterlockedAnd64_np(__int64 volatile * _Value, __int64 _Mask);
_InterlockedAnd8	char _InterlockedAnd8(char volatile * _Value, char _Mask);
_InterlockedAnd8_np	char _InterlockedAnd8_np(char volatile * _Value, char _Mask);
_interlockedbittestandreset	unsigned char _interlockedbittestandreset(long volatile *, long);
_interlockedbittestandreset64	unsigned char _interlockedbittestandreset64(__int64 volatile *, __int64);
_interlockedbittestandset	unsigned char _interlockedbittestandset(long volatile *, long);
_interlockedbittestandset64	unsigned char _interlockedbittestandset64(__int64 volatile *, __int64);
_InterlockedCompareExchange	long __cdecl _InterlockedCompareExchange(long volatile * _Destination, long _Exchange, long _Comparand);
_InterlockedCompareExchange_HLEAcquire	long _InterlockedCompareExchange_HLEAcquire(long volatile *,long,long);
_InterlockedCompareExchange_HLERelease	long _InterlockedCompareExchange_HLERelease(long volatile *,long,long);
_InterlockedCompareExchange_np	long _InterlockedCompareExchange_np(long volatile * _Destination, long _Exchange, long _Comparand);
_InterlockedCompareExchange128	unsigned char _InterlockedCompareExchange128(__int64 volatile * _Destination, __int64 _ExchangeHigh, __int64 _ExchangeLow, __int64 * _ComparandResult);
_InterlockedCompareExchange128_np	unsigned char _InterlockedCompareExchange128_np(__int64 volatile * _Destination, __int64 _ExchangeHigh, __int64 _ExchangeLow, __int64 * _ComparandResult);
_InterlockedCompareExchange16	short _InterlockedCompareExchange16(short volatile * _Destination, short _Exchange, short _Comparand);
_InterlockedCompareExchange16_np	short _InterlockedCompareExchange16_np(short volatile * _Destination, short _Exchange, short _Comparand);
_InterlockedCompareExchange64	__int64 _InterlockedCompareExchange64(__int64 volatile * _Destination, __int64 _Exchange, __int64 _Comparand);
_InterlockedCompareExchange64_HLEAcquire	__int64 _InterlockedCompareExchange64_HLEAcquire(__int64 volatile *,__int64,__int64);
_InterlockedCompareExchange64_HLERelease	__int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *,__int64,__int64);
_InterlockedCompareExchange64_np	__int64 _InterlockedCompareExchange64_np(__int64 volatile * _Destination, __int64 _Exchange, __int64 _Comparand);
_InterlockedCompareExchange8	char _InterlockedCompareExchange8(char volatile * _Destination, char _Exchange, char _Comparand);
_InterlockedCompareExchangePointer	void * _InterlockedCompareExchangePointer(void * volatile * _Destination, void * _Exchange, void * _Comparand);
_InterlockedCompareExchangePointer_HLEAcquire	void _InterlockedCompareExchangePointer_HLEAcquire(void volatile ,void ,void *);
_InterlockedCompareExchangePointer_HLERelease	void _InterlockedCompareExchangePointer_HLERelease(void volatile ,void ,void *);
_InterlockedCompareExchangePointer_np	void * _InterlockedCompareExchangePointer_np(void * volatile * _Destination, void * _Exchange, void * _Comparand);
_InterlockedDecrement	long __cdecl _InterlockedDecrement(long volatile * _Addend);
_InterlockedDecrement16	short _InterlockedDecrement16(short volatile * _Addend);
_InterlockedDecrement64	__int64 _InterlockedDecrement64(__int64 volatile * _Addend);
_InterlockedExchange	long __cdecl _InterlockedExchange(long volatile * _Target, long _Value);
_InterlockedExchange16	short _InterlockedExchange16(short volatile * _Target, short _Value);
_InterlockedExchange64	__int64 _InterlockedExchange64(__int64 volatile * _Target, __int64 _Value);
_InterlockedExchange8	char _InterlockedExchange8(char volatile * _Target, char _Value);
_InterlockedExchangeAdd	long __cdecl _InterlockedExchangeAdd(long volatile * _Addend, long _Value);
_InterlockedExchangeAdd_HLEAcquire	long _InterlockedExchangeAdd_HLEAcquire(long volatile *,long);
_InterlockedExchangeAdd_HLERelease	long _InterlockedExchangeAdd_HLERelease(long volatile *,long);
_InterlockedExchangeAdd16	short _InterlockedExchangeAdd16(short volatile * _Addend, short _Value);
_InterlockedExchangeAdd64	__int64 _InterlockedExchangeAdd64(__int64 volatile * _Addend, __int64 _Value);
_InterlockedExchangeAdd64_HLEAcquire	__int64 _InterlockedExchangeAdd64_HLEAcquire(__int64 volatile *,__int64);
_InterlockedExchangeAdd64_HLERelease	__int64 _InterlockedExchangeAdd64_HLERelease(__int64 volatile *,__int64);
_InterlockedExchangeAdd8	char _InterlockedExchangeAdd8(char volatile * _Addend, char _Value);
_InterlockedExchangePointer	void * _InterlockedExchangePointer(void * volatile * _Target, void * _Value);
_InterlockedIncrement	long __cdecl _InterlockedIncrement(long volatile * _Addend);
_InterlockedIncrement16	short _InterlockedIncrement16(short volatile * _Addend);
_InterlockedIncrement64	__int64 _InterlockedIncrement64(__int64 volatile * _Addend);
_InterlockedOr	long _InterlockedOr(long volatile * _Value, long _Mask);
_InterlockedOr_np	long _InterlockedOr_np(long volatile * _Value, long _Mask);
_InterlockedOr16	short _InterlockedOr16(short volatile * _Value, short _Mask);
_InterlockedOr16_np	short _InterlockedOr16_np(short volatile * _Value, short _Mask);
_InterlockedOr64	__int64 _InterlockedOr64(__int64 volatile * _Value, __int64 _Mask);
_InterlockedOr64_np	__int64 _InterlockedOr64_np(__int64 volatile * _Value, __int64 _Mask);
_InterlockedOr8	char _InterlockedOr8(char volatile * _Value, char _Mask);
_InterlockedOr8_np	char _InterlockedOr8_np(char volatile * _Value, char _Mask);
_InterlockedXor	long _InterlockedXor(long volatile * _Value, long _Mask);
_InterlockedXor_np	long _InterlockedXor_np(long volatile * _Value, long _Mask);
_InterlockedXor16	short _InterlockedXor16(short volatile * _Value, short _Mask);
_InterlockedXor16_np	short _InterlockedXor16_np(short volatile * _Value, short _Mask);
_InterlockedXor64	__int64 _InterlockedXor64(__int64 volatile * _Value, __int64 _Mask);
_InterlockedXor64_np	__int64 _InterlockedXor64_np(__int64 volatile * _Value, __int64 _Mask);
_InterlockedXor8	char _InterlockedXor8(char volatile * _Value, char _Mask);
_InterlockedXor8_np	char _InterlockedXor8_np(char volatile * _Value, char _Mask);
_invpcid	void __cdecl _invpcid(unsigned int, void * );
_lrotl	unsigned long __cdecl _lrotl(unsigned long, int);
_lrotr	unsigned long __cdecl _lrotr(unsigned long, int);
_lzcnt_u32	unsigned int _lzcnt_u32(unsigned int);
_lzcnt_u32	unsigned int _lzcnt_u32(unsigned int);
_lzcnt_u64	unsigned __int64 _lzcnt_u64(unsigned __int64);
_lzcnt_u64	unsigned __int64 _lzcnt_u64(unsigned __int64);
_m_empty	void _m_empty(void);
_m_from_int	__m64 _m_from_int(int _I);
_m_maskmovq	void _m_maskmovq(__m64, __m64, char *);
_m_packssdw	__m64 _m_packssdw(__m64 _MM1, __m64 _MM2);
_m_packsswb	__m64 _m_packsswb(__m64 _MM1, __m64 _MM2);
_m_packuswb	__m64 _m_packuswb(__m64 _MM1, __m64 _MM2);
_m_paddb	__m64 _m_paddb(__m64 _MM1, __m64 _MM2);
_m_paddd	__m64 _m_paddd(__m64 _MM1, __m64 _MM2);
_m_paddsb	__m64 _m_paddsb(__m64 _MM1, __m64 _MM2);
_m_paddsw	__m64 _m_paddsw(__m64 _MM1, __m64 _MM2);
_m_paddusb	__m64 _m_paddusb(__m64 _MM1, __m64 _MM2);
_m_paddusw	__m64 _m_paddusw(__m64 _MM1, __m64 _MM2);
_m_paddw	__m64 _m_paddw(__m64 _MM1, __m64 _MM2);
_m_pand	__m64 _m_pand(__m64 _MM1, __m64 _MM2);
_m_pandn	__m64 _m_pandn(__m64 _MM1, __m64 _MM2);
_m_pavgb	__m64 _m_pavgb(__m64, __m64);
_m_pavgw	__m64 _m_pavgw(__m64, __m64);
_m_pcmpeqb	__m64 _m_pcmpeqb(__m64 _MM1, __m64 _MM2);
_m_pcmpeqd	__m64 _m_pcmpeqd(__m64 _MM1, __m64 _MM2);
_m_pcmpeqw	__m64 _m_pcmpeqw(__m64 _MM1, __m64 _MM2);
_m_pcmpgtb	__m64 _m_pcmpgtb(__m64 _MM1, __m64 _MM2);
_m_pcmpgtd	__m64 _m_pcmpgtd(__m64 _MM1, __m64 _MM2);
_m_pcmpgtw	__m64 _m_pcmpgtw(__m64 _MM1, __m64 _MM2);
_m_pextrw	int _m_pextrw(__m64, int);
_m_pinsrw	__m64 _m_pinsrw(__m64, int, int);
_m_pmaddwd	__m64 _m_pmaddwd(__m64 _MM1, __m64 _MM2);
_m_pmaxsw	__m64 _m_pmaxsw(__m64, __m64);
_m_pmaxub	__m64 _m_pmaxub(__m64, __m64);
_m_pminsw	__m64 _m_pminsw(__m64, __m64);
_m_pminub	__m64 _m_pminub(__m64, __m64);
_m_pmovmskb	int _m_pmovmskb(__m64);
_m_pmulhuw	__m64 _m_pmulhuw(__m64, __m64);
_m_pmulhw	__m64 _m_pmulhw(__m64 _MM1, __m64 _MM2);
_m_pmullw	__m64 _m_pmullw(__m64 _MM1, __m64 _MM2);
_m_por	__m64 _m_por(__m64 _MM1, __m64 _MM2);
_m_prefetch	void _m_prefetch(void *);
_m_prefetchw	void _m_prefetchw(volatile const void *);
_m_psadbw	__m64 _m_psadbw(__m64, __m64);
_m_pshufw	__m64 _m_pshufw(__m64, int);
_m_pslld	__m64 _m_pslld(__m64 _M, __m64 _Count);
_m_pslldi	__m64 _m_pslldi(__m64 _M, int _Count);
_m_psllq	__m64 _m_psllq(__m64 _M, __m64 _Count);
_m_psllqi	__m64 _m_psllqi(__m64 _M, int _Count);
_m_psllw	__m64 _m_psllw(__m64 _M, __m64 _Count);
_m_psllwi	__m64 _m_psllwi(__m64 _M, int _Count);
_m_psrad	__m64 _m_psrad(__m64 _M, __m64 _Count);
_m_psradi	__m64 _m_psradi(__m64 _M, int _Count);
_m_psraw	__m64 _m_psraw(__m64 _M, __m64 _Count);
_m_psrawi	__m64 _m_psrawi(__m64 _M, int _Count);
_m_psrld	__m64 _m_psrld(__m64 _M, __m64 _Count);
_m_psrldi	__m64 _m_psrldi(__m64 _M, int _Count);
_m_psrlq	__m64 _m_psrlq(__m64 _M, __m64 _Count);
_m_psrlqi	__m64 _m_psrlqi(__m64 _M, int _Count);
_m_psrlw	__m64 _m_psrlw(__m64 _M, __m64 _Count);
_m_psrlwi	__m64 _m_psrlwi(__m64 _M, int _Count);
_m_psubb	__m64 _m_psubb(__m64 _MM1, __m64 _MM2);
_m_psubd	__m64 _m_psubd(__m64 _MM1, __m64 _MM2);
_m_psubsb	__m64 _m_psubsb(__m64 _MM1, __m64 _MM2);
_m_psubsw	__m64 _m_psubsw(__m64 _MM1, __m64 _MM2);
_m_psubusb	__m64 _m_psubusb(__m64 _MM1, __m64 _MM2);
_m_psubusw	__m64 _m_psubusw(__m64 _MM1, __m64 _MM2);
_m_psubw	__m64 _m_psubw(__m64 _MM1, __m64 _MM2);
_m_punpckhbw	__m64 _m_punpckhbw(__m64 _MM1, __m64 _MM2);
_m_punpckhdq	__m64 _m_punpckhdq(__m64 _MM1, __m64 _MM2);
_m_punpckhwd	__m64 _m_punpckhwd(__m64 _MM1, __m64 _MM2);
_m_punpcklbw	__m64 _m_punpcklbw(__m64 _MM1, __m64 _MM2);
_m_punpckldq	__m64 _m_punpckldq(__m64 _MM1, __m64 _MM2);
_m_punpcklwd	__m64 _m_punpcklwd(__m64 _MM1, __m64 _MM2);
_m_pxor	__m64 _m_pxor(__m64 _MM1, __m64 _MM2);
_m_to_int	int _m_to_int(__m64 _M);
_mm_abs_epi16	__m128i _mm_abs_epi16(__m128i a);
_mm_abs_epi16	__m128i _mm_abs_epi16(__m128i);
_mm_abs_epi32	__m128i _mm_abs_epi32(__m128i a);
_mm_abs_epi32	__m128i _mm_abs_epi32(__m128i);
_mm_abs_epi8	__m128i _mm_abs_epi8(__m128i a);
_mm_abs_epi8	__m128i _mm_abs_epi8(__m128i);
_mm_abs_pi16	__m64 _mm_abs_pi16(__m64 a);
_mm_abs_pi16	__m64 _mm_abs_pi16(__m64);
_mm_abs_pi32	__m64 _mm_abs_pi32(__m64 a);
_mm_abs_pi32	__m64 _mm_abs_pi32(__m64);
_mm_abs_pi8	__m64 _mm_abs_pi8(__m64 a);
_mm_abs_pi8	__m64 _mm_abs_pi8(__m64);
_mm_add_epi16	__m128i _mm_add_epi16(__m128i _A, __m128i _B);
_mm_add_epi16	__m128i _mm_add_epi16(__m128i, __m128i);
_mm_add_epi32	__m128i _mm_add_epi32(__m128i _A, __m128i _B);
_mm_add_epi32	__m128i _mm_add_epi32(__m128i, __m128i);
_mm_add_epi64	__m128i _mm_add_epi64(__m128i _A, __m128i _B);
_mm_add_epi64	__m128i _mm_add_epi64(__m128i, __m128i);
_mm_add_epi8	__m128i _mm_add_epi8(__m128i _A, __m128i _B);
_mm_add_epi8	__m128i _mm_add_epi8(__m128i, __m128i);
_mm_add_pd	__m128d _mm_add_pd(__m128d _A, __m128d _B);
_mm_add_pd	__m128d _mm_add_pd(__m128d, __m128d);
_mm_add_ps	__m128 _mm_add_ps(__m128 _A, __m128 _B);
_mm_add_ps	__m128 _mm_add_ps(__m128, __m128);
_mm_add_sd	__m128d _mm_add_sd(__m128d _A, __m128d _B);
_mm_add_sd	__m128d _mm_add_sd(__m128d, __m128d);
_mm_add_si64	__m64 _mm_add_si64(__m64 _A, __m64 _B);
_mm_add_ss	__m128 _mm_add_ss(__m128 _A, __m128 _B);
_mm_add_ss	__m128 _mm_add_ss(__m128, __m128);
_mm_adds_epi16	__m128i _mm_adds_epi16(__m128i _A, __m128i _B);
_mm_adds_epi16	__m128i _mm_adds_epi16(__m128i, __m128i);
_mm_adds_epi8	__m128i _mm_adds_epi8(__m128i _A, __m128i _B);
_mm_adds_epi8	__m128i _mm_adds_epi8(__m128i, __m128i);
_mm_adds_epu16	__m128i _mm_adds_epu16(__m128i _A, __m128i _B);
_mm_adds_epu16	__m128i _mm_adds_epu16(__m128i, __m128i);
_mm_adds_epu8	__m128i _mm_adds_epu8(__m128i _A, __m128i _B);
_mm_adds_epu8	__m128i _mm_adds_epu8(__m128i, __m128i);
_mm_addsub_pd	__m128d _mm_addsub_pd(__m128d a, __m128d b);
_mm_addsub_pd	__m128d _mm_addsub_pd(__m128d, __m128d);
_mm_addsub_ps	__m128 _mm_addsub_ps(__m128 a, __m128 b);
_mm_addsub_ps	__m128 _mm_addsub_ps(__m128, __m128);
_mm_aesdec_si128	__m128i _mm_aesdec_si128(__m128i v, __m128i rkey);
_mm_aesdeclast_si128	__m128i _mm_aesdeclast_si128(__m128i v, __m128i rkey);
_mm_aesenc_si128	__m128i _mm_aesenc_si128(__m128i v, __m128i rkey);
_mm_aesenclast_si128	__m128i _mm_aesenclast_si128(__m128i v, __m128i rkey);
_mm_aesimc_si128	__m128i _mm_aesimc_si128(__m128i v);
