Liste der intrinsischen Funktionen für x64 (amd64)

In diesem Dokument werden systeminterne Elemente aufgeführt, die der Microsoft C++-Compiler unterstützt, wenn x64 (auch als amd64 bezeichnet) gezielt ist.

Weitere Informationen über einzelne systeminterne Funktionen finden Sie in diesen Ressourcen, je nach Prozessor, den Sie als Ziel bestimmen.

x64 systeminterne

Die folgende Tabelle enthält die systeminternen Funktionen, die auf x64-Prozessoren verfügbar sind. In der Spalte "Technologie" ist die erforderliche Anweisungssatzunterstützung aufgeführt. Verwenden Sie das __cpuid systeminterne System, um die Unterstützung von Anweisungssatz zur Laufzeit zu ermitteln. Wenn sich zwei Einträge in einer Zeile befinden, stellen sie verschiedene Einstiegspunkte für dieselbe systeminterne Funktion dar. [Makro] gibt an, dass der Prototyp ein Makro ist. Der für den Funktionsprototyp erforderliche Header ist in der Spalte „Header“ aufgeführt. Die intrin.h Kopfzeile enthält sowohl die ammintrin.h Einfachheit als auch immintrin.h die Einfachheit.

Systeminterner Name Technologie Header Funktionsprototyp
_addcarry_u16 intrin.h unsigned char _addcarry_u16(unsigned char, unsigned short, unsigned short, unsigned short *);
_addcarry_u32 intrin.h unsigned char _addcarry_u32(unsigned char, unsigned int, unsigned int, unsigned int *);
_addcarry_u64 intrin.h unsigned char _addcarry_u64(unsigned char, unsigned __int64, unsigned __int64, unsigned __int64 *);
_addcarry_u8 intrin.h unsigned char _addcarry_u8(unsigned char, unsigned char, unsigned char, unsigned char *);
_addcarryx_u32 ADX immintrin.h unsigned char _addcarryx_u32(unsigned char, unsigned int, unsigned int, unsigned int *);
_addcarryx_u64 ADX immintrin.h unsigned char _addcarryx_u64(unsigned char, unsigned __int64, unsigned __int64, unsigned __int64 *);
__addgsbyte intrin.h void __addgsbyte(unsigned long, unsigned char);
__addgsdword intrin.h void __addgsdword(unsigned long, unsigned int);
__addgsqword intrin.h void __addgsqword(unsigned long, unsigned __int64);
__addgsword intrin.h void __addgsword(unsigned long, unsigned short);
_AddressOfReturnAddress intrin.h void * _AddressOfReturnAddress(void);
_andn_u32 BMI ammintrin.h unsigned int _andn_u32(unsigned int, unsigned int);
_andn_u64 BMI ammintrin.h unsigned __int64 _andn_u64(unsigned __int64, unsigned __int64);
_bextr_u32 BMI ammintrin.h, immintrin.h unsigned int _bextr_u32(unsigned int, unsigned int, unsigned int);
_bextr_u64 BMI ammintrin.h, immintrin.h unsigned __int64 _bextr_u64(unsigned __int64, unsigned int, unsigned int);
_bextri_u32 ABM ammintrin.h unsigned int _bextri_u32(unsigned int, unsigned int);
_bextri_u64 ABM ammintrin.h unsigned __int64 _bextri_u64(unsigned __int64, unsigned int);
_BitScanForward intrin.h unsigned char _BitScanForward(unsigned long*, unsigned long);
_BitScanForward64 intrin.h unsigned char _BitScanForward64(unsigned long*, unsigned __int64);
_BitScanReverse intrin.h unsigned char _BitScanReverse(unsigned long*, unsigned long);
_BitScanReverse64 intrin.h unsigned char _BitScanReverse64(unsigned long*, unsigned __int64);
_bittest intrin.h unsigned char _bittest(long const *, long);
_bittest64 intrin.h unsigned char _bittest64(__int64 const *, __int64);
_bittestandcomplement intrin.h unsigned char _bittestandcomplement(long *, long);
_bittestandcomplement64 intrin.h unsigned char _bittestandcomplement64(__int64 *, __int64);
_bittestandreset intrin.h unsigned char _bittestandreset(long *, long);
_bittestandreset64 intrin.h unsigned char _bittestandreset64(__int64 *, __int64);
_bittestandset intrin.h unsigned char _bittestandset(long *, long);
_bittestandset64 intrin.h unsigned char _bittestandset64(__int64 *, __int64);
_blcfill_u32 ABM ammintrin.h unsigned int _blcfill_u32(unsigned int);
_blcfill_u64 ABM ammintrin.h unsigned __int64 _blcfill_u64(unsigned __int64);
_blci_u32 ABM ammintrin.h unsigned int _blci_u32(unsigned int);
_blci_u64 ABM ammintrin.h unsigned __int64 _blci_u64(unsigned __int64);
_blcic_u32 ABM ammintrin.h unsigned int _blcic_u32(unsigned int);
_blcic_u64 ABM ammintrin.h unsigned __int64 _blcic_u64(unsigned __int64);
_blcmsk_u32 ABM ammintrin.h unsigned int _blcmsk_u32(unsigned int);
_blcmsk_u64 ABM ammintrin.h unsigned __int64 _blcmsk_u64(unsigned __int64);
_blcs_u32 ABM ammintrin.h unsigned int _blcs_u32(unsigned int);
_blcs_u64 ABM ammintrin.h unsigned __int64 _blcs_u64(unsigned __int64);
_blsfill_u32 ABM ammintrin.h unsigned int _blsfill_u32(unsigned int);
_blsfill_u64 ABM ammintrin.h unsigned __int64 _blsfill_u64(unsigned __int64);
_blsi_u32 BMI ammintrin.h, immintrin.h unsigned int _blsi_u32(unsigned int);
_blsi_u64 BMI ammintrin.h, immintrin.h unsigned __int64 _blsi_u64(unsigned __int64);
_blsic_u32 ABM ammintrin.h unsigned int _blsic_u32(unsigned int);
_blsic_u64 ABM ammintrin.h unsigned __int64 _blsic_u64(unsigned __int64);
_blsmsk_u32 BMI ammintrin.h, immintrin.h unsigned int _blsmsk_u32(unsigned int);
_blsmsk_u64 BMI ammintrin.h, immintrin.h unsigned __int64 _blsmsk_u64(unsigned __int64);
_blsr_u32 BMI ammintrin.h, immintrin.h unsigned int _blsr_u32(unsigned int);
_blsr_u64 BMI ammintrin.h, immintrin.h unsigned __int64 _blsr_u64(unsigned __int64);
_bzhi_u32 BMI immintrin.h unsigned int _bzhi_u32(unsigned int, unsigned int);
_bzhi_u64 BMI immintrin.h unsigned __int64 _bzhi_u64(unsigned __int64, unsigned int);
_castf32_u32 immintrin.h unsigned __int32 _castf32_u32 (float);
_castf64_u64 immintrin.h unsigned __int64 _castf64_u64 (double);
_castu32_f32 immintrin.h float _castu32_f32 (unsigned __int32);
_castu64_f64 immintrin.h double _castu64_f64 (unsigned __int64 a);
__check_isa_support immintrin.h bool __check_isa_support(unsigned int, unsigned int);
__check_arch_support immintrin.h bool __check_arch_support(unsigned int, unsigned int);
_clac SMAP intrin.h void _clac(void);
__cpuid intrin.h void __cpuid(int *, int);
__cpuidex intrin.h void __cpuidex(int *, int, int);
__debugbreak intrin.h void __debugbreak(void);
_disable intrin.h void _disable(void);
_div128 intrin.h __int64 _div128(__int64, __int64, __int64, __int64 *);
_div64 intrin.h int _div64(__int64, int, int*);
__emul intrin.h __int64 [pascal/cdecl] __emul(int, int);
__emulu intrin.h unsigned __int64 [pascal/cdecl]__emulu(unsigned int, unsigned int);
_enable intrin.h void _enable(void);
__fastfail intrin.h void __fastfail(unsigned int);
__faststorefence intrin.h void __faststorefence(void);
_fxrstor FXSR immintrin.h void _fxrstor(void const*);
_fxrstor64 FXSR immintrin.h void _fxrstor64(void const*);
_fxsave FXSR immintrin.h void _fxsave(void*);
_fxsave64 FXSR immintrin.h void _fxsave64(void*);
__getcallerseflags intrin.h (unsigned int __getcallerseflags());
__halt intrin.h void __halt(void);
__inbyte intrin.h unsigned char __inbyte(unsigned short);
__inbytestring intrin.h void __inbytestring(unsigned short, unsigned char *, unsigned long);
__incgsbyte intrin.h void __incgsbyte(unsigned long);
__incgsdword intrin.h void __incgsdword(unsigned long);
__incgsqword intrin.h void __incgsqword(unsigned long);
__incgsword intrin.h void __incgsword(unsigned long);
__indword intrin.h unsigned long __indword(unsigned short);
__indwordstring intrin.h void __indwordstring(unsigned short, unsigned long *, unsigned long);
__int2c intrin.h void __int2c(void);
_InterlockedAnd intrin.h long _InterlockedAnd(long volatile *, long);
_InterlockedAnd_HLEAcquire HLE immintrin.h long _InterlockedAnd_HLEAcquire(long volatile *, long);
_InterlockedAnd_HLERelease HLE immintrin.h long _InterlockedAnd_HLERelease(long volatile *, long);
_InterlockedAnd_np intrin.h long _InterlockedAnd_np(long *, long);
_InterlockedAnd16 intrin.h short _InterlockedAnd16(short volatile *, short);
_InterlockedAnd16_np intrin.h short _InterlockedAnd16_np(short *, short);
_InterlockedAnd64 intrin.h __int64 _InterlockedAnd64(__int64 volatile *, __int64);
_InterlockedAnd64_HLEAcquire HLE immintrin.h __int64 _InterlockedAnd64_HLEAcquire(__int64 volatile *, __int64);
_InterlockedAnd64_HLERelease HLE immintrin.h __int64 _InterlockedAnd64_HLERelease(__int64 volatile *, __int64);
_InterlockedAnd64_np intrin.h __int64 _InterlockedAnd64_np(__int64 *, __int64);
_InterlockedAnd8 intrin.h char _InterlockedAnd8(char volatile *, char);
_InterlockedAnd8_np intrin.h char _InterlockedAnd8_np(char *, char);
_interlockedbittestandreset intrin.h unsigned char _interlockedbittestandreset(long *, long);
_interlockedbittestandreset_HLEAcquire HLE immintrin.h unsigned char _interlockedbittestandreset_HLEAcquire(long *, long);
_interlockedbittestandreset_HLERelease HLE immintrin.h unsigned char _interlockedbittestandreset_HLERelease(long *, long);
_interlockedbittestandreset64 intrin.h unsigned char _interlockedbittestandreset64(__int64 *, __int64);
_interlockedbittestandreset64_HLEAcquire HLE immintrin.h unsigned char _interlockedbittestandreset64_HLEAcquire(__int64 *, __int64);
_interlockedbittestandreset64_HLERelease HLE immintrin.h unsigned char _interlockedbittestandreset64_HLERelease(__int64 *, __int64);
_interlockedbittestandset intrin.h unsigned char _interlockedbittestandset(long *, long);
_interlockedbittestandset_HLEAcquire HLE immintrin.h unsigned char _interlockedbittestandset_HLEAcquire(long *, long);
_interlockedbittestandset_HLERelease HLE immintrin.h unsigned char _interlockedbittestandset_HLERelease(long *, long);
_interlockedbittestandset64 intrin.h unsigned char _interlockedbittestandset64(__int64 *, __int64);
_interlockedbittestandset64_HLEAcquire HLE immintrin.h unsigned char _interlockedbittestandset64_HLEAcquire(__int64 *, __int64);
_interlockedbittestandset64_HLERelease HLE immintrin.h unsigned char _interlockedbittestandset64_HLERelease(__int64 *, __int64);
_InterlockedCompareExchange intrin.h long _InterlockedCompareExchange (long volatile *, long, long);
_InterlockedCompareExchange_HLEAcquire HLE immintrin.h long _InterlockedCompareExchange_HLEAcquire(long volatile *, long, long);
_InterlockedCompareExchange_HLERelease HLE immintrin.h long _InterlockedCompareExchange_HLERelease(long volatile *, long, long);
_InterlockedCompareExchange_np intrin.h long _InterlockedCompareExchange_np (long *, long, long);
_InterlockedCompareExchange128 intrin.h unsigned char _InterlockedCompareExchange128(__int64 volatile *, __int64, __int64, __int64*);
_InterlockedCompareExchange128_np intrin.h unsigned char _InterlockedCompareExchange128(__int64 volatile *, __int64, __int64, __int64*);
_InterlockedCompareExchange16 intrin.h short _InterlockedCompareExchange16(short volatile *, short, short);
_InterlockedCompareExchange16_np intrin.h short _InterlockedCompareExchange16_np(short volatile *, short, short);
_InterlockedCompareExchange64 intrin.h __int64 _InterlockedCompareExchange64(__int64 volatile *, __int64, __int64);
_InterlockedCompareExchange64_HLEAcquire HLE immintrin.h __int64 _InterlockedCompareExchange64_HLEAcquire(__int64 volatile *, __int64, __int64);
_InterlockedCompareExchange64_HLERelease HLE immintrin.h __int64 _InterlockedCompareExchange64_HLERelease(__int64 volatile *, __int64, __int64);
_InterlockedCompareExchange64_np intrin.h __int64 _InterlockedCompareExchange64_np(__int64 *, __int64, __int64);
_InterlockedCompareExchange8 intrin.h char _InterlockedCompareExchange8(char volatile *, char, char);
_InterlockedCompareExchangePointer intrin.h void *_InterlockedCompareExchangePointer (void *volatile *, void *, void *);
_InterlockedCompareExchangePointer_HLEAcquire HLE immintrin.h void *_InterlockedCompareExchangePointer_HLEAcquire(void *volatile *, void *, void *);
_InterlockedCompareExchangePointer_HLERelease HLE immintrin.h void *_InterlockedCompareExchangePointer_HLERelease(void *volatile *, void *, void *);
_InterlockedCompareExchangePointer_np intrin.h void *_InterlockedCompareExchangePointer_np(void **, void *, void *);
_InterlockedDecrement intrin.h long _InterlockedDecrement(long volatile *);
_InterlockedDecrement16 intrin.h short _InterlockedDecrement16(short volatile *);
_InterlockedDecrement64 intrin.h __int64 _InterlockedDecrement64(__int64 volatile *);
