| 1 | //===-- Implementation header for exp10m1f16 --------------------*- 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_EXP10M1F16_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_EXP10M1F16_H |
| 11 | |
| 12 | #include "include/llvm-libc-macros/float16-macros.h" |
| 13 | |
| 14 | #ifdef LIBC_TYPES_HAS_FLOAT16 |
| 15 | |
| 16 | #include "exp10f16_utils.h" |
| 17 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 18 | #include "src/__support/FPUtil/FPBits.h" |
| 19 | #include "src/__support/FPUtil/PolyEval.h" |
| 20 | #include "src/__support/FPUtil/cast.h" |
| 21 | #include "src/__support/FPUtil/except_value_utils.h" |
| 22 | #include "src/__support/FPUtil/multiply_add.h" |
| 23 | #include "src/__support/FPUtil/rounding_mode.h" |
| 24 | #include "src/__support/common.h" |
| 25 | #include "src/__support/macros/config.h" |
| 26 | #include "src/__support/macros/optimization.h" |
| 27 | #include "src/__support/macros/properties/cpu_features.h" |
| 28 | |
| 29 | namespace LIBC_NAMESPACE_DECL { |
| 30 | |
| 31 | namespace math { |
| 32 | |
| 33 | LIBC_INLINE constexpr float16 exp10m1f16(float16 x) { |
| 34 | |
| 35 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 36 | constexpr fputil::ExceptValues<float16, 3> EXP10M1F16_EXCEPTS_LO = {.values: { |
| 37 | // (input, RZ output, RU offset, RD offset, RN offset) |
| 38 | // x = 0x1.5c4p-4, exp10m1f16(x) = 0x1.bacp-3 (RZ) |
| 39 | {.input: 0x2d71U, .rnd_towardzero_result: 0x32ebU, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
| 40 | // x = -0x1.5ep-13, exp10m1f16(x) = -0x1.92cp-12 (RZ) |
| 41 | {.input: 0x8978U, .rnd_towardzero_result: 0x8e4bU, .rnd_upward_offset: 0U, .rnd_downward_offset: 1U, .rnd_tonearest_offset: 0U}, |
| 42 | // x = -0x1.e2p-10, exp10m1f16(x) = -0x1.14cp-8 (RZ) |
| 43 | {.input: 0x9788U, .rnd_towardzero_result: 0x9c53U, .rnd_upward_offset: 0U, .rnd_downward_offset: 1U, .rnd_tonearest_offset: 0U}, |
| 44 | }}; |
| 45 | |
| 46 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 47 | constexpr size_t N_EXP10M1F16_EXCEPTS_HI = 3; |
| 48 | #else |
| 49 | constexpr size_t N_EXP10M1F16_EXCEPTS_HI = 6; |
| 50 | #endif |
| 51 | |
| 52 | constexpr fputil::ExceptValues<float16, N_EXP10M1F16_EXCEPTS_HI> |
| 53 | EXP10M1F16_EXCEPTS_HI = {.values: { |
| 54 | // (input, RZ output, RU offset, RD offset, RN offset) |
| 55 | // x = 0x1.8f4p-2, exp10m1f16(x) = 0x1.744p+0 (RZ) |
| 56 | {.input: 0x363dU, .rnd_towardzero_result: 0x3dd1U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
| 57 | // x = 0x1.95cp-2, exp10m1f16(x) = 0x1.7d8p+0 (RZ) |
| 58 | {.input: 0x3657U, .rnd_towardzero_result: 0x3df6U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
| 59 | // x = 0x1.d04p-2, exp10m1f16(x) = 0x1.d7p+0 (RZ) |
| 60 | {.input: 0x3741U, .rnd_towardzero_result: 0x3f5cU, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 61 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 62 | // x = 0x1.0cp+1, exp10m1f16(x) = 0x1.ec4p+6 (RZ) |
| 63 | {.input: 0x4030U, .rnd_towardzero_result: 0x57b1U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 64 | // x = 0x1.1b8p+1, exp10m1f16(x) = 0x1.45cp+7 (RZ) |
| 65 | {.input: 0x406eU, .rnd_towardzero_result: 0x5917U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 66 | // x = 0x1.2f4p+2, exp10m1f16(x) = 0x1.ab8p+15 (RZ) |
| 67 | {.input: 0x44bdU, .rnd_towardzero_result: 0x7aaeU, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 68 | #endif |
| 69 | }}; |
| 70 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 71 | |
| 72 | using FPBits = fputil::FPBits<float16>; |
| 73 | FPBits x_bits(x); |
| 74 | |
| 75 | uint16_t x_u = x_bits.uintval(); |
| 76 | uint16_t x_abs = x_u & 0x7fffU; |
| 77 | |
| 78 | // When |x| <= 2^(-3), or |x| >= 11 * log10(2), or x is NaN. |
| 79 | if (LIBC_UNLIKELY(x_abs <= 0x3000U || x_abs >= 0x429fU)) { |
| 80 | // exp10m1(NaN) = NaN |
| 81 | if (x_bits.is_nan()) { |
| 82 | if (x_bits.is_signaling_nan()) { |
| 83 | fputil::raise_except_if_required(FE_INVALID); |
| 84 | return FPBits::quiet_nan().get_val(); |
| 85 | } |
| 86 | |
| 87 | return x; |
| 88 | } |
| 89 | |
| 90 | // When x >= 16 * log10(2). |
| 91 | if (x_u >= 0x44d1U && x_bits.is_pos()) { |
| 92 | // exp10m1(+inf) = +inf |
| 93 | if (x_bits.is_inf()) |
