|
|
|
| 1 |
|
/** |
| 2 |
|
*** Copyright (C) 1985-1999 Intel Corporation. All rights reserved. |
| 3 |
|
*** |
| 4 |
|
*** The information and source code contained herein is the exclusive |
| 5 |
|
*** property of Intel Corporation and may not be disclosed, examined |
| 6 |
|
*** or reproduced in whole or in part without explicit written authorization |
| 7 |
|
*** from the company. |
| 8 |
|
*** |
| 9 |
|
**/ |
| 10 |
|
|
| 11 |
|
/* |
| 12 |
|
* Definitions and declarations for use with compiler intrinsics. |
| 13 |
|
*/ |
| 14 |
|
|
| 15 |
|
#ifndef _MMINTRIN_H_INCLUDED |
| 16 |
|
#define _MMINTRIN_H_INCLUDED |
| 17 |
|
#ifndef __midl |
| 18 |
|
|
| 19 |
|
#include <crtdefs.h> |
| 20 |
|
|
| 21 |
|
#if defined(_M_CEE_PURE) |
| 22 |
|
#error ERROR: MM intrinsics not supported in the pure mode! |
| 23 |
|
#else |
| 24 |
|
|
| 25 |
|
#if defined __cplusplus |
| 26 |
|
extern "C" { /* Begin "C" */ |
| 27 |
|
/* Intrinsics use C name-mangling. |
| 28 |
|
*/ |
| 29 |
|
#endif /* __cplusplus */ |
| 30 |
|
|
| 31 |
|
typedef union __declspec(intrin_type) _CRT_ALIGN(8) __m64 |
| 32 |
|
{ |
| 33 |
|
unsigned __int64 m64_u64; |
| 34 |
|
float m64_f32[2]; |
| 35 |
|
__int8 m64_i8[8]; |
| 36 |
|
__int16 m64_i16[4]; |
| 37 |
|
__int32 m64_i32[2]; |
| 38 |
|
__int64 m64_i64; |
| 39 |
|
unsigned __int8 m64_u8[8]; |
| 40 |
|
unsigned __int16 m64_u16[4]; |
| 41 |
|
unsigned __int32 m64_u32[2]; |
| 42 |
|
} __m64; |
| 43 |
|
|
| 44 |
|
/* General support intrinsics */ |
| 45 |
|
void _m_empty(void); |
| 46 |
|
__m64 _m_from_int(int _I); |
| 47 |
|
int _m_to_int(__m64 _M); |
| 48 |
|
__m64 _m_packsswb(__m64 _MM1, __m64 _MM2); |
| 49 |
|
__m64 _m_packssdw(__m64 _MM1, __m64 _MM2); |
| 50 |
|
__m64 _m_packuswb(__m64 _MM1, __m64 _MM2); |
| 51 |
|
__m64 _m_punpckhbw(__m64 _MM1, __m64 _MM2); |
| 52 |
|
__m64 _m_punpckhwd(__m64 _MM1, __m64 _MM2); |
| 53 |
|
__m64 _m_punpckhdq(__m64 _MM1, __m64 _MM2); |
| 54 |
|
__m64 _m_punpcklbw(__m64 _MM1, __m64 _MM2); |
| 55 |
|
__m64 _m_punpcklwd(__m64 _MM1, __m64 _MM2); |
| 56 |
|
__m64 _m_punpckldq(__m64 _MM1, __m64 _MM2); |
| 57 |
|
|
| 58 |
|
/* Packed arithmetic intrinsics */ |
| 59 |
|
__m64 _m_paddb(__m64 _MM1, __m64 _MM2); |
| 60 |
|
__m64 _m_paddw(__m64 _MM1, __m64 _MM2); |
| 61 |
|
__m64 _m_paddd(__m64 _MM1, __m64 _MM2); |
| 62 |
|
__m64 _m_paddsb(__m64 _MM1, __m64 _MM2); |
| 63 |
|
__m64 _m_paddsw(__m64 _MM1, __m64 _MM2); |
| 64 |
|
__m64 _m_paddusb(__m64 _MM1, __m64 _MM2); |
| 65 |
|
__m64 _m_paddusw(__m64 _MM1, __m64 _MM2); |
| 66 |
|
