| 1 | //===-- Implementation header for acosbf16 ----------------------*- 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_ACOSBF16_H |
| 10 | #define LLVM_LIBC_SRC___SUPPORT_MATH_ACOSBF16_H |
| 11 | |
| 12 | #include "src/__support/FPUtil/FEnvImpl.h" |
| 13 | #include "src/__support/FPUtil/FPBits.h" |
| 14 | #include "src/__support/FPUtil/bfloat16.h" |
| 15 | #include "src/__support/FPUtil/cast.h" |
| 16 | #include "src/__support/FPUtil/multiply_add.h" |
| 17 | #include "src/__support/FPUtil/sqrt.h" |
| 18 | #include "src/__support/macros/optimization.h" |
| 19 | #include "src/__support/math/inv_trigf_utils.h" |
| 20 | |
| 21 | namespace LIBC_NAMESPACE_DECL { |
| 22 | namespace math { |
| 23 | |
| 24 | LIBC_INLINE bfloat16 acosbf16(bfloat16 x) { |
| 25 | // Generated by Sollya using the following command: |
| 26 | // > display = hexadecimal; |
| 27 | // > round(pi/2, SG, RN); |
| 28 | constexpr float PI_2 = 0x1.921fb6p0f; |
| 29 | // > round(pi, SG, RN); |
| 30 | constexpr float PI = 0x1.921fb6p1f; |
| 31 | |
| 32 | using FPBits = fputil::FPBits<bfloat16>; |
| 33 | FPBits xbits(x); |
| 34 | |
| 35 | uint16_t x_u = xbits.uintval(); |
| 36 | uint16_t x_abs = x_u & 0x7fff; |
| 37 | bool sign = (x_u >> 15); |
| 38 | float xf = x; |
| 39 | |
| 40 | float xf_abs = (xf < 0 ? -xf : xf); |
| 41 | float x_sq = xf_abs * xf_abs; |
| 42 | |
| 43 | // case 1: x <= 0.5 |
| 44 | if (x_abs <= 0x3F00) { |
| 45 | // |x| = {0} |
| 46 | if (LIBC_UNLIKELY(x_abs == 0)) |
| 47 | return fputil::cast<bfloat16>(x: PI_2); |
| 48 | |
| 49 | float xp = fputil::cast<float>(x: inv_trigf_utils_internal::asin_eval(xsq: x_sq)); |
| 50 | float result = xf * fputil::multiply_add(x: x_sq, y: xp, z: 1.0f); |
| 51 | return fputil::cast<bfloat16>(x: PI_2 - result); |
| 52 | } |
| 53 | |
| 54 | // case 2: 0.5< |x|<= 1.0 |
| 55 | if (x_abs <= 0x3F80) { |
| 56 | // |x| = {1} |
| 57 | if (x_abs == 0x3F80) { |
| 58 | if (sign) |
| 59 | return fputil::cast<bfloat16>(x: PI); |
| 60 | else |
| 61 | return FPBits::zero().get_val(); |
| 62 | } |
| 63 | |
| 64 | // using reduction for acos: |
| 65 | // acos(|x|) = 2*asin(sqrt((1 - |x|)/2)), |
| 66 | // and acos(x) = acos(|x|) for x >= 0, pi - acos(|x|) for x < 0 |
| 67 | float t = fputil::multiply_add<float>(x: xf_abs, y: -0.5f, z: 0.5f); |
| 68 | float t_sqrt = fputil::sqrt<float>(x: t); |
| 69 | // TODO: Use bfloat16 version for inv_trigf_utils_internals after they are |
| 70 | // available Tracking issue : |
| 71 | // https://github.com/llvm/llvm-project/issues/202079 |
| 72 | float tp = fputil::cast<float>(x: inv_trigf_utils_internal::asin_eval(xsq: t)); |
| 73 | float asin_sqrt_t = t_sqrt * (fputil::multiply_add(x: t, y: tp, z: 1.0f)); |
| 74 | |
| 75 | return fputil::cast<bfloat16>( |
| 76 | x: (sign) ? fputil::multiply_add(x: asin_sqrt_t, y: -2.0f, z: PI) |
| 77 | : 2.0f * asin_sqrt_t); |
| 78 | } |
| 79 | // case 3: NaN or |x| > 1 |
| 80 | // NaN |
| 81 | if (xbits.is_nan()) { |
| 82 | if (xbits.is_signaling_nan()) { |
| 83 | fputil::raise_except_if_required(FE_INVALID); |
| 84 | return FPBits::quiet_nan().get_val(); |
| 85 | } |
| 86 | return x; // quiet NaN |
| 87 | } |
| 88 | // |x| > 1 |
| 89 | fputil::raise_except_if_required(FE_INVALID); |
| 90 | fputil::set_errno_if_required(EDOM); // Domain is bounded |
| 91 | return FPBits::quiet_nan().get_val(); |
| 92 | } |
| 93 | |
| 94 | } // namespace math |
| 95 | } // namespace LIBC_NAMESPACE_DECL |
| 96 | |
| 97 | #endif // LLVM_LIBC_SRC___SUPPORT_MATH_ACOSBF16_H |
| 98 | |