xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_sinpi_3u1.c (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 /*
2  * Double-precision vector sinpi function.
3  *
4  * Copyright (c) 2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "mathlib.h"
9 #include "v_math.h"
10 #include "poly_advsimd_f64.h"
11 #include "pl_sig.h"
12 #include "pl_test.h"
13 
14 static const struct data
15 {
16   float64x2_t poly[10];
17 } data = {
18   /* Polynomial coefficients generated using Remez algorithm,
19      see sinpi.sollya for details.  */
20   .poly = { V2 (0x1.921fb54442d184p1), V2 (-0x1.4abbce625be53p2),
21 	    V2 (0x1.466bc6775ab16p1), V2 (-0x1.32d2cce62dc33p-1),
22 	    V2 (0x1.507834891188ep-4), V2 (-0x1.e30750a28c88ep-8),
23 	    V2 (0x1.e8f48308acda4p-12), V2 (-0x1.6fc0032b3c29fp-16),
24 	    V2 (0x1.af86ae521260bp-21), V2 (-0x1.012a9870eeb7dp-25) },
25 };
26 
27 #if WANT_SIMD_EXCEPT
28 # define TinyBound v_u64 (0x3bf0000000000000) /* asuint64(0x1p-64).  */
29 /* asuint64(0x1p64) - TinyBound.  */
30 # define Thresh v_u64 (0x07f0000000000000)
31 
32 static float64x2_t VPCS_ATTR NOINLINE
33 special_case (float64x2_t x, float64x2_t y, uint64x2_t odd, uint64x2_t cmp)
34 {
35   /* Fall back to scalar code.  */
36   y = vreinterpretq_f64_u64 (veorq_u64 (vreinterpretq_u64_f64 (y), odd));
37   return v_call_f64 (sinpi, x, y, cmp);
38 }
39 #endif
40 
41 /* Approximation for vector double-precision sinpi(x).
42    Maximum Error 3.05 ULP:
43   _ZGVnN2v_sinpi(0x1.d32750db30b4ap-2) got 0x1.fb295878301c7p-1
44 				      want 0x1.fb295878301cap-1.  */
45 float64x2_t VPCS_ATTR V_NAME_D1 (sinpi) (float64x2_t x)
46 {
47   const struct data *d = ptr_barrier (&data);
48 
49 #if WANT_SIMD_EXCEPT
50   uint64x2_t ir = vreinterpretq_u64_f64 (vabsq_f64 (x));
51   uint64x2_t cmp = vcgeq_u64 (vsubq_u64 (ir, TinyBound), Thresh);
52 
53   /* When WANT_SIMD_EXCEPT = 1, special lanes should be set to 0
54      to avoid them under/overflowing and throwing exceptions.  */
55   float64x2_t r = v_zerofy_f64 (x, cmp);
56 #else
57   float64x2_t r = x;
58 #endif
59 
60   /* If r is odd, the sign of the result should be inverted.  */
61   uint64x2_t odd
62       = vshlq_n_u64 (vreinterpretq_u64_s64 (vcvtaq_s64_f64 (r)), 63);
63 
64   /* r = x - rint(x). Range reduction to -1/2 .. 1/2.  */
65   r = vsubq_f64 (r, vrndaq_f64 (r));
66 
67   /* y = sin(r).  */
68   float64x2_t r2 = vmulq_f64 (r, r);
69   float64x2_t r4 = vmulq_f64 (r2, r2);
70   float64x2_t y = vmulq_f64 (v_pw_horner_9_f64 (r2, r4, d->poly), r);
71 
72 #if WANT_SIMD_EXCEPT
73   if (unlikely (v_any_u64 (cmp)))
74     return special_case (x, y, odd, cmp);
75 #endif
76 
77   return vreinterpretq_f64_u64 (veorq_u64 (vreinterpretq_u64_f64 (y), odd));
78 }
79 
80 PL_SIG (V, D, 1, sinpi, -0.9, 0.9)
81 PL_TEST_ULP (V_NAME_D1 (sinpi), 3.06)
82 PL_TEST_EXPECT_FENV (V_NAME_D1 (sinpi), WANT_SIMD_EXCEPT)
83 PL_TEST_SYM_INTERVAL (V_NAME_D1 (sinpi), 0, 0x1p-63, 5000)
84 PL_TEST_SYM_INTERVAL (V_NAME_D1 (sinpi), 0x1p-63, 0.5, 10000)
85 PL_TEST_SYM_INTERVAL (V_NAME_D1 (sinpi), 0.5, 0x1p51, 10000)
86 PL_TEST_SYM_INTERVAL (V_NAME_D1 (sinpi), 0x1p51, inf, 10000)
87