_mm_aeskeygenassist_si128	__m128i _mm_aeskeygenassist_si128(__m128i ckey, const int rcon);
_mm_alignr_epi8	__m128i _mm_alignr_epi8(__m128i a, __m128i b, int n);
_mm_alignr_epi8	__m128i _mm_alignr_epi8(__m128i, __m128i, int);
_mm_alignr_pi8	__m64 _mm_alignr_pi8(__m64 a, __m64 b, int n);
_mm_alignr_pi8	__m64 _mm_alignr_pi8(__m64, __m64, int);
_mm_and_pd	__m128d _mm_and_pd(__m128d _A, __m128d _B);
_mm_and_pd	__m128d _mm_and_pd(__m128d, __m128d);
_mm_and_ps	__m128 _mm_and_ps(__m128 _A, __m128 _B);
_mm_and_ps	__m128 _mm_and_ps(__m128, __m128);
_mm_and_si128	__m128i _mm_and_si128(__m128i _A, __m128i _B);
_mm_and_si128	__m128i _mm_and_si128(__m128i, __m128i);
_mm_andnot_pd	__m128d _mm_andnot_pd(__m128d _A, __m128d _B);
_mm_andnot_pd	__m128d _mm_andnot_pd(__m128d, __m128d);
_mm_andnot_ps	__m128 _mm_andnot_ps(__m128 _A, __m128 _B);
_mm_andnot_ps	__m128 _mm_andnot_ps(__m128, __m128);
_mm_andnot_si128	__m128i _mm_andnot_si128(__m128i _A, __m128i _B);
_mm_andnot_si128	__m128i _mm_andnot_si128(__m128i, __m128i);
_mm_avg_epu16	__m128i _mm_avg_epu16(__m128i _A, __m128i _B);
_mm_avg_epu16	__m128i _mm_avg_epu16(__m128i, __m128i);
_mm_avg_epu8	__m128i _mm_avg_epu8(__m128i _A, __m128i _B);
_mm_avg_epu8	__m128i _mm_avg_epu8(__m128i, __m128i);
_mm_blend_epi16	__m128i _mm_blend_epi16(__m128i v1, __m128i v2, const int mask);
_mm_blend_epi16	__m128i _mm_blend_epi16(__m128i, __m128i, int);
_mm_blend_epi32	__m128i __cdecl _mm_blend_epi32(__m128i, __m128i, const int);
_mm_blend_pd	__m128d _mm_blend_pd(__m128d v1, __m128d v2, const int mask);
_mm_blend_pd	__m128d _mm_blend_pd(__m128d, __m128d, int);
_mm_blend_ps	__m128 _mm_blend_ps(__m128 v1, __m128 v2, const int mask);
_mm_blend_ps	__m128 _mm_blend_ps(__m128, __m128, int);
_mm_blendv_epi8	__m128i _mm_blendv_epi8(__m128i v1, __m128i v2, __m128i mask);
_mm_blendv_epi8	__m128i _mm_blendv_epi8(__m128i, __m128i, __m128i);
_mm_blendv_pd	__m128d _mm_blendv_pd(__m128d v1, __m128d v2, __m128d v3);
_mm_blendv_pd	__m128d _mm_blendv_pd(__m128d, __m128d, __m128d);
_mm_blendv_ps	__m128 _mm_blendv_ps(__m128 v1, __m128 v2, __m128 v3);
_mm_blendv_ps	__m128 _mm_blendv_ps(__m128, __m128, __m128);
_mm_broadcast_ss	__m128 __cdecl _mm_broadcast_ss(float const *);
_mm_broadcastb_epi8	__m128i __cdecl _mm_broadcastb_epi8(__m128i);
_mm_broadcastd_epi32	__m128i __cdecl _mm_broadcastd_epi32(__m128i);
_mm_broadcastq_epi64	__m128i __cdecl _mm_broadcastq_epi64(__m128i);
_mm_broadcastsd_pd	__m128d __cdecl _mm_broadcastsd_pd(__m128d);
_mm_broadcastss_ps	__m128 __cdecl _mm_broadcastss_ps(__m128);
_mm_broadcastw_epi16	__m128i __cdecl _mm_broadcastw_epi16(__m128i);
_mm_castpd_ps	__m128 _mm_castpd_ps(__m128d);
_mm_castpd_si128	__m128i _mm_castpd_si128(__m128d);
_mm_castps_pd	__m128d _mm_castps_pd(__m128);
_mm_castps_si128	__m128i _mm_castps_si128(__m128);
_mm_castsi128_pd	__m128d _mm_castsi128_pd(__m128i);
_mm_castsi128_ps	__m128 _mm_castsi128_ps(__m128i);
_mm_clflush	void _mm_clflush(void const *);
_mm_clflush	void _mm_clflush(void const*_P);
_mm_clmulepi64_si128	__m128i _mm_clmulepi64_si128(__m128i v1, __m128i v2, const int imm8);
_mm_cmov_si128	__m128i _mm_cmov_si128(__m128i, __m128i, __m128i);
_mm_cmp_pd	__m128d __cdecl _mm_cmp_pd(__m128d, __m128d, const int);
_mm_cmp_ps	__m128 __cdecl _mm_cmp_ps(__m128, __m128, const int);
_mm_cmp_sd	__m128d __cdecl _mm_cmp_sd(__m128d, __m128d, const int);
_mm_cmp_ss	__m128 __cdecl _mm_cmp_ss(__m128, __m128, const int);
_mm_cmpeq_epi16	__m128i _mm_cmpeq_epi16(__m128i _A, __m128i _B);
_mm_cmpeq_epi16	__m128i _mm_cmpeq_epi16(__m128i, __m128i);
_mm_cmpeq_epi32	__m128i _mm_cmpeq_epi32(__m128i _A, __m128i _B);
_mm_cmpeq_epi32	__m128i _mm_cmpeq_epi32(__m128i, __m128i);
_mm_cmpeq_epi64	__m128i _mm_cmpeq_epi64(__m128i val1, __m128i val2);
_mm_cmpeq_epi64	__m128i _mm_cmpeq_epi64(__m128i, __m128i);
_mm_cmpeq_epi8	__m128i _mm_cmpeq_epi8(__m128i _A, __m128i _B);
_mm_cmpeq_epi8	__m128i _mm_cmpeq_epi8(__m128i, __m128i);
_mm_cmpeq_pd	__m128d _mm_cmpeq_pd(__m128d _A, __m128d _B);
_mm_cmpeq_pd	__m128d _mm_cmpeq_pd(__m128d, __m128d);
_mm_cmpeq_ps	__m128 _mm_cmpeq_ps(__m128 _A, __m128 _B);
_mm_cmpeq_ps	__m128 _mm_cmpeq_ps(__m128, __m128);
_mm_cmpeq_sd	__m128d _mm_cmpeq_sd(__m128d _A, __m128d _B);
_mm_cmpeq_sd	__m128d _mm_cmpeq_sd(__m128d, __m128d);
_mm_cmpeq_ss	__m128 _mm_cmpeq_ss(__m128 _A, __m128 _B);
_mm_cmpeq_ss	__m128 _mm_cmpeq_ss(__m128, __m128);
_mm_cmpestra	int _mm_cmpestra(__m128i a, int la, __m128i b, int lb, const int mode);
_mm_cmpestra	int _mm_cmpestra(__m128i, int, __m128i, int, int);
_mm_cmpestrc	int _mm_cmpestrc(__m128i a, int la, __m128i b, int lb, const int mode);
_mm_cmpestrc	int _mm_cmpestrc(__m128i, int, __m128i, int, int);
_mm_cmpestri	int _mm_cmpestri(__m128i a, int la, __m128i b, int lb, const int mode);
_mm_cmpestri	int _mm_cmpestri(__m128i, int, __m128i, int, int);
_mm_cmpestrm	__m128i _mm_cmpestrm(__m128i a, int la, __m128i b, int lb, const int mode);
_mm_cmpestrm	__m128i _mm_cmpestrm(__m128i, int, __m128i, int, int);
_mm_cmpestro	int _mm_cmpestro(__m128i a, int la, __m128i b, int lb, const int mode);
_mm_cmpestro	int _mm_cmpestro(__m128i, int, __m128i, int, int);
_mm_cmpestrs	int _mm_cmpestrs(__m128i a, int la, __m128i b, int lb, const int mode);
_mm_cmpestrs	int _mm_cmpestrs(__m128i, int, __m128i, int, int);
_mm_cmpestrz	int _mm_cmpestrz(__m128i a, int la, __m128i b, int lb, const int mode);
_mm_cmpestrz	int _mm_cmpestrz(__m128i, int, __m128i, int, int);
_mm_cmpge_pd	__m128d _mm_cmpge_pd(__m128d _A, __m128d _B);
_mm_cmpge_pd	__m128d _mm_cmpge_pd(__m128d, __m128d);
_mm_cmpge_ps	__m128 _mm_cmpge_ps(__m128 _A, __m128 _B);
_mm_cmpge_ps	__m128 _mm_cmpge_ps(__m128, __m128);
_mm_cmpge_sd	__m128d _mm_cmpge_sd(__m128d _A, __m128d _B);
_mm_cmpge_sd	__m128d _mm_cmpge_sd(__m128d, __m128d);
_mm_cmpge_ss	__m128 _mm_cmpge_ss(__m128 _A, __m128 _B);
_mm_cmpge_ss	__m128 _mm_cmpge_ss(__m128, __m128);
_mm_cmpgt_epi16	__m128i _mm_cmpgt_epi16(__m128i _A, __m128i _B);
_mm_cmpgt_epi16	__m128i _mm_cmpgt_epi16(__m128i, __m128i);
_mm_cmpgt_epi32	__m128i _mm_cmpgt_epi32(__m128i _A, __m128i _B);
_mm_cmpgt_epi32	__m128i _mm_cmpgt_epi32(__m128i, __m128i);
_mm_cmpgt_epi64	__m128i _mm_cmpgt_epi64(__m128i val1, __m128i val2);
_mm_cmpgt_epi64	__m128i _mm_cmpgt_epi64(__m128i, __m128i);
_mm_cmpgt_epi8	__m128i _mm_cmpgt_epi8(__m128i _A, __m128i _B);
_mm_cmpgt_epi8	__m128i _mm_cmpgt_epi8(__m128i, __m128i);
_mm_cmpgt_pd	__m128d _mm_cmpgt_pd(__m128d _A, __m128d _B);
_mm_cmpgt_pd	__m128d _mm_cmpgt_pd(__m128d, __m128d);
_mm_cmpgt_ps	__m128 _mm_cmpgt_ps(__m128 _A, __m128 _B);
_mm_cmpgt_ps	__m128 _mm_cmpgt_ps(__m128, __m128);
_mm_cmpgt_sd	__m128d _mm_cmpgt_sd(__m128d _A, __m128d _B);
_mm_cmpgt_sd	__m128d _mm_cmpgt_sd(__m128d, __m128d);
_mm_cmpgt_ss	__m128 _mm_cmpgt_ss(__m128 _A, __m128 _B);
_mm_cmpgt_ss	__m128 _mm_cmpgt_ss(__m128, __m128);
_mm_cmpistra	int _mm_cmpistra(__m128i a, __m128i b, const int mode);
_mm_cmpistra	int _mm_cmpistra(__m128i, __m128i, int);
_mm_cmpistrc	int _mm_cmpistrc(__m128i a, __m128i b, const int mode);
_mm_cmpistrc	int _mm_cmpistrc(__m128i, __m128i, int);
_mm_cmpistri	int _mm_cmpistri(__m128i a, __m128i b, const int mode);
_mm_cmpistri	int _mm_cmpistri(__m128i, __m128i, int);
_mm_cmpistrm	__m128i _mm_cmpistrm(__m128i a, __m128i b, const int mode);
_mm_cmpistrm	__m128i _mm_cmpistrm(__m128i, __m128i, int);
_mm_cmpistro	int _mm_cmpistro(__m128i a, __m128i b, const int mode);
_mm_cmpistro	int _mm_cmpistro(__m128i, __m128i, int);
_mm_cmpistrs	int _mm_cmpistrs(__m128i a, __m128i b, const int mode);
_mm_cmpistrs	int _mm_cmpistrs(__m128i, __m128i, int);
_mm_cmpistrz	int _mm_cmpistrz(__m128i a, __m128i b, const int mode);
_mm_cmpistrz	int _mm_cmpistrz(__m128i, __m128i, int);
_mm_cmple_pd	__m128d _mm_cmple_pd(__m128d _A, __m128d _B);
_mm_cmple_pd	__m128d _mm_cmple_pd(__m128d, __m128d);
_mm_cmple_ps	__m128 _mm_cmple_ps(__m128 _A, __m128 _B);
_mm_cmple_ps	__m128 _mm_cmple_ps(__m128, __m128);
_mm_cmple_sd	__m128d _mm_cmple_sd(__m128d _A, __m128d _B);
_mm_cmple_sd	__m128d _mm_cmple_sd(__m128d, __m128d);
_mm_cmple_ss	__m128 _mm_cmple_ss(__m128 _A, __m128 _B);
_mm_cmple_ss	__m128 _mm_cmple_ss(__m128, __m128);
_mm_cmplt_epi16	__m128i _mm_cmplt_epi16(__m128i _A, __m128i _B);
_mm_cmplt_epi16	__m128i _mm_cmplt_epi16(__m128i, __m128i);
_mm_cmplt_epi32	__m128i _mm_cmplt_epi32(__m128i _A, __m128i _B);
_mm_cmplt_epi32	__m128i _mm_cmplt_epi32(__m128i, __m128i);
_mm_cmplt_epi8	__m128i _mm_cmplt_epi8(__m128i _A, __m128i _B);
_mm_cmplt_epi8	__m128i _mm_cmplt_epi8(__m128i, __m128i);
_mm_cmplt_pd	__m128d _mm_cmplt_pd(__m128d _A, __m128d _B);
_mm_cmplt_pd	__m128d _mm_cmplt_pd(__m128d, __m128d);
_mm_cmplt_ps	__m128 _mm_cmplt_ps(__m128 _A, __m128 _B);
_mm_cmplt_ps	__m128 _mm_cmplt_ps(__m128, __m128);
_mm_cmplt_sd	__m128d _mm_cmplt_sd(__m128d _A, __m128d _B);
_mm_cmplt_sd	__m128d _mm_cmplt_sd(__m128d, __m128d);
_mm_cmplt_ss	__m128 _mm_cmplt_ss(__m128 _A, __m128 _B);
_mm_cmplt_ss	__m128 _mm_cmplt_ss(__m128, __m128);
_mm_cmpneq_pd	__m128d _mm_cmpneq_pd(__m128d _A, __m128d _B);
_mm_cmpneq_pd	__m128d _mm_cmpneq_pd(__m128d, __m128d);
_mm_cmpneq_ps	__m128 _mm_cmpneq_ps(__m128 _A, __m128 _B);
_mm_cmpneq_ps	__m128 _mm_cmpneq_ps(__m128, __m128);
_mm_cmpneq_sd	__m128d _mm_cmpneq_sd(__m128d _A, __m128d _B);
_mm_cmpneq_sd	__m128d _mm_cmpneq_sd(__m128d, __m128d);
_mm_cmpneq_ss	__m128 _mm_cmpneq_ss(__m128 _A, __m128 _B);
_mm_cmpneq_ss	__m128 _mm_cmpneq_ss(__m128, __m128);
_mm_cmpnge_pd	__m128d _mm_cmpnge_pd(__m128d _A, __m128d _B);
_mm_cmpnge_pd	__m128d _mm_cmpnge_pd(__m128d, __m128d);
_mm_cmpnge_ps	__m128 _mm_cmpnge_ps(__m128 _A, __m128 _B);
_mm_cmpnge_ps	__m128 _mm_cmpnge_ps(__m128, __m128);
_mm_cmpnge_sd	__m128d _mm_cmpnge_sd(__m128d _A, __m128d _B);
_mm_cmpnge_sd	__m128d _mm_cmpnge_sd(__m128d, __m128d);
_mm_cmpnge_ss	__m128 _mm_cmpnge_ss(__m128 _A, __m128 _B);
_mm_cmpnge_ss	__m128 _mm_cmpnge_ss(__m128, __m128);
_mm_cmpngt_pd	__m128d _mm_cmpngt_pd(__m128d _A, __m128d _B);
_mm_cmpngt_pd	__m128d _mm_cmpngt_pd(__m128d, __m128d);
_mm_cmpngt_ps	__m128 _mm_cmpngt_ps(__m128 _A, __m128 _B);
_mm_cmpngt_ps	__m128 _mm_cmpngt_ps(__m128, __m128);
_mm_cmpngt_sd	__m128d _mm_cmpngt_sd(__m128d _A, __m128d _B);
_mm_cmpngt_sd	__m128d _mm_cmpngt_sd(__m128d, __m128d);
_mm_cmpngt_ss	__m128 _mm_cmpngt_ss(__m128 _A, __m128 _B);
_mm_cmpngt_ss	__m128 _mm_cmpngt_ss(__m128, __m128);
_mm_cmpnle_pd	__m128d _mm_cmpnle_pd(__m128d _A, __m128d _B);
_mm_cmpnle_pd	__m128d _mm_cmpnle_pd(__m128d, __m128d);
_mm_cmpnle_ps	__m128 _mm_cmpnle_ps(__m128 _A, __m128 _B);
_mm_cmpnle_ps	__m128 _mm_cmpnle_ps(__m128, __m128);
_mm_cmpnle_sd	__m128d _mm_cmpnle_sd(__m128d _A, __m128d _B);
_mm_cmpnle_sd	__m128d _mm_cmpnle_sd(__m128d, __m128d);
_mm_cmpnle_ss	__m128 _mm_cmpnle_ss(__m128 _A, __m128 _B);
_mm_cmpnle_ss	__m128 _mm_cmpnle_ss(__m128, __m128);
_mm_cmpnlt_pd	__m128d _mm_cmpnlt_pd(__m128d _A, __m128d _B);
_mm_cmpnlt_pd	__m128d _mm_cmpnlt_pd(__m128d, __m128d);
_mm_cmpnlt_ps	__m128 _mm_cmpnlt_ps(__m128 _A, __m128 _B);
_mm_cmpnlt_ps	__m128 _mm_cmpnlt_ps(__m128, __m128);
_mm_cmpnlt_sd	__m128d _mm_cmpnlt_sd(__m128d _A, __m128d _B);
_mm_cmpnlt_sd	__m128d _mm_cmpnlt_sd(__m128d, __m128d);
_mm_cmpnlt_ss	__m128 _mm_cmpnlt_ss(__m128 _A, __m128 _B);
_mm_cmpnlt_ss	__m128 _mm_cmpnlt_ss(__m128, __m128);
_mm_cmpord_pd	__m128d _mm_cmpord_pd(__m128d _A, __m128d _B);
_mm_cmpord_pd	__m128d _mm_cmpord_pd(__m128d, __m128d);
_mm_cmpord_ps	__m128 _mm_cmpord_ps(__m128 _A, __m128 _B);
_mm_cmpord_ps	__m128 _mm_cmpord_ps(__m128, __m128);
_mm_cmpord_sd	__m128d _mm_cmpord_sd(__m128d _A, __m128d _B);
_mm_cmpord_sd	__m128d _mm_cmpord_sd(__m128d, __m128d);
_mm_cmpord_ss	__m128 _mm_cmpord_ss(__m128 _A, __m128 _B);
_mm_cmpord_ss	__m128 _mm_cmpord_ss(__m128, __m128);
_mm_cmpunord_pd	__m128d _mm_cmpunord_pd(__m128d _A, __m128d _B);
_mm_cmpunord_pd	__m128d _mm_cmpunord_pd(__m128d, __m128d);
_mm_cmpunord_ps	__m128 _mm_cmpunord_ps(__m128 _A, __m128 _B);
_mm_cmpunord_ps	__m128 _mm_cmpunord_ps(__m128, __m128);
_mm_cmpunord_sd	__m128d _mm_cmpunord_sd(__m128d _A, __m128d _B);
_mm_cmpunord_sd	__m128d _mm_cmpunord_sd(__m128d, __m128d);
_mm_cmpunord_ss	__m128 _mm_cmpunord_ss(__m128 _A, __m128 _B);
_mm_cmpunord_ss	__m128 _mm_cmpunord_ss(__m128, __m128);
_mm_com_epi16	__m128i _mm_com_epi16(__m128i, __m128i, int);
_mm_com_epi32	__m128i _mm_com_epi32(__m128i, __m128i, int);
_mm_com_epi64	__m128i _mm_com_epi64(__m128i, __m128i, int);
_mm_com_epi8	__m128i _mm_com_epi8(__m128i, __m128i, int);
