1 | //===-- Implementation of exp2m1f 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/exp2m1f.h" |
10 | #include "src/__support/FPUtil/FEnvImpl.h" |
11 | #include "src/__support/FPUtil/FPBits.h" |
12 | #include "src/__support/FPUtil/PolyEval.h" |
13 | #include "src/__support/FPUtil/except_value_utils.h" |
14 | #include "src/__support/FPUtil/multiply_add.h" |
15 | #include "src/__support/FPUtil/rounding_mode.h" |
16 | #include "src/__support/common.h" |
17 | #include "src/__support/macros/optimization.h" |
18 | #include "src/__support/macros/properties/cpu_features.h" |
19 | #include "src/errno/libc_errno.h" |
20 | |
21 | #include "explogxf.h" |
22 | |
23 | namespace LIBC_NAMESPACE { |
24 | |
25 | static constexpr size_t N_EXCEPTS_LO = 8; |
26 | |
27 | static constexpr fputil::ExceptValues<float, N_EXCEPTS_LO> EXP2M1F_EXCEPTS_LO = |
28 | {.values: { |
29 | // (input, RZ output, RU offset, RD offset, RN offset) |
30 | // x = 0x1.36dc8ep-36, exp2m1f(x) = 0x1.aef212p-37 (RZ) |
31 | {.input: 0x2d9b'6e47U, .rnd_towardzero_result: 0x2d57'7909U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
32 | // x = 0x1.224936p-19, exp2m1f(x) = 0x1.926c0ep-20 (RZ) |
33 | {.input: 0x3611'249bU, .rnd_towardzero_result: 0x35c9'3607U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
34 | // x = 0x1.d16d2p-20, exp2m1f(x) = 0x1.429becp-20 (RZ) |
35 | {.input: 0x35e8'b690U, .rnd_towardzero_result: 0x35a1'4df6U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
36 | // x = 0x1.17949ep-14, exp2m1f(x) = 0x1.8397p-15 (RZ) |
37 | {.input: 0x388b'ca4fU, .rnd_towardzero_result: 0x3841'cb80U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 1U}, |
38 | // x = -0x1.9c3e1ep-38, exp2m1f(x) = -0x1.1dbeacp-38 (RZ) |
39 | {.input: 0xacce'1f0fU, .rnd_towardzero_result: 0xac8e'df56U, .rnd_upward_offset: 0U, .rnd_downward_offset: 1U, .rnd_tonearest_offset: 0U}, |
40 | // x = -0x1.4d89b4p-32, exp2m1f(x) = -0x1.ce61b6p-33 (RZ) |
41 | {.input: 0xafa6'c4daU, .rnd_towardzero_result: 0xaf67'30dbU, .rnd_upward_offset: 0U, .rnd_downward_offset: 1U, .rnd_tonearest_offset: 1U}, |
42 | // x = -0x1.a6eac4p-10, exp2m1f(x) = -0x1.24fadap-10 (RZ) |
43 | {.input: 0xbad3'7562U, .rnd_towardzero_result: 0xba92'7d6dU, .rnd_upward_offset: 0U, .rnd_downward_offset: 1U, .rnd_tonearest_offset: 1U}, |
44 | // x = -0x1.e7526ep-6, exp2m1f(x) = -0x1.4e53dep-6 (RZ) |
45 | {.input: 0xbcf3'a937U, .rnd_towardzero_result: 0xbca7'29efU, .rnd_upward_offset: 0U, .rnd_downward_offset: 1U, .rnd_tonearest_offset: 1U}, |
46 | }}; |
47 | |
48 | static constexpr size_t N_EXCEPTS_HI = 3; |
49 | |
50 | static constexpr fputil::ExceptValues<float, N_EXCEPTS_HI> EXP2M1F_EXCEPTS_HI = |
51 | {.values: { |
52 | // (input, RZ output, RU offset, RD offset, RN offset) |
53 | // x = 0x1.16a972p-1, exp2m1f(x) = 0x1.d545b2p-2 (RZ) |
