xref: /aosp_15_r20/external/llvm-libc/src/math/generic/tanhf.cpp (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
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/config.h"
15 #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY
16 #include "src/__support/macros/properties/cpu_features.h"
17 #include "src/math/generic/explogxf.h"
18 
19 namespace LIBC_NAMESPACE_DECL {
20 
21 // 2^6 * log2(e)
22 constexpr double LOG2_E_EXP2_6 = ExpBase::LOG2_B * 2.0;
23 
24 LLVM_LIBC_FUNCTION(float, tanhf, (float x)) {
25   using FPBits = typename fputil::FPBits<float>;
26   FPBits xbits(x);
27   uint32_t x_abs = xbits.abs().uintval();
28 
29   const int sign_index = xbits.is_neg() ? 1 : 0;
30 
31   // When |x| >= 15, or x is inf or nan, or |x| <= 0.078125
32   if (LIBC_UNLIKELY((x_abs >= 0x4170'0000U) || (x_abs <= 0x3da0'0000U))) {
33     if (x_abs <= 0x3da0'0000U) {
34       // |x| <= 0.078125
35       if (LIBC_UNLIKELY(x_abs <= 0x3280'0000U)) {
36         // |x| <= 2^-26
37         return (x_abs != 0)
38                    ? static_cast<float>(x - 0x1.5555555555555p-2 * x * x * x)
39                    : x;
40       }
41 
42       const double TAYLOR[] = {-0x1.5555555555555p-2, 0x1.1111111111111p-3,
43                                -0x1.ba1ba1ba1ba1cp-5, 0x1.664f4882c10fap-6,
44                                -0x1.226e355e6c23dp-7};
45       double xdbl = x;
46       double x2 = xdbl * xdbl;
47       // Taylor polynomial.
48       double x4 = x2 * x2;
49       double c0 = x2 * TAYLOR[0];
50       double c1 = fputil::multiply_add(x2, TAYLOR[2], TAYLOR[1]);
51       double c2 = fputil::multiply_add(x2, TAYLOR[4], TAYLOR[3]);
52       double pe = fputil::polyeval(x4, c0, c1, c2);
53 
54       return static_cast<float>(fputil::multiply_add(xdbl, pe, xdbl));
55     }
56 
57     // |x| >= 15
58     if (LIBC_UNLIKELY(xbits.is_nan()))
59       return x + 1.0f; // sNaN to qNaN + signal
60 
61     constexpr float SIGNS[2][2] = {{1.0f, -0x1.0p-25f}, {-1.0f, 0x1.0p-25f}};
62 
63     if (LIBC_UNLIKELY(xbits.is_inf()))
64       return SIGNS[sign_index][0];
65 
66     return SIGNS[sign_index][0] + SIGNS[sign_index][1];
67   }
68 
69   // Range reduction: e^(2x) = 2^(hi + mid) * e^lo
70   // Let  k = round( x * 2^6 * log2(e)),
71   // So   k  = (hi + mid) * 2^5
72   // Then lo = 2x - (hi + mid) * log(2) = 2x - k * 2^-5 * log(2).
73 
74   double xd = static_cast<double>(x);
75   // k = round( x* 2^6 * log2(e) )
76   double k;
77   // mk = -k
78   int mk;
79 #ifdef LIBC_TARGET_CPU_HAS_NEAREST_INT
80   k = fputil::nearest_integer(xd * LOG2_E_EXP2_6);
81   mk = -static_cast<int>(k);
82 #else
83   constexpr double HALF_WAY[2] = {-0.5, 0.5};
84 
85   mk = static_cast<int>(
86       fputil::multiply_add(xd, -LOG2_E_EXP2_6, HALF_WAY[sign_index]));
87   k = static_cast<double>(-mk);
88 #endif // LIBC_TARGET_CPU_HAS_NEAREST_INT
89   // -hi = floor(-k * 2^(-MID_BITS))
90   // exp_mhi = shift -hi to the exponent field of double precision.
91   int64_t exp_mhi = static_cast<int64_t>(mk >> ExpBase::MID_BITS)
92                     << fputil::FPBits<double>::FRACTION_LEN;
93   // mh = 2^(-hi - mid)
94   int64_t mh_bits = ExpBase::EXP_2_MID[mk & ExpBase::MID_MASK] + exp_mhi;
95   double mh = fputil::FPBits<double>(uint64_t(mh_bits)).get_val();
96   // dx = lo/2 = x - (hi + mid) * log(2)/2 = x - k * 2^-6 * log(2)
97   double dx = fputil::multiply_add(
98       k, ExpBase::M_LOGB_2_LO * 0.5,
99       fputil::multiply_add(k, ExpBase::M_LOGB_2_HI * 0.5, xd));
100 
101   // > P = fpminimax(expm1(2*x)/x, 4, [|D...|], [-log(2)/128, log(2)/128]);
102   constexpr double COEFFS[] = {0x1.ffffffffe5bc8p0, 0x1.555555555cd67p0,
103                                0x1.5555c2a9b48b4p-1, 0x1.11112a0e34bdbp-2};
104 
105   double dx2 = dx * dx;
106   double c0 = fputil::multiply_add(dx, 2.0, 1.0);
107   double c1 = fputil::multiply_add(dx, COEFFS[1], COEFFS[0]);
108   double c2 = fputil::multiply_add(dx, COEFFS[3], COEFFS[2]);
109   double r = fputil::polyeval(dx2, c0, c1, c2);
110 
111   // tanh(x) = sinh(x) / cosh(x)
112   //         = (e^x - e^(-x)) / (e^x + e^(-x))
113   //         = (e^(2x) - 1) / (e^(2x) + 1)
114   //         = (2^(hi + mid) * e^lo - 1) / (2^(hi + mid) * e^lo + 1)
115   //         = (e^lo - 2^(-hi - mid)) / (e^lo + 2^(-hi - mid))
116   //         = (r - mh) / (r + mh)
117   return static_cast<float>((r - mh) / (r + mh));
118 }
119 
120 } // namespace LIBC_NAMESPACE_DECL
121