| 1 | //===-- Implementation header for acospif16 ---------------------*- 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_ACOSPIF16_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_ACOSPIF16_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 acospif16(float16 x) { |
| 29 | using FPBits = fputil::FPBits<float16>; |
| 30 | FPBits xbits(x); |
| 31 | |
| 32 | uint16_t x_u = xbits.uintval(); |
| 33 | uint16_t x_abs = x_u & 0x7fff; |
| 34 | uint16_t x_sign = x_u >> 15; |
| 35 | |
| 36 | // |x| > 0x1p0, |x| > 1, or x is NaN. |
| 37 | if (LIBC_UNLIKELY(x_abs > 0x3c00)) { |
| 38 | // acospif16(NaN) = NaN |
| 39 | if (xbits.is_nan()) { |
| 40 | if (xbits.is_signaling_nan()) { |
| 41 | fputil::raise_except_if_required(FE_INVALID); |
| 42 | return FPBits::quiet_nan().get_val(); |
| 43 | } |
| 44 | |
| 45 | return x; |
| 46 | } |
| 47 | |
| 48 | // 1 < |x| <= +inf |
| 49 | fputil::raise_except_if_required(FE_INVALID); |
| 50 | fputil::set_errno_if_required(EDOM); |
| 51 | |
| 52 | return FPBits::quiet_nan().get_val(); |
| 53 | } |
| 54 | |
| 55 | // |x| == 0x1p0, x is 1 or -1 |
| 56 | // if x is (-)1, return 1 |
| 57 | // if x is (+)1, return 0 |
| 58 | if (LIBC_UNLIKELY(x_abs == 0x3c00)) |
| 59 | return fputil::cast<float16>(x: x_sign ? 1.0f : 0.0f); |
| 60 | |
| 61 | float xf = x; |
| 62 | float xsq = xf * xf; |
| 63 | |
| 64 | // Degree-6 minimax polynomial coefficients of asin(x) generated by Sollya |
| 65 | // with: > P = fpminimax(asin(x)/(pi * x), [|0, 2, 4, 6, 8|], [|SG...|], [0, |
| 66 | // 0.5]); |
| 67 | constexpr float POLY_COEFFS[5] = {0x1.45f308p-2f, 0x1.b2900cp-5f, |
| 68 | 0x1.897e36p-6f, 0x1.9efafcp-7f, |
| 69 | 0x1.06d884p-6f}; |
| 70 | // |x| <= 0x1p-1, |x| <= 0.5 |
| 71 | if (x_abs <= 0x3800) { |
| 72 | // if x is 0, return 0.5 |
| 73 | if (LIBC_UNLIKELY(x_abs == 0)) |
| 74 | return fputil::cast<float16>(x: 0.5f); |
| 75 | |
| 76 | // Note that: acos(x) = pi/2 + asin(-x) = pi/2 - asin(x), then |
| 77 | // acospi(x) = 0.5 - asin(x)/pi |
| 78 | float interm = |
| 79 | fputil::polyeval(x: xsq, a0: POLY_COEFFS[0], a: POLY_COEFFS[1], a: POLY_COEFFS[2], |
| 80 | a: POLY_COEFFS[3], a: POLY_COEFFS[4]); |
| 81 | |
| 82 | return fputil::cast<float16>(x: fputil::multiply_add(x: -xf, y: interm, z: 0.5f)); |
| 83 | } |
| 84 | |
| 85 | // When |x| > 0.5, assume that 0.5 < |x| <= 1 |
| 86 | // |
| 87 | // Step-by-step range-reduction proof: |
| 88 | // 1: Let y = asin(x), such that, x = sin(y) |
| 89 | // 2: From complimentary angle identity: |
| 90 | // x = sin(y) = cos(pi/2 - y) |
| 91 | // 3: Let z = pi/2 - y, such that x = cos(z) |
| 92 | // 4: From double angle formula; cos(2A) = 1 - 2 * sin^2(A): |
| 93 | // z = 2A, z/2 = A |
| 94 | // cos(z) = 1 - 2 * sin^2(z/2) |
| 95 | // 5: Make sin(z/2) subject of the formula: |
| 96 | // sin(z/2) = sqrt((1 - cos(z))/2) |
| 97 | // 6: Recall [3]; x = cos(z). Therefore: |
| 98 | // sin(z/2) = sqrt((1 - x)/2) |
| 99 | // 7: Let u = (1 - x)/2 |
| 100 | // 8: Therefore: |
| 101 | // asin(sqrt(u)) = z/2 |
| 102 | // 2 * asin(sqrt(u)) = z |
| 103 | // 9: Recall [3]; z = pi/2 - y. Therefore: |
| 104 | // y = pi/2 - z |
| 105 | // y = pi/2 - 2 * asin(sqrt(u)) |
| 106 | // 10: Recall [1], y = asin(x). Therefore: |
| 107 | // asin(x) = pi/2 - 2 * asin(sqrt(u)) |
| 108 | // 11: Recall that: acos(x) = pi/2 + asin(-x) = pi/2 - asin(x) |
| 109 | // Therefore: |
| 110 | // acos(x) = pi/2 - (pi/2 - 2 * asin(sqrt(u))) |
| 111 | // acos(x) = 2 * asin(sqrt(u)) |
| 112 | // acospi(x) = 2 * (asin(sqrt(u)) / pi) |
| 113 | // |
| 114 | // THE RANGE REDUCTION, HOW? |
| 115 | // 12: Recall [7], u = (1 - x)/2 |
| 116 | // 13: Since 0.5 < x <= 1, therefore: |
| 117 | // 0 <= u <= 0.25 and 0 <= sqrt(u) <= 0.5 |
| 118 | // |
| 119 | // Hence, we can reuse the same [0, 0.5] domain polynomial approximation for |
| 120 | // Step [11] as `sqrt(u)` is in range. |
| 121 | // When -1 < x <= -0.5, the identity: |
| 122 | // acos(x) = pi - acos(-x) |
| 123 | // acospi(x) = 1 - acos(-x)/pi |
| 124 | // allows us to compute for the negative x value (lhs) |
| 125 | // with a positive x value instead (rhs). |
| 126 | |
| 127 | float xf_abs = (xf < 0 ? -xf : xf); |
| 128 | float u = fputil::multiply_add(x: -0.5f, y: xf_abs, z: 0.5f); |
| 129 | float sqrt_u = fputil::sqrt<float>(x: u); |
| 130 | |
| 131 | float asin_sqrt_u = |
| 132 | sqrt_u * fputil::polyeval(x: u, a0: POLY_COEFFS[0], a: POLY_COEFFS[1], |
| 133 | a: POLY_COEFFS[2], a: POLY_COEFFS[3], a: POLY_COEFFS[4]); |
| 134 | |
| 135 | // Same as acos(x), but devided the expression with pi |
| 136 | return fputil::cast<float16>( |
| 137 | x: x_sign ? fputil::multiply_add(x: -2.0f, y: asin_sqrt_u, z: 1.0f) |
| 138 | : 2.0f * asin_sqrt_u); |
| 139 | } |
| 140 | |
| 141 | } // namespace math |
| 142 | |
| 143 | } // namespace LIBC_NAMESPACE_DECL |
| 144 | |
| 145 | #endif // LIBC_TYPES_HAS_FLOAT16 |
| 146 | |
| 147 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_ACOSPIF16_H |
| 148 | |