| 1 | //===-- Implementation header for sin ---------------------------*- 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_SIN_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_SIN_H |
| 11 | |
| 12 | #include "range_reduction_double_common.h" |
| 13 | #include "sincos_eval.h" |
| 14 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 15 | #include "src/__support/FPUtil/FPBits.h" |
| 16 | #include "src/__support/FPUtil/double_double.h" |
| 17 | #include "src/__support/FPUtil/dyadic_float.h" |
| 18 | #include "src/__support/macros/config.h" |
| 19 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
| 20 | #include "src/__support/macros/properties/cpu_features.h" // LIBC_TARGET_CPU_HAS_FMA |
| 21 | |
| 22 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 23 | #include "range_reduction_double_fma.h" |
| 24 | #else |
| 25 | #include "range_reduction_double_nofma.h" |
| 26 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 27 | |
| 28 | namespace LIBC_NAMESPACE_DECL { |
| 29 | |
| 30 | namespace math { |
| 31 | |
| 32 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 33 | LIBC_INLINE double |
| 34 | sin_accurate(double x, uint16_t x_e, unsigned k, |
| 35 | const range_reduction_double_internal::LargeRangeReduction |
| 36 | &range_reduction_large) { |
| 37 | using namespace math::range_reduction_double_internal; |
| 38 | using FPBits = typename fputil::FPBits<double>; |
| 39 | |
| 40 | DFloat128 u_f128, sin_u, cos_u; |
| 41 | if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT)) |
| 42 | u_f128 = range_reduction_small_f128(x); |
| 43 | else |
| 44 | u_f128 = range_reduction_large.accurate(); |
| 45 | |
| 46 | math::sincos_eval_internal::sincos_eval(u: u_f128, sin_u, cos_u); |
| 47 | |
| 48 | auto get_sin_k = [](unsigned kk) -> DFloat128 { |
| 49 | unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63); |
| 50 | DFloat128 ans = SIN_K_PI_OVER_128_F128[idx]; |
| 51 | if (kk & 128) |
| 52 | ans.sign = Sign::NEG; |
| 53 | return ans; |
| 54 | }; |
| 55 | |
| 56 | // cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128). |
| 57 | DFloat128 sin_k_f128 = get_sin_k(k); |
| 58 | DFloat128 cos_k_f128 = get_sin_k(k + 64); |
| 59 | |
| 60 | // sin(x) = sin(k * pi/128 + u) |
| 61 | // = sin(u) * cos(k*pi/128) + cos(u) * sin(k*pi/128) |
| 62 | DFloat128 r = fputil::quick_add(a: fputil::quick_mul(a: sin_k_f128, b: cos_u), |
| 63 | b: fputil::quick_mul(a: cos_k_f128, b: sin_u)); |
| 64 | |
| 65 | // TODO: Add assertion if Ziv's accuracy tests fail in debug mode. |
| 66 | // https://github.com/llvm/llvm-project/issues/96452. |
| 67 | |
| 68 | return static_cast<double>(r); |
| 69 | } |
| 70 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 71 | |
| 72 | LIBC_INLINE double sin(double x) { |
| 73 | using namespace math::range_reduction_double_internal; |
| 74 | using FPBits = typename fputil::FPBits<double>; |
| 75 | FPBits xbits(x); |
| 76 | |
| 77 | uint16_t x_e = xbits.get_biased_exponent(); |
| 78 | |
| 79 | DoubleDouble y; |
| 80 | unsigned k = 0; |
| 81 | LargeRangeReduction range_reduction_large{}; |
| 82 | |
| 83 | // |x| < 2^16 |