_mm_com_epu16	__m128i _mm_com_epu16(__m128i, __m128i, int);
_mm_com_epu32	__m128i _mm_com_epu32(__m128i, __m128i, int);
_mm_com_epu64	__m128i _mm_com_epu64(__m128i, __m128i, int);
_mm_com_epu8	__m128i _mm_com_epu8(__m128i, __m128i, int);
_mm_comieq_sd	int _mm_comieq_sd(__m128d _A, __m128d _B);
_mm_comieq_sd	int _mm_comieq_sd(__m128d, __m128d);
_mm_comieq_ss	int _mm_comieq_ss(__m128 _A, __m128 _B);
_mm_comieq_ss	int _mm_comieq_ss(__m128, __m128);
_mm_comige_sd	int _mm_comige_sd(__m128d _A, __m128d _B);
_mm_comige_sd	int _mm_comige_sd(__m128d, __m128d);
_mm_comige_ss	int _mm_comige_ss(__m128 _A, __m128 _B);
_mm_comige_ss	int _mm_comige_ss(__m128, __m128);
_mm_comigt_sd	int _mm_comigt_sd(__m128d _A, __m128d _B);
_mm_comigt_sd	int _mm_comigt_sd(__m128d, __m128d);
_mm_comigt_ss	int _mm_comigt_ss(__m128 _A, __m128 _B);
_mm_comigt_ss	int _mm_comigt_ss(__m128, __m128);
_mm_comile_sd	int _mm_comile_sd(__m128d _A, __m128d _B);
_mm_comile_sd	int _mm_comile_sd(__m128d, __m128d);
_mm_comile_ss	int _mm_comile_ss(__m128 _A, __m128 _B);
_mm_comile_ss	int _mm_comile_ss(__m128, __m128);
_mm_comilt_sd	int _mm_comilt_sd(__m128d _A, __m128d _B);
_mm_comilt_sd	int _mm_comilt_sd(__m128d, __m128d);
_mm_comilt_ss	int _mm_comilt_ss(__m128 _A, __m128 _B);
_mm_comilt_ss	int _mm_comilt_ss(__m128, __m128);
_mm_comineq_sd	int _mm_comineq_sd(__m128d _A, __m128d _B);
_mm_comineq_sd	int _mm_comineq_sd(__m128d, __m128d);
_mm_comineq_ss	int _mm_comineq_ss(__m128 _A, __m128 _B);
_mm_comineq_ss	int _mm_comineq_ss(__m128, __m128);
_mm_crc32_u16	unsigned int _mm_crc32_u16(unsigned int crc, unsigned short v);
_mm_crc32_u16	unsigned int _mm_crc32_u16(unsigned int, unsigned short);
_mm_crc32_u32	unsigned int _mm_crc32_u32(unsigned int crc, unsigned int v);
_mm_crc32_u32	unsigned int _mm_crc32_u32(unsigned int, unsigned int);
_mm_crc32_u64	unsigned __int64 _mm_crc32_u64(unsigned __int64 crc, unsigned __int64 v);
_mm_crc32_u64	unsigned __int64 _mm_crc32_u64(unsigned __int64, unsigned __int64);
_mm_crc32_u8	unsigned int _mm_crc32_u8(unsigned int crc, unsigned char v);
_mm_crc32_u8	unsigned int _mm_crc32_u8(unsigned int, unsigned char);
_mm_cvt_pi2ps	__m128 _mm_cvt_pi2ps(__m128, __m64);
_mm_cvt_ps2pi	__m64 _mm_cvt_ps2pi(__m128 _A);
_mm_cvt_si2ss	__m128 _mm_cvt_si2ss(__m128, int);
_mm_cvt_si2ss	__m128 _mm_cvt_si2ss(__m128, int);
_mm_cvt_ss2si	int _mm_cvt_ss2si(__m128 _A);
_mm_cvt_ss2si	int _mm_cvt_ss2si(__m128);
_mm_cvtepi16_epi32	__m128i _mm_cvtepi16_epi32(__m128i shortValues);
_mm_cvtepi16_epi32	__m128i _mm_cvtepi16_epi32(__m128i);
_mm_cvtepi16_epi64	__m128i _mm_cvtepi16_epi64(__m128i shortValues);
_mm_cvtepi16_epi64	__m128i _mm_cvtepi16_epi64(__m128i);
_mm_cvtepi32_epi64	__m128i _mm_cvtepi32_epi64(__m128i intValues);
_mm_cvtepi32_epi64	__m128i _mm_cvtepi32_epi64(__m128i);
_mm_cvtepi32_pd	__m128d _mm_cvtepi32_pd(__m128i _A);
_mm_cvtepi32_pd	__m128d _mm_cvtepi32_pd(__m128i);
_mm_cvtepi32_ps	__m128 _mm_cvtepi32_ps(__m128i _A);
_mm_cvtepi32_ps	__m128 _mm_cvtepi32_ps(__m128i);
_mm_cvtepi8_epi16	__m128i _mm_cvtepi8_epi16(__m128i byteValues);
_mm_cvtepi8_epi16	__m128i _mm_cvtepi8_epi16(__m128i);
_mm_cvtepi8_epi32	__m128i _mm_cvtepi8_epi32(__m128i byteValues);
_mm_cvtepi8_epi32	__m128i _mm_cvtepi8_epi32(__m128i);
_mm_cvtepi8_epi64	__m128i _mm_cvtepi8_epi64(__m128i byteValues);
_mm_cvtepi8_epi64	__m128i _mm_cvtepi8_epi64(__m128i);
_mm_cvtepu16_epi32	__m128i _mm_cvtepu16_epi32(__m128i shortValues);
_mm_cvtepu16_epi32	__m128i _mm_cvtepu16_epi32(__m128i);
_mm_cvtepu16_epi64	__m128i _mm_cvtepu16_epi64(__m128i shortValues);
_mm_cvtepu16_epi64	__m128i _mm_cvtepu16_epi64(__m128i);
_mm_cvtepu32_epi64	__m128i _mm_cvtepu32_epi64(__m128i intValues);
_mm_cvtepu32_epi64	__m128i _mm_cvtepu32_epi64(__m128i);
_mm_cvtepu8_epi16	__m128i _mm_cvtepu8_epi16(__m128i byteValues);
_mm_cvtepu8_epi16	__m128i _mm_cvtepu8_epi16(__m128i);
_mm_cvtepu8_epi32	__m128i _mm_cvtepu8_epi32(__m128i byteValues);
_mm_cvtepu8_epi32	__m128i _mm_cvtepu8_epi32(__m128i);
_mm_cvtepu8_epi64	__m128i _mm_cvtepu8_epi64(__m128i shortValues);
_mm_cvtepu8_epi64	__m128i _mm_cvtepu8_epi64(__m128i);
_mm_cvtpd_epi32	__m128i _mm_cvtpd_epi32(__m128d _A);
_mm_cvtpd_epi32	__m128i _mm_cvtpd_epi32(__m128d);
_mm_cvtpd_pi32	__m64 _mm_cvtpd_pi32(__m128d _A);
_mm_cvtpd_ps	__m128 _mm_cvtpd_ps(__m128d _A);
_mm_cvtpd_ps	__m128 _mm_cvtpd_ps(__m128d);
_mm_cvtph_ps	__m128 __cdecl _mm_cvtph_ps(__m128i);
_mm_cvtpi16_ps	__m128 _mm_cvtpi16_ps(__m64 a);
_mm_cvtpi32_pd	__m128d _mm_cvtpi32_pd(__m64 _A);
_mm_cvtpi32x2_ps	__m128 _mm_cvtpi32x2_ps(__m64 a, __m64 b);
_mm_cvtpi8_ps	__m128 _mm_cvtpi8_ps(__m64 a);
_mm_cvtps_epi32	__m128i _mm_cvtps_epi32(__m128 _A);
_mm_cvtps_epi32	__m128i _mm_cvtps_epi32(__m128);
_mm_cvtps_pd	__m128d _mm_cvtps_pd(__m128 _A);
_mm_cvtps_pd	__m128d _mm_cvtps_pd(__m128);
_mm_cvtps_ph	__m128i __cdecl _mm_cvtps_ph(__m128 m1, const int imm);
_mm_cvtps_pi16	__m64 _mm_cvtps_pi16(__m128 a);
_mm_cvtps_pi8	__m64 _mm_cvtps_pi8(__m128 a);
_mm_cvtpu16_ps	__m128 _mm_cvtpu16_ps(__m64 a);
_mm_cvtpu8_ps	__m128 _mm_cvtpu8_ps(__m64 a);
_mm_cvtsd_f64	double _mm_cvtsd_f64(__m128d _A);
_mm_cvtsd_si32	int _mm_cvtsd_si32(__m128d _A);
_mm_cvtsd_si32	int _mm_cvtsd_si32(__m128d);
_mm_cvtsd_si64	__int64 _mm_cvtsd_si64(__m128d);
_mm_cvtsd_si64	__int64 _mm_cvtsd_si64(__m128d);
_mm_cvtsd_si64x	__int64 _mm_cvtsd_si64x(__m128d);
_mm_cvtsd_ss	__m128 _mm_cvtsd_ss(__m128 _A, __m128d _B);
_mm_cvtsd_ss	__m128 _mm_cvtsd_ss(__m128, __m128d);
_mm_cvtsi128_si32	int _mm_cvtsi128_si32(__m128i _A);
_mm_cvtsi128_si32	int _mm_cvtsi128_si32(__m128i);
_mm_cvtsi128_si64	__int64 _mm_cvtsi128_si64(__m128i);
_mm_cvtsi128_si64	__int64 _mm_cvtsi128_si64(__m128i);
_mm_cvtsi128_si64x	__int64 _mm_cvtsi128_si64x(__m128i);
_mm_cvtsi32_sd	__m128d _mm_cvtsi32_sd(__m128d _A, int _B);
_mm_cvtsi32_sd	__m128d _mm_cvtsi32_sd(__m128d, int);
_mm_cvtsi32_si128	__m128i _mm_cvtsi32_si128(int _A);
_mm_cvtsi32_si128	__m128i _mm_cvtsi32_si128(int);
_mm_cvtsi64_sd	__m128d _mm_cvtsi64_sd(__m128d, __int64);
_mm_cvtsi64_sd	__m128d _mm_cvtsi64_sd(__m128d, __int64);
_mm_cvtsi64_si128	__m128i _mm_cvtsi64_si128(__int64);
_mm_cvtsi64_si128	__m128i _mm_cvtsi64_si128(__int64);
_mm_cvtsi64_ss	__m128 _mm_cvtsi64_ss(__m128 _A, __int64 _B);
_mm_cvtsi64_ss	__m128 _mm_cvtsi64_ss(__m128, __int64);
_mm_cvtsi64x_sd	__m128d _mm_cvtsi64x_sd(__m128d, __int64);
_mm_cvtsi64x_si128	__m128i _mm_cvtsi64x_si128(__int64);
_mm_cvtsi64x_ss	__m128 _mm_cvtsi64x_ss(__m128, __int64);
_mm_cvtss_f32	float _mm_cvtss_f32(__m128 _A);
_mm_cvtss_sd	__m128d _mm_cvtss_sd(__m128d _A, __m128 _B);
_mm_cvtss_sd	__m128d _mm_cvtss_sd(__m128d, __m128);
_mm_cvtss_si64	__int64 _mm_cvtss_si64(__m128 _A);
_mm_cvtss_si64	__int64 _mm_cvtss_si64(__m128);
_mm_cvtss_si64x	__int64 _mm_cvtss_si64x(__m128);
_mm_cvtt_ps2pi	__m64 _mm_cvtt_ps2pi(__m128 _A);
_mm_cvtt_ss2si	int _mm_cvtt_ss2si(__m128 _A);
_mm_cvtt_ss2si	int _mm_cvtt_ss2si(__m128);
_mm_cvttpd_epi32	__m128i _mm_cvttpd_epi32(__m128d _A);
_mm_cvttpd_epi32	__m128i _mm_cvttpd_epi32(__m128d);
_mm_cvttpd_pi32	__m64 _mm_cvttpd_pi32(__m128d _A);
_mm_cvttps_epi32	__m128i _mm_cvttps_epi32(__m128 _A);
_mm_cvttps_epi32	__m128i _mm_cvttps_epi32(__m128);
_mm_cvttsd_si32	int _mm_cvttsd_si32(__m128d _A);
_mm_cvttsd_si32	int _mm_cvttsd_si32(__m128d);
_mm_cvttsd_si64	__int64 _mm_cvttsd_si64(__m128d);
_mm_cvttsd_si64	__int64 _mm_cvttsd_si64(__m128d);
_mm_cvttsd_si64x	__int64 _mm_cvttsd_si64x(__m128d);
_mm_cvttss_si64	__int64 _mm_cvttss_si64(__m128 _A);
_mm_cvttss_si64	__int64 _mm_cvttss_si64(__m128);
_mm_cvttss_si64x	__int64 _mm_cvttss_si64x(__m128);
_mm_div_pd	__m128d _mm_div_pd(__m128d _A, __m128d _B);
_mm_div_pd	__m128d _mm_div_pd(__m128d, __m128d);
_mm_div_ps	__m128 _mm_div_ps(__m128 _A, __m128 _B);
_mm_div_ps	__m128 _mm_div_ps(__m128, __m128);
_mm_div_sd	__m128d _mm_div_sd(__m128d _A, __m128d _B);
_mm_div_sd	__m128d _mm_div_sd(__m128d, __m128d);
_mm_div_ss	__m128 _mm_div_ss(__m128 _A, __m128 _B);
_mm_div_ss	__m128 _mm_div_ss(__m128, __m128);
_mm_dp_pd	__m128d _mm_dp_pd(__m128d val1, __m128d val2, const int mask);
_mm_dp_pd	__m128d _mm_dp_pd(__m128d, __m128d, int);
_mm_dp_ps	__m128 _mm_dp_ps(__m128 val1, __m128 val2, const int mask);
_mm_dp_ps	__m128 _mm_dp_ps(__m128, __m128, int);
_mm_extract_epi16	int _mm_extract_epi16(__m128i _A, int _Imm);
_mm_extract_epi16	int _mm_extract_epi16(__m128i, int);
_mm_extract_epi32	int _mm_extract_epi32(__m128i src, const int ndx);
_mm_extract_epi32	int _mm_extract_epi32(__m128i, int);
_mm_extract_epi64	__int64 _mm_extract_epi64(__m128i src, const int ndx);
_mm_extract_epi64	__int64 _mm_extract_epi64(__m128i, int);
_mm_extract_epi8	int _mm_extract_epi8(__m128i src, const int ndx);
_mm_extract_epi8	int _mm_extract_epi8(__m128i, int);
_mm_extract_ps	int _mm_extract_ps(__m128 src, const int ndx);
_mm_extract_ps	int _mm_extract_ps(__m128, int);
_mm_extract_si64	__m128i _mm_extract_si64(__m128i, __m128i);
_mm_extracti_si64	__m128i _mm_extracti_si64(__m128i, int, int);
_mm_fmadd_pd	__m128d __cdecl _mm_fmadd_pd(__m128d, __m128d, __m128d);
_mm_fmadd_ps	__m128 __cdecl _mm_fmadd_ps(__m128, __m128, __m128);
_mm_fmadd_sd	__m128d __cdecl _mm_fmadd_sd(__m128d, __m128d, __m128d);
_mm_fmadd_ss	__m128 __cdecl _mm_fmadd_ss(__m128, __m128, __m128);
_mm_fmaddsub_pd	__m128d __cdecl _mm_fmaddsub_pd(__m128d, __m128d, __m128d);
_mm_fmaddsub_ps	__m128 __cdecl _mm_fmaddsub_ps(__m128, __m128, __m128);
_mm_fmsub_pd	__m128d __cdecl _mm_fmsub_pd(__m128d, __m128d, __m128d);
_mm_fmsub_ps	__m128 __cdecl _mm_fmsub_ps(__m128, __m128, __m128);
_mm_fmsub_sd	__m128d __cdecl _mm_fmsub_sd(__m128d, __m128d, __m128d);
_mm_fmsub_ss	__m128 __cdecl _mm_fmsub_ss(__m128, __m128, __m128);
_mm_fmsubadd_pd	__m128d __cdecl _mm_fmsubadd_pd(__m128d, __m128d, __m128d);
_mm_fmsubadd_ps	__m128 __cdecl _mm_fmsubadd_ps(__m128, __m128, __m128);
_mm_fnmadd_pd	__m128d __cdecl _mm_fnmadd_pd(__m128d, __m128d, __m128d);
_mm_fnmadd_ps	__m128 __cdecl _mm_fnmadd_ps(__m128, __m128, __m128);
_mm_fnmadd_sd	__m128d __cdecl _mm_fnmadd_sd(__m128d, __m128d, __m128d);
_mm_fnmadd_ss	__m128 __cdecl _mm_fnmadd_ss(__m128, __m128, __m128);
_mm_fnmsub_pd	__m128d __cdecl _mm_fnmsub_pd(__m128d, __m128d, __m128d);
_mm_fnmsub_ps	__m128 __cdecl _mm_fnmsub_ps(__m128, __m128, __m128);
_mm_fnmsub_sd	__m128d __cdecl _mm_fnmsub_sd(__m128d, __m128d, __m128d);
_mm_fnmsub_ss	__m128 __cdecl _mm_fnmsub_ss(__m128, __m128, __m128);
_mm_frcz_pd	__m128d _mm_frcz_pd(__m128d);
_mm_frcz_ps	__m128 _mm_frcz_ps(__m128);
_mm_frcz_sd	__m128d _mm_frcz_sd(__m128d, __m128d);
_mm_frcz_ss	__m128 _mm_frcz_ss(__m128, __m128);
_mm_getcsr	unsigned int _mm_getcsr(void);
_mm_getcsr	unsigned int _mm_getcsr(void);
_mm_hadd_epi16	__m128i _mm_hadd_epi16(__m128i a, __m128i b);
_mm_hadd_epi16	__m128i _mm_hadd_epi16(__m128i, __m128i);
_mm_hadd_epi32	__m128i _mm_hadd_epi32(__m128i a, __m128i b);
_mm_hadd_epi32	__m128i _mm_hadd_epi32(__m128i, __m128i);
_mm_hadd_pd	__m128d _mm_hadd_pd(__m128d a, __m128d b);
_mm_hadd_pd	__m128d _mm_hadd_pd(__m128d, __m128d);
_mm_hadd_pi16	__m64 _mm_hadd_pi16(__m64 a, __m64 b);
_mm_hadd_pi16	__m64 _mm_hadd_pi16(__m64, __m64);
_mm_hadd_pi32	__m64 _mm_hadd_pi32(__m64 a, __m64 b);
_mm_hadd_pi32	__m64 _mm_hadd_pi32(__m64, __m64);
_mm_hadd_ps	__m128 _mm_hadd_ps(__m128 a, __m128 b);
_mm_hadd_ps	__m128 _mm_hadd_ps(__m128, __m128);
_mm_haddd_epi16	__m128i _mm_haddd_epi16(__m128i);
_mm_haddd_epi8	__m128i _mm_haddd_epi8(__m128i);
_mm_haddd_epu16	__m128i _mm_haddd_epu16(__m128i);
_mm_haddd_epu8	__m128i _mm_haddd_epu8(__m128i);
_mm_haddq_epi16	__m128i _mm_haddq_epi16(__m128i);
_mm_haddq_epi32	__m128i _mm_haddq_epi32(__m128i);
_mm_haddq_epi8	__m128i _mm_haddq_epi8(__m128i);
_mm_haddq_epu16	__m128i _mm_haddq_epu16(__m128i);
_mm_haddq_epu32	__m128i _mm_haddq_epu32(__m128i);
_mm_haddq_epu8	__m128i _mm_haddq_epu8(__m128i);
_mm_hadds_epi16	__m128i _mm_hadds_epi16(__m128i a, __m128i b);
_mm_hadds_epi16	__m128i _mm_hadds_epi16(__m128i, __m128i);
_mm_hadds_pi16	__m64 _mm_hadds_pi16(__m64 a, __m64 b);
_mm_hadds_pi16	__m64 _mm_hadds_pi16(__m64, __m64);
_mm_haddw_epi8	__m128i _mm_haddw_epi8(__m128i);
_mm_haddw_epu8	__m128i _mm_haddw_epu8(__m128i);
_mm_hsub_epi16	__m128i _mm_hsub_epi16(__m128i a, __m128i b);
_mm_hsub_epi16	__m128i _mm_hsub_epi16(__m128i, __m128i);
_mm_hsub_epi32	__m128i _mm_hsub_epi32(__m128i a, __m128i b);
_mm_hsub_epi32	__m128i _mm_hsub_epi32(__m128i, __m128i);
_mm_hsub_pd	__m128d _mm_hsub_pd(__m128d a, __m128d b);
