| 1 | //===----------------------------------------------------------------------===// |
| 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 | /// \file |
| 10 | /// General sinh/cosh functions for bfloat16. |
| 11 | /// |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_LIBC_SRC___SUPPORT_MATH_SINHCOSHBF16_UTILS_H |
| 15 | #define LLVM_LIBC_SRC___SUPPORT_MATH_SINHCOSHBF16_UTILS_H |
| 16 | |
| 17 | #include "src/__support/CPP/array.h" |
| 18 | #include "src/__support/FPUtil/FPBits.h" |
| 19 | #include "src/__support/FPUtil/multiply_add.h" |
| 20 | #include "src/__support/FPUtil/nearest_integer.h" |
| 21 | #include "src/__support/FPUtil/polyeval.h" |
| 22 | #include "src/__support/macros/attributes.h" |
| 23 | #include "src/__support/macros/config.h" |
| 24 | |
| 25 | namespace LIBC_NAMESPACE_DECL { |
| 26 | |
| 27 | namespace math { |
| 28 | |
| 29 | namespace sinhcoshbf16_internal { |
| 30 | |
| 31 | // Generated by Sollya with the following commands: |
| 32 | // > display = hexadecimal; |
| 33 | // > round(log2(exp(1)), SG, RN); |
| 34 | LIBC_INLINE_VAR constexpr float LOG2F_E = 0x1.715476p+0f; |
| 35 | |
| 36 | // Generated by Sollya with the following commands: |
| 37 | // > display = hexadecimal; |
| 38 | // > round(log(2), SG, RN); |
| 39 | LIBC_INLINE_VAR constexpr float LOGF_2 = 0x1.62e43p-1f; |
| 40 | |
| 41 | // Generated by Sollya with the following commands: |
| 42 | // > display = hexadecimal; |
| 43 | // > for i from 0 to 31 do printsingle(round(2^(i * 2^-5), SG, RN)); |
| 44 | LIBC_INLINE_VAR constexpr cpp::array<uint32_t, 32> EXP2_MID_5_BITS = { |
| 45 | 0x3f80'0000U, 0x3f82'cd87U, 0x3f85'aac3U, 0x3f88'980fU, 0x3f8b'95c2U, |
| 46 | 0x3f8e'a43aU, 0x3f91'c3d3U, 0x3f94'f4f0U, 0x3f98'37f0U, 0x3f9b'8d3aU, |
| 47 | 0x3f9e'f532U, 0x3fa2'7043U, 0x3fa5'fed7U, 0x3fa9'a15bU, 0x3fad'583fU, |
| 48 | 0x3fb1'23f6U, 0x3fb5'04f3U, 0x3fb8'fbafU, 0x3fbd'08a4U, 0x3fc1'2c4dU, |
| 49 | 0x3fc5'672aU, 0x3fc9'b9beU, 0x3fce'248cU, 0x3fd2'a81eU, 0x3fd7'44fdU, |
| 50 | 0x3fdb'fbb8U, 0x3fe0'ccdfU, 0x3fe5'b907U, 0x3fea'c0c7U, 0x3fef'e4baU, |
| 51 | 0x3ff5'257dU, 0x3ffa'83b3U, |
| 52 | }; |
| 53 | |
| 54 | // This function correctly calculates sinh(x) and cosh(x) by calculating exp(x) |
| 55 | // and exp(-x) simultaneously. |
| 56 | // To compute e^x, we perform the following range reduction: |
| 57 | // find hi, mid, lo such that: |
| 58 | // x = (hi + mid) * log(2) + lo, in which |
| 59 | // hi is an integer, |
| 60 | // 0 <= mid * 2^5 < 32 is an integer |
| 61 | // -2^(-5) <= lo * log2(e) <= 2^-5. |
| 62 | // In particular, |
| 63 | // hi + mid = round(x * log2(e) * 2^5) * 2^(-5). |
| 64 | // Then, |
| 65 | // e^x = 2^(hi + mid) * e^lo = 2^hi * 2^mid * e^lo. |
| 66 | // We store 2^mid in the lookup table EXP2_MID_5_BITS, and compute 2^hi * 2^mid |
