1234285Sdim/*===------------- avx512ifmavlintrin.h - IFMA intrinsics ------------------===
2234285Sdim *
3234285Sdim *
4234285Sdim * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
5234285Sdim * See https://llvm.org/LICENSE.txt for license information.
6234285Sdim * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7234285Sdim *
8234285Sdim *===-----------------------------------------------------------------------===
9234285Sdim */
10234285Sdim#ifndef __IMMINTRIN_H
11234285Sdim#error "Never use <avx512ifmavlintrin.h> directly; include <immintrin.h> instead."
12234285Sdim#endif
13234285Sdim
14280031Sdim#ifndef __IFMAVLINTRIN_H
15280031Sdim#define __IFMAVLINTRIN_H
16234285Sdim
17234285Sdim/* Define the default attributes for the functions in this file. */
18234285Sdim#define __DEFAULT_FN_ATTRS128                                                  \
19234285Sdim  __attribute__((__always_inline__, __nodebug__,                               \
20234285Sdim                 __target__("avx512ifma,avx512vl,no-evex512"),                 \
21234285Sdim                 __min_vector_width__(128)))
22234285Sdim#define __DEFAULT_FN_ATTRS256                                                  \
23234285Sdim  __attribute__((__always_inline__, __nodebug__,                               \
24234285Sdim                 __target__("avx512ifma,avx512vl,no-evex512"),                 \
25234285Sdim                 __min_vector_width__(256)))
26234285Sdim
27234285Sdim#define _mm_madd52hi_epu64(X, Y, Z)                                            \
28288943Sdim  ((__m128i)__builtin_ia32_vpmadd52huq128((__v2di)(X), (__v2di)(Y),            \
29288943Sdim                                          (__v2di)(Z)))
30288943Sdim
31288943Sdim#define _mm256_madd52hi_epu64(X, Y, Z)                                         \
32288943Sdim  ((__m256i)__builtin_ia32_vpmadd52huq256((__v4di)(X), (__v4di)(Y),            \
33288943Sdim                                          (__v4di)(Z)))
34234285Sdim
35234285Sdim#define _mm_madd52lo_epu64(X, Y, Z)                                            \
36276479Sdim  ((__m128i)__builtin_ia32_vpmadd52luq128((__v2di)(X), (__v2di)(Y),            \
37234285Sdim                                          (__v2di)(Z)))
38234285Sdim
39234285Sdim#define _mm256_madd52lo_epu64(X, Y, Z)                                         \
40276479Sdim  ((__m256i)__builtin_ia32_vpmadd52luq256((__v4di)(X), (__v4di)(Y),            \
41276479Sdim                                          (__v4di)(Z)))
42234285Sdim
43276479Sdimstatic __inline__ __m128i __DEFAULT_FN_ATTRS128
44234285Sdim_mm_mask_madd52hi_epu64 (__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y)
45296417Sdim{
46296417Sdim  return (__m128i)__builtin_ia32_selectq_128(__M,
47296417Sdim                                      (__v2di)_mm_madd52hi_epu64(__W, __X, __Y),
48296417Sdim                                      (__v2di)__W);
49234285Sdim}
50234285Sdim
51234285Sdimstatic __inline__ __m128i __DEFAULT_FN_ATTRS128
52276479Sdim_mm_maskz_madd52hi_epu64 (__mmask8 __M, __m128i __X, __m128i __Y, __m128i __Z)
53234285Sdim{
54234285Sdim  return (__m128i)__builtin_ia32_selectq_128(__M,
55276479Sdim                                      (__v2di)_mm_madd52hi_epu64(__X, __Y, __Z),
56234285Sdim                                      (__v2di)_mm_setzero_si128());
57234285Sdim}
58234285Sdim
59234285Sdimstatic __inline__ __m256i __DEFAULT_FN_ATTRS256
60234285Sdim_mm256_mask_madd52hi_epu64 (__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y)
61234285Sdim{
62234285Sdim  return (__m256i)__builtin_ia32_selectq_256(__M,
63                                   (__v4di)_mm256_madd52hi_epu64(__W, __X, __Y),
64                                   (__v4di)__W);
65}
66
67static __inline__ __m256i __DEFAULT_FN_ATTRS256
68_mm256_maskz_madd52hi_epu64 (__mmask8 __M, __m256i __X, __m256i __Y, __m256i __Z)
69{
70  return (__m256i)__builtin_ia32_selectq_256(__M,
71                                   (__v4di)_mm256_madd52hi_epu64(__X, __Y, __Z),
72                                   (__v4di)_mm256_setzero_si256());
73}
74
75static __inline__ __m128i __DEFAULT_FN_ATTRS128
76_mm_mask_madd52lo_epu64 (__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y)
77{
78  return (__m128i)__builtin_ia32_selectq_128(__M,
79                                      (__v2di)_mm_madd52lo_epu64(__W, __X, __Y),
80                                      (__v2di)__W);
81}
82
83static __inline__ __m128i __DEFAULT_FN_ATTRS128
84_mm_maskz_madd52lo_epu64 (__mmask8 __M, __m128i __X, __m128i __Y, __m128i __Z)
85{
86  return (__m128i)__builtin_ia32_selectq_128(__M,
87                                      (__v2di)_mm_madd52lo_epu64(__X, __Y, __Z),
88                                      (__v2di)_mm_setzero_si128());
89}
90
91static __inline__ __m256i __DEFAULT_FN_ATTRS256
92_mm256_mask_madd52lo_epu64 (__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y)
93{
94  return (__m256i)__builtin_ia32_selectq_256(__M,
95                                   (__v4di)_mm256_madd52lo_epu64(__W, __X, __Y),
96                                   (__v4di)__W);
97}
98
99static __inline__ __m256i __DEFAULT_FN_ATTRS256
100_mm256_maskz_madd52lo_epu64 (__mmask8 __M, __m256i __X, __m256i __Y, __m256i __Z)
101{
102  return (__m256i)__builtin_ia32_selectq_256(__M,
103                                   (__v4di)_mm256_madd52lo_epu64(__X, __Y, __Z),
104                                   (__v4di)_mm256_setzero_si256());
105}
106
107
108#undef __DEFAULT_FN_ATTRS128
109#undef __DEFAULT_FN_ATTRS256
110
111#endif
112