1 | //===-- Half-precision 10^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/exp10m1f16.h" |
10 | #include "expxf16.h" |
11 | #include "hdr/errno_macros.h" |
12 | #include "hdr/fenv_macros.h" |
13 | #include "src/__support/FPUtil/FEnvImpl.h" |
14 | #include "src/__support/FPUtil/FPBits.h" |
15 | #include "src/__support/FPUtil/PolyEval.h" |
16 | #include "src/__support/FPUtil/cast.h" |
17 | #include "src/__support/FPUtil/except_value_utils.h" |
18 | #include "src/__support/FPUtil/multiply_add.h" |
19 | #include "src/__support/FPUtil/rounding_mode.h" |
20 | #include "src/__support/common.h" |
21 | #include "src/__support/macros/config.h" |
22 | #include "src/__support/macros/optimization.h" |
23 | #include "src/__support/macros/properties/cpu_features.h" |
24 | |
25 | namespace LIBC_NAMESPACE_DECL { |
26 | |
27 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
28 | static constexpr fputil::ExceptValues<float16, 3> EXP10M1F16_EXCEPTS_LO = {{ |
29 | // (input, RZ output, RU offset, RD offset, RN offset) |
30 | // x = 0x1.5c4p-4, exp10m1f16(x) = 0x1.bacp-3 (RZ) |
31 | {0x2d71U, 0x32ebU, 1U, 0U, 0U}, |
32 | // x = -0x1.5ep-13, exp10m1f16(x) = -0x1.92cp-12 (RZ) |
33 | {0x8978U, 0x8e4bU, 0U, 1U, 0U}, |
34 | // x = -0x1.e2p-10, exp10m1f16(x) = -0x1.14cp-8 (RZ) |
35 | {0x9788U, 0x9c53U, 0U, 1U, 0U}, |
36 | }}; |
37 | |
38 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
39 | static constexpr size_t N_EXP10M1F16_EXCEPTS_HI = 3; |
40 | #else |
41 | static constexpr size_t N_EXP10M1F16_EXCEPTS_HI = 6; |
42 | #endif |
43 | |
44 | static constexpr fputil::ExceptValues<float16, N_EXP10M1F16_EXCEPTS_HI> |
45 | EXP10M1F16_EXCEPTS_HI = {{ |
46 | // (input, RZ output, RU offset, RD offset, RN offset) |
47 | // x = 0x1.8f4p-2, exp10m1f16(x) = 0x1.744p+0 (RZ) |
48 | {0x363dU, 0x3dd1U, 1U, 0U, 0U}, |
49 | // x = 0x1.95cp-2, exp10m1f16(x) = 0x1.7d8p+0 (RZ) |
50 | {0x3657U, 0x3df6U, 1U, 0U, 0U}, |
51 | // x = 0x1.d04p-2, exp10m1f16(x) = 0x1.d7p+0 (RZ) |
52 | {0x3741U, 0x3f5cU, 1U, 0U, 1U}, |
53 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
54 | // x = 0x1.0cp+1, exp10m1f16(x) = 0x1.ec4p+6 (RZ) |
55 | {0x4030U, 0x57b1U, 1U, 0U, 1U}, |
56 | // x = 0x1.1b8p+1, exp10m1f16(x) = 0x1.45cp+7 (RZ) |
57 | {0x406eU, 0x5917U, 1U, 0U, 1U}, |
58 | // x = 0x1.2f4p+2, exp10m1f16(x) = 0x1.ab8p+15 (RZ) |
59 | {0x44bdU, 0x7aaeU, 1U, 0U, 1U}, |
60 | #endif |
61 | }}; |
62 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
63 | |
64 | LLVM_LIBC_FUNCTION(float16, exp10m1f16, (float16 x)) { |
65 | using FPBits = fputil::FPBits<float16>; |
66 | FPBits x_bits(x); |
67 | |
68 | uint16_t x_u = x_bits.uintval(); |
69 | uint16_t x_abs = x_u & 0x7fffU; |
70 | |
71 | // When |x| <= 2^(-3), or |x| >= 11 * log10(2), or x is NaN. |
72 | if (LIBC_UNLIKELY(x_abs <= 0x3000U || x_abs >= 0x429fU)) { |
73 | // exp10m1(NaN) = NaN |
74 | if (x_bits.is_nan()) { |
75 | if (x_bits.is_signaling_nan()) { |
76 | fputil::raise_except_if_required(FE_INVALID); |
77 | return FPBits::quiet_nan().get_val(); |
78 | } |
79 | |
80 | return x; |
81 | } |
82 | |
83 | // When x >= 16 * log10(2). |
84 | if (x_u >= 0x44d1U && x_bits.is_pos()) { |
85 | // exp10m1(+inf) = +inf |