_mm_hsub_pd	__m128d _mm_hsub_pd(__m128d, __m128d);
_mm_hsub_pi16	__m64 _mm_hsub_pi16(__m64 a, __m64 b);
_mm_hsub_pi16	__m64 _mm_hsub_pi16(__m64, __m64);
_mm_hsub_pi32	__m64 _mm_hsub_pi32(__m64 a, __m64 b);
_mm_hsub_pi32	__m64 _mm_hsub_pi32(__m64, __m64);
_mm_hsub_ps	__m128 _mm_hsub_ps(__m128 a, __m128 b);
_mm_hsub_ps	__m128 _mm_hsub_ps(__m128, __m128);
_mm_hsubd_epi16	__m128i _mm_hsubd_epi16(__m128i);
_mm_hsubq_epi32	__m128i _mm_hsubq_epi32(__m128i);
_mm_hsubs_epi16	__m128i _mm_hsubs_epi16(__m128i a, __m128i b);
_mm_hsubs_epi16	__m128i _mm_hsubs_epi16(__m128i, __m128i);
_mm_hsubs_pi16	__m64 _mm_hsubs_pi16(__m64 a, __m64 b);
_mm_hsubs_pi16	__m64 _mm_hsubs_pi16(__m64, __m64);
_mm_hsubw_epi8	__m128i _mm_hsubw_epi8(__m128i);
_mm_i32gather_epi32	__m128i __cdecl _mm_i32gather_epi32(int const *, __m128i, const int );
_mm_i32gather_epi64	__m128i __cdecl _mm_i32gather_epi64(__int64 const *, __m128i, const int );
_mm_i32gather_pd	__m128d __cdecl _mm_i32gather_pd(double const *, __m128i, const int );
_mm_i32gather_ps	__m128 __cdecl _mm_i32gather_ps(float const *, __m128i, const int );
_mm_i64gather_epi32	__m128i __cdecl _mm_i64gather_epi32(int const *, __m128i, const int );
_mm_i64gather_epi64	__m128i __cdecl _mm_i64gather_epi64(__int64 const *, __m128i, const int );
_mm_i64gather_pd	__m128d __cdecl _mm_i64gather_pd(double const *, __m128i, const int );
_mm_i64gather_ps	__m128 __cdecl _mm_i64gather_ps(float const *, __m128i, const int );
_mm_insert_epi16	__m128i _mm_insert_epi16(__m128i _A, int _B, int _Imm);
_mm_insert_epi16	__m128i _mm_insert_epi16(__m128i, int, int);
_mm_insert_epi32	__m128i _mm_insert_epi32(__m128i dst, int s, const int ndx);
_mm_insert_epi32	__m128i _mm_insert_epi32(__m128i, int, int);
_mm_insert_epi64	__m128i _mm_insert_epi64(__m128i dst, __int64 s, const int ndx);
_mm_insert_epi64	__m128i _mm_insert_epi64(__m128i, __int64, int);
_mm_insert_epi8	__m128i _mm_insert_epi8(__m128i dst, int s, const int ndx);
_mm_insert_epi8	__m128i _mm_insert_epi8(__m128i, int, int);
_mm_insert_ps	__m128 _mm_insert_ps(__m128 dst, __m128 src, const int ndx);
_mm_insert_ps	__m128 _mm_insert_ps(__m128, __m128, int);
_mm_insert_si64	__m128i _mm_insert_si64(__m128i, __m128i);
_mm_inserti_si64	__m128i _mm_inserti_si64(__m128i, __m128i, int, int);
_mm_lddqu_si128	__m128i _mm_lddqu_si128(__m128i const *);
_mm_lddqu_si128	__m128i _mm_lddqu_si128(__m128i const *p);
_mm_lfence	void _mm_lfence(void);
_mm_lfence	void _mm_lfence(void);
_mm_load_pd	__m128d _mm_load_pd(double const *);
_mm_load_pd	__m128d _mm_load_pd(double const*_Dp);
_mm_load_ps	__m128 _mm_load_ps(float const *);
_mm_load_ps	__m128 _mm_load_ps(float const*_A);
_mm_load_ps1	__m128 _mm_load_ps1(float const *);
_mm_load_ps1	__m128 _mm_load_ps1(float const*_A);
_mm_load_sd	__m128d _mm_load_sd(double const *);
_mm_load_sd	__m128d _mm_load_sd(double const*_Dp);
_mm_load_si128	__m128i _mm_load_si128(__m128i const *);
_mm_load_si128	__m128i _mm_load_si128(__m128i const*_P);
_mm_load_ss	__m128 _mm_load_ss(float const *);
_mm_load_ss	__m128 _mm_load_ss(float const*_A);
_mm_load1_pd	__m128d _mm_load1_pd(double const *);
_mm_load1_pd	__m128d _mm_load1_pd(double const*_Dp);
_mm_loaddup_pd	__m128d _mm_loaddup_pd(double const * dp);
_mm_loaddup_pd	__m128d _mm_loaddup_pd(double const *);
_mm_loadh_pd	__m128d _mm_loadh_pd(__m128d _A, double const*_Dp);
_mm_loadh_pd	__m128d _mm_loadh_pd(__m128d, double const *);
_mm_loadh_pi	__m128 _mm_loadh_pi(__m128, __m64 const *);
_mm_loadh_pi	__m128 _mm_loadh_pi(__m128, __m64 const*);
_mm_loadl_epi64	__m128i _mm_loadl_epi64(__m128i const *);
_mm_loadl_epi64	__m128i _mm_loadl_epi64(__m128i const*_P);
_mm_loadl_pd	__m128d _mm_loadl_pd(__m128d _A, double const*_Dp);
_mm_loadl_pd	__m128d _mm_loadl_pd(__m128d, double const *);
_mm_loadl_pi	__m128 _mm_loadl_pi(__m128, __m64 const *);
_mm_loadl_pi	__m128 _mm_loadl_pi(__m128, __m64 const*);
_mm_loadr_pd	__m128d _mm_loadr_pd(double const *);
_mm_loadr_pd	__m128d _mm_loadr_pd(double const*_Dp);
_mm_loadr_ps	__m128 _mm_loadr_ps(float const *);
_mm_loadr_ps	__m128 _mm_loadr_ps(float const*_A);
_mm_loadu_pd	__m128d _mm_loadu_pd(double const *);
_mm_loadu_pd	__m128d _mm_loadu_pd(double const*_Dp);
_mm_loadu_ps	__m128 _mm_loadu_ps(float const *);
_mm_loadu_ps	__m128 _mm_loadu_ps(float const*_A);
_mm_loadu_si128	__m128i _mm_loadu_si128(__m128i const *);
_mm_loadu_si128	__m128i _mm_loadu_si128(__m128i const*_P);
_mm_macc_epi16	__m128i _mm_macc_epi16(__m128i, __m128i, __m128i);
_mm_macc_epi32	__m128i _mm_macc_epi32(__m128i, __m128i, __m128i);
_mm_macc_pd	__m128d _mm_macc_pd(__m128d, __m128d, __m128d);
_mm_macc_ps	__m128 _mm_macc_ps(__m128, __m128, __m128);
_mm_macc_sd	__m128d _mm_macc_sd(__m128d, __m128d, __m128d);
_mm_macc_ss	__m128 _mm_macc_ss(__m128, __m128, __m128);
_mm_maccd_epi16	__m128i _mm_maccd_epi16(__m128i, __m128i, __m128i);
_mm_macchi_epi32	__m128i _mm_macchi_epi32(__m128i, __m128i, __m128i);
_mm_macclo_epi32	__m128i _mm_macclo_epi32(__m128i, __m128i, __m128i);
_mm_maccs_epi16	__m128i _mm_maccs_epi16(__m128i, __m128i, __m128i);
_mm_maccs_epi32	__m128i _mm_maccs_epi32(__m128i, __m128i, __m128i);
_mm_maccsd_epi16	__m128i _mm_maccsd_epi16(__m128i, __m128i, __m128i);
_mm_maccshi_epi32	__m128i _mm_maccshi_epi32(__m128i, __m128i, __m128i);
_mm_maccslo_epi32	__m128i _mm_maccslo_epi32(__m128i, __m128i, __m128i);
_mm_madd_epi16	__m128i _mm_madd_epi16(__m128i _A, __m128i _B);
_mm_madd_epi16	__m128i _mm_madd_epi16(__m128i, __m128i);
_mm_maddd_epi16	__m128i _mm_maddd_epi16(__m128i, __m128i, __m128i);
_mm_maddsd_epi16	__m128i _mm_maddsd_epi16(__m128i, __m128i, __m128i);
_mm_maddsub_pd	__m128d _mm_maddsub_pd(__m128d, __m128d, __m128d);
_mm_maddsub_ps	__m128 _mm_maddsub_ps(__m128, __m128, __m128);
_mm_maddubs_epi16	__m128i _mm_maddubs_epi16(__m128i a, __m128i b);
_mm_maddubs_epi16	__m128i _mm_maddubs_epi16(__m128i, __m128i);
_mm_maddubs_pi16	__m64 _mm_maddubs_pi16(__m64 a, __m64 b);
_mm_maddubs_pi16	__m64 _mm_maddubs_pi16(__m64, __m64);
_mm_mask_i32gather_epi32	__m128i __cdecl _mm_mask_i32gather_epi32(__m128i, int const *, __m128i, __m128i, const int );
_mm_mask_i32gather_epi64	__m128i __cdecl _mm_mask_i32gather_epi64(__m128i, __int64 const *, __m128i, __m128i, const int );
_mm_mask_i32gather_pd	__m128d __cdecl _mm_mask_i32gather_pd(__m128d, double const *, __m128i, __m128d, const int );
_mm_mask_i32gather_ps	__m128 __cdecl _mm_mask_i32gather_ps(__m128, float const *, __m128i, __m128, const int );
_mm_mask_i64gather_epi32	__m128i __cdecl _mm_mask_i64gather_epi32(__m128i, int const *, __m128i, __m128i, const int );
_mm_mask_i64gather_epi64	__m128i __cdecl _mm_mask_i64gather_epi64(__m128i, __int64 const *, __m128i, __m128i, const int );
_mm_mask_i64gather_pd	__m128d __cdecl _mm_mask_i64gather_pd(__m128d, double const *, __m128i, __m128d, const int );
_mm_mask_i64gather_ps	__m128 __cdecl _mm_mask_i64gather_ps(__m128, float const *, __m128i, __m128, const int );
_mm_maskload_epi32	__m128i __cdecl _mm_maskload_epi32(int const *, __m128i );
_mm_maskload_epi64	__m128i __cdecl _mm_maskload_epi64(__int64 const *, __m128i );
_mm_maskload_pd	__m128d __cdecl _mm_maskload_pd(double const *, __m128i);
_mm_maskload_ps	__m128 __cdecl _mm_maskload_ps(float const *, __m128i);
_mm_maskmoveu_si128	void _mm_maskmoveu_si128(__m128i _D, __m128i _N, char *_P);
_mm_maskmoveu_si128	void _mm_maskmoveu_si128(__m128i, __m128i, char *);
_mm_maskstore_epi32	void __cdecl _mm_maskstore_epi32(int *, __m128i, __m128i );
_mm_maskstore_epi64	void __cdecl _mm_maskstore_epi64(__int64 *, __m128i, __m128i );
_mm_maskstore_pd	void __cdecl _mm_maskstore_pd(double *, __m128i, __m128d);
_mm_maskstore_ps	void __cdecl _mm_maskstore_ps(float *, __m128i, __m128);
_mm_max_epi16	__m128i _mm_max_epi16(__m128i _A, __m128i _B);
_mm_max_epi16	__m128i _mm_max_epi16(__m128i, __m128i);
_mm_max_epi32	__m128i _mm_max_epi32(__m128i val1, __m128i val2);
_mm_max_epi32	__m128i _mm_max_epi32(__m128i, __m128i);
_mm_max_epi8	__m128i _mm_max_epi8(__m128i val1, __m128i val2);
_mm_max_epi8	__m128i _mm_max_epi8(__m128i, __m128i);
_mm_max_epu16	__m128i _mm_max_epu16(__m128i val1, __m128i val2);
_mm_max_epu16	__m128i _mm_max_epu16(__m128i, __m128i);
_mm_max_epu32	__m128i _mm_max_epu32(__m128i val1, __m128i val2);
_mm_max_epu32	__m128i _mm_max_epu32(__m128i, __m128i);
_mm_max_epu8	__m128i _mm_max_epu8(__m128i _A, __m128i _B);
_mm_max_epu8	__m128i _mm_max_epu8(__m128i, __m128i);
_mm_max_pd	__m128d _mm_max_pd(__m128d _A, __m128d _B);
_mm_max_pd	__m128d _mm_max_pd(__m128d, __m128d);
_mm_max_ps	__m128 _mm_max_ps(__m128 _A, __m128 _B);
_mm_max_ps	__m128 _mm_max_ps(__m128, __m128);
_mm_max_sd	__m128d _mm_max_sd(__m128d _A, __m128d _B);
_mm_max_sd	__m128d _mm_max_sd(__m128d, __m128d);
_mm_max_ss	__m128 _mm_max_ss(__m128 _A, __m128 _B);
_mm_max_ss	__m128 _mm_max_ss(__m128, __m128);
_mm_mfence	void _mm_mfence(void);
_mm_mfence	void _mm_mfence(void);
_mm_min_epi16	__m128i _mm_min_epi16(__m128i _A, __m128i _B);
_mm_min_epi16	__m128i _mm_min_epi16(__m128i, __m128i);
_mm_min_epi32	__m128i _mm_min_epi32(__m128i val1, __m128i val2);
_mm_min_epi32	__m128i _mm_min_epi32(__m128i, __m128i);
_mm_min_epi8	__m128i _mm_min_epi8(__m128i val1, __m128i val2);
_mm_min_epi8	__m128i _mm_min_epi8(__m128i, __m128i);
_mm_min_epu16	__m128i _mm_min_epu16(__m128i val1, __m128i val2);
_mm_min_epu16	__m128i _mm_min_epu16(__m128i, __m128i);
_mm_min_epu32	__m128i _mm_min_epu32(__m128i val1, __m128i val2);
_mm_min_epu32	__m128i _mm_min_epu32(__m128i, __m128i);
_mm_min_epu8	__m128i _mm_min_epu8(__m128i _A, __m128i _B);
_mm_min_epu8	__m128i _mm_min_epu8(__m128i, __m128i);
_mm_min_pd	__m128d _mm_min_pd(__m128d _A, __m128d _B);
_mm_min_pd	__m128d _mm_min_pd(__m128d, __m128d);
_mm_min_ps	__m128 _mm_min_ps(__m128 _A, __m128 _B);
_mm_min_ps	__m128 _mm_min_ps(__m128, __m128);
_mm_min_sd	__m128d _mm_min_sd(__m128d _A, __m128d _B);
_mm_min_sd	__m128d _mm_min_sd(__m128d, __m128d);
_mm_min_ss	__m128 _mm_min_ss(__m128 _A, __m128 _B);
_mm_min_ss	__m128 _mm_min_ss(__m128, __m128);
_mm_minpos_epu16	__m128i _mm_minpos_epu16(__m128i shortValues);
_mm_minpos_epu16	__m128i _mm_minpos_epu16(__m128i);
_mm_monitor	void _mm_monitor(void const *, unsigned int, unsigned int);
_mm_monitor	void _mm_monitor(void const *p, unsigned extensions, unsigned hints);
_mm_move_epi64	__m128i _mm_move_epi64(__m128i _Q);
_mm_move_epi64	__m128i _mm_move_epi64(__m128i);
_mm_move_sd	__m128d _mm_move_sd(__m128d _A, __m128d _B);
_mm_move_sd	__m128d _mm_move_sd(__m128d, __m128d);
_mm_move_ss	__m128 _mm_move_ss(__m128 _A, __m128 _B);
_mm_move_ss	__m128 _mm_move_ss(__m128, __m128);
_mm_movedup_pd	__m128d _mm_movedup_pd(__m128d a);
_mm_movedup_pd	__m128d _mm_movedup_pd(__m128d);
_mm_movehdup_ps	__m128 _mm_movehdup_ps(__m128 a);
_mm_movehdup_ps	__m128 _mm_movehdup_ps(__m128);
_mm_movehl_ps	__m128 _mm_movehl_ps(__m128, __m128);
_mm_movehl_ps	__m128 _mm_movehl_ps(__m128, __m128);
_mm_moveldup_ps	__m128 _mm_moveldup_ps(__m128 a);
_mm_moveldup_ps	__m128 _mm_moveldup_ps(__m128);
_mm_movelh_ps	__m128 _mm_movelh_ps(__m128, __m128);
_mm_movelh_ps	__m128 _mm_movelh_ps(__m128, __m128);
_mm_movemask_epi8	int _mm_movemask_epi8(__m128i _A);
_mm_movemask_epi8	int _mm_movemask_epi8(__m128i);
_mm_movemask_pd	int _mm_movemask_pd(__m128d _A);
_mm_movemask_pd	int _mm_movemask_pd(__m128d);
_mm_movemask_ps	int _mm_movemask_ps(__m128 _A);
_mm_movemask_ps	int _mm_movemask_ps(__m128);
_mm_movepi64_pi64	__m64 _mm_movepi64_pi64(__m128i _Q);
_mm_movpi64_epi64	__m128i _mm_movpi64_epi64(__m64 _Q);
_mm_mpsadbw_epu8	__m128i _mm_mpsadbw_epu8(__m128i s1, __m128i s2, const int msk);
_mm_mpsadbw_epu8	__m128i _mm_mpsadbw_epu8(__m128i, __m128i, int);
_mm_msub_pd	__m128d _mm_msub_pd(__m128d, __m128d, __m128d);
_mm_msub_ps	__m128 _mm_msub_ps(__m128, __m128, __m128);
_mm_msub_sd	__m128d _mm_msub_sd(__m128d, __m128d, __m128d);
_mm_msub_ss	__m128 _mm_msub_ss(__m128, __m128, __m128);
_mm_msubadd_pd	__m128d _mm_msubadd_pd(__m128d, __m128d, __m128d);
_mm_msubadd_ps	__m128 _mm_msubadd_ps(__m128, __m128, __m128);
_mm_mul_epi32	__m128i _mm_mul_epi32(__m128i a, __m128i b);
_mm_mul_epi32	__m128i _mm_mul_epi32(__m128i, __m128i);
_mm_mul_epu32	__m128i _mm_mul_epu32(__m128i _A, __m128i _B);
_mm_mul_epu32	__m128i _mm_mul_epu32(__m128i, __m128i);
_mm_mul_pd	__m128d _mm_mul_pd(__m128d _A, __m128d _B);
_mm_mul_pd	__m128d _mm_mul_pd(__m128d, __m128d);
_mm_mul_ps	__m128 _mm_mul_ps(__m128 _A, __m128 _B);