_InterlockedExchange intrin.h long _InterlockedExchange(long volatile *, long);
_InterlockedExchange_HLEAcquire HLE immintrin.h long _InterlockedExchange_HLEAcquire(long volatile *, long);
_InterlockedExchange_HLERelease HLE immintrin.h long _InterlockedExchange_HLERelease(long volatile *, long);
_InterlockedExchange16 intrin.h short _InterlockedExchange16(short volatile *, short);
_InterlockedExchange64 intrin.h __int64 _InterlockedExchange64(__int64 volatile *, __int64);
_InterlockedExchange64_HLEAcquire HLE immintrin.h __int64 _InterlockedExchange64_HLEAcquire(__int64 volatile *, __int64);
_InterlockedExchange64_HLERelease HLE immintrin.h __int64 _InterlockedExchange64_HLERelease(__int64 volatile *, __int64);
_InterlockedExchange8 intrin.h char _InterlockedExchange8(char volatile *, char);
_InterlockedExchangeAdd intrin.h long _InterlockedExchangeAdd(long volatile *, long);
_InterlockedExchangeAdd_HLEAcquire HLE immintrin.h long _InterlockedExchangeAdd_HLEAcquire(long volatile *, long);
_InterlockedExchangeAdd_HLERelease HLE immintrin.h long _InterlockedExchangeAdd_HLERelease(long volatile *, long);
_InterlockedExchangeAdd16 intrin.h short _InterlockedExchangeAdd16(short volatile *, short);
_InterlockedExchangeAdd64 intrin.h __int64 _InterlockedExchangeAdd64(__int64 volatile *, __int64);
_InterlockedExchangeAdd64_HLEAcquire HLE immintrin.h __int64 _InterlockedExchangeAdd64_HLEAcquire(__int64 volatile *, __int64);
_InterlockedExchangeAdd64_HLERelease HLE immintrin.h __int64 _InterlockedExchangeAdd64_HLERelease(__int64 volatile *, __int64);
_InterlockedExchangeAdd8 intrin.h char _InterlockedExchangeAdd8(char volatile *, char);
_InterlockedExchangePointer intrin.h void * _InterlockedExchangePointer(void *volatile *, void *);
_InterlockedExchangePointer_HLEAcquire HLE immintrin.h void * _InterlockedExchangePointer_HLEAcquire(void *volatile *, void *);
_InterlockedExchangePointer_HLERelease HLE immintrin.h void * _InterlockedExchangePointer_HLERelease(void *volatile *, void *);
_InterlockedIncrement intrin.h long _InterlockedIncrement(long volatile *);
_InterlockedIncrement16 intrin.h short _InterlockedIncrement16(short volatile *);
_InterlockedIncrement64 intrin.h __int64 _InterlockedIncrement64(__int64 volatile *);
_InterlockedOr intrin.h long _InterlockedOr(long volatile *, long);
_InterlockedOr_HLEAcquire HLE immintrin.h long _InterlockedOr_HLEAcquire(long volatile *, long);
_InterlockedOr_HLERelease HLE immintrin.h long _InterlockedOr_HLERelease(long volatile *, long);
_InterlockedOr_np intrin.h long _InterlockedOr_np(long *, long);
_InterlockedOr16 intrin.h short _InterlockedOr16(short volatile *, short);
_InterlockedOr16_np intrin.h short _InterlockedOr16_np(short *, short);
_InterlockedOr64 intrin.h __int64 _InterlockedOr64(__int64 volatile *, __int64);
_InterlockedOr64_HLEAcquire HLE immintrin.h __int64 _InterlockedOr64_HLEAcquire(__int64 volatile *, __int64);
_InterlockedOr64_HLERelease HLE immintrin.h __int64 _InterlockedOr64_HLERelease(__int64 volatile *, __int64);
_InterlockedOr64_np intrin.h __int64 _InterlockedOr64_np(__int64 *, __int64);
_InterlockedOr8 intrin.h char _InterlockedOr8(char volatile *, char);
_InterlockedOr8_np intrin.h char _InterlockedOr8_np(char *, char);
_InterlockedXor intrin.h long _InterlockedXor(long volatile *, long);
_InterlockedXor_HLEAcquire HLE immintrin.h long _InterlockedXor_HLEAcquire(long volatile *, long);
_InterlockedXor_HLERelease HLE immintrin.h long _InterlockedXor_HLERelease(long volatile *, long);
_InterlockedXor_np intrin.h long _InterlockedXor_np(long *, long);
_InterlockedXor16 intrin.h short _InterlockedXor16(short volatile *, short);
_InterlockedXor16_np intrin.h short _InterlockedXor16_np(short *, short);
_InterlockedXor64 intrin.h __int64 _InterlockedXor64(__int64 volatile *, __int64);
_InterlockedXor64_HLEAcquire HLE immintrin.h __int64 _InterlockedXor64_HLEAcquire(__int64 volatile *, __int64);
_InterlockedXor64_HLERelease HLE immintrin.h __int64 _InterlockedXor64_HLERelease(__int64 volatile *, __int64);
_InterlockedXor64_np intrin.h __int64 _InterlockedXor64_np(__int64 *, __int64);
_InterlockedXor8 intrin.h char _InterlockedXor8(char volatile *, char);
_InterlockedXor8_np intrin.h char _InterlockedXor8_np(char *, char);
__invlpg intrin.h void __invlpg(void*);
_invpcid INVPCID immintrin.h void _invpcid(unsigned int, void *);
__inword intrin.h unsigned short __inword(unsigned short);
__inwordstring intrin.h void __inwordstring(unsigned short, unsigned short *, unsigned long);
_lgdt intrin.h void _lgdt(void*);
__lidt intrin.h void __lidt(void*);
__ll_lshift intrin.h unsigned __int64 [pascal/cdecl] __ll_lshift(unsigned __int64, int);
__ll_rshift intrin.h __int64 [pascal/cdecl] __ll_rshift(__int64, int);
__llwpcb LWP ammintrin.h void __llwpcb(void *);
_loadbe_i16 MOVBE immintrin.h short _loadbe_i16(void const*); [Makro]
_loadbe_i32 MOVBE immintrin.h int _loadbe_i32(void const*); [Makro]
_loadbe_i64 MOVBE immintrin.h __int64 _loadbe_i64(void const*); [Makro]
_load_be_u16 MOVBE immintrin.h unsigned short _load_be_u16(void const*); [Makro]
_load_be_u32 MOVBE immintrin.h unsigned int _load_be_u32(void const*); [Makro]
_load_be_u64 MOVBE immintrin.h unsigned __int64 _load_be_u64(void const*); [Makro]
__lwpins32 LWP ammintrin.h unsigned char __lwpins32(unsigned int, unsigned int, unsigned int);
__lwpins64 LWP ammintrin.h unsigned char __lwpins64(unsigned __int64, unsigned int, unsigned int);
__lwpval32 LWP ammintrin.h void __lwpval32(unsigned int, unsigned int, unsigned int);
__lwpval64 LWP ammintrin.h void __lwpval64(unsigned __int64, unsigned int, unsigned int);
__lzcnt LZCNT intrin.h unsigned int __lzcnt(unsigned int);
_lzcnt_u32 BMI ammintrin.h, immintrin.h unsigned int _lzcnt_u32(unsigned int);
_lzcnt_u64 BMI ammintrin.h, immintrin.h unsigned __int64 _lzcnt_u64(unsigned __int64);
__lzcnt16 LZCNT intrin.h unsigned short __lzcnt16(unsigned short);
__lzcnt64 LZCNT intrin.h unsigned __int64 __lzcnt64(unsigned __int64);
_m_prefetch 3DNOW intrin.h void _m_prefetch(void*);
_m_prefetchw 3DNOW intrin.h void _m_prefetchw(void*);
_mm_abs_epi16 SSSE3 intrin.h __m128i _mm_abs_epi16(__m128i);
_mm_abs_epi32 SSSE3 intrin.h __m128i _mm_abs_epi32(__m128i);
_mm_abs_epi8 SSSE3 intrin.h __m128i _mm_abs_epi8(__m128i);
_mm_add_epi16 SSE2 intrin.h __m128i _mm_add_epi16(__m128i, __m128i);
_mm_add_epi32 SSE2 intrin.h __m128i _mm_add_epi32(__m128i, __m128i);
_mm_add_epi64 SSE2 intrin.h __m128i _mm_add_epi64(__m128i, __m128i);
_mm_add_epi8 SSE2 intrin.h __m128i _mm_add_epi8(__m128i, __m128i);
_mm_add_pd SSE2 intrin.h __m128d _mm_add_pd(__m128d, __m128d);
_mm_add_ps SSE intrin.h __m128 _mm_add_ps(__m128, __m128);
_mm_add_sd SSE2 intrin.h __m128d _mm_add_sd(__m128d, __m128d);
_mm_add_ss SSE intrin.h __m128 _mm_add_ss(__m128, __m128);
_mm_adds_epi16 SSE2 intrin.h __m128i _mm_adds_epi16(__m128i, __m128i);
_mm_adds_epi8 SSE2 intrin.h __m128i _mm_adds_epi8(__m128i, __m128i);
_mm_adds_epu16 SSE2 intrin.h __m128i _mm_adds_epu16(__m128i, __m128i);
_mm_adds_epu8 SSE2 intrin.h __m128i _mm_adds_epu8(__m128i, __m128i);
_mm_addsub_pd SSE3 intrin.h __m128d _mm_addsub_pd(__m128d, __m128d);
_mm_addsub_ps SSE3 intrin.h __m128 _mm_addsub_ps(__m128, __m128);
_mm_aesdec_si128 AESNI immintrin.h __m128i _mm_aesdec_si128(__m128i, __m128i);
_mm_aesdeclast_si128 AESNI immintrin.h __m128i _mm_aesdeclast_si128(__m128i, __m128i);
_mm_aesenc_si128 AESNI immintrin.h __m128i _mm_aesenc_si128(__m128i, __m128i);
_mm_aesenclast_si128 AESNI immintrin.h __m128i _mm_aesenclast_si128(__m128i, __m128i);
_mm_aesimc_si128 AESNI immintrin.h __m128i _mm_aesimc_si128 (__m128i);
_mm_aeskeygenassist_si128 AESNI immintrin.h __m128i _mm_aeskeygenassist_si128 (__m128i, const int);
_mm_alignr_epi8 SSSE3 intrin.h __m128i _mm_alignr_epi8(__m128i, __m128i, int);
_mm_and_pd SSE2 intrin.h __m128d _mm_and_pd(__m128d, __m128d);
_mm_and_ps SSE intrin.h __m128 _mm_and_ps(__m128, __m128);
_mm_and_si128 SSE2 intrin.h __m128i _mm_and_si128(__m128i, __m128i);
_mm_andnot_pd SSE2 intrin.h __m128d _mm_andnot_pd(__m128d, __m128d);
_mm_andnot_ps SSE intrin.h __m128 _mm_andnot_ps(__m128, __m128);
_mm_andnot_si128 SSE2 intrin.h __m128i _mm_andnot_si128(__m128i, __m128i);
_mm_avg_epu16 SSE2 intrin.h __m128i _mm_avg_epu16(__m128i, __m128i);
_mm_avg_epu8 SSE2 intrin.h __m128i _mm_avg_epu8(__m128i, __m128i);
_mm_blend_epi16 SSE41 intrin.h __m128i _mm_blend_epi16 (__m128i, __m128i, const int);
_mm_blend_epi32 AVX2 immintrin.h __m128i _mm_blend_epi32(__m128i, __m128i, const int);
_mm_blend_pd SSE41 intrin.h __m128d _mm_blend_pd (__m128d, __m128d, const int);
_mm_blend_ps SSE41 intrin.h __m128 _mm_blend_ps (__m128, __m128, const int);
_mm_blendv_epi8 SSE41 intrin.h __m128i _mm_blendv_epi8 (__m128i, __m128i, __m128i);
_mm_blendv_pd SSE41 intrin.h __m128d _mm_blendv_pd(__m128d, __m128d, __m128d);
_mm_blendv_ps SSE41 intrin.h __m128 _mm_blendv_ps(__m128, __m128, __m128);
_mm_broadcast_ss AVX immintrin.h __m128 _mm_broadcast_ss(float const *);
_mm_broadcastb_epi8 AVX2 immintrin.h __m128i _mm_broadcastb_epi8(__m128i);
_mm_broadcastd_epi32 AVX2 immintrin.h __m128i _mm_broadcastd_epi32(__m128i);
_mm_broadcastq_epi64 AVX2 immintrin.h __m128i _mm_broadcastq_epi64(__m128i);
_mm_broadcastsd_pd AVX2 immintrin.h __m128d _mm_broadcastsd_pd(__m128d);
_mm_broadcastss_ps AVX2 immintrin.h __m128 _mm_broadcastss_ps(__m128);
_mm_broadcastw_epi16 AVX2 immintrin.h __m128i _mm_broadcastw_epi16(__m128i);
_mm_castpd_ps SSSE3 intrin.h __m128 _mm_castpd_ps(__m128d);
_mm_castpd_si128 SSSE3 intrin.h __m128i _mm_castpd_si128(__m128d);
_mm_castps_pd SSSE3 intrin.h __m128d _mm_castps_pd(__m128);
_mm_castps_si128 SSSE3 intrin.h __m128i _mm_castps_si128(__m128);
_mm_castsi128_pd SSSE3 intrin.h __m128d _mm_castsi128_pd(__m128i);
_mm_castsi128_ps SSSE3 intrin.h __m128 _mm_castsi128_ps(__m128i);
_mm_clflush SSE2 intrin.h void _mm_clflush(void const *);
_mm_clmulepi64_si128 PCLMULQDQ immintrin.h __m128i _mm_clmulepi64_si128 (__m128i, __m128i, const int);
_mm_cmov_si128 XOP ammintrin.h __m128i _mm_cmov_si128(__m128i, __m128i, __m128i);
_mm_cmp_pd AVX immintrin.h __m128d _mm_cmp_pd(__m128d, __m128d, const int);
_mm_cmp_ps AVX immintrin.h __m128 _mm_cmp_ps(__m128, __m128, const int);
_mm_cmp_sd AVX immintrin.h __m128d _mm_cmp_sd(__m128d, __m128d, const int);
_mm_cmp_ss AVX immintrin.h __m128 _mm_cmp_ss(__m128, __m128, const int);
_mm_cmpeq_epi16 SSE2 intrin.h __m128i _mm_cmpeq_epi16(__m128i, __m128i);
_mm_cmpeq_epi32 SSE2 intrin.h __m128i _mm_cmpeq_epi32(__m128i, __m128i);
_mm_cmpeq_epi64 SSE41 intrin.h __m128i _mm_cmpeq_epi64(__m128i, __m128i);
_mm_cmpeq_epi8 SSE2 intrin.h __m128i _mm_cmpeq_epi8(__m128i, __m128i);
_mm_cmpeq_pd SSE2 intrin.h __m128d _mm_cmpeq_pd(__m128d, __m128d);
_mm_cmpeq_ps SSE intrin.h __m128 _mm_cmpeq_ps(__m128, __m128);
_mm_cmpeq_sd SSE2 intrin.h __m128d _mm_cmpeq_sd(__m128d, __m128d);
_mm_cmpeq_ss SSE intrin.h __m128 _mm_cmpeq_ss(__m128, __m128);
_mm_cmpestra SSE42 intrin.h int _mm_cmpestra(__m128i, int, __m128i, int, const int);
_mm_cmpestrc SSE42 intrin.h int _mm_cmpestrc(__m128i, int, __m128i, int, const int);
_mm_cmpestri SSE42 intrin.h int _mm_cmpestri(__m128i, int, __m128i, int, const int);
_mm_cmpestrm SSE42 intrin.h __m128i _mm_cmpestrm(__m128i, int, __m128i, int, const int);
_mm_cmpestro SSE42 intrin.h int _mm_cmpestro(__m128i, int, __m128i, int, const int);
_mm_cmpestrs SSE42 intrin.h int _mm_cmpestrs(__m128i, int, __m128i, int, const int);
_mm_cmpestrz SSE42 intrin.h int _mm_cmpestrz(__m128i, int, __m128i, int, const int);
_mm_cmpge_pd SSE2 intrin.h __m128d _mm_cmpge_pd(__m128d, __m128d);
_mm_cmpge_ps SSE intrin.h __m128 _mm_cmpge_ps(__m128, __m128);
_mm_cmpge_sd SSE2 intrin.h __m128d _mm_cmpge_sd(__m128d, __m128d);
_mm_cmpge_ss SSE intrin.h __m128 _mm_cmpge_ss(__m128, __m128);
_mm_cmpgt_epi16 SSE2 intrin.h __m128i _mm_cmpgt_epi16(__m128i, __m128i);
_mm_cmpgt_epi32 SSE2 intrin.h __m128i _mm_cmpgt_epi32(__m128i, __m128i);
_mm_cmpgt_epi64 SSE42 intrin.h __m128i _mm_cmpgt_epi64(__m128i, __m128i);
_mm_cmpgt_epi8 SSE2 intrin.h __m128i _mm_cmpgt_epi8(__m128i, __m128i);
