| 1 | //===-- Implementation header for exp10f16 ----------------------*- 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_EXP10F16_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_EXP10F16_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/cast.h" |
| 20 | #include "src/__support/FPUtil/except_value_utils.h" |
| 21 | #include "src/__support/FPUtil/rounding_mode.h" |
| 22 | #include "src/__support/macros/config.h" |
| 23 | #include "src/__support/macros/optimization.h" |
| 24 | #include "src/__support/macros/properties/cpu_features.h" |
| 25 | |
| 26 | namespace LIBC_NAMESPACE_DECL { |
| 27 | |
| 28 | namespace math { |
| 29 | |
| 30 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 31 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 32 | LIBC_INLINE_VAR constexpr size_t N_EXP10F16_EXCEPTS = 5; |
| 33 | #else |
| 34 | LIBC_INLINE_VAR constexpr size_t N_EXP10F16_EXCEPTS = 8; |
| 35 | #endif |
| 36 | |
| 37 | LIBC_INLINE_VAR constexpr fputil::ExceptValues<float16, N_EXP10F16_EXCEPTS> |
| 38 | EXP10F16_EXCEPTS = {.values: { |
| 39 | // x = 0x1.8f4p-2, exp10f16(x) = 0x1.3ap+1 (RZ) |
| 40 | {.input: 0x363dU, .rnd_towardzero_result: 0x40e8U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 41 | // x = 0x1.95cp-2, exp10f16(x) = 0x1.3ecp+1 (RZ) |
| 42 | {.input: 0x3657U, .rnd_towardzero_result: 0x40fbU, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
| 43 | // x = -0x1.018p-4, exp10f16(x) = 0x1.bbp-1 (RZ) |
| 44 | {.input: 0xac06U, .rnd_towardzero_result: 0x3aecU, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
| 45 | // x = -0x1.c28p+0, exp10f16(x) = 0x1.1ccp-6 (RZ) |
| 46 | {.input: 0xbf0aU, .rnd_towardzero_result: 0x2473U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
| 47 | // x = -0x1.e1cp+1, exp10f16(x) = 0x1.694p-13 (RZ) |
| 48 | {.input: 0xc387U, .rnd_towardzero_result: 0x09a5U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
| 49 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 50 | // x = 0x1.0cp+1, exp10f16(x) = 0x1.f04p+6 (RZ) |
| 51 | {.input: 0x4030U, .rnd_towardzero_result: 0x57c1U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 52 | // x = 0x1.1b8p+1, exp10f16(x) = 0x1.47cp+7 (RZ) |
| 53 | {.input: 0x406eU, .rnd_towardzero_result: 0x591fU, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 54 | // x = 0x1.1b8p+2, exp10f16(x) = 0x1.a4p+14 (RZ) |
| 55 | {.input: 0x446eU, .rnd_towardzero_result: 0x7690U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
| 56 | #endif |
| 57 | }}; |
| 58 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 59 | |
| 60 | LIBC_INLINE constexpr float16 exp10f16(float16 x) { |
| 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| >= 5, or x is NaN. |
| 68 | if (LIBC_UNLIKELY(x_abs >= 0x4500U)) { |
| 69 | // exp10(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 | // When x >= 5. |
| 80 | if (x_bits.is_pos()) { |
| 81 | // exp10(+inf) = +inf |
| 82 | if (x_bits.is_inf()) |
| 83 | return FPBits::inf().get_val(); |
| 84 | |
| 85 | #ifdef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 86 | fputil::set_errno_if_required(ERANGE); |
| 87 | fputil::raise_except_if_required(FE_OVERFLOW); |
| 88 | return FPBits::inf().get_val(); |
| 89 | #else |
| 90 | switch (fputil::quick_get_round()) { |
| 91 | case FE_TONEAREST: |
| 92 | case FE_UPWARD: |
| 93 | fputil::set_errno_if_required(ERANGE); |
| 94 | fputil::raise_except_if_required(FE_OVERFLOW); |
| 95 | return FPBits::inf().get_val(); |
| 96 | default: |
| 97 | return FPBits::max_normal().get_val(); |
| 98 | } |
| 99 | #endif // LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 100 | } |
| 101 | |
| 102 | // When x <= -8. |
| 103 | if (x_u >= 0xc800U) { |
| 104 | // exp10(-inf) = +0 |
| 105 | if (x_bits.is_inf()) |
| 106 | return FPBits::zero().get_val(); |
| 107 | |
| 108 | fputil::set_errno_if_required(ERANGE); |
| 109 | fputil::raise_except_if_required(FE_UNDERFLOW | FE_INEXACT); |
| 110 | |
| 111 | #ifndef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 112 | if (fputil::fenv_is_round_up()) |
| 113 | return FPBits::min_subnormal().get_val(); |
| 114 | #endif |
| 115 | return FPBits::zero().get_val(); |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | // When x is 1, 2, 3, or 4. These are hard-to-round cases with exact results. |
| 120 | if (LIBC_UNLIKELY((x_u & ~(0x3c00U | 0x4000U | 0x4200U | 0x4400U)) == 0)) { |
| 121 | switch (x_u) { |
| 122 | case 0x3c00U: // x = 1.0f16 |
| 123 | return fputil::cast<float16>(x: 10.0); |
| 124 | case 0x4000U: // x = 2.0f16 |
| 125 | return fputil::cast<float16>(x: 100.0); |
| 126 | case 0x4200U: // x = 3.0f16 |
| 127 | return fputil::cast<float16>(x: 1'000.0); |
| 128 | case 0x4400U: // x = 4.0f16 |
| 129 | return fputil::cast<float16>(x: 10'000.0); |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 134 | if (auto r = EXP10F16_EXCEPTS.lookup(x_bits: x_u); LIBC_UNLIKELY(r.has_value())) |
| 135 | return r.value(); |
| 136 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 137 | |
| 138 | // 10^x = 2^((hi + mid) * log2(10)) * 10^lo |
| 139 | auto [exp2_hi_mid, exp10_lo] = exp10_range_reduction(x); |
| 140 | return fputil::cast<float16>(x: exp2_hi_mid * exp10_lo); |
| 141 | } |
| 142 | |
| 143 | } // namespace math |
| 144 | |
| 145 | } // namespace LIBC_NAMESPACE_DECL |
| 146 | |
| 147 | #endif // LIBC_TYPES_HAS_FLOAT16 |
| 148 | |
| 149 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_EXP10F16_H |
| 150 | |