_mm_mul_ps	__m128 _mm_mul_ps(__m128, __m128);
_mm_mul_sd	__m128d _mm_mul_sd(__m128d _A, __m128d _B);
_mm_mul_sd	__m128d _mm_mul_sd(__m128d, __m128d);
_mm_mul_ss	__m128 _mm_mul_ss(__m128 _A, __m128 _B);
_mm_mul_ss	__m128 _mm_mul_ss(__m128, __m128);
_mm_mul_su32	__m64 _mm_mul_su32(__m64 _A, __m64 _B);
_mm_mulhi_epi16	__m128i _mm_mulhi_epi16(__m128i _A, __m128i _B);
_mm_mulhi_epi16	__m128i _mm_mulhi_epi16(__m128i, __m128i);
_mm_mulhi_epu16	__m128i _mm_mulhi_epu16(__m128i _A, __m128i _B);
_mm_mulhi_epu16	__m128i _mm_mulhi_epu16(__m128i, __m128i);
_mm_mulhrs_epi16	__m128i _mm_mulhrs_epi16(__m128i a, __m128i b);
_mm_mulhrs_epi16	__m128i _mm_mulhrs_epi16(__m128i, __m128i);
_mm_mulhrs_pi16	__m64 _mm_mulhrs_pi16(__m64 a, __m64 b);
_mm_mulhrs_pi16	__m64 _mm_mulhrs_pi16(__m64, __m64);
_mm_mullo_epi16	__m128i _mm_mullo_epi16(__m128i _A, __m128i _B);
_mm_mullo_epi16	__m128i _mm_mullo_epi16(__m128i, __m128i);
_mm_mullo_epi32	__m128i _mm_mullo_epi32(__m128i a, __m128i b);
_mm_mullo_epi32	__m128i _mm_mullo_epi32(__m128i, __m128i);
_mm_mwait	void _mm_mwait(unsigned extensions, unsigned hints);
_mm_mwait	void _mm_mwait(unsigned int, unsigned int);
_mm_nmacc_pd	__m128d _mm_nmacc_pd(__m128d, __m128d, __m128d);
_mm_nmacc_ps	__m128 _mm_nmacc_ps(__m128, __m128, __m128);
_mm_nmacc_sd	__m128d _mm_nmacc_sd(__m128d, __m128d, __m128d);
_mm_nmacc_ss	__m128 _mm_nmacc_ss(__m128, __m128, __m128);
_mm_nmsub_pd	__m128d _mm_nmsub_pd(__m128d, __m128d, __m128d);
_mm_nmsub_ps	__m128 _mm_nmsub_ps(__m128, __m128, __m128);
_mm_nmsub_sd	__m128d _mm_nmsub_sd(__m128d, __m128d, __m128d);
_mm_nmsub_ss	__m128 _mm_nmsub_ss(__m128, __m128, __m128);
_mm_or_pd	__m128d _mm_or_pd(__m128d _A, __m128d _B);
_mm_or_pd	__m128d _mm_or_pd(__m128d, __m128d);
_mm_or_ps	__m128 _mm_or_ps(__m128 _A, __m128 _B);
_mm_or_ps	__m128 _mm_or_ps(__m128, __m128);
_mm_or_si128	__m128i _mm_or_si128(__m128i _A, __m128i _B);
_mm_or_si128	__m128i _mm_or_si128(__m128i, __m128i);
_mm_packs_epi16	__m128i _mm_packs_epi16(__m128i _A, __m128i _B);
_mm_packs_epi16	__m128i _mm_packs_epi16(__m128i, __m128i);
_mm_packs_epi32	__m128i _mm_packs_epi32(__m128i _A, __m128i _B);
_mm_packs_epi32	__m128i _mm_packs_epi32(__m128i, __m128i);
_mm_packus_epi16	__m128i _mm_packus_epi16(__m128i _A, __m128i _B);
_mm_packus_epi16	__m128i _mm_packus_epi16(__m128i, __m128i);
_mm_packus_epi32	__m128i _mm_packus_epi32(__m128i val1, __m128i val2);
_mm_packus_epi32	__m128i _mm_packus_epi32(__m128i, __m128i);
_mm_pause	void _mm_pause(void);
_mm_pause	void _mm_pause(void);
_mm_perm_epi8	__m128i _mm_perm_epi8(__m128i, __m128i, __m128i);
_mm_permute_pd	__m128d __cdecl _mm_permute_pd(__m128d, int);
_mm_permute_ps	__m128 __cdecl _mm_permute_ps(__m128, int);
_mm_permute2_pd	__m128d _mm_permute2_pd(__m128d, __m128d, __m128i, int);
_mm_permute2_ps	__m128 _mm_permute2_ps(__m128, __m128, __m128i, int);
_mm_permutevar_pd	__m128d __cdecl _mm_permutevar_pd(__m128d, __m128i);
_mm_permutevar_ps	__m128 __cdecl _mm_permutevar_ps(__m128, __m128i);
_mm_popcnt_u32	int _mm_popcnt_u32(unsigned int v);
_mm_popcnt_u32	int _mm_popcnt_u32(unsigned int);
_mm_popcnt_u64	__int64 _mm_popcnt_u64(unsigned __int64 v);
_mm_popcnt_u64	__int64 _mm_popcnt_u64(unsigned __int64);
_mm_prefetch	void _mm_prefetch(char const *, int);
_mm_prefetch	void _mm_prefetch(char const*_A, int _Sel);
_mm_rcp_ps	__m128 _mm_rcp_ps(__m128 _A);
_mm_rcp_ps	__m128 _mm_rcp_ps(__m128);
_mm_rcp_ss	__m128 _mm_rcp_ss(__m128 _A);
_mm_rcp_ss	__m128 _mm_rcp_ss(__m128);
_mm_rot_epi16	__m128i _mm_rot_epi16(__m128i, __m128i);
_mm_rot_epi32	__m128i _mm_rot_epi32(__m128i, __m128i);
_mm_rot_epi64	__m128i _mm_rot_epi64(__m128i, __m128i);
_mm_rot_epi8	__m128i _mm_rot_epi8(__m128i, __m128i);
_mm_roti_epi16	__m128i _mm_roti_epi16(__m128i, int);
_mm_roti_epi32	__m128i _mm_roti_epi32(__m128i, int);
_mm_roti_epi64	__m128i _mm_roti_epi64(__m128i, int);
_mm_roti_epi8	__m128i _mm_roti_epi8(__m128i, int);
_mm_round_pd	__m128d _mm_round_pd(__m128d val, int iRoundMode);
_mm_round_pd	__m128d _mm_round_pd(__m128d, int);
_mm_round_ps	__m128 _mm_round_ps(__m128 val, int iRoundMode);
_mm_round_ps	__m128 _mm_round_ps(__m128, int);
_mm_round_sd	__m128d _mm_round_sd(__m128d dst, __m128d val, int iRoundMode);
_mm_round_sd	__m128d _mm_round_sd(__m128d, __m128d, int);
_mm_round_ss	__m128 _mm_round_ss(__m128 dst, __m128 val, int iRoundMode);
_mm_round_ss	__m128 _mm_round_ss(__m128, __m128, int);
_mm_rsqrt_ps	__m128 _mm_rsqrt_ps(__m128 _A);
_mm_rsqrt_ps	__m128 _mm_rsqrt_ps(__m128);
_mm_rsqrt_ss	__m128 _mm_rsqrt_ss(__m128 _A);
_mm_rsqrt_ss	__m128 _mm_rsqrt_ss(__m128);
_mm_sad_epu8	__m128i _mm_sad_epu8(__m128i _A, __m128i _B);
_mm_sad_epu8	__m128i _mm_sad_epu8(__m128i, __m128i);
_mm_set_epi16	__m128i _mm_set_epi16(short _W7, short _W6, short _W5, short _W4, short _W3, short _W2, short _W1, short _W0);
_mm_set_epi16	__m128i _mm_set_epi16(short, short, short, short, short, short, short, short);
_mm_set_epi32	__m128i _mm_set_epi32(int _I3, int _I2, int _I1, int _I0);
_mm_set_epi32	__m128i _mm_set_epi32(int, int, int, int);
_mm_set_epi64	__m128i _mm_set_epi64(__m64 _Q1, __m64 _Q0);
_mm_set_epi64x	__m128i _mm_set_epi64x(__int64, __int64);
_mm_set_epi8	__m128i _mm_set_epi8(char _B15, char _B14, char _B13, char _B12, char _B11, char _B10, char _B9, char _B8, char _B7, char _B6, char _B5, char _B4, char _B3, char _B2, char _B1, char _B0);
_mm_set_epi8	__m128i _mm_set_epi8(char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char);
_mm_set_pd	__m128d _mm_set_pd(double _Z, double _Y);
_mm_set_pd	__m128d _mm_set_pd(double, double);
_mm_set_pi16	__m64 _mm_set_pi16(short _S3, short _S2, short _S1, short _S0);
_mm_set_pi32	__m64 _mm_set_pi32(int _I1, int _I0);
_mm_set_pi8	__m64 _mm_set_pi8(char _B7, char _B6, char _B5, char _B4, char _B3, char _B2, char _B1, char _B0);
_mm_set_ps	__m128 _mm_set_ps(float _A, float _B, float _C, float _D);
_mm_set_ps	__m128 _mm_set_ps(float, float, float, float);
_mm_set_ps1	__m128 _mm_set_ps1(float _A);
_mm_set_ps1	__m128 _mm_set_ps1(float);
_mm_set_sd	__m128d _mm_set_sd(double _W);
_mm_set_sd	__m128d _mm_set_sd(double);
_mm_set_ss	__m128 _mm_set_ss(float _A);
_mm_set_ss	__m128 _mm_set_ss(float);
_mm_set1_epi16	__m128i _mm_set1_epi16(short _W);
_mm_set1_epi16	__m128i _mm_set1_epi16(short);
_mm_set1_epi32	__m128i _mm_set1_epi32(int _I);
_mm_set1_epi32	__m128i _mm_set1_epi32(int);
_mm_set1_epi64	__m128i _mm_set1_epi64(__m64 _Q);
_mm_set1_epi64x	__m128i _mm_set1_epi64x(__int64);
_mm_set1_epi8	__m128i _mm_set1_epi8(char _B);
_mm_set1_epi8	__m128i _mm_set1_epi8(char);
_mm_set1_pd	__m128d _mm_set1_pd(double _A);
_mm_set1_pd	__m128d _mm_set1_pd(double);
_mm_set1_pi16	__m64 _mm_set1_pi16(short _S);
_mm_set1_pi32	__m64 _mm_set1_pi32(int _I);
_mm_set1_pi8	__m64 _mm_set1_pi8(char _B);
_mm_setcsr	void _mm_setcsr(unsigned int);
_mm_setcsr	void _mm_setcsr(unsigned int);
_mm_setl_epi64	__m128i _mm_setl_epi64(__m128i _Q);
_mm_setl_epi64	__m128i _mm_setl_epi64(__m128i);
_mm_setr_epi16	__m128i _mm_setr_epi16(short _W0, short _W1, short _W2, short _W3, short _W4, short _W5, short _W6, short _W7);
_mm_setr_epi16	__m128i _mm_setr_epi16(short, short, short, short, short, short, short, short);
_mm_setr_epi32	__m128i _mm_setr_epi32(int _I0, int _I1, int _I2, int _I3);
_mm_setr_epi32	__m128i _mm_setr_epi32(int, int, int, int);
_mm_setr_epi64	__m128i _mm_setr_epi64(__m64 _Q0, __m64 _Q1);
_mm_setr_epi8	__m128i _mm_setr_epi8(char _B15, char _B14, char _B13, char _B12, char _B11, char _B10, char _B9, char _B8, char _B7, char _B6, char _B5, char _B4, char _B3, char _B2, char _B1, char _B0);
_mm_setr_epi8	__m128i _mm_setr_epi8(char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char);
_mm_setr_pd	__m128d _mm_setr_pd(double _Y, double _Z);
_mm_setr_pd	__m128d _mm_setr_pd(double, double);
_mm_setr_pi16	__m64 _mm_setr_pi16(short _S3, short _S2, short _S1, short _S0);
_mm_setr_pi32	__m64 _mm_setr_pi32(int _I1, int _I0);
_mm_setr_pi8	__m64 _mm_setr_pi8(char _B7, char _B6, char _B5, char _B4, char _B3, char _B2, char _B1, char _B0);
_mm_setr_ps	__m128 _mm_setr_ps(float _A, float _B, float _C, float _D);
_mm_setr_ps	__m128 _mm_setr_ps(float, float, float, float);
_mm_setzero_pd	__m128d _mm_setzero_pd(void);
_mm_setzero_pd	__m128d _mm_setzero_pd(void);
_mm_setzero_ps	__m128 _mm_setzero_ps(void);
_mm_setzero_ps	__m128 _mm_setzero_ps(void);
_mm_setzero_si128	__m128i _mm_setzero_si128(void);
_mm_setzero_si128	__m128i _mm_setzero_si128(void);
_mm_setzero_si64	__m64 _mm_setzero_si64(void);
_mm_sfence	void _mm_sfence(void);
_mm_sfence	void _mm_sfence(void);
_mm_sha_epi16	__m128i _mm_sha_epi16(__m128i, __m128i);
_mm_sha_epi32	__m128i _mm_sha_epi32(__m128i, __m128i);
_mm_sha_epi64	__m128i _mm_sha_epi64(__m128i, __m128i);
_mm_sha_epi8	__m128i _mm_sha_epi8(__m128i, __m128i);
_mm_shl_epi16	__m128i _mm_shl_epi16(__m128i, __m128i);
_mm_shl_epi32	__m128i _mm_shl_epi32(__m128i, __m128i);
_mm_shl_epi64	__m128i _mm_shl_epi64(__m128i, __m128i);
_mm_shl_epi8	__m128i _mm_shl_epi8(__m128i, __m128i);
_mm_shuffle_epi32	__m128i _mm_shuffle_epi32(__m128i _A, int _Imm);
_mm_shuffle_epi32	__m128i _mm_shuffle_epi32(__m128i, int);
_mm_shuffle_epi8	__m128i _mm_shuffle_epi8(__m128i a, __m128i b);
_mm_shuffle_epi8	__m128i _mm_shuffle_epi8(__m128i, __m128i);
_mm_shuffle_pd	__m128d _mm_shuffle_pd(__m128d _A, __m128d _B, int _I);
_mm_shuffle_pd	__m128d _mm_shuffle_pd(__m128d, __m128d, int);
_mm_shuffle_pi8	__m64 _mm_shuffle_pi8(__m64 a, __m64 b);
_mm_shuffle_pi8	__m64 _mm_shuffle_pi8(__m64, __m64);
_mm_shuffle_ps	__m128 _mm_shuffle_ps(__m128 _A, __m128 _B, unsigned int _Imm8);
_mm_shuffle_ps	__m128 _mm_shuffle_ps(__m128, __m128, unsigned int);
_mm_shufflehi_epi16	__m128i _mm_shufflehi_epi16(__m128i _A, int _Imm);
_mm_shufflehi_epi16	__m128i _mm_shufflehi_epi16(__m128i, int);
_mm_shufflelo_epi16	__m128i _mm_shufflelo_epi16(__m128i _A, int _Imm);
_mm_shufflelo_epi16	__m128i _mm_shufflelo_epi16(__m128i, int);
_mm_sign_epi16	__m128i _mm_sign_epi16(__m128i a, __m128i b);
_mm_sign_epi16	__m128i _mm_sign_epi16(__m128i, __m128i);
_mm_sign_epi32	__m128i _mm_sign_epi32(__m128i a, __m128i b);
_mm_sign_epi32	__m128i _mm_sign_epi32(__m128i, __m128i);
_mm_sign_epi8	__m128i _mm_sign_epi8(__m128i a, __m128i b);
_mm_sign_epi8	__m128i _mm_sign_epi8(__m128i, __m128i);
_mm_sign_pi16	__m64 _mm_sign_pi16(__m64 a, __m64 b);
_mm_sign_pi16	__m64 _mm_sign_pi16(__m64, __m64);
_mm_sign_pi32	__m64 _mm_sign_pi32(__m64 a, __m64 b);
_mm_sign_pi32	__m64 _mm_sign_pi32(__m64, __m64);
_mm_sign_pi8	__m64 _mm_sign_pi8(__m64 a, __m64 b);
_mm_sign_pi8	__m64 _mm_sign_pi8(__m64, __m64);
_mm_sll_epi16	__m128i _mm_sll_epi16(__m128i _A, __m128i _Count);
_mm_sll_epi16	__m128i _mm_sll_epi16(__m128i, __m128i);
_mm_sll_epi32	__m128i _mm_sll_epi32(__m128i _A, __m128i _Count);
_mm_sll_epi32	__m128i _mm_sll_epi32(__m128i, __m128i);
_mm_sll_epi64	__m128i _mm_sll_epi64(__m128i _A, __m128i _Count);
_mm_sll_epi64	__m128i _mm_sll_epi64(__m128i, __m128i);
_mm_slli_epi16	__m128i _mm_slli_epi16(__m128i _A, int _Count);
_mm_slli_epi16	__m128i _mm_slli_epi16(__m128i, int);
_mm_slli_epi32	__m128i _mm_slli_epi32(__m128i _A, int _Count);
_mm_slli_epi32	__m128i _mm_slli_epi32(__m128i, int);
_mm_slli_epi64	__m128i _mm_slli_epi64(__m128i _A, int _Count);
_mm_slli_epi64	__m128i _mm_slli_epi64(__m128i, int);
_mm_slli_si128	__m128i _mm_slli_si128(__m128i _A, int _Imm);
_mm_slli_si128	__m128i _mm_slli_si128(__m128i, int);
_mm_sllv_epi32	__m128i __cdecl _mm_sllv_epi32(__m128i, __m128i);
_mm_sllv_epi64	__m128i __cdecl _mm_sllv_epi64(__m128i, __m128i);
_mm_sqrt_pd	__m128d _mm_sqrt_pd(__m128d _A);
_mm_sqrt_pd	__m128d _mm_sqrt_pd(__m128d);
_mm_sqrt_ps	__m128 _mm_sqrt_ps(__m128 _A);
_mm_sqrt_ps	__m128 _mm_sqrt_ps(__m128);
_mm_sqrt_sd	__m128d _mm_sqrt_sd(__m128d _A, __m128d _B);
_mm_sqrt_sd	__m128d _mm_sqrt_sd(__m128d, __m128d);
_mm_sqrt_ss	__m128 _mm_sqrt_ss(__m128 _A);
_mm_sqrt_ss	__m128 _mm_sqrt_ss(__m128);
_mm_sra_epi16	__m128i _mm_sra_epi16(__m128i _A, __m128i _Count);
_mm_sra_epi16	__m128i _mm_sra_epi16(__m128i, __m128i);
_mm_sra_epi32	__m128i _mm_sra_epi32(__m128i _A, __m128i _Count);
_mm_sra_epi32	__m128i _mm_sra_epi32(__m128i, __m128i);
_mm_srai_epi16	__m128i _mm_srai_epi16(__m128i _A, int _Count);
_mm_srai_epi16	__m128i _mm_srai_epi16(__m128i, int);
_mm_srai_epi32	__m128i _mm_srai_epi32(__m128i _A, int _Count);
_mm_srai_epi32	__m128i _mm_srai_epi32(__m128i, int);