_mm_cmpgt_pd SSE2 intrin.h __m128d _mm_cmpgt_pd(__m128d, __m128d);
_mm_cmpgt_ps SSE intrin.h __m128 _mm_cmpgt_ps(__m128, __m128);
_mm_cmpgt_sd SSE2 intrin.h __m128d _mm_cmpgt_sd(__m128d, __m128d);
_mm_cmpgt_ss SSE intrin.h __m128 _mm_cmpgt_ss(__m128, __m128);
_mm_cmpistra SSE42 intrin.h int _mm_cmpistra(__m128i, __m128i, const int);
_mm_cmpistrc SSE42 intrin.h int _mm_cmpistrc(__m128i, __m128i, const int);
_mm_cmpistri SSE42 intrin.h int _mm_cmpistri(__m128i, __m128i, const int);
_mm_cmpistrm SSE42 intrin.h __m128i _mm_cmpistrm(__m128i, __m128i, const int);
_mm_cmpistro SSE42 intrin.h int _mm_cmpistro(__m128i, __m128i, const int);
_mm_cmpistrs SSE42 intrin.h int _mm_cmpistrs(__m128i, __m128i, const int);
_mm_cmpistrz SSE42 intrin.h int _mm_cmpistrz(__m128i, __m128i, const int);
_mm_cmple_pd SSE2 intrin.h __m128d _mm_cmple_pd(__m128d, __m128d);
_mm_cmple_ps SSE intrin.h __m128 _mm_cmple_ps(__m128, __m128);
_mm_cmple_sd SSE2 intrin.h __m128d _mm_cmple_sd(__m128d, __m128d);
_mm_cmple_ss SSE intrin.h __m128 _mm_cmple_ss(__m128, __m128);
_mm_cmplt_epi16 SSE2 intrin.h __m128i _mm_cmplt_epi16(__m128i, __m128i);
_mm_cmplt_epi32 SSE2 intrin.h __m128i _mm_cmplt_epi32(__m128i, __m128i);
_mm_cmplt_epi8 SSE2 intrin.h __m128i _mm_cmplt_epi8(__m128i, __m128i);
_mm_cmplt_pd SSE2 intrin.h __m128d _mm_cmplt_pd(__m128d, __m128d);
_mm_cmplt_ps SSE intrin.h __m128 _mm_cmplt_ps(__m128, __m128);
_mm_cmplt_sd SSE2 intrin.h __m128d _mm_cmplt_sd(__m128d, __m128d);
_mm_cmplt_ss SSE intrin.h __m128 _mm_cmplt_ss(__m128, __m128);
_mm_cmpneq_pd SSE2 intrin.h __m128d _mm_cmpneq_pd(__m128d, __m128d);
_mm_cmpneq_ps SSE intrin.h __m128 _mm_cmpneq_ps(__m128, __m128);
_mm_cmpneq_sd SSE2 intrin.h __m128d _mm_cmpneq_sd(__m128d, __m128d);
_mm_cmpneq_ss SSE intrin.h __m128 _mm_cmpneq_ss(__m128, __m128);
_mm_cmpnge_pd SSE2 intrin.h __m128d _mm_cmpnge_pd(__m128d, __m128d);
_mm_cmpnge_ps SSE intrin.h __m128 _mm_cmpnge_ps(__m128, __m128);
_mm_cmpnge_sd SSE2 intrin.h __m128d _mm_cmpnge_sd(__m128d, __m128d);
_mm_cmpnge_ss SSE intrin.h __m128 _mm_cmpnge_ss(__m128, __m128);
_mm_cmpngt_pd SSE2 intrin.h __m128d _mm_cmpngt_pd(__m128d, __m128d);
_mm_cmpngt_ps SSE intrin.h __m128 _mm_cmpngt_ps(__m128, __m128);
_mm_cmpngt_sd SSE2 intrin.h __m128d _mm_cmpngt_sd(__m128d, __m128d);
_mm_cmpngt_ss SSE intrin.h __m128 _mm_cmpngt_ss(__m128, __m128);
_mm_cmpnle_pd SSE2 intrin.h __m128d _mm_cmpnle_pd(__m128d, __m128d);
_mm_cmpnle_ps SSE intrin.h __m128 _mm_cmpnle_ps(__m128, __m128);
_mm_cmpnle_sd SSE2 intrin.h __m128d _mm_cmpnle_sd(__m128d, __m128d);
_mm_cmpnle_ss SSE intrin.h __m128 _mm_cmpnle_ss(__m128, __m128);
_mm_cmpnlt_pd SSE2 intrin.h __m128d _mm_cmpnlt_pd(__m128d, __m128d);
_mm_cmpnlt_ps SSE intrin.h __m128 _mm_cmpnlt_ps(__m128, __m128);
_mm_cmpnlt_sd SSE2 intrin.h __m128d _mm_cmpnlt_sd(__m128d, __m128d);
_mm_cmpnlt_ss SSE intrin.h __m128 _mm_cmpnlt_ss(__m128, __m128);
_mm_cmpord_pd SSE2 intrin.h __m128d _mm_cmpord_pd(__m128d, __m128d);
_mm_cmpord_ps SSE intrin.h __m128 _mm_cmpord_ps(__m128, __m128);
_mm_cmpord_sd SSE2 intrin.h __m128d _mm_cmpord_sd(__m128d, __m128d);
_mm_cmpord_ss SSE intrin.h __m128 _mm_cmpord_ss(__m128, __m128);
_mm_cmpunord_pd SSE2 intrin.h __m128d _mm_cmpunord_pd(__m128d, __m128d);
_mm_cmpunord_ps SSE intrin.h __m128 _mm_cmpunord_ps(__m128, __m128);
_mm_cmpunord_sd SSE2 intrin.h __m128d _mm_cmpunord_sd(__m128d, __m128d);
_mm_cmpunord_ss SSE intrin.h __m128 _mm_cmpunord_ss(__m128, __m128);
_mm_com_epi16 XOP ammintrin.h __m128i _mm_com_epi16(__m128i, __m128i, int);
_mm_com_epi32 XOP ammintrin.h __m128i _mm_com_epi32(__m128i, __m128i, int);
_mm_com_epi64 XOP ammintrin.h __m128i _mm_com_epi32(__m128i, __m128i, int);
_mm_com_epi8 XOP ammintrin.h __m128i _mm_com_epi8(__m128i, __m128i, int);
_mm_com_epu16 XOP ammintrin.h __m128i _mm_com_epu16(__m128i, __m128i, int);
_mm_com_epu32 XOP ammintrin.h __m128i _mm_com_epu32(__m128i, __m128i, int);
_mm_com_epu64 XOP ammintrin.h __m128i _mm_com_epu32(__m128i, __m128i, int);
_mm_com_epu8 XOP ammintrin.h __m128i _mm_com_epu8(__m128i, __m128i, int);
_mm_comieq_sd SSE2 intrin.h int _mm_comieq_sd(__m128d, __m128d);
_mm_comieq_ss SSE intrin.h int _mm_comieq_ss(__m128, __m128);
_mm_comige_sd SSE2 intrin.h int _mm_comige_sd(__m128d, __m128d);
_mm_comige_ss SSE intrin.h int _mm_comige_ss(__m128, __m128);
_mm_comigt_sd SSE2 intrin.h int _mm_comigt_sd(__m128d, __m128d);
_mm_comigt_ss SSE intrin.h int _mm_comigt_ss(__m128, __m128);
_mm_comile_sd SSE2 intrin.h int _mm_comile_sd(__m128d, __m128d);
_mm_comile_ss SSE intrin.h int _mm_comile_ss(__m128, __m128);
_mm_comilt_sd SSE2 intrin.h int _mm_comilt_sd(__m128d, __m128d);
_mm_comilt_ss SSE intrin.h int _mm_comilt_ss(__m128, __m128);
_mm_comineq_sd SSE2 intrin.h int _mm_comineq_sd(__m128d, __m128d);
_mm_comineq_ss SSE intrin.h int _mm_comineq_ss(__m128, __m128);
_mm_crc32_u16 SSE42 intrin.h unsigned int _mm_crc32_u16(unsigned int, unsigned short);
_mm_crc32_u32 SSE42 intrin.h unsigned int _mm_crc32_u32(unsigned int, unsigned int);
_mm_crc32_u64 SSE42 intrin.h unsigned __int64 _mm_crc32_u64(unsigned __int64, unsigned __int64);
_mm_crc32_u8 SSE42 intrin.h unsigned int _mm_crc32_u8(unsigned int, unsigned char);
_mm_cvt_si2ss SSE intrin.h __m128 _mm_cvt_si2ss(__m128, int);
_mm_cvt_ss2si SSE intrin.h int _mm_cvt_ss2si(__m128);
_mm_cvtepi16_epi32 SSE41 intrin.h __m128i _mm_cvtepi16_epi32(__m128i);
_mm_cvtepi16_epi64 SSE41 intrin.h __m128i _mm_cvtepi16_epi64(__m128i);
_mm_cvtepi32_epi64 SSE41 intrin.h __m128i _mm_cvtepi32_epi64(__m128i);
_mm_cvtepi32_pd SSE2 intrin.h __m128d _mm_cvtepi32_pd(__m128i);
_mm_cvtepi32_ps SSE2 intrin.h __m128 _mm_cvtepi32_ps(__m128i);
_mm_cvtepi8_epi16 SSE41 intrin.h __m128i _mm_cvtepi8_epi16 (__m128i);
_mm_cvtepi8_epi32 SSE41 intrin.h __m128i _mm_cvtepi8_epi32 (__m128i);
_mm_cvtepi8_epi64 SSE41 intrin.h __m128i _mm_cvtepi8_epi64 (__m128i);
_mm_cvtepu16_epi32 SSE41 intrin.h __m128i _mm_cvtepu16_epi32(__m128i);
_mm_cvtepu16_epi64 SSE41 intrin.h __m128i _mm_cvtepu16_epi64(__m128i);
_mm_cvtepu32_epi64 SSE41 intrin.h __m128i _mm_cvtepu32_epi64(__m128i);
_mm_cvtepu8_epi16 SSE41 intrin.h __m128i _mm_cvtepu8_epi16 (__m128i);
_mm_cvtepu8_epi32 SSE41 intrin.h __m128i _mm_cvtepu8_epi32 (__m128i);
_mm_cvtepu8_epi64 SSE41 intrin.h __m128i _mm_cvtepu8_epi64 (__m128i);
_mm_cvtpd_epi32 SSE2 intrin.h __m128i _mm_cvtpd_epi32(__m128d);
_mm_cvtpd_ps SSE2 intrin.h __m128 _mm_cvtpd_ps(__m128d);
_mm_cvtph_ps F16C immintrin.h __m128 _mm_cvtph_ps(__m128i);
_mm_cvtps_epi32 SSE2 intrin.h __m128i _mm_cvtps_epi32(__m128);
_mm_cvtps_pd SSE2 intrin.h __m128d _mm_cvtps_pd(__m128);
_mm_cvtps_ph F16C immintrin.h __m128i _mm_cvtps_ph(__m128, const int);
_mm_cvtsd_f64 SSSE3 intrin.h double _mm_cvtsd_f64(__m128d);
_mm_cvtsd_si32 SSE2 intrin.h int _mm_cvtsd_si32(__m128d);
_mm_cvtsd_si64 SSE2 intrin.h __int64 _mm_cvtsd_si64(__m128d);
_mm_cvtsd_si64x SSE2 intrin.h __int64 _mm_cvtsd_si64x(__m128d);
_mm_cvtsd_ss SSE2 intrin.h __m128 _mm_cvtsd_ss(__m128, __m128d);
_mm_cvtsi128_si32 SSE2 intrin.h int _mm_cvtsi128_si32(__m128i);
_mm_cvtsi128_si64 SSE2 intrin.h __int64 _mm_cvtsi128_si64(__m128i);
_mm_cvtsi128_si64x SSE2 intrin.h __int64 _mm_cvtsi128_si64x(__m128i);
_mm_cvtsi32_sd SSE2 intrin.h __m128d _mm_cvtsi32_sd(__m128d, int);
_mm_cvtsi32_si128 SSE2 intrin.h __m128i _mm_cvtsi32_si128(int);
_mm_cvtsi64_sd SSE2 intrin.h __m128d _mm_cvtsi64_sd(__m128d, __int64);
_mm_cvtsi64_si128 SSE2 intrin.h __m128i _mm_cvtsi64_si128(__int64);
_mm_cvtsi64_ss SSE intrin.h __m128 _mm_cvtsi64_ss(__m128, __int64);
_mm_cvtsi64x_sd SSE2 intrin.h __m128d _mm_cvtsi64x_sd(__m128d, __int64);
_mm_cvtsi64x_si128 SSE2 intrin.h __m128i _mm_cvtsi64x_si128(__int64);
_mm_cvtsi64x_ss SSE2 intrin.h __m128 _mm_cvtsi64x_ss(__m128, __int64);
_mm_cvtss_f32 SSSE3 intrin.h float _mm_cvtss_f32(__m128);
_mm_cvtss_sd SSE2 intrin.h __m128d _mm_cvtss_sd(__m128d, __m128);
_mm_cvtss_si64 SSE intrin.h __int64 _mm_cvtss_si64(__m128);
_mm_cvtss_si64x SSE2 intrin.h __int64 _mm_cvtss_si64x(__m128);
_mm_cvtt_ss2si SSE intrin.h int _mm_cvtt_ss2si(__m128);
_mm_cvttpd_epi32 SSE2 intrin.h __m128i _mm_cvttpd_epi32(__m128d);
_mm_cvttps_epi32 SSE2 intrin.h __m128i _mm_cvttps_epi32(__m128);
_mm_cvttsd_si32 SSE2 intrin.h int _mm_cvttsd_si32(__m128d);
_mm_cvttsd_si64 SSE2 intrin.h __int64 _mm_cvttsd_si64(__m128d);
_mm_cvttsd_si64x SSE2 intrin.h __int64 _mm_cvttsd_si64x(__m128d);
_mm_cvttss_si64 SSE2 intrin.h __int64 _mm_cvttss_si64(__m128);
_mm_cvttss_si64x SSE2 intrin.h __int64 _mm_cvttss_si64x(__m128);
_mm_div_pd SSE2 intrin.h __m128d _mm_div_pd(__m128d, __m128d);
_mm_div_ps SSE intrin.h __m128 _mm_div_ps(__m128, __m128);
_mm_div_sd SSE2 intrin.h __m128d _mm_div_sd(__m128d, __m128d);
_mm_div_ss SSE intrin.h __m128 _mm_div_ss(__m128, __m128);
_mm_dp_pd SSE41 intrin.h __m128d _mm_dp_pd(__m128d, __m128d, const int);
_mm_dp_ps SSE41 intrin.h __m128 _mm_dp_ps(__m128, __m128, const int);
_mm_extract_epi16 SSE2 intrin.h int _mm_extract_epi16(__m128i, int);
_mm_extract_epi32 SSE41 intrin.h int _mm_extract_epi32(__m128i, const int);
_mm_extract_epi64 SSE41 intrin.h __int64 _mm_extract_epi64(__m128i, const int);
_mm_extract_epi8 SSE41 intrin.h int _mm_extract_epi8 (__m128i, const int);
_mm_extract_ps SSE41 intrin.h int _mm_extract_ps(__m128, const int);
_mm_extract_si64 SSE4a intrin.h __m128i _mm_extract_si64(__m128i, __m128i);
_mm_extracti_si64 SSE4a intrin.h __m128i _mm_extracti_si64(__m128i, int, int);
_mm_fmadd_pd FMA immintrin.h __m128d _mm_fmadd_pd (__m128d, __m128d, __m128d);
_mm_fmadd_ps FMA immintrin.h __m128 _mm_fmadd_ps (__m128, __m128, __m128);
_mm_fmadd_sd FMA immintrin.h __m128d _mm_fmadd_sd (__m128d, __m128d, __m128d);
_mm_fmadd_ss FMA immintrin.h __m128 _mm_fmadd_ss (__m128, __m128, __m128);
_mm_fmaddsub_pd FMA immintrin.h __m128d _mm_fmaddsub_pd (__m128d, __m128d, __m128d);
_mm_fmaddsub_ps FMA immintrin.h __m128 _mm_fmaddsub_ps (__m128, __m128, __m128);
_mm_fmsub_pd FMA immintrin.h __m128d _mm_fmsub_pd (__m128d, __m128d, __m128d);
_mm_fmsub_ps FMA immintrin.h __m128 _mm_fmsub_ps (__m128, __m128, __m128);
_mm_fmsub_sd FMA immintrin.h __m128d _mm_fmsub_sd (__m128d, __m128d, __m128d);
_mm_fmsub_ss FMA immintrin.h __m128 _mm_fmsub_ss (__m128, __m128, __m128);
_mm_fmsubadd_pd FMA immintrin.h __m128d _mm_fmsubadd_pd (__m128d, __m128d, __m128d);
_mm_fmsubadd_ps FMA immintrin.h __m128 _mm_fmsubadd_ps (__m128, __m128, __m128);
_mm_fnmadd_pd FMA immintrin.h __m128d _mm_fnmadd_pd (__m128d, __m128d, __m128d);
_mm_fnmadd_ps FMA immintrin.h __m128 _mm_fnmadd_ps (__m128, __m128, __m128);
_mm_fnmadd_sd FMA immintrin.h __m128d _mm_fnmadd_sd (__m128d, __m128d, __m128d);
_mm_fnmadd_ss FMA immintrin.h __m128 _mm_fnmadd_ss (__m128, __m128, __m128);
_mm_fnmsub_pd FMA immintrin.h __m128d _mm_fnmsub_pd (__m128d, __m128d, __m128d);
_mm_fnmsub_ps FMA immintrin.h __m128 _mm_fnmsub_ps (__m128, __m128, __m128);
_mm_fnmsub_sd FMA immintrin.h __m128d _mm_fnmsub_sd (__m128d, __m128d, __m128d);
_mm_fnmsub_ss FMA immintrin.h __m128 _mm_fnmsub_ss (__m128, __m128, __m128);
_mm_frcz_pd XOP ammintrin.h __m128d _mm_frcz_pd(__m128d);
_mm_frcz_ps XOP ammintrin.h __m128 _mm_frcz_ps(__m128);
_mm_frcz_sd XOP ammintrin.h __m128d _mm_frcz_sd(__m128d, __m128d);
_mm_frcz_ss XOP ammintrin.h __m128 _mm_frcz_ss(__m128, __m128);
_mm_getcsr SSE intrin.h unsigned int _mm_getcsr(void);
_mm_hadd_epi16 SSSE3 intrin.h __m128i _mm_hadd_epi16(__m128i, __m128i);
_mm_hadd_epi32 SSSE3 intrin.h __m128i _mm_hadd_epi32(__m128i, __m128i);
_mm_hadd_pd SSE3 intrin.h __m128d _mm_hadd_pd(__m128d, __m128d);
_mm_hadd_ps SSE3 intrin.h __m128 _mm_hadd_ps(__m128, __m128);
_mm_haddd_epi16 XOP ammintrin.h __m128i _mm_haddd_epi16(__m128i);
_mm_haddd_epi8 XOP ammintrin.h __m128i _mm_haddd_epi8(__m128i);
_mm_haddd_epu16 XOP ammintrin.h __m128i _mm_haddd_epu16(__m128i);
_mm_haddd_epu8 XOP ammintrin.h __m128i _mm_haddd_epu8(__m128i);
_mm_haddq_epi16 XOP ammintrin.h __m128i _mm_haddq_epi16(__m128i);
_mm_haddq_epi32 XOP ammintrin.h __m128i _mm_haddq_epi32(__m128i);
_mm_haddq_epi8 XOP ammintrin.h __m128i _mm_haddq_epi8(__m128i);
_mm_haddq_epu16 XOP ammintrin.h __m128i _mm_haddq_epu16(__m128i);
