| 1 |
/*===----------------- avxifmaintrin.h - IFMA intrinsics -------------------=== |
| 2 |
* |
| 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 |
| 6 |
* |
| 7 |
*===-----------------------------------------------------------------------=== |
| 8 |
*/ |
| 9 |
|
| 10 |
#ifndef __IMMINTRIN_H |
| 11 |
#error "Never use <avxifmaintrin.h> directly; include <immintrin.h> instead." |
| 12 |
#endif |
| 13 |
|
| 14 |
#ifndef __AVXIFMAINTRIN_H |
| 15 |
#define __AVXIFMAINTRIN_H |
| 16 |
|
| 17 |
/* Define the default attributes for the functions in this file. */ |
| 18 |
#define __DEFAULT_FN_ATTRS128 \ |
| 19 |
__attribute__((__always_inline__, __nodebug__, __target__("avxifma"), \ |
| 20 |
__min_vector_width__(128))) |
| 21 |
#define __DEFAULT_FN_ATTRS256 \ |
| 22 |
__attribute__((__always_inline__, __nodebug__, __target__("avxifma"), \ |
| 23 |
__min_vector_width__(256))) |
| 24 |
|
| 25 |
// must vex-encoding |
| 26 |
|
| 27 |
/// Multiply packed unsigned 52-bit integers in each 64-bit element of \a __Y |
| 28 |
/// and \a __Z to form a 104-bit intermediate result. Add the high 52-bit |
| 29 |
/// unsigned integer from the intermediate result with the corresponding |
| 30 |
/// unsigned 64-bit integer in \a __X, and store the results in \a dst. |
| 31 |
/// |
| 32 |
/// \headerfile <immintrin.h> |
| 33 |
/// |
| 34 |
/// \code |
| 35 |
/// __m128i |
| 36 |
/// _mm_madd52hi_avx_epu64 (__m128i __X, __m128i __Y, __m128i __Z) |
| 37 |
/// \endcode |
| 38 |
/// |
| 39 |
/// This intrinsic corresponds to the \c VPMADD52HUQ instruction. |
| 40 |
/// |
| 41 |
/// \return |
| 42 |
/// return __m128i dst. |
| 43 |
/// \param __X |
| 44 |
/// A 128-bit vector of [2 x i64] |
| 45 |
/// \param __Y |
| 46 |
/// A 128-bit vector of [2 x i64] |
| 47 |
/// \param __Z |
| 48 |
/// A 128-bit vector of [2 x i64] |
| 49 |
/// |
| 50 |
/// \code{.operation} |
| 51 |
/// FOR j := 0 to 1 |
| 52 |
/// i := j*64 |
| 53 |
/// tmp[127:0] := ZeroExtend64(__Y[i+51:i]) * ZeroExtend64(__Z[i+51:i]) |
| 54 |
/// dst[i+63:i] := __X[i+63:i] + ZeroExtend64(tmp[103:52]) |
| 55 |
/// ENDFOR |
| 56 |
/// dst[MAX:128] := 0 |
| 57 |
/// \endcode |
| 58 |
static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 59 |
_mm_madd52hi_avx_epu64(__m128i __X, __m128i __Y, __m128i __Z) { |
| 60 |
return (__m128i)__builtin_ia32_vpmadd52huq128((__v2di)__X, (__v2di)__Y, |
| 61 |
(__v2di)__Z); |
| 62 |
} |
| 63 |
|
| 64 |
/// Multiply packed unsigned 52-bit integers in each 64-bit element of \a __Y |
| 65 |
/// and \a __Z to form a 104-bit intermediate result. Add the high 52-bit |
| 66 |
/// unsigned integer from the intermediate result with the corresponding |
| 67 |
/// unsigned 64-bit integer in \a __X, and store the results in \a dst. |
| 68 |
/// |
| 69 |
/// \headerfile <immintrin.h> |
| 70 |
/// |
| 71 |
/// \code |
| 72 |
/// __m256i |
| 73 |
/// _mm256_madd52hi_avx_epu64 (__m256i __X, __m256i __Y, __m256i __Z) |
| 74 |
/// \endcode |
| 75 |
/// |
| 76 |
/// This intrinsic corresponds to the \c VPMADD52HUQ instruction. |
| 77 |
/// |
| 78 |
/// \return |
| 79 |
/// return __m256i dst. |
| 80 |
/// \param __X |
| 81 |
/// A 256-bit vector of [4 x i64] |
| 82 |
/// \param __Y |
| 83 |
/// A 256-bit vector of [4 x i64] |
| 84 |
/// \param __Z |
| 85 |
/// A 256-bit vector of [4 x i64] |
| 86 |
/// |
| 87 |
/// \code{.operation} |
| 88 |
/// FOR j := 0 to 3 |
| 89 |
/// i := j*64 |
| 90 |
/// tmp[127:0] := ZeroExtend64(__Y[i+51:i]) * ZeroExtend64(__Z[i+51:i]) |
| 91 |
