1 | //===-- Half-precision asinf16(x) 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/asinf16.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 | // Generated by Sollya using the following command: |
23 | // > round(pi/2, D, RN); |
24 | static constexpr float PI_2 = 0x1.921fb54442d18p0f; |
25 | |
26 | LLVM_LIBC_FUNCTION(float16, asinf16, (float16 x)) { |
27 | using FPBits = fputil::FPBits<float16>; |
28 | FPBits xbits(x); |
29 | |
30 | uint16_t x_u = xbits.uintval(); |
31 | uint16_t x_abs = x_u & 0x7fff; |
32 | float xf = x; |
33 | |
34 | // |x| > 0x1p0, |x| > 1, or x is NaN. |
35 | if (LIBC_UNLIKELY(x_abs > 0x3c00)) { |
36 | // asinf16(NaN) = NaN |
37 | if (xbits.is_nan()) { |
38 | if (xbits.is_signaling_nan()) { |
39 | fputil::raise_except_if_required(FE_INVALID); |
40 | return FPBits::quiet_nan().get_val(); |
41 | } |
42 | |
43 | return x; |
44 | } |
45 | |
46 | // 1 < |x| <= +/-inf |
47 | fputil::raise_except_if_required(FE_INVALID); |
48 | fputil::set_errno_if_required(EDOM); |
49 | |
50 | return FPBits::quiet_nan().get_val(); |
51 | } |
52 | |
53 | float xsq = xf * xf; |
54 | |
55 | // |x| <= 0x1p-1, |x| <= 0.5 |
56 | if (x_abs <= 0x3800) { |
57 | // asinf16(+/-0) = +/-0 |
58 | if (LIBC_UNLIKELY(x_abs == 0)) |
59 | return x; |
60 | |
61 | // Exhaustive tests show that, |
62 | // for |x| <= 0x1.878p-9, when: |
63 | // x > 0, and rounding upward, or |
64 | // x < 0, and rounding downward, then, |
65 | // asin(x) = x * 2^-11 + x |
66 | // else, in other rounding modes, |
67 | // asin(x) = x |
68 | if (LIBC_UNLIKELY(x_abs <= 0x1a1e)) { |
69 | int rounding = fputil::quick_get_round(); |
70 | |
71 | if ((xbits.is_pos() && rounding == FE_UPWARD) || |
72 | (xbits.is_neg() && rounding == FE_DOWNWARD)) |
73 | return fputil::cast<float16>(fputil::multiply_add(xf, 0x1.0p-11f, xf)); |
74 | return x; |
75 | } |
76 | |
77 | // Degree-6 minimax odd polynomial of asin(x) generated by Sollya with: |
78 | // > P = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8|], [|SG...|], [0, 0.5]); |
79 | float result = |
80 | fputil::polyeval(xsq, 0x1.000002p0f, 0x1.554c2ap-3f, 0x1.3541ccp-4f, |
81 | 0x1.43b2d6p-5f, 0x1.a0d73ep-5f); |
82 | return fputil::cast<float16>(xf * result); |
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 - 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 | // |
109 | // WHY? |
110 | // 11: Recall [7], u = (1 - x)/2 |
111 | // 12: Since 0.5 < x <= 1, therefore: |
112 | // 0 <= u <= 0.25 and 0 <= sqrt(u) <= 0.5 |
113 | // |
114 | // Hence, we can reuse the same [0, 0.5] domain polynomial approximation for |
115 | // Step [10] as `sqrt(u)` is in range. |
116 | |
117 | // 0x1p-1 < |x| <= 0x1p0, 0.5 < |x| <= 1.0 |
118 | float xf_abs = (xf < 0 ? -xf : xf); |
119 | float sign = (xbits.uintval() >> 15 == 1 ? -1.0 : 1.0); |
120 | float u = fputil::multiply_add(-0.5f, xf_abs, 0.5f); |
121 | float u_sqrt = fputil::sqrt<float>(u); |
122 | |
123 | // Degree-6 minimax odd polynomial of asin(x) generated by Sollya with: |
124 | // > P = fpminimax(asin(x)/x, [|0, 2, 4, 6, 8|], [|SG...|], [0, 0.5]); |
125 | float asin_sqrt_u = |
126 | u_sqrt * fputil::polyeval(u, 0x1.000002p0f, 0x1.554c2ap-3f, |
127 | 0x1.3541ccp-4f, 0x1.43b2d6p-5f, 0x1.a0d73ep-5f); |
128 | |
129 | return fputil::cast<float16>(sign * |
130 | fputil::multiply_add(-2.0f, asin_sqrt_u, PI_2)); |
131 | } |
132 | |
133 | } // namespace LIBC_NAMESPACE_DECL |
134 | |