xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/sve/tanpif.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1 /*
2  * Single-precision vector tanpif(x) function.
3  *
4  * Copyright (c) 2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "sv_math.h"
9 #include "test_defs.h"
10 #include "test_sig.h"
11 
12 const static struct v_tanpif_data
13 {
14   float c0, c2, c4, c6;
15   float c1, c3, c5, c7;
16 } tanpif_data = {
17   /* Coefficients for tan(pi * x).  */
18   .c0 = 0x1.921fb4p1f,	.c1 = 0x1.4abbcep3f,  .c2 = 0x1.466b8p5f,
19   .c3 = 0x1.461c72p7f,	.c4 = 0x1.42e9d4p9f,  .c5 = 0x1.69e2c4p11f,
20   .c6 = 0x1.e85558p11f, .c7 = 0x1.a52e08p16f,
21 };
22 
23 /* Approximation for single-precision vector tanpif(x)
24    The maximum error is 3.34 ULP:
25    _ZGVsMxv_tanpif(0x1.d6c09ap-2) got 0x1.f70aacp+2
26 				 want 0x1.f70aa6p+2.  */
SV_NAME_F1(tanpi)27 svfloat32_t SV_NAME_F1 (tanpi) (svfloat32_t x, const svbool_t pg)
28 {
29   const struct v_tanpif_data *d = ptr_barrier (&tanpif_data);
30   svfloat32_t odd_coeffs = svld1rq (pg, &d->c1);
31   svfloat32_t n = svrintn_x (pg, x);
32 
33   /* inf produces nan that propagates.  */
34   svfloat32_t xr = svsub_x (pg, x, n);
35   svfloat32_t ar = svabd_x (pg, x, n);
36   svbool_t flip = svcmpgt (pg, ar, 0.25f);
37   svfloat32_t r = svsel (flip, svsub_x (pg, sv_f32 (0.5f), ar), ar);
38 
39   svfloat32_t r2 = svmul_x (pg, r, r);
40   svfloat32_t r4 = svmul_x (pg, r2, r2);
41 
42   /* Order-7 Pairwise Horner.  */
43   svfloat32_t p01 = svmla_lane (sv_f32 (d->c0), r2, odd_coeffs, 0);
44   svfloat32_t p23 = svmla_lane (sv_f32 (d->c2), r2, odd_coeffs, 1);
45   svfloat32_t p45 = svmla_lane (sv_f32 (d->c4), r2, odd_coeffs, 2);
46   svfloat32_t p67 = svmla_lane (sv_f32 (d->c6), r2, odd_coeffs, 3);
47   svfloat32_t p = svmad_x (pg, p67, r4, p45);
48   p = svmad_x (pg, p, r4, p23);
49   p = svmad_x (pg, p, r4, p01);
50   svfloat32_t poly = svmul_x (pg, r, p);
51 
52   svfloat32_t poly_recip = svdiv_x (pg, sv_f32 (1.0), poly);
53   svfloat32_t y = svsel (flip, poly_recip, poly);
54 
55   svuint32_t sign
56       = sveor_x (pg, svreinterpret_u32 (xr), svreinterpret_u32 (ar));
57   return svreinterpret_f32 (svorr_x (pg, svreinterpret_u32 (y), sign));
58 }
59 
60 #if WANT_TRIGPI_TESTS
61 TEST_DISABLE_FENV (SV_NAME_F1 (tanpi))
62 TEST_ULP (SV_NAME_F1 (tanpi), 2.84)
63 TEST_SYM_INTERVAL (SV_NAME_F1 (tanpi), 0, 0x1p-31, 50000)
64 TEST_SYM_INTERVAL (SV_NAME_F1 (tanpi), 0x1p-31, 0.5, 100000)
65 TEST_SYM_INTERVAL (SV_NAME_F1 (tanpi), 0.5, 0x1p23f, 100000)
66 TEST_SYM_INTERVAL (SV_NAME_F1 (tanpi), 0x1p23f, inf, 100000)
67 #endif
68 CLOSE_SVE_ATTR
69