1 | //===-- Half-precision asinh(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/asinhf16.h" |
10 | #include "explogxf.h" |
11 | #include "hdr/fenv_macros.h" |
12 | #include "src/__support/FPUtil/FEnvImpl.h" |
13 | #include "src/__support/FPUtil/FPBits.h" |
14 | #include "src/__support/FPUtil/PolyEval.h" |
15 | #include "src/__support/FPUtil/cast.h" |
16 | #include "src/__support/FPUtil/except_value_utils.h" |
17 | #include "src/__support/FPUtil/multiply_add.h" |
18 | #include "src/__support/FPUtil/rounding_mode.h" |
19 | #include "src/__support/FPUtil/sqrt.h" |
20 | #include "src/__support/common.h" |
21 | #include "src/__support/macros/config.h" |
22 | #include "src/__support/macros/optimization.h" |
23 | |
24 | namespace LIBC_NAMESPACE_DECL { |
25 | |
26 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
27 | static constexpr size_t N_EXCEPTS = 8; |
28 | |
29 | static constexpr fputil::ExceptValues<float16, N_EXCEPTS> ASINHF16_EXCEPTS{{ |
30 | // (input, RZ output, RU offset, RD offset, RN offset) |
31 | |
32 | // x = 0x1.da4p-2, asinhf16(x) = 0x1.ca8p-2 (RZ) |
33 | {0x3769, 0x372a, 1, 0, 1}, |
34 | // x = 0x1.d6cp-1, asinhf16(x) = 0x1.a58p-1 (RZ) |
35 | {0x3b5b, 0x3a96, 1, 0, 0}, |
36 | // x = 0x1.c7cp+3, asinhf16(x) = 0x1.accp+1 (RZ) |
37 | {0x4b1f, 0x42b3, 1, 0, 0}, |
38 | // x = 0x1.26cp+4, asinhf16(x) = 0x1.cd8p+1 (RZ) |
39 | {0x4c9b, 0x4336, 1, 0, 1}, |
40 | // x = -0x1.da4p-2, asinhf16(x) = -0x1.ca8p-2 (RZ) |
41 | {0xb769, 0xb72a, 0, 1, 1}, |
42 | // x = -0x1.d6cp-1, asinhf16(x) = -0x1.a58p-1 (RZ) |
43 | {0xbb5b, 0xba96, 0, 1, 0}, |
44 | // x = -0x1.c7cp+3, asinhf16(x) = -0x1.accp+1 (RZ) |
45 | {0xcb1f, 0xc2b3, 0, 1, 0}, |
46 | // x = -0x1.26cp+4, asinhf16(x) = -0x1.cd8p+1 (RZ) |
47 | {0xcc9b, 0xc336, 0, 1, 1}, |
48 | }}; |
49 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
50 | |
51 | LLVM_LIBC_FUNCTION(float16, asinhf16, (float16 x)) { |
52 | using FPBits = fputil::FPBits<float16>; |
53 | FPBits xbits(x); |
54 | |
55 | uint16_t x_u = xbits.uintval(); |
56 | uint16_t x_abs = x_u & 0x7fff; |
57 | |
58 | if (LIBC_UNLIKELY(xbits.is_inf_or_nan())) { |
59 | if (xbits.is_signaling_nan()) { |
60 | fputil::raise_except_if_required(FE_INVALID); |
61 | return FPBits::quiet_nan().get_val(); |
62 | } |
63 | |
64 | return x; |
65 | } |
66 | |
67 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
68 | // Handle exceptional values |
69 | if (auto r = ASINHF16_EXCEPTS.lookup(x_u); LIBC_UNLIKELY(r.has_value())) |
70 | return r.value(); |
71 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
72 | |
73 | float xf = x; |
74 | const float SIGN[2] = {1.0f, -1.0f}; |
75 | float x_sign = SIGN[x_u >> 15]; |
76 | |
77 | // |x| <= 0.25 |
78 | if (LIBC_UNLIKELY(x_abs <= 0x3400)) { |
79 | // when |x| < 0x1.718p-5, asinhf16(x) = x. Adjust by 1 ULP for certain |
80 | // rounding types. |
81 | if (LIBC_UNLIKELY(x_abs < 0x29c6)) { |
82 | int rounding = fputil::quick_get_round(); |
83 | if ((rounding == FE_UPWARD || rounding == FE_TOWARDZERO) && xf < 0) |
84 | return fputil::cast<float16>(xf + 0x1p-24f); |
85 | if ((rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) && xf > 0) |
86 | return fputil::cast<float16>(xf - 0x1p-24f); |
87 | return fputil::cast<float16>(xf); |
88 | } |
89 | |
90 | float x_sq = xf * xf; |
91 | // Generated by Sollya with: |
92 | // > P = fpminimax(asinh(x)/x, [|0, 2, 4, 6, 8|], [|SG...|], [0, 2^-2]); |
93 | // The last coefficient 0x1.bd114ep-6f has been changed to 0x1.bd114ep-5f |
94 | // for better accuracy. |
95 | float p = fputil::polyeval(x_sq, 1.0f, -0x1.555552p-3f, 0x1.332f6ap-4f, |
96 | -0x1.6c53dep-5f, 0x1.bd114ep-5f); |
97 | |
98 | return fputil::cast<float16>(xf * p); |
99 | } |
100 | |
101 | // General case: asinh(x) = ln(x + sqrt(x^2 + 1)) |
102 | float sqrt_term = fputil::sqrt<float>(fputil::multiply_add(xf, xf, 1.0f)); |
103 | return fputil::cast<float16>( |
104 | x_sign * log_eval(fputil::multiply_add(xf, x_sign, sqrt_term))); |
105 | } |
106 | |
107 | } // namespace LIBC_NAMESPACE_DECL |
108 | |