1/*
2 * Single-precision SVE sinpi(x) 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 "sv_math.h"
10#include "pl_sig.h"
11#include "pl_test.h"
12#include "poly_sve_f32.h"
13
14static const struct data
15{
16  float poly[6];
17} data = {
18  /* Taylor series coefficents for sin(pi * x).  */
19  .poly = { 0x1.921fb6p1f, -0x1.4abbcep2f, 0x1.466bc6p1f, -0x1.32d2ccp-1f,
20	    0x1.50783p-4f, -0x1.e30750p-8f },
21};
22
23/* A fast SVE implementation of sinpif.
24   Maximum error 2.48 ULP:
25   _ZGVsMxv_sinpif(0x1.d062b6p-2) got 0x1.fa8c06p-1
26				 want 0x1.fa8c02p-1.  */
27svfloat32_t SV_NAME_F1 (sinpi) (svfloat32_t x, const svbool_t pg)
28{
29  const struct data *d = ptr_barrier (&data);
30
31  /* range reduction into -1/2 .. 1/2
32     with n = rint(x) and r = r - n.  */
33  svfloat32_t n = svrinta_x (pg, x);
34  svfloat32_t r = svsub_x (pg, x, n);
35
36  /* Result should be negated based on if n is odd or not.  */
37  svuint32_t intn = svreinterpret_u32 (svcvt_s32_x (pg, n));
38  svuint32_t sign = svlsl_z (pg, intn, 31);
39
40  /* y = sin(r).  */
41  svfloat32_t r2 = svmul_x (pg, r, r);
42  svfloat32_t y = sv_horner_5_f32_x (pg, r2, d->poly);
43  y = svmul_x (pg, y, r);
44
45  return svreinterpret_f32 (sveor_x (pg, svreinterpret_u32 (y), sign));
46}
47
48PL_SIG (SV, F, 1, sinpi, -0.9, 0.9)
49PL_TEST_ULP (SV_NAME_F1 (sinpi), 1.99)
50PL_TEST_SYM_INTERVAL (SV_NAME_F1 (sinpi), 0, 0x1p-31, 5000)
51PL_TEST_SYM_INTERVAL (SV_NAME_F1 (sinpi), 0x1p-31, 0.5, 10000)
52PL_TEST_SYM_INTERVAL (SV_NAME_F1 (sinpi), 0.5, 0x1p22f, 10000)
53PL_TEST_SYM_INTERVAL (SV_NAME_F1 (sinpi), 0x1p22f, inf, 10000)
54