1 /* Double-precision vector (SVE) erf function
3 Copyright (C) 2024 Free Software Foundation, Inc.
4 This file is part of the GNU C Library.
6 The GNU C Library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Lesser General Public
8 License as published by the Free Software Foundation; either
9 version 2.1 of the License, or (at your option) any later version.
11 The GNU C Library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
16 You should have received a copy of the GNU Lesser General Public
17 License along with the GNU C Library; if not, see
18 <https://www.gnu.org/licenses/>. */
22 static const struct data
25 double tenth
, two_over_five
, two_over_fifteen
;
26 double two_over_nine
, two_over_fortyfive
;
29 .third
= 0x1.5555555555556p
-2, /* used to compute 2/3 and 1/6 too. */
30 .two_over_fifteen
= 0x1.1111111111111p
-3,
31 .tenth
= -0x1.999999999999ap
-4,
32 .two_over_five
= -0x1.999999999999ap
-2,
33 .two_over_nine
= -0x1.c71c71c71c71cp
-3,
34 .two_over_fortyfive
= 0x1.6c16c16c16c17p
-5,
35 .max
= 5.9921875, /* 6 - 1/128. */
39 #define SignMask (0x8000000000000000)
41 /* Double-precision implementation of vector erf(x).
42 Approximation based on series expansion near x rounded to
43 nearest multiple of 1/128.
44 Let d = x - r, and scale = 2 / sqrt(pi) * exp(-r^2). For x near r,
45 erf(x) ~ erf(r) + scale * d * [
49 - 1/6 (r (2 r^2 - 3)) d^3
50 + 1/30 (4 r^4 - 12 r^2 + 3) d^4
51 - 1/90 (4 r^4 - 20 r^2 + 15) d^5
54 Maximum measure error: 2.29 ULP
55 _ZGVsMxv_erf(-0x1.00003c924e5d1p-8) got -0x1.20dd59132ebadp-8
56 want -0x1.20dd59132ebafp-8. */
57 svfloat64_t
SV_NAME_D1 (erf
) (svfloat64_t x
, const svbool_t pg
)
59 const struct data
*dat
= ptr_barrier (&data
);
61 /* |x| >= 6.0 - 1/128. Opposite conditions except none of them catch NaNs so
62 they can be used in lookup and BSLs to yield the expected results. */
63 svbool_t a_ge_max
= svacge (pg
, x
, dat
->max
);
64 svbool_t a_lt_max
= svaclt (pg
, x
, dat
->max
);
66 /* Set r to multiple of 1/128 nearest to |x|. */
67 svfloat64_t a
= svabs_x (pg
, x
);
68 svfloat64_t shift
= sv_f64 (dat
->shift
);
69 svfloat64_t z
= svadd_x (pg
, a
, shift
);
71 = svsub_x (pg
, svreinterpret_u64 (z
), svreinterpret_u64 (shift
));
73 /* Lookup without shortcut for small values but with predicate to avoid
74 segfault for large values and NaNs. */
75 svfloat64_t r
= svsub_x (pg
, z
, shift
);
76 svfloat64_t erfr
= svld1_gather_index (a_lt_max
, __sv_erf_data
.erf
, i
);
77 svfloat64_t scale
= svld1_gather_index (a_lt_max
, __sv_erf_data
.scale
, i
);
79 /* erf(x) ~ erf(r) + scale * d * poly (r, d). */
80 svfloat64_t d
= svsub_x (pg
, a
, r
);
81 svfloat64_t d2
= svmul_x (pg
, d
, d
);
82 svfloat64_t r2
= svmul_x (pg
, r
, r
);
84 /* poly (d, r) = 1 + p1(r) * d + p2(r) * d^2 + ... + p5(r) * d^5. */
86 svfloat64_t third
= sv_f64 (dat
->third
);
87 svfloat64_t twothird
= svmul_x (pg
, third
, 2.0);
88 svfloat64_t sixth
= svmul_x (pg
, third
, 0.5);
89 svfloat64_t p2
= svmls_x (pg
, third
, r2
, twothird
);
90 svfloat64_t p3
= svmad_x (pg
, r2
, third
, -0.5);
91 p3
= svmul_x (pg
, r
, p3
);
93 = svmla_x (pg
, sv_f64 (dat
->two_over_five
), r2
, dat
->two_over_fifteen
);
94 p4
= svmls_x (pg
, sv_f64 (dat
->tenth
), r2
, p4
);
96 = svmla_x (pg
, sv_f64 (dat
->two_over_nine
), r2
, dat
->two_over_fortyfive
);
97 p5
= svmla_x (pg
, sixth
, r2
, p5
);
98 p5
= svmul_x (pg
, r
, p5
);
100 svfloat64_t p34
= svmla_x (pg
, p3
, d
, p4
);
101 svfloat64_t p12
= svmla_x (pg
, p1
, d
, p2
);
102 svfloat64_t y
= svmla_x (pg
, p34
, d2
, p5
);
103 y
= svmla_x (pg
, p12
, d2
, y
);
105 y
= svmla_x (pg
, erfr
, scale
, svmls_x (pg
, d
, d2
, y
));
107 /* Solves the |x| = inf and NaN cases. */
108 y
= svsel (a_ge_max
, sv_f64 (1.0), y
);
111 svuint64_t ix
= svreinterpret_u64 (x
);
112 svuint64_t iy
= svreinterpret_u64 (y
);
113 svuint64_t sign
= svand_x (pg
, ix
, SignMask
);
114 return svreinterpret_f64 (svorr_x (pg
, sign
, iy
));