1 | //===-- Single-precision cos 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/cosf.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/except_value_utils.h" |
15 | #include "src/__support/FPUtil/multiply_add.h" |
16 | #include "src/__support/common.h" |
17 | #include "src/__support/macros/config.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 | namespace LIBC_NAMESPACE_DECL { |
22 | |
23 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
24 | // Exceptional cases for cosf. |
25 | static constexpr size_t N_EXCEPTS = 6; |
26 | |
27 | static constexpr fputil::ExceptValues<float, N_EXCEPTS> COSF_EXCEPTS{{ |
28 | // (inputs, RZ output, RU offset, RD offset, RN offset) |
29 | // x = 0x1.64a032p43, cos(x) = 0x1.9d4ba4p-1 (RZ) |
30 | {0x55325019, 0x3f4ea5d2, 1, 0, 0}, |
31 | // x = 0x1.4555p51, cos(x) = 0x1.115d7cp-1 (RZ) |
32 | {0x5922aa80, 0x3f08aebe, 1, 0, 1}, |
33 | // x = 0x1.48a858p54, cos(x) = 0x1.f48148p-2 (RZ) |
34 | {0x5aa4542c, 0x3efa40a4, 1, 0, 0}, |
35 | // x = 0x1.3170fp63, cos(x) = 0x1.fe2976p-1 (RZ) |
36 | {0x5f18b878, 0x3f7f14bb, 1, 0, 0}, |
37 | // x = 0x1.2b9622p67, cos(x) = 0x1.f0285cp-1 (RZ) |
38 | {0x6115cb11, 0x3f78142e, 1, 0, 1}, |
39 | // x = 0x1.ddebdep120, cos(x) = 0x1.114438p-1 (RZ) |
40 | {0x7beef5ef, 0x3f08a21c, 1, 0, 0}, |
41 | }}; |
42 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
43 | |
44 | LLVM_LIBC_FUNCTION(float, cosf, (float x)) { |
45 | using FPBits = typename fputil::FPBits<float>; |
46 | |
47 | FPBits xbits(x); |
48 | xbits.set_sign(Sign::POS); |
49 | |
50 | uint32_t x_abs = xbits.uintval(); |
51 | double xd = static_cast<double>(xbits.get_val()); |
52 | |
53 | // Range reduction: |
54 | // For |x| > pi/16, we perform range reduction as follows: |
55 | // Find k and y such that: |
56 | // x = (k + y) * pi/32 |
57 | // k is an integer |
58 | // |y| < 0.5 |
59 | // For small range (|x| < 2^45 when FMA instructions are available, 2^22 |
60 | // otherwise), this is done by performing: |
61 | // k = round(x * 32/pi) |
62 | // y = x * 32/pi - k |
63 | // For large range, we will omit all the higher parts of 16/pi such that the |
64 | // least significant bits of their full products with x are larger than 63, |
65 | // since cos((k + y + 64*i) * pi/32) = cos(x + i * 2pi) = cos(x). |
66 | // |
67 | // When FMA instructions are not available, we store the digits of 32/pi in |
68 | // chunks of 28-bit precision. This will make sure that the products: |
69 | // x * THIRTYTWO_OVER_PI_28[i] are all exact. |
70 | // When FMA instructions are available, we simply store the digits of 32/pi in |
71 | // chunks of doubles (53-bit of precision). |
72 | // So when multiplying by the largest values of single precision, the |
73 | // resulting output should be correct up to 2^(-208 + 128) ~ 2^-80. By the |
74 | // worst-case analysis of range reduction, |y| >= 2^-38, so this should give |
75 | // us more than 40 bits of accuracy. For the worst-case estimation of range |
76 | // reduction, see for instances: |
77 | // Elementary Functions by J-M. Muller, Chapter 11, |
78 | // Handbook of Floating-Point Arithmetic by J-M. Muller et. al., |
79 | // Chapter 10.2. |
80 | // |
81 | // Once k and y are computed, we then deduce the answer by the cosine of sum |
82 | // formula: |
83 | // cos(x) = cos((k + y)*pi/32) |
84 | // = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
85 | // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed |
86 | // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are |
87 | // computed using degree-7 and degree-6 minimax polynomials generated by |
88 | // Sollya respectively. |
89 | |
90 | // |x| < 0x1.0p-12f |
91 | if (LIBC_UNLIKELY(x_abs < 0x3980'0000U)) { |
92 | // When |x| < 2^-12, the relative error of the approximation cos(x) ~ 1 |
93 | // is: |
94 | // |cos(x) - 1| < |x^2 / 2| = 2^-25 < epsilon(1)/2. |
95 | // So the correctly rounded values of cos(x) are: |
96 | // = 1 - eps(x) if rounding mode = FE_TOWARDZERO or FE_DOWWARD, |
97 | // = 1 otherwise. |
98 | // To simplify the rounding decision and make it more efficient and to |
99 | // prevent compiler to perform constant folding, we use |
100 | // fma(x, -2^-25, 1) instead. |
101 | // Note: to use the formula 1 - 2^-25*x to decide the correct rounding, we |
102 | // do need fma(x, -2^-25, 1) to prevent underflow caused by -2^-25*x when |
103 | // |x| < 2^-125. For targets without FMA instructions, we simply use |
104 | // double for intermediate results as it is more efficient than using an |
105 | // emulated version of FMA. |
106 | #if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT) |
107 | return fputil::multiply_add(xbits.get_val(), -0x1.0p-25f, 1.0f); |
108 | #else |
109 | return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, 1.0)); |
110 | #endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT |
111 | } |
112 | |
113 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
114 | if (auto r = COSF_EXCEPTS.lookup(x_abs); LIBC_UNLIKELY(r.has_value())) |
115 | return r.value(); |
116 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
117 | |
118 | // x is inf or nan. |
119 | if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) { |
120 | if (xbits.is_signaling_nan()) { |
121 | fputil::raise_except_if_required(FE_INVALID); |
122 | return FPBits::quiet_nan().get_val(); |
123 | } |
124 | |
125 | if (x_abs == 0x7f80'0000U) { |
126 | fputil::set_errno_if_required(EDOM); |
127 | fputil::raise_except_if_required(FE_INVALID); |
128 | } |
129 | return x + FPBits::quiet_nan().get_val(); |
130 | } |
131 | |
132 | // Combine the results with the sine of sum formula: |
133 | // cos(x) = cos((k + y)*pi/32) |
134 | // = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
135 | // = cosm1_y * cos_k + sin_y * sin_k |
136 | // = (cosm1_y * cos_k + cos_k) + sin_y * sin_k |
137 | double sin_k, cos_k, sin_y, cosm1_y; |
138 | |
139 | sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y); |
140 | |
141 | return static_cast<float>(fputil::multiply_add( |
142 | sin_y, -sin_k, fputil::multiply_add(cosm1_y, cos_k, cos_k))); |
143 | } |
144 | |
145 | } // namespace LIBC_NAMESPACE_DECL |
146 | |