xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_erff_2u.c (revision 6580f5c38dd5b01aeeaed16b370f1a12423437f0)
1 /*
2  * Single-precision vector erf(x) function.
3  *
4  * Copyright (c) 2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "v_math.h"
9 #include "pl_sig.h"
10 #include "pl_test.h"
11 
12 static const struct data
13 {
14   float32x4_t max, shift, third;
15 #if WANT_SIMD_EXCEPT
16   float32x4_t tiny_bound, scale_minus_one;
17 #endif
18 } data = {
19   .max = V4 (3.9375), /* 4 - 8/128.  */
20   .shift = V4 (0x1p16f),
21   .third = V4 (0x1.555556p-2f), /* 1/3.  */
22 #if WANT_SIMD_EXCEPT
23   .tiny_bound = V4 (0x1p-62f),
24   .scale_minus_one = V4 (0x1.06eba8p-3f), /* scale - 1.0.  */
25 #endif
26 };
27 
28 #define AbsMask 0x7fffffff
29 
30 struct entry
31 {
32   float32x4_t erf;
33   float32x4_t scale;
34 };
35 
36 static inline struct entry
37 lookup (uint32x4_t i)
38 {
39   struct entry e;
40   float64_t t0 = *((float64_t *) (__erff_data.tab + i[0]));
41   float64_t t1 = *((float64_t *) (__erff_data.tab + i[1]));
42   float64_t t2 = *((float64_t *) (__erff_data.tab + i[2]));
43   float64_t t3 = *((float64_t *) (__erff_data.tab + i[3]));
44   float32x4_t e1 = vreinterpretq_f32_f64 ((float64x2_t){ t0, t1 });
45   float32x4_t e2 = vreinterpretq_f32_f64 ((float64x2_t){ t2, t3 });
46   e.erf = vuzp1q_f32 (e1, e2);
47   e.scale = vuzp2q_f32 (e1, e2);
48   return e;
49 }
50 
51 /* Single-precision implementation of vector erf(x).
52    Approximation based on series expansion near x rounded to
53    nearest multiple of 1/128.
54    Let d = x - r, and scale = 2 / sqrt(pi) * exp(-r^2). For x near r,
55 
56    erf(x) ~ erf(r) + scale * d * [1 - r * d - 1/3 * d^2]
57 
58    Values of erf(r) and scale are read from lookup tables.
59    For |x| > 3.9375, erf(|x|) rounds to 1.0f.
60 
61    Maximum error: 1.93 ULP
62      _ZGVnN4v_erff(0x1.c373e6p-9) got 0x1.fd686cp-9
63 				 want 0x1.fd6868p-9.  */
64 float32x4_t VPCS_ATTR V_NAME_F1 (erf) (float32x4_t x)
65 {
66   const struct data *dat = ptr_barrier (&data);
67 
68 #if WANT_SIMD_EXCEPT
69   /* |x| < 2^-62.  */
70   uint32x4_t cmp = vcaltq_f32 (x, dat->tiny_bound);
71   float32x4_t xm = x;
72   /* If any lanes are special, mask them with 1 and retain a copy of x to allow
73      special case handler to fix special lanes later. This is only necessary if
74      fenv exceptions are to be triggered correctly.  */
75   if (unlikely (v_any_u32 (cmp)))
76     x = vbslq_f32 (cmp, v_f32 (1), x);
77 #endif
78 
79   float32x4_t a = vabsq_f32 (x);
80   uint32x4_t a_gt_max = vcgtq_f32 (a, dat->max);
81 
82   /* Lookup erf(r) and scale(r) in tables, e.g. set erf(r) to 0 and scale to
83      2/sqrt(pi), when x reduced to r = 0.  */
84   float32x4_t shift = dat->shift;
85   float32x4_t z = vaddq_f32 (a, shift);
86 
87   uint32x4_t i
88       = vsubq_u32 (vreinterpretq_u32_f32 (z), vreinterpretq_u32_f32 (shift));
89   i = vminq_u32 (i, v_u32 (512));
90   struct entry e = lookup (i);
91 
92   float32x4_t r = vsubq_f32 (z, shift);
93 
94   /* erf(x) ~ erf(r) + scale * d * (1 - r * d - 1/3 * d^2).  */
95   float32x4_t d = vsubq_f32 (a, r);
96   float32x4_t d2 = vmulq_f32 (d, d);
97   float32x4_t y = vfmaq_f32 (r, dat->third, d);
98   y = vfmaq_f32 (e.erf, e.scale, vfmsq_f32 (d, d2, y));
99 
100   /* Solves the |x| = inf case.  */
101   y = vbslq_f32 (a_gt_max, v_f32 (1.0f), y);
102 
103   /* Copy sign.  */
104   y = vbslq_f32 (v_u32 (AbsMask), y, x);
105 
106 #if WANT_SIMD_EXCEPT
107   if (unlikely (v_any_u32 (cmp)))
108     return vbslq_f32 (cmp, vfmaq_f32 (xm, dat->scale_minus_one, xm), y);
109 #endif
110   return y;
111 }
112 
113 PL_SIG (V, F, 1, erf, -4.0, 4.0)
114 PL_TEST_ULP (V_NAME_F1 (erf), 1.43)
115 PL_TEST_EXPECT_FENV (V_NAME_F1 (erf), WANT_SIMD_EXCEPT)
116 PL_TEST_SYM_INTERVAL (V_NAME_F1 (erf), 0, 3.9375, 40000)
117 PL_TEST_SYM_INTERVAL (V_NAME_F1 (erf), 3.9375, inf, 40000)
118 PL_TEST_SYM_INTERVAL (V_NAME_F1 (erf), 0, inf, 40000)
119