xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/v_logf.c (revision a90b9d0159070121c221b966469c3e36d912bf82)
1 /*
2  * Single-precision vector log 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   uint32x4_t min_norm;
14   uint16x8_t special_bound;
15   float32x4_t poly[7];
16   float32x4_t ln2, tiny_bound;
17   uint32x4_t off, mantissa_mask;
18 } data = {
19   /* 3.34 ulp error.  */
20   .poly = { V4 (-0x1.3e737cp-3f), V4 (0x1.5a9aa2p-3f), V4 (-0x1.4f9934p-3f),
21 	    V4 (0x1.961348p-3f), V4 (-0x1.00187cp-2f), V4 (0x1.555d7cp-2f),
22 	    V4 (-0x1.ffffc8p-2f) },
23   .ln2 = V4 (0x1.62e43p-1f),
24   .tiny_bound = V4 (0x1p-126),
25   .min_norm = V4 (0x00800000),
26   .special_bound = V8 (0x7f00), /* asuint32(inf) - min_norm.  */
27   .off = V4 (0x3f2aaaab),	/* 0.666667.  */
28   .mantissa_mask = V4 (0x007fffff)
29 };
30 
31 #define P(i) d->poly[7 - i]
32 
33 static float32x4_t VPCS_ATTR NOINLINE
34 special_case (float32x4_t x, float32x4_t y, float32x4_t r2, float32x4_t p,
35 	      uint16x4_t cmp)
36 {
37   /* Fall back to scalar code.  */
38   return v_call_f32 (logf, x, vfmaq_f32 (p, y, r2), vmovl_u16 (cmp));
39 }
40 
41 float32x4_t VPCS_ATTR V_NAME_F1 (log) (float32x4_t x)
42 {
43   const struct data *d = ptr_barrier (&data);
44   float32x4_t n, p, q, r, r2, y;
45   uint32x4_t u;
46   uint16x4_t cmp;
47 
48   u = vreinterpretq_u32_f32 (x);
49   cmp = vcge_u16 (vsubhn_u32 (u, d->min_norm),
50 		  vget_low_u16 (d->special_bound));
51 
52   /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3.  */
53   u = vsubq_u32 (u, d->off);
54   n = vcvtq_f32_s32 (
55       vshrq_n_s32 (vreinterpretq_s32_u32 (u), 23)); /* signextend.  */
56   u = vandq_u32 (u, d->mantissa_mask);
57   u = vaddq_u32 (u, d->off);
58   r = vsubq_f32 (vreinterpretq_f32_u32 (u), v_f32 (1.0f));
59 
60   /* y = log(1+r) + n*ln2.  */
61   r2 = vmulq_f32 (r, r);
62   /* n*ln2 + r + r2*(P1 + r*P2 + r2*(P3 + r*P4 + r2*(P5 + r*P6 + r2*P7))).  */
63   p = vfmaq_f32 (P (5), P (6), r);
64   q = vfmaq_f32 (P (3), P (4), r);
65   y = vfmaq_f32 (P (1), P (2), r);
66   p = vfmaq_f32 (p, P (7), r2);
67   q = vfmaq_f32 (q, p, r2);
68   y = vfmaq_f32 (y, q, r2);
69   p = vfmaq_f32 (r, d->ln2, n);
70 
71   if (unlikely (v_any_u16h (cmp)))
72     return special_case (x, y, r2, p, cmp);
73   return vfmaq_f32 (p, y, r2);
74 }
75