1 /*
2 * Core approximation for single-precision vector sincos
3 *
4 * Copyright (c) 2023-2024, Arm Limited.
5 * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6 */
7
8 #include "v_math.h"
9
10 const static struct v_sincosf_data
11 {
12 float32x4_t poly_sin[3], poly_cos[3], pio2[3], inv_pio2, shift, range_val;
13 } v_sincosf_data = {
14 .poly_sin = { /* Generated using Remez, odd coeffs only, in [-pi/4, pi/4]. */
15 V4 (-0x1.555546p-3), V4 (0x1.11076p-7), V4 (-0x1.994eb4p-13) },
16 .poly_cos = { /* Generated using Remez, even coeffs only, in [-pi/4, pi/4]. */
17 V4 (0x1.55554ap-5), V4 (-0x1.6c0c1ap-10), V4 (0x1.99e0eep-16) },
18 .pio2 = { V4 (0x1.921fb6p+0f), V4 (-0x1.777a5cp-25f), V4 (-0x1.ee59dap-50f) },
19 .inv_pio2 = V4 (0x1.45f306p-1f),
20 .shift = V4 (0x1.8p23),
21 .range_val = V4 (0x1p20),
22 };
23
24 static inline uint32x4_t
check_ge_rangeval(float32x4_t x,const struct v_sincosf_data * d)25 check_ge_rangeval (float32x4_t x, const struct v_sincosf_data *d)
26 {
27 return vcagtq_f32 (x, d->range_val);
28 }
29
30 /* Single-precision vector function allowing calculation of both sin and cos in
31 one function call, using shared argument reduction and separate low-order
32 polynomials.
33 Worst-case error for sin is 1.67 ULP:
34 v_sincosf_sin(0x1.c704c4p+19) got 0x1.fff698p-5 want 0x1.fff69cp-5
35 Worst-case error for cos is 1.81 ULP:
36 v_sincosf_cos(0x1.e506fp+19) got -0x1.ffec6ep-6 want -0x1.ffec72p-6. */
37 static inline float32x4x2_t
v_sincosf_inline(float32x4_t x,const struct v_sincosf_data * d)38 v_sincosf_inline (float32x4_t x, const struct v_sincosf_data *d)
39 {
40 /* n = rint ( x / (pi/2) ). */
41 float32x4_t shift = d->shift;
42 float32x4_t q = vfmaq_f32 (shift, x, d->inv_pio2);
43 q = vsubq_f32 (q, shift);
44 int32x4_t n = vcvtq_s32_f32 (q);
45
46 /* Reduce x such that r is in [ -pi/4, pi/4 ]. */
47 float32x4_t r = x;
48 r = vfmsq_f32 (r, q, d->pio2[0]);
49 r = vfmsq_f32 (r, q, d->pio2[1]);
50 r = vfmsq_f32 (r, q, d->pio2[2]);
51
52 /* Approximate sin(r) ~= r + r^3 * poly_sin(r^2). */
53 float32x4_t r2 = vmulq_f32 (r, r), r3 = vmulq_f32 (r, r2);
54 float32x4_t s = vfmaq_f32 (d->poly_sin[1], r2, d->poly_sin[2]);
55 s = vfmaq_f32 (d->poly_sin[0], r2, s);
56 s = vfmaq_f32 (r, r3, s);
57
58 /* Approximate cos(r) ~= 1 - (r^2)/2 + r^4 * poly_cos(r^2). */
59 float32x4_t r4 = vmulq_f32 (r2, r2);
60 float32x4_t p = vfmaq_f32 (d->poly_cos[1], r2, d->poly_cos[2]);
61 float32x4_t c = vfmaq_f32 (v_f32 (-0.5), r2, d->poly_cos[0]);
62 c = vfmaq_f32 (c, r4, p);
63 c = vfmaq_f32 (v_f32 (1), c, r2);
64
65 /* If odd quadrant, swap cos and sin. */
66 uint32x4_t swap = vtstq_u32 (vreinterpretq_u32_s32 (n), v_u32 (1));
67 float32x4_t ss = vbslq_f32 (swap, c, s);
68 float32x4_t cc = vbslq_f32 (swap, s, c);
69
70 /* Fix signs according to quadrant.
71 ss = asfloat(asuint(ss) ^ ((n & 2) << 30))
72 cc = asfloat(asuint(cc) & (((n + 1) & 2) << 30)). */
73 uint32x4_t sin_sign
74 = vshlq_n_u32 (vandq_u32 (vreinterpretq_u32_s32 (n), v_u32 (2)), 30);
75 uint32x4_t cos_sign = vshlq_n_u32 (
76 vandq_u32 (vreinterpretq_u32_s32 (vaddq_s32 (n, v_s32 (1))), v_u32 (2)),
77 30);
78 ss = vreinterpretq_f32_u32 (
79 veorq_u32 (vreinterpretq_u32_f32 (ss), sin_sign));
80 cc = vreinterpretq_f32_u32 (
81 veorq_u32 (vreinterpretq_u32_f32 (cc), cos_sign));
82
83 return (float32x4x2_t){ ss, cc };
84 }
85