_mm_haddq_epu32 XOP ammintrin.h __m128i _mm_haddq_epu32(__m128i);
_mm_haddq_epu8 XOP ammintrin.h __m128i _mm_haddq_epu8(__m128i);
_mm_hadds_epi16 SSSE3 intrin.h __m128i _mm_hadds_epi16(__m128i, __m128i);
_mm_haddw_epi8 XOP ammintrin.h __m128i _mm_haddw_epi8(__m128i);
_mm_haddw_epu8 XOP ammintrin.h __m128i _mm_haddw_epu8(__m128i);
_mm_hsub_epi16 SSSE3 intrin.h __m128i _mm_hsub_epi16(__m128i, __m128i);
_mm_hsub_epi32 SSSE3 intrin.h __m128i _mm_hsub_epi32(__m128i, __m128i);
_mm_hsub_pd SSE3 intrin.h __m128d _mm_hsub_pd(__m128d, __m128d);
_mm_hsub_ps SSE3 intrin.h __m128 _mm_hsub_ps(__m128, __m128);
_mm_hsubd_epi16 XOP ammintrin.h __m128i _mm_hsubd_epi16(__m128i);
_mm_hsubq_epi32 XOP ammintrin.h __m128i _mm_hsubq_epi32(__m128i);
_mm_hsubs_epi16 SSSE3 intrin.h __m128i _mm_hsubs_epi16(__m128i, __m128i);
_mm_hsubw_epi8 XOP ammintrin.h __m128i _mm_hsubw_epi8(__m128i);
_mm_i32gather_epi32 AVX2 immintrin.h __m128i _mm_i32gather_epi32(int const *, __m128i, const int);
_mm_i32gather_epi64 AVX2 immintrin.h __m128i _mm_i32gather_epi64(__int64 const *, __m128i, const int);
_mm_i32gather_pd AVX2 immintrin.h __m128d _mm_i32gather_pd(double const *, __m128i, const int);
_mm_i32gather_ps AVX2 immintrin.h __m128 _mm_i32gather_ps(float const *, __m128i, const int);
_mm_i64gather_epi32 AVX2 immintrin.h __m128i _mm_i64gather_epi32(int const *, __m128i, const int);
_mm_i64gather_epi64 AVX2 immintrin.h __m128i _mm_i64gather_epi64(__int64 const *, __m128i, const int);
_mm_i64gather_pd AVX2 immintrin.h __m128d _mm_i64gather_pd(double const *, __m128i, const int);
_mm_i64gather_ps AVX2 immintrin.h __m128 _mm_i64gather_ps(float const *, __m128i, const int);
_mm_insert_epi16 SSE2 intrin.h __m128i _mm_insert_epi16(__m128i, int, int);
_mm_insert_epi32 SSE41 intrin.h __m128i _mm_insert_epi32(__m128i, int, const int);
_mm_insert_epi64 SSE41 intrin.h __m128i _mm_insert_epi64(__m128i, __int64, const int);
_mm_insert_epi8 SSE41 intrin.h __m128i _mm_insert_epi8 (__m128i, int, const int);
_mm_insert_ps SSE41 intrin.h __m128 _mm_insert_ps(__m128, __m128, const int);
_mm_insert_si64 SSE4a intrin.h __m128i _mm_insert_si64(__m128i, __m128i);
_mm_inserti_si64 SSE4a intrin.h __m128i _mm_inserti_si64(__m128i, __m128i, int, int);
_mm_lddqu_si128 SSE3 intrin.h __m128i _mm_lddqu_si128(__m128i const*);
_mm_lfence SSE2 intrin.h void _mm_lfence(void);
_mm_load_pd SSE2 intrin.h __m128d _mm_load_pd(double*);
_mm_load_ps SSE intrin.h __m128 _mm_load_ps(float*);
_mm_load_ps1 SSE intrin.h __m128 _mm_load_ps1(float*);
_mm_load_sd SSE2 intrin.h __m128d _mm_load_sd(double*);
_mm_load_si128 SSE2 intrin.h __m128i _mm_load_si128(__m128i*);
_mm_load_ss SSE intrin.h __m128 _mm_load_ss(float*);
_mm_load1_pd SSE2 intrin.h __m128d _mm_load1_pd(double*);
_mm_loaddup_pd SSE3 intrin.h __m128d _mm_loaddup_pd(double const*);
_mm_loadh_pd SSE2 intrin.h __m128d _mm_loadh_pd(__m128d, double*);
_mm_loadh_pi SSE intrin.h __m128 _mm_loadh_pi(__m128, __m64*);
_mm_loadl_epi64 SSE2 intrin.h __m128i _mm_loadl_epi64(__m128i*);
_mm_loadl_pd SSE2 intrin.h __m128d _mm_loadl_pd(__m128d, double*);
_mm_loadl_pi SSE intrin.h __m128 _mm_loadl_pi(__m128, __m64*);
_mm_loadr_pd SSE2 intrin.h __m128d _mm_loadr_pd(double*);
_mm_loadr_ps SSE intrin.h __m128 _mm_loadr_ps(float*);
_mm_loadu_pd SSE2 intrin.h __m128d _mm_loadu_pd(double*);
_mm_loadu_ps SSE intrin.h __m128 _mm_loadu_ps(float*);
_mm_loadu_si128 SSE2 intrin.h __m128i _mm_loadu_si128(__m128i*);
_mm_macc_epi16 XOP ammintrin.h __m128i _mm_macc_epi16(__m128i, __m128i, __m128i);
_mm_macc_epi32 XOP ammintrin.h __m128i _mm_macc_epi32(__m128i, __m128i, __m128i);
_mm_macc_pd FMA4 ammintrin.h __m128d _mm_macc_pd(__m128d, __m128d, __m128d);
_mm_macc_ps FMA4 ammintrin.h __m128 _mm_macc_ps(__m128, __m128, __m128);
_mm_macc_sd FMA4 ammintrin.h __m128d _mm_macc_sd(__m128d, __m128d, __m128d);
_mm_macc_ss FMA4 ammintrin.h __m128 _mm_macc_ss(__m128, __m128, __m128);
_mm_maccd_epi16 XOP ammintrin.h __m128i _mm_maccd_epi16(__m128i, __m128i, __m128i);
_mm_macchi_epi32 XOP ammintrin.h __m128i _mm_macchi_epi32(__m128i, __m128i, __m128i);
_mm_macclo_epi32 XOP ammintrin.h __m128i _mm_macclo_epi32(__m128i, __m128i, __m128i);
_mm_maccs_epi16 XOP ammintrin.h __m128i _mm_maccs_epi16(__m128i, __m128i, __m128i);
_mm_maccs_epi32 XOP ammintrin.h __m128i _mm_maccs_epi32(__m128i, __m128i, __m128i);
_mm_maccsd_epi16 XOP ammintrin.h __m128i _mm_maccsd_epi16(__m128i, __m128i, __m128i);
_mm_maccshi_epi32 XOP ammintrin.h __m128i _mm_maccshi_epi32(__m128i, __m128i, __m128i);
_mm_maccslo_epi32 XOP ammintrin.h __m128i _mm_maccslo_epi32(__m128i, __m128i, __m128i);
_mm_madd_epi16 SSE2 intrin.h __m128i _mm_madd_epi16(__m128i, __m128i);
_mm_maddd_epi16 XOP ammintrin.h __m128i _mm_maddd_epi16(__m128i, __m128i, __m128i);
_mm_maddsd_epi16 XOP ammintrin.h __m128i _mm_maddsd_epi16(__m128i, __m128i, __m128i);
_mm_maddsub_pd FMA4 ammintrin.h __m128d _mm_maddsub_pd(__m128d, __m128d, __m128d);
_mm_maddsub_ps FMA4 ammintrin.h __m128 _mm_maddsub_ps(__m128, __m128, __m128);
_mm_maddubs_epi16 SSSE3 intrin.h __m128i _mm_maddubs_epi16(__m128i, __m128i);
_mm_mask_i32gather_epi32 AVX2 immintrin.h __m128i _mm_mask_i32gather_epi32(__m128i, int const *, __m128i, __m128i, const int);
_mm_mask_i32gather_epi64 AVX2 immintrin.h __m128i _mm_mask_i32gather_epi64(__m128i, __int64 const *, __m128i, __m128i, const int);
_mm_mask_i32gather_pd AVX2 immintrin.h __m128d _mm_mask_i32gather_pd(__m128d, double const *, __m128i, __m128d, const int);
_mm_mask_i32gather_ps AVX2 immintrin.h __m128 _mm_mask_i32gather_ps(__m128, float const *, __m128i, __m128, const int);
_mm_mask_i64gather_epi32 AVX2 immintrin.h __m128i _mm_mask_i64gather_epi32(__m128i, int const *, __m128i, __m128i, const int);
_mm_mask_i64gather_epi64 AVX2 immintrin.h __m128i _mm_mask_i64gather_epi64(__m128i, __int64 const *, __m128i, __m128i, const int);
_mm_mask_i64gather_pd AVX2 immintrin.h __m128d _mm_mask_i64gather_pd(__m128d, double const *, __m128i, __m128d, const int);
_mm_mask_i64gather_ps AVX2 immintrin.h __m128 _mm_mask_i64gather_ps(__m128, float const *, __m128i, __m128, const int);
_mm_maskload_epi32 AVX2 immintrin.h __m128i _mm_maskload_epi32(int const *, __m128i);
_mm_maskload_epi64 AVX2 immintrin.h __m128i _mm_maskload_epi64(__int64 const *, __m128i);
_mm_maskload_pd AVX immintrin.h __m128d _mm_maskload_pd(double const *, __m128i);
_mm_maskload_ps AVX immintrin.h __m128 _mm_maskload_ps(float const *, __m128i);
_mm_maskmoveu_si128 SSE2 intrin.h void _mm_maskmoveu_si128(__m128i, __m128i, char*);
_mm_maskstore_epi32 AVX2 immintrin.h void _mm_maskstore_epi32(int *, __m128i, __m128i);
_mm_maskstore_epi64 AVX2 immintrin.h void _mm_maskstore_epi64(__int64 *, __m128i, __m128i);
_mm_maskstore_pd AVX immintrin.h void _mm_maskstore_pd(double *, __m128i, __m128d);
_mm_maskstore_ps AVX immintrin.h void _mm_maskstore_ps(float *, __m128i, __m128);
_mm_max_epi16 SSE2 intrin.h __m128i _mm_max_epi16(__m128i, __m128i);
_mm_max_epi32 SSE41 intrin.h __m128i _mm_max_epi32(__m128i, __m128i);
_mm_max_epi8 SSE41 intrin.h __m128i _mm_max_epi8 (__m128i, __m128i);
_mm_max_epu16 SSE41 intrin.h __m128i _mm_max_epu16(__m128i, __m128i);
_mm_max_epu32 SSE41 intrin.h __m128i _mm_max_epu32(__m128i, __m128i);
_mm_max_epu8 SSE2 intrin.h __m128i _mm_max_epu8(__m128i, __m128i);
_mm_max_pd SSE2 intrin.h __m128d _mm_max_pd(__m128d, __m128d);
_mm_max_ps SSE intrin.h __m128 _mm_max_ps(__m128, __m128);
_mm_max_sd SSE2 intrin.h __m128d _mm_max_sd(__m128d, __m128d);
_mm_max_ss SSE intrin.h __m128 _mm_max_ss(__m128, __m128);
_mm_mfence SSE2 intrin.h void _mm_mfence(void);
_mm_min_epi16 SSE2 intrin.h __m128i _mm_min_epi16(__m128i, __m128i);
_mm_min_epi32 SSE41 intrin.h __m128i _mm_min_epi32(__m128i, __m128i);
_mm_min_epi8 SSE41 intrin.h __m128i _mm_min_epi8 (__m128i, __m128i);
_mm_min_epu16 SSE41 intrin.h __m128i _mm_min_epu16(__m128i, __m128i);
_mm_min_epu32 SSE41 intrin.h __m128i _mm_min_epu32(__m128i, __m128i);
_mm_min_epu8 SSE2 intrin.h __m128i _mm_min_epu8(__m128i, __m128i);
_mm_min_pd SSE2 intrin.h __m128d _mm_min_pd(__m128d, __m128d);
_mm_min_ps SSE intrin.h __m128 _mm_min_ps(__m128, __m128);
_mm_min_sd SSE2 intrin.h __m128d _mm_min_sd(__m128d, __m128d);
_mm_min_ss SSE intrin.h __m128 _mm_min_ss(__m128, __m128);
_mm_minpos_epu16 SSE41 intrin.h __m128i _mm_minpos_epu16(__m128i);
_mm_monitor SSE3 intrin.h void _mm_monitor(void const*, unsigned int, unsigned int);
_mm_move_epi64 SSE2 intrin.h __m128i _mm_move_epi64(__m128i);
_mm_move_sd SSE2 intrin.h __m128d _mm_move_sd(__m128d, __m128d);
_mm_move_ss SSE intrin.h __m128 _mm_move_ss(__m128, __m128);
_mm_movedup_pd SSE3 intrin.h __m128d _mm_movedup_pd(__m128d);
_mm_movehdup_ps SSE3 intrin.h __m128 _mm_movehdup_ps(__m128);
_mm_movehl_ps SSE intrin.h __m128 _mm_movehl_ps(__m128, __m128);
_mm_moveldup_ps SSE3 intrin.h __m128 _mm_moveldup_ps(__m128);
_mm_movelh_ps SSE intrin.h __m128 _mm_movelh_ps(__m128, __m128);
_mm_movemask_epi8 SSE2 intrin.h int _mm_movemask_epi8(__m128i);
_mm_movemask_pd SSE2 intrin.h int _mm_movemask_pd(__m128d);
_mm_movemask_ps SSE intrin.h int _mm_movemask_ps(__m128);
_mm_mpsadbw_epu8 SSE41 intrin.h __m128i _mm_mpsadbw_epu8(__m128i, __m128i, const int);
_mm_msub_pd FMA4 ammintrin.h __m128d _mm_msub_pd(__m128d, __m128d, __m128d);
_mm_msub_ps FMA4 ammintrin.h __m128 _mm_msub_ps(__m128, __m128, __m128);
_mm_msub_sd FMA4 ammintrin.h __m128d _mm_msub_sd(__m128d, __m128d, __m128d);
_mm_msub_ss FMA4 ammintrin.h __m128 _mm_msub_ss(__m128, __m128, __m128);
_mm_msubadd_pd FMA4 ammintrin.h __m128d _mm_msubadd_pd(__m128d, __m128d, __m128d);
_mm_msubadd_ps FMA4 ammintrin.h __m128 _mm_msubadd_ps(__m128, __m128, __m128);
_mm_mul_epi32 SSE41 intrin.h __m128i _mm_mul_epi32(__m128i, __m128i);
_mm_mul_epu32 SSE2 intrin.h __m128i _mm_mul_epu32(__m128i, __m128i);
_mm_mul_pd SSE2 intrin.h __m128d _mm_mul_pd(__m128d, __m128d);
_mm_mul_ps SSE intrin.h __m128 _mm_mul_ps(__m128, __m128);
_mm_mul_sd SSE2 intrin.h __m128d _mm_mul_sd(__m128d, __m128d);
_mm_mul_ss SSE intrin.h __m128 _mm_mul_ss(__m128, __m128);
_mm_mulhi_epi16 SSE2 intrin.h __m128i _mm_mulhi_epi16(__m128i, __m128i);
_mm_mulhi_epu16 SSE2 intrin.h __m128i _mm_mulhi_epu16(__m128i, __m128i);
_mm_mulhrs_epi16 SSSE3 intrin.h __m128i _mm_mulhrs_epi16(__m128i, __m128i);
_mm_mullo_epi16 SSE2 intrin.h __m128i _mm_mullo_epi16(__m128i, __m128i);
_mm_mullo_epi32 SSE41 intrin.h __m128i _mm_mullo_epi32(__m128i, __m128i);
_mm_mwait SSE3 intrin.h void _mm_mwait(unsigned int, unsigned int);
_mm_nmacc_pd FMA4 ammintrin.h __m128d _mm_nmacc_pd(__m128d, __m128d, __m128d);
_mm_nmacc_ps FMA4 ammintrin.h __m128 _mm_nmacc_ps(__m128, __m128, __m128);
_mm_nmacc_sd FMA4 ammintrin.h __m128d _mm_nmacc_sd(__m128d, __m128d, __m128d);
_mm_nmacc_ss FMA4 ammintrin.h __m128 _mm_nmacc_ss(__m128, __m128, __m128);
_mm_nmsub_pd FMA4 ammintrin.h __m128d _mm_nmsub_pd(__m128d, __m128d, __m128d);
_mm_nmsub_ps FMA4 ammintrin.h __m128 _mm_nmsub_ps(__m128, __m128, __m128);
_mm_nmsub_sd FMA4 ammintrin.h __m128d _mm_nmsub_sd(__m128d, __m128d, __m128d);
_mm_nmsub_ss FMA4 ammintrin.h __m128 _mm_nmsub_ss(__m128, __m128, __m128);
_mm_or_pd SSE2 intrin.h __m128d _mm_or_pd(__m128d, __m128d);
_mm_or_ps SSE intrin.h __m128 _mm_or_ps(__m128, __m128);
_mm_or_si128 SSE2 intrin.h __m128i _mm_or_si128(__m128i, __m128i);
_mm_packs_epi16 SSE2 intrin.h __m128i _mm_packs_epi16(__m128i, __m128i);
_mm_packs_epi32 SSE2 intrin.h __m128i _mm_packs_epi32(__m128i, __m128i);
_mm_packus_epi16 SSE2 intrin.h __m128i _mm_packus_epi16(__m128i, __m128i);
_mm_packus_epi32 SSE41 intrin.h __m128i _mm_packus_epi32(__m128i, __m128i);
_mm_pause SSE2 intrin.h void _mm_pause(void);
_mm_perm_epi8 XOP ammintrin.h __m128i _mm_perm_epi8(__m128i, __m128i, __m128i);
_mm_permute_pd AVX immintrin.h __m128d _mm_permute_pd(__m128d, int);
_mm_permute_ps AVX immintrin.h __m128 _mm_permute_ps(__m128, int);
_mm_permute2_pd XOP ammintrin.h __m128d _mm_permute2_pd(__m128d, __m128d, __m128i, int);
_mm_permute2_ps XOP ammintrin.h __m128 _mm_permute2_ps(__m128, __m128, __m128i, int);
_mm_permutevar_pd AVX immintrin.h __m128d _mm_permutevar_pd(__m128d, __m128i);
