1 | //===-- Half-precision 2^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/exp2m1f16.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, 6> EXP2M1F16_EXCEPTS_LO = {{ |
29 | // (input, RZ output, RU offset, RD offset, RN offset) |
30 | // x = 0x1.cf4p-13, exp2m1f16(x) = 0x1.41p-13 (RZ) |
31 | {0x0b3dU, 0x0904U, 1U, 0U, 1U}, |
32 | // x = 0x1.4fcp-12, exp2m1f16(x) = 0x1.d14p-13 (RZ) |
33 | {0x0d3fU, 0x0b45U, 1U, 0U, 1U}, |
34 | // x = 0x1.63p-11, exp2m1f16(x) = 0x1.ec4p-12 (RZ) |
35 | {0x118cU, 0x0fb1U, 1U, 0U, 0U}, |
36 | // x = 0x1.6fp-7, exp2m1f16(x) = 0x1.fe8p-8 (RZ) |
37 | {0x21bcU, 0x1ffaU, 1U, 0U, 1U}, |
38 | // x = -0x1.c6p-10, exp2m1f16(x) = -0x1.3a8p-10 (RZ) |
39 | {0x9718U, 0x94eaU, 0U, 1U, 0U}, |
40 | // x = -0x1.cfcp-10, exp2m1f16(x) = -0x1.414p-10 (RZ) |
41 | {0x973fU, 0x9505U, 0U, 1U, 0U}, |
42 | }}; |
43 | |
44 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
45 | static constexpr size_t N_EXP2M1F16_EXCEPTS_HI = 6; |
46 | #else |
47 | static constexpr size_t N_EXP2M1F16_EXCEPTS_HI = 7; |
48 | #endif |
49 | |
50 | static constexpr fputil::ExceptValues<float16, N_EXP2M1F16_EXCEPTS_HI> |
51 | EXP2M1F16_EXCEPTS_HI = {{ |
52 | // (input, RZ output, RU offset, RD offset, RN offset) |
53 | // x = 0x1.e58p-3, exp2m1f16(x) = 0x1.6dcp-3 (RZ) |
54 | {0x3396U, 0x31b7U, 1U, 0U, 0U}, |
55 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
56 | // x = 0x1.2e8p-2, exp2m1f16(x) = 0x1.d14p-3 (RZ) |
57 | {0x34baU, 0x3345U, 1U, 0U, 0U}, |
58 | #endif |
59 | // x = 0x1.ad8p-2, exp2m1f16(x) = 0x1.598p-2 (RZ) |
60 | {0x36b6U, 0x3566U, 1U, 0U, 0U}, |
61 | #ifdef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
62 | // x = 0x1.edcp-2, exp2m1f16(x) = 0x1.964p-2 (RZ) |
63 | {0x37b7U, 0x3659U, 1U, 0U, 1U}, |
64 | #endif |
65 | // x = -0x1.804p-3, exp2m1f16(x) = -0x1.f34p-4 (RZ) |
66 | {0xb201U, 0xafcdU, 0U, 1U, 1U}, |
67 | // x = -0x1.f3p-3, exp2m1f16(x) = -0x1.3e4p-3 (RZ) |
68 | {0xb3ccU, 0xb0f9U, 0U, 1U, 0U}, |
69 | // x = -0x1.294p-1, exp2m1f16(x) = -0x1.53p-2 (RZ) |
70 | {0xb8a5U, 0xb54cU, 0U, 1U, 1U}, |
71 | #ifndef LIBC_TARGET_CPU_HAS_FMA_FLOAT |
72 | // x = -0x1.a34p-1, exp2m1f16(x) = -0x1.bb4p-2 (RZ) |
73 | {0xba8dU, 0xb6edU, 0U, 1U, 1U}, |
74 | #endif |
75 | }}; |
76 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
77 | |
78 | LLVM_LIBC_FUNCTION(float16, exp2m1f16, (float16 x)) { |
79 | using FPBits = fputil::FPBits<float16>; |
80 | FPBits x_bits(x); |
81 | |
82 | uint16_t x_u = x_bits.uintval(); |
83 | uint16_t x_abs = x_u & 0x7fffU; |
84 | |
85 | // When |x| <= 2^(-3), or |x| >= 11, or x is NaN. |
86 | if (LIBC_UNLIKELY(x_abs <= 0x3000U || x_abs >= 0x4980U)) { |
87 | // exp2m1(NaN) = NaN |
88 | if (x_bits.is_nan()) { |
