| 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 \ |
| 12 | "Never use <avx512ifmavlintrin.h> directly; include <immintrin.h> instead." |
| 13 | #endif |
| 14 | |
| 15 | #ifndef __IFMAVLINTRIN_H |
| 16 | #define __IFMAVLINTRIN_H |
| 17 | |
| 18 | /* Define the default attributes for the functions in this file. */ |
| 19 | #if defined(__cplusplus) && (__cplusplus >= 201103L) |
| 20 | #define __DEFAULT_FN_ATTRS128 \ |
| 21 | __attribute__((__always_inline__, __nodebug__, \ |
| 22 | __target__("avx512ifma,avx512vl"), \ |
| 23 | __min_vector_width__(128))) constexpr |
| 24 | #define __DEFAULT_FN_ATTRS256 \ |
| 25 | __attribute__((__always_inline__, __nodebug__, \ |
| 26 | __target__("avx512ifma,avx512vl"), \ |
| 27 | __min_vector_width__(256))) constexpr |
| 28 | #else |
| 29 | #define __DEFAULT_FN_ATTRS128 \ |
| 30 | __attribute__((__always_inline__, __nodebug__, \ |
| 31 | __target__("avx512ifma,avx512vl"), \ |
| 32 | __min_vector_width__(128))) |
| 33 | #define __DEFAULT_FN_ATTRS256 \ |
| 34 | __attribute__((__always_inline__, __nodebug__, \ |
| 35 | __target__("avx512ifma,avx512vl"), \ |
| 36 | __min_vector_width__(256))) |
| 37 | #endif |
| 38 | |
| 39 | #if !(defined(__AVXIFMA__) || defined(__AVX512IFMA__)) |
| 40 | #define _mm_madd52hi_epu64(X, Y, Z) \ |
| 41 | ((__m128i)__builtin_ia32_vpmadd52huq128((__v2di)(X), (__v2di)(Y), \ |
| 42 | (__v2di)(Z))) |
| 43 | |
| 44 | #define _mm256_madd52hi_epu64(X, Y, Z) \ |
| 45 | ((__m256i)__builtin_ia32_vpmadd52huq256((__v4di)(X), (__v4di)(Y), \ |
| 46 | (__v4di)(Z))) |
| 47 | |
| 48 | #define _mm_madd52lo_epu64(X, Y, Z) \ |
| 49 | ((__m128i)__builtin_ia32_vpmadd52luq128((__v2di)(X), (__v2di)(Y), \ |
| 50 | (__v2di)(Z))) |
| 51 | |
| 52 | #define _mm256_madd52lo_epu64(X, Y, Z) \ |
| 53 | ((__m256i)__builtin_ia32_vpmadd52luq256((__v4di)(X), (__v4di)(Y), \ |
| 54 | (__v4di)(Z))) |
| 55 | #endif |
| 56 | |
| 57 | #if defined(__AVX512IFMA__) |
| 58 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 59 | _mm_madd52hi_epu64(__m128i __X, __m128i __Y, __m128i __Z) { |
| 60 | return (__m128i)__builtin_ia32_vpmadd52huq128((__v2di)__X, (__v2di)__Y, |
| 61 | (__v2di)__Z); |
| 62 | } |
| 63 | |
| 64 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 65 | _mm256_madd52hi_epu64(__m256i __X, __m256i __Y, __m256i __Z) { |
| 66 | return (__m256i)__builtin_ia32_vpmadd52huq256((__v4di)__X, (__v4di)__Y, |
| 67 | (__v4di)__Z); |
| 68 | } |
| 69 | |
| 70 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 71 | _mm_madd52lo_epu64(__m128i __X, __m128i __Y, __m128i __Z) { |
| 72 | return (__m128i)__builtin_ia32_vpmadd52luq128((__v2di)__X, (__v2di)__Y, |
| 73 | (__v2di)__Z); |
| 74 | } |
| 75 | |
| 76 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 77 | _mm256_madd52lo_epu64(__m256i __X, __m256i __Y, __m256i __Z) { |
| 78 | return (__m256i)__builtin_ia32_vpmadd52luq256((__v4di)__X, (__v4di)__Y, |
| 79 | (__v4di)__Z); |
| 80 | } |
| 81 | #endif |
| 82 | |
| 83 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 84 | _mm_mask_madd52hi_epu64(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y) { |
| 85 | return (__m128i)__builtin_ia32_selectq_128( |
| 86 | __M, (__v2di)__builtin_ia32_vpmadd52huq128(__W, __X, __Y), (__v2di)__W); |
| 87 | } |
| 88 | |
| 89 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 90 | _mm_maskz_madd52hi_epu64(__mmask8 __M, __m128i __X, __m128i __Y, __m128i __Z) { |
| 91 | return (__m128i)__builtin_ia32_selectq_128( |
| 92 | __M, (__v2di)__builtin_ia32_vpmadd52huq128(__X, __Y, __Z), |
| 93 | (__v2di)_mm_setzero_si128()); |
| 94 | } |
| 95 | |
| 96 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_madd52hi_epu64( |
| 97 | __m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) { |
| 98 | return (__m256i)__builtin_ia32_selectq_256( |
| 99 | __M, (__v4di)__builtin_ia32_vpmadd52huq256(__W, __X, __Y), (__v4di)__W); |
| 100 | } |
| 101 | |
| 102 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_madd52hi_epu64( |
| 103 | __mmask8 __M, __m256i __X, __m256i __Y, __m256i __Z) { |
| 104 | return (__m256i)__builtin_ia32_selectq_256( |
| 105 | __M, (__v4di)__builtin_ia32_vpmadd52huq256(__X, __Y, __Z), |
| 106 | (__v4di)_mm256_setzero_si256()); |
| 107 | } |
| 108 | |
| 109 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 110 | _mm_mask_madd52lo_epu64(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y) { |
| 111 | return (__m128i)__builtin_ia32_selectq_128( |
| 112 | __M, (__v2di)__builtin_ia32_vpmadd52luq128(__W, __X, __Y), (__v2di)__W); |
| 113 | } |
| 114 | |
| 115 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 116 | _mm_maskz_madd52lo_epu64(__mmask8 __M, __m128i __X, __m128i __Y, __m128i __Z) { |
| 117 | return (__m128i)__builtin_ia32_selectq_128( |
| 118 | __M, (__v2di)__builtin_ia32_vpmadd52luq128(__X, __Y, __Z), |
| 119 | (__v2di)_mm_setzero_si128()); |
| 120 | } |
| 121 | |
| 122 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_madd52lo_epu64( |
| 123 | __m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) { |
| 124 | return (__m256i)__builtin_ia32_selectq_256( |
| 125 | __M, (__v4di)__builtin_ia32_vpmadd52luq256(__W, __X, __Y), (__v4di)__W); |
| 126 | } |
| 127 | |
| 128 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_madd52lo_epu64( |
| 129 | __mmask8 __M, __m256i __X, __m256i __Y, __m256i __Z) { |
| 130 | return (__m256i)__builtin_ia32_selectq_256( |
| 131 | __M, (__v4di)__builtin_ia32_vpmadd52luq256(__X, __Y, __Z), |
| 132 | (__v4di)_mm256_setzero_si256()); |
| 133 | } |
| 134 | |
| 135 | #undef __DEFAULT_FN_ATTRS128 |
| 136 | #undef __DEFAULT_FN_ATTRS256 |
| 137 | |
| 138 | #endif |