| 1 | //===-- Implementation header for expm1f16 ----------------------*- C++ -*-===// |
| 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 | #ifndef LLVM_LIBC_SRC___SUPPORT_MATH_EXPM1F16_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_EXPM1F16_H |
| 11 | |
| 12 | #include "include/llvm-libc-macros/float16-macros.h" |
| 13 | |
| 14 | #ifdef LIBC_TYPES_HAS_FLOAT16 |
| 15 | |
| 16 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 17 | #include "src/__support/FPUtil/FPBits.h" |
| 18 | #include "src/__support/FPUtil/PolyEval.h" |
| 19 | #include "src/__support/FPUtil/cast.h" |
| 20 | #include "src/__support/FPUtil/except_value_utils.h" |
| 21 | #include "src/__support/FPUtil/multiply_add.h" |
| 22 | #include "src/__support/FPUtil/rounding_mode.h" |
| 23 | #include "src/__support/common.h" |
| 24 | #include "src/__support/macros/config.h" |
| 25 | #include "src/__support/macros/optimization.h" |
| 26 | #include "src/__support/math/expxf16_utils.h" |
| 27 | |
| 28 | namespace LIBC_NAMESPACE_DECL { |
| 29 | |
| 30 | namespace math { |
| 31 | |
| 32 | LIBC_INLINE constexpr float16 expm1f16(float16 x) { |
| 33 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 34 | constexpr fputil::ExceptValues<float16, 1> EXPM1F16_EXCEPTS_LO = {.values: { |
| 35 | // (input, RZ output, RU offset, RD offset, RN offset) |
| 36 | // x = 0x1.564p-5, expm1f16(x) = 0x1.5d4p-5 (RZ) |
| 37 | {.input: 0x2959U, .rnd_towardzero_result: 0x2975U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 38 | }}; |
| 39 | |
| 40 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 41 | constexpr size_t N_EXPM1F16_EXCEPTS_HI = 2; |
| 42 | #else |
| 43 | constexpr size_t N_EXPM1F16_EXCEPTS_HI = 3; |
| 44 | #endif |
| 45 | |
| 46 | constexpr fputil::ExceptValues<float16, N_EXPM1F16_EXCEPTS_HI> |
| 47 | EXPM1F16_EXCEPTS_HI = {.values: { |
| 48 | // (input, RZ output, RU offset, RD offset, RN offset) |
| 49 | // x = 0x1.c34p+0, expm1f16(x) = 0x1.34cp+2 (RZ) |
| 50 | {.input: 0x3f0dU, .rnd_towardzero_result: 0x44d3U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 51 | // x = -0x1.e28p-3, expm1f16(x) = -0x1.adcp-3 (RZ) |
| 52 | {.input: 0xb38aU, .rnd_towardzero_result: 0xb2b7U, .rnd_upward_offset: 0U, .rnd_downward_offset: 1U, .rnd_tonearest_offset: 1U}, |
| 53 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 54 | // x = 0x1.a08p-3, exp10m1f(x) = 0x1.cdcp-3 (RZ) |
| 55 | {.input: 0x3282U, .rnd_towardzero_result: 0x3337U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
| 56 | #endif |
| 57 | }}; |
| 58 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 59 | |
| 60 | using namespace math::expxf16_internal; |
| 61 | using FPBits = fputil::FPBits<float16>; |
| 62 | FPBits x_bits(x); |
| 63 | |
| 64 | uint16_t x_u = x_bits.uintval(); |
| 65 | uint16_t x_abs = x_u & 0x7fffU; |
| 66 | |
| 67 | // When |x| <= 2^(-3), or |x| >= -11 * log(2), or x is NaN. |
| 68 | if (LIBC_UNLIKELY(x_abs <= 0x3000U || x_abs >= 0x47a0U)) { |
| 69 | // expm1(NaN) = NaN |
| 70 | if (x_bits.is_nan()) { |
| 71 | if (x_bits.is_signaling_nan()) { |
| 72 | fputil::raise_except_if_required(FE_INVALID); |
| 73 | return FPBits::quiet_nan().get_val(); |
| 74 | } |
| 75 | |
