1 | //===-- Single-precision sincos 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/sincosf.h" |
10 | #include "sincosf_utils.h" |
11 | #include "src/__support/FPUtil/FEnvImpl.h" |
12 | #include "src/__support/FPUtil/FPBits.h" |
13 | #include "src/__support/FPUtil/multiply_add.h" |
14 | #include "src/__support/FPUtil/rounding_mode.h" |
15 | #include "src/__support/common.h" |
16 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
17 | #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA |
18 | |
19 | #include <errno.h> |
20 | |
21 | namespace LIBC_NAMESPACE { |
22 | |
23 | // Exceptional values |
24 | static constexpr int N_EXCEPTS = 6; |
25 | |
26 | static constexpr uint32_t EXCEPT_INPUTS[N_EXCEPTS] = { |
27 | 0x46199998, // x = 0x1.33333p13 x |
28 | 0x55325019, // x = 0x1.64a032p43 x |
29 | 0x5922aa80, // x = 0x1.4555p51 x |
30 | 0x5f18b878, // x = 0x1.3170fp63 x |
31 | 0x6115cb11, // x = 0x1.2b9622p67 x |
32 | 0x7beef5ef, // x = 0x1.ddebdep120 x |
33 | }; |
34 | |
35 | static constexpr uint32_t EXCEPT_OUTPUTS_SIN[N_EXCEPTS][4] = { |
36 | {0xbeb1fa5d, 0, 1, 0}, // x = 0x1.33333p13, sin(x) = -0x1.63f4bap-2 (RZ) |
37 | {0xbf171adf, 0, 1, 1}, // x = 0x1.64a032p43, sin(x) = -0x1.2e35bep-1 (RZ) |
38 | {0xbf587521, 0, 1, 1}, // x = 0x1.4555p51, sin(x) = -0x1.b0ea42p-1 (RZ) |
39 | {0x3dad60f6, 1, 0, 1}, // x = 0x1.3170fp63, sin(x) = 0x1.5ac1ecp-4 (RZ) |
40 | {0xbe7cc1e0, 0, 1, 1}, // x = 0x1.2b9622p67, sin(x) = -0x1.f983cp-3 (RZ) |
41 | {0xbf587d1b, 0, 1, 1}, // x = 0x1.ddebdep120, sin(x) = -0x1.b0fa36p-1 (RZ) |
42 | }; |
43 | |
44 | static constexpr uint32_t EXCEPT_OUTPUTS_COS[N_EXCEPTS][4] = { |
45 | {0xbf70090b, 0, 1, 0}, // x = 0x1.33333p13, cos(x) = -0x1.e01216p-1 (RZ) |
46 | {0x3f4ea5d2, 1, 0, 0}, // x = 0x1.64a032p43, cos(x) = 0x1.9d4ba4p-1 (RZ) |
47 | {0x3f08aebe, 1, 0, 1}, // x = 0x1.4555p51, cos(x) = 0x1.115d7cp-1 (RZ) |
48 | {0x3f7f14bb, 1, 0, 0}, // x = 0x1.3170fp63, cos(x) = 0x1.fe2976p-1 (RZ) |
49 | {0x3f78142e, 1, 0, 1}, // x = 0x1.2b9622p67, cos(x) = 0x1.f0285cp-1 (RZ) |
50 | {0x3f08a21c, 1, 0, 0}, // x = 0x1.ddebdep120, cos(x) = 0x1.114438p-1 (RZ) |
51 | }; |
52 | |
53 | LLVM_LIBC_FUNCTION(void, sincosf, (float x, float *sinp, float *cosp)) { |
54 | using FPBits = typename fputil::FPBits<float>; |
55 | FPBits xbits(x); |
56 | |
57 | uint32_t x_abs = xbits.uintval() & 0x7fff'ffffU; |
58 | double xd = static_cast<double>(x); |
59 | |
60 | // Range reduction: |
61 | // For |x| >= 2^-12, we perform range reduction as follows: |
62 | // Find k and y such that: |
63 | // x = (k + y) * pi/32 |
64 | // k is an integer |
65 | // |y| < 0.5 |
66 | // For small range (|x| < 2^45 when FMA instructions are available, 2^22 |
67 | // otherwise), this is done by performing: |
68 | // k = round(x * 32/pi) |
69 | // y = x * 32/pi - k |
70 | // For large range, we will omit all the higher parts of 32/pi such that the |
71 | // least significant bits of their full products with x are larger than 63, |