86 | if (x_bits.is_inf()) |
87 | return FPBits::inf().get_val(); |
88 | |
89 | switch (fputil::quick_get_round()) { |
90 | case FE_TONEAREST: |
91 | case FE_UPWARD: |
92 | fputil::set_errno_if_required(ERANGE); |
93 | fputil::raise_except_if_required(FE_OVERFLOW | FE_INEXACT); |
94 | return FPBits::inf().get_val(); |
95 | default: |
96 | return FPBits::max_normal().get_val(); |
97 | } |
98 | } |
99 | |
100 | // When x < -11 * log10(2). |
101 | if (x_u > 0xc29fU) { |
102 | // exp10m1(-inf) = -1 |
103 | if (x_bits.is_inf()) |
104 | return FPBits::one(Sign::NEG).get_val(); |
105 | |
106 | // When x >= -0x1.ce4p+1, round(10^x - 1, HP, RN) = -0x1.ffcp-1. |
107 | if (x_u <= 0xc339U) { |
108 | return fputil::round_result_slightly_down( |
109 | fputil::cast<float16>(-0x1.ffcp-1)); |
110 | } |
111 | |
112 | // When x < -0x1.ce4p+1, round(10^x - 1, HP, RN) = -1. |
113 | switch (fputil::quick_get_round()) { |
114 | case FE_TONEAREST: |
115 | case FE_DOWNWARD: |
116 | return FPBits::one(Sign::NEG).get_val(); |
117 | default: |
118 | return fputil::cast<float16>(-0x1.ffcp-1); |
119 | } |
120 | } |
121 | |
122 | // When |x| <= 2^(-3). |
123 | if (x_abs <= 0x3000U) { |
124 | if (LIBC_UNLIKELY(x_abs == 0)) |
125 | return x; |
126 | |
127 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
128 | if (auto r = EXP10M1F16_EXCEPTS_LO.lookup(x_u); |
129 | LIBC_UNLIKELY(r.has_value())) |
130 | return r.value(); |
131 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
132 | |
133 | float xf = x; |
134 | // Degree-5 minimax polynomial generated by Sollya with the following |
135 | // commands: |
136 | // > display = hexadecimal; |
137 | // > P = fpminimax((10^x - 1)/x, 4, [|SG...|], [-2^-3, 2^-3]); |
138 | // > x * P; |
139 | return fputil::cast<float16>( |
140 | xf * fputil::polyeval(xf, 0x1.26bb1cp+1f, 0x1.5351c8p+1f, |
141 | 0x1.04704p+1f, 0x1.2ce084p+0f, 0x1.14a6bep-1f)); |
142 | } |
143 | } |
144 | |
145 | // When x is 1, 2, or 3. These are hard-to-round cases with exact results. |
146 | // 10^4 - 1 = 9'999 is not exactly representable as a float16, but luckily the |
147 | // polynomial approximation gives the correct result for x = 4 in all |
148 | // rounding modes. |
149 | if (LIBC_UNLIKELY((x_u & ~(0x3c00U | 0x4000U | 0x4200U | 0x4400U)) == 0)) { |
150 | switch (x_u) { |
151 | case 0x3c00U: // x = 1.0f16 |
152 | return fputil::cast<float16>(9.0); |
153 | case 0x4000U: // x = 2.0f16 |
154 | return fputil::cast<float16>(99.0); |
155 | case 0x4200U: // x = 3.0f16 |
156 | return fputil::cast<float16>(999.0); |
157 | } |
158 | } |
159 | |
160 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
161 | if (auto r = EXP10M1F16_EXCEPTS_HI.lookup(x_u); LIBC_UNLIKELY(r.has_value())) |
162 | return r.value(); |
163 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
164 | |
165 | // exp10(x) = exp2((hi + mid) * log2(10)) * exp10(lo) |
166 | auto [exp2_hi_mid, exp10_lo] = exp10_range_reduction(x); |
167 | // exp10m1(x) = exp2((hi + mid) * log2(lo)) * exp10(lo) - 1 |
168 | return fputil::cast<float16>( |
169 | fputil::multiply_add(exp2_hi_mid, exp10_lo, -1.0f)); |
170 | } |
171 | |
172 | } // namespace LIBC_NAMESPACE_DECL |
173 | |