1 | //===-- Half-precision acospi function ------------------------------------===// |
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 | #include "src/math/acospif16.h" |
10 | #include "hdr/errno_macros.h" |
11 | #include "hdr/fenv_macros.h" |
12 | #include "src/__support/FPUtil/FEnvImpl.h" |
13 | #include "src/__support/FPUtil/FPBits.h" |
14 | #include "src/__support/FPUtil/PolyEval.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 | |
20 | namespace LIBC_NAMESPACE_DECL { |
21 | |
22 | LLVM_LIBC_FUNCTION(float16, acospif16, (float16 x)) { |
23 | using FPBits = fputil::FPBits<float16>; |
24 | FPBits xbits(x); |
25 | |
26 | uint16_t x_u = xbits.uintval(); |
27 | uint16_t x_abs = x_u & 0x7fff; |
28 | uint16_t x_sign = x_u >> 15; |
29 | |
30 | // |x| > 0x1p0, |x| > 1, or x is NaN. |
31 | if (LIBC_UNLIKELY(x_abs > 0x3c00)) { |
32 | // acospif16(NaN) = NaN |
33 | if (xbits.is_nan()) { |
34 | if (xbits.is_signaling_nan()) { |
35 | fputil::raise_except_if_required(FE_INVALID); |
36 | return FPBits::quiet_nan().get_val(); |
37 | } |
38 | |
39 | return x; |
40 | } |
41 | |
42 | // 1 < |x| <= +inf |
43 | fputil::raise_except_if_required(FE_INVALID); |
44 | fputil::set_errno_if_required(EDOM); |
45 | |
46 | return FPBits::quiet_nan().get_val(); |
47 | } |
48 | |
49 | // |x| == 0x1p0, x is 1 or -1 |
50 | // if x is (-)1, return 1 |
51 | // if x is (+)1, return 0 |
52 | if (LIBC_UNLIKELY(x_abs == 0x3c00)) |
53 | return fputil::cast<float16>(x_sign ? 1.0f : 0.0f); |
54 | |
55 | float xf = x; |
56 | float xsq = xf * xf; |
57 | |
58 | // Degree-6 minimax polynomial coefficients of asin(x) generated by Sollya |
59 | // with: > P = fpminimax(asin(x)/(pi * x), [|0, 2, 4, 6, 8|], [|SG...|], [0, |
60 | // 0.5]); |
61 | constexpr float POLY_COEFFS[5] = {0x1.45f308p-2f, 0x1.b2900cp-5f, |
62 | 0x1.897e36p-6f, 0x1.9efafcp-7f, |
63 | 0x1.06d884p-6f}; |
64 | // |x| <= 0x1p-1, |x| <= 0.5 |
65 | if (x_abs <= 0x3800) { |
66 | // if x is 0, return 0.5 |
67 | if (LIBC_UNLIKELY(x_abs == 0)) |
68 | return fputil::cast<float16>(0.5f); |
69 | |
70 | // Note that: acos(x) = pi/2 + asin(-x) = pi/2 - asin(x), then |
71 | // acospi(x) = 0.5 - asin(x)/pi |
72 | float interm = |
73 | fputil::polyeval(xsq, POLY_COEFFS[0], POLY_COEFFS[1], POLY_COEFFS[2], |
74 | POLY_COEFFS[3], POLY_COEFFS[4]); |
75 | |
76 | return fputil::cast<float16>(fputil::multiply_add(-xf, interm, 0.5f)); |
77 | } |
78 | |
79 | // When |x| > 0.5, assume that 0.5 < |x| <= 1 |
80 | // |
81 | // Step-by-step range-reduction proof: |
82 | // 1: Let y = asin(x), such that, x = sin(y) |
83 | // 2: From complimentary angle identity: |
84 | // x = sin(y) = cos(pi/2 - y) |
85 | // 3: Let z = pi/2 - y, such that x = cos(z) |
86 | // 4: From double angle formula; cos(2A) = 1 - 2 * sin^2(A): |
87 | // z = 2A, z/2 = A |
88 | // cos(z) = 1 - 2 * sin^2(z/2) |
89 | // 5: Make sin(z/2) subject of the formula: |
90 | // sin(z/2) = sqrt((1 - cos(z))/2) |
91 | // 6: Recall [3]; x = cos(z). Therefore: |
92 | // sin(z/2) = sqrt((1 - x)/2) |
93 | // 7: Let u = (1 - x)/2 |
94 | // 8: Therefore: |
95 | // asin(sqrt(u)) = z/2 |
96 | // 2 * asin(sqrt(u)) = z |
97 | // 9: Recall [3]; z = pi/2 - y. Therefore: |
98 | // y = pi/2 - z |
99 | // y = pi/2 - 2 * asin(sqrt(u)) |
100 | // 10: Recall [1], y = asin(x). Therefore: |
101 | // asin(x) = pi/2 - 2 * asin(sqrt(u)) |
102 | // 11: Recall that: acos(x) = pi/2 + asin(-x) = pi/2 - asin(x) |
103 | // Therefore: |
104 | // acos(x) = pi/2 - (pi/2 - 2 * asin(sqrt(u))) |
105 | // acos(x) = 2 * asin(sqrt(u)) |
106 | // acospi(x) = 2 * (asin(sqrt(u)) / pi) |
107 | // |
108 | // THE RANGE REDUCTION, HOW? |
109 | // 12: Recall [7], u = (1 - x)/2 |
110 | // 13: Since 0.5 < x <= 1, therefore: |
111 | // 0 <= u <= 0.25 and 0 <= sqrt(u) <= 0.5 |
112 | // |
113 | // Hence, we can reuse the same [0, 0.5] domain polynomial approximation for |
114 | // Step [11] as `sqrt(u)` is in range. |
115 | // When -1 < x <= -0.5, the identity: |
116 | // acos(x) = pi - acos(-x) |
117 | // acospi(x) = 1 - acos(-x)/pi |
118 | // allows us to compute for the negative x value (lhs) |
119 | // with a positive x value instead (rhs). |
120 | |
121 | float xf_abs = (xf < 0 ? -xf : xf); |
122 | float u = fputil::multiply_add(-0.5f, xf_abs, 0.5f); |
123 | float sqrt_u = fputil::sqrt<float>(u); |
124 | |
125 | float asin_sqrt_u = |
126 | sqrt_u * fputil::polyeval(u, POLY_COEFFS[0], POLY_COEFFS[1], |
127 | POLY_COEFFS[2], POLY_COEFFS[3], POLY_COEFFS[4]); |
128 | |
129 | // Same as acos(x), but devided the expression with pi |
130 | return fputil::cast<float16>( |
131 | x_sign ? fputil::multiply_add(-2.0f, asin_sqrt_u, 1.0f) |
132 | : 2.0f * asin_sqrt_u); |
133 | } |
134 | } // namespace LIBC_NAMESPACE_DECL |
135 | |