72 | // since: |
73 | // sin((k + y + 64*i) * pi/32) = sin(x + i * 2pi) = sin(x), and |
74 | // cos((k + y + 64*i) * pi/32) = cos(x + i * 2pi) = cos(x). |
75 | // |
76 | // When FMA instructions are not available, we store the digits of 32/pi in |
77 | // chunks of 28-bit precision. This will make sure that the products: |
78 | // x * THIRTYTWO_OVER_PI_28[i] are all exact. |
79 | // When FMA instructions are available, we simply store the digits of326/pi in |
80 | // chunks of doubles (53-bit of precision). |
81 | // So when multiplying by the largest values of single precision, the |
82 | // resulting output should be correct up to 2^(-208 + 128) ~ 2^-80. By the |
83 | // worst-case analysis of range reduction, |y| >= 2^-38, so this should give |
84 | // us more than 40 bits of accuracy. For the worst-case estimation of range |
85 | // reduction, see for instances: |
86 | // Elementary Functions by J-M. Muller, Chapter 11, |
87 | // Handbook of Floating-Point Arithmetic by J-M. Muller et. al., |
88 | // Chapter 10.2. |
89 | // |
90 | // Once k and y are computed, we then deduce the answer by the sine and cosine |
91 | // of sum formulas: |
92 | // sin(x) = sin((k + y)*pi/32) |
93 | // = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32) |
94 | // cos(x) = cos((k + y)*pi/32) |
95 | // = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
96 | // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed |
97 | // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are |
98 | // computed using degree-7 and degree-6 minimax polynomials generated by |
99 | // Sollya respectively. |
100 | |
101 | // |x| < 0x1.0p-12f |
102 | if (LIBC_UNLIKELY(x_abs < 0x3980'0000U)) { |
103 | if (LIBC_UNLIKELY(x_abs == 0U)) { |
104 | // For signed zeros. |
105 | *sinp = x; |
106 | *cosp = 1.0f; |
107 | return; |
108 | } |
109 | // When |x| < 2^-12, the relative errors of the approximations |
110 | // sin(x) ~ x, cos(x) ~ 1 |
111 | // are: |
112 | // |sin(x) - x| / |sin(x)| < |x^3| / (6|x|) |
113 | // = x^2 / 6 |
114 | // < 2^-25 |
115 | // < epsilon(1)/2. |
116 | // |cos(x) - 1| < |x^2 / 2| = 2^-25 < epsilon(1)/2. |
117 | // So the correctly rounded values of sin(x) and cos(x) are: |
118 | // sin(x) = x - sign(x)*eps(x) if rounding mode = FE_TOWARDZERO, |
119 | // or (rounding mode = FE_UPWARD and x is |
120 | // negative), |
121 | // = x otherwise. |
122 | // cos(x) = 1 - eps(x) if rounding mode = FE_TOWARDZERO or FE_DOWWARD, |
123 | // = 1 otherwise. |
124 | // To simplify the rounding decision and make it more efficient and to |
125 | // prevent compiler to perform constant folding, we use |
126 | // sin(x) = fma(x, -2^-25, x), |
127 | // cos(x) = fma(x*0.5f, -x, 1) |
128 | // instead. |
129 | // Note: to use the formula x - 2^-25*x to decide the correct rounding, we |
130 | // do need fma(x, -2^-25, x) to prevent underflow caused by -2^-25*x when |
