xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_asinf_2u5.c (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 /*
2  * Single-precision vector 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 "v_math.h"
9 #include "poly_advsimd_f32.h"
10 #include "pl_sig.h"
11 #include "pl_test.h"
12 
13 static const struct data
14 {
15   float32x4_t poly[5];
16   float32x4_t pi_over_2f;
17 } data = {
18   /* Polynomial approximation of  (asin(sqrt(x)) - sqrt(x)) / (x * sqrt(x))  on
19      [ 0x1p-24 0x1p-2 ] order = 4 rel error: 0x1.00a23bbp-29 .  */
20   .poly = { V4 (0x1.55555ep-3), V4 (0x1.33261ap-4), V4 (0x1.70d7dcp-5),
21 	    V4 (0x1.b059dp-6), V4 (0x1.3af7d8p-5) },
22   .pi_over_2f = V4 (0x1.921fb6p+0f),
23 };
24 
25 #define AbsMask 0x7fffffff
26 #define Half 0x3f000000
27 #define One 0x3f800000
28 #define Small 0x39800000 /* 2^-12.  */
29 
30 #if WANT_SIMD_EXCEPT
31 static float32x4_t VPCS_ATTR NOINLINE
32 special_case (float32x4_t x, float32x4_t y, uint32x4_t special)
33 {
34   return v_call_f32 (asinf, x, y, special);
35 }
36 #endif
37 
38 /* Single-precision implementation of vector asin(x).
39 
40    For |x| < Small, approximate asin(x) by x. Small = 2^-12 for correct
41    rounding. If WANT_SIMD_EXCEPT = 0, Small = 0 and we proceed with the
42    following approximation.
43 
44    For |x| in [Small, 0.5], use order 4 polynomial P such that the final
45    approximation is an odd polynomial: asin(x) ~ x + x^3 P(x^2).
46 
47     The largest observed error in this region is 0.83 ulps,
48       _ZGVnN4v_asinf (0x1.ea00f4p-2) got 0x1.fef15ep-2 want 0x1.fef15cp-2.
49 
50     For |x| in [0.5, 1.0], use same approximation with a change of variable
51 
52     asin(x) = pi/2 - (y + y * z * P(z)), with  z = (1-x)/2 and y = sqrt(z).
53 
54    The largest observed error in this region is 2.41 ulps,
55      _ZGVnN4v_asinf (0x1.00203ep-1) got 0x1.0c3a64p-1 want 0x1.0c3a6p-1.  */
56 float32x4_t VPCS_ATTR V_NAME_F1 (asin) (float32x4_t x)
57 {
58   const struct data *d = ptr_barrier (&data);
59 
60   uint32x4_t ix = vreinterpretq_u32_f32 (x);
61   uint32x4_t ia = vandq_u32 (ix, v_u32 (AbsMask));
62 
63 #if WANT_SIMD_EXCEPT
64   /* Special values need to be computed with scalar fallbacks so
65      that appropriate fp exceptions are raised.  */
66   uint32x4_t special
67       = vcgtq_u32 (vsubq_u32 (ia, v_u32 (Small)), v_u32 (One - Small));
68   if (unlikely (v_any_u32 (special)))
69     return special_case (x, x, v_u32 (0xffffffff));
70 #endif
71 
72   float32x4_t ax = vreinterpretq_f32_u32 (ia);
73   uint32x4_t a_lt_half = vcltq_u32 (ia, v_u32 (Half));
74 
75   /* Evaluate polynomial Q(x) = y + y * z * P(z) with
76      z = x ^ 2 and y = |x|            , if |x| < 0.5
77      z = (1 - |x|) / 2 and y = sqrt(z), if |x| >= 0.5.  */
78   float32x4_t z2 = vbslq_f32 (a_lt_half, vmulq_f32 (x, x),
79 			      vfmsq_n_f32 (v_f32 (0.5), ax, 0.5));
80   float32x4_t z = vbslq_f32 (a_lt_half, ax, vsqrtq_f32 (z2));
81 
82   /* Use a single polynomial approximation P for both intervals.  */
83   float32x4_t p = v_horner_4_f32 (z2, d->poly);
84   /* Finalize polynomial: z + z * z2 * P(z2).  */
85   p = vfmaq_f32 (z, vmulq_f32 (z, z2), p);
86 
87   /* asin(|x|) = Q(|x|)         , for |x| < 0.5
88 	       = pi/2 - 2 Q(|x|), for |x| >= 0.5.  */
89   float32x4_t y
90       = vbslq_f32 (a_lt_half, p, vfmsq_n_f32 (d->pi_over_2f, p, 2.0));
91 
92   /* Copy sign.  */
93   return vbslq_f32 (v_u32 (AbsMask), y, x);
94 }
95 
96 PL_SIG (V, F, 1, asin, -1.0, 1.0)
97 PL_TEST_ULP (V_NAME_F1 (asin), 1.91)
98 PL_TEST_EXPECT_FENV (V_NAME_F1 (asin), WANT_SIMD_EXCEPT)
99 PL_TEST_INTERVAL (V_NAME_F1 (asin), 0, 0x1p-12, 5000)
100 PL_TEST_INTERVAL (V_NAME_F1 (asin), 0x1p-12, 0.5, 50000)
101 PL_TEST_INTERVAL (V_NAME_F1 (asin), 0.5, 1.0, 50000)
102 PL_TEST_INTERVAL (V_NAME_F1 (asin), 1.0, 0x1p11, 50000)
103 PL_TEST_INTERVAL (V_NAME_F1 (asin), 0x1p11, inf, 20000)
104 PL_TEST_INTERVAL (V_NAME_F1 (asin), -0, -inf, 20000)
105