1 | //===-- Single-precision e^x - 1 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/expm1f.h" |
10 | #include "common_constants.h" // Lookup tables EXP_M1 and EXP_M2. |
11 | #include "src/__support/FPUtil/BasicOperations.h" |
12 | #include "src/__support/FPUtil/FEnvImpl.h" |
13 | #include "src/__support/FPUtil/FMA.h" |
14 | #include "src/__support/FPUtil/FPBits.h" |
15 | #include "src/__support/FPUtil/PolyEval.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" // LIBC_TARGET_CPU_HAS_FMA |
22 | |
23 | #include <errno.h> |
24 | |
25 | namespace LIBC_NAMESPACE { |
26 | |
27 | LLVM_LIBC_FUNCTION(float, expm1f, (float x)) { |
28 | using FPBits = typename fputil::FPBits<float>; |
29 | FPBits xbits(x); |
30 | |
31 | uint32_t x_u = xbits.uintval(); |
32 | uint32_t x_abs = x_u & 0x7fff'ffffU; |
33 | |
34 | // Exceptional value |
35 | if (LIBC_UNLIKELY(x_u == 0x3e35'bec5U)) { // x = 0x1.6b7d8ap-3f |
36 | int round_mode = fputil::quick_get_round(); |
37 | if (round_mode == FE_TONEAREST || round_mode == FE_UPWARD) |
38 | return 0x1.8dbe64p-3f; |
39 | return 0x1.8dbe62p-3f; |
40 | } |
41 | |
42 | #if !defined(LIBC_TARGET_CPU_HAS_FMA) |
43 | if (LIBC_UNLIKELY(x_u == 0xbdc1'c6cbU)) { // x = -0x1.838d96p-4f |
44 | int round_mode = fputil::quick_get_round(); |
45 | if (round_mode == FE_TONEAREST || round_mode == FE_DOWNWARD) |
46 | return -0x1.71c884p-4f; |
47 | return -0x1.71c882p-4f; |
48 | } |
49 | #endif // LIBC_TARGET_CPU_HAS_FMA |
50 | |
51 | // When |x| > 25*log(2), or nan |
52 | if (LIBC_UNLIKELY(x_abs >= 0x418a'a123U)) { |
53 | // x < log(2^-25) |
54 | if (xbits.is_neg()) { |
55 | // exp(-Inf) = 0 |
56 | if (xbits.is_inf()) |
57 | return -1.0f; |
58 | // exp(nan) = nan |
59 | if (xbits.is_nan()) |
60 | return x; |
61 | int round_mode = fputil::quick_get_round(); |
62 | if (round_mode == FE_UPWARD || round_mode == FE_TOWARDZERO) |
63 | return -0x1.ffff'fep-1f; // -1.0f + 0x1.0p-24f |
64 | return -1.0f; |
65 | } else { |
66 | // x >= 89 or nan |
67 | if (xbits.uintval() >= 0x42b2'0000) { |
68 | if (xbits.uintval() < 0x7f80'0000U) { |
69 | int rounding = fputil::quick_get_round(); |
70 | if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) |
71 | return FPBits::max_normal().get_val(); |
72 | |
73 | fputil::set_errno_if_required(ERANGE); |
74 | fputil::raise_except_if_required(FE_OVERFLOW); |
75 | } |
76 | return x + FPBits::inf().get_val(); |
77 | } |
78 | } |
79 | } |
80 | |
81 | // |x| < 2^-4 |
82 | if (x_abs < 0x3d80'0000U) { |
83 | // |x| < 2^-25 |
84 | if (x_abs < 0x3300'0000U) { |
85 | // x = -0.0f |
86 | if (LIBC_UNLIKELY(xbits.uintval() == 0x8000'0000U)) |
87 | return x; |
88 | // When |x| < 2^-25, the relative error of the approximation e^x - 1 ~ x |
89 | // is: |
90 | // |(e^x - 1) - x| / |e^x - 1| < |x^2| / |x| |
91 | // = |x| |
92 | // < 2^-25 |
93 | // < epsilon(1)/2. |
94 | // So the correctly rounded values of expm1(x) are: |
95 | // = x + eps(x) if rounding mode = FE_UPWARD, |
96 | // or (rounding mode = FE_TOWARDZERO and x is |
97 | // negative), |
98 | // = x otherwise. |
99 | // To simplify the rounding decision and make it more efficient, we use |
100 | // fma(x, x, x) ~ x + x^2 instead. |
