1 | //===-- Half-precision tanh(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/tanhf16.h" |
10 | #include "expxf16.h" |
11 | #include "hdr/fenv_macros.h" |
12 | #include "src/__support/CPP/array.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/nearest_integer.h" |
20 | #include "src/__support/FPUtil/rounding_mode.h" |
21 | #include "src/__support/common.h" |
22 | #include "src/__support/macros/config.h" |
23 | #include "src/__support/macros/optimization.h" |
24 | |
25 | namespace LIBC_NAMESPACE_DECL { |
26 | |
27 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
28 | static constexpr fputil::ExceptValues<float16, 2> TANHF16_EXCEPTS = {{ |
29 | // x = 0x1.f54p+0, tanhf16(x) = 0x1.ecp-1 (RZ) |
30 | {0x3fd5U, 0x3bb0U, 1U, 0U, 0U}, |
31 | // x = -0x1.f54p+0, tanhf16(x) = -0x1.ecp-1 (RZ) |
32 | {0xbfd5U, 0xbbb0U, 0U, 1U, 0U}, |
33 | }}; |
34 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
35 | |
36 | LLVM_LIBC_FUNCTION(float16, tanhf16, (float16 x)) { |
37 | using FPBits = fputil::FPBits<float16>; |
38 | FPBits x_bits(x); |
39 | |
40 | uint16_t x_u = x_bits.uintval(); |
41 | uint16_t x_abs = x_u & 0x7fffU; |
42 | |
43 | // When -2^(-14) <= x <= -2^(-9), or |x| <= 0x1.d2p-4, |
44 | // or |x| >= atanh(1 - 2^(-11)), or x is NaN. |
45 | if (LIBC_UNLIKELY(x_abs <= 0x2f48U || x_abs >= 0x4429U)) { |
46 | // tanh(NaN) = NaN |
47 | if (x_bits.is_nan()) { |
48 | if (x_bits.is_signaling_nan()) { |
49 | fputil::raise_except_if_required(FE_INVALID); |
50 | return FPBits::quiet_nan().get_val(); |
51 | } |
52 | |
53 | return x; |
54 | } |
55 | |
56 | // When -2^(-14) <= x <= -2^(-9). |
57 | if (x_u >= 0x8400U && x_u <= 0x9800U) { |
58 | switch (fputil::quick_get_round()) { |
59 | case FE_TONEAREST: |
60 | case FE_DOWNWARD: |
61 | return x; |
62 | default: |
63 | return FPBits(static_cast<uint16_t>(x_u - 1U)).get_val(); |
64 | } |
65 | } |
66 | |
67 | // When |x| <= 0x1.d2p-4. |
68 | if (x_abs <= 0x2f48U) { |
69 | if (LIBC_UNLIKELY(x_abs == 0)) |
70 | return x; |
71 | |
72 | float xf = x; |
73 | float xf_sq = xf * xf; |
74 | // Degree-7 Taylor expansion generated by Sollya with the following |
75 | // commands: |
76 | // > taylor(tanh(x), 7, 0); |
77 | // > display = hexadecimal; |
78 | // > // For each coefficient: |
79 | // > round(/* put coefficient here */, SG, RN); |
80 | return fputil::cast<float16>( |
81 | xf * fputil::polyeval(xf_sq, 0x1p+0f, -0x1.555556p-2f, 0x1.111112p-3f, |
82 | -0x1.ba1ba2p-5f)); |
83 | } |
84 | |
85 | // tanh(+/-inf) = +/-1 |
86 | if (x_bits.is_inf()) |
87 | return FPBits::one(x_bits.sign()).get_val(); |
88 | |
89 | // When |x| >= atanh(1 - 2^(-11)). |
90 | fputil::raise_except_if_required(FE_INEXACT); |
91 | |
92 | int rounding_mode = fputil::quick_get_round(); |
