2 * Single-precision vector cos function.
4 * Copyright (c) 2019, Arm Limited.
5 * SPDX-License-Identifier: MIT
12 static const float Poly[] = {
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 InvPi v_f32 (0x1.45f306p-2f)
28 #define Shift v_f32 (0x1.8p+23f)
29 #define AbsMask v_u32 (0x7fffffff)
30 #define HalfPi v_f32 (0x1.921fb6p0f)
34 specialcase (v_f32_t x, v_f32_t y, v_u32_t cmp)
36 /* Fall back to scalar code. */
37 return v_call_f32 (cosf, x, y, cmp);
42 V_NAME(cosf) (v_f32_t x)
47 r = v_as_f32_u32 (v_as_u32_f32 (x) & AbsMask);
48 cmp = v_cond_u32 (v_as_u32_f32 (r) >= v_as_u32_f32 (RangeVal));
50 /* n = rint((|x|+pi/2)/pi) - 0.5 */
51 n = v_fma_f32 (InvPi, r + HalfPi, Shift);
52 odd = v_as_u32_f32 (n) << 31;
56 /* r = |x| - n*pi (range reduction into -pi/2 .. pi/2) */
57 r = v_fma_f32 (-Pi1, n, r);
58 r = v_fma_f32 (-Pi2, n, r);
59 r = v_fma_f32 (-Pi3, n, r);
63 y = v_fma_f32 (A9, r2, A7);
64 y = v_fma_f32 (y, r2, A5);
65 y = v_fma_f32 (y, r2, A3);
66 y = v_fma_f32 (y * r2, r, r);
69 y = v_as_f32_u32 (v_as_u32_f32 (y) ^ odd);
71 if (unlikely (v_any_u32 (cmp)))
72 return specialcase (x, y, cmp);