| 1 | //===-- Implementation header for lgammaf16 ---------------------*- 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_LGAMMAF16_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_LGAMMAF16_H |
| 11 | |
| 12 | #include "include/llvm-libc-macros/float16-macros.h" |
| 13 | |
| 14 | #ifdef LIBC_TYPES_HAS_FLOAT16 |
| 15 | |
| 16 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 17 | #include "src/__support/FPUtil/FPBits.h" |
| 18 | #include "src/__support/FPUtil/NearestIntegerOperations.h" |
| 19 | #include "src/__support/FPUtil/PolyEval.h" |
| 20 | #include "src/__support/FPUtil/cast.h" |
| 21 | #include "src/__support/FPUtil/multiply_add.h" |
| 22 | #include "src/__support/macros/config.h" |
| 23 | #include "src/__support/macros/optimization.h" |
| 24 | #include "src/__support/macros/properties/types.h" |
| 25 | #include "src/__support/math/gamma_util.h" |
| 26 | |
| 27 | namespace LIBC_NAMESPACE_DECL { |
| 28 | |
| 29 | namespace math { |
| 30 | |
| 31 | LIBC_INLINE float16 lgammaf16(float16 x) { |
| 32 | using namespace gamma_internal; |
| 33 | using FPBits = fputil::FPBits<float16>; |
| 34 | |
| 35 | FPBits xbits(x); |
| 36 | uint16_t x_abs = xbits.abs().uintval(); |
| 37 | |
| 38 | // Handle NaN and Inf |
| 39 | if (LIBC_UNLIKELY(x_abs >= 0x7c00u)) { |
| 40 | if (x_abs == 0x7c00u) |
| 41 | return FPBits::inf().get_val(); |
| 42 | if (xbits.is_signaling_nan()) { |
| 43 | fputil::raise_except_if_required(FE_INVALID); |
| 44 | return FPBits::quiet_nan().get_val(); |
| 45 | } |
| 46 | return x; |
| 47 | } |
| 48 | |
| 49 | // +-0 pole error |
| 50 | if (LIBC_UNLIKELY(x_abs == 0)) { |
| 51 | fputil::raise_except_if_required(FE_DIVBYZERO); |
| 52 | fputil::set_errno_if_required(ERANGE); |
| 53 | return FPBits::inf().get_val(); |
| 54 | } |
| 55 | |
| 56 | if (LIBC_UNLIKELY(is_integer(x))) { |
| 57 | if (xbits.is_neg()) { |
| 58 | // pole -> +Inf |
| 59 | fputil::raise_except_if_required(FE_DIVBYZERO); |
| 60 | fputil::set_errno_if_required(ERANGE); |
| 61 | return FPBits::inf().get_val(); |
| 62 | } |
| 63 | // lgamma(1) = lgamma(2) = 0 |
| 64 | if (x_abs == 0x3c00u || x_abs == 0x4000u) |
| 65 | return FPBits::zero().get_val(); |
| 66 | } |
| 67 | |
| 68 | double xd = fputil::cast<double>(x); |
| 69 | double abs_xd = xd < 0.0 ? -xd : xd; |
| 70 | double lgamma_val; |
| 71 | |
| 72 | if (LIBC_UNLIKELY(x_abs < 0x3948u)) { |
| 73 | // Small path: |x| < 0.66015625 |
| 74 | // For t = |x|: lgamma(t) = t*P(u) - log(t), where P(u) approximates |
| 75 | // h(t) = (lgamma(t) + log(t)) / t, u = (t - MID_S) / HW_S in [-1, 1] |