89 | if (x_bits.is_signaling_nan()) { |
90 | fputil::raise_except_if_required(FE_INVALID); |
91 | return FPBits::quiet_nan().get_val(); |
92 | } |
93 | |
94 | return x; |
95 | } |
96 | |
97 | // When x >= 16. |
98 | if (x_u >= 0x4c00 && x_bits.is_pos()) { |
99 | // exp2m1(+inf) = +inf |
100 | if (x_bits.is_inf()) |
101 | return FPBits::inf().get_val(); |
102 | |
103 | switch (fputil::quick_get_round()) { |
104 | case FE_TONEAREST: |
105 | case FE_UPWARD: |
106 | fputil::set_errno_if_required(ERANGE); |
107 | fputil::raise_except_if_required(FE_OVERFLOW | FE_INEXACT); |
108 | return FPBits::inf().get_val(); |
109 | default: |
110 | return FPBits::max_normal().get_val(); |
111 | } |
112 | } |
113 | |
114 | // When x < -11. |
115 | if (x_u > 0xc980U) { |
116 | // exp2m1(-inf) = -1 |
117 | if (x_bits.is_inf()) |
118 | return FPBits::one(Sign::NEG).get_val(); |
119 | |
120 | // When -12 < x < -11, round(2^x - 1, HP, RN) = -0x1.ffcp-1. |
121 | if (x_u < 0xca00U) |
122 | return fputil::round_result_slightly_down( |
123 | fputil::cast<float16>(-0x1.ffcp-1)); |
124 | |
125 | // When x <= -12, round(2^x - 1, HP, RN) = -1. |
126 | switch (fputil::quick_get_round()) { |
127 | case FE_TONEAREST: |
128 | case FE_DOWNWARD: |
129 | return FPBits::one(Sign::NEG).get_val(); |
130 | default: |
131 | return fputil::cast<float16>(-0x1.ffcp-1); |
132 | } |
133 | } |
134 | |
135 | // When |x| <= 2^(-3). |
136 | if (x_abs <= 0x3000U) { |
137 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
138 | if (auto r = EXP2M1F16_EXCEPTS_LO.lookup(x_u); |
139 | LIBC_UNLIKELY(r.has_value())) |
140 | return r.value(); |
141 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
142 | |
143 | float xf = x; |
144 | // Degree-5 minimax polynomial generated by Sollya with the following |
145 | // commands: |
146 | // > display = hexadecimal; |
147 | // > P = fpminimax((2^x - 1)/x, 4, [|SG...|], [-2^-3, 2^-3]); |
148 | // > x * P; |
149 | return fputil::cast<float16>( |
150 | xf * fputil::polyeval(xf, 0x1.62e43p-1f, 0x1.ebfbdep-3f, |
151 | 0x1.c6af88p-5f, 0x1.3b45d6p-7f, |
152 | 0x1.641e7cp-10f)); |
153 | } |
154 | } |
155 | |
156 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
157 | if (auto r = EXP2M1F16_EXCEPTS_HI.lookup(x_u); LIBC_UNLIKELY(r.has_value())) |
158 | return r.value(); |
159 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
160 | |
161 | // exp2(x) = exp2(hi + mid) * exp2(lo) |
162 | auto [exp2_hi_mid, exp2_lo] = exp2_range_reduction(x); |
163 | // exp2m1(x) = exp2(hi + mid) * exp2(lo) - 1 |
164 | return fputil::cast<float16>( |
165 | fputil::multiply_add(exp2_hi_mid, exp2_lo, -1.0f)); |
166 | } |
167 | |
168 | } // namespace LIBC_NAMESPACE_DECL |
169 | |