1 /*
2  * Single-precision vector e^x function.
3  *
4  * Copyright (c) 2019, Arm Limited.
5  * SPDX-License-Identifier: MIT
6  */
7 
8 #include "mathlib.h"
9 #include "v_math.h"
10 #if V_SUPPORTED
11 
12 static const float Poly[] = {
13   /*  maxerr: 0.36565 +0.5 ulp.  */
14   0x1.6a6000p-10f,
15   0x1.12718ep-7f,
16   0x1.555af0p-5f,
17   0x1.555430p-3f,
18   0x1.fffff4p-2f,
19 };
20 #define C0 v_f32 (Poly[0])
21 #define C1 v_f32 (Poly[1])
22 #define C2 v_f32 (Poly[2])
23 #define C3 v_f32 (Poly[3])
24 #define C4 v_f32 (Poly[4])
25 
26 #define Shift v_f32 (0x1.8p23f)
27 #define InvLn2 v_f32 (0x1.715476p+0f)
28 #define Ln2hi v_f32 (0x1.62e4p-1f)
29 #define Ln2lo v_f32 (0x1.7f7d1cp-20f)
30 
31 VPCS_ATTR
32 static v_f32_t
33 specialcase (v_f32_t poly, v_f32_t n, v_u32_t e, v_f32_t absn)
34 {
35   /* 2^n may overflow, break it up into s1*s2.  */
36   v_u32_t b = v_cond_u32 (n <= v_f32 (0.0f)) & v_u32 (0x83000000);
37   v_f32_t s1 = v_as_f32_u32 (v_u32 (0x7f000000) + b);
38   v_f32_t s2 = v_as_f32_u32 (e - b);
39   v_u32_t cmp = v_cond_u32 (absn > v_f32 (192.0f));
40   v_f32_t r1 = s1 * s1;
41   v_f32_t r0 = poly * s1 * s2;
42   return v_as_f32_u32 ((cmp & v_as_u32_f32 (r1)) | (~cmp & v_as_u32_f32 (r0)));
43 }
44 
45 VPCS_ATTR
46 v_f32_t
47 V_NAME(expf_1u) (v_f32_t x)
48 {
49   v_f32_t n, r, scale, poly, absn, z;
50   v_u32_t cmp, e;
51 
52   /* exp(x) = 2^n * poly(r), with poly(r) in [1/sqrt(2),sqrt(2)]
53      x = ln2*n + r, with r in [-ln2/2, ln2/2].  */
54 #if 1
55   z = v_fma_f32 (x, InvLn2, Shift);
56   n = z - Shift;
57   r = v_fma_f32 (n, -Ln2hi, x);
58   r = v_fma_f32 (n, -Ln2lo, r);
59   e = v_as_u32_f32 (z) << 23;
60 #else
61   z = x * InvLn2;
62   n = v_round_f32 (z);
63   r = v_fma_f32 (n, -Ln2hi, x);
64   r = v_fma_f32 (n, -Ln2lo, r);
65   e = v_as_u32_s32 (v_round_s32 (z)) << 23;
66 #endif
67   scale = v_as_f32_u32 (e + v_u32 (0x3f800000));
68   absn = v_abs_f32 (n);
69   cmp = v_cond_u32 (absn > v_f32 (126.0f));
70   poly = v_fma_f32 (C0, r, C1);
71   poly = v_fma_f32 (poly, r, C2);
72   poly = v_fma_f32 (poly, r, C3);
73   poly = v_fma_f32 (poly, r, C4);
74   poly = v_fma_f32 (poly, r, v_f32 (1.0f));
75   poly = v_fma_f32 (poly, r, v_f32 (1.0f));
76   if (unlikely (v_any_u32 (cmp)))
77     return specialcase (poly, n, e, absn);
78   return scale * poly;
79 }
80 #endif
81