54 | {.input: 0x3f0b'54b9U, .rnd_towardzero_result: 0x3eea'a2d9U, .rnd_upward_offset: 1U, .rnd_downward_offset: 0U, .rnd_tonearest_offset: 0U}, |
55 | // x = -0x1.9f12acp-5, exp2m1f(x) = -0x1.1ab68cp-5 (RZ) |
56 | {.input: 0xbd4f'8956U, .rnd_towardzero_result: 0xbd0d'5b46U, .rnd_upward_offset: 0U, .rnd_downward_offset: 1U, .rnd_tonearest_offset: 0U}, |
57 | // x = -0x1.de7b9cp-5, exp2m1f(x) = -0x1.4508f4p-5 (RZ) |
58 | {.input: 0xbd6f'3dceU, .rnd_towardzero_result: 0xbd22'847aU, .rnd_upward_offset: 0U, .rnd_downward_offset: 1U, .rnd_tonearest_offset: 1U}, |
59 | }}; |
60 | |
61 | LLVM_LIBC_FUNCTION(float, exp2m1f, (float x)) { |
62 | using FPBits = fputil::FPBits<float>; |
63 | FPBits xbits(x); |
64 | |
65 | uint32_t x_u = xbits.uintval(); |
66 | uint32_t x_abs = x_u & 0x7fff'ffffU; |
67 | |
68 | // When |x| >= 128, or x is nan, or |x| <= 2^-5 |
69 | if (LIBC_UNLIKELY(x_abs >= 0x4300'0000U || x_abs <= 0x3d00'0000U)) { |
70 | // |x| <= 2^-5 |
71 | if (x_abs <= 0x3d00'0000U) { |
72 | if (auto r = EXP2M1F_EXCEPTS_LO.lookup(x_bits: x_u); LIBC_UNLIKELY(r.has_value())) |
73 | return r.value(); |
74 | |
75 | // Minimax polynomial generated by Sollya with: |
76 | // > display = hexadecimal; |
77 | // > fpminimax((2^x - 1)/x, 5, [|D...|], [-2^-5, 2^-5]); |
78 | constexpr double COEFFS[] = { |
79 | 0x1.62e42fefa39f3p-1, 0x1.ebfbdff82c57bp-3, 0x1.c6b08d6f2d7aap-5, |
80 | 0x1.3b2ab6fc92f5dp-7, 0x1.5d897cfe27125p-10, 0x1.43090e61e6af1p-13}; |
81 | double xd = x; |
82 | double xsq = xd * xd; |
83 | double c0 = fputil::multiply_add(x: xd, y: COEFFS[1], z: COEFFS[0]); |
84 | double c1 = fputil::multiply_add(x: xd, y: COEFFS[3], z: COEFFS[2]); |
85 | double c2 = fputil::multiply_add(x: xd, y: COEFFS[5], z: COEFFS[4]); |
86 | double p = fputil::polyeval(x: xsq, a0: c0, a: c1, a: c2); |
87 | return static_cast<float>(p * xd); |
88 | } |
89 | |
90 | // x >= 128, or x is nan |
91 | if (xbits.is_pos()) { |
92 | if (xbits.is_finite()) { |
93 | int rounding = fputil::quick_get_round(); |
94 | if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) |
95 | return FPBits::max_normal().get_val(); |
96 | |
97 | fputil::set_errno_if_required(ERANGE); |
98 | fputil::raise_except_if_required(FE_OVERFLOW); |
99 | } |
100 | |
101 | // x >= 128 and 2^x - 1 rounds to +inf, or x is +inf or nan |
102 | return x + FPBits::inf().get_val(); |
103 | } |
104 | } |
105 | |
106 | if (LIBC_UNLIKELY(x <= -25.0f)) { |
107 | // 2^(-inf) - 1 = -1 |
108 | if (xbits.is_inf()) |
109 | return -1.0f; |
110 | // 2^nan - 1 = nan |
111 | if (xbits.is_nan()) |
112 | return x; |
113 | |
114 | int rounding = fputil::quick_get_round(); |
115 | if (rounding == FE_UPWARD || rounding == FE_TOWARDZERO) |
116 | return -0x1.ffff'fep-1f; // -1.0f + 0x1.0p-24f |
117 | |
118 | fputil::set_errno_if_required(ERANGE); |
119 | fputil::raise_except_if_required(FE_UNDERFLOW); |
120 | return -1.0f; |
121 | } |
122 | |