| 76 | return x; |
| 77 | } |
| 78 | |
| 79 | // expm1(+/-0) = +/-0 |
| 80 | if (x_abs == 0) |
| 81 | return x; |
| 82 | |
| 83 | // When x >= 16 * log(2). |
| 84 | if (x_bits.is_pos() && x_abs >= 0x498cU) { |
| 85 | // expm1(+inf) = +inf |
| 86 | if (x_bits.is_inf()) |
| 87 | return FPBits::inf().get_val(); |
| 88 | |
| 89 | #ifdef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 90 | fputil::set_errno_if_required(ERANGE); |
| 91 | fputil::raise_except_if_required(FE_OVERFLOW | FE_INEXACT); |
| 92 | return FPBits::inf().get_val(); |
| 93 | #else |
| 94 | switch (fputil::quick_get_round()) { |
| 95 | case FE_TONEAREST: |
| 96 | case FE_UPWARD: |
| 97 | fputil::set_errno_if_required(ERANGE); |
| 98 | fputil::raise_except_if_required(FE_OVERFLOW | FE_INEXACT); |
| 99 | return FPBits::inf().get_val(); |
| 100 | default: |
| 101 | return FPBits::max_normal().get_val(); |
| 102 | } |
| 103 | #endif // LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 104 | } |
| 105 | |
| 106 | // When x <= -11 * log(2). |
| 107 | if (x_u >= 0xc7a0U) { |
| 108 | // expm1(-inf) = -1 |
| 109 | if (x_bits.is_inf()) |
| 110 | return FPBits::one(sign: Sign::NEG).get_val(); |
| 111 | |
| 112 | // When x > -0x1.0ap+3, round(expm1(x), HP, RN) = -1. |
| 113 | if (x_u > 0xc828U) |
| 114 | return fputil::round_result_slightly_up( |
| 115 | value_rn: FPBits::one(sign: Sign::NEG).get_val()); |
| 116 | // When x <= -0x1.0ap+3, round(expm1(x), HP, RN) = -0x1.ffcp-1. |
| 117 | return fputil::round_result_slightly_down( |
| 118 | value_rn: fputil::cast<float16>(x: -0x1.ffcp-1)); |
| 119 | } |
| 120 | |
| 121 | // When 0 < |x| <= 2^(-3). |
| 122 | if (x_abs <= 0x3000U && !x_bits.is_zero()) { |
| 123 | |
| 124 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 125 | if (auto r = EXPM1F16_EXCEPTS_LO.lookup(x_bits: x_u); |
| 126 | LIBC_UNLIKELY(r.has_value())) |
| 127 | return r.value(); |
| 128 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 129 | |
| 130 | float xf = x; |
| 131 | // Degree-5 minimax polynomial generated by Sollya with the following |
| 132 | // commands: |
| 133 | // > display = hexadecimal; |
| 134 | // > P = fpminimax(expm1(x)/x, 4, [|SG...|], [-2^-3, 2^-3]); |
| 135 | // > x * P; |
| 136 | return fputil::cast<float16>( |
| 137 | x: xf * fputil::polyeval(x: xf, a0: 0x1p+0f, a: 0x1.fffff8p-2f, a: 0x1.555556p-3f, |
| 138 | a: 0x1.55905ep-5f, a: 0x1.1124c2p-7f)); |
| 139 | } |
| 140 | } |
| 141 | |
| 142 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 143 | if (auto r = EXPM1F16_EXCEPTS_HI.lookup(x_bits: x_u); LIBC_UNLIKELY(r.has_value())) |
| 144 | return r.value(); |
| 145 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 146 | |
| 147 | // exp(x) = exp(hi + mid) * exp(lo) |
| 148 | auto [exp_hi_mid, exp_lo] = exp_range_reduction(x); |
| 149 | // expm1(x) = exp(hi + mid) * exp(lo) - 1 |
| 150 | return fputil::cast<float16>(x: fputil::multiply_add(x: exp_hi_mid, y: exp_lo, z: -1.0f)); |
| 151 | } |
| 152 | |
| 153 | } // namespace math |
| 154 | |
| 155 | } // namespace LIBC_NAMESPACE_DECL |
| 156 | |
| 157 | #endif // LIBC_TYPES_HAS_FLOAT16 |
| 158 | |
| 159 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_EXPM1F16_H |
| 160 | |