| 1 | //===-- Implementation header for cosf --------------------------*- C++ -*-===// |
| 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 | #ifndef LLVM_LIBC_SRC___SUPPORT_MATH_COSF_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_COSF_H |
| 11 | |
| 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/macros/config.h" |
| 17 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
| 18 | #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA |
| 19 | |
| 20 | #if defined(LIBC_MATH_HAS_SKIP_ACCURATE_PASS) && \ |
| 21 | defined(LIBC_MATH_HAS_INTERMEDIATE_COMP_IN_FLOAT) && \ |
| 22 | defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT) |
| 23 | |
| 24 | #include "sincosf_float_eval.h" |
| 25 | |
| 26 | namespace LIBC_NAMESPACE_DECL { |
| 27 | namespace math { |
| 28 | |
| 29 | LIBC_INLINE float cosf(float x) { |
| 30 | return sincosf_float_eval::sincosf_eval</*IS_SIN*/ false>(x); |
| 31 | } |
| 32 | |
| 33 | } // namespace math |
| 34 | } // namespace LIBC_NAMESPACE_DECL |
| 35 | |
| 36 | #else // !LIBC_MATH_HAS_INTERMEDIATE_COMP_IN_FLOAT |
| 37 | |
| 38 | #include "sincosf_utils.h" |
| 39 | |
| 40 | namespace LIBC_NAMESPACE_DECL { |
| 41 | |
| 42 | namespace math { |
| 43 | |
| 44 | LIBC_INLINE float cosf(float x) { |
| 45 | using namespace sincosf_utils_internal; |
| 46 | |
| 47 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 48 | // Exceptional cases for cosf. |
| 49 | constexpr size_t N_EXCEPTS = 6; |
| 50 | |
| 51 | constexpr fputil::ExceptValues<float, N_EXCEPTS> COSF_EXCEPTS{.values: { |
| 52 | // (inputs, RZ output, RU offset, RD offset, RN offset) |
| 53 | // x = 0x1.64a032p43, cos(x) = 0x1.9d4ba4p-1 (RZ) |
| 54 | {.input: 0x55325019, .rnd_towardzero_result: 0x3f4ea5d2, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 0}, |
| 55 | // x = 0x1.4555p51, cos(x) = 0x1.115d7cp-1 (RZ) |
| 56 | {.input: 0x5922aa80, .rnd_towardzero_result: 0x3f08aebe, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
| 57 | // x = 0x1.48a858p54, cos(x) = 0x1.f48148p-2 (RZ) |
| 58 | {.input: 0x5aa4542c, .rnd_towardzero_result: 0x3efa40a4, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 0}, |
| 59 | // x = 0x1.3170fp63, cos(x) = 0x1.fe2976p-1 (RZ) |
| 60 | {.input: 0x5f18b878, .rnd_towardzero_result: 0x3f7f14bb, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 0}, |
| 61 | // x = 0x1.2b9622p67, cos(x) = 0x1.f0285cp-1 (RZ) |
| 62 | {.input: 0x6115cb11, .rnd_towardzero_result: 0x3f78142e, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 1}, |
| 63 | // x = 0x1.ddebdep120, cos(x) = 0x1.114438p-1 (RZ) |
| 64 | {.input: 0x7beef5ef, .rnd_towardzero_result: 0x3f08a21c, .rnd_upward_offset: 1, .rnd_downward_offset: 0, .rnd_tonearest_offset: 0}, |
| 65 | }}; |
| 66 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 67 | |
| 68 | using FPBits = typename fputil::FPBits<float>; |
| 69 | |
| 70 | FPBits xbits(x); |
| 71 | xbits.set_sign(Sign::POS); |
| 72 | |
| 73 | uint32_t x_abs = xbits.uintval(); |
| 74 | |
| 75 | // Range reduction: |
| 76 | // For |x| > pi/16, we perform range reduction as follows: |
| 77 | // Find k and y such that: |
| 78 | // x = (k + y) * pi/32 |
| 79 | // k is an integer |
| 80 | // |y| < 0.5 |
| 81 | // For small range (|x| < 2^45 when FMA instructions are available, 2^22 |
| 82 | // otherwise), this is done by performing: |
| 83 | // k = round(x * 32/pi) |
| 84 | // y = x * 32/pi - k |
| 85 | // For large range, we will omit all the higher parts of 16/pi such that the |
| 86 | // least significant bits of their full products with x are larger than 63, |
| 87 | // since cos((k + y + 64*i) * pi/32) = cos(x + i * 2pi) = cos(x). |
| 88 | // |
| 89 | // When FMA instructions are not available, we store the digits of 32/pi in |
| 90 | // chunks of 28-bit precision. This will make sure that the products: |
| 91 | // x * THIRTYTWO_OVER_PI_28[i] are all exact. |
| 92 | // When FMA instructions are available, we simply store the digits of 32/pi in |
| 93 | // chunks of doubles (53-bit of precision). |