_mm_srav_epi32	__m128i __cdecl _mm_srav_epi32(__m128i, __m128i);
_mm_srl_epi16	__m128i _mm_srl_epi16(__m128i _A, __m128i _Count);
_mm_srl_epi16	__m128i _mm_srl_epi16(__m128i, __m128i);
_mm_srl_epi32	__m128i _mm_srl_epi32(__m128i _A, __m128i _Count);
_mm_srl_epi32	__m128i _mm_srl_epi32(__m128i, __m128i);
_mm_srl_epi64	__m128i _mm_srl_epi64(__m128i _A, __m128i _Count);
_mm_srl_epi64	__m128i _mm_srl_epi64(__m128i, __m128i);
_mm_srli_epi16	__m128i _mm_srli_epi16(__m128i _A, int _Count);
_mm_srli_epi16	__m128i _mm_srli_epi16(__m128i, int);
_mm_srli_epi32	__m128i _mm_srli_epi32(__m128i _A, int _Count);
_mm_srli_epi32	__m128i _mm_srli_epi32(__m128i, int);
_mm_srli_epi64	__m128i _mm_srli_epi64(__m128i _A, int _Count);
_mm_srli_epi64	__m128i _mm_srli_epi64(__m128i, int);
_mm_srli_si128	__m128i _mm_srli_si128(__m128i _A, int _Imm);
_mm_srli_si128	__m128i _mm_srli_si128(__m128i, int);
_mm_srlv_epi32	__m128i __cdecl _mm_srlv_epi32(__m128i, __m128i);
_mm_srlv_epi64	__m128i __cdecl _mm_srlv_epi64(__m128i, __m128i);
_mm_store_pd	void _mm_store_pd(double *, __m128d);
_mm_store_pd	void _mm_store_pd(double *_Dp, __m128d _A);
_mm_store_ps	void _mm_store_ps(float *, __m128);
_mm_store_ps	void _mm_store_ps(float *_V, __m128 _A);
_mm_store_ps1	void _mm_store_ps1(float *, __m128);
_mm_store_ps1	void _mm_store_ps1(float *_V, __m128 _A);
_mm_store_sd	void _mm_store_sd(double *, __m128d);
_mm_store_sd	void _mm_store_sd(double *_Dp, __m128d _A);
_mm_store_si128	void _mm_store_si128(__m128i *, __m128i);
_mm_store_si128	void _mm_store_si128(__m128i *_P, __m128i _B);
_mm_store_ss	void _mm_store_ss(float *, __m128);
_mm_store_ss	void _mm_store_ss(float *_V, __m128 _A);
_mm_store1_pd	void _mm_store1_pd(double *, __m128d);
_mm_store1_pd	void _mm_store1_pd(double *_Dp, __m128d _A);
_mm_storeh_pd	void _mm_storeh_pd(double *, __m128d);
_mm_storeh_pd	void _mm_storeh_pd(double *_Dp, __m128d _A);
_mm_storeh_pi	void _mm_storeh_pi(__m64 *, __m128);
_mm_storeh_pi	void _mm_storeh_pi(__m64 *, __m128);
_mm_storel_epi64	void _mm_storel_epi64(__m128i *, __m128i);
_mm_storel_epi64	void _mm_storel_epi64(__m128i *_P, __m128i _Q);
_mm_storel_pd	void _mm_storel_pd(double *, __m128d);
_mm_storel_pd	void _mm_storel_pd(double *_Dp, __m128d _A);
_mm_storel_pi	void _mm_storel_pi(__m64 *, __m128);
_mm_storel_pi	void _mm_storel_pi(__m64 *, __m128);
_mm_storer_pd	void _mm_storer_pd(double *, __m128d);
_mm_storer_pd	void _mm_storer_pd(double *_Dp, __m128d _A);
_mm_storer_ps	void _mm_storer_ps(float *, __m128);
_mm_storer_ps	void _mm_storer_ps(float *_V, __m128 _A);
_mm_storeu_pd	void _mm_storeu_pd(double *, __m128d);
_mm_storeu_pd	void _mm_storeu_pd(double *_Dp, __m128d _A);
_mm_storeu_ps	void _mm_storeu_ps(float *, __m128);
_mm_storeu_ps	void _mm_storeu_ps(float *_V, __m128 _A);
_mm_storeu_si128	void _mm_storeu_si128(__m128i *, __m128i);
_mm_storeu_si128	void _mm_storeu_si128(__m128i *_P, __m128i _B);
_mm_stream_load_si128	__m128i _mm_stream_load_si128(__m128i *);
_mm_stream_load_si128	__m128i _mm_stream_load_si128(__m128i* v1);
_mm_stream_pd	void _mm_stream_pd(double *, __m128d);
_mm_stream_pd	void _mm_stream_pd(double *_Dp, __m128d _A);
_mm_stream_pi	void _mm_stream_pi(__m64 *, __m64);
_mm_stream_ps	void _mm_stream_ps(float *, __m128);
_mm_stream_ps	void _mm_stream_ps(float *, __m128);
_mm_stream_sd	void _mm_stream_sd(double *, __m128d);
_mm_stream_si128	void _mm_stream_si128(__m128i *, __m128i);
_mm_stream_si128	void _mm_stream_si128(__m128i *_P, __m128i _A);
_mm_stream_si32	void _mm_stream_si32(int *, int);
_mm_stream_si32	void _mm_stream_si32(int *_P, int _I);
_mm_stream_si64x	void _mm_stream_si64x(__int64 *, __int64);
_mm_stream_ss	void _mm_stream_ss(float *, __m128);
_mm_sub_epi16	__m128i _mm_sub_epi16(__m128i _A, __m128i _B);
_mm_sub_epi16	__m128i _mm_sub_epi16(__m128i, __m128i);
_mm_sub_epi32	__m128i _mm_sub_epi32(__m128i _A, __m128i _B);
_mm_sub_epi32	__m128i _mm_sub_epi32(__m128i, __m128i);
_mm_sub_epi64	__m128i _mm_sub_epi64(__m128i _A, __m128i _B);
_mm_sub_epi64	__m128i _mm_sub_epi64(__m128i, __m128i);
_mm_sub_epi8	__m128i _mm_sub_epi8(__m128i _A, __m128i _B);
_mm_sub_epi8	__m128i _mm_sub_epi8(__m128i, __m128i);
_mm_sub_pd	__m128d _mm_sub_pd(__m128d _A, __m128d _B);
_mm_sub_pd	__m128d _mm_sub_pd(__m128d, __m128d);
_mm_sub_ps	__m128 _mm_sub_ps(__m128 _A, __m128 _B);
_mm_sub_ps	__m128 _mm_sub_ps(__m128, __m128);
_mm_sub_sd	__m128d _mm_sub_sd(__m128d _A, __m128d _B);
_mm_sub_sd	__m128d _mm_sub_sd(__m128d, __m128d);
_mm_sub_si64	__m64 _mm_sub_si64(__m64 _A, __m64 _B);
_mm_sub_ss	__m128 _mm_sub_ss(__m128 _A, __m128 _B);
_mm_sub_ss	__m128 _mm_sub_ss(__m128, __m128);
_mm_subs_epi16	__m128i _mm_subs_epi16(__m128i _A, __m128i _B);
_mm_subs_epi16	__m128i _mm_subs_epi16(__m128i, __m128i);
_mm_subs_epi8	__m128i _mm_subs_epi8(__m128i _A, __m128i _B);
_mm_subs_epi8	__m128i _mm_subs_epi8(__m128i, __m128i);
_mm_subs_epu16	__m128i _mm_subs_epu16(__m128i _A, __m128i _B);
_mm_subs_epu16	__m128i _mm_subs_epu16(__m128i, __m128i);
_mm_subs_epu8	__m128i _mm_subs_epu8(__m128i _A, __m128i _B);
_mm_subs_epu8	__m128i _mm_subs_epu8(__m128i, __m128i);
_mm_testc_pd	int __cdecl _mm_testc_pd(__m128d, __m128d);
_mm_testc_ps	int __cdecl _mm_testc_ps(__m128, __m128);
_mm_testc_si128	int _mm_testc_si128(__m128i mask, __m128i val);
_mm_testc_si128	int _mm_testc_si128(__m128i, __m128i);
_mm_testnzc_pd	int __cdecl _mm_testnzc_pd(__m128d, __m128d);
_mm_testnzc_ps	int __cdecl _mm_testnzc_ps(__m128, __m128);
_mm_testnzc_si128	int _mm_testnzc_si128(__m128i mask, __m128i s2);
_mm_testnzc_si128	int _mm_testnzc_si128(__m128i, __m128i);
_mm_testz_pd	int __cdecl _mm_testz_pd(__m128d, __m128d);
_mm_testz_ps	int __cdecl _mm_testz_ps(__m128, __m128);
_mm_testz_si128	int _mm_testz_si128(__m128i mask, __m128i val);
_mm_testz_si128	int _mm_testz_si128(__m128i, __m128i);
_mm_ucomieq_sd	int _mm_ucomieq_sd(__m128d _A, __m128d _B);
_mm_ucomieq_sd	int _mm_ucomieq_sd(__m128d, __m128d);
_mm_ucomieq_ss	int _mm_ucomieq_ss(__m128 _A, __m128 _B);
_mm_ucomieq_ss	int _mm_ucomieq_ss(__m128, __m128);
_mm_ucomige_sd	int _mm_ucomige_sd(__m128d _A, __m128d _B);
_mm_ucomige_sd	int _mm_ucomige_sd(__m128d, __m128d);
_mm_ucomige_ss	int _mm_ucomige_ss(__m128 _A, __m128 _B);
_mm_ucomige_ss	int _mm_ucomige_ss(__m128, __m128);
_mm_ucomigt_sd	int _mm_ucomigt_sd(__m128d _A, __m128d _B);
_mm_ucomigt_sd	int _mm_ucomigt_sd(__m128d, __m128d);
_mm_ucomigt_ss	int _mm_ucomigt_ss(__m128 _A, __m128 _B);
_mm_ucomigt_ss	int _mm_ucomigt_ss(__m128, __m128);
_mm_ucomile_sd	int _mm_ucomile_sd(__m128d _A, __m128d _B);
_mm_ucomile_sd	int _mm_ucomile_sd(__m128d, __m128d);
_mm_ucomile_ss	int _mm_ucomile_ss(__m128 _A, __m128 _B);
_mm_ucomile_ss	int _mm_ucomile_ss(__m128, __m128);
_mm_ucomilt_sd	int _mm_ucomilt_sd(__m128d _A, __m128d _B);
_mm_ucomilt_sd	int _mm_ucomilt_sd(__m128d, __m128d);
_mm_ucomilt_ss	int _mm_ucomilt_ss(__m128 _A, __m128 _B);
_mm_ucomilt_ss	int _mm_ucomilt_ss(__m128, __m128);
_mm_ucomineq_sd	int _mm_ucomineq_sd(__m128d _A, __m128d _B);
_mm_ucomineq_sd	int _mm_ucomineq_sd(__m128d, __m128d);
_mm_ucomineq_ss	int _mm_ucomineq_ss(__m128 _A, __m128 _B);
_mm_ucomineq_ss	int _mm_ucomineq_ss(__m128, __m128);
_mm_unpackhi_epi16	__m128i _mm_unpackhi_epi16(__m128i _A, __m128i _B);
_mm_unpackhi_epi16	__m128i _mm_unpackhi_epi16(__m128i, __m128i);
_mm_unpackhi_epi32	__m128i _mm_unpackhi_epi32(__m128i _A, __m128i _B);
_mm_unpackhi_epi32	__m128i _mm_unpackhi_epi32(__m128i, __m128i);
_mm_unpackhi_epi64	__m128i _mm_unpackhi_epi64(__m128i _A, __m128i _B);
_mm_unpackhi_epi64	__m128i _mm_unpackhi_epi64(__m128i, __m128i);
_mm_unpackhi_epi8	__m128i _mm_unpackhi_epi8(__m128i _A, __m128i _B);
_mm_unpackhi_epi8	__m128i _mm_unpackhi_epi8(__m128i, __m128i);
_mm_unpackhi_pd	__m128d _mm_unpackhi_pd(__m128d _A, __m128d _B);
_mm_unpackhi_pd	__m128d _mm_unpackhi_pd(__m128d, __m128d);
_mm_unpackhi_ps	__m128 _mm_unpackhi_ps(__m128 _A, __m128 _B);
_mm_unpackhi_ps	__m128 _mm_unpackhi_ps(__m128, __m128);
_mm_unpacklo_epi16	__m128i _mm_unpacklo_epi16(__m128i _A, __m128i _B);
_mm_unpacklo_epi16	__m128i _mm_unpacklo_epi16(__m128i, __m128i);
_mm_unpacklo_epi32	__m128i _mm_unpacklo_epi32(__m128i _A, __m128i _B);
_mm_unpacklo_epi32	__m128i _mm_unpacklo_epi32(__m128i, __m128i);
_mm_unpacklo_epi64	__m128i _mm_unpacklo_epi64(__m128i _A, __m128i _B);
_mm_unpacklo_epi64	__m128i _mm_unpacklo_epi64(__m128i, __m128i);
_mm_unpacklo_epi8	__m128i _mm_unpacklo_epi8(__m128i _A, __m128i _B);
_mm_unpacklo_epi8	__m128i _mm_unpacklo_epi8(__m128i, __m128i);
_mm_unpacklo_pd	__m128d _mm_unpacklo_pd(__m128d _A, __m128d _B);
_mm_unpacklo_pd	__m128d _mm_unpacklo_pd(__m128d, __m128d);
_mm_unpacklo_ps	__m128 _mm_unpacklo_ps(__m128 _A, __m128 _B);
_mm_unpacklo_ps	__m128 _mm_unpacklo_ps(__m128, __m128);
_mm_xor_pd	__m128d _mm_xor_pd(__m128d _A, __m128d _B);
_mm_xor_pd	__m128d _mm_xor_pd(__m128d, __m128d);
_mm_xor_ps	__m128 _mm_xor_ps(__m128 _A, __m128 _B);
_mm_xor_ps	__m128 _mm_xor_ps(__m128, __m128);
_mm_xor_si128	__m128i _mm_xor_si128(__m128i _A, __m128i _B);
_mm_xor_si128	__m128i _mm_xor_si128(__m128i, __m128i);
_mm256_abs_epi16	__m256i __cdecl _mm256_abs_epi16(__m256i);
_mm256_abs_epi32	__m256i __cdecl _mm256_abs_epi32(__m256i);
_mm256_abs_epi8	__m256i __cdecl _mm256_abs_epi8(__m256i);
_mm256_add_epi16	__m256i __cdecl _mm256_add_epi16(__m256i, __m256i);
_mm256_add_epi32	__m256i __cdecl _mm256_add_epi32(__m256i, __m256i);
_mm256_add_epi64	__m256i __cdecl _mm256_add_epi64(__m256i, __m256i);
_mm256_add_epi8	__m256i __cdecl _mm256_add_epi8(__m256i, __m256i);
_mm256_add_pd	__m256d __cdecl _mm256_add_pd(__m256d, __m256d);
_mm256_add_ps	__m256 __cdecl _mm256_add_ps(__m256, __m256);
_mm256_adds_epi16	__m256i __cdecl _mm256_adds_epi16(__m256i, __m256i);
_mm256_adds_epi8	__m256i __cdecl _mm256_adds_epi8(__m256i, __m256i);
_mm256_adds_epu16	__m256i __cdecl _mm256_adds_epu16(__m256i, __m256i);
_mm256_adds_epu8	__m256i __cdecl _mm256_adds_epu8(__m256i, __m256i);
_mm256_addsub_pd	__m256d __cdecl _mm256_addsub_pd(__m256d, __m256d);
_mm256_addsub_ps	__m256 __cdecl _mm256_addsub_ps(__m256, __m256);
_mm256_alignr_epi8	__m256i __cdecl _mm256_alignr_epi8(__m256i, __m256i, const int);
_mm256_and_pd	__m256d __cdecl _mm256_and_pd(__m256d, __m256d);
_mm256_and_ps	__m256 __cdecl _mm256_and_ps(__m256, __m256);
_mm256_and_si256	__m256i __cdecl _mm256_and_si256(__m256i, __m256i);
_mm256_andnot_pd	__m256d __cdecl _mm256_andnot_pd(__m256d, __m256d);
_mm256_andnot_ps	__m256 __cdecl _mm256_andnot_ps(__m256, __m256);
_mm256_andnot_si256	__m256i __cdecl _mm256_andnot_si256(__m256i, __m256i);
_mm256_avg_epu16	__m256i __cdecl _mm256_avg_epu16(__m256i, __m256i);
_mm256_avg_epu8	__m256i __cdecl _mm256_avg_epu8(__m256i, __m256i);
_mm256_blend_epi16	__m256i __cdecl _mm256_blend_epi16(__m256i, __m256i, const int);
_mm256_blend_epi32	__m256i __cdecl _mm256_blend_epi32(__m256i,__m256i, const int);
_mm256_blend_pd	__m256d __cdecl _mm256_blend_pd(__m256d, __m256d, const int);
_mm256_blend_ps	__m256 __cdecl _mm256_blend_ps(__m256, __m256, const int);
_mm256_blendv_epi8	__m256i __cdecl _mm256_blendv_epi8(__m256i, __m256i, __m256i);
_mm256_blendv_pd	__m256d __cdecl _mm256_blendv_pd(__m256d, __m256d, __m256d);
_mm256_blendv_ps	__m256 __cdecl _mm256_blendv_ps(__m256, __m256, __m256);
_mm256_broadcast_pd	__m256d __cdecl _mm256_broadcast_pd(__m128d const *);
_mm256_broadcast_ps	__m256 __cdecl _mm256_broadcast_ps(__m128 const *);
_mm256_broadcast_sd	__m256d __cdecl _mm256_broadcast_sd(double const *);
_mm256_broadcast_ss	__m256 __cdecl _mm256_broadcast_ss(float const *);
_mm256_broadcastb_epi8	__m256i __cdecl _mm256_broadcastb_epi8(__m128i);
_mm256_broadcastd_epi32	__m256i __cdecl _mm256_broadcastd_epi32(__m128i);
_mm256_broadcastq_epi64	__m256i __cdecl _mm256_broadcastq_epi64(__m128i);
_mm256_broadcastsd_pd	__m256d __cdecl _mm256_broadcastsd_pd(__m128d);
_mm256_broadcastsi128_si256	__m256i __cdecl _mm256_broadcastsi128_si256(__m128i);
_mm256_broadcastss_ps	__m256 __cdecl _mm256_broadcastss_ps(__m128);
_mm256_broadcastw_epi16	__m256i __cdecl _mm256_broadcastw_epi16(__m128i);
_mm256_castpd_ps	__m256 __cdecl _mm256_castpd_ps(__m256d);
_mm256_castpd_si256	__m256i __cdecl _mm256_castpd_si256(__m256d);
_mm256_castpd128_pd256	__m256d __cdecl _mm256_castpd128_pd256(__m128d);
_mm256_castpd256_pd128	__m128d __cdecl _mm256_castpd256_pd128(__m256d);
_mm256_castps_pd	__m256d __cdecl _mm256_castps_pd(__m256);
