| 1 | //===----------------------------------------------------------------------===// |
| 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 | /// \file |
| 10 | /// Common header for multiply-add and multiply-sub implementations. |
| 11 | /// |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_LIBC_SRC___SUPPORT_FPUTIL_MULTIPLY_ADD_H |
| 15 | #define LLVM_LIBC_SRC___SUPPORT_FPUTIL_MULTIPLY_ADD_H |
| 16 | |
| 17 | #include "src/__support/CPP/type_traits.h" |
| 18 | #include "src/__support/common.h" |
| 19 | #include "src/__support/macros/config.h" |
| 20 | #include "src/__support/macros/optimization.h" |
| 21 | #include "src/__support/macros/properties/architectures.h" |
| 22 | #include "src/__support/macros/properties/compiler.h" |
| 23 | #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA |
| 24 | |
| 25 | #if defined(LIBC_TARGET_CPU_HAS_FMA) |
| 26 | #if !__has_builtin(__builtin_elementwise_fma) && defined(LIBC_COMPILER_IS_MSVC) |
| 27 | #if defined(LIBC_TARGET_ARCH_IS_X86) |
| 28 | #include <immintrin.h> |
| 29 | #elif defined(LIBC_TARGET_ARCH_IS_AARCH64) |
| 30 | #include <arm64_neon.h> |
| 31 | #endif |
| 32 | #endif |
| 33 | #endif // LIBC_TARGET_CPU_HAS_FMA |
| 34 | |
| 35 | namespace LIBC_NAMESPACE_DECL { |
| 36 | namespace fputil { |
| 37 | |
| 38 | namespace internal { |
| 39 | |
| 40 | template <typename T> struct HasHardwareFma : cpp::false_type {}; |
| 41 | |
| 42 | template <typename T> T fma_impl(T, T, T) = delete; |
| 43 | template <typename T> T fms_impl(T, T, T) = delete; |
| 44 | template <typename T> T fnma_impl(T, T, T) = delete; |
| 45 | |
| 46 | // Hardware FMA helpers use builtins directly instead of including FMA.h to |
| 47 | // avoid a circular dependency: |
| 48 | // multiply_add.h -> FMA.h -> generic/FMA.h -> dyadic_float.h. |
| 49 | #if defined(LIBC_TARGET_CPU_HAS_FMA) |
| 50 | |
| 51 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 52 | template <> struct HasHardwareFma<float> : cpp::true_type {}; |
| 53 | |
| 54 | LIBC_ALWAYS_INLINE float fma_impl(float x, float y, float z) { |
| 55 | #if __has_builtin(__builtin_elementwise_fma) |
| 56 | return __builtin_elementwise_fma(x, y, z); |
| 57 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_X86) |
| 58 | return _mm_cvtss_f32( |
| 59 | _mm_fmadd_ss(_mm_set_ss(x), _mm_set_ss(y), _mm_set_ss(z))); |
| 60 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_AARCH64) |
| 61 | return vget_lane_f32(vfma_f32(vdup_n_f32(z), vdup_n_f32(x), vdup_n_f32(y)), |
| 62 | 0); |
| 63 | #else |
| 64 | return __builtin_fmaf(x, y, z); |
| 65 | #endif |
| 66 | } |
| 67 | |
| 68 | LIBC_ALWAYS_INLINE float fms_impl(float x, float y, float z) { |
| 69 | #if __has_builtin(__builtin_elementwise_fma) |
| 70 | return __builtin_elementwise_fma(x, y, -z); |
| 71 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_X86) |
| 72 | return _mm_cvtss_f32( |
| 73 | _mm_fmsub_ss(_mm_set_ss(x), _mm_set_ss(y), _mm_set_ss(z))); |
| 74 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_AARCH64) |
| 75 | return vget_lane_f32(vfma_f32(vdup_n_f32(-z), vdup_n_f32(x), vdup_n_f32(y)), |
| 76 | 0); |
