1 /*===---- fma4intrin.h - FMA4 intrinsics -----------------------------------===
3 * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 * See https://llvm.org/LICENSE.txt for license information.
5 * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 *===-----------------------------------------------------------------------===
11 #error "Never use <fma4intrin.h> directly; include <x86intrin.h> instead."
14 #ifndef __FMA4INTRIN_H
15 #define __FMA4INTRIN_H
17 #include <pmmintrin.h>
19 /* Define the default attributes for the functions in this file. */
20 #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("fma4"), __min_vector_width__(128)))
21 #define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("fma4"), __min_vector_width__(256)))
23 static __inline__ __m128 __DEFAULT_FN_ATTRS128
24 _mm_macc_ps(__m128 __A
, __m128 __B
, __m128 __C
)
26 return (__m128
)__builtin_ia32_vfmaddps((__v4sf
)__A
, (__v4sf
)__B
, (__v4sf
)__C
);
29 static __inline__ __m128d __DEFAULT_FN_ATTRS128
30 _mm_macc_pd(__m128d __A
, __m128d __B
, __m128d __C
)
32 return (__m128d
)__builtin_ia32_vfmaddpd((__v2df
)__A
, (__v2df
)__B
, (__v2df
)__C
);
35 static __inline__ __m128 __DEFAULT_FN_ATTRS128
36 _mm_macc_ss(__m128 __A
, __m128 __B
, __m128 __C
)
38 return (__m128
)__builtin_ia32_vfmaddss((__v4sf
)__A
, (__v4sf
)__B
, (__v4sf
)__C
);
41 static __inline__ __m128d __DEFAULT_FN_ATTRS128
42 _mm_macc_sd(__m128d __A
, __m128d __B
, __m128d __C
)
44 return (__m128d
)__builtin_ia32_vfmaddsd((__v2df
)__A
, (__v2df
)__B
, (__v2df
)__C
);
47 static __inline__ __m128 __DEFAULT_FN_ATTRS128
48 _mm_msub_ps(__m128 __A
, __m128 __B
, __m128 __C
)
50 return (__m128
)__builtin_ia32_vfmaddps((__v4sf
)__A
, (__v4sf
)__B
, -(__v4sf
)__C
);
53 static __inline__ __m128d __DEFAULT_FN_ATTRS128
54 _mm_msub_pd(__m128d __A
, __m128d __B
, __m128d __C
)
56 return (__m128d
)__builtin_ia32_vfmaddpd((__v2df
)__A
, (__v2df
)__B
, -(__v2df
)__C
);
59 static __inline__ __m128 __DEFAULT_FN_ATTRS128
60 _mm_msub_ss(__m128 __A
, __m128 __B
, __m128 __C
)
62 return (__m128
)__builtin_ia32_vfmaddss((__v4sf
)__A
, (__v4sf
)__B
, -(__v4sf
)__C
);
65 static __inline__ __m128d __DEFAULT_FN_ATTRS128
66 _mm_msub_sd(__m128d __A
, __m128d __B
, __m128d __C
)
68 return (__m128d
)__builtin_ia32_vfmaddsd((__v2df
)__A
, (__v2df
)__B
, -(__v2df
)__C
);
71 static __inline__ __m128 __DEFAULT_FN_ATTRS128
72 _mm_nmacc_ps(__m128 __A
, __m128 __B
, __m128 __C
)
74 return (__m128
)__builtin_ia32_vfmaddps(-(__v4sf
)__A
, (__v4sf
)__B
, (__v4sf
)__C
);
77 static __inline__ __m128d __DEFAULT_FN_ATTRS128
78 _mm_nmacc_pd(__m128d __A
, __m128d __B
, __m128d __C
)
80 return (__m128d
)__builtin_ia32_vfmaddpd(-(__v2df
)__A
, (__v2df
)__B
, (__v2df
)__C
);
83 static __inline__ __m128 __DEFAULT_FN_ATTRS128
84 _mm_nmacc_ss(__m128 __A
, __m128 __B
, __m128 __C
)
86 return (__m128
)__builtin_ia32_vfmaddss(-(__v4sf
)__A
, (__v4sf
)__B
, (__v4sf
)__C
);
89 static __inline__ __m128d __DEFAULT_FN_ATTRS128
90 _mm_nmacc_sd(__m128d __A
, __m128d __B
, __m128d __C
)
92 return (__m128d
)__builtin_ia32_vfmaddsd(-(__v2df
)__A
, (__v2df
)__B
, (__v2df
)__C
);
95 static __inline__ __m128 __DEFAULT_FN_ATTRS128
96 _mm_nmsub_ps(__m128 __A
, __m128 __B
, __m128 __C
)
98 return (__m128
)__builtin_ia32_vfmaddps(-(__v4sf
)__A
, (__v4sf
)__B
, -(__v4sf
)__C
);
101 static __inline__ __m128d __DEFAULT_FN_ATTRS128
102 _mm_nmsub_pd(__m128d __A
, __m128d __B
, __m128d __C
)
104 return (__m128d
)__builtin_ia32_vfmaddpd(-(__v2df
)__A
, (__v2df
)__B
, -(__v2df
)__C
);
107 static __inline__ __m128 __DEFAULT_FN_ATTRS128
108 _mm_nmsub_ss(__m128 __A
, __m128 __B
, __m128 __C
)
110 return (__m128
)__builtin_ia32_vfmaddss(-(__v4sf