| 76 | // For x < 0: lgamma(-t) = log(pi) - log(sin(pi*t)) - lgamma(t) - log(t) |
| 77 | // = log(pi) - lg_ln(lg_sinpi(t)) - t*P(u) |
| 78 | // Degree-12 Chebyshev approximation for h in u; max lgamma error ~1.5e-13 |
| 79 | // Coefficients generated by DCT at 13 Chebyshev nodes on |
| 80 | // [1e-5, 0.66015625]. |
| 81 | // Evaluated via Clenshaw: sum_{k=0}^{12} CHEB_H[k] * T_k(u) |
| 82 | constexpr double MID_S = 0x1.52014f8b588e3p-2; |
| 83 | constexpr double HW_S = 0x1.51feb074a771dp-2; |
| 84 | constexpr double CHEB_H[13] = { |
| 85 | -0x1.6ab0351d4e02fp-2, 0x1.ac55f136e49cep-3, -0x1.9fc5280cd41bfp-7, |
| 86 | 0x1.1825d37c81c1ep-10, -0x1.ae2f107c31c9ep-14, 0x1.6174bdf95e9abp-17, |
| 87 | -0x1.2e6abd22bc4ecp-20, 0x1.09b66d4f14b06p-23, -0x1.dbb938f189d8ap-27, |
| 88 | 0x1.afcb748e20e4ep-30, -0x1.8c221a9d89d8ap-33, 0x1.6e708e49b3aeep-36, |
| 89 | -0x1.50accec4ec4ecp-39}; |
| 90 | // Clenshaw's algorithm: p = sum_k CHEB_H[k] * T_k(u_s) |
| 91 | double u_s = (abs_xd - MID_S) / HW_S, u_s2 = 2.0 * u_s; |
| 92 | double bh1 = CHEB_H[12], bh2 = 0.0; |
| 93 | for (int k = 11; k >= 1; k--) { |
| 94 | double bh0 = fputil::multiply_add(x: u_s2, y: bh1, z: CHEB_H[k] - bh2); |
| 95 | bh2 = bh1; |
| 96 | bh1 = bh0; |
| 97 | } |
| 98 | double poly_h = fputil::multiply_add(x: u_s, y: bh1, z: CHEB_H[0] - bh2); |
| 99 | // abs_xd * poly_h = lgamma(abs_xd) + log(abs_xd) |
| 100 | double poly_val = abs_xd * poly_h; |
| 101 | if (xbits.is_neg()) |
| 102 | lgamma_val = 0x1.250d048e7a1bdp+0 - lg_ln(x: lg_sinpi(x: abs_xd)) - poly_val; |
| 103 | else |
| 104 | lgamma_val = poly_val - lg_ln(x: abs_xd); |
| 105 | } else if (x_abs > 0x42BEu) { // 3.3710938 |
| 106 | double log_abs_xd = lg_ln(x: abs_xd); |
| 107 | // lgamma(x) = (x-0.5)*log(x) - x + log(2*pi)/2 + Bernoulli corrections |
| 108 | // Stirling base: (x-0.5)*(log(x)-1) + (log(2*pi)-1)/2 |
| 109 | lgamma_val = (abs_xd - 0.5) * (log_abs_xd - 1.0) + 0x1.acfe390c97d69p-2; |
| 110 | |
| 111 | double inv_x = 1.0 / abs_xd, inv_x2 = inv_x * inv_x; |
| 112 | |
| 113 | if (x_abs > 0x64AEu) { |
| 114 | // |x| > 1198 1-term Stirling-Bernoulli: 1/(12*x) |
| 115 | lgamma_val += inv_x * 0x1.5555555555555p-4; |
| 116 | } else if (x_abs > 0x549Eu) { |
| 117 | // |x| in (73.875, 1198] 2-term correction |
| 118 | // Minimax polynomial in 1/x^2 for R(x) = lgamma(x) - stirling_base(x), |
| 119 | // where stirling_base(x) = (x-0.5)*log(x) - x + log(2*pi)/2 |
| 120 | // Exact Bernoulli: B_2/(1*2) = 1/12, B_4/(3*4) = -1/360 |
| 121 | // Generated by Sollya with: |
| 122 | // > stir = proc(x) { return lgamma(x)-((x-0.5)*log(x)-x+log(2*pi)/2); |
