1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
|
/*
* Double-precision SVE sinpi(x) function.
*
* Copyright (c) 2023, Arm Limited.
* SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
*/
#include "mathlib.h"
#include "sv_math.h"
#include "pl_sig.h"
#include "pl_test.h"
#include "poly_sve_f64.h"
static const struct data
{
double poly[10];
} data = {
/* Polynomial coefficients generated using Remez algorithm,
see sinpi.sollya for details. */
.poly = { 0x1.921fb54442d184p1, -0x1.4abbce625be53p2, 0x1.466bc6775ab16p1,
-0x1.32d2cce62dc33p-1, 0x1.507834891188ep-4, -0x1.e30750a28c88ep-8,
0x1.e8f48308acda4p-12, -0x1.6fc0032b3c29fp-16,
0x1.af86ae521260bp-21, -0x1.012a9870eeb7dp-25 },
};
/* A fast SVE implementation of sinpi.
Maximum error 3.10 ULP:
_ZGVsMxv_sinpi(0x1.df1a14f1b235p-2) got 0x1.fd64f541606cp-1
want 0x1.fd64f541606c3p-1. */
svfloat64_t SV_NAME_D1 (sinpi) (svfloat64_t x, const svbool_t pg)
{
const struct data *d = ptr_barrier (&data);
/* range reduction into -1/2 .. 1/2)
with n = rint(x) and r = r - n. */
svfloat64_t n = svrinta_x (pg, x);
svfloat64_t r = svsub_x (pg, x, n);
/* Result should be negated based on if n is odd or not. */
svuint64_t intn = svreinterpret_u64 (svcvt_s64_x (pg, n));
svuint64_t sign = svlsl_z (pg, intn, 63);
/* y = sin(r). */
svfloat64_t r2 = svmul_x (pg, r, r);
svfloat64_t r4 = svmul_x (pg, r2, r2);
svfloat64_t y = sv_pw_horner_9_f64_x (pg, r2, r4, d->poly);
y = svmul_x (pg, y, r);
return svreinterpret_f64 (sveor_x (pg, svreinterpret_u64 (y), sign));
}
PL_SIG (SV, D, 1, sinpi, -0.9, 0.9)
PL_TEST_ULP (SV_NAME_D1 (sinpi), 2.61)
PL_TEST_SYM_INTERVAL (SV_NAME_D1 (sinpi), 0, 0x1p-63, 5000)
PL_TEST_SYM_INTERVAL (SV_NAME_D1 (sinpi), 0x1p-63, 0.5, 10000)
PL_TEST_SYM_INTERVAL (SV_NAME_D1 (sinpi), 0.5, 0x1p51, 10000)
PL_TEST_SYM_INTERVAL (SV_NAME_D1 (sinpi), 0x1p51, inf, 10000)
|
