xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/v_exp2f.c (revision 6580f5c38dd5b01aeeaed16b370f1a12423437f0)
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 struct data
12 {
13   float32x4_t poly[5];
14   uint32x4_t exponent_bias;
15 #if !WANT_SIMD_EXCEPT
16   float32x4_t special_bound, scale_thresh;
17 #endif
18 } data = {
19   /* maxerr: 1.962 ulp.  */
20   .poly = { V4 (0x1.59977ap-10f), V4 (0x1.3ce9e4p-7f), V4 (0x1.c6bd32p-5f),
21 	    V4 (0x1.ebf9bcp-3f), V4 (0x1.62e422p-1f) },
22   .exponent_bias = V4 (0x3f800000),
23 #if !WANT_SIMD_EXCEPT
24   .special_bound = V4 (126.0f),
25   .scale_thresh = V4 (192.0f),
26 #endif
27 };
28 
29 #define C(i) d->poly[i]
30 
31 #if WANT_SIMD_EXCEPT
32 
33 # define TinyBound v_u32 (0x20000000)	  /* asuint (0x1p-63).  */
34 # define BigBound v_u32 (0x42800000)	  /* asuint (0x1p6).  */
35 # define SpecialBound v_u32 (0x22800000) /* BigBound - TinyBound.  */
36 
37 static float32x4_t VPCS_ATTR NOINLINE
38 special_case (float32x4_t x, float32x4_t y, uint32x4_t cmp)
39 {
40   /* If fenv exceptions are to be triggered correctly, fall back to the scalar
41      routine for special lanes.  */
42   return v_call_f32 (exp2f, x, y, cmp);
43 }
44 
45 #else
46 
47 # define SpecialOffset v_u32 (0x82000000)
48 # define SpecialBias v_u32 (0x7f000000)
49 
50 static float32x4_t VPCS_ATTR NOINLINE
51 special_case (float32x4_t poly, float32x4_t n, uint32x4_t e, uint32x4_t cmp1,
52 	      float32x4_t scale, const struct data *d)
53 {
54   /* 2^n may overflow, break it up into s1*s2.  */
55   uint32x4_t b = vandq_u32 (vclezq_f32 (n), SpecialOffset);
56   float32x4_t s1 = vreinterpretq_f32_u32 (vaddq_u32 (b, SpecialBias));
57   float32x4_t s2 = vreinterpretq_f32_u32 (vsubq_u32 (e, b));
58   uint32x4_t cmp2 = vcagtq_f32 (n, d->scale_thresh);
59   float32x4_t r2 = vmulq_f32 (s1, s1);
60   float32x4_t r1 = vmulq_f32 (vfmaq_f32 (s2, poly, s2), s1);
61   /* Similar to r1 but avoids double rounding in the subnormal range.  */
62   float32x4_t r0 = vfmaq_f32 (scale, poly, scale);
63   float32x4_t r = vbslq_f32 (cmp1, r1, r0);
64   return vbslq_f32 (cmp2, r2, r);
65 }
66 
67 #endif
68 
69 float32x4_t VPCS_ATTR V_NAME_F1 (exp2) (float32x4_t x)
70 {
71   const struct data *d = ptr_barrier (&data);
72   float32x4_t n, r, r2, scale, p, q, poly;
73   uint32x4_t cmp, e;
74 
75 #if WANT_SIMD_EXCEPT
76   /* asuint(|x|) - TinyBound >= BigBound - TinyBound.  */
77   uint32x4_t ia = vreinterpretq_u32_f32 (vabsq_f32 (x));
78   cmp = vcgeq_u32 (vsubq_u32 (ia, TinyBound), SpecialBound);
79   float32x4_t xm = x;
80   /* If any lanes are special, mask them with 1 and retain a copy of x to allow
81      special_case to fix special lanes later. This is only necessary if fenv
82      exceptions are to be triggered correctly.  */
83   if (unlikely (v_any_u32 (cmp)))
84     x = vbslq_f32 (cmp, v_f32 (1), x);
85 #endif
86 
87     /* exp2(x) = 2^n (1 + poly(r)), with 1 + poly(r) in [1/sqrt(2),sqrt(2)]
88        x = n + r, with r in [-1/2, 1/2].  */
89   n = vrndaq_f32 (x);
90   r = vsubq_f32 (x, n);
91   e = vshlq_n_u32 (vreinterpretq_u32_s32 (vcvtaq_s32_f32 (x)), 23);
92   scale = vreinterpretq_f32_u32 (vaddq_u32 (e, d->exponent_bias));
93 
94 #if !WANT_SIMD_EXCEPT
95   cmp = vcagtq_f32 (n, d->special_bound);
96 #endif
97 
98   r2 = vmulq_f32 (r, r);
99   p = vfmaq_f32 (C (1), C (0), r);
100   q = vfmaq_f32 (C (3), C (2), r);
101   q = vfmaq_f32 (q, p, r2);
102   p = vmulq_f32 (C (4), r);
103   poly = vfmaq_f32 (p, q, r2);
104 
105   if (unlikely (v_any_u32 (cmp)))
106 #if WANT_SIMD_EXCEPT
107     return special_case (xm, vfmaq_f32 (scale, poly, scale), cmp);
108 #else
109     return special_case (poly, n, e, cmp, scale, d);
110 #endif
111 
112   return vfmaq_f32 (scale, poly, scale);
113 }
114