blob: 694801b3e8dd4bc0b8639ca0a75c13be19523f71 [file] [log] [blame]
Logan Chien2833ffb2018-10-09 10:03:24 +08001/*===---- fma4intrin.h - FMA4 intrinsics -----------------------------------===
2 *
Logan Chiendf4f7662019-09-04 16:45:23 -07003 * 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
Logan Chien2833ffb2018-10-09 10:03:24 +08006 *
7 *===-----------------------------------------------------------------------===
8 */
9
10#ifndef __X86INTRIN_H
11#error "Never use <fma4intrin.h> directly; include <x86intrin.h> instead."
12#endif
13
14#ifndef __FMA4INTRIN_H
15#define __FMA4INTRIN_H
16
17#include <pmmintrin.h>
18
19/* Define the default attributes for the functions in this file. */
Logan Chien55afb0a2018-10-15 10:42:14 +080020#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)))
Logan Chien2833ffb2018-10-09 10:03:24 +080022
Logan Chien55afb0a2018-10-15 10:42:14 +080023static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080024_mm_macc_ps(__m128 __A, __m128 __B, __m128 __C)
25{
26 return (__m128)__builtin_ia32_vfmaddps((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
27}
28
Logan Chien55afb0a2018-10-15 10:42:14 +080029static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080030_mm_macc_pd(__m128d __A, __m128d __B, __m128d __C)
31{
32 return (__m128d)__builtin_ia32_vfmaddpd((__v2df)__A, (__v2df)__B, (__v2df)__C);
33}
34
Logan Chien55afb0a2018-10-15 10:42:14 +080035static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080036_mm_macc_ss(__m128 __A, __m128 __B, __m128 __C)
37{
38 return (__m128)__builtin_ia32_vfmaddss((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
39}
40
Logan Chien55afb0a2018-10-15 10:42:14 +080041static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080042_mm_macc_sd(__m128d __A, __m128d __B, __m128d __C)
43{
44 return (__m128d)__builtin_ia32_vfmaddsd((__v2df)__A, (__v2df)__B, (__v2df)__C);
45}
46
Logan Chien55afb0a2018-10-15 10:42:14 +080047static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080048_mm_msub_ps(__m128 __A, __m128 __B, __m128 __C)
49{
Logan Chien55afb0a2018-10-15 10:42:14 +080050 return (__m128)__builtin_ia32_vfmaddps((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +080051}
52
Logan Chien55afb0a2018-10-15 10:42:14 +080053static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080054_mm_msub_pd(__m128d __A, __m128d __B, __m128d __C)
55{
Logan Chien55afb0a2018-10-15 10:42:14 +080056 return (__m128d)__builtin_ia32_vfmaddpd((__v2df)__A, (__v2df)__B, -(__v2df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +080057}
58
Logan Chien55afb0a2018-10-15 10:42:14 +080059static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080060_mm_msub_ss(__m128 __A, __m128 __B, __m128 __C)
61{
Logan Chien55afb0a2018-10-15 10:42:14 +080062 return (__m128)__builtin_ia32_vfmaddss((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +080063}
64
Logan Chien55afb0a2018-10-15 10:42:14 +080065static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080066_mm_msub_sd(__m128d __A, __m128d __B, __m128d __C)
67{
Logan Chien55afb0a2018-10-15 10:42:14 +080068 return (__m128d)__builtin_ia32_vfmaddsd((__v2df)__A, (__v2df)__B, -(__v2df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +080069}
70
Logan Chien55afb0a2018-10-15 10:42:14 +080071static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080072_mm_nmacc_ps(__m128 __A, __m128 __B, __m128 __C)
73{
Logan Chien55afb0a2018-10-15 10:42:14 +080074 return (__m128)__builtin_ia32_vfmaddps(-(__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +080075}
76
Logan Chien55afb0a2018-10-15 10:42:14 +080077static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080078_mm_nmacc_pd(__m128d __A, __m128d __B, __m128d __C)
79{
Logan Chien55afb0a2018-10-15 10:42:14 +080080 return (__m128d)__builtin_ia32_vfmaddpd(-(__v2df)__A, (__v2df)__B, (__v2df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +080081}
