xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/advsimd/sin.c (revision dd21556857e8d40f66bf5ad54754d9d52669ebf7)
1 /*
2  * Double-precision vector sin function.
3  *
4  * Copyright (c) 2019-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "test_defs.h"
9 #include "test_sig.h"
10 #include "mathlib.h"
11 #include "v_math.h"
12 
13 static const struct data
14 {
15   float64x2_t poly[7];
16   float64x2_t range_val, inv_pi, pi_1, pi_2, pi_3;
17 } data = {
18   .poly = { V2 (-0x1.555555555547bp-3), V2 (0x1.1111111108a4dp-7),
19 	    V2 (-0x1.a01a019936f27p-13), V2 (0x1.71de37a97d93ep-19),
20 	    V2 (-0x1.ae633919987c6p-26), V2 (0x1.60e277ae07cecp-33),
21 	    V2 (-0x1.9e9540300a1p-41) },
22 
23   .range_val = V2 (0x1p23),
24   .inv_pi = V2 (0x1.45f306dc9c883p-2),
25   .pi_1 = V2 (0x1.921fb54442d18p+1),
26   .pi_2 = V2 (0x1.1a62633145c06p-53),
27   .pi_3 = V2 (0x1.c1cd129024e09p-106),
28 };
29 
30 #if WANT_SIMD_EXCEPT
31 /* asuint64(0x1p-253)), below which multiply by inv_pi underflows.  */
32 # define TinyBound v_u64 (0x3020000000000000)
33 /* RangeVal - TinyBound.  */
34 # define Thresh v_u64 (0x1160000000000000)
35 #endif
36 
37 #define C(i) d->poly[i]
38 
39 static float64x2_t VPCS_ATTR NOINLINE
40 special_case (float64x2_t x, float64x2_t y, uint64x2_t odd, uint64x2_t cmp)
41 {
42   y = vreinterpretq_f64_u64 (veorq_u64 (vreinterpretq_u64_f64 (y), odd));
43   return v_call_f64 (sin, x, y, cmp);
44 }
45 
46 /* Vector (AdvSIMD) sin approximation.
47    Maximum observed error in [-pi/2, pi/2], where argument is not reduced,
48    is 2.87 ULP:
49    _ZGVnN2v_sin (0x1.921d5c6a07142p+0) got 0x1.fffffffa7dc02p-1
50 				      want 0x1.fffffffa7dc05p-1
51    Maximum observed error in the entire non-special domain ([-2^23, 2^23])
52    is 3.22 ULP:
53    _ZGVnN2v_sin (0x1.5702447b6f17bp+22) got 0x1.ffdcd125c84fbp-3
54 				       want 0x1.ffdcd125c84f8p-3.  */
55 float64x2_t VPCS_ATTR V_NAME_D1 (sin) (float64x2_t x)
56 {
57   const struct data *d = ptr_barrier (&data);
58   float64x2_t n, r, r2, r3, r4, y, t1, t2, t3;
59   uint64x2_t odd, cmp;
60 
61 #if WANT_SIMD_EXCEPT
62   /* Detect |x| <= TinyBound or |x| >= RangeVal. If fenv exceptions are to be
63      triggered correctly, set any special lanes to 1 (which is neutral w.r.t.
64      fenv). These lanes will be fixed by special-case handler later.  */
65   uint64x2_t ir = vreinterpretq_u64_f64 (vabsq_f64 (x));
66   cmp = vcgeq_u64 (vsubq_u64 (ir, TinyBound), Thresh);
67   r = vreinterpretq_f64_u64 (vbicq_u64 (vreinterpretq_u64_f64 (x), cmp));
68 #else
69   r = x;
70   cmp = vcageq_f64 (x, d->range_val);
71 #endif
72 
73   /* n = rint(|x|/pi).  */
74   n = vrndaq_f64 (vmulq_f64 (r, d->inv_pi));
75   odd = vshlq_n_u64 (vreinterpretq_u64_s64 (vcvtq_s64_f64 (n)), 63);
76 
77   /* r = |x| - n*pi  (range reduction into -pi/2 .. pi/2).  */
78   r = vfmsq_f64 (r, d->pi_1, n);
79   r = vfmsq_f64 (r, d->pi_2, n);
80   r = vfmsq_f64 (r, d->pi_3, n);
81 
82   /* sin(r) poly approx.  */
83   r2 = vmulq_f64 (r, r);
84   r3 = vmulq_f64 (r2, r);
85   r4 = vmulq_f64 (r2, r2);
86 
87   t1 = vfmaq_f64 (C (4), C (5), r2);
88   t2 = vfmaq_f64 (C (2), C (3), r2);
89   t3 = vfmaq_f64 (C (0), C (1), r2);
90 
91   y = vfmaq_f64 (t1, C (6), r4);
92   y = vfmaq_f64 (t2, y, r4);
93   y = vfmaq_f64 (t3, y, r4);
94   y = vfmaq_f64 (r, y, r3);
95 
96   if (unlikely (v_any_u64 (cmp)))
97     return special_case (x, y, odd, cmp);
98   return vreinterpretq_f64_u64 (veorq_u64 (vreinterpretq_u64_f64 (y), odd));
99 }
100 
101 TEST_SIG (V, D, 1, sin, -3.1, 3.1)
102 TEST_ULP (V_NAME_D1 (sin), 3.0)
103 TEST_DISABLE_FENV_IF_NOT (V_NAME_D1 (sin), WANT_SIMD_EXCEPT)
104 TEST_SYM_INTERVAL (V_NAME_D1 (sin), 0, 0x1p23, 500000)
105 TEST_SYM_INTERVAL (V_NAME_D1 (sin), 0x1p23, inf, 10000)
106