xref: /aosp_15_r20/external/llvm-libc/src/math/generic/cosf.cpp (revision 71db0c75aadcf003ffe3238005f61d7618a3fead)
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 // Exceptional cases for cosf.
24 static constexpr size_t N_EXCEPTS = 6;
25 
26 static constexpr fputil::ExceptValues<float, N_EXCEPTS> COSF_EXCEPTS{{
27     // (inputs, RZ output, RU offset, RD offset, RN offset)
28     // x = 0x1.64a032p43, cos(x) = 0x1.9d4ba4p-1 (RZ)
29     {0x55325019, 0x3f4ea5d2, 1, 0, 0},
30     // x = 0x1.4555p51, cos(x) = 0x1.115d7cp-1 (RZ)
31     {0x5922aa80, 0x3f08aebe, 1, 0, 1},
32     // x = 0x1.48a858p54, cos(x) = 0x1.f48148p-2 (RZ)
33     {0x5aa4542c, 0x3efa40a4, 1, 0, 0},
34     // x = 0x1.3170fp63, cos(x) = 0x1.fe2976p-1 (RZ)
35     {0x5f18b878, 0x3f7f14bb, 1, 0, 0},
36     // x = 0x1.2b9622p67, cos(x) = 0x1.f0285cp-1 (RZ)
37     {0x6115cb11, 0x3f78142e, 1, 0, 1},
38     // x = 0x1.ddebdep120, cos(x) = 0x1.114438p-1 (RZ)
39     {0x7beef5ef, 0x3f08a21c, 1, 0, 0},
40 }};
41 
42 LLVM_LIBC_FUNCTION(float, cosf, (float x)) {
43   using FPBits = typename fputil::FPBits<float>;
44 
45   FPBits xbits(x);
46   xbits.set_sign(Sign::POS);
47 
48   uint32_t x_abs = xbits.uintval();
49   double xd = static_cast<double>(xbits.get_val());
50 
51   // Range reduction:
52   // For |x| > pi/16, we perform range reduction as follows:
53   // Find k and y such that:
54   //   x = (k + y) * pi/32
55   //   k is an integer
56   //   |y| < 0.5
57   // For small range (|x| < 2^45 when FMA instructions are available, 2^22
58   // otherwise), this is done by performing:
59   //   k = round(x * 32/pi)
60   //   y = x * 32/pi - k
61   // For large range, we will omit all the higher parts of 16/pi such that the
62   // least significant bits of their full products with x are larger than 63,
63   // since cos((k + y + 64*i) * pi/32) = cos(x + i * 2pi) = cos(x).
64   //
65   // When FMA instructions are not available, we store the digits of 32/pi in
66   // chunks of 28-bit precision.  This will make sure that the products:
67   //   x * THIRTYTWO_OVER_PI_28[i] are all exact.
68   // When FMA instructions are available, we simply store the digits of 32/pi in
69   // chunks of doubles (53-bit of precision).
70   // So when multiplying by the largest values of single precision, the
71   // resulting output should be correct up to 2^(-208 + 128) ~ 2^-80.  By the
72   // worst-case analysis of range reduction, |y| >= 2^-38, so this should give
73   // us more than 40 bits of accuracy. For the worst-case estimation of range
74   // reduction, see for instances:
75   //   Elementary Functions by J-M. Muller, Chapter 11,
76   //   Handbook of Floating-Point Arithmetic by J-M. Muller et. al.,
77   //   Chapter 10.2.
78   //
79   // Once k and y are computed, we then deduce the answer by the cosine of sum
80   // formula:
81   //   cos(x) = cos((k + y)*pi/32)
82   //          = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32)
83   // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed
84   // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are
85   // computed using degree-7 and degree-6 minimax polynomials generated by
86   // Sollya respectively.
87 
88   // |x| < 0x1.0p-12f
89   if (LIBC_UNLIKELY(x_abs < 0x3980'0000U)) {
90     // When |x| < 2^-12, the relative error of the approximation cos(x) ~ 1
91     // is:
92     //   |cos(x) - 1| < |x^2 / 2| = 2^-25 < epsilon(1)/2.
93     // So the correctly rounded values of cos(x) are:
94     //   = 1 - eps(x) if rounding mode = FE_TOWARDZERO or FE_DOWWARD,
95     //   = 1 otherwise.
96     // To simplify the rounding decision and make it more efficient and to
97     // prevent compiler to perform constant folding, we use
98     //   fma(x, -2^-25, 1) instead.
99     // Note: to use the formula 1 - 2^-25*x to decide the correct rounding, we
100     // do need fma(x, -2^-25, 1) to prevent underflow caused by -2^-25*x when
101     // |x| < 2^-125. For targets without FMA instructions, we simply use
102     // double for intermediate results as it is more efficient than using an
103     // emulated version of FMA.
104 #if defined(LIBC_TARGET_CPU_HAS_FMA)
105     return fputil::multiply_add(xbits.get_val(), -0x1.0p-25f, 1.0f);
106 #else
107     return static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, 1.0));
108 #endif // LIBC_TARGET_CPU_HAS_FMA
109   }
110 
111   if (auto r = COSF_EXCEPTS.lookup(x_abs); LIBC_UNLIKELY(r.has_value()))
112     return r.value();
113 
114   // x is inf or nan.
115   if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) {
116     if (x_abs == 0x7f80'0000U) {
117       fputil::set_errno_if_required(EDOM);
118       fputil::raise_except_if_required(FE_INVALID);
119     }
120     return x + FPBits::quiet_nan().get_val();
121   }
122 
123   // Combine the results with the sine of sum formula:
124   //   cos(x) = cos((k + y)*pi/32)
125   //          = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32)
126   //          = cosm1_y * cos_k + sin_y * sin_k
127   //          = (cosm1_y * cos_k + cos_k) + sin_y * sin_k
128   double sin_k, cos_k, sin_y, cosm1_y;
129 
130   sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y);
131 
132   return static_cast<float>(fputil::multiply_add(
133       sin_y, -sin_k, fputil::multiply_add(cosm1_y, cos_k, cos_k)));
134 }
135 
136 } // namespace LIBC_NAMESPACE_DECL
137