1 | //===-- Half-precision log2(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 | #include "src/math/log2f16.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/common.h" |
20 | #include "src/__support/macros/config.h" |
21 | #include "src/__support/macros/optimization.h" |
22 | #include "src/__support/macros/properties/cpu_features.h" |
23 | |
24 | namespace LIBC_NAMESPACE_DECL { |
25 | |
26 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
27 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
28 | static constexpr size_t N_LOG2F16_EXCEPTS = 2; |
29 | #else |
30 | static constexpr size_t N_LOG2F16_EXCEPTS = 9; |
31 | #endif |
32 | |
33 | static constexpr fputil::ExceptValues<float16, N_LOG2F16_EXCEPTS> |
34 | LOG2F16_EXCEPTS = {{ |
35 | // (input, RZ output, RU offset, RD offset, RN offset) |
36 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
37 | // x = 0x1.224p-1, log2f16(x) = -0x1.a34p-1 (RZ) |
38 | {0x3889U, 0xba8dU, 0U, 1U, 0U}, |
39 | // x = 0x1.e34p-1, log2f16(x) = -0x1.558p-4 (RZ) |
40 | {0x3b8dU, 0xad56U, 0U, 1U, 0U}, |
41 | #endif |
42 | // x = 0x1.e8cp-1, log2f16(x) = -0x1.128p-4 (RZ) |
43 | {0x3ba3U, 0xac4aU, 0U, 1U, 0U}, |
44 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
45 | // x = 0x1.f98p-1, log2f16(x) = -0x1.2ep-6 (RZ) |
46 | {0x3be6U, 0xa4b8U, 0U, 1U, 0U}, |
47 | // x = 0x1.facp-1, log2f16(x) = -0x1.e7p-7 (RZ) |
48 | {0x3bebU, 0xa39cU, 0U, 1U, 1U}, |
49 | #endif |
50 | // x = 0x1.fb4p-1, log2f16(x) = -0x1.b88p-7 (RZ) |
51 | {0x3bedU, 0xa2e2U, 0U, 1U, 1U}, |
52 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
53 | // x = 0x1.fecp-1, log2f16(x) = -0x1.cep-9 (RZ) |
54 | {0x3bfbU, 0x9b38U, 0U, 1U, 1U}, |
55 | // x = 0x1.ffcp-1, log2f16(x) = -0x1.714p-11 (RZ) |
56 | {0x3bffU, 0x91c5U, 0U, 1U, 1U}, |
57 | // x = 0x1.224p+0, log2f16(x) = 0x1.72cp-3 (RZ) |
58 | {0x3c89U, 0x31cbU, 1U, 0U, 1U}, |
59 | #endif |
60 | }}; |
61 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
62 | |
63 | LLVM_LIBC_FUNCTION(float16, log2f16, (float16 x)) { |
64 | using FPBits = fputil::FPBits<float16>; |
65 | FPBits x_bits(x); |
66 | |
67 | uint16_t x_u = x_bits.uintval(); |
68 | |
69 | // If x <= 0, or x is 1, or x is +inf, or x is NaN. |
70 | if (LIBC_UNLIKELY(x_u == 0U || x_u == 0x3c00U || x_u >= 0x7c00U)) { |
71 | // log2(NaN) = NaN |
72 | if (x_bits.is_nan()) { |
73 | if (x_bits.is_signaling_nan()) { |
74 | fputil::raise_except_if_required(FE_INVALID); |
75 | return FPBits::quiet_nan().get_val(); |
76 | } |
77 | |
78 | return x; |
79 | } |
80 | |
81 | // log2(+/-0) = −inf |
82 | if ((x_u & 0x7fffU) == 0U) { |
83 | fputil::raise_except_if_required(FE_DIVBYZERO); |
84 | return FPBits::inf(Sign::NEG).get_val(); |
85 | } |
86 | |
87 | if (x_u == 0x3c00U) |
88 | return FPBits::zero().get_val(); |
89 | |
90 | // When x < 0. |
91 | if (x_u > 0x8000U) { |
92 | fputil::set_errno_if_required(EDOM); |
93 | fputil::raise_except_if_required(FE_INVALID); |
94 | return FPBits::quiet_nan().get_val(); |
95 | } |
96 | |
97 | // log2(+inf) = +inf |
98 | return FPBits::inf().get_val(); |
99 | } |
100 | |
101 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
102 | if (auto r = LOG2F16_EXCEPTS.lookup(x_u); LIBC_UNLIKELY(r.has_value())) |
103 | return r.value(); |
104 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
105 | |
106 | // To compute log2(x), we perform the following range reduction: |
107 | // x = 2^m * 1.mant, |
108 | // log2(x) = m + log2(1.mant). |
109 | // To compute log2(1.mant), let f be the highest 6 bits including the hidden |
110 | // bit, and d be the difference (1.mant - f), i.e., the remaining 5 bits of |
111 | // the mantissa, then: |
112 | // log2(1.mant) = log2(f) + log2(1.mant / f) |
113 | // = log2(f) + log2(1 + d/f) |
114 | // since d/f is sufficiently small. |
115 | // We store log2(f) and 1/f in the lookup tables LOG2F_F and ONE_OVER_F_F |
116 | // respectively. |
117 | |
118 | int m = -FPBits::EXP_BIAS; |
119 | |
120 | // When x is subnormal, normalize it. |
121 | if ((x_u & FPBits::EXP_MASK) == 0U) { |
122 | // Can't pass an integer to fputil::cast directly. |
123 | constexpr float NORMALIZE_EXP = 1U << FPBits::FRACTION_LEN; |
124 | x_bits = FPBits(x_bits.get_val() * fputil::cast<float16>(NORMALIZE_EXP)); |
125 | x_u = x_bits.uintval(); |
126 | m -= FPBits::FRACTION_LEN; |
127 | } |
128 | |
129 | uint16_t mant = x_bits.get_mantissa(); |
130 | // Leading 10 - 5 = 5 bits of the mantissa. |
131 | int f = mant >> 5; |
132 | // Unbiased exponent. |
133 | m += x_u >> FPBits::FRACTION_LEN; |
134 | |
135 | // Set bits to 1.mant instead of 2^m * 1.mant. |
136 | x_bits.set_biased_exponent(FPBits::EXP_BIAS); |
137 | float mant_f = x_bits.get_val(); |
138 | // v = 1.mant * 1/f - 1 = d/f |
139 | float v = fputil::multiply_add(mant_f, ONE_OVER_F_F[f], -1.0f); |
140 | |
141 | // Degree-3 minimax polynomial generated by Sollya with the following |
142 | // commands: |
143 | // > display = hexadecimal; |
144 | // > P = fpminimax(log2(1 + x)/x, 2, [|SG...|], [-2^-5, 2^-5]); |
145 | // > x * P; |
146 | float log2p1_d_over_f = |
147 | v * fputil::polyeval(v, 0x1.715476p+0f, -0x1.71771ap-1f, 0x1.ecb38ep-2f); |
148 | // log2(1.mant) = log2(f) + log2(1 + d/f) |
149 | float log2_1_mant = LOG2F_F[f] + log2p1_d_over_f; |
150 | return fputil::cast<float16>(static_cast<float>(m) + log2_1_mant); |
151 | } |
152 | |
153 | } // namespace LIBC_NAMESPACE_DECL |
154 | |