_mm_permutevar_ps AVX immintrin.h __m128 _mm_permutevar_ps(__m128, __m128i);
_mm_popcnt_u32 POPCNT intrin.h int _mm_popcnt_u32(unsigned int);
_mm_popcnt_u64 POPCNT intrin.h __int64 _mm_popcnt_u64(unsigned __int64);
_mm_prefetch SSE intrin.h void _mm_prefetch(char*, int);
_mm_rcp_ps SSE intrin.h __m128 _mm_rcp_ps(__m128);
_mm_rcp_ss SSE intrin.h __m128 _mm_rcp_ss(__m128);
_mm_rot_epi16 XOP ammintrin.h __m128i _mm_rot_epi16(__m128i, __m128i);
_mm_rot_epi32 XOP ammintrin.h __m128i _mm_rot_epi32(__m128i, __m128i);
_mm_rot_epi64 XOP ammintrin.h __m128i _mm_rot_epi64(__m128i, __m128i);
_mm_rot_epi8 XOP ammintrin.h __m128i _mm_rot_epi8(__m128i, __m128i);
_mm_roti_epi16 XOP ammintrin.h __m128i _mm_rot_epi16(__m128i, int);
_mm_roti_epi32 XOP ammintrin.h __m128i _mm_rot_epi32(__m128i, int);
_mm_roti_epi64 XOP ammintrin.h __m128i _mm_rot_epi64(__m128i, int);
_mm_roti_epi8 XOP ammintrin.h __m128i _mm_rot_epi8(__m128i, int);
_mm_round_pd SSE41 intrin.h __m128d _mm_round_pd(__m128d, const int);
_mm_round_ps SSE41 intrin.h __m128 _mm_round_ps(__m128, const int);
_mm_round_sd SSE41 intrin.h __m128d _mm_round_sd(__m128d, __m128d, const int);
_mm_round_ss SSE41 intrin.h __m128 _mm_round_ss(__m128, __m128, const int);
_mm_rsqrt_ps SSE intrin.h __m128 _mm_rsqrt_ps(__m128);
_mm_rsqrt_ss SSE intrin.h __m128 _mm_rsqrt_ss(__m128);
_mm_sad_epu8 SSE2 intrin.h __m128i _mm_sad_epu8(__m128i, __m128i);
_mm_set_epi16 SSE2 intrin.h __m128i _mm_set_epi16(short, short, short, short, short, short, short, short);
_mm_set_epi32 SSE2 intrin.h __m128i _mm_set_epi32(int, int, int, int);
_mm_set_epi64x SSE2 intrin.h __m128i _mm_set_epi64x(__int64, __int64);
_mm_set_epi8 SSE2 intrin.h __m128i _mm_set_epi8(char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char);
_mm_set_pd SSE2 intrin.h __m128d _mm_set_pd(double, double);
_mm_set_ps SSE intrin.h __m128 _mm_set_ps(float, float, float, float);
_mm_set_ps1 SSE intrin.h __m128 _mm_set_ps1(float);
_mm_set_sd SSE2 intrin.h __m128d _mm_set_sd(double);
_mm_set_ss SSE intrin.h __m128 _mm_set_ss(float);
_mm_set1_epi16 SSE2 intrin.h __m128i _mm_set1_epi16(short);
_mm_set1_epi32 SSE2 intrin.h __m128i _mm_set1_epi32(int);
_mm_set1_epi64x SSE2 intrin.h __m128i _mm_set1_epi64x(__int64);
_mm_set1_epi8 SSE2 intrin.h __m128i _mm_set1_epi8(char);
_mm_set1_pd SSE2 intrin.h __m128d _mm_set1_pd(double);
_mm_setcsr SSE intrin.h void _mm_setcsr(unsigned int);
_mm_setl_epi64 SSE2 intrin.h __m128i _mm_setl_epi64(__m128i);
_mm_setr_epi16 SSE2 intrin.h __m128i _mm_setr_epi16(short, short, short, short, short, short, short, short);
_mm_setr_epi32 SSE2 intrin.h __m128i _mm_setr_epi32(int, int, int, int);
_mm_setr_epi8 SSE2 intrin.h __m128i _mm_setr_epi8(char, char, char, char, char, char, char, char, char, char, char, char, char, char, char, char);
_mm_setr_pd SSE2 intrin.h __m128d _mm_setr_pd(double, double);
_mm_setr_ps SSE intrin.h __m128 _mm_setr_ps(float, float, float, float);
_mm_setzero_pd SSE2 intrin.h __m128d _mm_setzero_pd(void);
_mm_setzero_ps SSE intrin.h __m128 _mm_setzero_ps(void);
_mm_setzero_si128 SSE2 intrin.h __m128i _mm_setzero_si128(void);
_mm_sfence SSE intrin.h void _mm_sfence(void);
_mm_sha_epi16 XOP ammintrin.h __m128i _mm_sha_epi16(__m128i, __m128i);
_mm_sha_epi32 XOP ammintrin.h __m128i _mm_sha_epi32(__m128i, __m128i);
_mm_sha_epi64 XOP ammintrin.h __m128i _mm_sha_epi64(__m128i, __m128i);
_mm_sha_epi8 XOP ammintrin.h __m128i _mm_sha_epi8(__m128i, __m128i);
_mm_shl_epi16 XOP ammintrin.h __m128i _mm_shl_epi16(__m128i, __m128i);
_mm_shl_epi32 XOP ammintrin.h __m128i _mm_shl_epi32(__m128i, __m128i);
_mm_shl_epi64 XOP ammintrin.h __m128i _mm_shl_epi64(__m128i, __m128i);
_mm_shl_epi8 XOP ammintrin.h __m128i _mm_shl_epi8(__m128i, __m128i);
_mm_shuffle_epi32 SSE2 intrin.h __m128i _mm_shuffle_epi32(__m128i, int);
_mm_shuffle_epi8 SSSE3 intrin.h __m128i _mm_shuffle_epi8(__m128i, __m128i);
_mm_shuffle_pd SSE2 intrin.h __m128d _mm_shuffle_pd(__m128d, __m128d, int);
_mm_shuffle_ps SSE intrin.h __m128 _mm_shuffle_ps(__m128, __m128, unsigned int);
_mm_shufflehi_epi16 SSE2 intrin.h __m128i _mm_shufflehi_epi16(__m128i, int);
_mm_shufflelo_epi16 SSE2 intrin.h __m128i _mm_shufflelo_epi16(__m128i, int);
_mm_sign_epi16 SSSE3 intrin.h __m128i _mm_sign_epi16(__m128i, __m128i);
_mm_sign_epi32 SSSE3 intrin.h __m128i _mm_sign_epi32(__m128i, __m128i);
_mm_sign_epi8 SSSE3 intrin.h __m128i _mm_sign_epi8(__m128i, __m128i);
_mm_sll_epi16 SSE2 intrin.h __m128i _mm_sll_epi16(__m128i, __m128i);
_mm_sll_epi32 SSE2 intrin.h __m128i _mm_sll_epi32(__m128i, __m128i);
_mm_sll_epi64 SSE2 intrin.h __m128i _mm_sll_epi64(__m128i, __m128i);
_mm_slli_epi16 SSE2 intrin.h __m128i _mm_slli_epi16(__m128i, int);
_mm_slli_epi32 SSE2 intrin.h __m128i _mm_slli_epi32(__m128i, int);
_mm_slli_epi64 SSE2 intrin.h __m128i _mm_slli_epi64(__m128i, int);
_mm_slli_si128 SSE2 intrin.h __m128i _mm_slli_si128(__m128i, int);
_mm_sllv_epi32 AVX2 immintrin.h __m128i _mm_sllv_epi32(__m128i, __m128i);
_mm_sllv_epi64 AVX2 immintrin.h __m128i _mm_sllv_epi64(__m128i, __m128i);
_mm_sqrt_pd SSE2 intrin.h __m128d _mm_sqrt_pd(__m128d);
_mm_sqrt_ps SSE intrin.h __m128 _mm_sqrt_ps(__m128);
_mm_sqrt_sd SSE2 intrin.h __m128d _mm_sqrt_sd(__m128d, __m128d);
_mm_sqrt_ss SSE intrin.h __m128 _mm_sqrt_ss(__m128);
_mm_sra_epi16 SSE2 intrin.h __m128i _mm_sra_epi16(__m128i, __m128i);
_mm_sra_epi32 SSE2 intrin.h __m128i _mm_sra_epi32(__m128i, __m128i);
_mm_srai_epi16 SSE2 intrin.h __m128i _mm_srai_epi16(__m128i, int);
_mm_srai_epi32 SSE2 intrin.h __m128i _mm_srai_epi32(__m128i, int);
_mm_srav_epi32 AVX2 immintrin.h __m128i _mm_srav_epi32(__m128i, __m128i);
_mm_srl_epi16 SSE2 intrin.h __m128i _mm_srl_epi16(__m128i, __m128i);
_mm_srl_epi32 SSE2 intrin.h __m128i _mm_srl_epi32(__m128i, __m128i);
_mm_srl_epi64 SSE2 intrin.h __m128i _mm_srl_epi64(__m128i, __m128i);
_mm_srli_epi16 SSE2 intrin.h __m128i _mm_srli_epi16(__m128i, int);
_mm_srli_epi32 SSE2 intrin.h __m128i _mm_srli_epi32(__m128i, int);
_mm_srli_epi64 SSE2 intrin.h __m128i _mm_srli_epi64(__m128i, int);
_mm_srli_si128 SSE2 intrin.h __m128i _mm_srli_si128(__m128i, int);
_mm_srlv_epi32 AVX2 immintrin.h __m128i _mm_srlv_epi32(__m128i, __m128i);
_mm_srlv_epi64 AVX2 immintrin.h __m128i _mm_srlv_epi64(__m128i, __m128i);
_mm_store_pd SSE2 intrin.h void _mm_store_pd(double*, __m128d);
_mm_store_ps SSE intrin.h void _mm_store_ps(float*, __m128);
_mm_store_ps1 SSE intrin.h void _mm_store_ps1(float*, __m128);
_mm_store_sd SSE2 intrin.h void _mm_store_sd(double*, __m128d);
_mm_store_si128 SSE2 intrin.h void _mm_store_si128(__m128i*, __m128i);
_mm_store_ss SSE intrin.h void _mm_store_ss(float*, __m128);
_mm_store1_pd SSE2 intrin.h void _mm_store1_pd(double*, __m128d);
_mm_storeh_pd SSE2 intrin.h void _mm_storeh_pd(double*, __m128d);
_mm_storeh_pi SSE intrin.h void _mm_storeh_pi(__m64*, __m128);
_mm_storel_epi64 SSE2 intrin.h void _mm_storel_epi64(__m128i*, __m128i);
_mm_storel_pd SSE2 intrin.h void _mm_storel_pd(double*, __m128d);
_mm_storel_pi SSE intrin.h void _mm_storel_pi(__m64*, __m128);
_mm_storer_pd SSE2 intrin.h void _mm_storer_pd(double*, __m128d);
_mm_storer_ps SSE intrin.h void _mm_storer_ps(float*, __m128);
_mm_storeu_pd SSE2 intrin.h void _mm_storeu_pd(double*, __m128d);
_mm_storeu_ps SSE intrin.h void _mm_storeu_ps(float*, __m128);
_mm_storeu_si128 SSE2 intrin.h void _mm_storeu_si128(__m128i*, __m128i);
_mm_stream_load_si128 SSE41 intrin.h __m128i _mm_stream_load_si128(__m128i*);
_mm_stream_pd SSE2 intrin.h void _mm_stream_pd(double*, __m128d);
_mm_stream_ps SSE intrin.h void _mm_stream_ps(float*, __m128);
_mm_stream_sd SSE4a intrin.h void _mm_stream_sd(double*, __m128d);
_mm_stream_si128 SSE2 intrin.h void _mm_stream_si128(__m128i*, __m128i);
_mm_stream_si32 SSE2 intrin.h void _mm_stream_si32(int*, int);
_mm_stream_si64x SSE2 intrin.h void _mm_stream_si64x(__int64 *, __int64);
_mm_stream_ss SSE4a intrin.h void _mm_stream_ss(float*, __m128);
_mm_sub_epi16 SSE2 intrin.h __m128i _mm_sub_epi16(__m128i, __m128i);
_mm_sub_epi32 SSE2 intrin.h __m128i _mm_sub_epi32(__m128i, __m128i);
_mm_sub_epi64 SSE2 intrin.h __m128i _mm_sub_epi64(__m128i, __m128i);
_mm_sub_epi8 SSE2 intrin.h __m128i _mm_sub_epi8(__m128i, __m128i);
_mm_sub_pd SSE2 intrin.h __m128d _mm_sub_pd(__m128d, __m128d);
_mm_sub_ps SSE intrin.h __m128 _mm_sub_ps(__m128, __m128);
_mm_sub_sd SSE2 intrin.h __m128d _mm_sub_sd(__m128d, __m128d);
_mm_sub_ss SSE intrin.h __m128 _mm_sub_ss(__m128, __m128);
_mm_subs_epi16 SSE2 intrin.h __m128i _mm_subs_epi16(__m128i, __m128i);
_mm_subs_epi8 SSE2 intrin.h __m128i _mm_subs_epi8(__m128i, __m128i);
_mm_subs_epu16 SSE2 intrin.h __m128i _mm_subs_epu16(__m128i, __m128i);
_mm_subs_epu8 SSE2 intrin.h __m128i _mm_subs_epu8(__m128i, __m128i);
_mm_testc_pd AVX immintrin.h int _mm_testc_pd(__m128d, __m128d);
_mm_testc_ps AVX immintrin.h int _mm_testc_ps(__m128, __m128);
_mm_testc_si128 SSE41 intrin.h int _mm_testc_si128(__m128i, __m128i);
_mm_testnzc_pd AVX immintrin.h int _mm_testnzc_pd(__m128d, __m128d);
_mm_testnzc_ps AVX immintrin.h int _mm_testnzc_ps(__m128, __m128);
_mm_testnzc_si128 SSE41 intrin.h int _mm_testnzc_si128(__m128i, __m128i);
_mm_testz_pd AVX immintrin.h int _mm_testz_pd(__m128d, __m128d);
_mm_testz_ps AVX immintrin.h int _mm_testz_ps(__m128, __m128);
_mm_testz_si128 SSE41 intrin.h int _mm_testz_si128(__m128i, __m128i);
_mm_ucomieq_sd SSE2 intrin.h int _mm_ucomieq_sd(__m128d, __m128d);
_mm_ucomieq_ss SSE intrin.h int _mm_ucomieq_ss(__m128, __m128);
_mm_ucomige_sd SSE2 intrin.h int _mm_ucomige_sd(__m128d, __m128d);
_mm_ucomige_ss SSE intrin.h int _mm_ucomige_ss(__m128, __m128);
_mm_ucomigt_sd SSE2 intrin.h int _mm_ucomigt_sd(__m128d, __m128d);
_mm_ucomigt_ss SSE intrin.h int _mm_ucomigt_ss(__m128, __m128);
_mm_ucomile_sd SSE2 intrin.h int _mm_ucomile_sd(__m128d, __m128d);
_mm_ucomile_ss SSE intrin.h int _mm_ucomile_ss(__m128, __m128);
_mm_ucomilt_sd SSE2 intrin.h int _mm_ucomilt_sd(__m128d, __m128d);
_mm_ucomilt_ss SSE intrin.h int _mm_ucomilt_ss(__m128, __m128);
_mm_ucomineq_sd SSE2 intrin.h int _mm_ucomineq_sd(__m128d, __m128d);
_mm_ucomineq_ss SSE intrin.h int _mm_ucomineq_ss(__m128, __m128);
_mm_unpackhi_epi16 SSE2 intrin.h __m128i _mm_unpackhi_epi16(__m128i, __m128i);
_mm_unpackhi_epi32 SSE2 intrin.h __m128i _mm_unpackhi_epi32(__m128i, __m128i);
_mm_unpackhi_epi64 SSE2 intrin.h __m128i _mm_unpackhi_epi64(__m128i, __m128i);
_mm_unpackhi_epi8 SSE2 intrin.h __m128i _mm_unpackhi_epi8(__m128i, __m128i);
_mm_unpackhi_pd SSE2 intrin.h __m128d _mm_unpackhi_pd(__m128d, __m128d);
_mm_unpackhi_ps SSE intrin.h __m128 _mm_unpackhi_ps(__m128, __m128);
_mm_unpacklo_epi16 SSE2 intrin.h __m128i _mm_unpacklo_epi16(__m128i, __m128i);
_mm_unpacklo_epi32 SSE2 intrin.h __m128i _mm_unpacklo_epi32(__m128i, __m128i);
_mm_unpacklo_epi64 SSE2 intrin.h __m128i _mm_unpacklo_epi64(__m128i, __m128i);
_mm_unpacklo_epi8 SSE2 intrin.h __m128i _mm_unpacklo_epi8(__m128i, __m128i);
_mm_unpacklo_pd SSE2 intrin.h __m128d _mm_unpacklo_pd(__m128d, __m128d);
_mm_unpacklo_ps SSE intrin.h __m128 _mm_unpacklo_ps(__m128, __m128);
_mm_xor_pd SSE2 intrin.h __m128d _mm_xor_pd(__m128d, __m128d);
_mm_xor_ps SSE intrin.h __m128 _mm_xor_ps(__m128, __m128);
_mm_xor_si128 SSE2 intrin.h __m128i _mm_xor_si128(__m128i, __m128i);
_mm256_abs_epi16 AVX2 immintrin.h __m256i _mm256_abs_epi16(__m256i);
_mm256_abs_epi32 AVX2 immintrin.h __m256i _mm256_abs_epi32(__m256i);
_mm256_abs_epi8 AVX2 immintrin.h __m256i _mm256_abs_epi8(__m256i);
_mm256_add_epi16 AVX2 immintrin.h __m256i _mm256_add_epi16(__m256i, __m256i);
_mm256_add_epi32 AVX2 immintrin.h __m256i _mm256_add_epi32(__m256i, __m256i);
_mm256_add_epi64 AVX2 immintrin.h __m256i _mm256_add_epi64(__m256i, __m256i);
_mm256_add_epi8 AVX2 immintrin.h __m256i _mm256_add_epi8(__m256i, __m256i);
_mm256_add_pd AVX immintrin.h __m256d _mm256_add_pd(__m256d, __m256d);
_mm256_add_ps AVX immintrin.h __m256 _mm256_add_ps(__m256, __m256);
_mm256_adds_epi16 AVX2 immintrin.h __m256i _mm256_adds_epi16(__m256i, __m256i);
_mm256_adds_epi8 AVX2 immintrin.h __m256i _mm256_adds_epi8(__m256i, __m256i);
_mm256_adds_epu16 AVX2 immintrin.h __m256i _mm256_adds_epu16(__m256i, __m256i);
