| 1 | //===-- Common utils for exp10f ---------------------------------*- 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_EXP10F_UTILS_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_EXP10F_UTILS_H |
| 11 | |
| 12 | #include "src/__support/FPUtil/FPBits.h" |
| 13 | #include "src/__support/FPUtil/PolyEval.h" |
| 14 | #include "src/__support/FPUtil/nearest_integer.h" |
| 15 | #include "src/__support/macros/config.h" |
| 16 | |
| 17 | namespace LIBC_NAMESPACE_DECL { |
| 18 | |
| 19 | struct ExpBase { |
| 20 | // Base = e |
| 21 | static constexpr int MID_BITS = 5; |
| 22 | static constexpr int MID_MASK = (1 << MID_BITS) - 1; |
| 23 | // log2(e) * 2^5 |
| 24 | static constexpr double LOG2_B = 0x1.71547652b82fep+0 * (1 << MID_BITS); |
| 25 | // High and low parts of -log(2) * 2^(-5) |
| 26 | static constexpr double M_LOGB_2_HI = -0x1.62e42fefa0000p-1 / (1 << MID_BITS); |
| 27 | static constexpr double M_LOGB_2_LO = |
| 28 | -0x1.cf79abc9e3b3ap-40 / (1 << MID_BITS); |
| 29 | // Look up table for bit fields of 2^(i/32) for i = 0..31, generated by Sollya |
| 30 | // with: |
| 31 | // > for i from 0 to 31 do printdouble(round(2^(i/32), D, RN)); |
| 32 | static constexpr int64_t EXP_2_MID[1 << MID_BITS] = { |
| 33 | 0x3ff0000000000000, 0x3ff059b0d3158574, 0x3ff0b5586cf9890f, |
| 34 | 0x3ff11301d0125b51, 0x3ff172b83c7d517b, 0x3ff1d4873168b9aa, |
| 35 | 0x3ff2387a6e756238, 0x3ff29e9df51fdee1, 0x3ff306fe0a31b715, |
| 36 | 0x3ff371a7373aa9cb, 0x3ff3dea64c123422, 0x3ff44e086061892d, |
| 37 | 0x3ff4bfdad5362a27, 0x3ff5342b569d4f82, 0x3ff5ab07dd485429, |
| 38 | 0x3ff6247eb03a5585, 0x3ff6a09e667f3bcd, 0x3ff71f75e8ec5f74, |
| 39 | 0x3ff7a11473eb0187, 0x3ff82589994cce13, 0x3ff8ace5422aa0db, |
| 40 | 0x3ff93737b0cdc5e5, 0x3ff9c49182a3f090, 0x3ffa5503b23e255d, |
| 41 | 0x3ffae89f995ad3ad, 0x3ffb7f76f2fb5e47, 0x3ffc199bdd85529c, |
| 42 | 0x3ffcb720dcef9069, 0x3ffd5818dcfba487, 0x3ffdfc97337b9b5f, |
| 43 | 0x3ffea4afa2a490da, 0x3fff50765b6e4540, |
| 44 | }; |
| 45 | |
| 46 | // Approximating e^dx with degree-5 minimax polynomial generated by Sollya: |
| 47 | // > Q = fpminimax(expm1(x)/x, 4, [|1, D...|], [-log(2)/64, log(2)/64]); |
| 48 | // Then: |
| 49 | // e^dx ~ P(dx) = 1 + dx + COEFFS[0] * dx^2 + ... + COEFFS[3] * dx^5. |
| 50 | static constexpr double COEFFS[4] = { |
| 51 | 0x1.ffffffffe5bc8p-2, 0x1.555555555cd67p-3, 0x1.5555c2a9b48b4p-5, |
| 52 | 0x1.11112a0e34bdbp-7}; |
| 53 | |
| 54 | LIBC_INLINE static double powb_lo(double dx) { |
| 55 | using fputil::multiply_add; |
| 56 | double dx2 = dx * dx; |
| 57 | double c0 = 1.0 + dx; |
| 58 | // c1 = COEFFS[0] + COEFFS[1] * dx |
| 59 | double c1 = multiply_add(x: dx, y: ExpBase::COEFFS[1], z: ExpBase::COEFFS[0]); |
| 60 | // c2 = COEFFS[2] + COEFFS[3] * dx |
| 61 | double c2 = multiply_add(x: dx, y: ExpBase::COEFFS[3], z: ExpBase::COEFFS[2]); |
| 62 | // r = c4 + c5 * dx^4 |
| 63 | // = 1 + dx + COEFFS[0] * dx^2 + ... + COEFFS[5] * dx^7 |
| 64 | return fputil::polyeval(x: dx2, a0: c0, a: c1, a: c2); |
| 65 | } |
| 66 | }; |
| 67 | |
| 68 | struct Exp10Base : public ExpBase { |
| 69 | // log2(10) * 2^5 |
| 70 | static constexpr double LOG2_B = 0x1.a934f0979a371p1 * (1 << MID_BITS); |
| 71 | // High and low parts of -log10(2) * 2^(-5). |
| 72 | // Notice that since |x * log2(10)| < 150: |
| 73 | // |k| = |round(x * log2(10) * 2^5)| < 2^8 * 2^5 = 2^13 |
| 74 | // So when the FMA instructions are not available, in order for the product |
| 75 | // k * M_LOGB_2_HI |
| 76 | // to be exact, we only store the high part of log10(2) up to 38 bits |
| 77 | // (= 53 - 15) of precision. |
| 78 | // It is generated by Sollya with: |
| 79 | // > round(log10(2), 44, RN); |
| 80 | static constexpr double M_LOGB_2_HI = -0x1.34413509f8p-2 / (1 << MID_BITS); |
