xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/sve/tanpi.c (revision dd21556857e8d40f66bf5ad54754d9d52669ebf7)
1 /*
2  * Double-precision vector tanpi(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_sig.h"
10 #include "test_defs.h"
11 
12 const static struct v_tanpi_data
13 {
14   double c0, c2, c4, c6, c8, c10, c12;
15   double c1, c3, c5, c7, c9, c11, c13, c14;
16 } tanpi_data = {
17   /* Coefficents for tan(pi * x) computed with fpminimax
18      on [ 0x1p-1022 0x1p-2 ]
19      approx rel error: 0x1.7eap-55
20      approx abs error: 0x1.7eap-55.  */
21   .c0 = 0x1.921fb54442d18p1, /* pi.  */
22   .c1 = 0x1.4abbce625be52p3,	.c2 = 0x1.466bc6775b0f9p5,
23   .c3 = 0x1.45fff9b426f5ep7,	.c4 = 0x1.45f4730dbca5cp9,
24   .c5 = 0x1.45f3265994f85p11,	.c6 = 0x1.45f4234b330cap13,
25   .c7 = 0x1.45dca11be79ebp15,	.c8 = 0x1.47283fc5eea69p17,
26   .c9 = 0x1.3a6d958cdefaep19,	.c10 = 0x1.927896baee627p21,
27   .c11 = -0x1.89333f6acd922p19, .c12 = 0x1.5d4e912bb8456p27,
28   .c13 = -0x1.a854d53ab6874p29, .c14 = 0x1.1b76de7681424p32,
29 };
30 
31 /* Approximation for double-precision vector tanpi(x)
32    The maximum error is 3.06 ULP:
33    _ZGVsMxv_tanpi(0x1.0a4a07dfcca3ep-1) got -0x1.fa30112702c98p+3
34 				       want -0x1.fa30112702c95p+3.  */
35 svfloat64_t SV_NAME_D1 (tanpi) (svfloat64_t x, const svbool_t pg)
36 {
37   const struct v_tanpi_data *d = ptr_barrier (&tanpi_data);
38 
39   svfloat64_t n = svrintn_x (pg, x);
40 
41   /* inf produces nan that propagates.  */
42   svfloat64_t xr = svsub_x (pg, x, n);
43   svfloat64_t ar = svabd_x (pg, x, n);
44   svbool_t flip = svcmpgt (pg, ar, 0.25);
45   svfloat64_t r = svsel (flip, svsubr_x (pg, ar, 0.5), ar);
46 
47   /* Order-14 pairwise Horner.  */
48   svfloat64_t r2 = svmul_x (pg, r, r);
49   svfloat64_t r4 = svmul_x (pg, r2, r2);
50 
51   svfloat64_t c_1_3 = svld1rq (pg, &d->c1);
52   svfloat64_t c_5_7 = svld1rq (pg, &d->c5);
53   svfloat64_t c_9_11 = svld1rq (pg, &d->c9);
54   svfloat64_t c_13_14 = svld1rq (pg, &d->c13);
55   svfloat64_t p01 = svmla_lane (sv_f64 (d->c0), r2, c_1_3, 0);
56   svfloat64_t p23 = svmla_lane (sv_f64 (d->c2), r2, c_1_3, 1);
57   svfloat64_t p45 = svmla_lane (sv_f64 (d->c4), r2, c_5_7, 0);
58   svfloat64_t p67 = svmla_lane (sv_f64 (d->c6), r2, c_5_7, 1);
59   svfloat64_t p89 = svmla_lane (sv_f64 (d->c8), r2, c_9_11, 0);
60   svfloat64_t p1011 = svmla_lane (sv_f64 (d->c10), r2, c_9_11, 1);
61   svfloat64_t p1213 = svmla_lane (sv_f64 (d->c12), r2, c_13_14, 0);
62 
63   svfloat64_t p = svmla_lane (p1213, r4, c_13_14, 1);
64   p = svmad_x (pg, p, r4, p1011);
65   p = svmad_x (pg, p, r4, p89);
66   p = svmad_x (pg, p, r4, p67);
67   p = svmad_x (pg, p, r4, p45);
68   p = svmad_x (pg, p, r4, p23);
69   p = svmad_x (pg, p, r4, p01);
70   p = svmul_x (pg, r, p);
71 
72   svfloat64_t p_recip = svdivr_x (pg, p, 1.0);
73   svfloat64_t y = svsel (flip, p_recip, p);
74 
75   svuint64_t sign
76       = sveor_x (pg, svreinterpret_u64 (xr), svreinterpret_u64 (ar));
77   return svreinterpret_f64 (svorr_x (pg, svreinterpret_u64 (y), sign));
78 }
79 
80 #if WANT_TRIGPI_TESTS
81 TEST_DISABLE_FENV (SV_NAME_D1 (tanpi))
82 TEST_ULP (SV_NAME_D1 (tanpi), 2.57)
83 TEST_SYM_INTERVAL (SV_NAME_D1 (tanpi), 0, 0x1p-31, 50000)
84 TEST_SYM_INTERVAL (SV_NAME_D1 (tanpi), 0x1p-31, 0.5, 50000)
85 TEST_SYM_INTERVAL (SV_NAME_D1 (tanpi), 0.5, 1.0, 200000)
86 TEST_SYM_INTERVAL (SV_NAME_D1 (tanpi), 1.0, 0x1p23, 50000)
87 TEST_SYM_INTERVAL (SV_NAME_D1 (tanpi), 0x1p23, inf, 50000)
88 #endif
89 CLOSE_SVE_ATTR
90