1 | //===-- Single-precision tanh 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/tanhf.h" |
10 | #include "src/__support/FPUtil/FPBits.h" |
11 | #include "src/__support/FPUtil/PolyEval.h" |
12 | #include "src/__support/FPUtil/multiply_add.h" |
13 | #include "src/__support/FPUtil/nearest_integer.h" |
14 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
15 | #include "src/__support/macros/properties/cpu_features.h" |
16 | #include "src/math/generic/explogxf.h" |
17 | |
18 | namespace LIBC_NAMESPACE { |
19 | |
20 | // 2^6 * log2(e) |
21 | constexpr double LOG2_E_EXP2_6 = ExpBase::LOG2_B * 2.0; |
22 | |
23 | LLVM_LIBC_FUNCTION(float, tanhf, (float x)) { |
24 | using FPBits = typename fputil::FPBits<float>; |
25 | FPBits xbits(x); |
26 | uint32_t x_abs = xbits.abs().uintval(); |
27 | |
28 | const int sign_index = xbits.is_neg() ? 1 : 0; |
29 | |
30 | // When |x| >= 15, or x is inf or nan, or |x| <= 0.078125 |
31 | if (LIBC_UNLIKELY((x_abs >= 0x4170'0000U) || (x_abs <= 0x3da0'0000U))) { |
32 | if (x_abs <= 0x3da0'0000U) { |
33 | // |x| <= 0.078125 |
34 | if (LIBC_UNLIKELY(x_abs <= 0x3280'0000U)) { |
35 | // |x| <= 2^-26 |
36 | return (x_abs != 0) |
37 | ? static_cast<float>(x - 0x1.5555555555555p-2 * x * x * x) |
38 | : x; |
39 | } |
40 | |
41 | const double TAYLOR[] = {-0x1.5555555555555p-2, 0x1.1111111111111p-3, |
42 | -0x1.ba1ba1ba1ba1cp-5, 0x1.664f4882c10fap-6, |
43 | -0x1.226e355e6c23dp-7}; |
44 | double xdbl = x; |
45 | double x2 = xdbl * xdbl; |
46 | // Taylor polynomial. |
47 | double x4 = x2 * x2; |
48 | double c0 = x2 * TAYLOR[0]; |
49 | double c1 = fputil::multiply_add(x: x2, y: TAYLOR[2], z: TAYLOR[1]); |
50 | double c2 = fputil::multiply_add(x: x2, y: TAYLOR[4], z: TAYLOR[3]); |
51 | double pe = fputil::polyeval(x: x4, a0: c0, a: c1, a: c2); |
52 | |
53 | return static_cast<float>(fputil::multiply_add(x: xdbl, y: pe, z: xdbl)); |
54 | } |
55 | |
56 | // |x| >= 15 |
57 | if (LIBC_UNLIKELY(xbits.is_nan())) |
58 | return x + 1.0f; // sNaN to qNaN + signal |
59 | |
60 | constexpr float SIGNS[2][2] = {{1.0f, -0x1.0p-25f}, {-1.0f, 0x1.0p-25f}}; |
61 | |
62 | if (LIBC_UNLIKELY(xbits.is_inf())) |
63 | return SIGNS[sign_index][0]; |
64 | |
65 | return SIGNS[sign_index][0] + SIGNS[sign_index][1]; |
66 | } |
67 | |
68 | // Range reduction: e^(2x) = 2^(hi + mid) * e^lo |
69 | // Let k = round( x * 2^6 * log2(e)), |
70 | // So k = (hi + mid) * 2^5 |
71 | // Then lo = 2x - (hi + mid) * log(2) = 2x - k * 2^-5 * log(2). |
72 | |
73 | double xd = static_cast<double>(x); |
74 | // k = round( x* 2^6 * log2(e) ) |
75 | double k; |
76 | // mk = -k |
77 | int mk; |
78 | #ifdef LIBC_TARGET_CPU_HAS_NEAREST_INT |
79 | k = fputil::nearest_integer(x: xd * LOG2_E_EXP2_6); |
80 | mk = -static_cast<int>(k); |
81 | #else |
82 | constexpr double HALF_WAY[2] = {-0.5, 0.5}; |
83 | |
84 | mk = static_cast<int>( |
85 | fputil::multiply_add(xd, -LOG2_E_EXP2_6, HALF_WAY[sign_index])); |
86 | k = static_cast<double>(-mk); |
87 | #endif // LIBC_TARGET_CPU_HAS_NEAREST_INT |
88 | // -hi = floor(-k * 2^(-MID_BITS)) |
89 | // exp_mhi = shift -hi to the exponent field of double precision. |
90 | int64_t exp_mhi = static_cast<int64_t>(mk >> ExpBase::MID_BITS) |
91 | << fputil::FPBits<double>::FRACTION_LEN; |
92 | // mh = 2^(-hi - mid) |
93 | int64_t mh_bits = ExpBase::EXP_2_MID[mk & ExpBase::MID_MASK] + exp_mhi; |
94 | double mh = fputil::FPBits<double>(uint64_t(mh_bits)).get_val(); |
95 | // dx = lo/2 = x - (hi + mid) * log(2)/2 = x - k * 2^-6 * log(2) |
96 | double dx = fputil::multiply_add( |
97 | x: k, y: ExpBase::M_LOGB_2_LO * 0.5, |
98 | z: fputil::multiply_add(x: k, y: ExpBase::M_LOGB_2_HI * 0.5, z: xd)); |
99 | |
100 | // > P = fpminimax(expm1(2*x)/x, 4, [|D...|], [-log(2)/128, log(2)/128]); |
101 | constexpr double COEFFS[] = {0x1.ffffffffe5bc8p0, 0x1.555555555cd67p0, |
102 | 0x1.5555c2a9b48b4p-1, 0x1.11112a0e34bdbp-2}; |
103 | |
104 | double dx2 = dx * dx; |
105 | double c0 = fputil::multiply_add(x: dx, y: 2.0, z: 1.0); |
106 | double c1 = fputil::multiply_add(x: dx, y: COEFFS[1], z: COEFFS[0]); |
107 | double c2 = fputil::multiply_add(x: dx, y: COEFFS[3], z: COEFFS[2]); |
108 | double r = fputil::polyeval(x: dx2, a0: c0, a: c1, a: c2); |
109 | |
110 | // tanh(x) = sinh(x) / cosh(x) |
111 | // = (e^x - e^(-x)) / (e^x + e^(-x)) |
112 | // = (e^(2x) - 1) / (e^(2x) + 1) |
113 | // = (2^(hi + mid) * e^lo - 1) / (2^(hi + mid) * e^lo + 1) |
114 | // = (e^lo - 2^(-hi - mid)) / (e^lo + 2^(-hi - mid)) |
115 | // = (r - mh) / (r + mh) |
116 | return static_cast<float>((r - mh) / (r + mh)); |
117 | } |
118 | |
119 | } // namespace LIBC_NAMESPACE |
120 | |