xref: /freebsd/contrib/arm-optimized-routines/pl/math/asin_3u.c (revision e9ac41698b2f322d55ccf9da50a3596edb2c1800)
1 /*
2  * Double-precision asin(x) function.
3  *
4  * Copyright (c) 2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "poly_scalar_f64.h"
9 #include "math_config.h"
10 #include "pl_sig.h"
11 #include "pl_test.h"
12 
13 #define AbsMask (0x7fffffffffffffff)
14 #define Half (0x3fe0000000000000)
15 #define One (0x3ff0000000000000)
16 #define PiOver2 (0x1.921fb54442d18p+0)
17 #define Small (0x3e50000000000000) /* 2^-26.  */
18 #define Small16 (0x3e50)
19 #define QNaN (0x7ff8)
20 
21 /* Fast implementation of double-precision asin(x) based on polynomial
22    approximation.
23 
24    For x < Small, approximate asin(x) by x. Small = 2^-26 for correct rounding.
25 
26    For x in [Small, 0.5], use an order 11 polynomial P such that the final
27    approximation is an odd polynomial: asin(x) ~ x + x^3 P(x^2).
28 
29    The largest observed error in this region is 1.01 ulps,
30    asin(0x1.da9735b5a9277p-2) got 0x1.ed78525a927efp-2
31 			     want 0x1.ed78525a927eep-2.
32 
33    No cheap approximation can be obtained near x = 1, since the function is not
34    continuously differentiable on 1.
35 
36    For x in [0.5, 1.0], we use a method based on a trigonometric identity
37 
38      asin(x) = pi/2 - acos(x)
39 
40    and a generalized power series expansion of acos(y) near y=1, that reads as
41 
42      acos(y)/sqrt(2y) ~ 1 + 1/12 * y + 3/160 * y^2 + ... (1)
43 
44    The Taylor series of asin(z) near z = 0, reads as
45 
46      asin(z) ~ z + z^3 P(z^2) = z + z^3 * (1/6 + 3/40 z^2 + ...).
47 
48    Therefore, (1) can be written in terms of P(y/2) or even asin(y/2)
49 
50      acos(y) ~ sqrt(2y) (1 + y/2 * P(y/2)) = 2 * sqrt(y/2) (1 + y/2 * P(y/2)
51 
52    Hence, if we write z = (1-x)/2, z is near 0 when x approaches 1 and
53 
54      asin(x) ~ pi/2 - acos(x) ~ pi/2 - 2 * sqrt(z) (1 + z * P(z)).
55 
56    The largest observed error in this region is 2.69 ulps,
57    asin(0x1.044ac9819f573p-1) got 0x1.110d7e85fdd5p-1
58 			     want 0x1.110d7e85fdd53p-1.  */
59 double
60 asin (double x)
61 {
62   uint64_t ix = asuint64 (x);
63   uint64_t ia = ix & AbsMask;
64   uint64_t ia16 = ia >> 48;
65   double ax = asdouble (ia);
66   uint64_t sign = ix & ~AbsMask;
67 
68   /* Special values and invalid range.  */
69   if (unlikely (ia16 == QNaN))
70     return x;
71   if (ia > One)
72     return __math_invalid (x);
73   if (ia16 < Small16)
74     return x;
75 
76   /* Evaluate polynomial Q(x) = y + y * z * P(z) with
77      z2 = x ^ 2         and z = |x|     , if |x| < 0.5
78      z2 = (1 - |x|) / 2 and z = sqrt(z2), if |x| >= 0.5.  */
79   double z2 = ax < 0.5 ? x * x : fma (-0.5, ax, 0.5);
80   double z = ax < 0.5 ? ax : sqrt (z2);
81 
82   /* Use a single polynomial approximation P for both intervals.  */
83   double z4 = z2 * z2;
84   double z8 = z4 * z4;
85   double z16 = z8 * z8;
86   double p = estrin_11_f64 (z2, z4, z8, z16, __asin_poly);
87 
88   /* Finalize polynomial: z + z * z2 * P(z2).  */
89   p = fma (z * z2, p, z);
90 
91   /* asin(|x|) = Q(|x|)         , for |x| < 0.5
92 	       = pi/2 - 2 Q(|x|), for |x| >= 0.5.  */
93   double y = ax < 0.5 ? p : fma (-2.0, p, PiOver2);
94 
95   /* Copy sign.  */
96   return asdouble (asuint64 (y) | sign);
97 }
98 
99 PL_SIG (S, D, 1, asin, -1.0, 1.0)
100 PL_TEST_ULP (asin, 2.19)
101 PL_TEST_INTERVAL (asin, 0, Small, 5000)
102 PL_TEST_INTERVAL (asin, Small, 0.5, 50000)
103 PL_TEST_INTERVAL (asin, 0.5, 1.0, 50000)
104 PL_TEST_INTERVAL (asin, 1.0, 0x1p11, 50000)
105 PL_TEST_INTERVAL (asin, 0x1p11, inf, 20000)
106 PL_TEST_INTERVAL (asin, -0, -inf, 20000)
107