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/config.h" |
21 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
22 | #include "src/__support/macros/properties/cpu_features.h" |
23 | |
24 | #include "explogxf.h" |
25 | |
26 | namespace LIBC_NAMESPACE_DECL { |
27 | namespace generic { |
28 | |
29 | LIBC_INLINE float exp2f(float x) { |
30 | using FPBits = typename fputil::FPBits<float>; |
31 | FPBits xbits(x); |
32 | |
33 | uint32_t x_u = xbits.uintval(); |
34 | uint32_t x_abs = x_u & 0x7fff'ffffU; |
35 | |
36 | // When |x| >= 128, or x is nan, or |x| <= 2^-5 |
37 | if (LIBC_UNLIKELY(x_abs >= 0x4300'0000U || x_abs <= 0x3d00'0000U)) { |
38 | // |x| <= 2^-5 |
39 | if (x_abs <= 0x3d00'0000) { |
40 | // |x| < 2^-25 |
41 | if (LIBC_UNLIKELY(x_abs <= 0x3280'0000U)) { |
42 | return 1.0f + x; |
43 | } |
44 | |
45 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
46 | constexpr uint32_t EXVAL1 = 0x3b42'9d37U; |
47 | constexpr uint32_t EXVAL2 = 0xbcf3'a937U; |
48 | constexpr uint32_t EXVAL_MASK = EXVAL1 & EXVAL2; |
49 | |
50 | // Check exceptional values. |
51 | if (LIBC_UNLIKELY((x_u & EXVAL_MASK) == EXVAL_MASK)) { |
52 | if (LIBC_UNLIKELY(x_u == EXVAL1)) { // x = 0x1.853a6ep-9f |
53 | return fputil::round_result_slightly_down(0x1.00870ap+0f); |
54 | } else if (LIBC_UNLIKELY(x_u == EXVAL2)) { // x = -0x1.e7526ep-6f |
55 | return fputil::round_result_slightly_down(0x1.f58d62p-1f); |
56 | } |
57 | } |
58 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
59 | |
60 | // Minimax polynomial generated by Sollya with: |
61 | // > P = fpminimax((2^x - 1)/x, 5, [|D...|], [-2^-5, 2^-5]); |
62 | constexpr double COEFFS[] = { |
63 | 0x1.62e42fefa39f3p-1, 0x1.ebfbdff82c57bp-3, 0x1.c6b08d6f2d7aap-5, |
64 | 0x1.3b2ab6fc92f5dp-7, 0x1.5d897cfe27125p-10, 0x1.43090e61e6af1p-13}; |
65 | double xd = static_cast<double>(x); |
66 | double xsq = xd * xd; |
67 | double c0 = fputil::multiply_add(xd, COEFFS[1], COEFFS[0]); |
68 | double c1 = fputil::multiply_add(xd, COEFFS[3], COEFFS[2]); |
69 | double c2 = fputil::multiply_add(xd, COEFFS[5], COEFFS[4]); |
70 | double p = fputil::polyeval(xsq, c0, c1, c2); |
71 | double r = fputil::multiply_add(p, xd, 1.0); |
72 | return static_cast<float>(r); |
73 | } |
74 | |
75 | // x >= 128 |
76 | if (xbits.is_pos()) { |
77 | // x is finite |
78 | if (x_u < 0x7f80'0000U) { |
79 | int rounding = fputil::quick_get_round(); |
80 | if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) |
81 | return FPBits::max_normal().get_val(); |
82 | |
83 | fputil::set_errno_if_required(ERANGE); |
84 | fputil::raise_except_if_required(FE_OVERFLOW); |
85 | } |
86 | // x is +inf or nan |
87 | return x + FPBits::inf().get_val(); |
88 | } |
89 | // x <= -150 |
90 | if (x_u >= 0xc316'0000U) { |
91 | // exp(-Inf) = 0 |
92 | if (xbits.is_inf()) |
93 | return 0.0f; |
94 | // exp(nan) = nan |
95 | if (xbits.is_nan()) |
96 | return x; |
97 | if (fputil::fenv_is_round_up()) |