| 123 | // }; > P = fpminimax(stir(x), [|1,3|], [|D...|], [73.875, 1198]); |
| 124 | constexpr double BERN2[2] = {0x1.555555547fbadp-4, -0x1.6c0fd270c465p-9}; |
| 125 | lgamma_val += inv_x * fputil::polyeval(x: inv_x2, a0: BERN2[0], a: BERN2[1]); |
| 126 | } else if (x_abs > 0x4955u) { |
| 127 | // |x| in (10.664, 73.875] 4-term correction |
| 128 | // Minimax polynomial in 1/x^2 for R(x); exact Bernoulli terms are |
| 129 | // 1/12, -1/360, 1/1260, -1/1680 |
| 130 | // Generated by Sollya with: |
| 131 | // > display = hexadecimal; |
| 132 | // > stir = proc(x) { return lgamma(x)-((x-0.5)*log(x)-x+log(2*pi)/2); |
| 133 | // }; > P = fpminimax(stir(x), [|1,3,5,7|], [|D...|], [10.664, 73.875]); |
| 134 | constexpr double BERN4[4] = {0x1.555555554de0bp-4, -0x1.6c16bdc45944fp-9, |
| 135 | 0x1.a0077f300ecb3p-11, |
| 136 | -0x1.2e9cfff3b29c2p-11}; |
| 137 | double inv_x4 = inv_x2 * inv_x2; |
| 138 | lgamma_val += |
| 139 | inv_x * fputil::multiply_add( |
| 140 | x: inv_x4, y: fputil::multiply_add(x: inv_x2, y: BERN4[3], z: BERN4[2]), |
| 141 | z: fputil::multiply_add(x: inv_x2, y: BERN4[1], z: BERN4[0])); |
| 142 | } else { |
| 143 | // |x| in (3.373, 10.664]: 8-term correction. |
| 144 | // Degree-7 minimax polynomial in 1/x^2 for R(x), generated by Sollya |
| 145 | // with: |
| 146 | // > stir = proc(x) { return lgamma(x)-((x-0.5)*log(x)-x+log(2*pi)/2); |
| 147 | // }; > P = fpminimax(stir(x), [|1,3,5,7,9,11,13,15|], [|D...|], |
| 148 | // [3.373, 10.664]); |
| 149 | constexpr double BERN8[8] = { |
| 150 | 0x1.5555555551286p-4, -0x1.6c16c0e7c4cf4p-9, 0x1.a0193267fe6f2p-11, |
| 151 | -0x1.37e87ec19cb45p-11, 0x1.b40011dfff081p-11, -0x1.c16c8946b19b6p-10, |
| 152 | 0x1.e9f47ace150d8p-9, -0x1.4f5843a71a338p-8}; |
| 153 | double inv_x4 = inv_x2 * inv_x2, inv_x8 = inv_x4 * inv_x4; |
| 154 | double p01 = fputil::multiply_add(x: inv_x2, y: BERN8[1], z: BERN8[0]); |
| 155 | double p23 = fputil::multiply_add(x: inv_x2, y: BERN8[3], z: BERN8[2]); |
| 156 | double p45 = fputil::multiply_add(x: inv_x2, y: BERN8[5], z: BERN8[4]); |
| 157 | double p67 = fputil::multiply_add(x: inv_x2, y: BERN8[7], z: BERN8[6]); |
| 158 | double p03 = fputil::multiply_add(x: inv_x4, y: p23, z: p01); |
| 159 | double p47 = fputil::multiply_add(x: inv_x4, y: p67, z: p45); |
| 160 | lgamma_val += inv_x * fputil::multiply_add(x: inv_x8, y: p47, z: p03); |
| 161 | } |
| 162 | |
| 163 | if (xbits.is_neg()) { |
| 164 | // Reflection formula |
| 165 | // lgamma(x) = log(pi) - log|x| - log|sin(pi*frac_x)| - lgamma|x| |
| 166 | double frac_x = xd - fputil::floor(x: xd); |
| 167 | lgamma_val = 0x1.250d048e7a1bdp+0 - lgamma_val - log_abs_xd; |