)__A
, (__v4sf
)__B
, -(__v4sf
)__C
);
113 static __inline__ __m128d __DEFAULT_FN_ATTRS128
114 _mm_nmsub_sd(__m128d __A
, __m128d __B
, __m128d __C
)
116 return (__m128d
)__builtin_ia32_vfmaddsd(-(__v2df
)__A
, (__v2df
)__B
, -(__v2df
)__C
);
119 static __inline__ __m128 __DEFAULT_FN_ATTRS128
120 _mm_maddsub_ps(__m128 __A
, __m128 __B
, __m128 __C
)
122 return (__m128
)__builtin_ia32_vfmaddsubps((__v4sf
)__A
, (__v4sf
)__B
, (__v4sf
)__C
);
125 static __inline__ __m128d __DEFAULT_FN_ATTRS128
126 _mm_maddsub_pd(__m128d __A
, __m128d __B
, __m128d __C
)
128 return (__m128d
)__builtin_ia32_vfmaddsubpd((__v2df
)__A
, (__v2df
)__B
, (__v2df
)__C
);
131 static __inline__ __m128 __DEFAULT_FN_ATTRS128
132 _mm_msubadd_ps(__m128 __A
, __m128 __B
, __m128 __C
)
134 return (__m128
)__builtin_ia32_vfmaddsubps((__v4sf
)__A
, (__v4sf
)__B
, -(__v4sf
)__C
);
137 static __inline__ __m128d __DEFAULT_FN_ATTRS128
138 _mm_msubadd_pd(__m128d __A
, __m128d __B
, __m128d __C
)
140 return (__m128d
)__builtin_ia32_vfmaddsubpd((__v2df
)__A
, (__v2df
)__B
, -(__v2df
)__C
);
143 static __inline__ __m256 __DEFAULT_FN_ATTRS256
144 _mm256_macc_ps(__m256 __A
, __m256 __B
, __m256 __C
)
146 return (__m256
)__builtin_ia32_vfmaddps256((__v8sf
)__A
, (__v8sf
)__B
, (__v8sf
)__C
);
149 static __inline__ __m256d __DEFAULT_FN_ATTRS256
150 _mm256_macc_pd(__m256d __A
, __m256d __B
, __m256d __C
)
152 return (__m256d
)__builtin_ia32_vfmaddpd256((__v4df
)__A
, (__v4df
)__B
, (__v4df
)__C
);
155 static __inline__ __m256 __DEFAULT_FN_ATTRS256
156 _mm256_msub_ps(__m256 __A
, __m256 __B
, __m256 __C
)
158 return (__m256
)__builtin_ia32_vfmaddps256((__v8sf
)__A
, (__v8sf
)__B
, -(__v8sf
)__C
);
161 static __inline__ __m256d __DEFAULT_FN_ATTRS256
162 _mm256_msub_pd(__m256d __A
, __m256d __B
, __m256d __C
)
164 return (__m256d
)__builtin_ia32_vfmaddpd256((__v4df
)__A
, (__v4df
)__B
, -(__v4df
)__C
);
167 static __inline__ __m256 __DEFAULT_FN_ATTRS256
168 _mm256_nmacc_ps(__m256 __A
, __m256 __B
, __m256 __C
)
170 return (__m256
)__builtin_ia32_vfmaddps256(-(__v8sf
)__A
, (__v8sf
)__B
, (__v8sf
)__C
);
173 static __inline__ __m256d __DEFAULT_FN_ATTRS256
174 _mm256_nmacc_pd(__m256d __A
, __m256d __B
, __m256d __C
)
176 return (__m256d
)__builtin_ia32_vfmaddpd256(-(__v4df
)__A
, (__v4df
)__B
, (__v4df
)__C
);
179 static __inline__ __m256 __DEFAULT_FN_ATTRS256
180 _mm256_nmsub_ps(__m256 __A
, __m256 __B
, __m256 __C
)
182 return (__m256
)__builtin_ia32_vfmaddps256(-(__v8sf
)__A
, (__v8sf
)__B
, -(__v8sf
)__C
);
185 static __inline__ __m256d __DEFAULT_FN_ATTRS256
186 _mm256_nmsub_pd(__m256d __A
, __m256d __B
, __m256d __C
)
188 return (__m256d
)__builtin_ia32_vfmaddpd256(-(__v4df
)__A
, (__v4df
)__B
, -(__v4df
)__C
);
191 static __inline__ __m256 __DEFAULT_FN_ATTRS256
192 _mm256_maddsub_ps(__m256 __A
, __m256 __B
, __m256 __C
)
194 return (__m256
)__builtin_ia32_vfmaddsubps256((__v8sf
)__A
, (__v8sf
)__B
, (__v8sf
)__C
);
197 static __inline__ __m256d __DEFAULT_FN_ATTRS256
198 _mm256_maddsub_pd(__m256d __A
, __m256d __B
, __m256d __C
)
200 return (__m256d
)__builtin_ia32_vfmaddsubpd256((__v4df
)__A
, (__v4df
)__B
, (__v4df
)__C
);
203 static __inline__ __m256 __DEFAULT_FN_ATTRS256
204 _mm256_msubadd_ps(__m256 __A
, __m256 __B
, __m256 __C
)
206 return (__m256
)__builtin_ia32_vfmaddsubps256((__v8sf
)__A
, (__v8sf
)__B
, -(__v8sf
)__C
);
209 static __inline__ __m256d __DEFAULT_FN_ATTRS256
210 _mm256_msubadd_pd(__m256d __A
, __m256d __B
, __m256d __C
)
212 return (__m256d
)__builtin_ia32_vfmaddsubpd256((__v4df
)__A
, (__v4df
)__B
, -(__v4df
)__C
);
215 #undef __DEFAULT_FN_ATTRS128
216 #undef __DEFAULT_FN_ATTRS256
218 #endif /* __FMA4INTRIN_H */
