1 /*
2 * Single-precision vector sin function.
3 *
4 * Copyright (c) 2019-2022, Arm Limited.
5 * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6 */
7
8 #include "mathlib.h"
9 #include "v_math.h"
10 #if V_SUPPORTED
11
12 static const float Poly[] = {
13 /* 1.886 ulp error */
14 0x1.5b2e76p-19f,
15 -0x1.9f42eap-13f,
16 0x1.110df4p-7f,
17 -0x1.555548p-3f,
18 };
19 #define Pi1 v_f32 (0x1.921fb6p+1f)
20 #define Pi2 v_f32 (-0x1.777a5cp-24f)
21 #define Pi3 v_f32 (-0x1.ee59dap-49f)
22 #define A3 v_f32 (Poly[3])
23 #define A5 v_f32 (Poly[2])
24 #define A7 v_f32 (Poly[1])
25 #define A9 v_f32 (Poly[0])
26 #define RangeVal v_f32 (0x1p20f)
27 #define TinyBound v_f32 (0x1p-61f)
28 #define InvPi v_f32 (0x1.45f306p-2f)
29 #define Shift v_f32 (0x1.8p+23f)
30 #define AbsMask v_u32 (0x7fffffff)
31
32 VPCS_ATTR
33 static v_f32_t
specialcase(v_f32_t x,v_f32_t y,v_u32_t cmp)34 specialcase (v_f32_t x, v_f32_t y, v_u32_t cmp)
35 {
36 /* Fall back to scalar code. */
37 return v_call_f32 (sinf, x, y, cmp);
38 }
39
40 VPCS_ATTR
41 v_f32_t
V_NAME(sinf)42 V_NAME(sinf) (v_f32_t x)
43 {
44 v_f32_t n, r, r2, y;
45 v_u32_t sign, odd, cmp, ir;
46
47 ir = v_as_u32_f32 (x) & AbsMask;
48 r = v_as_f32_u32 (ir);
49 sign = v_as_u32_f32 (x) & ~AbsMask;
50
51 #if WANT_SIMD_EXCEPT
52 cmp = v_cond_u32 ((ir - v_as_u32_f32 (TinyBound)
53 >= v_as_u32_f32 (RangeVal) - v_as_u32_f32 (TinyBound)));
54 if (unlikely (v_any_u32 (cmp)))
55 /* If fenv exceptions are to be triggered correctly, set any special lanes
56 to 1 (which is neutral w.r.t. fenv). These lanes will be fixed by
57 specialcase later. */
58 r = v_sel_f32 (cmp, v_f32 (1), r);
59 #else
60 cmp = v_cond_u32 (ir >= v_as_u32_f32 (RangeVal));
61 #endif
62
63 /* n = rint(|x|/pi) */
64 n = v_fma_f32 (InvPi, r, Shift);
65 odd = v_as_u32_f32 (n) << 31;
66 n -= Shift;
67
68 /* r = |x| - n*pi (range reduction into -pi/2 .. pi/2) */
69 r = v_fma_f32 (-Pi1, n, r);
70 r = v_fma_f32 (-Pi2, n, r);
71 r = v_fma_f32 (-Pi3, n, r);
72
73 /* y = sin(r) */
74 r2 = r * r;
75 y = v_fma_f32 (A9, r2, A7);
76 y = v_fma_f32 (y, r2, A5);
77 y = v_fma_f32 (y, r2, A3);
78 y = v_fma_f32 (y * r2, r, r);
79
80 /* sign fix */
81 y = v_as_f32_u32 (v_as_u32_f32 (y) ^ sign ^ odd);
82
83 if (unlikely (v_any_u32 (cmp)))
84 return specialcase (x, y, cmp);
85 return y;
86 }
87 VPCS_ALIAS
88 #endif
89