| 168 | lgamma_val -= lg_ln(x: lg_sinpi(x: frac_x)); |
| 169 | } |
| 170 | } else { |
| 171 | // Medium path: |x| in [0.66015625, 3.373046875] |
| 172 | // lgamma(|x|) = (|x|-1)*(|x|-2) * P(u), where P(u) approximates |
| 173 | // g(t) = lgamma(t) / ((t-1)*(t-2)), u = (t - MID_M) / HW_M in [-1, 1] |
| 174 | // Degree-20 Chebyshev approximation for g in u; max lgamma error ~1.4e-10 |
| 175 | // Coefficients generated by DCT at 21 Chebyshev nodes on |
| 176 | // [0.66015625, 3.373046875]. Evaluated via Clenshaw: sum_{k=0}^{20} |
| 177 | // CHEB_G[k] * T_k(u) |
| 178 | constexpr double MID_M = 0x1.0220000000000p+1; |
| 179 | constexpr double HW_M = 0x1.5b40000000000p+0; |
| 180 | constexpr double CHEB_G[21] = { |
| 181 | 0x1.d690d4de8b355p-2, -0x1.53439726f7d8bp-3, 0x1.568fc6653c026p-5, |
| 182 | -0x1.8d074f73f36efp-7, 0x1.eefebf4ca8e86p-9, -0x1.4275913d85ff5p-10, |
| 183 | 0x1.b075f0d96e492p-12, -0x1.27ed35f357339p-13, 0x1.9b12cbe71a800p-15, |
| 184 | -0x1.20c949c2829b5p-16, 0x1.996eb95241555p-18, -0x1.2460d6c60f800p-19, |
| 185 | 0x1.a42297aae1862p-21, -0x1.2f6c7d28230a5p-22, 0x1.b8394aac30c31p-24, |
| 186 | -0x1.409726ce1110ep-25, 0x1.d484ac42db6dbp-27, -0x1.574dae55861abp-28, |
| 187 | 0x1.f782605555555p-30, -0x1.6cefb8f65aa2ep-31, 0x1.d9f268f3cf3cfp-33}; |
| 188 | // Clenshaw's algorithm: poly_g = sum_k CHEB_G[k] * T_k(u_m) |
| 189 | double u_m = (abs_xd - MID_M) / HW_M, u_m2 = 2.0 * u_m; |
| 190 | double bg1 = CHEB_G[20], bg2 = 0.0; |
| 191 | for (int k = 19; k >= 1; k--) { |
| 192 | double bg0 = fputil::multiply_add(x: u_m2, y: bg1, z: CHEB_G[k] - bg2); |
| 193 | bg2 = bg1; |
| 194 | bg1 = bg0; |
| 195 | } |
| 196 | double poly_g = fputil::multiply_add(x: u_m, y: bg1, z: CHEB_G[0] - bg2); |
| 197 | lgamma_val = (abs_xd - 1.0) * (abs_xd - 2.0) * poly_g; |
| 198 | |
| 199 | if (xbits.is_neg()) { |
| 200 | // Reflection |
| 201 | // (x) = log(pi) - lgamma|x| - log(|x|*|sin(pi*frac_x)|) |
| 202 | double frac_x = xd - fputil::floor(x: xd); |
| 203 | lgamma_val = 0x1.250d048e7a1bdp+0 - lgamma_val; |
| 204 | lgamma_val -= lg_ln(x: lg_sinpi(x: frac_x) * abs_xd); |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | float16 result = fputil::cast<float16>(x: lgamma_val); |
| 209 | if (LIBC_UNLIKELY(fputil::FPBits<float16>(result).is_inf())) |
| 210 | fputil::set_errno_if_required(ERANGE); |
| 211 | return result; |
| 212 | } |
| 213 | |
| 214 | } // namespace math |
| 215 | |
| 216 | } // namespace LIBC_NAMESPACE_DECL |
| 217 | |
| 218 | #endif // LIBC_TYPES_HAS_FLOAT16 |
| 219 | |
| 220 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_LGAMMAF16_H |
| 221 | |