98 | return FPBits::min_subnormal().get_val(); |
99 | if (x != 0.0f) { |
100 | fputil::set_errno_if_required(ERANGE); |
101 | fputil::raise_except_if_required(FE_UNDERFLOW); |
102 | } |
103 | return 0.0f; |
104 | } |
105 | } |
106 | |
107 | // For -150 < x < 128, to compute 2^x, we perform the following range |
108 | // reduction: find hi, mid, lo such that: |
109 | // x = hi + mid + lo, in which |
110 | // hi is an integer, |
111 | // 0 <= mid * 2^5 < 32 is an integer |
112 | // -2^(-6) <= lo <= 2^-6. |
113 | // In particular, |
114 | // hi + mid = round(x * 2^5) * 2^(-5). |
115 | // Then, |
116 | // 2^x = 2^(hi + mid + lo) = 2^hi * 2^mid * 2^lo. |
117 | // 2^mid is stored in the lookup table of 32 elements. |
118 | // 2^lo is computed using a degree-5 minimax polynomial |
119 | // generated by Sollya. |
120 | // We perform 2^hi * 2^mid by simply add hi to the exponent field |
121 | // of 2^mid. |
122 | |
123 | // kf = (hi + mid) * 2^5 = round(x * 2^5) |
124 | float kf; |
125 | int k; |
126 | #ifdef LIBC_TARGET_CPU_HAS_NEAREST_INT |
127 | kf = fputil::nearest_integer(x * 32.0f); |
128 | k = static_cast<int>(kf); |
129 | #else |
130 | constexpr float HALF[2] = {0.5f, -0.5f}; |
131 | k = static_cast<int>(fputil::multiply_add(x, 32.0f, HALF[x < 0.0f])); |
132 | kf = static_cast<float>(k); |
133 | #endif // LIBC_TARGET_CPU_HAS_NEAREST_INT |
134 | |
135 | // dx = lo = x - (hi + mid) = x - kf * 2^(-5) |
136 | double dx = fputil::multiply_add(-0x1.0p-5f, kf, x); |
137 | |
138 | // hi = floor(kf * 2^(-4)) |
139 | // exp_hi = shift hi to the exponent field of double precision. |
140 | int64_t exp_hi = |
141 | static_cast<int64_t>(static_cast<uint64_t>(k >> ExpBase::MID_BITS) |
142 | << fputil::FPBits<double>::FRACTION_LEN); |
143 | // mh = 2^hi * 2^mid |
144 | // mh_bits = bit field of mh |
145 | int64_t mh_bits = ExpBase::EXP_2_MID[k & ExpBase::MID_MASK] + exp_hi; |
146 | double mh = fputil::FPBits<double>(uint64_t(mh_bits)).get_val(); |
147 | |
148 | // Degree-5 polynomial approximating (2^x - 1)/x generating by Sollya with: |
149 | // > P = fpminimax((2^x - 1)/x, 5, [|D...|], [-1/32. 1/32]); |
150 | constexpr double COEFFS[5] = {0x1.62e42fefa39efp-1, 0x1.ebfbdff8131c4p-3, |
151 | 0x1.c6b08d7061695p-5, 0x1.3b2b1bee74b2ap-7, |
152 | 0x1.5d88091198529p-10}; |
153 | double dx_sq = dx * dx; |
154 | double c1 = fputil::multiply_add(dx, COEFFS[0], 1.0); |
155 | double c2 = fputil::multiply_add(dx, COEFFS[2], COEFFS[1]); |
156 | double c3 = fputil::multiply_add(dx, COEFFS[4], COEFFS[3]); |
157 | double p = fputil::multiply_add(dx_sq, c3, c2); |
158 | // 2^x = 2^(hi + mid + lo) |
159 | // = 2^(hi + mid) * 2^lo |
160 | // ~ mh * (1 + lo * P(lo)) |
161 | // = mh + (mh*lo) * P(lo) |
162 | return static_cast<float>(fputil::multiply_add(p, dx_sq * mh, c1 * mh)); |
163 | } |
164 | |
165 | } // namespace generic |
166 | } // namespace LIBC_NAMESPACE_DECL |
167 | |
168 | #endif // LLVM_LIBC_SRC_MATH_GENERIC_EXP2F_IMPL_H |
169 | |