/// dst[i+63:i] := __X[i+63:i] + ZeroExtend64(tmp[103:52]) |
| 92 |
/// ENDFOR |
| 93 |
/// dst[MAX:256] := 0 |
| 94 |
/// \endcode |
| 95 |
static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 96 |
_mm256_madd52hi_avx_epu64(__m256i __X, __m256i __Y, __m256i __Z) { |
| 97 |
return (__m256i)__builtin_ia32_vpmadd52huq256((__v4di)__X, (__v4di)__Y, |
| 98 |
(__v4di)__Z); |
| 99 |
} |
| 100 |
|
| 101 |
/// Multiply packed unsigned 52-bit integers in each 64-bit element of \a __Y |
| 102 |
/// and \a __Z to form a 104-bit intermediate result. Add the low 52-bit |
| 103 |
/// unsigned integer from the intermediate result with the corresponding |
| 104 |
/// unsigned 64-bit integer in \a __X, and store the results in \a dst. |
| 105 |
/// |
| 106 |
/// \headerfile <immintrin.h> |
| 107 |
/// |
| 108 |
/// \code |
| 109 |
/// __m128i |
| 110 |
/// _mm_madd52lo_avx_epu64 (__m128i __X, __m128i __Y, __m128i __Z) |
| 111 |
/// \endcode |
| 112 |
/// |
| 113 |
/// This intrinsic corresponds to the \c VPMADD52LUQ instruction. |
| 114 |
/// |
| 115 |
/// \return |
| 116 |
/// return __m128i dst. |
| 117 |
/// \param __X |
| 118 |
/// A 128-bit vector of [2 x i64] |
| 119 |
/// \param __Y |
| 120 |
/// A 128-bit vector of [2 x i64] |
| 121 |
/// \param __Z |
| 122 |
/// A 128-bit vector of [2 x i64] |
| 123 |
/// |
| 124 |
/// \code{.operation} |
| 125 |
/// FOR j := 0 to 1 |
| 126 |
/// i := j*64 |
| 127 |
/// tmp[127:0] := ZeroExtend64(__Y[i+51:i]) * ZeroExtend64(__Z[i+51:i]) |
| 128 |
/// dst[i+63:i] := __X[i+63:i] + ZeroExtend64(tmp[51:0]) |
| 129 |
/// ENDFOR |
| 130 |
/// dst[MAX:128] := 0 |
| 131 |
/// \endcode |
| 132 |
static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 133 |
_mm_madd52lo_avx_epu64(__m128i __X, __m128i __Y, __m128i __Z) { |
| 134 |
return (__m128i)__builtin_ia32_vpmadd52luq128((__v2di)__X, (__v2di)__Y, |
| 135 |
(__v2di)__Z); |
| 136 |
} |
| 137 |
|
| 138 |
/// Multiply packed unsigned 52-bit integers in each 64-bit element of \a __Y |
| 139 |
/// and \a __Z to form a 104-bit intermediate result. Add the low 52-bit |
| 140 |
/// unsigned integer from the intermediate result with the corresponding |
| 141 |
/// unsigned 64-bit integer in \a __X, and store the results in \a dst. |
| 142 |
/// |
| 143 |
/// \headerfile <immintrin.h> |
| 144 |
/// |
| 145 |
/// \code |
| 146 |
/// __m256i |
| 147 |
/// _mm256_madd52lo_avx_epu64 (__m256i __X, __m256i __Y, __m256i __Z) |
| 148 |
/// \endcode |
| 149 |
/// |
| 150 |
/// This intrinsic corresponds to the \c VPMADD52LUQ instruction. |
| 151 |
/// |
| 152 |
/// \return |
| 153 |
/// return __m256i dst. |
| 154 |
/// \param __X |
| 155 |
/// A 256-bit vector of [4 x i64] |
| 156 |
/// \param __Y |
| 157 |
/// A 256-bit vector of [4 x i64] |
| 158 |
/// \param __Z |
| 159 |
/// A 256-bit vector of [4 x i64] |
| 160 |
/// |
| 161 |
/// \code{.operation} |
| 162 |
/// FOR j := 0 to 3 |
| 163 |
/// i := j*64 |
| 164 |
/// tmp[127:0] := ZeroExtend64(__Y[i+51:i]) * ZeroExtend64(__Z[i+51:i]) |
| 165 |
/// dst[i+63:i] := __X[i+63:i] + ZeroExtend64(tmp[51:0]) |
| 166 |
/// ENDFOR |
| 167 |
/// dst[MAX:256] := 0 |
| 168 |
/// \endcode |
| 169 |
static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 170 |
_mm256_madd52lo_avx_epu64(__m256i __X, __m256i __Y, __m256i __Z) { |
| 171 |
return (__m256i)__builtin_ia32_vpmadd52luq256((__v4di)__X, (__v4di)__Y, |
| 172 |
(__v4di)__Z); |
| 173 |
} |
| 174 |
#undef __DEFAULT_FN_ATTRS128 |
| 175 |
#undef __DEFAULT_FN_ATTRS256 |
| 176 |
|
| 177 |
#endif // __AVXIFMAINTRIN_H |