123 | if (auto r = EXP2M1F_EXCEPTS_HI.lookup(x_bits: x_u); LIBC_UNLIKELY(r.has_value())) |
124 | return r.value(); |
125 | |
126 | // For -25 < x < 128, to compute 2^x, we perform the following range |
127 | // reduction: find hi, mid, lo such that: |
128 | // x = hi + mid + lo, in which: |
129 | // hi is an integer, |
130 | // 0 <= mid * 2^5 < 32 is an integer, |
131 | // -2^(-6) <= lo <= 2^(-6). |
132 | // In particular, |
133 | // hi + mid = round(x * 2^5) * 2^(-5). |
134 | // Then, |
135 | // 2^x = 2^(hi + mid + lo) = 2^hi * 2^mid * 2^lo. |
136 | // 2^mid is stored in the lookup table of 32 elements. |
137 | // 2^lo is computed using a degree-4 minimax polynomial generated by Sollya. |
138 | // We perform 2^hi * 2^mid by simply add hi to the exponent field of 2^mid. |
139 | |
140 | // kf = (hi + mid) * 2^5 = round(x * 2^5) |
141 | float kf; |
142 | int k; |
143 | #ifdef LIBC_TARGET_CPU_HAS_NEAREST_INT |
144 | kf = fputil::nearest_integer(x: x * 32.0f); |
145 | k = static_cast<int>(kf); |
146 | #else |
147 | constexpr float HALF[2] = {0.5f, -0.5f}; |
148 | k = static_cast<int>(fputil::multiply_add(x, 32.0f, HALF[x < 0.0f])); |
149 | kf = static_cast<float>(k); |
150 | #endif // LIBC_TARGET_CPU_HAS_NEAREST_INT |
151 | |
152 | // lo = x - (hi + mid) = x - kf * 2^(-5) |
153 | double lo = fputil::multiply_add(x: -0x1.0p-5f, y: kf, z: x); |
154 | |
155 | // hi = floor(kf * 2^(-4)) |
156 | // exp2_hi = shift hi to the exponent field of double precision. |
157 | int64_t exp2_hi = |
158 | static_cast<int64_t>(static_cast<uint64_t>(k >> ExpBase::MID_BITS) |
159 | << fputil::FPBits<double>::FRACTION_LEN); |
160 | // mh = 2^hi * 2^mid |
161 | // mh_bits = bit field of mh |
162 | int64_t mh_bits = ExpBase::EXP_2_MID[k & ExpBase::MID_MASK] + exp2_hi; |
163 | double mh = fputil::FPBits<double>(static_cast<uint64_t>(mh_bits)).get_val(); |
164 | |
165 | // Degree-4 polynomial approximating (2^x - 1)/x generated by Sollya with: |
166 | // > display = hexadecimal; |
167 | // > fpminimax((2^x - 1)/x, 4, [|D...|], [-2^-6, 2^-6]); |
168 | constexpr double COEFFS[5] = {0x1.62e42fefa39efp-1, 0x1.ebfbdff8131c4p-3, |
169 | 0x1.c6b08d7061695p-5, 0x1.3b2b1bee74b2ap-7, |
170 | 0x1.5d88091198529p-10}; |
171 | double lo_sq = lo * lo; |
172 | double c1 = fputil::multiply_add(x: lo, y: COEFFS[0], z: 1.0); |
173 | double c2 = fputil::multiply_add(x: lo, y: COEFFS[2], z: COEFFS[1]); |
174 | double c3 = fputil::multiply_add(x: lo, y: COEFFS[4], z: COEFFS[3]); |
175 | double exp2_lo = fputil::polyeval(x: lo_sq, a0: c1, a: c2, a: c3); |
176 | // 2^x - 1 = 2^(hi + mid + lo) - 1 |
177 | // = 2^(hi + mid) * 2^lo - 1 |
178 | // ~ mh * (1 + lo * P(lo)) - 1 |
179 | // = mh * exp2_lo - 1 |
180 | return static_cast<float>(fputil::multiply_add(x: exp2_lo, y: mh, z: -1.0)); |
181 | } |
182 | |
183 | } // namespace LIBC_NAMESPACE |
184 | |