| 81 | // > round(log10(2) - 0x1.34413509f8p-2, D, RN); |
| 82 | static constexpr double M_LOGB_2_LO = 0x1.80433b83b532ap-44 / (1 << MID_BITS); |
| 83 | |
| 84 | // Approximating 10^dx with degree-5 minimax polynomial generated by Sollya: |
| 85 | // > Q = fpminimax((10^x - 1)/x, 4, [|D...|], [-log10(2)/2^6, log10(2)/2^6]); |
| 86 | // Then: |
| 87 | // 10^dx ~ P(dx) = 1 + COEFFS[0] * dx + ... + COEFFS[4] * dx^5. |
| 88 | static constexpr double COEFFS[5] = {0x1.26bb1bbb55515p1, 0x1.53524c73bd3eap1, |
| 89 | 0x1.0470591dff149p1, 0x1.2bd7c0a9fbc4dp0, |
| 90 | 0x1.1429e74a98f43p-1}; |
| 91 | |
| 92 | LIBC_INLINE static double powb_lo(double dx) { |
| 93 | using fputil::multiply_add; |
| 94 | double dx2 = dx * dx; |
| 95 | // c0 = 1 + COEFFS[0] * dx |
| 96 | double c0 = multiply_add(x: dx, y: Exp10Base::COEFFS[0], z: 1.0); |
| 97 | // c1 = COEFFS[1] + COEFFS[2] * dx |
| 98 | double c1 = multiply_add(x: dx, y: Exp10Base::COEFFS[2], z: Exp10Base::COEFFS[1]); |
| 99 | // c2 = COEFFS[3] + COEFFS[4] * dx |
| 100 | double c2 = multiply_add(x: dx, y: Exp10Base::COEFFS[4], z: Exp10Base::COEFFS[3]); |
| 101 | // r = c0 + dx^2 * (c1 + c2 * dx^2) |
| 102 | // = c0 + c1 * dx^2 + c2 * dx^4 |
| 103 | // = 1 + COEFFS[0] * dx + ... + COEFFS[4] * dx^5. |
| 104 | return fputil::polyeval(x: dx2, a0: c0, a: c1, a: c2); |
| 105 | } |
| 106 | }; |
| 107 | |
| 108 | // Output of range reduction for exp_b: (2^(mid + hi), lo) |
| 109 | // where: |
| 110 | // b^x = 2^(mid + hi) * b^lo |
| 111 | struct exp_b_reduc_t { |
| 112 | double mh; // 2^(mid + hi) |
| 113 | double lo; |
| 114 | }; |
| 115 | |
| 116 | // The function correctly calculates b^x value with at least float precision |
| 117 | // in a limited range. |
| 118 | // Range reduction: |
| 119 | // b^x = 2^(hi + mid) * b^lo |
| 120 | // where: |
| 121 | // x = (hi + mid) * log_b(2) + lo |
| 122 | // hi is an integer, |
| 123 | // 0 <= mid * 2^MID_BITS < 2^MID_BITS is an integer |
| 124 | // -2^(-MID_BITS - 1) <= lo * log2(b) <= 2^(-MID_BITS - 1) |
| 125 | // Base class needs to provide the following constants: |
| 126 | // - MID_BITS : number of bits after decimal points used for mid |
| 127 | // - MID_MASK : 2^MID_BITS - 1, mask to extract mid bits |
| 128 | // - LOG2_B : log2(b) * 2^MID_BITS for scaling |
| 129 | // - M_LOGB_2_HI : high part of -log_b(2) * 2^(-MID_BITS) |
| 130 | // - M_LOGB_2_LO : low part of -log_b(2) * 2^(-MID_BITS) |
| 131 | // - EXP_2_MID : look up table for bit fields of 2^mid |
| 132 | // Return: |
| 133 | // { 2^(hi + mid), lo } |
| 134 | template <class Base> |
| 135 | LIBC_INLINE constexpr exp_b_reduc_t exp_b_range_reduc(float x) { |
| 136 | double xd = static_cast<double>(x); |
| 137 | // kd = round((hi + mid) * log2(b) * 2^MID_BITS) |
| 138 | double kd = fputil::nearest_integer(Base::LOG2_B * xd); |
| 139 | // k = round((hi + mid) * log2(b) * 2^MID_BITS) |
| 140 | int k = static_cast<int>(kd); |
| 141 | // hi = floor(kd * 2^(-MID_BITS)) |
| 142 | // exp_hi = shift hi to the exponent field of double precision. |
| 143 | uint64_t exp_hi = static_cast<uint64_t>(k >> Base::MID_BITS) |
| 144 | << fputil::FPBits<double>::FRACTION_LEN; |
| 145 | // mh = 2^hi * 2^mid |
| 146 | // mh_bits = bit field of mh |
| 147 | uint64_t mh_bits = Base::EXP_2_MID[k & Base::MID_MASK] + exp_hi; |
| 148 | double mh = fputil::FPBits<double>(mh_bits).get_val(); |
| 149 | // dx = lo = x - (hi + mid) * log(2) |
| 150 | double dx = fputil::multiply_add( |
| 151 | kd, Base::M_LOGB_2_LO, fputil::multiply_add(kd, Base::M_LOGB_2_HI, xd)); |
| 152 | return {.mh: mh, .lo: dx}; |
| 153 | } |
| 154 | |
| 155 | } // namespace LIBC_NAMESPACE_DECL |
| 156 | |
| 157 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_EXP10F_UTILS_H |
| 158 | |