| 84 | if (LIBC_LIKELY(x_e < FPBits::EXP_BIAS + FAST_PASS_EXPONENT)) { |
| 85 | // |x| < 2^-4 |
| 86 | if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 4)) { |
| 87 | // |x| < 2^-26, |sin(x) - x| < ulp(x)/2. |
| 88 | if (LIBC_UNLIKELY(x_e < FPBits::EXP_BIAS - 26)) { |
| 89 | // Signed zeros. |
| 90 | if (LIBC_UNLIKELY(x == 0.0)) |
| 91 | return x + x; // Make sure it works with FTZ/DAZ. |
| 92 | |
| 93 | #ifdef LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 94 | return fputil::multiply_add(x, -0x1.0p-54, x); |
| 95 | #else |
| 96 | if (LIBC_UNLIKELY(x_e < 4)) { |
| 97 | #ifndef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 98 | int rounding_mode = fputil::quick_get_round(); |
| 99 | if (rounding_mode == FE_TOWARDZERO || |
| 100 | (xbits.sign() == Sign::POS && rounding_mode == FE_DOWNWARD) || |
| 101 | (xbits.sign() == Sign::NEG && rounding_mode == FE_UPWARD)) |
| 102 | return FPBits(xbits.uintval() - 1).get_val(); |
| 103 | #endif // !LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 104 | } |
| 105 | return fputil::multiply_add(x, y: -0x1.0p-54, z: x); |
| 106 | #endif // LIBC_TARGET_CPU_HAS_FMA_DOUBLE |
| 107 | } |
| 108 | // No range reduction needed. |
| 109 | |
| 110 | // Use degree-9 polynomial approximation: |
| 111 | // sin(x) ~ x + a1 * x^3 + a2 * x^5 + a3 * x^7 + a4 * x^9 |
| 112 | // ~ x + x^3 * Q(x^2). |
| 113 | // > P = fpminimax(sin(x)/x, [|0, 2, 4, 6, 8|], [|1, D...|], [0, 2^-4]); |
| 114 | // > dirtyinfnorm((sin(x) - x*P)/sin(x), [-2^-4, 2^-4]); |
| 115 | // 0x1.3c2e...p-69 |
| 116 | // > P; |
| 117 | constexpr double COEFFS[] = {-0x1.5555555555555p-3, 0x1.111111110f491p-7, |
| 118 | -0x1.a01a00e16af3ep-13, |
| 119 | 0x1.71c24233f1bafp-19}; |
| 120 | double x_sq = x * x; |
| 121 | double c0 = fputil::multiply_add(x: x_sq, y: COEFFS[1], z: COEFFS[0]); |
| 122 | double c1 = fputil::multiply_add(x: x_sq, y: COEFFS[3], z: COEFFS[2]); |
| 123 | double x4 = x_sq * x_sq; |
| 124 | double x3 = x * x_sq; |
| 125 | double r_lo = fputil::multiply_add(x: x4, y: c1, z: c0) * x3; |
| 126 | |
| 127 | #ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 128 | return x + r_lo; |
| 129 | #else |
| 130 | // Overall errors <= 2 * ulp(x^3/6) + |x| * 2^-68. |
| 131 | double err = fputil::multiply_add(x: x_sq, y: 0x1.0p-53, z: 0x1.0p-68); |
| 132 | double r_lo_u = fputil::multiply_add(x, y: err, z: r_lo); |
| 133 | double r_lo_l = fputil::multiply_add(x: -x, y: err, z: r_lo); |
| 134 | double r_upper = x + r_lo_u; |
| 135 | double r_lower = x + r_lo_l; |
| 136 | |
| 137 | if (LIBC_LIKELY(r_upper == r_lower)) |
| 138 | return r_upper; |
| 139 | |
| 140 | k = range_reduction_small(x, u&: y); |
| 141 | return sin_accurate(x, x_e, k, range_reduction_large); |
| 142 | #endif // LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 143 | } else { |
| 144 | // Small range reduction. |
| 145 | k = range_reduction_small(x, u&: y); |
| 146 | } |
| 147 | } else { |
| 148 | // Inf or NaN |
| 149 | if (LIBC_UNLIKELY(x_e > 2 * FPBits::EXP_BIAS)) { |
| 150 | // sin(+-Inf) = NaN |
| 151 | if (xbits.is_signaling_nan()) { |
| 152 | fputil::raise_except_if_required(FE_INVALID); |
