1 | //===-- Single-precision 2^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 | #ifndef LLVM_LIBC_SRC_MATH_GENERIC_EXP2F_IMPL_H |
10 | #define LLVM_LIBC_SRC_MATH_GENERIC_EXP2F_IMPL_H |
11 | |
12 | #include "src/__support/FPUtil/FEnvImpl.h" |
13 | #include "src/__support/FPUtil/FPBits.h" |
14 | #include "src/__support/FPUtil/PolyEval.h" |
15 | #include "src/__support/FPUtil/except_value_utils.h" |
16 | #include "src/__support/FPUtil/multiply_add.h" |
17 | #include "src/__support/FPUtil/nearest_integer.h" |
18 | #include "src/__support/FPUtil/rounding_mode.h" |
19 | #include "src/__support/common.h" |
20 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
21 | #include "src/__support/macros/properties/cpu_features.h" |
22 | |
23 | #include <errno.h> |
24 | |
25 | #include "explogxf.h" |
26 | |
27 | namespace LIBC_NAMESPACE::generic { |
28 | |
29 | LIBC_INLINE float exp2f(float x) { |
30 | constexpr uint32_t EXVAL1 = 0x3b42'9d37U; |
31 | constexpr uint32_t EXVAL2 = 0xbcf3'a937U; |
32 | constexpr uint32_t EXVAL_MASK = EXVAL1 & EXVAL2; |
33 | |
34 | using FPBits = typename fputil::FPBits<float>; |
35 | FPBits xbits(x); |
36 | |
37 | uint32_t x_u = xbits.uintval(); |
38 | uint32_t x_abs = x_u & 0x7fff'ffffU; |
39 | |
40 | // When |x| >= 128, or x is nan, or |x| <= 2^-5 |
41 | if (LIBC_UNLIKELY(x_abs >= 0x4300'0000U || x_abs <= 0x3d00'0000U)) { |
42 | // |x| <= 2^-5 |
43 | if (x_abs <= 0x3d00'0000) { |
44 | // |x| < 2^-25 |
45 | if (LIBC_UNLIKELY(x_abs <= 0x3280'0000U)) { |
46 | return 1.0f + x; |
47 | } |
48 | |
49 | // Check exceptional values. |
50 | if (LIBC_UNLIKELY((x_u & EXVAL_MASK) == EXVAL_MASK)) { |
51 | if (LIBC_UNLIKELY(x_u == EXVAL1)) { // x = 0x1.853a6ep-9f |
52 | return fputil::round_result_slightly_down(value_rn: 0x1.00870ap+0f); |
53 | } else if (LIBC_UNLIKELY(x_u == EXVAL2)) { // x = -0x1.e7526ep-6f |
54 | return fputil::round_result_slightly_down(value_rn: 0x1.f58d62p-1f); |
55 | } |
56 | } |
57 | |
58 | // Minimax polynomial generated by Sollya with: |
59 | // > P = fpminimax((2^x - 1)/x, 5, [|D...|], [-2^-5, 2^-5]); |
60 | constexpr double COEFFS[] = { |
61 | 0x1.62e42fefa39f3p-1, 0x1.ebfbdff82c57bp-3, 0x1.c6b08d6f2d7aap-5, |
62 | 0x1.3b2ab6fc92f5dp-7, 0x1.5d897cfe27125p-10, 0x1.43090e61e6af1p-13}; |
63 | double xd = static_cast<double>(x); |
64 | double xsq = xd * xd; |
65 | double c0 = fputil::multiply_add(x: xd, y: COEFFS[1], z: COEFFS[0]); |
66 | double c1 = fputil::multiply_add(x: xd, y: COEFFS[3], z: COEFFS[2]); |
67 | double c2 = fputil::multiply_add(x: xd, y: COEFFS[5], z: COEFFS[4]); |
68 | double p = fputil::polyeval(x: xsq, a0: c0, a: c1, a: c2); |
69 | double r = fputil::multiply_add(x: p, y: xd, z: 1.0); |
70 | return static_cast<float>(r); |
71 | } |
72 | |
73 | // x >= 128 |
74 | if (xbits.is_pos()) { |
75 | // x is finite |
76 | if (x_u < 0x7f80'0000U) { |
77 | int rounding = fputil::quick_get_round(); |
78 | if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) |
79 | return FPBits::max_normal().get_val(); |
80 | |
81 | fputil::set_errno_if_required(ERANGE); |
82 | fputil::raise_except_if_required(FE_OVERFLOW); |
83 | } |
84 | // x is +inf or nan |
85 | return x + FPBits::inf().get_val(); |
86 | } |
87 | // x <= -150 |