_mm256_adds_epu8 AVX2 immintrin.h __m256i _mm256_adds_epu8(__m256i, __m256i);
_mm256_addsub_pd AVX immintrin.h __m256d _mm256_addsub_pd(__m256d, __m256d);
_mm256_addsub_ps AVX immintrin.h __m256 _mm256_addsub_ps(__m256, __m256);
_mm256_alignr_epi8 AVX2 immintrin.h __m256i _mm256_alignr_epi8(__m256i, __m256i, const int);
_mm256_and_pd AVX immintrin.h __m256d _mm256_and_pd(__m256d, __m256d);
_mm256_and_ps AVX immintrin.h __m256 _mm256_and_ps(__m256, __m256);
_mm256_and_si256 AVX2 immintrin.h __m256i _mm256_and_si256(__m256i, __m256i);
_mm256_andnot_pd AVX immintrin.h __m256d _mm256_andnot_pd(__m256d, __m256d);
_mm256_andnot_ps AVX immintrin.h __m256 _mm256_andnot_ps(__m256, __m256);
_mm256_andnot_si256 AVX2 immintrin.h __m256i _mm256_andnot_si256(__m256i, __m256i);
_mm256_avg_epu16 AVX2 immintrin.h __m256i _mm256_avg_epu16(__m256i, __m256i);
_mm256_avg_epu8 AVX2 immintrin.h __m256i _mm256_avg_epu8(__m256i, __m256i);
_mm256_blend_epi16 AVX2 immintrin.h __m256i _mm256_blend_epi16(__m256i, __m256i, const int);
_mm256_blend_epi32 AVX2 immintrin.h __m256i _mm256_blend_epi32(__m256i, __m256i, const int);
_mm256_blend_pd AVX immintrin.h __m256d _mm256_blend_pd(__m256d, __m256d, const int);
_mm256_blend_ps AVX immintrin.h __m256 _mm256_blend_ps(__m256, __m256, const int);
_mm256_blendv_epi8 AVX2 immintrin.h __m256i _mm256_blendv_epi8(__m256i, __m256i, __m256i);
_mm256_blendv_pd AVX immintrin.h __m256d _mm256_blendv_pd(__m256d, __m256d, __m256d);
_mm256_blendv_ps AVX immintrin.h __m256 _mm256_blendv_ps(__m256, __m256, __m256);
_mm256_broadcast_pd AVX immintrin.h __m256d _mm256_broadcast_pd(__m128d const *);
_mm256_broadcast_ps AVX immintrin.h __m256 _mm256_broadcast_ps(__m128 const *);
_mm256_broadcast_sd AVX immintrin.h __m256d _mm256_broadcast_sd(double const *);
_mm256_broadcast_ss AVX immintrin.h __m256 _mm256_broadcast_ss(float const *);
_mm256_broadcastb_epi8 AVX2 immintrin.h __m256i _mm256_broadcastb_epi8 (__m128i);
_mm256_broadcastd_epi32 AVX2 immintrin.h __m256i _mm256_broadcastd_epi32(__m128i);
_mm256_broadcastq_epi64 AVX2 immintrin.h __m256i _mm256_broadcastq_epi64(__m128i);
_mm256_broadcastsd_pd AVX2 immintrin.h __m256d _mm256_broadcastsd_pd(__m128d);
_mm256_broadcastsi128_si256 AVX2 immintrin.h __m256i _mm256_broadcastsi128_si256(__m128i);
_mm256_broadcastss_ps AVX2 immintrin.h __m256 _mm256_broadcastss_ps(__m128);
_mm256_broadcastw_epi16 AVX2 immintrin.h __m256i _mm256_broadcastw_epi16(__m128i);
_mm256_castpd_ps AVX immintrin.h __m256 _mm256_castpd_ps(__m256d);
_mm256_castpd_si256 AVX immintrin.h __m256i _mm256_castpd_si256(__m256d);
_mm256_castpd128_pd256 AVX immintrin.h __m256d _mm256_castpd128_pd256(__m128d);
_mm256_castpd256_pd128 AVX immintrin.h __m128d _mm256_castpd256_pd128(__m256d);
_mm256_castps_pd AVX immintrin.h __m256d _mm256_castps_pd(__m256);
_mm256_castps_si256 AVX immintrin.h __m256i _mm256_castps_si256(__m256);
_mm256_castps128_ps256 AVX immintrin.h __m256 _mm256_castps128_ps256(__m128);
_mm256_castps256_ps128 AVX immintrin.h __m128 _mm256_castps256_ps128(__m256);
_mm256_castsi128_si256 AVX immintrin.h __m256i _mm256_castsi128_si256(__m128i);
_mm256_castsi256_pd AVX immintrin.h __m256d _mm256_castsi256_pd(__m256i);
_mm256_castsi256_ps AVX immintrin.h __m256 _mm256_castsi256_ps(__m256i);
_mm256_castsi256_si128 AVX immintrin.h __m128i _mm256_castsi256_si128(__m256i);
_mm256_cmov_si256 XOP ammintrin.h __m256i _mm256_cmov_si256(__m256i, __m256i, __m256i);
_mm256_cmp_pd AVX immintrin.h __m256d _mm256_cmp_pd(__m256d, __m256d, const int);
_mm256_cmp_ps AVX immintrin.h __m256 _mm256_cmp_ps(__m256, __m256, const int);
_mm256_cmpeq_epi16 AVX2 immintrin.h __m256i _mm256_cmpeq_epi16(__m256i, __m256i);
_mm256_cmpeq_epi32 AVX2 immintrin.h __m256i _mm256_cmpeq_epi32(__m256i, __m256i);
_mm256_cmpeq_epi64 AVX2 immintrin.h __m256i _mm256_cmpeq_epi64(__m256i, __m256i);
_mm256_cmpeq_epi8 AVX2 immintrin.h __m256i _mm256_cmpeq_epi8(__m256i, __m256i);
_mm256_cmpgt_epi16 AVX2 immintrin.h __m256i _mm256_cmpgt_epi16(__m256i, __m256i);
_mm256_cmpgt_epi32 AVX2 immintrin.h __m256i _mm256_cmpgt_epi32(__m256i, __m256i);
_mm256_cmpgt_epi64 AVX2 immintrin.h __m256i _mm256_cmpgt_epi64(__m256i, __m256i);
_mm256_cmpgt_epi8 AVX2 immintrin.h __m256i _mm256_cmpgt_epi8(__m256i, __m256i);
_mm256_cvtepi16_epi32 AVX2 immintrin.h __m256i _mm256_cvtepi16_epi32(__m128i);
_mm256_cvtepi16_epi64 AVX2 immintrin.h __m256i _mm256_cvtepi16_epi64(__m128i);
_mm256_cvtepi32_epi64 AVX2 immintrin.h __m256i _mm256_cvtepi32_epi64(__m128i);
_mm256_cvtepi32_pd AVX immintrin.h __m256d _mm256_cvtepi32_pd(__m128i);
_mm256_cvtepi32_ps AVX immintrin.h __m256 _mm256_cvtepi32_ps(__m256i);
_mm256_cvtepi8_epi16 AVX2 immintrin.h __m256i _mm256_cvtepi8_epi16(__m128i);
_mm256_cvtepi8_epi32 AVX2 immintrin.h __m256i _mm256_cvtepi8_epi32(__m128i);
_mm256_cvtepi8_epi64 AVX2 immintrin.h __m256i _mm256_cvtepi8_epi64(__m128i);
_mm256_cvtepu16_epi32 AVX2 immintrin.h __m256i _mm256_cvtepu16_epi32(__m128i);
_mm256_cvtepu16_epi64 AVX2 immintrin.h __m256i _mm256_cvtepu16_epi64(__m128i);
_mm256_cvtepu32_epi64 AVX2 immintrin.h __m256i _mm256_cvtepu32_epi64(__m128i);
_mm256_cvtepu8_epi16 AVX2 immintrin.h __m256i _mm256_cvtepu8_epi16(__m128i);
_mm256_cvtepu8_epi32 AVX2 immintrin.h __m256i _mm256_cvtepu8_epi32(__m128i);
_mm256_cvtepu8_epi64 AVX2 immintrin.h __m256i _mm256_cvtepu8_epi64(__m128i);
_mm256_cvtpd_epi32 AVX immintrin.h __m128i _mm256_cvtpd_epi32(__m256d);
_mm256_cvtpd_ps AVX immintrin.h __m128 _mm256_cvtpd_ps(__m256d);
_mm256_cvtph_ps F16C immintrin.h __m256 _mm256_cvtph_ps(__m128i);
_mm256_cvtps_epi32 AVX immintrin.h __m256i _mm256_cvtps_epi32(__m256);
_mm256_cvtps_pd AVX immintrin.h __m256d _mm256_cvtps_pd(__m128);
_mm256_cvtps_ph F16C immintrin.h __m128i _mm256_cvtps_ph(__m256, const int);
_mm256_cvttpd_epi32 AVX immintrin.h __m128i _mm256_cvttpd_epi32(__m256d);
_mm256_cvttps_epi32 AVX immintrin.h __m256i _mm256_cvttps_epi32(__m256);
_mm256_div_pd AVX immintrin.h __m256d _mm256_div_pd(__m256d, __m256d);
_mm256_div_ps AVX immintrin.h __m256 _mm256_div_ps(__m256, __m256);
_mm256_dp_ps AVX immintrin.h __m256 _mm256_dp_ps(__m256, __m256, const int);
_mm256_extractf128_pd AVX immintrin.h __m128d _mm256_extractf128_pd(__m256d, const int);
_mm256_extractf128_ps AVX immintrin.h __m128 _mm256_extractf128_ps(__m256, const int);
_mm256_extractf128_si256 AVX immintrin.h __m128i _mm256_extractf128_si256(__m256i, const int);
_mm256_extracti128_si256 AVX2 immintrin.h __m128i _mm256_extracti128_si256(__m256i, int);
_mm256_fmadd_pd FMA immintrin.h __m256d _mm256_fmadd_pd (__m256d, __m256d, __m256d);
_mm256_fmadd_ps FMA immintrin.h __m256 _mm256_fmadd_ps (__m256, __m256, __m256);
_mm256_fmaddsub_pd FMA immintrin.h __m256d _mm256_fmaddsub_pd (__m256d, __m256d, __m256d);
_mm256_fmaddsub_ps FMA immintrin.h __m256 _mm256_fmaddsub_ps (__m256, __m256, __m256);
_mm256_fmsub_pd FMA immintrin.h __m256d _mm256_fmsub_pd (__m256d, __m256d, __m256d);
_mm256_fmsub_ps FMA immintrin.h __m256 _mm256_fmsub_ps (__m256, __m256, __m256);
_mm256_fmsubadd_pd FMA immintrin.h __m256d _mm256_fmsubadd_pd (__m256d, __m256d, __m256d);
_mm256_fmsubadd_ps FMA immintrin.h __m256 _mm256_fmsubadd_ps (__m256, __m256, __m256);
_mm256_fnmadd_pd FMA immintrin.h __m256d _mm256_fnmadd_pd (__m256d, __m256d, __m256d);
_mm256_fnmadd_ps FMA immintrin.h __m256 _mm256_fnmadd_ps (__m256, __m256, __m256);
_mm256_fnmsub_pd FMA immintrin.h __m256d _mm256_fnmsub_pd (__m256d, __m256d, __m256d);
_mm256_fnmsub_ps FMA immintrin.h __m256 _mm256_fnmsub_ps (__m256, __m256, __m256);
_mm256_frcz_pd XOP ammintrin.h __m256d _mm256_frcz_pd(__m256d);
_mm256_frcz_ps XOP ammintrin.h __m256 _mm256_frcz_ps(__m256);
_mm256_hadd_epi16 AVX2 immintrin.h __m256i _mm256_hadd_epi16(__m256i, __m256i);
_mm256_hadd_epi32 AVX2 immintrin.h __m256i _mm256_hadd_epi32(__m256i, __m256i);
_mm256_hadd_pd AVX immintrin.h __m256d _mm256_hadd_pd(__m256d, __m256d);
_mm256_hadd_ps AVX immintrin.h __m256 _mm256_hadd_ps(__m256, __m256);
_mm256_hadds_epi16 AVX2 immintrin.h __m256i _mm256_hadds_epi16(__m256i, __m256i);
_mm256_hsub_epi16 AVX2 immintrin.h __m256i _mm256_hsub_epi16(__m256i, __m256i);
_mm256_hsub_epi32 AVX2 immintrin.h __m256i _mm256_hsub_epi32(__m256i, __m256i);
_mm256_hsub_pd AVX immintrin.h __m256d _mm256_hsub_pd(__m256d, __m256d);
_mm256_hsub_ps AVX immintrin.h __m256 _mm256_hsub_ps(__m256, __m256);
_mm256_hsubs_epi16 AVX2 immintrin.h __m256i _mm256_hsubs_epi16(__m256i, __m256i);
_mm256_i32gather_epi32 AVX2 immintrin.h __m256i _mm256_i32gather_epi32(int const *, __m256i, const int);
_mm256_i32gather_epi64 AVX2 immintrin.h __m256i _mm256_i32gather_epi64(__int64 const *, __m128i, const int);
_mm256_i32gather_pd AVX2 immintrin.h __m256d _mm256_i32gather_pd(double const *, __m128i, const int);
_mm256_i32gather_ps AVX2 immintrin.h __m256 _mm256_i32gather_ps(float const *, __m256i, const int);
_mm256_i64gather_epi32 AVX2 immintrin.h __m256i _mm256_i64gather_epi32(int const *, __m256i, const int);
_mm256_i64gather_epi64 AVX2 immintrin.h __m256i _mm256_i64gather_epi64(__int64 const *, __m256i, const int);
_mm256_i64gather_pd AVX2 immintrin.h __m256d _mm256_i64gather_pd(double const *, __m256i, const int);
_mm256_i64gather_ps AVX2 immintrin.h __m128 _mm256_i64gather_ps(float const *, __m256i, const int);
_mm256_insertf128_pd AVX immintrin.h __m256d _mm256_insertf128_pd(__m256d, __m128d, int);
_mm256_insertf128_ps AVX immintrin.h __m256 _mm256_insertf128_ps(__m256, __m128, int);
_mm256_insertf128_si256 AVX immintrin.h __m256i _mm256_insertf128_si256(__m256i, __m128i, int);
_mm256_inserti128_si256 AVX2 immintrin.h __m256i _mm256_inserti128_si256(__m256i, __m128i, int);
_mm256_lddqu_si256 AVX immintrin.h __m256i _mm256_lddqu_si256(__m256i *);
_mm256_load_pd AVX immintrin.h __m256d _mm256_load_pd(double const *);
_mm256_load_ps AVX immintrin.h __m256 _mm256_load_ps(float const *);
_mm256_load_si256 AVX immintrin.h __m256i _mm256_load_si256(__m256i *);
_mm256_loadu_pd AVX immintrin.h __m256d _mm256_loadu_pd(double const *);
_mm256_loadu_ps AVX immintrin.h __m256 _mm256_loadu_ps(float const *);
_mm256_loadu_si256 AVX immintrin.h __m256i _mm256_loadu_si256(__m256i *);
_mm256_macc_pd FMA4 ammintrin.h __m256d _mm_macc_pd(__m256d, __m256d, __m256d);
_mm256_macc_ps FMA4 ammintrin.h __m256 _mm_macc_ps(__m256, __m256, __m256);
_mm256_madd_epi16 AVX2 immintrin.h __m256i _mm256_madd_epi16(__m256i, __m256i);
_mm256_maddsub_pd FMA4 ammintrin.h __m256d _mm_maddsub_pd(__m256d, __m256d, __m256d);
_mm256_maddsub_ps FMA4 ammintrin.h __m256 _mm_maddsub_ps(__m256, __m256, __m256);
_mm256_maddubs_epi16 AVX2 immintrin.h __m256i _mm256_maddubs_epi16(__m256i, __m256i);
_mm256_mask_i32gather_epi32 AVX2 immintrin.h __m256i _mm256_mask_i32gather_epi32(__m256i, int const *, __m256i, __m256i, const int);
_mm256_mask_i32gather_epi64 AVX2 immintrin.h __m256i _mm256_mask_i32gather_epi64(__m256i, __int64 const *, __m128i, __m256i, const int);
_mm256_mask_i32gather_pd AVX2 immintrin.h __m256d _mm256_mask_i32gather_pd(__m256d, double const *, __m128i, __m256d, const int);
_mm256_mask_i32gather_ps AVX2 immintrin.h __m256 _mm256_mask_i32gather_ps(__m256, float const *, __m256i, __m256, const int);
_mm256_mask_i64gather_epi32 AVX2 immintrin.h __m128i _mm256_mask_i64gather_epi32(__m128i, int const *, __m256i, __m128i, const int);
_mm256_mask_i64gather_epi64 AVX2 immintrin.h __m256i _mm256_mask_i64gather_epi64(__m256i, __int64 const *, __m256i, __m256i, const int);
_mm256_mask_i64gather_pd AVX2 immintrin.h __m256d _mm256_mask_i64gather_pd(__m256d, double const *, __m256i, __m256d, const int);
_mm256_mask_i64gather_ps AVX2 immintrin.h __m128 _mm256_mask_i64gather_ps(__m128, float const *, __m256i, __m128, const int);
_mm256_maskload_epi32 AVX2 immintrin.h __m256i _mm256_maskload_epi32(int const *, __m256i);
_mm256_maskload_epi64 AVX2 immintrin.h __m256i _mm256_maskload_epi64(__int64 const *, __m256i);
_mm256_maskload_pd AVX immintrin.h __m256d _mm256_maskload_pd(double const *, __m256i);
_mm256_maskload_ps AVX immintrin.h __m256 _mm256_maskload_ps(float const *, __m256i);
