1 | //===-- Half-precision 10^x function --------------------------------------===// |
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 | #include "src/math/exp10f16.h" |
10 | #include "expxf16.h" |
11 | #include "hdr/errno_macros.h" |
12 | #include "hdr/fenv_macros.h" |
13 | #include "src/__support/CPP/array.h" |
14 | #include "src/__support/FPUtil/FEnvImpl.h" |
15 | #include "src/__support/FPUtil/FPBits.h" |
16 | #include "src/__support/FPUtil/PolyEval.h" |
17 | #include "src/__support/FPUtil/cast.h" |
18 | #include "src/__support/FPUtil/except_value_utils.h" |
19 | #include "src/__support/FPUtil/multiply_add.h" |
20 | #include "src/__support/FPUtil/nearest_integer.h" |
21 | #include "src/__support/FPUtil/rounding_mode.h" |
22 | #include "src/__support/common.h" |
23 | #include "src/__support/macros/config.h" |
24 | #include "src/__support/macros/optimization.h" |
25 | #include "src/__support/macros/properties/cpu_features.h" |
26 | |
27 | namespace LIBC_NAMESPACE_DECL { |
28 | |
29 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
30 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
31 | static constexpr size_t N_EXP10F16_EXCEPTS = 5; |
32 | #else |
33 | static constexpr size_t N_EXP10F16_EXCEPTS = 8; |
34 | #endif |
35 | |
36 | static constexpr fputil::ExceptValues<float16, N_EXP10F16_EXCEPTS> |
37 | EXP10F16_EXCEPTS = {{ |
38 | // x = 0x1.8f4p-2, exp10f16(x) = 0x1.3ap+1 (RZ) |
39 | {0x363dU, 0x40e8U, 1U, 0U, 1U}, |
40 | // x = 0x1.95cp-2, exp10f16(x) = 0x1.3ecp+1 (RZ) |
41 | {0x3657U, 0x40fbU, 1U, 0U, 0U}, |
42 | // x = -0x1.018p-4, exp10f16(x) = 0x1.bbp-1 (RZ) |
43 | {0xac06U, 0x3aecU, 1U, 0U, 0U}, |
44 | // x = -0x1.c28p+0, exp10f16(x) = 0x1.1ccp-6 (RZ) |
45 | {0xbf0aU, 0x2473U, 1U, 0U, 0U}, |
46 | // x = -0x1.e1cp+1, exp10f16(x) = 0x1.694p-13 (RZ) |
47 | {0xc387U, 0x09a5U, 1U, 0U, 0U}, |
48 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
49 | // x = 0x1.0cp+1, exp10f16(x) = 0x1.f04p+6 (RZ) |
50 | {0x4030U, 0x57c1U, 1U, 0U, 1U}, |
51 | // x = 0x1.1b8p+1, exp10f16(x) = 0x1.47cp+7 (RZ) |
52 | {0x406eU, 0x591fU, 1U, 0U, 1U}, |
53 | // x = 0x1.1b8p+2, exp10f16(x) = 0x1.a4p+14 (RZ) |
54 | {0x446eU, 0x7690U, 1U, 0U, 1U}, |
55 | #endif |
56 | }}; |
57 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
58 | |
59 | LLVM_LIBC_FUNCTION(float16, exp10f16, (float16 x)) { |
60 | using FPBits = fputil::FPBits<float16>; |
61 | FPBits x_bits(x); |
62 | |
63 | uint16_t x_u = x_bits.uintval(); |
64 | uint16_t x_abs = x_u & 0x7fffU; |
65 | |
66 | // When |x| >= 5, or x is NaN. |
67 | if (LIBC_UNLIKELY(x_abs >= 0x4500U)) { |
68 | // exp10(NaN) = NaN |
69 | if (x_bits.is_nan()) { |
70 | if (x_bits.is_signaling_nan()) { |
71 | fputil::raise_except_if_required(FE_INVALID); |
72 | return FPBits::quiet_nan().get_val(); |
73 | } |
74 | |
75 | return x; |
76 | } |
77 | |
78 | // When x >= 5. |
79 | if (x_bits.is_pos()) { |
80 | // exp10(+inf) = +inf |
81 | if (x_bits.is_inf()) |
82 | return FPBits::inf().get_val(); |
83 | |
84 | switch (fputil::quick_get_round()) { |
85 | case FE_TONEAREST: |
86 | case FE_UPWARD: |
87 | fputil::set_errno_if_required(ERANGE); |
88 | fputil::raise_except_if_required(FE_OVERFLOW); |
89 | return FPBits::inf().get_val(); |
90 | default: |
91 | return FPBits::max_normal().get_val(); |
92 | } |
93 | } |
94 | |
95 | // When x <= -8. |
96 | if (x_u >= 0xc800U) { |
97 | // exp10(-inf) = +0 |
98 | if (x_bits.is_inf()) |
99 | return FPBits::zero().get_val(); |
100 | |
101 | fputil::set_errno_if_required(ERANGE); |
102 | fputil::raise_except_if_required(FE_UNDERFLOW | FE_INEXACT); |
103 | |
104 | if (fputil::fenv_is_round_up()) |
105 | return FPBits::min_subnormal().get_val(); |
106 | return FPBits::zero().get_val(); |
107 | } |
108 | } |
109 | |
110 | // When x is 1, 2, 3, or 4. These are hard-to-round cases with exact results. |
111 | if (LIBC_UNLIKELY((x_u & ~(0x3c00U | 0x4000U | 0x4200U | 0x4400U)) == 0)) { |
112 | switch (x_u) { |
113 | case 0x3c00U: // x = 1.0f16 |
114 | return fputil::cast<float16>(10.0); |
115 | case 0x4000U: // x = 2.0f16 |
116 | return fputil::cast<float16>(100.0); |
117 | case 0x4200U: // x = 3.0f16 |
118 | return fputil::cast<float16>(1'000.0); |
119 | case 0x4400U: // x = 4.0f16 |
120 | return fputil::cast<float16>(10'000.0); |
121 | } |
122 | } |
123 | |
124 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
125 | if (auto r = EXP10F16_EXCEPTS.lookup(x_u); LIBC_UNLIKELY(r.has_value())) |
126 | return r.value(); |
127 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
128 | |
129 | // 10^x = 2^((hi + mid) * log2(10)) * 10^lo |
130 | auto [exp2_hi_mid, exp10_lo] = exp10_range_reduction(x); |
131 | return fputil::cast<float16>(exp2_hi_mid * exp10_lo); |
132 | } |
133 | |
134 | } // namespace LIBC_NAMESPACE_DECL |
135 | |