1//===----------------------------------------------------------------------===//
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/// \file
10/// Double-precision implementation of exp2f.
11///
12//===----------------------------------------------------------------------===//
13
14#ifndef LLVM_LIBC_SRC___SUPPORT_MATH_EXP2F_DOUBLE_EVAL_H
15#define LLVM_LIBC_SRC___SUPPORT_MATH_EXP2F_DOUBLE_EVAL_H
16
17#include "src/__support/FPUtil/FPBits.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/polyeval.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#include "src/__support/math/exp10f_utils.h"
27
28namespace LIBC_NAMESPACE_DECL {
29namespace math {
30namespace double_eval {
31
32LIBC_INLINE float exp2f(float x) {
33 using FPBits = fputil::FPBits<float>;
34 FPBits xbits(x);
35
36 uint32_t x_u = xbits.uintval();
37 uint32_t x_abs = x_u & 0x7fff'ffffU;
38
39 // |x| < 2^-25
40 if (LIBC_UNLIKELY(x_abs <= 0x3280'0000U))
41 return 1.0f + x;
42
43 auto exp2f_mid = [&]() -> float {
44 // kf = (hi + mid) * 2^5 = round(x * 2^5)
45 float kf = 0;
46 int k = 0;
47#ifdef LIBC_TARGET_CPU_HAS_NEAREST_INT
48 kf = fputil::nearest_integer(x * 32.0f);
49 k = static_cast<int>(kf);
50#else // !LIBC_TARGET_CPU_HAS_NEAREST_INT
51 constexpr float HALF[2] = {0.5f, -0.5f};
52 k = static_cast<int>(fputil::multiply_add(x, y: 32.0f, z: HALF[x < 0.0f]));
53 kf = static_cast<float>(k);
54#endif // LIBC_TARGET_CPU_HAS_NEAREST_INT
55
56 // dx = lo = x - (hi + mid) = x - kf * 2^(-5)
57 double dx = fputil::sub_multiply(x, y: 0x1.0p-5f, z: kf);
58
59 // hi = floor(kf * 2^(-4))
60 // exp_hi = shift hi to the exponent field of double precision.
61 int64_t exp_hi =
62 static_cast<int64_t>(static_cast<uint64_t>(k >> ExpBase::MID_BITS)
63 << fputil::FPBits<double>::FRACTION_LEN);
64 // mh = 2^hi * 2^mid
65 // mh_bits = bit field of mh
66 int64_t mh_bits = ExpBase::EXP_2_MID[k & ExpBase::MID_MASK] + exp_hi;
67 double mh = fputil::FPBits<double>(uint64_t(mh_bits)).get_val();
68
69 // Degree-5 polynomial approximating (2^x - 1)/x generated by Sollya.
70 constexpr double COEFFS[5] = {0x1.62e42fefa39efp-1, 0x1.ebfbdff8131c4p-3,
71 0x1.c6b08d7061695p-5, 0x1.3b2b1bee74b2ap-7,
72 0x1.5d88091198529p-10};
73 double dx_sq = dx * dx;
74 double c1 = fputil::multiply_add(x: dx, y: COEFFS[0], z: 1.0);
75 double c2 = fputil::multiply_add(x: dx, y: COEFFS[2], z: COEFFS[1]);
76 double c3 = fputil::multiply_add(x: dx, y: COEFFS[4], z: COEFFS[3]);
77 double p = fputil::multiply_add(x: dx_sq, y: c3, z: c2);
78 // 2^x = 2^(hi + mid + lo)
79 // = 2^(hi + mid) * 2^lo
80 // ~ mh * (1 + lo * P(lo))
81 // = mh + (mh*lo) * P(lo)
82 return static_cast<float>(fputil::multiply_add(x: p, y: dx_sq * mh, z: c1 * mh));
83 };
84
85 if (LIBC_LIKELY(x_abs > 0x3d00'0000U && x_abs < 0x4300'0000U))
86 return exp2f_mid();
87
88 // When |x| >= 128, or x is nan, or |x| <= 2^-5
89 if (LIBC_UNLIKELY(x_abs >= 0x4300'0000U || x_abs <= 0x3d00'0000U)) {
90 // |x| <= 2^-5
91 if (x_abs <= 0x3d00'0000) {
92#ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS
93 constexpr uint32_t EXVAL1 = 0x3b42'9d37U;
94 constexpr uint32_t EXVAL2 = 0xbcf3'a937U;
95 constexpr uint32_t EXVAL_MASK = EXVAL1 & EXVAL2;
96
97 // Check exceptional values.