82
Logan Chien55afb0a2018-10-15 10:42:14 +080083static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080084_mm_nmacc_ss(__m128 __A, __m128 __B, __m128 __C)
85{
Logan Chien55afb0a2018-10-15 10:42:14 +080086 return (__m128)__builtin_ia32_vfmaddss(-(__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +080087}
88
Logan Chien55afb0a2018-10-15 10:42:14 +080089static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080090_mm_nmacc_sd(__m128d __A, __m128d __B, __m128d __C)
91{
Logan Chien55afb0a2018-10-15 10:42:14 +080092 return (__m128d)__builtin_ia32_vfmaddsd(-(__v2df)__A, (__v2df)__B, (__v2df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +080093}
94
Logan Chien55afb0a2018-10-15 10:42:14 +080095static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +080096_mm_nmsub_ps(__m128 __A, __m128 __B, __m128 __C)
97{
Logan Chien55afb0a2018-10-15 10:42:14 +080098 return (__m128)__builtin_ia32_vfmaddps(-(__v4sf)__A, (__v4sf)__B, -(__v4sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +080099}
100
Logan Chien55afb0a2018-10-15 10:42:14 +0800101static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +0800102_mm_nmsub_pd(__m128d __A, __m128d __B, __m128d __C)
103{
Logan Chien55afb0a2018-10-15 10:42:14 +0800104 return (__m128d)__builtin_ia32_vfmaddpd(-(__v2df)__A, (__v2df)__B, -(__v2df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800105}
106
Logan Chien55afb0a2018-10-15 10:42:14 +0800107static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +0800108_mm_nmsub_ss(__m128 __A, __m128 __B, __m128 __C)
109{
Logan Chien55afb0a2018-10-15 10:42:14 +0800110 return (__m128)__builtin_ia32_vfmaddss(-(__v4sf)__A, (__v4sf)__B, -(__v4sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800111}
112
Logan Chien55afb0a2018-10-15 10:42:14 +0800113static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +0800114_mm_nmsub_sd(__m128d __A, __m128d __B, __m128d __C)
115{
Logan Chien55afb0a2018-10-15 10:42:14 +0800116 return (__m128d)__builtin_ia32_vfmaddsd(-(__v2df)__A, (__v2df)__B, -(__v2df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800117}
118
Logan Chien55afb0a2018-10-15 10:42:14 +0800119static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +0800120_mm_maddsub_ps(__m128 __A, __m128 __B, __m128 __C)
121{
122 return (__m128)__builtin_ia32_vfmaddsubps((__v4sf)__A, (__v4sf)__B, (__v4sf)__C);
123}
124
Logan Chien55afb0a2018-10-15 10:42:14 +0800125static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +0800126_mm_maddsub_pd(__m128d __A, __m128d __B, __m128d __C)
127{
128 return (__m128d)__builtin_ia32_vfmaddsubpd((__v2df)__A, (__v2df)__B, (__v2df)__C);
129}
130
Logan Chien55afb0a2018-10-15 10:42:14 +0800131static __inline__ __m128 __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +0800132_mm_msubadd_ps(__m128 __A, __m128 __B, __m128 __C)
133{
Logan Chien55afb0a2018-10-15 10:42:14 +0800134 return (__m128)__builtin_ia32_vfmaddsubps((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800135}
136
Logan Chien55afb0a2018-10-15 10:42:14 +0800137static __inline__ __m128d __DEFAULT_FN_ATTRS128
Logan Chien2833ffb2018-10-09 10:03:24 +0800138_mm_msubadd_pd(__m128d __A, __m128d __B, __m128d __C)
139{
Logan Chien55afb0a2018-10-15 10:42:14 +0800140 return (__m128d)__builtin_ia32_vfmaddsubpd((__v2df)__A, (__v2df)__B, -(__v2df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800141}
142
Logan Chien55afb0a2018-10-15 10:42:14 +0800143static __inline__ __m256 __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800144_mm256_macc_ps(__m256 __A, __m256 __B, __m256 __C)
145{
146 return (__m256)__builtin_ia32_vfmaddps256((__v8sf)__A, (__v8sf)__B, (__v8sf)__C);
147}
148
Logan Chien55afb0a2018-10-15 10:42:14 +0800149static __inline__ __m256d __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800150_mm256_macc_pd(__m256d __A, __m256d __B, __m256d __C)
151{
