1 | //===-- Half-precision e^x - 1 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/expm1f16.h" |
10 | #include "expxf16.h" |
11 | #include "hdr/errno_macros.h" |
12 | #include "hdr/fenv_macros.h" |
13 | #include "src/__support/FPUtil/FEnvImpl.h" |
14 | #include "src/__support/FPUtil/FPBits.h" |
15 | #include "src/__support/FPUtil/PolyEval.h" |
16 | #include "src/__support/FPUtil/cast.h" |
17 | #include "src/__support/FPUtil/except_value_utils.h" |
18 | #include "src/__support/FPUtil/multiply_add.h" |
19 | #include "src/__support/FPUtil/rounding_mode.h" |
20 | #include "src/__support/common.h" |
21 | #include "src/__support/macros/config.h" |
22 | #include "src/__support/macros/optimization.h" |
23 | |
24 | namespace LIBC_NAMESPACE_DECL { |
25 | |
26 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
27 | static constexpr fputil::ExceptValues<float16, 1> EXPM1F16_EXCEPTS_LO = {{ |
28 | // (input, RZ output, RU offset, RD offset, RN offset) |
29 | // x = 0x1.564p-5, expm1f16(x) = 0x1.5d4p-5 (RZ) |
30 | {0x2959U, 0x2975U, 1U, 0U, 1U}, |
31 | }}; |
32 | |
33 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
34 | static constexpr size_t N_EXPM1F16_EXCEPTS_HI = 2; |
35 | #else |
36 | static constexpr size_t N_EXPM1F16_EXCEPTS_HI = 3; |
37 | #endif |
38 | |
39 | static constexpr fputil::ExceptValues<float16, N_EXPM1F16_EXCEPTS_HI> |
40 | EXPM1F16_EXCEPTS_HI = {{ |
41 | // (input, RZ output, RU offset, RD offset, RN offset) |
42 | // x = 0x1.c34p+0, expm1f16(x) = 0x1.34cp+2 (RZ) |
43 | {0x3f0dU, 0x44d3U, 1U, 0U, 1U}, |
44 | // x = -0x1.e28p-3, expm1f16(x) = -0x1.adcp-3 (RZ) |
45 | {0xb38aU, 0xb2b7U, 0U, 1U, 1U}, |
46 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
47 | // x = 0x1.a08p-3, exp10m1f(x) = 0x1.cdcp-3 (RZ) |
48 | {0x3282U, 0x3337U, 1U, 0U, 0U}, |
49 | #endif |
50 | }}; |
51 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
52 | |
53 | LLVM_LIBC_FUNCTION(float16, expm1f16, (float16 x)) { |
54 | using FPBits = fputil::FPBits<float16>; |
55 | FPBits x_bits(x); |
56 | |
57 | uint16_t x_u = x_bits.uintval(); |
58 | uint16_t x_abs = x_u & 0x7fffU; |
59 | |
60 | // When |x| <= 2^(-3), or |x| >= -11 * log(2), or x is NaN. |
61 | if (LIBC_UNLIKELY(x_abs <= 0x3000U || x_abs >= 0x47a0U)) { |
62 | // expm1(NaN) = NaN |
63 | if (x_bits.is_nan()) { |
64 | if (x_bits.is_signaling_nan()) { |
65 | fputil::raise_except_if_required(FE_INVALID); |
66 | return FPBits::quiet_nan().get_val(); |
67 | } |
68 | |
69 | return x; |
70 | } |
71 | |
72 | // expm1(+/-0) = +/-0 |
73 | if (x_abs == 0) |
74 | return x; |
75 | |
76 | // When x >= 16 * log(2). |
77 | if (x_bits.is_pos() && x_abs >= 0x498cU) { |
78 | // expm1(+inf) = +inf |
79 | if (x_bits.is_inf()) |
80 | return FPBits::inf().get_val(); |
81 | |
82 | switch (fputil::quick_get_round()) { |
83 | case FE_TONEAREST: |
84 | case FE_UPWARD: |
85 | fputil::set_errno_if_required(ERANGE); |
86 | fputil::raise_except_if_required(FE_OVERFLOW | FE_INEXACT); |
87 | return FPBits::inf().get_val(); |
88 | default: |
89 | return FPBits::max_normal().get_val(); |
90 | } |
91 | } |
92 | |
93 | // When x <= -11 * log(2). |
94 | if (x_u >= 0xc7a0U) { |
95 | // expm1(-inf) = -1 |
96 | if (x_bits.is_inf()) |
97 | return FPBits::one(Sign::NEG).get_val(); |
98 | |
99 | // When x > -0x1.0ap+3, round(expm1(x), HP, RN) = -1. |
100 | if (x_u > 0xc828U) |
101 | return fputil::round_result_slightly_up( |
102 | FPBits::one(Sign::NEG).get_val()); |
103 | // When x <= -0x1.0ap+3, round(expm1(x), HP, RN) = -0x1.ffcp-1. |
104 | return fputil::round_result_slightly_down( |
105 | fputil::cast<float16>(-0x1.ffcp-1)); |
106 | } |
107 | |
108 | // When 0 < |x| <= 2^(-3). |
109 | if (x_abs <= 0x3000U && !x_bits.is_zero()) { |
110 | |
111 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
112 | if (auto r = EXPM1F16_EXCEPTS_LO.lookup(x_u); |
113 | LIBC_UNLIKELY(r.has_value())) |
114 | return r.value(); |
115 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
116 | |
117 | float xf = x; |
118 | // Degree-5 minimax polynomial generated by Sollya with the following |
119 | // commands: |
120 | // > display = hexadecimal; |
121 | // > P = fpminimax(expm1(x)/x, 4, [|SG...|], [-2^-3, 2^-3]); |
122 | // > x * P; |
123 | return fputil::cast<float16>( |
124 | xf * fputil::polyeval(xf, 0x1p+0f, 0x1.fffff8p-2f, 0x1.555556p-3f, |
125 | 0x1.55905ep-5f, 0x1.1124c2p-7f)); |
126 | } |
127 | } |
128 | |
129 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
130 | if (auto r = EXPM1F16_EXCEPTS_HI.lookup(x_u); LIBC_UNLIKELY(r.has_value())) |
131 | return r.value(); |
132 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
133 | |
134 | // exp(x) = exp(hi + mid) * exp(lo) |
135 | auto [exp_hi_mid, exp_lo] = exp_range_reduction(x); |
136 | // expm1(x) = exp(hi + mid) * exp(lo) - 1 |
137 | return fputil::cast<float16>(fputil::multiply_add(exp_hi_mid, exp_lo, -1.0f)); |
138 | } |
139 | |
140 | } // namespace LIBC_NAMESPACE_DECL |
141 | |