xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/sve/erff.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1 /*
2  * Single-precision vector erf(x) function.
3  *
4  * Copyright (c) 2023-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "sv_math.h"
9 #include "test_sig.h"
10 #include "test_defs.h"
11 
12 static const struct data
13 {
14   float min, max, scale, shift, third;
15 } data = {
16   .min = 0x1.cp-7f,	   /* 1/64 - 1/512.  */
17   .max = 3.9375,	   /* 4 - 8/128.  */
18   .scale = 0x1.20dd76p+0f, /* 2/sqrt(pi).  */
19   .shift = 0x1p16f,
20   .third = 0x1.555556p-2f, /* 1/3.  */
21 };
22 
23 #define SignMask (0x80000000)
24 
25 /* Single-precision implementation of vector erf(x).
26    Approximation based on series expansion near x rounded to
27    nearest multiple of 1/128.
28    Let d = x - r, and scale = 2 / sqrt(pi) * exp(-r^2). For x near r,
29 
30    erf(x) ~ erf(r) + scale * d * [1 - r * d - 1/3 * d^2]
31 
32    Values of erf(r) and scale are read from lookup tables.
33    For |x| < 0x1.cp-7, the algorithm sets r = 0, erf(r) = 0, and scale = 2 /
34    sqrt(pi), so it simply boils down to a Taylor series expansion near 0. For
35    |x| > 3.9375, erf(|x|) rounds to 1.0f.
36 
37    Maximum error on each interval:
38    - [0, 0x1.cp-7]: 1.93 ULP
39      _ZGVsMxv_erff(0x1.c373e6p-9) got 0x1.fd686cp-9 want 0x1.fd6868p-9
40    - [0x1.cp-7, 4.0]: 1.26 ULP
41      _ZGVsMxv_erff(0x1.1d002ep+0) got 0x1.c4eb9ap-1 want 0x1.c4eb98p-1.  */
SV_NAME_F1(erf)42 svfloat32_t SV_NAME_F1 (erf) (svfloat32_t x, const svbool_t pg)
43 {
44   const struct data *dat = ptr_barrier (&data);
45 
46   /* |x| > 1/64 - 1/512.  */
47   svbool_t a_gt_min = svacgt (pg, x, dat->min);
48 
49   /* |x| >= 4.0 - 8/128.  */
50   svbool_t a_ge_max = svacge (pg, x, dat->max);
51   svfloat32_t a = svabs_x (pg, x);
52 
53   svfloat32_t shift = sv_f32 (dat->shift);
54   svfloat32_t z = svadd_x (pg, a, shift);
55   svuint32_t i = svand_x (pg, svreinterpret_u32 (z), 0xfff);
56   i = svadd_x (pg, i, i);
57 
58   /* r and erf(r) set to 0 for |x| below min.  */
59   svfloat32_t r = svsub_z (a_gt_min, z, shift);
60   svfloat32_t erfr
61       = svld1_gather_index (a_gt_min, &__v_erff_data.tab[0].erf, i);
62 
63   /* scale set to 2/sqrt(pi) for |x| below min.  */
64   svfloat32_t scale
65       = svld1_gather_index (a_gt_min, &__v_erff_data.tab[0].scale, i);
66   scale = svsel (a_gt_min, scale, sv_f32 (dat->scale));
67 
68   /* erf(x) ~ erf(r) + scale * d * (1 - r * d + 1/3 * d^2).  */
69   svfloat32_t d = svsub_x (pg, a, r);
70   svfloat32_t d2 = svmul_x (pg, d, d);
71   svfloat32_t y = svmla_x (pg, r, d, dat->third);
72   y = svmla_x (pg, erfr, scale, svmls_x (pg, d, d2, y));
73 
74   /* Solves the |x| = inf case.  */
75   y = svsel (a_ge_max, sv_f32 (1.0f), y);
76 
77   /* Copy sign.  */
78   svuint32_t ix = svreinterpret_u32 (x);
79   svuint32_t iy = svreinterpret_u32 (y);
80   svuint32_t sign = svand_x (pg, ix, SignMask);
81   return svreinterpret_f32 (svorr_x (pg, sign, iy));
82 }
83 
84 TEST_SIG (SV, F, 1, erf, -4.0, 4.0)
85 TEST_ULP (SV_NAME_F1 (erf), 1.43)
86 TEST_DISABLE_FENV (SV_NAME_F1 (erf))
87 TEST_SYM_INTERVAL (SV_NAME_F1 (erf), 0, 0x1.cp-7, 40000)
88 TEST_SYM_INTERVAL (SV_NAME_F1 (erf), 0x1.cp-7, 3.9375, 40000)
89 TEST_SYM_INTERVAL (SV_NAME_F1 (erf), 3.9375, inf, 40000)
90 TEST_SYM_INTERVAL (SV_NAME_F1 (erf), 0, inf, 4000)
91 CLOSE_SVE_ATTR
92