| 94 | // So when multiplying by the largest values of single precision, the |
| 95 | // resulting output should be correct up to 2^(-208 + 128) ~ 2^-80. By the |
| 96 | // worst-case analysis of range reduction, |y| >= 2^-38, so this should give |
| 97 | // us more than 40 bits of accuracy. For the worst-case estimation of range |
| 98 | // reduction, see for instances: |
| 99 | // Elementary Functions by J-M. Muller, Chapter 11, |
| 100 | // Handbook of Floating-Point Arithmetic by J-M. Muller et. al., |
| 101 | // Chapter 10.2. |
| 102 | // |
| 103 | // Once k and y are computed, we then deduce the answer by the cosine of sum |
| 104 | // formula: |
| 105 | // cos(x) = cos((k + y)*pi/32) |
| 106 | // = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 107 | // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed |
| 108 | // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are |
| 109 | // computed using degree-7 and degree-6 minimax polynomials generated by |
| 110 | // Sollya respectively. |
| 111 | |
| 112 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 113 | // |x| < 0x1.0p-12f |
| 114 | if (LIBC_UNLIKELY(x_abs < 0x3980'0000U)) { |
| 115 | // When |x| < 2^-12, the relative error of the approximation cos(x) ~ 1 |
| 116 | // is: |
| 117 | // |cos(x) - 1| < |x^2 / 2| = 2^-25 < epsilon(1)/2. |
| 118 | // So the correctly rounded values of cos(x) are: |
| 119 | // = 1 - eps(x) if rounding mode = FE_TOWARDZERO or FE_DOWWARD, |
| 120 | // = 1 otherwise. |
| 121 | // To simplify the rounding decision and make it more efficient and to |
| 122 | // prevent compiler to perform constant folding, we use |
| 123 | // fma(x, -2^-25, 1) instead. |
| 124 | // Note: to use the formula 1 - 2^-25*x to decide the correct rounding, we |
| 125 | // do need fma(x, -2^-25, 1) to prevent underflow caused by -2^-25*x when |
| 126 | // |x| < 2^-125. For targets without FMA instructions, we simply use |
| 127 | // double for intermediate results as it is more efficient than using an |
| 128 | // emulated version of FMA. |
| 129 | #if defined(LIBC_TARGET_CPU_HAS_FMA_FLOAT) |
| 130 | return fputil::multiply_add(xbits.get_val(), -0x1.0p-25f, 1.0f); |
| 131 | #else // !LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 132 | double xd = static_cast<double>(xbits.get_val()); |
| 133 | return static_cast<float>(fputil::multiply_add(x: xd, y: -0x1.0p-25, z: 1.0)); |
| 134 | #endif // LIBC_TARGET_CPU_HAS_FMA_FLOAT |
| 135 | } |
| 136 | |
| 137 | if (auto r = COSF_EXCEPTS.lookup(x_bits: x_abs); LIBC_UNLIKELY(r.has_value())) |
| 138 | return r.value(); |
| 139 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 140 | |
| 141 | // x is inf or nan. |
| 142 | if (LIBC_UNLIKELY(x_abs >= 0x7f80'0000U)) { |
| 143 | if (xbits.is_signaling_nan()) { |
| 144 | fputil::raise_except_if_required(FE_INVALID); |
| 145 | return FPBits::quiet_nan().get_val(); |
| 146 | } |
| 147 | |
| 148 | if (x_abs == 0x7f80'0000U) { |
| 149 | fputil::set_errno_if_required(EDOM); |
| 150 | fputil::raise_except_if_required(FE_INVALID); |
| 151 | } |
| 152 | return x + FPBits::quiet_nan().get_val(); |
| 153 | } |
| 154 | |
| 155 | double xd = static_cast<double>(xbits.get_val()); |
| 156 | // Combine the results with the sine of sum formula: |
| 157 | // cos(x) = cos((k + y)*pi/32) |
| 158 | // = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 159 | // = cosm1_y * cos_k + sin_y * sin_k |
| 160 | // = (cosm1_y * cos_k + cos_k) + sin_y * sin_k |
| 161 | double sin_k = 0, cos_k = 0, sin_y = 0, cosm1_y = 0; |
| 162 | |
| 163 | sincosf_eval(xd, x_abs, sin_k, cos_k, sin_y, cosm1_y); |
| 164 | |
| 165 | return static_cast<float>(fputil::multiply_add( |
| 166 | x: sin_y, y: -sin_k, z: fputil::multiply_add(x: cosm1_y, y: cos_k, z: cos_k))); |
| 167 | } |
| 168 | |
| 169 | } // namespace math |
| 170 | |
| 171 | } // namespace LIBC_NAMESPACE_DECL |
| 172 | |
| 173 | #endif // LIBC_MATH_HAS_INTERMEDIATE_COMP_IN_FLOAT |
| 174 | |
| 175 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_COSF_H |
| 176 | |