| 67 | // by adding hi to the exponent field of 2^mid. |
| 68 | // e^lo is computed using a degree-3 minimax polynomial generated by Sollya: |
| 69 | // e^lo ~ P(lo) |
| 70 | // = 1 + lo + c2 * lo^2 + ... + c5 * lo^5 |
| 71 | // = (1 + c2*lo^2 + c4*lo^4) + lo * (1 + c3*lo^2 + c5*lo^4) |
| 72 | // = P_even + lo * P_odd |
| 73 | // To compute e^(-x), notice that: |
| 74 | // e^(-x) = 2^(-(hi + mid)) * e^(-lo) |
| 75 | // ~ 2^(-(hi + mid)) * P(-lo) |
| 76 | // = 2^(-(hi + mid)) * (P_even - lo * P_odd) |
| 77 | // So: |
| 78 | // sinh(x) = (e^x - e^(-x)) / 2 |
| 79 | // ~ 0.5 * (2^(hi + mid) * (P_even + lo * P_odd) - |
| 80 | // 2^(-(hi + mid)) * (P_even - lo * P_odd)) |
| 81 | // = 0.5 * (P_even * (2^(hi + mid) - 2^(-(hi + mid))) + |
| 82 | // lo * P_odd * (2^(hi + mid) + 2^(-(hi + mid)))) |
| 83 | // And similarly: |
| 84 | // cosh(x) = (e^x + e^(-x)) / 2 |
| 85 | // ~ 0.5 * (P_even * (2^(hi + mid) + 2^(-(hi + mid))) + |
| 86 | // lo * P_odd * (2^(hi + mid) - 2^(-(hi + mid)))) |
| 87 | // The main point of these formulas is that the expensive part of calculating |
| 88 | // the polynomials approximating lower parts of e^x and e^(-x) is shared and |
| 89 | // only done once. |
| 90 | template <bool IsSinh> LIBC_INLINE float eval_sinh_or_cosh(float xf) { |
| 91 | float kf = fputil::nearest_integer(x: xf * (LOG2F_E * 0x1.0p+5f)); |
| 92 | int x_hi_mid_p = static_cast<int>(kf); |
| 93 | int x_hi_mid_m = -x_hi_mid_p; |
| 94 | |
| 95 | unsigned x_hi_p = static_cast<unsigned>(x_hi_mid_p) >> 5; |
| 96 | unsigned x_hi_m = static_cast<unsigned>(x_hi_mid_m) >> 5; |
| 97 | unsigned x_mid_p = static_cast<unsigned>(x_hi_mid_p) & 0x1f; |
| 98 | unsigned x_mid_m = static_cast<unsigned>(x_hi_mid_m) & 0x1f; |
| 99 | |
| 100 | uint32_t exp2_hi_mid_bits_p = |
| 101 | EXP2_MID_5_BITS[x_mid_p] + |
| 102 | static_cast<uint32_t>(x_hi_p << fputil::FPBits<float>::FRACTION_LEN); |
| 103 | uint32_t exp2_hi_mid_bits_m = |
| 104 | EXP2_MID_5_BITS[x_mid_m] + |
| 105 | static_cast<uint32_t>(x_hi_m << fputil::FPBits<float>::FRACTION_LEN); |
| 106 | // exp2_hi_mid_p = 2^(hi + mid) |
| 107 | float exp2_hi_mid_p = fputil::FPBits<float>(exp2_hi_mid_bits_p).get_val(); |
| 108 | // exp2_hi_mid_m = 2^(-(hi + mid)) |
| 109 | float exp2_hi_mid_m = fputil::FPBits<float>(exp2_hi_mid_bits_m).get_val(); |
| 110 | |
| 111 | // exp2_hi_mid_sum = 2^(hi + mid) + 2^(-(hi + mid)) |
| 112 | float exp2_hi_mid_sum = exp2_hi_mid_p + exp2_hi_mid_m; |
| 113 | // exp2_hi_mid_diff = 2^(hi + mid) - 2^(-(hi + mid)) |
| 114 | float exp2_hi_mid_diff = exp2_hi_mid_p - exp2_hi_mid_m; |
| 115 | |
| 116 | // lo = x - (hi + mid) = round(x * log2(e) * 2^5) * log(2) * (-2^(-5)) + x |
| 117 | float lo = fputil::sub_multiply(x: xf, y: kf, z: LOGF_2 * 0x1.0p-5f); |
| 118 | float lo_sq = lo * lo; |
| 119 | |