152 return (__m256d)__builtin_ia32_vfmaddpd256((__v4df)__A, (__v4df)__B, (__v4df)__C);
153}
154
Logan Chien55afb0a2018-10-15 10:42:14 +0800155static __inline__ __m256 __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800156_mm256_msub_ps(__m256 __A, __m256 __B, __m256 __C)
157{
Logan Chien55afb0a2018-10-15 10:42:14 +0800158 return (__m256)__builtin_ia32_vfmaddps256((__v8sf)__A, (__v8sf)__B, -(__v8sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800159}
160
Logan Chien55afb0a2018-10-15 10:42:14 +0800161static __inline__ __m256d __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800162_mm256_msub_pd(__m256d __A, __m256d __B, __m256d __C)
163{
Logan Chien55afb0a2018-10-15 10:42:14 +0800164 return (__m256d)__builtin_ia32_vfmaddpd256((__v4df)__A, (__v4df)__B, -(__v4df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800165}
166
Logan Chien55afb0a2018-10-15 10:42:14 +0800167static __inline__ __m256 __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800168_mm256_nmacc_ps(__m256 __A, __m256 __B, __m256 __C)
169{
Logan Chien55afb0a2018-10-15 10:42:14 +0800170 return (__m256)__builtin_ia32_vfmaddps256(-(__v8sf)__A, (__v8sf)__B, (__v8sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800171}
172
Logan Chien55afb0a2018-10-15 10:42:14 +0800173static __inline__ __m256d __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800174_mm256_nmacc_pd(__m256d __A, __m256d __B, __m256d __C)
175{
Logan Chien55afb0a2018-10-15 10:42:14 +0800176 return (__m256d)__builtin_ia32_vfmaddpd256(-(__v4df)__A, (__v4df)__B, (__v4df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800177}
178
Logan Chien55afb0a2018-10-15 10:42:14 +0800179static __inline__ __m256 __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800180_mm256_nmsub_ps(__m256 __A, __m256 __B, __m256 __C)
181{
Logan Chien55afb0a2018-10-15 10:42:14 +0800182 return (__m256)__builtin_ia32_vfmaddps256(-(__v8sf)__A, (__v8sf)__B, -(__v8sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800183}
184
Logan Chien55afb0a2018-10-15 10:42:14 +0800185static __inline__ __m256d __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800186_mm256_nmsub_pd(__m256d __A, __m256d __B, __m256d __C)
187{
Logan Chien55afb0a2018-10-15 10:42:14 +0800188 return (__m256d)__builtin_ia32_vfmaddpd256(-(__v4df)__A, (__v4df)__B, -(__v4df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800189}
190
Logan Chien55afb0a2018-10-15 10:42:14 +0800191static __inline__ __m256 __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800192_mm256_maddsub_ps(__m256 __A, __m256 __B, __m256 __C)
193{
194 return (__m256)__builtin_ia32_vfmaddsubps256((__v8sf)__A, (__v8sf)__B, (__v8sf)__C);
195}
196
Logan Chien55afb0a2018-10-15 10:42:14 +0800197static __inline__ __m256d __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800198_mm256_maddsub_pd(__m256d __A, __m256d __B, __m256d __C)
199{
200 return (__m256d)__builtin_ia32_vfmaddsubpd256((__v4df)__A, (__v4df)__B, (__v4df)__C);
201}
202
Logan Chien55afb0a2018-10-15 10:42:14 +0800203static __inline__ __m256 __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800204_mm256_msubadd_ps(__m256 __A, __m256 __B, __m256 __C)
205{
Logan Chien55afb0a2018-10-15 10:42:14 +0800206 return (__m256)__builtin_ia32_vfmaddsubps256((__v8sf)__A, (__v8sf)__B, -(__v8sf)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800207}
208
Logan Chien55afb0a2018-10-15 10:42:14 +0800209static __inline__ __m256d __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800210_mm256_msubadd_pd(__m256d __A, __m256d __B, __m256d __C)
211{
Logan Chien55afb0a2018-10-15 10:42:14 +0800212 return (__m256d)__builtin_ia32_vfmaddsubpd256((__v4df)__A, (__v4df)__B, -(__v4df)__C);
Logan Chien2833ffb2018-10-09 10:03:24 +0800213}
214
Logan Chien55afb0a2018-10-15 10:42:14 +0800215#undef __DEFAULT_FN_ATTRS128
216#undef __DEFAULT_FN_ATTRS256
Logan Chien2833ffb2018-10-09 10:03:24 +0800217
218#endif /* __FMA4INTRIN_H */