2 * Copyright 2023 Siemens
4 * The authors hereby grant permission to use, copy, modify, distribute,
5 * and license this software and its documentation for any purpose, provided
6 * that existing copyright notices are retained in all copies and that this
7 * notice is included verbatim in any distributions. No written agreement,
8 * license, or royalty fee is required for any of the authorized uses.
9 * Modifications to this software may be copyrighted by their authors
10 * and need not follow the licensing terms described here, provided that
11 * the new terms are clearly indicated on the first page of each file where
16 * Copyright (c) 1994-2009 Red Hat, Inc. All rights reserved.
18 * This copyrighted material is made available to anyone wishing to use,
19 * modify, copy, or redistribute it subject to the terms and conditions
20 * of the BSD License. This program is distributed in the hope that
21 * it will be useful, but WITHOUT ANY WARRANTY expressed or implied,
22 * including the implied warranties of MERCHANTABILITY or FITNESS FOR
23 * A PARTICULAR PURPOSE. A copy of this license is available at
24 * http://www.opensource.org/licenses. Any Red Hat trademarks that are
25 * incorporated in the source code or documentation are not subject to
26 * the BSD License and may only be used or replicated with the express
27 * permission of Red Hat, Inc.
30 /******************************************************************
31 * The following routines are coded directly from the algorithms
32 * and coefficients given in "Software Manual for the Elementary
33 * Functions" by William J. Cody, Jr. and William Waite, Prentice
35 ******************************************************************/
37 /* Based on newlib/libm/mathfp/s_exp.c in Newlib. */
39 #include "amdgcnmach.h"
41 v64si
v64df_ispos (v64df
);
42 v64si
v64df_numtest (v64df
);
44 static const double INV_LN2
= 1.4426950408889634074;
45 static const double LN2
= 0.6931471805599453094172321;
46 static const double p
[] = { 0.25, 0.75753180159422776666e-2,
47 0.31555192765684646356e-4 };
48 static const double q
[] = { 0.5, 0.56817302698551221787e-1,
49 0.63121894374398504557e-3,
50 0.75104028399870046114e-6 };
52 #if defined (__has_builtin) && __has_builtin (__builtin_gcn_ldexpv)
54 DEF_VD_MATH_FUNC (v64df
, exp
, v64df x
)
56 FUNCTION_INIT (v64df
);
58 v64si num_type
= v64df_numtest (x
);
59 VECTOR_IF (num_type
== NAN
, cond
)
61 VECTOR_RETURN (x
, cond
);
62 VECTOR_ELSEIF (num_type
== INF
, cond
)
64 VECTOR_RETURN (VECTOR_MERGE (VECTOR_INIT (z_infinity
.d
),
68 VECTOR_ELSEIF (num_type
== 0, cond
)
69 VECTOR_RETURN (VECTOR_INIT (1.0), cond
);
72 /* Check for out of bounds. */
73 VECTOR_IF ((x
> BIGX
) | (x
< SMALLX
), cond
)
75 VECTOR_RETURN (x
, cond
);
78 /* Check for a value too small to calculate. */
79 VECTOR_RETURN (VECTOR_INIT (1.0),
80 (-z_rooteps_f
< x
) & (x
< z_rooteps_f
));
82 /* Calculate the exponent. */
83 v64si Nneg
= __builtin_convertvector (x
* INV_LN2
- 0.5, v64si
);
84 v64si Npos
= __builtin_convertvector (x
* INV_LN2
+ 0.5, v64si
);
85 v64si N
= VECTOR_MERGE (Nneg
, Npos
, x
< 0.0);
87 /* Construct the mantissa. */
88 v64df g
= x
- __builtin_convertvector (N
, v64df
) * LN2
;
90 v64df P
= g
* ((p
[2] * z
+ p
[1]) * z
+ p
[0]);
91 v64df Q
= ((q
[3] * z
+ q
[2]) * z
+ q
[1]) * z
+ q
[0];
92 v64df R
= 0.5 + P
/ (Q
- P
);
94 /* Return the floating point value. */
96 VECTOR_RETURN (__builtin_gcn_ldexpv (R
, N
), NO_COND
);
101 DEF_VARIANTS (exp
, df
, df
)