| 77 | #else |
| 78 | return __builtin_fmaf(x, y, -z); |
| 79 | #endif |
| 80 | } |
| 81 | |
| 82 | LIBC_ALWAYS_INLINE float fnma_impl(float x, float y, float z) { |
| 83 | #if __has_builtin(__builtin_elementwise_fma) |
| 84 | return __builtin_elementwise_fma(-x, y, z); |
| 85 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_X86) |
| 86 | return _mm_cvtss_f32( |
| 87 | _mm_fnmadd_ss(_mm_set_ss(x), _mm_set_ss(y), _mm_set_ss(z))); |
| 88 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_AARCH64) |
| 89 | return vget_lane_f32(vfms_f32(vdup_n_f32(z), vdup_n_f32(x), vdup_n_f32(y)), |
| 90 | 0); |
| 91 | #else |
| 92 | return __builtin_fmaf(-x, y, z); |
| 93 | #endif |
| 94 | } |
| 95 | #endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 96 | |
| 97 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 98 | template <> struct HasHardwareFma<double> : cpp::true_type {}; |
| 99 | |
| 100 | LIBC_ALWAYS_INLINE double fma_impl(double x, double y, double z) { |
| 101 | #if __has_builtin(__builtin_elementwise_fma) |
| 102 | return __builtin_elementwise_fma(x, y, z); |
| 103 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_X86) |
| 104 | return _mm_cvtsd_f64( |
| 105 | _mm_fmadd_sd(_mm_set_sd(x), _mm_set_sd(y), _mm_set_sd(z))); |
| 106 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_AARCH64) |
| 107 | return vget_lane_f64(vfma_f64(vdup_n_f64(z), vdup_n_f64(x), vdup_n_f64(y)), |
| 108 | 0); |
| 109 | #else |
| 110 | return __builtin_fma(x, y, z); |
| 111 | #endif |
| 112 | } |
| 113 | |
| 114 | LIBC_ALWAYS_INLINE double fms_impl(double x, double y, double z) { |
| 115 | #if __has_builtin(__builtin_elementwise_fma) |
| 116 | return __builtin_elementwise_fma(x, y, -z); |
| 117 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_X86) |
| 118 | return _mm_cvtsd_f64( |
| 119 | _mm_fmsub_sd(_mm_set_sd(x), _mm_set_sd(y), _mm_set_sd(z))); |
| 120 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_AARCH64) |
| 121 | return vget_lane_f64(vfma_f64(vdup_n_f64(-z), vdup_n_f64(x), vdup_n_f64(y)), |
| 122 | 0); |
| 123 | #else |
| 124 | return __builtin_fma(x, y, -z); |
| 125 | #endif |
| 126 | } |
| 127 | |
| 128 | LIBC_ALWAYS_INLINE double fnma_impl(double x, double y, double z) { |
| 129 | #if __has_builtin(__builtin_elementwise_fma) |
| 130 | return __builtin_elementwise_fma(-x, y, z); |
| 131 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_X86) |
| 132 | return _mm_cvtsd_f64( |
| 133 | _mm_fnmadd_sd(_mm_set_sd(x), _mm_set_sd(y), _mm_set_sd(z))); |
| 134 | #elif defined(LIBC_COMPILER_IS_MSVC) && defined(LIBC_TARGET_ARCH_IS_AARCH64) |
| 135 | return vget_lane_f64(vfms_f64(vdup_n_f64(z), vdup_n_f64(x), vdup_n_f64(y)), |
| 136 | 0); |
| 137 | #else |
| 138 | return __builtin_fma(-x, y, z); |
| 139 | #endif |
| 140 | } |
| 141 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 142 | |
| 143 | #endif // LIBC_TARGET_CPU_HAS_FMA |
| 144 | |
| 145 | } // namespace internal |
| 146 | |
| 147 | // Compute x * y + z, using hardware FMA when available. |
| 148 | template <typename T> |
| 149 | LIBC_INLINE constexpr cpp::enable_if_t<(sizeof(T) > sizeof(double)), T> |
