1 | //===-- Half-precision cospif 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/cospif16.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/multiply_add.h" |
17 | #include "src/__support/macros/optimization.h" |
18 | |
19 | namespace LIBC_NAMESPACE_DECL { |
20 | |
21 | LLVM_LIBC_FUNCTION(float16, cospif16, (float16 x)) { |
22 | using FPBits = typename fputil::FPBits<float16>; |
23 | FPBits xbits(x); |
24 | |
25 | uint16_t x_u = xbits.uintval(); |
26 | uint16_t x_abs = x_u & 0x7fff; |
27 | float xf = x; |
28 | |
29 | // Range reduction: |
30 | // For |x| > 1/32, we perform range reduction as follows: |
31 | // Find k and y such that: |
32 | // x = (k + y) * 1/32 |
33 | // k is an integer |
34 | // |y| < 0.5 |
35 | // |
36 | // This is done by performing: |
37 | // k = round(x * 32) |
38 | // y = x * 32 - k |
39 | // |
40 | // Once k and y are computed, we then deduce the answer by the cosine of sum |
41 | // formula: |
42 | // cos(x * pi) = cos((k + y) * pi/32) |
43 | // = cos(k * pi/32) * cos(y * pi/32) + |
44 | // sin(y * pi/32) * sin(k * pi/32) |
45 | |
46 | // For signed zeros |
47 | if (LIBC_UNLIKELY(x_abs == 0U)) |
48 | return fputil::cast<float16>(1.0f); |
49 | |
50 | // Numbers greater or equal to 2^10 are integers, or infinity, or NaN |
51 | if (LIBC_UNLIKELY(x_abs >= 0x6400)) { |
52 | if (LIBC_UNLIKELY(x_abs <= 0x67FF)) |
53 | return fputil::cast<float16>((x_abs & 0x1) ? -1.0f : 1.0f); |
54 | |
55 | // Check for NaN or infintiy values |
56 | if (LIBC_UNLIKELY(x_abs >= 0x7c00)) { |
57 | if (xbits.is_signaling_nan()) { |
58 | fputil::raise_except_if_required(FE_INVALID); |
59 | return FPBits::quiet_nan().get_val(); |
60 | } |
61 | // If value is equal to infinity |
62 | if (x_abs == 0x7c00) { |
63 | fputil::set_errno_if_required(EDOM); |
64 | fputil::raise_except_if_required(FE_INVALID); |
65 | } |
66 | |
67 | return x + FPBits::quiet_nan().get_val(); |
68 | } |
69 | |
70 | return fputil::cast<float16>(1.0f); |
71 | } |
72 | |
73 | float sin_k, cos_k, sin_y, cosm1_y; |
74 | sincospif16_eval(xf, sin_k, cos_k, sin_y, cosm1_y); |
75 | |
76 | if (LIBC_UNLIKELY(sin_y == 0 && cos_k == 0)) |
77 | return fputil::cast<float16>(0.0f); |
78 | |
79 | // Since, cosm1_y = cos_y - 1, therefore: |
80 | // cos(x * pi) = cos_k(cosm1_y) + cos_k - sin_k * sin_y |
81 | return fputil::cast<float16>(fputil::multiply_add( |
82 | cos_k, cosm1_y, fputil::multiply_add(-sin_k, sin_y, cos_k))); |
83 | } |
84 | |
85 | } // namespace LIBC_NAMESPACE_DECL |
86 | |