1 | //===-- Single-precision sin 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/sinf.h" |
10 | #include "sincosf_utils.h" |
11 | #include "src/__support/FPUtil/BasicOperations.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/multiply_add.h" |
16 | #include "src/__support/FPUtil/rounding_mode.h" |
17 | #include "src/__support/common.h" |
18 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
19 | #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA |
20 | |
21 | #include <errno.h> |
22 | |
23 | #if defined(LIBC_TARGET_CPU_HAS_FMA) |
24 | #include "range_reduction_fma.h" |
25 | #else |
26 | #include "range_reduction.h" |
27 | #endif |
28 | |
29 | namespace LIBC_NAMESPACE { |
30 | |
31 | LLVM_LIBC_FUNCTION(float, sinf, (float x)) { |
32 | using FPBits = typename fputil::FPBits<float>; |
33 | FPBits xbits(x); |
34 | |
35 | uint32_t x_u = xbits.uintval(); |
36 | uint32_t x_abs = x_u & 0x7fff'ffffU; |
37 | double xd = static_cast<double>(x); |
38 | |
39 | // Range reduction: |
40 | // For |x| > pi/32, we perform range reduction as follows: |
41 | // Find k and y such that: |
42 | // x = (k + y) * pi/32 |
43 | // k is an integer |
44 | // |y| < 0.5 |
45 | // For small range (|x| < 2^45 when FMA instructions are available, 2^22 |
46 | // otherwise), this is done by performing: |
47 | // k = round(x * 32/pi) |
48 | // y = x * 32/pi - k |
49 | // For large range, we will omit all the higher parts of 32/pi such that the |
50 | // least significant bits of their full products with x are larger than 63, |
51 | // since sin((k + y + 64*i) * pi/32) = sin(x + i * 2pi) = sin(x). |
52 | // |
53 | // When FMA instructions are not available, we store the digits of 32/pi in |
54 | // chunks of 28-bit precision. This will make sure that the products: |
55 | // x * THIRTYTWO_OVER_PI_28[i] are all exact. |
56 | // When FMA instructions are available, we simply store the digits of 32/pi in |
57 | // chunks of doubles (53-bit of precision). |
58 | // So when multiplying by the largest values of single precision, the |
59 | // resulting output should be correct up to 2^(-208 + 128) ~ 2^-80. By the |
60 | // worst-case analysis of range reduction, |y| >= 2^-38, so this should give |
61 | // us more than 40 bits of accuracy. For the worst-case estimation of range |
62 | // reduction, see for instances: |
63 | // Elementary Functions by J-M. Muller, Chapter 11, |
64 | // Handbook of Floating-Point Arithmetic by J-M. Muller et. al., |
65 | // Chapter 10.2. |
66 | // |
67 | // Once k and y are computed, we then deduce the answer by the sine of sum |
68 | // formula: |
69 | // sin(x) = sin((k + y)*pi/32) |
70 | // = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32) |
71 | // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..31 are precomputed |
72 | // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are |
73 | // computed using degree-7 and degree-6 minimax polynomials generated by |
74 | // Sollya respectively. |
75 | |
76 | // |x| <= pi/16 |
77 | if (LIBC_UNLIKELY(x_abs <= 0x3e49'0fdbU)) { |
78 | |
79 | // |x| < 0x1.d12ed2p-12f |
80 | if (LIBC_UNLIKELY(x_abs < 0x39e8'9769U)) { |
81 | if (LIBC_UNLIKELY(x_abs == 0U)) { |
82 | // For signed zeros. |
83 | return x; |
84 | } |
85 | // When |x| < 2^-12, the relative error of the approximation sin(x) ~ x |
86 | // is: |
87 | // |sin(x) - x| / |sin(x)| < |x^3| / (6|x|) |
88 | // = x^2 / 6 |
89 | // < 2^-25 |
90 | // < epsilon(1)/2. |
91 | // So the correctly rounded values of sin(x) are: |
92 | // = x - sign(x)*eps(x) if rounding mode = FE_TOWARDZERO, |
93 | // or (rounding mode = FE_UPWARD and x is |
94 | // negative), |
95 | // = x otherwise. |
96 | // To simplify the rounding decision and make it more efficient, we use |
97 | // fma(x, -2^-25, x) instead. |
98 | // An exhaustive test shows that this formula work correctly for all |
99 | // rounding modes up to |x| < 0x1.c555dep-11f. |
100 | // Note: to use the formula x - 2^-25*x to decide the correct rounding, we |
101 | // do need fma(x, -2^-25, x) to prevent underflow caused by -2^-25*x when |
102 | // |x| < 2^-125. For targets without FMA instructions, we simply use |
103 | // double for intermediate results as it is more efficient than using an |
104 | // emulated version of FMA. |
105 | #if defined(LIBC_TARGET_CPU_HAS_FMA) |
106 | return fputil::multiply_add(x, y: -0x1.0p-25f, z: x); |
107 | #else |
108 | return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, xd)); |
109 | #endif // LIBC_TARGET_CPU_HAS_FMA |
110 | } |
111 | |
112 | // |x| < pi/16. |
113 | double xsq = xd * xd; |
114 | |
115 | // Degree-9 polynomial approximation: |
116 | // sin(x) ~ x + a_3 x^3 + a_5 x^5 + a_7 x^7 + a_9 x^9 |
117 | // = x (1 + a_3 x^2 + ... + a_9 x^8) |
118 | // = x * P(x^2) |
119 | // generated by Sollya with the following commands: |
120 | // > display = hexadecimal; |
121 | // > Q = fpminimax(sin(x)/x, [|0, 2, 4, 6, 8|], [|1, D...|], [0, pi/16]); |
122 | double result = |
123 | fputil::polyeval(x: xsq, a0: 1.0, a: -0x1.55555555554c6p-3, a: 0x1.1111111085e65p-7, |
124 | a: -0x1.a019f70fb4d4fp-13, a: 0x1.718d179815e74p-19); |
125 | return static_cast<float>(xd * result); |
126 | } |
127 | |
128 | if (LIBC_UNLIKELY(x_abs == 0x4619'9998U)) { // x = 0x1.33333p13 |
129 | float r = -0x1.63f4bap-2f; |
130 | int rounding = fputil::quick_get_round(); |
131 | if ((rounding == FE_DOWNWARD && xbits.is_pos()) || |
132 | (rounding == FE_UPWARD && xbits.is_neg())) |
133 | r = -0x1.63f4bcp-2f; |
134 | return xbits.is_neg() ? -r : r; |
135 | } |
136 | |
137 | if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) { |
138 | if (x_abs == 0x7f80'0000U) { |
139 | fputil::set_errno_if_required(EDOM); |
140 | fputil::raise_except_if_required(FE_INVALID); |
141 | } |
142 | return x + FPBits::quiet_nan().get_val(); |
143 | } |
144 | |
145 | // Combine the results with the sine of sum formula: |
146 | // sin(x) = sin((k + y)*pi/32) |
147 | // = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32) |
148 | // = sin_y * cos_k + (1 + cosm1_y) * sin_k |
149 | // = sin_y * cos_k + (cosm1_y * sin_k + sin_k) |
150 | double sin_k, cos_k, sin_y, cosm1_y; |
151 | |
152 | sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y); |
153 | |
154 | return static_cast<float>(fputil::multiply_add( |
155 | x: sin_y, y: cos_k, z: fputil::multiply_add(x: cosm1_y, y: sin_k, z: sin_k))); |
156 | } |
157 | |
158 | } // namespace LIBC_NAMESPACE |
159 | |