_mm256_castps_si256	__m256i __cdecl _mm256_castps_si256(__m256);
_mm256_castps128_ps256	__m256 __cdecl _mm256_castps128_ps256(__m128);
_mm256_castps256_ps128	__m128 __cdecl _mm256_castps256_ps128(__m256);
_mm256_castsi128_si256	__m256i __cdecl _mm256_castsi128_si256(__m128i);
_mm256_castsi256_pd	__m256d __cdecl _mm256_castsi256_pd(__m256i);
_mm256_castsi256_ps	__m256 __cdecl _mm256_castsi256_ps(__m256i);
_mm256_castsi256_si128	__m128i __cdecl _mm256_castsi256_si128(__m256i);
_mm256_cmov_si256	__m256i _mm256_cmov_si256(__m256i, __m256i, __m256i);
_mm256_cmp_pd	__m256d __cdecl _mm256_cmp_pd(__m256d, __m256d, const int);
_mm256_cmp_ps	__m256 __cdecl _mm256_cmp_ps(__m256, __m256, const int);
_mm256_cmpeq_epi16	__m256i __cdecl _mm256_cmpeq_epi16(__m256i, __m256i);
_mm256_cmpeq_epi32	__m256i __cdecl _mm256_cmpeq_epi32(__m256i, __m256i);
_mm256_cmpeq_epi64	__m256i __cdecl _mm256_cmpeq_epi64(__m256i, __m256i);
_mm256_cmpeq_epi8	__m256i __cdecl _mm256_cmpeq_epi8(__m256i, __m256i);
_mm256_cmpgt_epi16	__m256i __cdecl _mm256_cmpgt_epi16(__m256i, __m256i);
_mm256_cmpgt_epi32	__m256i __cdecl _mm256_cmpgt_epi32(__m256i, __m256i);
_mm256_cmpgt_epi64	__m256i __cdecl _mm256_cmpgt_epi64(__m256i, __m256i);
_mm256_cmpgt_epi8	__m256i __cdecl _mm256_cmpgt_epi8(__m256i, __m256i);
_mm256_cvtepi16_epi32	__m256i __cdecl _mm256_cvtepi16_epi32(__m128i);
_mm256_cvtepi16_epi64	__m256i __cdecl _mm256_cvtepi16_epi64(__m128i);
_mm256_cvtepi32_epi64	__m256i __cdecl _mm256_cvtepi32_epi64(__m128i);
_mm256_cvtepi32_pd	__m256d __cdecl _mm256_cvtepi32_pd(__m128i);
_mm256_cvtepi32_ps	__m256 __cdecl _mm256_cvtepi32_ps(__m256i);
_mm256_cvtepi8_epi16	__m256i __cdecl _mm256_cvtepi8_epi16(__m128i);
_mm256_cvtepi8_epi32	__m256i __cdecl _mm256_cvtepi8_epi32(__m128i);
_mm256_cvtepi8_epi64	__m256i __cdecl _mm256_cvtepi8_epi64(__m128i);
_mm256_cvtepu16_epi32	__m256i __cdecl _mm256_cvtepu16_epi32(__m128i);
_mm256_cvtepu16_epi64	__m256i __cdecl _mm256_cvtepu16_epi64(__m128i);
_mm256_cvtepu32_epi64	__m256i __cdecl _mm256_cvtepu32_epi64(__m128i);
_mm256_cvtepu8_epi16	__m256i __cdecl _mm256_cvtepu8_epi16(__m128i);
_mm256_cvtepu8_epi32	__m256i __cdecl _mm256_cvtepu8_epi32(__m128i);
_mm256_cvtepu8_epi64	__m256i __cdecl _mm256_cvtepu8_epi64(__m128i);
_mm256_cvtpd_epi32	__m128i __cdecl _mm256_cvtpd_epi32(__m256d);
_mm256_cvtpd_ps	__m128 __cdecl _mm256_cvtpd_ps(__m256d);
_mm256_cvtph_ps	__m256 __cdecl _mm256_cvtph_ps(__m128i);
_mm256_cvtps_epi32	__m256i __cdecl _mm256_cvtps_epi32(__m256);
_mm256_cvtps_pd	__m256d __cdecl _mm256_cvtps_pd(__m128);
_mm256_cvtps_ph	__m128i __cdecl _mm256_cvtps_ph(__m256, int);
_mm256_cvttpd_epi32	__m128i __cdecl _mm256_cvttpd_epi32(__m256d);
_mm256_cvttps_epi32	__m256i __cdecl _mm256_cvttps_epi32(__m256);
_mm256_div_pd	__m256d __cdecl _mm256_div_pd(__m256d, __m256d);
_mm256_div_ps	__m256 __cdecl _mm256_div_ps(__m256, __m256);
_mm256_dp_ps	__m256 __cdecl _mm256_dp_ps(__m256, __m256, const int);
_mm256_extractf128_pd	__m128d __cdecl _mm256_extractf128_pd(__m256d, const int);
_mm256_extractf128_ps	__m128 __cdecl _mm256_extractf128_ps(__m256, const int);
_mm256_extractf128_si256	__m128i __cdecl _mm256_extractf128_si256(__m256i, const int);
_mm256_extracti128_si256	__m128i __cdecl _mm256_extracti128_si256(__m256i, const int);
_mm256_fmadd_pd	__m256d __cdecl _mm256_fmadd_pd(__m256d, __m256d, __m256d);
_mm256_fmadd_ps	__m256 __cdecl _mm256_fmadd_ps(__m256, __m256, __m256);
_mm256_fmaddsub_pd	__m256d __cdecl _mm256_fmaddsub_pd(__m256d, __m256d, __m256d);
_mm256_fmaddsub_ps	__m256 __cdecl _mm256_fmaddsub_ps(__m256, __m256, __m256);
_mm256_fmsub_pd	__m256d __cdecl _mm256_fmsub_pd(__m256d, __m256d, __m256d);
_mm256_fmsub_ps	__m256 __cdecl _mm256_fmsub_ps(__m256, __m256, __m256);
_mm256_fmsubadd_pd	__m256d __cdecl _mm256_fmsubadd_pd(__m256d, __m256d, __m256d);
_mm256_fmsubadd_ps	__m256 __cdecl _mm256_fmsubadd_ps(__m256, __m256, __m256);
_mm256_fnmadd_pd	__m256d __cdecl _mm256_fnmadd_pd(__m256d, __m256d, __m256d);
_mm256_fnmadd_ps	__m256 __cdecl _mm256_fnmadd_ps(__m256, __m256, __m256);
_mm256_fnmsub_pd	__m256d __cdecl _mm256_fnmsub_pd(__m256d, __m256d, __m256d);
_mm256_fnmsub_ps	__m256 __cdecl _mm256_fnmsub_ps(__m256, __m256, __m256);
_mm256_frcz_pd	__m256d _mm256_frcz_pd(__m256d);
_mm256_frcz_ps	__m256 _mm256_frcz_ps(__m256);
_mm256_hadd_epi16	__m256i __cdecl _mm256_hadd_epi16(__m256i, __m256i);
_mm256_hadd_epi32	__m256i __cdecl _mm256_hadd_epi32(__m256i, __m256i);
_mm256_hadd_pd	__m256d __cdecl _mm256_hadd_pd(__m256d, __m256d);
_mm256_hadd_ps	__m256 __cdecl _mm256_hadd_ps(__m256, __m256);
_mm256_hadds_epi16	__m256i __cdecl _mm256_hadds_epi16(__m256i, __m256i);
_mm256_hsub_epi16	__m256i __cdecl _mm256_hsub_epi16(__m256i, __m256i);
_mm256_hsub_epi32	__m256i __cdecl _mm256_hsub_epi32(__m256i, __m256i);
_mm256_hsub_pd	__m256d __cdecl _mm256_hsub_pd(__m256d, __m256d);
_mm256_hsub_ps	__m256 __cdecl _mm256_hsub_ps(__m256, __m256);
_mm256_hsubs_epi16	__m256i __cdecl _mm256_hsubs_epi16(__m256i, __m256i);
_mm256_i32gather_epi32	__m256i __cdecl _mm256_i32gather_epi32(int const *, __m256i, const int );
_mm256_i32gather_epi64	__m256i __cdecl _mm256_i32gather_epi64(__int64 const *, __m128i, const int );
_mm256_i32gather_pd	__m256d __cdecl _mm256_i32gather_pd(double const *, __m128i, const int );
_mm256_i32gather_ps	__m256 __cdecl _mm256_i32gather_ps(float const *, __m256i, const int );
_mm256_i64gather_epi32	__m128i __cdecl _mm256_i64gather_epi32(int const *, __m256i, const int );
_mm256_i64gather_epi64	__m256i __cdecl _mm256_i64gather_epi64(__int64 const *, __m256i, const int );
_mm256_i64gather_pd	__m256d __cdecl _mm256_i64gather_pd(double const *, __m256i, const int );
_mm256_i64gather_ps	__m128 __cdecl _mm256_i64gather_ps(float const *, __m256i, const int );
_mm256_insertf128_pd	__m256d __cdecl _mm256_insertf128_pd(__m256d, __m128d, int);
_mm256_insertf128_ps	__m256 __cdecl _mm256_insertf128_ps(__m256, __m128, int);
_mm256_insertf128_si256	__m256i __cdecl _mm256_insertf128_si256(__m256i, __m128i, int);
_mm256_inserti128_si256	__m256i __cdecl _mm256_inserti128_si256(__m256i, __m128i, const int);
_mm256_lddqu_si256	__m256i __cdecl _mm256_lddqu_si256(__m256i const *);
_mm256_load_pd	__m256d __cdecl _mm256_load_pd(double const *);
_mm256_load_ps	__m256 __cdecl _mm256_load_ps(float const *);
_mm256_load_si256	__m256i __cdecl _mm256_load_si256(__m256i const *);
_mm256_loadu_pd	__m256d __cdecl _mm256_loadu_pd(double const *);
_mm256_loadu_ps	__m256 __cdecl _mm256_loadu_ps(float const *);
_mm256_loadu_si256	__m256i __cdecl _mm256_loadu_si256(__m256i const *);
_mm256_macc_pd	__m256d _mm256_macc_pd(__m256d, __m256d, __m256d);
_mm256_macc_ps	__m256 _mm256_macc_ps(__m256, __m256, __m256);
_mm256_madd_epi16	__m256i __cdecl _mm256_madd_epi16(__m256i, __m256i);
_mm256_maddsub_pd	__m256d _mm256_maddsub_pd(__m256d, __m256d, __m256d);
_mm256_maddsub_ps	__m256 _mm256_maddsub_ps(__m256, __m256, __m256);
_mm256_maddubs_epi16	__m256i __cdecl _mm256_maddubs_epi16(__m256i, __m256i);
_mm256_mask_i32gather_epi32	__m256i __cdecl _mm256_mask_i32gather_epi32(__m256i, int const *, __m256i, __m256i, const int );
_mm256_mask_i32gather_epi64	__m256i __cdecl _mm256_mask_i32gather_epi64(__m256i, __int64 const *, __m128i, __m256i, const int );
_mm256_mask_i32gather_pd	__m256d __cdecl _mm256_mask_i32gather_pd(__m256d, double const *, __m128i, __m256d, const int );
_mm256_mask_i32gather_ps	__m256 __cdecl _mm256_mask_i32gather_ps(__m256, float const *, __m256i, __m256, const int );
_mm256_mask_i64gather_epi32	__m128i __cdecl _mm256_mask_i64gather_epi32(__m128i, int const *, __m256i, __m128i, const int );
_mm256_mask_i64gather_epi64	__m256i __cdecl _mm256_mask_i64gather_epi64(__m256i, __int64 const *, __m256i, __m256i, const int );
_mm256_mask_i64gather_pd	__m256d __cdecl _mm256_mask_i64gather_pd(__m256d, double const *, __m256i, __m256d, const int );
_mm256_mask_i64gather_ps	__m128 __cdecl _mm256_mask_i64gather_ps(__m128, float const *, __m256i, __m128, const int );
_mm256_maskload_epi32	__m256i __cdecl _mm256_maskload_epi32(int const *, __m256i );
_mm256_maskload_epi64	__m256i __cdecl _mm256_maskload_epi64(__int64 const *, __m256i );
_mm256_maskload_pd	__m256d __cdecl _mm256_maskload_pd(double const *, __m256i);
_mm256_maskload_ps	__m256 __cdecl _mm256_maskload_ps(float const *, __m256i);
_mm256_maskstore_epi32	void __cdecl _mm256_maskstore_epi32(int *, __m256i, __m256i );
_mm256_maskstore_epi64	void __cdecl _mm256_maskstore_epi64(__int64 *, __m256i, __m256i );
_mm256_maskstore_pd	void __cdecl _mm256_maskstore_pd(double *, __m256i, __m256d);
_mm256_maskstore_ps	void __cdecl _mm256_maskstore_ps(float *, __m256i, __m256);
_mm256_max_epi16	__m256i __cdecl _mm256_max_epi16(__m256i, __m256i);
_mm256_max_epi32	__m256i __cdecl _mm256_max_epi32(__m256i, __m256i);
_mm256_max_epi8	__m256i __cdecl _mm256_max_epi8(__m256i, __m256i);
_mm256_max_epu16	__m256i __cdecl _mm256_max_epu16(__m256i, __m256i);
_mm256_max_epu32	__m256i __cdecl _mm256_max_epu32(__m256i, __m256i);
_mm256_max_epu8	__m256i __cdecl _mm256_max_epu8(__m256i, __m256i);
_mm256_max_pd	__m256d __cdecl _mm256_max_pd(__m256d, __m256d);
_mm256_max_ps	__m256 __cdecl _mm256_max_ps(__m256, __m256);
_mm256_min_epi16	__m256i __cdecl _mm256_min_epi16(__m256i, __m256i);
_mm256_min_epi32	__m256i __cdecl _mm256_min_epi32(__m256i, __m256i);
_mm256_min_epi8	__m256i __cdecl _mm256_min_epi8(__m256i, __m256i);
_mm256_min_epu16	__m256i __cdecl _mm256_min_epu16(__m256i, __m256i);
_mm256_min_epu32	__m256i __cdecl _mm256_min_epu32(__m256i, __m256i);
_mm256_min_epu8	__m256i __cdecl _mm256_min_epu8(__m256i, __m256i);
_mm256_min_pd	__m256d __cdecl _mm256_min_pd(__m256d, __m256d);
_mm256_min_ps	__m256 __cdecl _mm256_min_ps(__m256, __m256);
_mm256_movedup_pd	__m256d __cdecl _mm256_movedup_pd(__m256d);
_mm256_movehdup_ps	__m256 __cdecl _mm256_movehdup_ps(__m256);
_mm256_moveldup_ps	__m256 __cdecl _mm256_moveldup_ps(__m256);
_mm256_movemask_epi8	int __cdecl _mm256_movemask_epi8(__m256i);
_mm256_movemask_pd	int __cdecl _mm256_movemask_pd(__m256d);
_mm256_movemask_ps	int __cdecl _mm256_movemask_ps(__m256);
_mm256_mpsadbw_epu8	__m256i __cdecl _mm256_mpsadbw_epu8(__m256i, __m256i, const int);
_mm256_msub_pd	__m256d _mm256_msub_pd(__m256d, __m256d, __m256d);
_mm256_msub_ps	__m256 _mm256_msub_ps(__m256, __m256, __m256);
_mm256_msubadd_pd	__m256d _mm256_msubadd_pd(__m256d, __m256d, __m256d);
_mm256_msubadd_ps	__m256 _mm256_msubadd_ps(__m256, __m256, __m256);
_mm256_mul_epi32	__m256i __cdecl _mm256_mul_epi32(__m256i, __m256i);
_mm256_mul_epu32	__m256i __cdecl _mm256_mul_epu32(__m256i, __m256i);
_mm256_mul_pd	__m256d __cdecl _mm256_mul_pd(__m256d, __m256d);
_mm256_mul_ps	__m256 __cdecl _mm256_mul_ps(__m256, __m256);
_mm256_mulhi_epi16	__m256i __cdecl _mm256_mulhi_epi16(__m256i, __m256i);
_mm256_mulhi_epu16	__m256i __cdecl _mm256_mulhi_epu16(__m256i, __m256i);
_mm256_mulhrs_epi16	__m256i __cdecl _mm256_mulhrs_epi16(__m256i, __m256i);
_mm256_mullo_epi16	__m256i __cdecl _mm256_mullo_epi16(__m256i, __m256i);
_mm256_mullo_epi32	__m256i __cdecl _mm256_mullo_epi32(__m256i, __m256i);
_mm256_nmacc_pd	__m256d _mm256_nmacc_pd(__m256d, __m256d, __m256d);
_mm256_nmacc_ps	__m256 _mm256_nmacc_ps(__m256, __m256, __m256);
_mm256_nmsub_pd	__m256d _mm256_nmsub_pd(__m256d, __m256d, __m256d);
_mm256_nmsub_ps	__m256 _mm256_nmsub_ps(__m256, __m256, __m256);
_mm256_or_pd	__m256d __cdecl _mm256_or_pd(__m256d, __m256d);
_mm256_or_ps	__m256 __cdecl _mm256_or_ps(__m256, __m256);
_mm256_or_si256	__m256i __cdecl _mm256_or_si256(__m256i, __m256i);
_mm256_packs_epi16	__m256i __cdecl _mm256_packs_epi16(__m256i, __m256i);
_mm256_packs_epi32	__m256i __cdecl _mm256_packs_epi32(__m256i, __m256i);
_mm256_packus_epi16	__m256i __cdecl _mm256_packus_epi16(__m256i, __m256i);
_mm256_packus_epi32	__m256i __cdecl _mm256_packus_epi32(__m256i, __m256i);
_mm256_permute_pd	__m256d __cdecl _mm256_permute_pd(__m256d, int);
_mm256_permute_ps	__m256 __cdecl _mm256_permute_ps(__m256, int);
_mm256_permute2_pd	__m256d _mm256_permute2_pd(__m256d, __m256d, __m256i, int);
_mm256_permute2_ps	__m256 _mm256_permute2_ps(__m256, __m256, __m256i, int);
_mm256_permute2f128_pd	__m256d __cdecl _mm256_permute2f128_pd(__m256d, __m256d, int);
_mm256_permute2f128_ps	__m256 __cdecl _mm256_permute2f128_ps(__m256, __m256, int);
_mm256_permute2f128_si256	__m256i __cdecl _mm256_permute2f128_si256(__m256i, __m256i, int);