| 120 | // Degree-3 minimax polynomial generated by Sollya with the following |
| 121 | // commands: |
| 122 | // > display = hexadecimal; |
| 123 | // > P = fpminimax(expm1(x)/x, 2, [|SG...|], [-2^-5, 2^-5]); |
| 124 | // > 1 + x * P; |
| 125 | constexpr cpp::array<float, 4> COEFFS = {0x1p+0f, 0x1p+0f, 0x1.0004p-1f, |
| 126 | 0x1.555778p-3f}; |
| 127 | float half_p_odd = |
| 128 | fputil::polyeval(x: lo_sq, a0: COEFFS[1] * 0.5f, a: COEFFS[3] * 0.5f); |
| 129 | float half_p_even = |
| 130 | fputil::polyeval(x: lo_sq, a0: COEFFS[0] * 0.5f, a: COEFFS[2] * 0.5f); |
| 131 | |
| 132 | // sinh(x) = lo * (0.5 * P_odd * (2^(hi + mid) + 2^(-(hi + mid)))) + |
| 133 | // (0.5 * P_even * (2^(hi + mid) - 2^(-(hi + mid)))) |
| 134 | if constexpr (IsSinh) |
| 135 | return fputil::multiply_add(x: lo, y: half_p_odd * exp2_hi_mid_sum, |
| 136 | z: half_p_even * exp2_hi_mid_diff); |
| 137 | // cosh(x) = lo * (0.5 * P_odd * (2^(hi + mid) - 2^(-(hi + mid)))) + |
| 138 | // (0.5 * P_even * (2^(hi + mid) + 2^(-(hi + mid)))) |
| 139 | return fputil::multiply_add(x: lo, y: half_p_odd * exp2_hi_mid_diff, |
| 140 | z: half_p_even * exp2_hi_mid_sum); |
| 141 | } |
| 142 | |
| 143 | // Calculate e^x / 2 |
| 144 | LIBC_INLINE float exp_half(float xf) { |
| 145 | float kf = fputil::nearest_integer(x: xf * (LOG2F_E * 0x1.0p+5f)); |
| 146 | int x_hi_mid_p = static_cast<int>(kf); |
| 147 | |
| 148 | unsigned x_hi_p = (static_cast<unsigned>(x_hi_mid_p) >> 5) - 1; |
| 149 | unsigned x_mid_p = static_cast<unsigned>(x_hi_mid_p) & 0x1f; |
| 150 | |
| 151 | uint32_t exp2_hi_mid_bits_p = |
| 152 | EXP2_MID_5_BITS[x_mid_p] + |
| 153 | static_cast<uint32_t>(x_hi_p << fputil::FPBits<float>::FRACTION_LEN); |
| 154 | |
| 155 | // exp2_hi_mid_p = 2^(hi + mid - 1) |
| 156 | float exp2_hi_mid_p = fputil::FPBits<float>(exp2_hi_mid_bits_p).get_val(); |
| 157 | |
| 158 | // lo = x - (hi + mid) = round(x * log2(e) * 2^5) * log(2) * (-2^(-5)) + x |
| 159 | float lo = fputil::sub_multiply(x: xf, y: kf, z: LOGF_2 * 0x1.0p-5f); |
| 160 | float lo_sq = lo * lo; |
| 161 | |
| 162 | constexpr cpp::array<float, 4> COEFFS = {0x1p+0f, 0x1p+0f, 0x1.0004p-1f, |
| 163 | 0x1.555778p-3f}; |
| 164 | |
| 165 | float p0 = fputil::multiply_add(x: lo, y: COEFFS[1], z: COEFFS[0]); |
| 166 | float p1 = fputil::multiply_add(x: lo, y: COEFFS[3], z: COEFFS[2]); |
| 167 | |
| 168 | float exp_lo = fputil::multiply_add(x: lo_sq, y: p1, z: p0); |
| 169 | |
| 170 | // exp(x) / 2 = 2^(hi + mid - 1) * exp(lo). |
| 171 | float result = exp2_hi_mid_p * exp_lo; |
| 172 | return result; |
| 173 | } |
| 174 | |
| 175 | } // namespace sinhcoshbf16_internal |
| 176 | |
| 177 | } // namespace math |
| 178 | |
| 179 | } // namespace LIBC_NAMESPACE_DECL |
| 180 | |
| 181 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_SINHCOSHBF16_UTILS_H |
| 182 | |