_mm256_maskstore_epi32 AVX2 immintrin.h void _mm256_maskstore_epi32(int *, __m256i, __m256i);
_mm256_maskstore_epi64 AVX2 immintrin.h void _mm256_maskstore_epi64(__int64 *, __m256i, __m256i);
_mm256_maskstore_pd AVX immintrin.h void _mm256_maskstore_pd(double *, __m256i, __m256d);
_mm256_maskstore_ps AVX immintrin.h void _mm256_maskstore_ps(float *, __m256i, __m256);
_mm256_max_epi16 AVX2 immintrin.h __m256i _mm256_max_epi16(__m256i, __m256i);
_mm256_max_epi32 AVX2 immintrin.h __m256i _mm256_max_epi32(__m256i, __m256i);
_mm256_max_epi8 AVX2 immintrin.h __m256i _mm256_max_epi8(__m256i, __m256i);
_mm256_max_epu16 AVX2 immintrin.h __m256i _mm256_max_epu16(__m256i, __m256i);
_mm256_max_epu32 AVX2 immintrin.h __m256i _mm256_max_epu32(__m256i, __m256i);
_mm256_max_epu8 AVX2 immintrin.h __m256i _mm256_max_epu8(__m256i, __m256i);
_mm256_max_pd AVX immintrin.h __m256d _mm256_max_pd(__m256d, __m256d);
_mm256_max_ps AVX immintrin.h __m256 _mm256_max_ps(__m256, __m256);
_mm256_min_epi16 AVX2 immintrin.h __m256i _mm256_min_epi16(__m256i, __m256i);
_mm256_min_epi32 AVX2 immintrin.h __m256i _mm256_min_epi32(__m256i, __m256i);
_mm256_min_epi8 AVX2 immintrin.h __m256i _mm256_min_epi8(__m256i, __m256i);
_mm256_min_epu16 AVX2 immintrin.h __m256i _mm256_min_epu16(__m256i, __m256i);
_mm256_min_epu32 AVX2 immintrin.h __m256i _mm256_min_epu32(__m256i, __m256i);
_mm256_min_epu8 AVX2 immintrin.h __m256i _mm256_min_epu8(__m256i, __m256i);
_mm256_min_pd AVX immintrin.h __m256d _mm256_min_pd(__m256d, __m256d);
_mm256_min_ps AVX immintrin.h __m256 _mm256_min_ps(__m256, __m256);
_mm256_movedup_pd AVX immintrin.h __m256d _mm256_movedup_pd(__m256d);
_mm256_movehdup_ps AVX immintrin.h __m256 _mm256_movehdup_ps(__m256);
_mm256_moveldup_ps AVX immintrin.h __m256 _mm256_moveldup_ps(__m256);
_mm256_movemask_epi8 AVX2 immintrin.h int _mm256_movemask_epi8(__m256i);
_mm256_movemask_pd AVX immintrin.h int _mm256_movemask_pd(__m256d);
_mm256_movemask_ps AVX immintrin.h int _mm256_movemask_ps(__m256);
_mm256_mpsadbw_epu8 AVX2 immintrin.h __m256i _mm256_mpsadbw_epu8(__m256i, __m256i, const int);
_mm256_msub_pd FMA4 ammintrin.h __m256d _mm_msub_pd(__m256d, __m256d, __m256d);
_mm256_msub_ps FMA4 ammintrin.h __m256 _mm_msub_ps(__m256, __m256, __m256);
_mm256_msubadd_pd FMA4 ammintrin.h __m256d _mm_msubadd_pd(__m256d, __m256d, __m256d);
_mm256_msubadd_ps FMA4 ammintrin.h __m256 _mm_msubadd_ps(__m256, __m256, __m256);
_mm256_mul_epi32 AVX2 immintrin.h __m256i _mm256_mul_epi32(__m256i, __m256i);
_mm256_mul_epu32 AVX2 immintrin.h __m256i _mm256_mul_epu32(__m256i, __m256i);
_mm256_mul_pd AVX immintrin.h __m256d _mm256_mul_pd(__m256d, __m256d);
_mm256_mul_ps AVX immintrin.h __m256 _mm256_mul_ps(__m256, __m256);
_mm256_mulhi_epi16 AVX2 immintrin.h __m256i _mm256_mulhi_epi16(__m256i, __m256i);
_mm256_mulhi_epu16 AVX2 immintrin.h __m256i _mm256_mulhi_epu16(__m256i, __m256i);
_mm256_mulhrs_epi16 AVX2 immintrin.h __m256i _mm256_mulhrs_epi16(__m256i, __m256i);
_mm256_mullo_epi16 AVX2 immintrin.h __m256i _mm256_mullo_epi16(__m256i, __m256i);
_mm256_mullo_epi32 AVX2 immintrin.h __m256i _mm256_mullo_epi32(__m256i, __m256i);
_mm256_nmacc_pd FMA4 ammintrin.h __m256d _mm_nmacc_pd(__m256d, __m256d, __m256d);
_mm256_nmacc_ps FMA4 ammintrin.h __m256 _mm_nmacc_ps(__m256, __m256, __m256);
_mm256_nmsub_pd FMA4 ammintrin.h __m256d _mm_nmsub_pd(__m256d, __m256d, __m256d);
_mm256_nmsub_ps FMA4 ammintrin.h __m256 _mm_nmsub_ps(__m256, __m256, __m256);
_mm256_or_pd AVX immintrin.h __m256d _mm256_or_pd(__m256d, __m256d);
_mm256_or_ps AVX immintrin.h __m256 _mm256_or_ps(__m256, __m256);
_mm256_or_si256 AVX2 immintrin.h __m256i _mm256_or_si256(__m256i, __m256i);
_mm256_packs_epi16 AVX2 immintrin.h __m256i _mm256_packs_epi16(__m256i, __m256i);
_mm256_packs_epi32 AVX2 immintrin.h __m256i _mm256_packs_epi32(__m256i, __m256i);
_mm256_packus_epi16 AVX2 immintrin.h __m256i _mm256_packus_epi16(__m256i, __m256i);
_mm256_packus_epi32 AVX2 immintrin.h __m256i _mm256_packus_epi32(__m256i, __m256i);
_mm256_permute_pd AVX immintrin.h __m256d _mm256_permute_pd(__m256d, int);
_mm256_permute_ps AVX immintrin.h __m256 _mm256_permute_ps(__m256, int);
_mm256_permute2_pd XOP ammintrin.h __m256d _mm256_permute2_pd(__m256d, __m256d, __m256i, int);
_mm256_permute2_ps XOP ammintrin.h __m256 _mm256_permute2_ps(__m256, __m256, __m256i, int);
_mm256_permute2f128_pd AVX immintrin.h __m256d _mm256_permute2f128_pd(__m256d, __m256d, int);
_mm256_permute2f128_ps AVX immintrin.h __m256 _mm256_permute2f128_ps(__m256, __m256, int);
_mm256_permute2f128_si256 AVX immintrin.h __m256i _mm256_permute2f128_si256(__m256i, __m256i, int);
_mm256_permute2x128_si256 AVX2 immintrin.h __m256i _mm256_permute2x128_si256(__m256i, __m256i, const int);
_mm256_permute4x64_epi64 AVX2 immintrin.h __m256i _mm256_permute4x64_epi64 (__m256i, const int);
_mm256_permute4x64_pd AVX2 immintrin.h __m256d _mm256_permute4x64_pd(__m256d, const int);
_mm256_permutevar_pd AVX immintrin.h __m256d _mm256_permutevar_pd(__m256d, __m256i);
_mm256_permutevar_ps AVX immintrin.h __m256 _mm256_permutevar_ps(__m256, __m256i);
_mm256_permutevar8x32_epi32 AVX2 immintrin.h __m256i _mm256_permutevar8x32_epi32(__m256i, __m256i);
_mm256_permutevar8x32_ps AVX2 immintrin.h __m256 _mm256_permutevar8x32_ps (__m256, __m256i);
_mm256_rcp_ps AVX immintrin.h __m256 _mm256_rcp_ps(__m256);
_mm256_round_pd AVX immintrin.h __m256d _mm256_round_pd(__m256d, int);
_mm256_round_ps AVX immintrin.h __m256 _mm256_round_ps(__m256, int);
_mm256_rsqrt_ps AVX immintrin.h __m256 _mm256_rsqrt_ps(__m256);
_mm256_sad_epu8 AVX2 immintrin.h __m256i _mm256_sad_epu8(__m256i, __m256i);
_mm256_set_epi16 AVX immintrin.h (__m256i _mm256_set_epi16(short, short, short, short, short, short, short, short, short, short, short, short, short, short, short, short);
_mm256_set_epi32 AVX immintrin.h __m256i _mm256_set_epi32(int, int, int, int, int, int, int, int);
_mm256_set_epi64x AVX immintrin.h __m256i _mm256_set_epi64x(long long, long long, long long, long long);
_mm256_set_epi8 AVX immintrin.h __m256i _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 AVX immintrin.h __m256d _mm256_set_pd(double, double, double, double);
_mm256_set_ps AVX immintrin.h __m256 _mm256_set_ps(float, float, float, float, float, float, float, float);
_mm256_set1_epi16 AVX immintrin.h __m256i _mm256_set1_epi16(short);
_mm256_set1_epi32 AVX immintrin.h __m256i _mm256_set1_epi32(int);
_mm256_set1_epi64x AVX immintrin.h __m256i _mm256_set1_epi64x(long long);
_mm256_set1_epi8 AVX immintrin.h __m256i _mm256_set1_epi8(char);
_mm256_set1_pd AVX immintrin.h __m256d _mm256_set1_pd(double);
_mm256_set1_ps AVX immintrin.h __m256 _mm256_set1_ps(float);
_mm256_setr_epi16 AVX immintrin.h (__m256i _mm256_setr_epi16(short, short, short, short, short, short, short, short, short, short, short, short, short, short, short, short);
_mm256_setr_epi32 AVX immintrin.h __m256i _mm256_setr_epi32(int, int, int, int, int, int, int, int);
_mm256_setr_epi64x AVX immintrin.h __m256i _mm256_setr_epi64x(long long, long long, long long, long long);
_mm256_setr_epi8 AVX immintrin.h (__m256i _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 AVX immintrin.h __m256d _mm256_setr_pd(double, double, double, double);
_mm256_setr_ps AVX immintrin.h __m256 _mm256_setr_ps(float, float, float, float, float, float, float, float);
_mm256_setzero_pd AVX immintrin.h __m256d _mm256_setzero_pd(void);
_mm256_setzero_ps AVX immintrin.h __m256 _mm256_setzero_ps(void);
_mm256_setzero_si256 AVX immintrin.h __m256i _mm256_setzero_si256(void);
_mm256_shuffle_epi32 AVX2 immintrin.h __m256i _mm256_shuffle_epi32(__m256i, const int);
_mm256_shuffle_epi8 AVX2 immintrin.h __m256i _mm256_shuffle_epi8(__m256i, __m256i);
_mm256_shuffle_pd AVX immintrin.h __m256d _mm256_shuffle_pd(__m256d, __m256d, const int);
_mm256_shuffle_ps AVX immintrin.h __m256 _mm256_shuffle_ps(__m256, __m256, const int);
_mm256_shufflehi_epi16 AVX2 immintrin.h __m256i _mm256_shufflehi_epi16(__m256i, const int);
_mm256_shufflelo_epi16 AVX2 immintrin.h __m256i _mm256_shufflelo_epi16(__m256i, const int);
_mm256_sign_epi16 AVX2 immintrin.h __m256i _mm256_sign_epi16(__m256i, __m256i);
_mm256_sign_epi32 AVX2 immintrin.h __m256i _mm256_sign_epi32(__m256i, __m256i);
_mm256_sign_epi8 AVX2 immintrin.h __m256i _mm256_sign_epi8(__m256i, __m256i);
_mm256_sll_epi16 AVX2 immintrin.h __m256i _mm256_sll_epi16(__m256i, __m128i);
_mm256_sll_epi32 AVX2 immintrin.h __m256i _mm256_sll_epi32(__m256i, __m128i);
_mm256_sll_epi64 AVX2 immintrin.h __m256i _mm256_sll_epi64(__m256i, __m128i);
_mm256_slli_epi16 AVX2 immintrin.h __m256i _mm256_slli_epi16(__m256i, int);
_mm256_slli_epi32 AVX2 immintrin.h __m256i _mm256_slli_epi32(__m256i, int);
_mm256_slli_epi64 AVX2 immintrin.h __m256i _mm256_slli_epi64(__m256i, int);
_mm256_slli_si256 AVX2 immintrin.h __m256i _mm256_slli_si256(__m256i, int);
_mm256_sllv_epi32 AVX2 immintrin.h __m256i _mm256_sllv_epi32(__m256i, __m256i);
_mm256_sllv_epi64 AVX2 immintrin.h __m256i _mm256_sllv_epi64(__m256i, __m256i);
_mm256_sqrt_pd AVX immintrin.h __m256d _mm256_sqrt_pd(__m256d);
_mm256_sqrt_ps AVX immintrin.h __m256 _mm256_sqrt_ps(__m256);
_mm256_sra_epi16 AVX2 immintrin.h __m256i _mm256_sra_epi16(__m256i, __m128i);
_mm256_sra_epi32 AVX2 immintrin.h __m256i _mm256_sra_epi32(__m256i, __m128i);
_mm256_srai_epi16 AVX2 immintrin.h __m256i _mm256_srai_epi16(__m256i, int);
_mm256_srai_epi32 AVX2 immintrin.h __m256i _mm256_srai_epi32(__m256i, int);
_mm256_srav_epi32 AVX2 immintrin.h __m256i _mm256_srav_epi32(__m256i, __m256i);
_mm256_srl_epi16 AVX2 immintrin.h __m256i _mm256_srl_epi16(__m256i, __m128i);
_mm256_srl_epi32 AVX2 immintrin.h __m256i _mm256_srl_epi32(__m256i, __m128i);
_mm256_srl_epi64 AVX2 immintrin.h __m256i _mm256_srl_epi64(__m256i, __m128i);
_mm256_srli_epi16 AVX2 immintrin.h __m256i _mm256_srli_epi16(__m256i, int);
_mm256_srli_epi32 AVX2 immintrin.h __m256i _mm256_srli_epi32(__m256i, int);
_mm256_srli_epi64 AVX2 immintrin.h __m256i _mm256_srli_epi64(__m256i, int);
_mm256_srli_si256 AVX2 immintrin.h __m256i _mm256_srli_si256(__m256i, int);
_mm256_srlv_epi32 AVX2 immintrin.h __m256i _mm256_srlv_epi32(__m256i, __m256i);
_mm256_srlv_epi64 AVX2 immintrin.h __m256i _mm256_srlv_epi64(__m256i, __m256i);
_mm256_store_pd AVX immintrin.h void _mm256_store_pd(double *, __m256d);
_mm256_store_ps AVX immintrin.h void _mm256_store_ps(float *, __m256);
_mm256_store_si256 AVX immintrin.h void _mm256_store_si256(__m256i *, __m256i);
_mm256_storeu_pd AVX immintrin.h void _mm256_storeu_pd(double *, __m256d);
_mm256_storeu_ps AVX immintrin.h void _mm256_storeu_ps(float *, __m256);
_mm256_storeu_si256 AVX immintrin.h void _mm256_storeu_si256(__m256i *, __m256i);
_mm256_stream_load_si256 AVX2 immintrin.h __m256i _mm256_stream_load_si256(__m256i const *);
_mm256_stream_pd AVX immintrin.h void __mm256_stream_pd(double *, __m256d);
_mm256_stream_ps AVX immintrin.h void _mm256_stream_ps(float *, __m256);
_mm256_stream_si256 AVX immintrin.h void __mm256_stream_si256(__m256i *, __m256i);
_mm256_sub_epi16 AVX2 immintrin.h __m256i _mm256_sub_epi16(__m256i, __m256i);
_mm256_sub_epi32 AVX2 immintrin.h __m256i _mm256_sub_epi32(__m256i, __m256i);
_mm256_sub_epi64 AVX2 immintrin.h __m256i _mm256_sub_epi64(__m256i, __m256i);
_mm256_sub_epi8 AVX2 immintrin.h __m256i _mm256_sub_epi8(__m256i, __m256i);
_mm256_sub_pd AVX immintrin.h __m256d _mm256_sub_pd(__m256d, __m256d);
_mm256_sub_ps AVX immintrin.h __m256 _mm256_sub_ps(__m256, __m256);
_mm256_subs_epi16 AVX2 immintrin.h __m256i _mm256_subs_epi16(__m256i, __m256i);
_mm256_subs_epi8 AVX2 immintrin.h __m256i _mm256_subs_epi8(__m256i, __m256i);
_mm256_subs_epu16 AVX2 immintrin.h __m256i _mm256_subs_epu16(__m256i, __m256i);
_mm256_subs_epu8 AVX2 immintrin.h __m256i _mm256_subs_epu8(__m256i, __m256i);
_mm256_testc_pd AVX immintrin.h int _mm256_testc_pd(__m256d, __m256d);
_mm256_testc_ps AVX immintrin.h int _mm256_testc_ps(__m256, __m256);
_mm256_testc_si256 AVX immintrin.h int _mm256_testc_si256(__m256i, __m256i);
_mm256_testnzc_pd AVX immintrin.h int _mm256_testnzc_pd(__m256d, __m256d);