__m64 _m_psubb(__m64 _MM1, __m64 _MM2); |
| 67 |
|
__m64 _m_psubw(__m64 _MM1, __m64 _MM2); |
| 68 |
|
__m64 _m_psubd(__m64 _MM1, __m64 _MM2); |
| 69 |
|
__m64 _m_psubsb(__m64 _MM1, __m64 _MM2); |
| 70 |
|
__m64 _m_psubsw(__m64 _MM1, __m64 _MM2); |
| 71 |
|
__m64 _m_psubusb(__m64 _MM1, __m64 _MM2); |
| 72 |
|
__m64 _m_psubusw(__m64 _MM1, __m64 _MM2); |
| 73 |
|
__m64 _m_pmaddwd(__m64 _MM1, __m64 _MM2); |
| 74 |
|
__m64 _m_pmulhw(__m64 _MM1, __m64 _MM2); |
| 75 |
|
__m64 _m_pmullw(__m64 _MM1, __m64 _MM2); |
| 76 |
|
|
| 77 |
|
/* Shift intrinsics */ |
| 78 |
|
__m64 _m_psllw(__m64 _M, __m64 _Count); |
| 79 |
|
__m64 _m_psllwi(__m64 _M, int _Count); |
| 80 |
|
__m64 _m_pslld(__m64 _M, __m64 _Count); |
| 81 |
|
__m64 _m_pslldi(__m64 _M, int _Count); |
| 82 |
|
__m64 _m_psllq(__m64 _M, __m64 _Count); |
| 83 |
|
__m64 _m_psllqi(__m64 _M, int _Count); |
| 84 |
|
__m64 _m_psraw(__m64 _M, __m64 _Count); |
| 85 |
|
__m64 _m_psrawi(__m64 _M, int _Count); |
| 86 |
|
__m64 _m_psrad(__m64 _M, __m64 _Count); |
| 87 |
|
__m64 _m_psradi(__m64 _M, int _Count); |
| 88 |
|
__m64 _m_psrlw(__m64 _M, __m64 _Count); |
| 89 |
|
__m64 _m_psrlwi(__m64 _M, int _Count); |
| 90 |
|
__m64 _m_psrld(__m64 _M, __m64 _Count); |
| 91 |
|
__m64 _m_psrldi(__m64 _M, int _Count); |
| 92 |
|
__m64 _m_psrlq(__m64 _M, __m64 _Count); |
| 93 |
|
__m64 _m_psrlqi(__m64 _M, int _Count); |
| 94 |
|
|
| 95 |
|
/* Logical intrinsics */ |
| 96 |
|
__m64 _m_pand(__m64 _MM1, __m64 _MM2); |
| 97 |
|
__m64 _m_pandn(__m64 _MM1, __m64 _MM2); |
| 108 |
|
|
| 109 |
|
/* Utility intrinsics */ |
| 110 |
|
__m64 _mm_setzero_si64(void); |
| 111 |
|
__m64 _mm_set_pi32(int _I1, int _I0); |
| 112 |
|
__m64 _mm_set_pi16(short _S3, short _S2, short _S1, short _S0); |
| 113 |
|
__m64 _mm_set_pi8(char _B7, char _B6, char _B5, char _B4, |
| 114 |
|
char _B3, char _B2, char _B1, char _B0); |
| 115 |
|
__m64 _mm_set1_pi32(int _I); |
| 116 |
|
__m64 _mm_set1_pi16(short _S); |
| 117 |
|
__m64 _mm_set1_pi8(char _B); |
| 118 |
|
__m64 _mm_setr_pi32(int _I1, int _I0); |
| 119 |
|
__m64 _mm_setr_pi16(short _S3, short _S2, short _S1, short _S0); |
| 120 |
|
__m64 _mm_setr_pi8(char _B7, char _B6, char _B5, char _B4, |
| 121 |
|
char _B3, char _B2, char _B1, char _B0); |
| 122 |
|
|
| 123 |
|
/* Alternate intrinsic name definitions */ |
| 124 |
|
#define _mm_empty _m_empty |
| 125 |
|
#define _mm_cvtsi32_si64 _m_from_int |
| 126 |
|
#define _mm_cvtsi64_si32 _m_to_int |
| 127 |
|
#define _mm_packs_pi16 _m_packsswb |
| 128 |
|
#define _mm_packs_pi32 _m_packssdw |