| 150 | multiply_add(const T &x, const T &y, const T &z) { |
| 151 | T prod = x * y; |
| 152 | return prod + z; |
| 153 | } |
| 154 | |
| 155 | template <typename T> |
| 156 | LIBC_ALWAYS_INLINE constexpr cpp::enable_if_t<(sizeof(T) <= sizeof(double)), T> |
| 157 | multiply_add(T x, T y, T z) { |
| 158 | if constexpr (internal::HasHardwareFma<T>::value) { |
| 159 | if (!cpp::is_constant_evaluated()) |
| 160 | return internal::fma_impl(x, y, z); |
| 161 | } |
| 162 | T prod = x * y; |
| 163 | return prod + z; |
| 164 | } |
| 165 | |
| 166 | // Compute x * y - z, using hardware FMA when available. |
| 167 | template <typename T> |
| 168 | LIBC_INLINE constexpr cpp::enable_if_t<(sizeof(T) > sizeof(double)), T> |
| 169 | multiply_sub(const T &x, const T &y, const T &z) { |
| 170 | T prod = x * y; |
| 171 | return prod - z; |
| 172 | } |
| 173 | |
| 174 | template <typename T> |
| 175 | LIBC_ALWAYS_INLINE constexpr cpp::enable_if_t<(sizeof(T) <= sizeof(double)), T> |
| 176 | multiply_sub(T x, T y, T z) { |
| 177 | if constexpr (internal::HasHardwareFma<T>::value) { |
| 178 | if (!cpp::is_constant_evaluated()) |
| 179 | return internal::fms_impl(x, y, z); |
| 180 | } |
| 181 | T prod = x * y; |
| 182 | return prod - z; |
| 183 | } |
| 184 | |
| 185 | // Compute x - y * z, using hardware FMA when available. |
| 186 | // Without FMA, when directed rounding modes (FE_UPWARD / FE_DOWNWARD) may be |
| 187 | // active, evaluate x + ((-y) * z) instead of x - (y * z) for floating-point |
| 188 | // types so that (-y) * z is rounded in the same direction as the outer sum |
| 189 | // when y or z is constant (e.g., for x = min_denormal and 0 < z < 1 in |
| 190 | // FE_UPWARD, RU(x + RU((-x) * z)) = x, whereas RU(x - RU(x * z)) = 0). |
| 191 | template <typename T> |
| 192 | LIBC_INLINE constexpr cpp::enable_if_t<(sizeof(T) > sizeof(double)), T> |
| 193 | sub_multiply(const T &x, const T &y, const T &z) { |
| 194 | #ifndef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 195 | if constexpr (cpp::is_floating_point_v<T>) { |
| 196 | T prod = (-y) * z; |
| 197 | return x + prod; |
| 198 | } |
| 199 | #endif // !LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 200 | T prod = y * z; |
| 201 | return x - prod; |
| 202 | } |
| 203 | |
| 204 | template <typename T> |
| 205 | LIBC_ALWAYS_INLINE constexpr cpp::enable_if_t<(sizeof(T) <= sizeof(double)), T> |
| 206 | sub_multiply(T x, T y, T z) { |
| 207 | if constexpr (internal::HasHardwareFma<T>::value) { |
| 208 | if (!cpp::is_constant_evaluated()) |
| 209 | return internal::fnma_impl(y, z, x); |
| 210 | } |
| 211 | #ifndef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 212 | if constexpr (cpp::is_floating_point_v<T>) { |
| 213 | T prod = (-y) * z; |
| 214 | return x + prod; |
| 215 | } |
| 216 | #endif // !LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 217 | T prod = y * z; |
| 218 | return x - prod; |
| 219 | } |
| 220 | |
| 221 | } // namespace fputil |
| 222 | } // namespace LIBC_NAMESPACE_DECL |
| 223 | |
| 224 | #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_MULTIPLY_ADD_H |
| 225 | |