_mm256_testnzc_ps AVX immintrin.h int _mm256_testnzc_ps(__m256, __m256);
_mm256_testnzc_si256 AVX immintrin.h int _mm256_testnzc_si256(__m256i, __m256i);
_mm256_testz_pd AVX immintrin.h int _mm256_testz_pd(__m256d, __m256d);
_mm256_testz_ps AVX immintrin.h int _mm256_testz_ps(__m256, __m256);
_mm256_testz_si256 AVX immintrin.h int _mm256_testz_si256(__m256i, __m256i);
_mm256_unpackhi_epi16 AVX2 immintrin.h __m256i _mm256_unpackhi_epi16(__m256i, __m256i);
_mm256_unpackhi_epi32 AVX2 immintrin.h __m256i _mm256_unpackhi_epi32(__m256i, __m256i);
_mm256_unpackhi_epi64 AVX2 immintrin.h __m256i _mm256_unpackhi_epi64(__m256i, __m256i);
_mm256_unpackhi_epi8 AVX2 immintrin.h __m256i _mm256_unpackhi_epi8(__m256i, __m256i);
_mm256_unpackhi_pd AVX immintrin.h __m256d _mm256_unpackhi_pd(__m256d, __m256d);
_mm256_unpackhi_ps AVX immintrin.h __m256 _mm256_unpackhi_ps(__m256, __m256);
_mm256_unpacklo_epi16 AVX2 immintrin.h __m256i _mm256_unpacklo_epi16(__m256i, __m256i);
_mm256_unpacklo_epi32 AVX2 immintrin.h __m256i _mm256_unpacklo_epi32(__m256i, __m256i);
_mm256_unpacklo_epi64 AVX2 immintrin.h __m256i _mm256_unpacklo_epi64(__m256i, __m256i);
_mm256_unpacklo_epi8 AVX2 immintrin.h __m256i _mm256_unpacklo_epi8(__m256i, __m256i);
_mm256_unpacklo_pd AVX immintrin.h __m256d _mm256_unpacklo_pd(__m256d, __m256d);
_mm256_unpacklo_ps AVX immintrin.h __m256 _mm256_unpacklo_ps(__m256, __m256);
_mm256_xor_pd AVX immintrin.h __m256d _mm256_xor_pd(__m256d, __m256d);
_mm256_xor_ps AVX immintrin.h __m256 _mm256_xor_ps(__m256, __m256);
_mm256_xor_si256 AVX2 immintrin.h __m256i _mm256_xor_si256(__m256i, __m256i);
_mm256_zeroall AVX immintrin.h void _mm256_zeroall(void);
_mm256_zeroupper AVX immintrin.h void _mm256_zeroupper(void);
__movsb intrin.h VOID __movsb(unsigned char *, unsigned char const *, size_t);
__movsd intrin.h VOID __movsd(unsigned long *, unsigned long const *, size_t);
__movsq intrin.h VOID __movsq(unsigned __int64 *, unsigned __int64 const *, size_t);
__movsw intrin.h VOID __movsw(unsigned short *, unsigned short const *, size_t);
_mul128 intrin.h __int64 _mul128(__int64, __int64, __int64 *);
__mulh intrin.h __int64 __mulh(__int64, __int64);
_mulx_u32 BMI immintrin.h unsigned int _mulx_u32(unsigned int, unsigned int, unsigned int*);
_mulx_u64 BMI immintrin.h unsigned __int64 _mulx_u64(unsigned __int64, unsigned __int64, unsigned __int64*);
__nop intrin.h void __nop(void);
__nvreg_restore_fence intrin.h void __nvreg_restore_fence(void);
__nvreg_save_fence intrin.h void __nvreg_save_fence(void);
__outbyte intrin.h void __outbyte(unsigned short, unsigned char);
__outbytestring intrin.h void __outbytestring(unsigned short, unsigned char *, unsigned long);
__outdword intrin.h void __outdword(unsigned short, unsigned long);
__outdwordstring intrin.h void __outdwordstring(unsigned short, unsigned long *, unsigned long);
__outword intrin.h void __outword(unsigned short, unsigned short);
__outwordstring intrin.h void __outwordstring(unsigned short, unsigned short *, unsigned long);
_pdep_u32 BMI immintrin.h unsigned int _pdep_u32(unsigned int, unsigned int);
_pdep_u64 BMI immintrin.h unsigned __int64 _pdep_u64(unsigned __int64, unsigned __int64);
_pext_u32 BMI immintrin.h unsigned int _pext_u32(unsigned int, unsigned int);
_pext_u64 BMI immintrin.h unsigned __int64 _pext_u64(unsigned __int64, unsigned __int64);
__popcnt POPCNT intrin.h unsigned int __popcnt(unsigned int);
__popcnt16 POPCNT intrin.h unsigned short __popcnt16(unsigned short);
__popcnt64 POPCNT intrin.h unsigned __int64 __popcnt64(unsigned __int64);
_rdrand16_step RDRAND immintrin.h int _rdrand16_step(unsigned short *);
_rdrand32_step RDRAND immintrin.h int _rdrand32_step(unsigned int *);
_rdrand64_step RDRAND immintrin.h int _rdrand64_step(unsigned __int64 *);
_rdseed16_step RDSEED immintrin.h int _rdseed16_step(unsigned short *);
_rdseed32_step RDSEED immintrin.h int _rdseed32_step(unsigned int *);
_rdseed64_step RDSEED immintrin.h int _rdseed64_step(unsigned __int64 *);
__rdtsc intrin.h unsigned __int64 __rdtsc(void);
__rdtscp RDTSCP intrin.h unsigned __int64 __rdtscp(unsigned int*);
_ReadBarrier intrin.h void _ReadBarrier(void);
__readcr0 intrin.h unsigned __int64 __readcr0(void);
__readcr2 intrin.h unsigned __int64 __readcr2(void);
__readcr3 intrin.h unsigned __int64 __readcr3(void);
__readcr4 intrin.h unsigned __int64 __readcr4(void);
__readcr8 intrin.h unsigned __int64 __readcr8(void);
__readdr intrin.h unsigned __int64 __readdr(unsigned);
__readeflags intrin.h unsigned __int64 __readeflags(void);
_readfsbase_u32 FSGSBASE immintrin.h unsigned int _readfsbase_u32(void);
_readfsbase_u64 FSGSBASE immintrin.h unsigned __int64 _readfsbase_u64(void);
_readgsbase_u32 FSGSBASE immintrin.h unsigned int _readgsbase_u32(void);
_readgsbase_u64 FSGSBASE immintrin.h unsigned __int64 _readgsbase_u64(void);
__readgsbyte intrin.h unsigned char __readgsbyte(unsigned long);
__readgsdword intrin.h unsigned long __readgsdword(unsigned long);
__readgsqword intrin.h unsigned __int64 __readgsqword(unsigned long);
__readgsword intrin.h unsigned short __readgsword(unsigned long);
__readmsr intrin.h unsigned __int64 __readmsr(unsigned long);
__readpmc intrin.h unsigned __int64 __readpmc(unsigned long);
_ReadWriteBarrier intrin.h void _ReadWriteBarrier(void);
_ReturnAddress intrin.h void * _ReturnAddress(void);
_rorx_u32 BMI immintrin.h unsigned int _rorx_u32(unsigned int, const unsigned int);
_rorx_u64 BMI immintrin.h unsigned __int64 _rorx_u64(unsigned __int64, const unsigned int);
_rotl16 intrin.h unsigned short _rotl16(unsigned short, unsigned char);
_rotl8 intrin.h unsigned char _rotl8(unsigned char, unsigned char);
_rotr16 intrin.h unsigned short _rotr16(unsigned short, unsigned char);
_rotr8 intrin.h unsigned char _rotr8(unsigned char, unsigned char);
_rsm intrin.h void _rsm(void);
_sarx_i32 BMI immintrin.h int _sarx_i32(int, unsigned int);
_sarx_i64 BMI immintrin.h __int64 _sarx_i64(__int64, unsigned int);
__segmentlimit intrin.h unsigned long __segmentlimit(unsigned long);
_sgdt intrin.h void _sgdt(void*);
__shiftleft128 intrin.h unsigned __int64 __shiftleft128(unsigned __int64, unsigned __int64, unsigned char);
__shiftright128 intrin.h unsigned __int64 __shiftright128(unsigned __int64, unsigned __int64, unsigned char);
_shlx_u32 BMI immintrin.h unsigned int _shlx_u32(unsigned int, unsigned int);
_shlx_u64 BMI immintrin.h unsigned __int64 _shlx_u64(unsigned __int64, unsigned int);
_shrx_u32 BMI immintrin.h unsigned int _shrx_u32(unsigned int, unsigned int);
_shrx_u64 BMI immintrin.h unsigned __int64 _shrx_u64(unsigned __int64, unsigned int);
__sidt intrin.h void __sidt(void*);
__slwpcb LWP ammintrin.h void *__slwpcb(void);
_stac SMAP intrin.h void _stac(void);
_storebe_i16 MOVBE immintrin.h void _storebe_i16(void *, short); [Makro]
_storebe_i32 MOVBE immintrin.h void _storebe_i32(void *, int); [Makro]
_storebe_i64 MOVBE immintrin.h void _storebe_i64(void *, __int64); [Makro]
_store_be_u16 MOVBE immintrin.h void _store_be_u16(void *, unsigned short); [Makro]
_store_be_u32 MOVBE immintrin.h void _store_be_u32(void *, unsigned int); [Makro]
_store_be_u64 MOVBE immintrin.h void _store_be_u64(void *, unsigned __int64); [Makro]
_Store_HLERelease HLE immintrin.h void _Store_HLERelease(long volatile *, long);
_Store64_HLERelease HLE immintrin.h void _Store64_HLERelease(__int64 volatile *, __int64);
_StorePointer_HLERelease HLE immintrin.h void _StorePointer_HLERelease(void * volatile *, void *);
__stosb intrin.h void __stosb(unsigned char *, unsigned char, size_t);
__stosd intrin.h void __stosd(unsigned long *, unsigned long, size_t);
__stosq intrin.h void __stosq(unsigned __int64 *, unsigned __int64, size_t);
__stosw intrin.h void __stosw(unsigned short *, unsigned short, size_t);
_subborrow_u16 intrin.h unsigned char _subborrow_u16(unsigned char, unsigned short, unsigned short, unsigned short *);
_subborrow_u32 intrin.h unsigned char _subborrow_u32(unsigned char, unsigned int, unsigned int, unsigned int *);
_subborrow_u64 intrin.h unsigned char _subborrow_u64(unsigned char, unsigned __int64, unsigned __int64, unsigned __int64 *);
_subborrow_u8 intrin.h unsigned char _subborrow_u8(unsigned char, unsigned char, unsigned char, unsigned char *);
__svm_clgi intrin.h void __svm_clgi(void);
__svm_invlpga intrin.h void __svm_invlpga(void*, int);
__svm_skinit intrin.h void __svm_skinit(int);
__svm_stgi intrin.h void __svm_stgi(void);
__svm_vmload intrin.h void __svm_vmload(size_t);
__svm_vmrun intrin.h void __svm_vmrun(size_t);
__svm_vmsave intrin.h void __svm_vmsave(size_t);
_t1mskc_u32 ABM ammintrin.h unsigned int _t1mskc_u32(unsigned int);
_t1mskc_u64 ABM ammintrin.h unsigned __int64 _t1mskc_u64(unsigned __int64);
_tzcnt_u32 BMI ammintrin.h, immintrin.h unsigned int _tzcnt_u32(unsigned int);
_tzcnt_u64 BMI ammintrin.h, immintrin.h unsigned __int64 _tzcnt_u64(unsigned __int64);
_tzmsk_u32 ABM ammintrin.h unsigned int _tzmsk_u32(unsigned int);
_tzmsk_u64 ABM ammintrin.h unsigned __int64 _tzmsk_u64(unsigned __int64);
__ud2 intrin.h void __ud2(void);
_udiv128 intrin.h unsigned __int64 _udiv128(unsigned __int64, unsigned __int64, unsigned __int64, unsigned __int64 *);
_udiv64 intrin.h unsigned int _udiv64(unsigned __int64, unsigned int, unsigned int*);
__ull_rshift intrin.h unsigned __int64 [pascal/cdecl] __ull_rshift(unsigned __int64, int);
_umul128 intrin.h unsigned __int64 _umul128(unsigned __int64, unsigned __int64, unsigned __int64 *);
__umulh intrin.h unsigned __int64 __umulh(unsigned __int64, unsigned __int64);
__vmx_off intrin.h void __vmx_off(void);
__vmx_on intrin.h unsigned char __vmx_on(unsigned __int64*);
__vmx_vmclear intrin.h unsigned char __vmx_vmclear(unsigned __int64*);
__vmx_vmlaunch intrin.h unsigned char __vmx_vmlaunch(void);
__vmx_vmptrld intrin.h unsigned char __vmx_vmptrld(unsigned __int64*);
__vmx_vmptrst intrin.h void __vmx_vmptrst(unsigned __int64 *);
__vmx_vmread intrin.h unsigned char __vmx_vmread(size_t, size_t*);
__vmx_vmresume intrin.h unsigned char __vmx_vmresume(void);
__vmx_vmwrite intrin.h unsigned char __vmx_vmwrite(size_t, size_t);
__wbinvd intrin.h void __wbinvd(void);
_WriteBarrier intrin.h void _WriteBarrier(void);
__writecr0 intrin.h void __writecr0(unsigned __int64);
__writecr3 intrin.h void __writecr3(unsigned __int64);
__writecr4 intrin.h void __writecr4(unsigned __int64);
__writecr8 intrin.h void __writecr8(unsigned __int64);
__writedr intrin.h void __writedr(unsigned, unsigned __int64);
__writeeflags intrin.h void __writeeflags(unsigned __int64);
_writefsbase_u32 FSGSBASE immintrin.h void _writefsbase_u32(unsigned int);
_writefsbase_u64 FSGSBASE immintrin.h void _writefsbase_u64(unsigned __int64);
_writegsbase_u32 FSGSBASE immintrin.h void _writegsbase_u32(unsigned int);
_writegsbase_u64 FSGSBASE immintrin.h void _writegsbase_u64(unsigned __int64);
__writegsbyte intrin.h void __writegsbyte(unsigned long, unsigned char);
__writegsdword intrin.h void __writegsdword(unsigned long, unsigned long);
__writegsqword intrin.h void __writegsqword(unsigned long, unsigned __int64);
__writegsword intrin.h void __writegsword(unsigned long, unsigned short);
__writemsr intrin.h void __writemsr(unsigned long, unsigned __int64);
_xabort RTM immintrin.h void _xabort(unsigned int);
_xbegin RTM immintrin.h unsigned _xbegin(void);
_xend RTM immintrin.h void _xend(void);
_xgetbv XSAVE immintrin.h unsigned __int64 _xgetbv(unsigned int);
_xrstor XSAVE immintrin.h void _xrstor(void const*, unsigned __int64);
_xrstor64 XSAVE immintrin.h void _xrstor64(void const*, unsigned __int64);
_xsave XSAVE immintrin.h void _xsave(void*, unsigned __int64);
_xsave64 XSAVE immintrin.h void _xsave64(void*, unsigned __int64);
_xsaveopt XSAVEOPT immintrin.h void _xsaveopt(void*, unsigned __int64);
_xsaveopt64 XSAVEOPT immintrin.h void _xsaveopt64(void*, unsigned __int64);
_xsetbv XSAVE immintrin.h void _xsetbv(unsigned int, unsigned __int64);
_xtest XTEST immintrin.h unsigned char _xtest(void);

Siehe auch

Intrinsische Compilerfunktionen
Systeminterne ARM-Funktionen
Systeminterne ARM64-Funktionen
x86 systeminterne Daten