88 | if (x_u >= 0xc316'0000U) { |
89 | // exp(-Inf) = 0 |
90 | if (xbits.is_inf()) |
91 | return 0.0f; |
92 | // exp(nan) = nan |
93 | if (xbits.is_nan()) |
94 | return x; |
95 | if (fputil::fenv_is_round_up()) |
96 | return FPBits::min_subnormal().get_val(); |
97 | if (x != 0.0f) { |
98 | fputil::set_errno_if_required(ERANGE); |
99 | fputil::raise_except_if_required(FE_UNDERFLOW); |
100 | } |
101 | return 0.0f; |
102 | } |
103 | } |
104 | |
105 | // For -150 < x < 128, to compute 2^x, we perform the following range |
106 | // reduction: find hi, mid, lo such that: |
107 | // x = hi + mid + lo, in which |
108 | // hi is an integer, |
109 | // 0 <= mid * 2^5 < 32 is an integer |
110 | // -2^(-6) <= lo <= 2^-6. |
111 | // In particular, |
112 | // hi + mid = round(x * 2^5) * 2^(-5). |
113 | // Then, |
114 | // 2^x = 2^(hi + mid + lo) = 2^hi * 2^mid * 2^lo. |
115 | // 2^mid is stored in the lookup table of 32 elements. |
116 | // 2^lo is computed using a degree-5 minimax polynomial |
117 | // generated by Sollya. |
118 | // We perform 2^hi * 2^mid by simply add hi to the exponent field |
119 | // of 2^mid. |
120 | |
121 | // kf = (hi + mid) * 2^5 = round(x * 2^5) |
122 | float kf; |
123 | int k; |
124 | #ifdef LIBC_TARGET_CPU_HAS_NEAREST_INT |
125 | kf = fputil::nearest_integer(x: x * 32.0f); |
126 | k = static_cast<int>(kf); |
127 | #else |
128 | constexpr float HALF[2] = {0.5f, -0.5f}; |
129 | k = static_cast<int>(fputil::multiply_add(x, 32.0f, HALF[x < 0.0f])); |
130 | kf = static_cast<float>(k); |
131 | #endif // LIBC_TARGET_CPU_HAS_NEAREST_INT |
132 | |
133 | // dx = lo = x - (hi + mid) = x - kf * 2^(-5) |
134 | double dx = fputil::multiply_add(x: -0x1.0p-5f, y: kf, z: x); |
135 | |
136 | // hi = floor(kf * 2^(-4)) |
137 | // exp_hi = shift hi to the exponent field of double precision. |
138 | int64_t exp_hi = |
139 | static_cast<int64_t>(static_cast<uint64_t>(k >> ExpBase::MID_BITS) |
140 | << fputil::FPBits<double>::FRACTION_LEN); |
141 | // mh = 2^hi * 2^mid |
142 | // mh_bits = bit field of mh |
143 | int64_t mh_bits = ExpBase::EXP_2_MID[k & ExpBase::MID_MASK] + exp_hi; |
144 | double mh = fputil::FPBits<double>(uint64_t(mh_bits)).get_val(); |
145 | |
146 | // Degree-5 polynomial approximating (2^x - 1)/x generating by Sollya with: |
147 | // > P = fpminimax((2^x - 1)/x, 5, [|D...|], [-1/32. 1/32]); |
148 | constexpr double COEFFS[5] = {0x1.62e42fefa39efp-1, 0x1.ebfbdff8131c4p-3, |
149 | 0x1.c6b08d7061695p-5, 0x1.3b2b1bee74b2ap-7, |
150 | 0x1.5d88091198529p-10}; |
151 | double dx_sq = dx * dx; |
152 | double c1 = fputil::multiply_add(x: dx, y: COEFFS[0], z: 1.0); |
153 | double c2 = fputil::multiply_add(x: dx, y: COEFFS[2], z: COEFFS[1]); |
154 | double c3 = fputil::multiply_add(x: dx, y: COEFFS[4], z: COEFFS[3]); |
155 | double p = fputil::multiply_add(x: dx_sq, y: c3, z: c2); |
156 | // 2^x = 2^(hi + mid + lo) |
157 | // = 2^(hi + mid) * 2^lo |
158 | // ~ mh * (1 + lo * P(lo)) |
159 | // = mh + (mh*lo) * P(lo) |
160 | return static_cast<float>(fputil::multiply_add(x: p, y: dx_sq * mh, z: c1 * mh)); |
161 | } |
162 | |
163 | } // namespace LIBC_NAMESPACE::generic |
164 | |
165 | #endif // LLVM_LIBC_SRC_MATH_GENERIC_EXP2F_IMPL_H |
166 | |