xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/v_exp2f_1u.c (revision 96190b4fef3b4a0cc3ca0606b0c4e3e69a5e6717)
1 /*
2  * Single-precision vector 2^x function.
3  *
4  * Copyright (c) 2019-2023, 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 
11 static const float Poly[] = {
12   /*  maxerr: 0.878 ulp.  */
13   0x1.416b5ep-13f, 0x1.5f082ep-10f, 0x1.3b2dep-7f, 0x1.c6af7cp-5f, 0x1.ebfbdcp-3f, 0x1.62e43p-1f
14 };
15 #define C0 v_f32 (Poly[0])
16 #define C1 v_f32 (Poly[1])
17 #define C2 v_f32 (Poly[2])
18 #define C3 v_f32 (Poly[3])
19 #define C4 v_f32 (Poly[4])
20 #define C5 v_f32 (Poly[5])
21 
22 #define Shift v_f32 (0x1.8p23f)
23 #define InvLn2 v_f32 (0x1.715476p+0f)
24 #define Ln2hi v_f32 (0x1.62e4p-1f)
25 #define Ln2lo v_f32 (0x1.7f7d1cp-20f)
26 
27 static float32x4_t VPCS_ATTR NOINLINE
28 specialcase (float32x4_t poly, float32x4_t n, uint32x4_t e, float32x4_t absn)
29 {
30   /* 2^n may overflow, break it up into s1*s2.  */
31   uint32x4_t b = (n <= v_f32 (0.0f)) & v_u32 (0x83000000);
32   float32x4_t s1 = vreinterpretq_f32_u32 (v_u32 (0x7f000000) + b);
33   float32x4_t s2 = vreinterpretq_f32_u32 (e - b);
34   uint32x4_t cmp = absn > v_f32 (192.0f);
35   float32x4_t r1 = s1 * s1;
36   float32x4_t r0 = poly * s1 * s2;
37   return vreinterpretq_f32_u32 ((cmp & vreinterpretq_u32_f32 (r1))
38 				| (~cmp & vreinterpretq_u32_f32 (r0)));
39 }
40 
41 float32x4_t VPCS_ATTR
42 _ZGVnN4v_exp2f_1u (float32x4_t x)
43 {
44   float32x4_t n, r, scale, poly, absn;
45   uint32x4_t cmp, e;
46 
47   /* exp2(x) = 2^n * poly(r), with poly(r) in [1/sqrt(2),sqrt(2)]
48      x = n + r, with r in [-1/2, 1/2].  */
49 #if 0
50   float32x4_t z;
51   z = x + Shift;
52   n = z - Shift;
53   r = x - n;
54   e = vreinterpretq_u32_f32 (z) << 23;
55 #else
56   n = vrndaq_f32 (x);
57   r = x - n;
58   e = vreinterpretq_u32_s32 (vcvtaq_s32_f32 (x)) << 23;
59 #endif
60   scale = vreinterpretq_f32_u32 (e + v_u32 (0x3f800000));
61   absn = vabsq_f32 (n);
62   cmp = absn > v_f32 (126.0f);
63   poly = vfmaq_f32 (C1, C0, r);
64   poly = vfmaq_f32 (C2, poly, r);
65   poly = vfmaq_f32 (C3, poly, r);
66   poly = vfmaq_f32 (C4, poly, r);
67   poly = vfmaq_f32 (C5, poly, r);
68   poly = vfmaq_f32 (v_f32 (1.0f), poly, r);
69   if (unlikely (v_any_u32 (cmp)))
70     return specialcase (poly, n, e, absn);
71   return scale * poly;
72 }
73