| 1 | //===-- Implementation header for asinf16 -----------------------*- 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_ASINF16_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_ASINF16_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/PolyEval.h" |
| 19 | #include "src/__support/FPUtil/cast.h" |
| 20 | #include "src/__support/FPUtil/multiply_add.h" |
| 21 | #include "src/__support/FPUtil/sqrt.h" |
| 22 | #include "src/__support/macros/optimization.h" |
| 23 | |
| 24 | namespace LIBC_NAMESPACE_DECL { |
| 25 | |
| 26 | namespace math { |
| 27 | |
| 28 | LIBC_INLINE constexpr float16 asinf16(float16 x) { |
| 29 | |
| 30 | // Generated by Sollya using the following command: |
| 31 | // > round(pi/2, D, RN); |
| 32 | constexpr float PI_2 = 0x1.921fb54442d18p0f; |
| 33 | |
| 34 | using FPBits = fputil::FPBits<float16>; |
| 35 | FPBits xbits(x); |
| 36 | |
| 37 | uint16_t x_u = xbits.uintval(); |
| 38 | uint16_t x_abs = x_u & 0x7fff; |
| 39 | float xf = x; |
| 40 | |
| 41 | // |x| > 0x1p0, |x| > 1, or x is NaN. |
| 42 | if (LIBC_UNLIKELY(x_abs > 0x3c00)) { |
| 43 | // asinf16(NaN) = NaN |
| 44 | if (xbits.is_nan()) { |
| 45 | if (xbits.is_signaling_nan()) { |
| 46 | fputil::raise_except_if_required(FE_INVALID); |
| 47 | return FPBits::quiet_nan().get_val(); |
| 48 | } |
| 49 | |
| 50 | return x; |
| 51 | } |
| 52 | |
| 53 | // 1 < |x| <= +/-inf |
| 54 | fputil::raise_except_if_required(FE_INVALID); |
| 55 | fputil::set_errno_if_required(EDOM); |
| 56 | |
| 57 | return FPBits::quiet_nan().get_val(); |
| 58 | } |
| 59 | |
| 60 | float xsq = xf * xf; |
| 61 | |
| 62 | // |x| <= 0x1p-1, |x| <= 0.5 |
| 63 | if (x_abs <= 0x3800) { |
| 64 | // asinf16(+/-0) = +/-0 |
| 65 | if (LIBC_UNLIKELY(x_abs == 0)) |
| 66 | return x; |
| 67 | |
| 68 | // Exhaustive tests show that, |
| 69 | // for |x| <= 0x1.878p-9, when: |
| 70 | // x > 0, and rounding upward, or |
| 71 | // x < 0, and rounding downward, then, |
| 72 | // asin(x) = x * 2^-11 + x |
| 73 | // else, in other rounding modes, |
| 74 | // asin(x) = x |
| 75 | if (LIBC_UNLIKELY(x_abs <= 0x1a1e)) { |
| 76 | #ifndef LIBC_MATH_HAS_ASSUME_ROUND_NEAREST_ONLY |
| 77 | int rounding = fputil::quick_get_round(); |
| 78 | |
| 79 | if ((xbits.is_pos() && rounding == FE_UPWARD) || |
| 80 | (xbits.is_neg() && rounding == FE_DOWNWARD)) |
| 81 | return fputil::cast<float16>(x: fputil::multiply_add(x: xf, y: 0x1.0p-11f, z: xf)); |
| 82 | #endif |
| 83 | return x; |
| 84 | } |
| 85 | |
| 86 | // Degree-6 minimax odd polynomial of asin(x) generated by Sollya with: |
| 87 | // > P = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8|], [|SG...|], [0, 0.5]); |
| 88 | float result = |
| 89 | fputil::polyeval(x: xsq, a0: 0x1.000002p0f, a: 0x1.554c2ap-3f, a: 0x1.3541ccp-4f, |
| 90 | a: 0x1.43b2d6p-5f, a: 0x1.a0d73ep-5f); |
| 91 | return fputil::cast<float16>(x: xf * result); |
| 92 | } |
| 93 | |
| 94 | // When |x| > 0.5, assume that 0.5 < |x| <= 1, |
| 95 | // |
| 96 | // Step-by-step range-reduction proof: |
| 97 | // 1: Let y = asin(x), such that, x = sin(y) |
| 98 | // 2: From complimentary angle identity: |
| 99 | // x = sin(y) = cos(pi/2 - y) |
| 100 | // 3: Let z = pi/2 - y, such that x = cos(z) |
| 101 | // 4: From double angle formula; cos(2A) = 1 - sin^2(A): |
| 102 | // z = 2A, z/2 = A |
| 103 | // cos(z) = 1 - 2 * sin^2(z/2) |
| 104 | // 5: Make sin(z/2) subject of the formula: |
| 105 | // sin(z/2) = sqrt((1 - cos(z))/2) |
| 106 | // 6: Recall [3]; x = cos(z). Therefore: |
| 107 | // sin(z/2) = sqrt((1 - x)/2) |
| 108 | // 7: Let u = (1 - x)/2 |
| 109 | // 8: Therefore: |
| 110 | // asin(sqrt(u)) = z/2 |
| 111 | // 2 * asin(sqrt(u)) = z |
| 112 | // 9: Recall [3], z = pi/2 - y. Therefore: |
| 113 | // y = pi/2 - z |
| 114 | // y = pi/2 - 2 * asin(sqrt(u)) |
| 115 | // 10: Recall [1], y = asin(x). Therefore: |
| 116 | // asin(x) = pi/2 - 2 * asin(sqrt(u)) |
| 117 | // |
| 118 | // WHY? |
| 119 | // 11: Recall [7], u = (1 - x)/2 |
| 120 | // 12: Since 0.5 < x <= 1, therefore: |
| 121 | // 0 <= u <= 0.25 and 0 <= sqrt(u) <= 0.5 |
| 122 | // |
| 123 | // Hence, we can reuse the same [0, 0.5] domain polynomial approximation for |
| 124 | // Step [10] as `sqrt(u)` is in range. |
| 125 | |
| 126 | // 0x1p-1 < |x| <= 0x1p0, 0.5 < |x| <= 1.0 |
| 127 | float xf_abs = (xf < 0 ? -xf : xf); |
| 128 | float sign = (xbits.uintval() >> 15 == 1 ? -1.0 : 1.0); |
| 129 | float u = fputil::multiply_add(x: -0.5f, y: xf_abs, z: 0.5f); |
| 130 | float u_sqrt = fputil::sqrt<float>(x: u); |
| 131 | |
| 132 | // Degree-6 minimax odd polynomial of asin(x) generated by Sollya with: |
| 133 | // > P = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8|], [|SG...|], [0, 0.5]); |
| 134 | float asin_sqrt_u = |
| 135 | u_sqrt * fputil::polyeval(x: u, a0: 0x1.000002p0f, a: 0x1.554c2ap-3f, |
| 136 | a: 0x1.3541ccp-4f, a: 0x1.43b2d6p-5f, a: 0x1.a0d73ep-5f); |
| 137 | |
| 138 | return fputil::cast<float16>(x: sign * |
| 139 | fputil::multiply_add(x: -2.0f, y: asin_sqrt_u, z: PI_2)); |
| 140 | } |
| 141 | |
| 142 | } // namespace math |
| 143 | |
| 144 | } // namespace LIBC_NAMESPACE_DECL |
| 145 | |
| 146 | #endif // LIBC_TYPES_HAS_FLOAT16 |
| 147 | |
| 148 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_ASINF16_H |
| 149 | |