Mercurial > hg > CbC > CbC_gcc
comparison gcc/config/arm/arm_neon.h @ 131:84e7813d76e9
gcc-8.2
| author | mir3636 |
|---|---|
| date | Thu, 25 Oct 2018 07:37:49 +0900 |
| parents | 04ced10e8804 |
| children | 1830386684a0 |
comparison
equal
deleted
inserted
replaced
| 111:04ced10e8804 | 131:84e7813d76e9 |
|---|---|
| 1 /* ARM NEON intrinsics include file. | 1 /* ARM NEON intrinsics include file. |
| 2 | 2 |
| 3 Copyright (C) 2006-2017 Free Software Foundation, Inc. | 3 Copyright (C) 2006-2018 Free Software Foundation, Inc. |
| 4 Contributed by CodeSourcery. | 4 Contributed by CodeSourcery. |
| 5 | 5 |
| 6 This file is part of GCC. | 6 This file is part of GCC. |
| 7 | 7 |
| 8 GCC is free software; you can redistribute it and/or modify it | 8 GCC is free software; you can redistribute it and/or modify it |
| 18032 return __rv; | 18032 return __rv; |
| 18033 } | 18033 } |
| 18034 | 18034 |
| 18035 #endif | 18035 #endif |
| 18036 | 18036 |
| 18037 /* AdvSIMD Dot Product intrinsics. */ | |
| 18038 | |
| 18039 #if __ARM_ARCH >= 8 | |
| 18040 #pragma GCC push_options | |
| 18041 #pragma GCC target ("arch=armv8.2-a+dotprod") | |
| 18042 | |
| 18043 __extension__ extern __inline uint32x2_t | |
| 18044 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18045 vdot_u32 (uint32x2_t __r, uint8x8_t __a, uint8x8_t __b) | |
| 18046 { | |
| 18047 return __builtin_neon_udotv8qi_uuuu (__r, __a, __b); | |
| 18048 } | |
| 18049 | |
| 18050 __extension__ extern __inline uint32x4_t | |
| 18051 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18052 vdotq_u32 (uint32x4_t __r, uint8x16_t __a, uint8x16_t __b) | |
| 18053 { | |
| 18054 return __builtin_neon_udotv16qi_uuuu (__r, __a, __b); | |
| 18055 } | |
| 18056 | |
| 18057 __extension__ extern __inline int32x2_t | |
| 18058 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18059 vdot_s32 (int32x2_t __r, int8x8_t __a, int8x8_t __b) | |
| 18060 { | |
| 18061 return __builtin_neon_sdotv8qi (__r, __a, __b); | |
| 18062 } | |
| 18063 | |
| 18064 __extension__ extern __inline int32x4_t | |
| 18065 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18066 vdotq_s32 (int32x4_t __r, int8x16_t __a, int8x16_t __b) | |
| 18067 { | |
| 18068 return __builtin_neon_sdotv16qi (__r, __a, __b); | |
| 18069 } | |
| 18070 | |
| 18071 __extension__ extern __inline uint32x2_t | |
| 18072 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18073 vdot_lane_u32 (uint32x2_t __r, uint8x8_t __a, uint8x8_t __b, const int __index) | |
| 18074 { | |
| 18075 return __builtin_neon_udot_lanev8qi_uuuus (__r, __a, __b, __index); | |
| 18076 } | |
| 18077 | |
| 18078 __extension__ extern __inline uint32x4_t | |
| 18079 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18080 vdotq_lane_u32 (uint32x4_t __r, uint8x16_t __a, uint8x8_t __b, | |
| 18081 const int __index) | |
| 18082 { | |
| 18083 return __builtin_neon_udot_lanev16qi_uuuus (__r, __a, __b, __index); | |
| 18084 } | |
| 18085 | |
| 18086 __extension__ extern __inline int32x2_t | |
| 18087 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18088 vdot_lane_s32 (int32x2_t __r, int8x8_t __a, int8x8_t __b, const int __index) | |
| 18089 { | |
| 18090 return __builtin_neon_sdot_lanev8qi (__r, __a, __b, __index); | |
| 18091 } | |
| 18092 | |
| 18093 __extension__ extern __inline int32x4_t | |
| 18094 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18095 vdotq_lane_s32 (int32x4_t __r, int8x16_t __a, int8x8_t __b, const int __index) | |
| 18096 { | |
| 18097 return __builtin_neon_sdot_lanev16qi (__r, __a, __b, __index); | |
| 18098 } | |
| 18099 | |
| 18100 #pragma GCC pop_options | |
| 18101 #endif | |
| 18102 | |
| 18103 #if defined (__ARM_FP16_FORMAT_IEEE) || defined (__ARM_FP16_FORMAT_ALTERNATIVE) | |
| 18104 #pragma GCC push_options | |
| 18105 #pragma GCC target ("arch=armv8.2-a+fp16fml") | |
| 18106 | |
| 18107 __extension__ extern __inline float32x2_t | |
| 18108 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18109 vfmlal_low_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b) | |
| 18110 { | |
| 18111 return __builtin_neon_vfmal_lowv2sf (__r, __a, __b); | |
| 18112 } | |
| 18113 | |
| 18114 __extension__ extern __inline float32x2_t | |
| 18115 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18116 vfmlsl_low_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b) | |
