xref: /llvm-project/libc/src/math/generic/tanf16.cpp (revision ecf4f95c4f55eea0830659654fa264189773a423)
1 //===-- Half-precision tan(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/tanf16.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 constexpr size_t N_EXCEPTS = 9;
23 
24 constexpr fputil::ExceptValues<float16, N_EXCEPTS> TANF16_EXCEPTS{{
25     // (input, RZ output, RU offset, RD offset, RN offset)
26     {0x2894, 0x2894, 1, 0, 1},
27     {0x3091, 0x3099, 1, 0, 0},
28     {0x3098, 0x30a0, 1, 0, 0},
29     {0x55ed, 0x3911, 1, 0, 0},
30     {0x607b, 0xc638, 0, 1, 1},
31     {0x674e, 0x3b7d, 1, 0, 0},
32     {0x6807, 0x4014, 1, 0, 1},
33     {0x6f4d, 0xbe19, 0, 1, 1},
34     {0x7330, 0xcb62, 0, 1, 0},
35 }};
36 
37 LLVM_LIBC_FUNCTION(float16, tanf16, (float16 x)) {
38   using FPBits = fputil::FPBits<float16>;
39   FPBits xbits(x);
40 
41   uint16_t x_u = xbits.uintval();
42   uint16_t x_abs = x_u & 0x7fff;
43   bool x_sign = x_u >> 15;
44   float xf = x;
45 
46   // Handle exceptional values
47   if (auto r = TANF16_EXCEPTS.lookup_odd(x_abs, x_sign);
48       LIBC_UNLIKELY(r.has_value()))
49     return r.value();
50 
51   // |x| <= 0x1.d1p-5
52   if (LIBC_UNLIKELY(x_abs <= 0x2b44)) {
53     // |x| <= 0x1.398p-11
54     if (LIBC_UNLIKELY(x_abs <= 0x10e6)) {
55       // tan(+/-0) = +/-0
56       if (LIBC_UNLIKELY(x_abs == 0))
57         return x;
58 
59       int rounding = fputil::quick_get_round();
60 
61       // Exhaustive tests show that, when:
62       // x > 0, and rounding upward or
63       // x < 0, and rounding downward then,
64       // tan(x) = x * 2^-11 + x
65       if ((xbits.is_pos() && rounding == FE_UPWARD) ||
66           (xbits.is_neg() && rounding == FE_DOWNWARD))
67         return fputil::cast<float16>(fputil::multiply_add(xf, 0x1.0p-11f, xf));
68       return x;
69     }
70 
71     float xsq = xf * xf;
72 
73     // Degree-6 minimax odd polynomial of tan(x) generated by Sollya with:
74     // > P = fpminimax(tan(x)/x, [|0, 2, 4, 6|], [|1, SG...|], [0, pi/32]);
75     float result = fputil::polyeval(xsq, 0x1p0f, 0x1.555556p-2f, 0x1.110ee4p-3f,
76                                     0x1.be80f6p-5f);
77 
78     return fputil::cast<float16>(xf * result);
79   }
80 
81   // tan(+/-inf) = NaN, and tan(NaN) = NaN
82   if (LIBC_UNLIKELY(x_abs >= 0x7c00)) {
83     // x = +/-inf
84     if (x_abs == 0x7c00) {
85       fputil::set_errno_if_required(EDOM);
86       fputil::raise_except_if_required(FE_INVALID);
87     }
88 
89     return x + FPBits::quiet_nan().get_val();
90   }
91 
92   // Range reduction:
93   // For |x| > pi/32, we perform range reduction as follows:
94   // Find k and y such that:
95   //   x = (k + y) * pi/32;
96   //   k is an integer, |y| < 0.5
97   //
98   // This is done by performing:
99   //   k = round(x * 32/pi)
100   //   y = x * 32/pi - k
101   //
102   // Once k and y are computed, we then deduce the answer by the formula:
103   // tan(x) = sin(x) / cos(x)
104   // 	    = (sin_y * cos_k + cos_y * sin_k) / (cos_y * cos_k - sin_y * sin_k)
105   float sin_k, cos_k, sin_y, cosm1_y;
106   sincosf16_eval(xf, sin_k, cos_k, sin_y, cosm1_y);
107 
108   // Note that, cosm1_y = cos_y - 1:
109   using fputil::multiply_add;
110   return fputil::cast<float16>(
111       multiply_add(sin_y, cos_k, multiply_add(cosm1_y, sin_k, sin_k)) /
112       multiply_add(sin_y, -sin_k, multiply_add(cosm1_y, cos_k, cos_k)));
113 }
114 
115 } // namespace LIBC_NAMESPACE_DECL
116