| 153 | return FPBits::quiet_nan().get_val(); |
| 154 | } |
| 155 | |
| 156 | if (xbits.get_mantissa() == 0) { |
| 157 | fputil::set_errno_if_required(EDOM); |
| 158 | fputil::raise_except_if_required(FE_INVALID); |
| 159 | } |
| 160 | return x + FPBits::quiet_nan().get_val(); |
| 161 | } |
| 162 | |
| 163 | // Large range reduction. |
| 164 | k = range_reduction_large.fast(x, u&: y); |
| 165 | } |
| 166 | |
| 167 | DoubleDouble sin_y, cos_y; |
| 168 | |
| 169 | [[maybe_unused]] double err = |
| 170 | math::sincos_eval_internal::sincos_eval(u: y, sin_u&: sin_y, cos_u&: cos_y); |
| 171 | |
| 172 | // Look up sin(k * pi/128) and cos(k * pi/128) |
| 173 | #ifdef LIBC_MATH_HAS_SMALL_TABLES |
| 174 | // Memory saving versions. Use 65-entry table. |
| 175 | auto get_idx_dd = [](unsigned kk) -> DoubleDouble { |
| 176 | unsigned idx = (kk & 64) ? 64 - (kk & 63) : (kk & 63); |
| 177 | DoubleDouble ans = SIN_K_PI_OVER_128[idx]; |
| 178 | if (kk & 128) { |
| 179 | ans.hi = -ans.hi; |
| 180 | ans.lo = -ans.lo; |
| 181 | } |
| 182 | return ans; |
| 183 | }; |
| 184 | DoubleDouble sin_k = get_idx_dd(k); |
| 185 | DoubleDouble cos_k = get_idx_dd(k + 64); |
| 186 | #else |
| 187 | // Fast look up version, but needs 256-entry table. |
| 188 | // cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128). |
| 189 | DoubleDouble sin_k = SIN_K_PI_OVER_128[k & 255]; |
| 190 | DoubleDouble cos_k = SIN_K_PI_OVER_128[(k + 64) & 255]; |
| 191 | #endif |
| 192 | |
| 193 | // After range reduction, k = round(x * 128 / pi) and y = x - k * (pi / 128). |
| 194 | // So k is an integer and -pi / 256 <= y <= pi / 256. |
| 195 | // Then sin(x) = sin((k * pi/128 + y) |
| 196 | // = sin(y) * cos(k*pi/128) + cos(y) * sin(k*pi/128) |
| 197 | DoubleDouble sin_k_cos_y = fputil::quick_mult(a: cos_y, b: sin_k); |
| 198 | DoubleDouble cos_k_sin_y = fputil::quick_mult(a: sin_y, b: cos_k); |
| 199 | // When k != 0 mod 128, |
| 200 | // |sin( k * pi/128 )| > pi/128 - epsilon > |y| >= |sin(y)|, |
| 201 | // and cos(y) > 1 - pi/128. So we can use Fast2Sum for the addition: |
| 202 | // sin(y) * cos(k*pi/128) + cos(y) * sin(k*pi/128). |
| 203 | DoubleDouble rr = fputil::exact_add(a: sin_k_cos_y.hi, b: cos_k_sin_y.hi); |
| 204 | rr.lo += sin_k_cos_y.lo + cos_k_sin_y.lo; |
| 205 | |
| 206 | #ifdef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 207 | return rr.hi + rr.lo; |
| 208 | #else |
| 209 | // Accurate test and pass for correctly rounded implementation. |
| 210 | |
| 211 | double rlp = rr.lo + err; |
| 212 | double rlm = rr.lo - err; |
| 213 | |
| 214 | double r_upper = rr.hi + rlp; // (rr.lo + ERR); |
| 215 | double r_lower = rr.hi + rlm; // (rr.lo - ERR); |
| 216 | |
| 217 | // Ziv's rounding test. |
| 218 | if (LIBC_LIKELY(r_upper == r_lower)) |
| 219 | return r_upper; |
| 220 | |
| 221 | return sin_accurate(x, x_e, k, range_reduction_large); |
| 222 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
| 223 | } |
| 224 | |
| 225 | } // namespace math |
| 226 | |
| 227 | } // namespace LIBC_NAMESPACE_DECL |
| 228 | |
| 229 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_SIN_H |
| 230 | |