101 | // Note: to use the formula x + x^2 to decide the correct rounding, we |
102 | // do need fma(x, x, x) to prevent underflow caused by x*x when |x| < |
103 | // 2^-76. For targets without FMA instructions, we simply use double for |
104 | // intermediate results as it is more efficient than using an emulated |
105 | // version of FMA. |
106 | #if defined(LIBC_TARGET_CPU_HAS_FMA) |
107 | return fputil::fma(x, y: x, z: x); |
108 | #else |
109 | double xd = x; |
110 | return static_cast<float>(fputil::multiply_add(xd, xd, xd)); |
111 | #endif // LIBC_TARGET_CPU_HAS_FMA |
112 | } |
113 | |
114 | constexpr double COEFFS[] = {0x1p-1, |
115 | 0x1.55555555557ddp-3, |
116 | 0x1.55555555552fap-5, |
117 | 0x1.111110fcd58b7p-7, |
118 | 0x1.6c16c1717660bp-10, |
119 | 0x1.a0241f0006d62p-13, |
120 | 0x1.a01e3f8d3c06p-16}; |
121 | |
122 | // 2^-25 <= |x| < 2^-4 |
123 | double xd = static_cast<double>(x); |
124 | double xsq = xd * xd; |
125 | // Degree-8 minimax polynomial generated by Sollya with: |
126 | // > display = hexadecimal; |
127 | // > P = fpminimax((expm1(x) - x)/x^2, 6, [|D...|], [-2^-4, 2^-4]); |
128 | |
129 | double c0 = fputil::multiply_add(x: xd, y: COEFFS[1], z: COEFFS[0]); |
130 | double c1 = fputil::multiply_add(x: xd, y: COEFFS[3], z: COEFFS[2]); |
131 | double c2 = fputil::multiply_add(x: xd, y: COEFFS[5], z: COEFFS[4]); |
132 | |
133 | double r = fputil::polyeval(x: xsq, a0: c0, a: c1, a: c2, a: COEFFS[6]); |
134 | return static_cast<float>(fputil::multiply_add(x: r, y: xsq, z: xd)); |
135 | } |
136 | |
137 | // For -18 < x < 89, to compute expm1(x), we perform the following range |
138 | // reduction: find hi, mid, lo such that: |
139 | // x = hi + mid + lo, in which |
140 | // hi is an integer, |
141 | // mid * 2^7 is an integer |
142 | // -2^(-8) <= lo < 2^-8. |
143 | // In particular, |
144 | // hi + mid = round(x * 2^7) * 2^(-7). |
145 | // Then, |
146 | // expm1(x) = exp(hi + mid + lo) - 1 = exp(hi) * exp(mid) * exp(lo) - 1. |
147 | // We store exp(hi) and exp(mid) in the lookup tables EXP_M1 and EXP_M2 |
148 | // respectively. exp(lo) is computed using a degree-4 minimax polynomial |
149 | // generated by Sollya. |
150 | |
151 | // x_hi = hi + mid. |
152 | float kf = fputil::nearest_integer(x: x * 0x1.0p7f); |
153 | int x_hi = static_cast<int>(kf); |
154 | // Subtract (hi + mid) from x to get lo. |
155 | double xd = static_cast<double>(fputil::multiply_add(x: kf, y: -0x1.0p-7f, z: x)); |
156 | x_hi += 104 << 7; |
157 | // hi = x_hi >> 7 |
158 | double exp_hi = EXP_M1[x_hi >> 7]; |
159 | // lo = x_hi & 0x0000'007fU; |
160 | double exp_mid = EXP_M2[x_hi & 0x7f]; |
161 | double exp_hi_mid = exp_hi * exp_mid; |
162 | // Degree-4 minimax polynomial generated by Sollya with the following |
163 | // commands: |
164 | // > display = hexadecimal; |
165 | // > Q = fpminimax(expm1(x)/x, 3, [|D...|], [-2^-8, 2^-8]); |
166 | // > Q; |
167 | double exp_lo = |
168 | fputil::polyeval(x: xd, a0: 0x1.0p0, a: 0x1.ffffffffff777p-1, a: 0x1.000000000071cp-1, |
169 | a: 0x1.555566668e5e7p-3, a: 0x1.55555555ef243p-5); |
170 | return static_cast<float>(fputil::multiply_add(x: exp_hi_mid, y: exp_lo, z: -1.0)); |
171 | } |
172 | |
173 | } // namespace LIBC_NAMESPACE |
174 | |