| 18117 { | |
| 18118 return __builtin_neon_vfmsl_lowv2sf (__r, __a, __b); | |
| 18119 } | |
| 18120 | |
| 18121 __extension__ extern __inline float32x2_t | |
| 18122 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18123 vfmlal_high_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b) | |
| 18124 { | |
| 18125 return __builtin_neon_vfmal_highv2sf (__r, __a, __b); | |
| 18126 } | |
| 18127 | |
| 18128 __extension__ extern __inline float32x2_t | |
| 18129 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18130 vfmlsl_high_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b) | |
| 18131 { | |
| 18132 return __builtin_neon_vfmsl_highv2sf (__r, __a, __b); | |
| 18133 } | |
| 18134 | |
| 18135 __extension__ extern __inline float32x4_t | |
| 18136 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18137 vfmlalq_low_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b) | |
| 18138 { | |
| 18139 return __builtin_neon_vfmal_lowv4sf (__r, __a, __b); | |
| 18140 } | |
| 18141 | |
| 18142 __extension__ extern __inline float32x4_t | |
| 18143 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18144 vfmlslq_low_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b) | |
| 18145 { | |
| 18146 return __builtin_neon_vfmsl_lowv4sf (__r, __a, __b); | |
| 18147 } | |
| 18148 | |
| 18149 __extension__ extern __inline float32x4_t | |
| 18150 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18151 vfmlalq_high_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b) | |
| 18152 { | |
| 18153 return __builtin_neon_vfmal_highv4sf (__r, __a, __b); | |
| 18154 } | |
| 18155 | |
| 18156 __extension__ extern __inline float32x4_t | |
| 18157 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18158 vfmlslq_high_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b) | |
| 18159 { | |
| 18160 return __builtin_neon_vfmsl_highv4sf (__r, __a, __b); | |
| 18161 } | |
| 18162 | |
| 18163 __extension__ extern __inline float32x2_t | |
| 18164 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18165 vfmlal_lane_low_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b, | |
| 18166 const int __index) | |
| 18167 { | |
| 18168 __builtin_arm_lane_check (4, __index); | |
| 18169 return __builtin_neon_vfmal_lane_lowv2sf (__r, __a, __b, __index); | |
| 18170 } | |
| 18171 | |
| 18172 __extension__ extern __inline float32x2_t | |
| 18173 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18174 vfmlal_lane_high_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b, | |
| 18175 const int __index) | |
| 18176 { | |
| 18177 __builtin_arm_lane_check (4, __index); | |
| 18178 return __builtin_neon_vfmal_lane_highv2sf (__r, __a, __b, __index); | |
| 18179 } | |
| 18180 | |
| 18181 __extension__ extern __inline float32x4_t | |
| 18182 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18183 vfmlalq_laneq_low_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b, | |
| 18184 const int __index) | |
| 18185 { | |
| 18186 __builtin_arm_lane_check (8, __index); | |
| 18187 return __builtin_neon_vfmal_lane_lowv4sf (__r, __a, __b, __index); | |
| 18188 } | |
| 18189 | |
| 18190 __extension__ extern __inline float32x4_t | |
| 18191 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18192 vfmlalq_lane_low_u32 (float32x4_t __r, float16x8_t __a, float16x4_t __b, | |
| 18193 const int __index) | |
| 18194 { | |
| 18195 __builtin_arm_lane_check (4, __index); | |
| 18196 return __builtin_neon_vfmal_lane_lowv4hfv4sf (__r, __a, __b, __index); | |
| 18197 } | |
| 18198 | |
| 18199 __extension__ extern __inline float32x2_t | |
| 18200 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18201 vfmlal_laneq_low_u32 (float32x2_t __r, float16x4_t __a, float16x8_t __b, | |
| 18202 const int __index) | |
| 18203 { | |
| 18204 __builtin_arm_lane_check (8, __index); | |
| 18205 return __builtin_neon_vfmal_lane_lowv8hfv2sf (__r, __a, __b, __index); | |
| 18206 } | |
| 18207 | |
| 18208 __extension__ extern __inline float32x4_t | |
| 18209 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18210 vfmlalq_laneq_high_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b, | |
| 18211 const int __index) | |
| 18212 { | |
| 18213 __builtin_arm_lane_check (8, __index); | |
| 18214 return __builtin_neon_vfmal_lane_highv4sf (__r, __a, __b, __index); | |
| 18215 } | |