| 129 |
|
#define _mm_packs_pu16 _m_packuswb |
| 130 |
|
#define _mm_unpackhi_pi8 _m_punpckhbw |
| 131 |
|
#define _mm_unpackhi_pi16 _m_punpckhwd |
| 132 |
|
#define _mm_unpackhi_pi32 _m_punpckhdq |
| 133 |
|
#define _mm_unpacklo_pi8 _m_punpcklbw |
| 134 |
|
#define _mm_unpacklo_pi16 _m_punpcklwd |
| 135 |
|
#define _mm_unpacklo_pi32 _m_punpckldq |
| 136 |
|
#define _mm_add_pi8 _m_paddb |
| 137 |
|
#define _mm_add_pi16 _m_paddw |
| 138 |
|
#define _mm_add_pi32 _m_paddd |
| 139 |
|
#define _mm_adds_pi8 _m_paddsb |
| 140 |
|
#define _mm_adds_pi16 _m_paddsw |
| 141 |
|
#define _mm_adds_pu8 _m_paddusb |
| 142 |
|
#define _mm_adds_pu16 _m_paddusw |
| 143 |
|
#define _mm_sub_pi8 _m_psubb |
| 144 |
|
#define _mm_sub_pi16 _m_psubw |
| 145 |
|
#define _mm_sub_pi32 _m_psubd |
| 146 |
|
#define _mm_subs_pi8 _m_psubsb |
| 147 |
|
#define _mm_subs_pi16 _m_psubsw |
| 148 |
|
#define _mm_subs_pu8 _m_psubusb |
| 149 |
|
#define _mm_subs_pu16 _m_psubusw |
| 150 |
|
#define _mm_madd_pi16 _m_pmaddwd |
| 151 |
|
#define _mm_mulhi_pi16 _m_pmulhw |
| 152 |
|
#define _mm_mullo_pi16 _m_pmullw |
| 153 |
|
#define _mm_sll_pi16 _m_psllw |
| 154 |
|
#define _mm_slli_pi16 _m_psllwi |
| 155 |
|
#define _mm_sll_pi32 _m_pslld |
| 156 |
|
#define _mm_slli_pi32 _m_pslldi |
| 157 |
|
#define _mm_sll_si64 _m_psllq |
| 158 |
|
#define _mm_slli_si64 _m_psllqi |
| 159 |
|
#define _mm_sra_pi16 _m_psraw |
| 160 |
|
#define _mm_srai_pi16 _m_psrawi |
| 161 |
|
#define _mm_sra_pi32 _m_psrad |
| 162 |
|
#define _mm_srai_pi32 _m_psradi |
| 163 |
|
#define _mm_srl_pi16 _m_psrlw |
| 164 |
|
#define _mm_srli_pi16 _m_psrlwi |
| 165 |
|
#define _mm_srl_pi32 _m_psrld |
| 166 |
|
#define _mm_srli_pi32 _m_psrldi |
| 167 |
|
#define _mm_srl_si64 _m_psrlq |
| 168 |
|
#define _mm_srli_si64 _m_psrlqi |
| 169 |
|
#define _mm_and_si64 _m_pand |
| 170 |
|
#define _mm_andnot_si64 _m_pandn |
| 171 |
|
#define _mm_or_si64 _m_por |
| 172 |
|
#define _mm_xor_si64 _m_pxor |
| 173 |
|
#define _mm_cmpeq_pi8 _m_pcmpeqb |
| 174 |
|
#define _mm_cmpeq_pi16 _m_pcmpeqw |
| 175 |
|
#define _mm_cmpeq_pi32 _m_pcmpeqd |
| 176 |
|
#define _mm_cmpgt_pi8 _m_pcmpgtb |
| 177 |
|
#define _mm_cmpgt_pi16 _m_pcmpgtw |
| 178 |
|
#define _mm_cmpgt_pi32 _m_pcmpgtd |
| 179 |
|
|
| 180 |
|
|
| 181 |
|
#if defined __cplusplus |
| 182 |
|
}; /* End "C" */ |
| 183 |
|
#endif /* __cplusplus */ |
| 184 |
|
|
| 185 |
|
#endif /* defined(_M_CEE_PURE) */ |
| 186 |
|
#endif |
| 187 |
|
#endif /* _MMINTRIN_H_INCLUDED */ |
| 188 |
|
|
| 189 |
|
|
|
|
|