131 | // |x| < 2^-125. For targets without FMA instructions, we simply use |
132 | // double for intermediate results as it is more efficient than using an |
133 | // emulated version of FMA. |
134 | #if defined(LIBC_TARGET_CPU_HAS_FMA) |
135 | *sinp = fputil::multiply_add(x, y: -0x1.0p-25f, z: x); |
136 | *cosp = fputil::multiply_add(x: FPBits(x_abs).get_val(), y: -0x1.0p-25f, z: 1.0f); |
137 | #else |
138 | *sinp = static_cast<float>(fputil::multiply_add(xd, -0x1.0p-25, xd)); |
139 | *cosp = static_cast<float>(fputil::multiply_add( |
140 | static_cast<double>(FPBits(x_abs).get_val()), -0x1.0p-25, 1.0)); |
141 | #endif // LIBC_TARGET_CPU_HAS_FMA |
142 | return; |
143 | } |
144 | |
145 | // x is inf or nan. |
146 | if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) { |
147 | if (x_abs == 0x7f80'0000U) { |
148 | fputil::set_errno_if_required(EDOM); |
149 | fputil::raise_except_if_required(FE_INVALID); |
150 | } |
151 | *sinp = FPBits::quiet_nan().get_val(); |
152 | *cosp = *sinp; |
153 | return; |
154 | } |
155 | |
156 | // Check exceptional values. |
157 | for (int i = 0; i < N_EXCEPTS; ++i) { |
158 | if (LIBC_UNLIKELY(x_abs == EXCEPT_INPUTS[i])) { |
159 | uint32_t s = EXCEPT_OUTPUTS_SIN[i][0]; // FE_TOWARDZERO |
160 | uint32_t c = EXCEPT_OUTPUTS_COS[i][0]; // FE_TOWARDZERO |
161 | bool x_sign = x < 0; |
162 | switch (fputil::quick_get_round()) { |
163 | case FE_UPWARD: |
164 | s += x_sign ? EXCEPT_OUTPUTS_SIN[i][2] : EXCEPT_OUTPUTS_SIN[i][1]; |
165 | c += EXCEPT_OUTPUTS_COS[i][1]; |
166 | break; |
167 | case FE_DOWNWARD: |
168 | s += x_sign ? EXCEPT_OUTPUTS_SIN[i][1] : EXCEPT_OUTPUTS_SIN[i][2]; |
169 | c += EXCEPT_OUTPUTS_COS[i][2]; |
170 | break; |
171 | case FE_TONEAREST: |
172 | s += EXCEPT_OUTPUTS_SIN[i][3]; |
173 | c += EXCEPT_OUTPUTS_COS[i][3]; |
174 | break; |
175 | } |
176 | *sinp = x_sign ? -FPBits(s).get_val() : FPBits(s).get_val(); |
177 | *cosp = FPBits(c).get_val(); |
178 | |
179 | return; |
180 | } |
181 | } |
182 | |
183 | // Combine the results with the sine and cosine of sum formulas: |
184 | // sin(x) = sin((k + y)*pi/32) |
185 | // = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32) |
186 | // = sin_y * cos_k + (1 + cosm1_y) * sin_k |
187 | // = sin_y * cos_k + (cosm1_y * sin_k + sin_k) |
188 | // cos(x) = cos((k + y)*pi/32) |
189 | // = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
190 | // = cosm1_y * cos_k + sin_y * sin_k |
191 | // = (cosm1_y * cos_k + cos_k) + sin_y * sin_k |
192 | double sin_k, cos_k, sin_y, cosm1_y; |
193 | |
194 | sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y); |
195 | |
196 | *sinp = static_cast<float>(fputil::multiply_add( |
197 | x: sin_y, y: cos_k, z: fputil::multiply_add(x: cosm1_y, y: sin_k, z: sin_k))); |
198 | *cosp = static_cast<float>(fputil::multiply_add( |
199 | x: sin_y, y: -sin_k, z: fputil::multiply_add(x: cosm1_y, y: cos_k, z: cos_k))); |
200 | } |
201 | |
202 | } // namespace LIBC_NAMESPACE |
203 | |