93 | if ((rounding_mode == FE_TONEAREST && x_abs >= 0x4482U) || |
94 | (rounding_mode == FE_UPWARD && x_bits.is_pos()) || |
95 | (rounding_mode == FE_DOWNWARD && x_bits.is_neg())) { |
96 | return FPBits::one(x_bits.sign()).get_val(); |
97 | } |
98 | if (x_bits.is_pos()) |
99 | return fputil::cast<float16>(0x1.ffcp-1); |
100 | return fputil::cast<float16>(-0x1.ffcp-1); |
101 | } |
102 | |
103 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
104 | if (auto r = TANHF16_EXCEPTS.lookup(x_u); LIBC_UNLIKELY(r.has_value())) |
105 | return r.value(); |
106 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
107 | |
108 | // For atanh(-1 + 2^(-11)) < x < atanh(1 - 2^(-11)), to compute tanh(x), we |
109 | // perform the following range reduction: find hi, mid, lo, such that: |
110 | // x = (hi + mid) * log(2) * 0.5 + lo, in which |
111 | // hi is an integer, |
112 | // mid * 2^5 is an integer, |
113 | // -2^(-5) <= lo < 2^(-5). |
114 | // In particular, |
115 | // hi + mid = round(x * log2(e) * 2 * 2^5) * 2^(-5). |
116 | // Then, |
117 | // tanh(x) = sinh(x)/cosh(x) |
118 | // = (e^x - e^(-x)) / (e^x + e^(-x)) |
119 | // = (e^(2x) - 1) / (e^(2x) + 1) |
120 | // = (2^(hi + mid) * e^(2*lo) - 1) / (2^(hi + mid) * e^(2*lo) + 1) |
121 | // = (e^(2*lo) - 2^(-hi - mid)) / (e^(2*lo) + 2^(-hi - mid)) |
122 | // We store 2^(-mid) in the lookup table EXP2_MID_5_BITS, and compute |
123 | // 2^(-hi - mid) by adding -hi to the exponent field of 2^(-mid). |
124 | // e^lo is computed using a degree-3 minimax polynomial generated by Sollya. |
125 | |
126 | float xf = x; |
127 | float kf = fputil::nearest_integer(xf * (LOG2F_E * 2.0f * 0x1.0p+5f)); |
128 | int x_hi_mid = -static_cast<int>(kf); |
129 | unsigned x_hi = static_cast<unsigned>(x_hi_mid) >> 5; |
130 | unsigned x_mid = static_cast<unsigned>(x_hi_mid) & 0x1f; |
131 | // lo = x - (hi + mid) |
132 | // = round(x * log2(e) * 2 * 2^5) * log(2) * 0.5 * (-2^(-5)) + x |
133 | float lo = fputil::multiply_add(kf, LOGF_2 * 0.5f * -0x1.0p-5f, xf); |
134 | |
135 | uint32_t exp2_hi_mid_bits = |
136 | EXP2_MID_5_BITS[x_mid] + |
137 | static_cast<uint32_t>(x_hi << fputil::FPBits<float>::FRACTION_LEN); |
138 | // exp2_hi_mid = 2^(-hi - mid) |
139 | float exp2_hi_mid = fputil::FPBits<float>(exp2_hi_mid_bits).get_val(); |
140 | // Degree-3 minimax polynomial generated by Sollya with the following |
141 | // commands: |
142 | // > display = hexadecimal; |
143 | // > P = fpminimax(expm1(2*x)/x, 2, [|SG...|], [-2^-5, 2^-5]); |
144 | // > 1 + x * P; |
145 | float exp_2lo = |
146 | fputil::polyeval(lo, 0x1p+0f, 0x1p+1f, 0x1.001p+1f, 0x1.555ddep+0f); |
147 | return fputil::cast<float16>((exp_2lo - exp2_hi_mid) / |
148 | (exp_2lo + exp2_hi_mid)); |
149 | } |
150 | |
151 | } // namespace LIBC_NAMESPACE_DECL |
152 | |