98 if (LIBC_UNLIKELY((x_u & EXVAL_MASK) == EXVAL_MASK)) {
99 if (LIBC_UNLIKELY(x_u == EXVAL1)) { // x = 0x1.853a6ep-9f
100 return fputil::round_result_slightly_down(value_rn: 0x1.00870ap+0f);
101 } else if (LIBC_UNLIKELY(x_u == EXVAL2)) { // x = -0x1.e7526ep-6f
102 return fputil::round_result_slightly_down(value_rn: 0x1.f58d62p-1f);
103 }
104 }
105#endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS
106
107 // Minimax polynomial generated by Sollya with:
108 // > P = fpminimax((2^x - 1)/x, 5, [|D...|], [-2^-5, 2^-5]);
109 constexpr double COEFFS[] = {
110 0x1.62e42fefa39f3p-1, 0x1.ebfbdff82c57bp-3, 0x1.c6b08d6f2d7aap-5,
111 0x1.3b2ab6fc92f5dp-7, 0x1.5d897cfe27125p-10, 0x1.43090e61e6af1p-13};
112 double xd = static_cast<double>(x);
113 double xsq = xd * xd;
114 double c0 = fputil::multiply_add(x: xd, y: COEFFS[1], z: COEFFS[0]);
115 double c1 = fputil::multiply_add(x: xd, y: COEFFS[3], z: COEFFS[2]);
116 double c2 = fputil::multiply_add(x: xd, y: COEFFS[5], z: COEFFS[4]);
117 double p = fputil::polyeval(x: xsq, a0: c0, a: c1, a: c2);
118 double r = fputil::multiply_add(x: p, y: xd, z: 1.0);
119 return static_cast<float>(r);
120 }
121
122 // x >= 128
123 if (xbits.is_pos()) {
124 // x is finite
125 if (x_u < 0x7f80'0000U)
126 return fputil::set_overflow<float>();
127 // x is +inf or nan
128 return x + FPBits::inf().get_val();
129 }
130 // x <= -150
131 if (x_u >= 0xc316'0000U) {
132 // exp(-Inf) = 0
133 if (xbits.is_inf())
134 return 0.0f;
135 // exp(nan) = nan
136 if (xbits.is_nan())
137 return x;
138 return fputil::set_underflow<float>();
139 }
140 }
141
142 // For -150 < x < 128, to compute 2^x, we perform the following range
143 // reduction: find hi, mid, lo such that:
144 // x = hi + mid + lo, in which
145 // hi is an integer,
146 // 0 <= mid * 2^5 < 32 is an integer
147 // -2^(-6) <= lo <= 2^-6.
148 // In particular,
149 // hi + mid = round(x * 2^5) * 2^(-5).
150 // Then,
151 // 2^x = 2^(hi + mid + lo) = 2^hi * 2^mid * 2^lo.
152 // 2^mid is stored in the lookup table of 32 elements.
153 // 2^lo is computed using a degree-5 minimax polynomial
154 // generated by Sollya.
155 // We perform 2^hi * 2^mid by simply add hi to the exponent field
156 // of 2^mid.
157 return exp2f_mid();
158}
159
160} // namespace double_eval
161} // namespace math
162} // namespace LIBC_NAMESPACE_DECL
163
164#endif // LLVM_LIBC_SRC___SUPPORT_MATH_EXP2F_DOUBLE_EVAL_H
165