| 18216 | |
| 18217 __extension__ extern __inline float32x4_t | |
| 18218 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18219 vfmlalq_lane_high_u32 (float32x4_t __r, float16x8_t __a, float16x4_t __b, | |
| 18220 const int __index) | |
| 18221 { | |
| 18222 __builtin_arm_lane_check (4, __index); | |
| 18223 return __builtin_neon_vfmal_lane_highv4hfv4sf (__r, __a, __b, __index); | |
| 18224 } | |
| 18225 | |
| 18226 __extension__ extern __inline float32x2_t | |
| 18227 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18228 vfmlal_laneq_high_u32 (float32x2_t __r, float16x4_t __a, float16x8_t __b, | |
| 18229 const int __index) | |
| 18230 { | |
| 18231 __builtin_arm_lane_check (8, __index); | |
| 18232 return __builtin_neon_vfmal_lane_highv8hfv2sf (__r, __a, __b, __index); | |
| 18233 } | |
| 18234 | |
| 18235 __extension__ extern __inline float32x2_t | |
| 18236 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18237 vfmlsl_lane_low_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b, | |
| 18238 const int __index) | |
| 18239 { | |
| 18240 __builtin_arm_lane_check (4, __index); | |
| 18241 return __builtin_neon_vfmsl_lane_lowv2sf (__r, __a, __b, __index); | |
| 18242 } | |
| 18243 | |
| 18244 __extension__ extern __inline float32x2_t | |
| 18245 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18246 vfmlsl_lane_high_u32 (float32x2_t __r, float16x4_t __a, float16x4_t __b, | |
| 18247 const int __index) | |
| 18248 { | |
| 18249 __builtin_arm_lane_check (4, __index); | |
| 18250 return __builtin_neon_vfmsl_lane_highv2sf (__r, __a, __b, __index); | |
| 18251 } | |
| 18252 | |
| 18253 __extension__ extern __inline float32x4_t | |
| 18254 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18255 vfmlslq_laneq_low_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b, | |
| 18256 const int __index) | |
| 18257 { | |
| 18258 __builtin_arm_lane_check (8, __index); | |
| 18259 return __builtin_neon_vfmsl_lane_lowv4sf (__r, __a, __b, __index); | |
| 18260 } | |
| 18261 | |
| 18262 __extension__ extern __inline float32x4_t | |
| 18263 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18264 vfmlslq_lane_low_u32 (float32x4_t __r, float16x8_t __a, float16x4_t __b, | |
| 18265 const int __index) | |
| 18266 { | |
| 18267 __builtin_arm_lane_check (4, __index); | |
| 18268 return __builtin_neon_vfmsl_lane_lowv4hfv4sf (__r, __a, __b, __index); | |
| 18269 } | |
| 18270 | |
| 18271 __extension__ extern __inline float32x2_t | |
| 18272 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18273 vfmlsl_laneq_low_u32 (float32x2_t __r, float16x4_t __a, float16x8_t __b, | |
| 18274 const int __index) | |
| 18275 { | |
| 18276 __builtin_arm_lane_check (8, __index); | |
| 18277 return __builtin_neon_vfmsl_lane_lowv8hfv2sf (__r, __a, __b, __index); | |
| 18278 } | |
| 18279 | |
| 18280 __extension__ extern __inline float32x4_t | |
| 18281 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18282 vfmlslq_laneq_high_u32 (float32x4_t __r, float16x8_t __a, float16x8_t __b, | |
| 18283 const int __index) | |
| 18284 { | |
| 18285 __builtin_arm_lane_check (8, __index); | |
| 18286 return __builtin_neon_vfmsl_lane_highv4sf (__r, __a, __b, __index); | |
| 18287 } | |
| 18288 | |
| 18289 __extension__ extern __inline float32x4_t | |
| 18290 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18291 vfmlslq_lane_high_u32 (float32x4_t __r, float16x8_t __a, float16x4_t __b, | |
| 18292 const int __index) | |
| 18293 { | |
| 18294 __builtin_arm_lane_check (4, __index); | |
| 18295 return __builtin_neon_vfmsl_lane_highv4hfv4sf (__r, __a, __b, __index); | |
| 18296 } | |
| 18297 | |
| 18298 __extension__ extern __inline float32x2_t | |
| 18299 __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) | |
| 18300 vfmlsl_laneq_high_u32 (float32x2_t __r, float16x4_t __a, float16x8_t __b, | |
| 18301 const int __index) | |
| 18302 { | |
| 18303 __builtin_arm_lane_check (8, __index); | |
| 18304 return __builtin_neon_vfmsl_lane_highv8hfv2sf (__r, __a, __b, __index); | |
| 18305 } | |
| 18306 | |
| 18307 #pragma GCC pop_options | |
| 18308 #endif | |
| 18309 | |
| 18037 #ifdef __cplusplus | 18310 #ifdef __cplusplus |
| 18038 } | 18311 } |
| 18039 #endif | 18312 #endif |
| 18040 | 18313 |
| 18041 #pragma GCC pop_options | 18314 #pragma GCC pop_options |