1 | //===-- Half-precision sin(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/sinf16.h" |
10 | #include "hdr/errno_macros.h" |
11 | #include "hdr/fenv_macros.h" |
12 | #include "sincosf16_utils.h" |
13 | #include "src/__support/FPUtil/FEnvImpl.h" |
14 | #include "src/__support/FPUtil/FPBits.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/macros/optimization.h" |
19 | |
20 | namespace LIBC_NAMESPACE_DECL { |
21 | |
22 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
23 | constexpr size_t N_EXCEPTS = 4; |
24 | |
25 | constexpr fputil::ExceptValues<float16, N_EXCEPTS> SINF16_EXCEPTS{{ |
26 | // (input, RZ output, RU offset, RD offset, RN offset) |
27 | {0x2b45, 0x2b43, 1, 0, 1}, |
28 | {0x585c, 0x3ba3, 1, 0, 1}, |
29 | {0x5cb0, 0xbbff, 0, 1, 0}, |
30 | {0x51f5, 0xb80f, 0, 1, 0}, |
31 | }}; |
32 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
33 | |
34 | LLVM_LIBC_FUNCTION(float16, sinf16, (float16 x)) { |
35 | using FPBits = fputil::FPBits<float16>; |
36 | FPBits xbits(x); |
37 | |
38 | uint16_t x_u = xbits.uintval(); |
39 | uint16_t x_abs = x_u & 0x7fff; |
40 | float xf = x; |
41 | |
42 | // Range reduction: |
43 | // For |x| > pi/32, we perform range reduction as follows: |
44 | // Find k and y such that: |
45 | // x = (k + y) * pi/32 |
46 | // k is an integer, |y| < 0.5 |
47 | // |
48 | // This is done by performing: |
49 | // k = round(x * 32/pi) |
50 | // y = x * 32/pi - k |
51 | // |
52 | // Once k and y are computed, we then deduce the answer by the sine of sum |
53 | // formula: |
54 | // sin(x) = sin((k + y) * pi/32) |
55 | // = sin(k * pi/32) * cos(y * pi/32) + |
56 | // sin(y * pi/32) * cos(k * pi/32) |
57 | |
58 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
59 | // Handle exceptional values |
60 | bool x_sign = x_u >> 15; |
61 | |
62 | if (auto r = SINF16_EXCEPTS.lookup_odd(x_abs, x_sign); |
63 | LIBC_UNLIKELY(r.has_value())) |
64 | return r.value(); |
65 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
66 | |
67 | int rounding = fputil::quick_get_round(); |
68 | |
69 | // Exhaustive tests show that for |x| <= 0x1.f4p-11, 1ULP rounding errors |
70 | // occur. To fix this, the following apply: |
71 | if (LIBC_UNLIKELY(x_abs <= 0x13d0)) { |
72 | // sin(+/-0) = +/-0 |
73 | if (LIBC_UNLIKELY(x_abs == 0U)) |
74 | return x; |
75 | |
76 | // When x > 0, and rounding upward, sin(x) == x. |
77 | // When x < 0, and rounding downward, sin(x) == x. |
78 | if ((rounding == FE_UPWARD && xbits.is_pos()) || |
79 | (rounding == FE_DOWNWARD && xbits.is_neg())) |
80 | return x; |
81 | |
82 | // When x < 0, and rounding upward, sin(x) == (x - 1ULP) |
83 | if (rounding == FE_UPWARD && xbits.is_neg()) { |
84 | x_u--; |
85 | return FPBits(x_u).get_val(); |
86 | } |
87 | } |
88 | |
89 | if (xbits.is_inf_or_nan()) { |
90 | if (xbits.is_signaling_nan()) { |
91 | fputil::raise_except_if_required(FE_INVALID); |
92 | return FPBits::quiet_nan().get_val(); |
93 | } |
94 | |
95 | if (xbits.is_inf()) { |
96 | fputil::set_errno_if_required(EDOM); |
97 | fputil::raise_except_if_required(FE_INVALID); |
98 | } |
99 | |
100 | return x + FPBits::quiet_nan().get_val(); |
101 | } |
102 | |
103 | float sin_k, cos_k, sin_y, cosm1_y; |
104 | sincosf16_eval(xf, sin_k, cos_k, sin_y, cosm1_y); |
105 | |
106 | if (LIBC_UNLIKELY(sin_y == 0 && sin_k == 0)) |
107 | return FPBits::zero(xbits.sign()).get_val(); |
108 | |
109 | // Since, cosm1_y = cos_y - 1, therefore: |
110 | // sin(x) = cos_k * sin_y + sin_k + (cosm1_y * sin_k) |
111 | return fputil::cast<float16>(fputil::multiply_add( |
112 | sin_y, cos_k, fputil::multiply_add(cosm1_y, sin_k, sin_k))); |
113 | } |
114 | |
115 | } // namespace LIBC_NAMESPACE_DECL |
116 | |