| 94 | return FPBits::inf().get_val(); |
| 95 | |
| 96 | #ifdef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 97 | fputil::set_errno_if_required(ERANGE); |
| 98 | fputil::raise_except_if_required(FE_OVERFLOW | FE_INEXACT); |
| 99 | return FPBits::inf().get_val(); |
| 100 | #else |
| 101 | switch (fputil::quick_get_round()) { |
| 102 | case FE_TONEAREST: |
| 103 | case FE_UPWARD: |
| 104 | fputil::set_errno_if_required(ERANGE); |
| 105 | fputil::raise_except_if_required(FE_OVERFLOW | FE_INEXACT); |
| 106 | return FPBits::inf().get_val(); |
| 107 | default: |
| 108 | return FPBits::max_normal().get_val(); |
| 109 | } |
| 110 | #endif // LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 111 | } |
| 112 | |
| 113 | // When x < -11 * log10(2). |
| 114 | if (x_u > 0xc29fU) { |
| 115 | // exp10m1(-inf) = -1 |
| 116 | if (x_bits.is_inf()) |
| 117 | return FPBits::one(sign: Sign::NEG).get_val(); |
| 118 | |
| 119 | // When x >= -0x1.ce4p+1, round(10^x - 1, HP, RN) = -0x1.ffcp-1. |
| 120 | if (x_u <= 0xc339U) { |
| 121 | return fputil::round_result_slightly_down( |
| 122 | value_rn: fputil::cast<float16>(x: -0x1.ffcp-1)); |
| 123 | } |
| 124 | |
| 125 | // When x < -0x1.ce4p+1, round(10^x - 1, HP, RN) = -1. |
| 126 | #ifdef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 127 | fputil::raise_except_if_required(FE_INEXACT); |
| 128 | return FPBits::one(Sign::NEG).get_val(); |
| 129 | #else |
| 130 | switch (fputil::quick_get_round()) { |
| 131 | case FE_TONEAREST: |
| 132 | case FE_DOWNWARD: |
| 133 | fputil::raise_except_if_required(FE_INEXACT); |
| 134 | return FPBits::one(sign: Sign::NEG).get_val(); |
| 135 | default: |
| 136 | fputil::raise_except_if_required(FE_INEXACT); |
| 137 | return fputil::cast<float16>(x: -0x1.ffcp-1); |
| 138 | } |
| 139 | #endif // LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 140 | } |
| 141 | |
| 142 | // When |x| <= 2^(-3). |
| 143 | if (x_abs <= 0x3000U) { |
| 144 | if (LIBC_UNLIKELY(x_abs == 0)) |
| 145 | return x; |
| 146 | |
| 147 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 148 | if (auto r = EXP10M1F16_EXCEPTS_LO.lookup(x_bits: x_u); |
| 149 | LIBC_UNLIKELY(r.has_value())) |
| 150 | return r.value(); |
| 151 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 152 | |
| 153 | float xf = x; |
| 154 | // Degree-5 minimax polynomial generated by Sollya with the following |
| 155 | // commands: |
| 156 | // > display = hexadecimal; |
| 157 | // > P = fpminimax((10^x - 1)/x, 4, [|SG...|], [-2^-3, 2^-3]); |
| 158 | // > x * P; |
| 159 | return fputil::cast<float16>( |
| 160 | x: xf * fputil::polyeval(x: xf, a0: 0x1.26bb1cp+1f, a: 0x1.5351c8p+1f, |
| 161 | a: 0x1.04704p+1f, a: 0x1.2ce084p+0f, a: 0x1.14a6bep-1f)); |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | // When x is 1, 2, or 3. These are hard-to-round cases with exact results. |
| 166 | // 10^4 - 1 = 9'999 is not exactly representable as a float16, but luckily |
| 167 | // the polynomial approximation gives the correct result for x = 4 in all |
| 168 | // rounding modes. |
| 169 | if (LIBC_UNLIKELY((x_u & ~(0x3c00U | 0x4000U | 0x4200U | 0x4400U)) == 0)) { |
| 170 | switch (x_u) { |
| 171 | case 0x3c00U: // x = 1.0f16 |
| 172 | return fputil::cast<float16>(x: 9.0); |
| 173 | case 0x4000U: // x = 2.0f16 |
| 174 | return fputil::cast<float16>(x: 99.0); |
| 175 | case 0x4200U: // x = 3.0f16 |
| 176 | return fputil::cast<float16>(x: 999.0); |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 181 | if (auto r = EXP10M1F16_EXCEPTS_HI.lookup(x_bits: x_u); LIBC_UNLIKELY(r.has_value())) |
| 182 | return r.value(); |
| 183 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 184 | |
| 185 | // exp10(x) = exp2((hi + mid) * log2(10)) * exp10(lo) |
| 186 | auto [exp2_hi_mid, exp10_lo] = exp10_range_reduction(x); |
| 187 | // exp10m1(x) = exp2((hi + mid) * log2(lo)) * exp10(lo) - 1 |
| 188 | return fputil::cast<float16>( |
| 189 | x: fputil::multiply_add(x: exp2_hi_mid, y: exp10_lo, z: -1.0f)); |
| 190 | } |
| 191 | |
| 192 | } // namespace math |
| 193 | |
| 194 | } // namespace LIBC_NAMESPACE_DECL |
| 195 | |
| 196 | #endif // LIBC_TYPES_HAS_FLOAT16 |
| 197 | |
| 198 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_EXP10M1F16_H |
| 199 | |