1 /*===------------- avx512ifmavlintrin.h - IFMA intrinsics ------------------===
2  *
3  *
4  * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
5  * See https://llvm.org/LICENSE.txt for license information.
6  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7  *
8  *===-----------------------------------------------------------------------===
9  */
10 #ifndef __IMMINTRIN_H
11 #error "Never use <avx512ifmavlintrin.h> directly; include <immintrin.h> instead."
12 #endif
13 
14 #ifndef __IFMAVLINTRIN_H
15 #define __IFMAVLINTRIN_H
16 
17 /* Define the default attributes for the functions in this file. */
18 #define __DEFAULT_FN_ATTRS128                                                  \
19   __attribute__((__always_inline__, __nodebug__,                               \
20                  __target__("avx512ifma,avx512vl,no-evex512"),                 \
21                  __min_vector_width__(128)))
22 #define __DEFAULT_FN_ATTRS256                                                  \
23   __attribute__((__always_inline__, __nodebug__,                               \
24                  __target__("avx512ifma,avx512vl,no-evex512"),                 \
25                  __min_vector_width__(256)))
26 
27 #define _mm_madd52hi_epu64(X, Y, Z)                                            \
28   ((__m128i)__builtin_ia32_vpmadd52huq128((__v2di)(X), (__v2di)(Y),            \
29                                           (__v2di)(Z)))
30 
31 #define _mm256_madd52hi_epu64(X, Y, Z)                                         \
32   ((__m256i)__builtin_ia32_vpmadd52huq256((__v4di)(X), (__v4di)(Y),            \
33                                           (__v4di)(Z)))
34 
35 #define _mm_madd52lo_epu64(X, Y, Z)                                            \
36   ((__m128i)__builtin_ia32_vpmadd52luq128((__v2di)(X), (__v2di)(Y),            \
37                                           (__v2di)(Z)))
38 
39 #define _mm256_madd52lo_epu64(X, Y, Z)                                         \
40   ((__m256i)__builtin_ia32_vpmadd52luq256((__v4di)(X), (__v4di)(Y),            \
41                                           (__v4di)(Z)))
42 
43 static __inline__ __m128i __DEFAULT_FN_ATTRS128
_mm_mask_madd52hi_epu64(__m128i __W,__mmask8 __M,__m128i __X,__m128i __Y)44 _mm_mask_madd52hi_epu64 (__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y)
45 {
46   return (__m128i)__builtin_ia32_selectq_128(__M,
47                                       (__v2di)_mm_madd52hi_epu64(__W, __X, __Y),
48                                       (__v2di)__W);
49 }
50 
51 static __inline__ __m128i __DEFAULT_FN_ATTRS128
_mm_maskz_madd52hi_epu64(__mmask8 __M,__m128i __X,__m128i __Y,__m128i __Z)52 _mm_maskz_madd52hi_epu64 (__mmask8 __M, __m128i __X, __m128i __Y, __m128i __Z)
53 {
54   return (__m128i)__builtin_ia32_selectq_128(__M,
55                                       (__v2di)_mm_madd52hi_epu64(__X, __Y, __Z),
56                                       (__v2di)_mm_setzero_si128());
57 }
58 
59 static __inline__ __m256i __DEFAULT_FN_ATTRS256
_mm256_mask_madd52hi_epu64(__m256i __W,__mmask8 __M,__m256i __X,__m256i __Y)60 _mm256_mask_madd52hi_epu64 (__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y)
61 {
62   return (__m256i)__builtin_ia32_selectq_256(__M,
63                                    (__v4di)_mm256_madd52hi_epu64(__W, __X, __Y),
64                                    (__v4di)__W);
65 }
66 
67 static __inline__ __m256i __DEFAULT_FN_ATTRS256
_mm256_maskz_madd52hi_epu64(__mmask8 __M,__m256i __X,__m256i __Y,__m256i __Z)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 
75 static __inline__ __m128i __DEFAULT_FN_ATTRS128
_mm_mask_madd52lo_epu64(__m128i __W,__mmask8 __M,__m128i __X,__m128i __Y)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 
83 static __inline__ __m128i __DEFAULT_FN_ATTRS128
_mm_maskz_madd52lo_epu64(__mmask8 __M,__m128i __X,__m128i __Y,__m128i __Z)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 
91 static __inline__ __m256i __DEFAULT_FN_ATTRS256
_mm256_mask_madd52lo_epu64(__m256i __W,__mmask8 __M,__m256i __X,__m256i __Y)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 
99 static __inline__ __m256i __DEFAULT_FN_ATTRS256
_mm256_maskz_madd52lo_epu64(__mmask8 __M,__m256i __X,__m256i __Y,__m256i __Z)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