1*f3087befSAndrew Turner /*
2*f3087befSAndrew Turner * Single-precision vector sinpi function.
3*f3087befSAndrew Turner *
4*f3087befSAndrew Turner * Copyright (c) 2023-2024, Arm Limited.
5*f3087befSAndrew Turner * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*f3087befSAndrew Turner */
7*f3087befSAndrew Turner
8*f3087befSAndrew Turner #include "mathlib.h"
9*f3087befSAndrew Turner #include "v_math.h"
10*f3087befSAndrew Turner #include "v_poly_f32.h"
11*f3087befSAndrew Turner #include "test_sig.h"
12*f3087befSAndrew Turner #include "test_defs.h"
13*f3087befSAndrew Turner
14*f3087befSAndrew Turner static const struct data
15*f3087befSAndrew Turner {
16*f3087befSAndrew Turner float32x4_t poly[6];
17*f3087befSAndrew Turner } data = {
18*f3087befSAndrew Turner /* Taylor series coefficents for sin(pi * x). */
19*f3087befSAndrew Turner .poly = { V4 (0x1.921fb6p1f), V4 (-0x1.4abbcep2f), V4 (0x1.466bc6p1f),
20*f3087befSAndrew Turner V4 (-0x1.32d2ccp-1f), V4 (0x1.50783p-4f), V4 (-0x1.e30750p-8f) },
21*f3087befSAndrew Turner };
22*f3087befSAndrew Turner
23*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT
24*f3087befSAndrew Turner # define TinyBound v_u32 (0x30000000) /* asuint32(0x1p-31f). */
25*f3087befSAndrew Turner # define Thresh v_u32 (0x1f000000) /* asuint32(0x1p31f) - TinyBound. */
26*f3087befSAndrew Turner
27*f3087befSAndrew Turner static float32x4_t VPCS_ATTR NOINLINE
special_case(float32x4_t x,float32x4_t y,uint32x4_t odd,uint32x4_t cmp)28*f3087befSAndrew Turner special_case (float32x4_t x, float32x4_t y, uint32x4_t odd, uint32x4_t cmp)
29*f3087befSAndrew Turner {
30*f3087befSAndrew Turner /* Fall back to scalar code. */
31*f3087befSAndrew Turner y = vreinterpretq_f32_u32 (veorq_u32 (vreinterpretq_u32_f32 (y), odd));
32*f3087befSAndrew Turner return v_call_f32 (arm_math_sinpif, x, y, cmp);
33*f3087befSAndrew Turner }
34*f3087befSAndrew Turner #endif
35*f3087befSAndrew Turner
36*f3087befSAndrew Turner /* Approximation for vector single-precision sinpi(x)
37*f3087befSAndrew Turner Maximum Error 3.03 ULP:
38*f3087befSAndrew Turner _ZGVnN4v_sinpif(0x1.c597ccp-2) got 0x1.f7cd56p-1
39*f3087befSAndrew Turner want 0x1.f7cd5p-1. */
V_NAME_F1(sinpi)40*f3087befSAndrew Turner float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (sinpi) (float32x4_t x)
41*f3087befSAndrew Turner {
42*f3087befSAndrew Turner const struct data *d = ptr_barrier (&data);
43*f3087befSAndrew Turner
44*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT
45*f3087befSAndrew Turner uint32x4_t ir = vreinterpretq_u32_f32 (vabsq_f32 (x));
46*f3087befSAndrew Turner uint32x4_t cmp = vcgeq_u32 (vsubq_u32 (ir, TinyBound), Thresh);
47*f3087befSAndrew Turner
48*f3087befSAndrew Turner /* When WANT_SIMD_EXCEPT = 1, special lanes should be set to 0
49*f3087befSAndrew Turner to avoid them under/overflowing and throwing exceptions. */
50*f3087befSAndrew Turner float32x4_t r = v_zerofy_f32 (x, cmp);
51*f3087befSAndrew Turner #else
52*f3087befSAndrew Turner float32x4_t r = x;
53*f3087befSAndrew Turner #endif
54*f3087befSAndrew Turner
55*f3087befSAndrew Turner /* If r is odd, the sign of the result should be inverted. */
56*f3087befSAndrew Turner uint32x4_t odd
57*f3087befSAndrew Turner = vshlq_n_u32 (vreinterpretq_u32_s32 (vcvtaq_s32_f32 (r)), 31);
58*f3087befSAndrew Turner
59*f3087befSAndrew Turner /* r = x - rint(x). Range reduction to -1/2 .. 1/2. */
60*f3087befSAndrew Turner r = vsubq_f32 (r, vrndaq_f32 (r));
61*f3087befSAndrew Turner
62*f3087befSAndrew Turner /* Pairwise Horner approximation for y = sin(r * pi). */
63*f3087befSAndrew Turner float32x4_t r2 = vmulq_f32 (r, r);
64*f3087befSAndrew Turner float32x4_t r4 = vmulq_f32 (r2, r2);
65*f3087befSAndrew Turner float32x4_t y = vmulq_f32 (v_pw_horner_5_f32 (r2, r4, d->poly), r);
66*f3087befSAndrew Turner
67*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT
68*f3087befSAndrew Turner if (unlikely (v_any_u32 (cmp)))
69*f3087befSAndrew Turner return special_case (x, y, odd, cmp);
70*f3087befSAndrew Turner #endif
71*f3087befSAndrew Turner
72*f3087befSAndrew Turner return vreinterpretq_f32_u32 (veorq_u32 (vreinterpretq_u32_f32 (y), odd));
73*f3087befSAndrew Turner }
74*f3087befSAndrew Turner
75*f3087befSAndrew Turner HALF_WIDTH_ALIAS_F1 (sinpi)
76*f3087befSAndrew Turner
77*f3087befSAndrew Turner #if WANT_TRIGPI_TESTS
78*f3087befSAndrew Turner TEST_ULP (V_NAME_F1 (sinpi), 2.54)
79*f3087befSAndrew Turner TEST_DISABLE_FENV_IF_NOT (V_NAME_F1 (sinpi), WANT_SIMD_EXCEPT)
80*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (sinpi), 0, 0x1p-31, 5000)
81*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (sinpi), 0x1p-31, 0.5, 10000)
82*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (sinpi), 0.5, 0x1p31f, 10000)
83*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (sinpi), 0x1p31f, inf, 10000)
84*f3087befSAndrew Turner #endif
85