_mm256_permute2x128_si256	__m256i __cdecl _mm256_permute2x128_si256(__m256i, __m256i, const int);
_mm256_permute4x64_epi64	__m256i __cdecl _mm256_permute4x64_epi64(__m256i, const int);
_mm256_permute4x64_pd	__m256d __cdecl _mm256_permute4x64_pd(__m256d, const int);
_mm256_permutevar_pd	__m256d __cdecl _mm256_permutevar_pd(__m256d, __m256i);
_mm256_permutevar_ps	__m256 __cdecl _mm256_permutevar_ps(__m256, __m256i);
_mm256_permutevar8x32_epi32	__m256i __cdecl _mm256_permutevar8x32_epi32(__m256i, __m256i);
_mm256_permutevar8x32_ps	__m256 __cdecl _mm256_permutevar8x32_ps(__m256, __m256i);
_mm256_rcp_ps	__m256 __cdecl _mm256_rcp_ps(__m256);
_mm256_round_pd	__m256d __cdecl _mm256_round_pd(__m256d, int);
_mm256_round_ps	__m256 __cdecl _mm256_round_ps(__m256, int);
_mm256_rsqrt_ps	__m256 __cdecl _mm256_rsqrt_ps(__m256);
_mm256_sad_epu8	__m256i __cdecl _mm256_sad_epu8(__m256i, __m256i);
_mm256_set_epi16	__m256i __cdecl _mm256_set_epi16(short, short, short, short, short, short, short, short, short, short, short, short, short, short, short, short);
_mm256_set_epi32	__m256i __cdecl _mm256_set_epi32(int, int, int, int, int, int, int, int);
_mm256_set_epi64x	__m256i __cdecl _mm256_set_epi64x(__int64, __int64, __int64, __int64);
_mm256_set_epi8	__m256i __cdecl _mm256_set_epi8(char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char);
_mm256_set_pd	__m256d __cdecl _mm256_set_pd(double, double, double, double);
_mm256_set_ps	__m256 __cdecl _mm256_set_ps(float, float, float, float, float, float, float, float);
_mm256_set1_epi16	__m256i __cdecl _mm256_set1_epi16(short);
_mm256_set1_epi32	__m256i __cdecl _mm256_set1_epi32(int);
_mm256_set1_epi64x	__m256i __cdecl _mm256_set1_epi64x(long long);
_mm256_set1_epi8	__m256i __cdecl _mm256_set1_epi8(char);
_mm256_set1_pd	__m256d __cdecl _mm256_set1_pd(double);
_mm256_set1_ps	__m256 __cdecl _mm256_set1_ps(float);
_mm256_setr_epi16	__m256i __cdecl _mm256_setr_epi16(short, short, short, short, short, short, short, short, short, short, short, short, short, short, short, short);
_mm256_setr_epi32	__m256i __cdecl _mm256_setr_epi32(int, int, int, int, int, int, int, int);
_mm256_setr_epi64x	__m256i __cdecl _mm256_setr_epi64x(__int64, __int64, __int64, __int64);
_mm256_setr_epi8	__m256i __cdecl _mm256_setr_epi8(char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char);
_mm256_setr_pd	__m256d __cdecl _mm256_setr_pd(double, double, double, double);
_mm256_setr_ps	__m256 __cdecl _mm256_setr_ps(float, float, float, float, float, float, float, float);
_mm256_setzero_pd	__m256d __cdecl _mm256_setzero_pd(void);
_mm256_setzero_ps	__m256 __cdecl _mm256_setzero_ps(void);
_mm256_setzero_si256	__m256i __cdecl _mm256_setzero_si256(void);
_mm256_shuffle_epi32	__m256i __cdecl _mm256_shuffle_epi32(__m256i, const int);
_mm256_shuffle_epi8	__m256i __cdecl _mm256_shuffle_epi8(__m256i, __m256i);
_mm256_shuffle_pd	__m256d __cdecl _mm256_shuffle_pd(__m256d, __m256d, const int);
_mm256_shuffle_ps	__m256 __cdecl _mm256_shuffle_ps(__m256, __m256, const int);
_mm256_shufflehi_epi16	__m256i __cdecl _mm256_shufflehi_epi16(__m256i, const int);
_mm256_shufflelo_epi16	__m256i __cdecl _mm256_shufflelo_epi16(__m256i, const int);
_mm256_sign_epi16	__m256i __cdecl _mm256_sign_epi16(__m256i, __m256i);
_mm256_sign_epi32	__m256i __cdecl _mm256_sign_epi32(__m256i, __m256i);
_mm256_sign_epi8	__m256i __cdecl _mm256_sign_epi8(__m256i, __m256i);
_mm256_sll_epi16	__m256i __cdecl _mm256_sll_epi16(__m256i, __m128i);
_mm256_sll_epi32	__m256i __cdecl _mm256_sll_epi32(__m256i, __m128i);
_mm256_sll_epi64	__m256i __cdecl _mm256_sll_epi64(__m256i, __m128i);
_mm256_slli_epi16	__m256i __cdecl _mm256_slli_epi16(__m256i, int);
_mm256_slli_epi32	__m256i __cdecl _mm256_slli_epi32(__m256i, int);
_mm256_slli_epi64	__m256i __cdecl _mm256_slli_epi64(__m256i, int);
_mm256_slli_si256	__m256i __cdecl _mm256_slli_si256(__m256i, const int);
_mm256_sllv_epi32	__m256i __cdecl _mm256_sllv_epi32(__m256i, __m256i);
_mm256_sllv_epi64	__m256i __cdecl _mm256_sllv_epi64(__m256i, __m256i);
_mm256_sqrt_pd	__m256d __cdecl _mm256_sqrt_pd(__m256d);
_mm256_sqrt_ps	__m256 __cdecl _mm256_sqrt_ps(__m256);
_mm256_sra_epi16	__m256i __cdecl _mm256_sra_epi16(__m256i, __m128i);
_mm256_sra_epi32	__m256i __cdecl _mm256_sra_epi32(__m256i, __m128i);
_mm256_srai_epi16	__m256i __cdecl _mm256_srai_epi16(__m256i, int);
_mm256_srai_epi32	__m256i __cdecl _mm256_srai_epi32(__m256i, int);
_mm256_srav_epi32	__m256i __cdecl _mm256_srav_epi32(__m256i, __m256i);
_mm256_srl_epi16	__m256i __cdecl _mm256_srl_epi16(__m256i, __m128i);
_mm256_srl_epi32	__m256i __cdecl _mm256_srl_epi32(__m256i, __m128i);
_mm256_srl_epi64	__m256i __cdecl _mm256_srl_epi64(__m256i, __m128i);
_mm256_srli_epi16	__m256i __cdecl _mm256_srli_epi16(__m256i, int);
_mm256_srli_epi32	__m256i __cdecl _mm256_srli_epi32(__m256i, int);
_mm256_srli_epi64	__m256i __cdecl _mm256_srli_epi64(__m256i, int);
_mm256_srli_si256	__m256i __cdecl _mm256_srli_si256(__m256i, const int);
_mm256_srlv_epi32	__m256i __cdecl _mm256_srlv_epi32(__m256i, __m256i);
_mm256_srlv_epi64	__m256i __cdecl _mm256_srlv_epi64(__m256i, __m256i);
_mm256_store_pd	void __cdecl _mm256_store_pd(double *, __m256d);
_mm256_store_ps	void __cdecl _mm256_store_ps(float *, __m256);
_mm256_store_si256	void __cdecl _mm256_store_si256(__m256i *, __m256i);
_mm256_storeu_pd	void __cdecl _mm256_storeu_pd(double *, __m256d);
_mm256_storeu_ps	void __cdecl _mm256_storeu_ps(float *, __m256);
_mm256_storeu_si256	void __cdecl _mm256_storeu_si256(__m256i *, __m256i);
_mm256_stream_load_si256	__m256i __cdecl _mm256_stream_load_si256(__m256i const *);
_mm256_stream_pd	void __cdecl _mm256_stream_pd(double *, __m256d);
_mm256_stream_ps	void __cdecl _mm256_stream_ps(float *, __m256);
_mm256_stream_si256	void __cdecl _mm256_stream_si256(__m256i *, __m256i);
_mm256_sub_epi16	__m256i __cdecl _mm256_sub_epi16(__m256i, __m256i);
_mm256_sub_epi32	__m256i __cdecl _mm256_sub_epi32(__m256i, __m256i);
_mm256_sub_epi64	__m256i __cdecl _mm256_sub_epi64(__m256i, __m256i);
_mm256_sub_epi8	__m256i __cdecl _mm256_sub_epi8(__m256i, __m256i);
_mm256_sub_pd	__m256d __cdecl _mm256_sub_pd(__m256d, __m256d);
_mm256_sub_ps	__m256 __cdecl _mm256_sub_ps(__m256, __m256);
_mm256_subs_epi16	__m256i __cdecl _mm256_subs_epi16(__m256i, __m256i);
_mm256_subs_epi8	__m256i __cdecl _mm256_subs_epi8(__m256i, __m256i);
_mm256_subs_epu16	__m256i __cdecl _mm256_subs_epu16(__m256i, __m256i);
_mm256_subs_epu8	__m256i __cdecl _mm256_subs_epu8(__m256i, __m256i);
_mm256_testc_pd	int __cdecl _mm256_testc_pd(__m256d, __m256d);
_mm256_testc_ps	int __cdecl _mm256_testc_ps(__m256, __m256);
_mm256_testc_si256	int __cdecl _mm256_testc_si256(__m256i, __m256i);
_mm256_testnzc_pd	int __cdecl _mm256_testnzc_pd(__m256d, __m256d);
_mm256_testnzc_ps	int __cdecl _mm256_testnzc_ps(__m256, __m256);
_mm256_testnzc_si256	int __cdecl _mm256_testnzc_si256(__m256i, __m256i);
_mm256_testz_pd	int __cdecl _mm256_testz_pd(__m256d, __m256d);
_mm256_testz_ps	int __cdecl _mm256_testz_ps(__m256, __m256);
_mm256_testz_si256	int __cdecl _mm256_testz_si256(__m256i, __m256i);
_mm256_unpackhi_epi16	__m256i __cdecl _mm256_unpackhi_epi16(__m256i, __m256i);
_mm256_unpackhi_epi32	__m256i __cdecl _mm256_unpackhi_epi32(__m256i, __m256i);
_mm256_unpackhi_epi64	__m256i __cdecl _mm256_unpackhi_epi64(__m256i, __m256i);
_mm256_unpackhi_epi8	__m256i __cdecl _mm256_unpackhi_epi8(__m256i, __m256i);
_mm256_unpackhi_pd	__m256d __cdecl _mm256_unpackhi_pd(__m256d, __m256d);
_mm256_unpackhi_ps	__m256 __cdecl _mm256_unpackhi_ps(__m256, __m256);
_mm256_unpacklo_epi16	__m256i __cdecl _mm256_unpacklo_epi16(__m256i, __m256i);
_mm256_unpacklo_epi32	__m256i __cdecl _mm256_unpacklo_epi32(__m256i, __m256i);
_mm256_unpacklo_epi64	__m256i __cdecl _mm256_unpacklo_epi64(__m256i, __m256i);
_mm256_unpacklo_epi8	__m256i __cdecl _mm256_unpacklo_epi8(__m256i, __m256i);
_mm256_unpacklo_pd	__m256d __cdecl _mm256_unpacklo_pd(__m256d, __m256d);
_mm256_unpacklo_ps	__m256 __cdecl _mm256_unpacklo_ps(__m256, __m256);
_mm256_xor_pd	__m256d __cdecl _mm256_xor_pd(__m256d, __m256d);
_mm256_xor_ps	__m256 __cdecl _mm256_xor_ps(__m256, __m256);
_mm256_xor_si256	__m256i __cdecl _mm256_xor_si256(__m256i, __m256i);
_mm256_zeroall	void __cdecl _mm256_zeroall(void);
_mm256_zeroupper	void __cdecl _mm256_zeroupper(void);
_mul128	__int64 _mul128(__int64 _Multiplier, __int64 _Multiplicand, __int64 * _HighProduct);
_mulx_u32	unsigned int _mulx_u32(unsigned int, unsigned int, unsigned int * );
_mulx_u64	unsigned __int64 _mulx_u64(unsigned __int64, unsigned __int64, unsigned __int64 * );
_pdep_u32	unsigned int _pdep_u32(unsigned int, unsigned int );
_pdep_u64	unsigned __int64 _pdep_u64(unsigned __int64, unsigned __int64 );
_pext_u32	unsigned int _pext_u32(unsigned int, unsigned int );
_pext_u64	unsigned __int64 _pext_u64(unsigned __int64, unsigned __int64 );
_rdrand16_step	int __cdecl _rdrand16_step(unsigned short *);
_rdrand32_step	int __cdecl _rdrand32_step(unsigned int *);
_rdrand64_step	int __cdecl _rdrand64_step(unsigned __int64 *);
_ReadBarrier	void _ReadBarrier(void);
_readfsbase_u32	unsigned int __cdecl _readfsbase_u32();
_readfsbase_u64	unsigned __int64 __cdecl _readfsbase_u64();
_readgsbase_u32	unsigned int __cdecl _readgsbase_u32();
_readgsbase_u64	unsigned __int64 __cdecl _readgsbase_u64();
_ReadWriteBarrier	void _ReadWriteBarrier(void);
_ReturnAddress	void * _ReturnAddress(void);
_rorx_u32	unsigned int _rorx_u32(unsigned int, const unsigned int );
_rorx_u64	unsigned __int64 _rorx_u64(unsigned __int64, const unsigned int );
_rotl	unsigned int __cdecl _rotl(unsigned int _Value, int _Shift);
_rotl16	unsigned short _rotl16(unsigned short _Value, unsigned char _Shift);
_rotl64	unsigned __int64 __cdecl _rotl64(unsigned __int64 _Value, int _Shift);
_rotl8	unsigned char _rotl8(unsigned char _Value, unsigned char _Shift);
_rotr	unsigned int __cdecl _rotr(unsigned int _Value, int _Shift);
_rotr16	unsigned short _rotr16(unsigned short _Value, unsigned char _Shift);
_rotr64	unsigned __int64 __cdecl _rotr64(unsigned __int64 _Value, int _Shift);
_rotr8	unsigned char _rotr8(unsigned char _Value, unsigned char _Shift);
_sarx_i32	int _sarx_i32(int, unsigned int );
_sarx_i64	__int64 _sarx_i64(__int64, unsigned int );
_setjmp	int __cdecl _setjmp(jmp_buf);
_setjmpex	int __cdecl _setjmpex(jmp_buf);
_shlx_u32	unsigned int _shlx_u32(unsigned int, unsigned int );
_shlx_u64	unsigned __int64 _shlx_u64(unsigned __int64, unsigned int );
_shrx_u32	unsigned int _shrx_u32(unsigned int, unsigned int );
_shrx_u64	unsigned __int64 _shrx_u64(unsigned __int64, unsigned int );
_Store_HLERelease	void _Store_HLERelease(long volatile *,long);
_Store64_HLERelease	void _Store64_HLERelease(__int64 volatile *,__int64);
_StorePointer_HLERelease	void _StorePointer_HLERelease(void * volatile ,void );
_t1mskc_u32	unsigned int _t1mskc_u32(unsigned int);
_t1mskc_u64	unsigned __int64 _t1mskc_u64(unsigned __int64);
_tzcnt_u32	unsigned int _tzcnt_u32(unsigned int);
_tzcnt_u32	unsigned int _tzcnt_u32(unsigned int);
_tzcnt_u64	unsigned __int64 _tzcnt_u64(unsigned __int64);
_tzcnt_u64	unsigned __int64 _tzcnt_u64(unsigned __int64);
_tzmsk_u32	unsigned int _tzmsk_u32(unsigned int);
_tzmsk_u64	unsigned __int64 _tzmsk_u64(unsigned __int64);
_umul128	unsigned __int64 _umul128(unsigned __int64 _Multiplier, unsigned __int64 _Multiplicand, unsigned __int64 * _HighProduct);
_WriteBarrier	void _WriteBarrier(void);
_writefsbase_u32	void __cdecl _writefsbase_u32(unsigned int);
_writefsbase_u64	void __cdecl _writefsbase_u64(unsigned __int64);
_writegsbase_u32	void __cdecl _writegsbase_u32(unsigned int);
_writegsbase_u64	void __cdecl _writegsbase_u64(unsigned __int64);
_xabort	void _xabort(const unsigned int imm);
_xbegin	unsigned __int32 _xbegin(void);
_xend	void _xend(void);
_xgetbv	unsigned __int64 __cdecl _xgetbv(unsigned int);
_xrstor	void __cdecl _xrstor(void const *, unsigned __int64);
_xrstor64	void __cdecl _xrstor64(void const *, unsigned __int64);
_xsave	void __cdecl _xsave(void *, unsigned __int64);
_xsave64	void __cdecl _xsave64(void *, unsigned __int64);
_xsaveopt	void __cdecl _xsaveopt(void *, unsigned __int64);
_xsaveopt64	void __cdecl _xsaveopt64(void *, unsigned __int64);
_xsetbv	void __cdecl _xsetbv(unsigned int, unsigned __int64);
_xtest	unsigned char _xtest(void);

Share via

x64 (amd64) Intrinsics List

See Also

Reference

Additional resources