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_logarithm.c in Newlib. */
39 #include "amdgcnmach.h"
41 v64si
v64df_finite (v64df
);
42 v64si
v64df_isnan (v64df
);
44 static const double a
[] = { -0.64124943423745581147e+02,
45 0.16383943563021534222e+02,
46 -0.78956112887481257267 };
47 static const double b
[] = { -0.76949932108494879777e+03,
48 0.31203222091924532844e+03,
49 -0.35667977739034646171e+02 };
50 static const double C1
= 22713.0 / 32768.0;
51 static const double C2
= 1.428606820309417232e-06;
53 #if defined (__has_builtin) \
54 && __has_builtin (__builtin_gcn_frexpv_mant) \
55 && __has_builtin (__builtin_gcn_frexpv_exp) \
57 DEF_VD_MATH_FUNC (v64df, log, v64df x)
59 FUNCTION_INIT (v64df
);
61 /* Check for domain/range errors here. */
62 VECTOR_IF (x
== 0.0, cond
)
64 VECTOR_RETURN (VECTOR_INIT (-z_infinity
.d
), cond
);
65 VECTOR_ELSEIF (x
< 0.0, cond
)
67 VECTOR_RETURN (VECTOR_INIT (z_notanum
.d
), cond
);
68 VECTOR_ELSEIF (__builtin_convertvector (~v64df_finite (x
), v64di
), cond
)
69 VECTOR_RETURN (VECTOR_MERGE (VECTOR_INIT (z_notanum
.d
),
70 VECTOR_INIT (z_infinity
.d
),
75 /* Get the exponent and mantissa where x = f * 2^N. */
76 v64df f
= __builtin_gcn_frexpv_mant (x
);
77 v64si N
= __builtin_gcn_frexpv_exp (x
);
81 VECTOR_IF (f
> __SQRT_HALF
, cond
)
82 VECTOR_COND_MOVE (z
, (z
- 0.5) / (f
* 0.5 + 0.5), cond
);
84 VECTOR_COND_MOVE (N
, N
- 1, cond
);
85 VECTOR_COND_MOVE (z
, z
/ (z
* 0.5 + 0.5), cond
);
90 /* Use Newton's method with 4 terms. */
91 z
+= z
* w
* ((a
[2] * w
+ a
[1]) * w
+ a
[0]) / (((w
+ b
[2]) * w
+ b
[1]) * w
+ b
[0]);
93 v64df Nf
= __builtin_convertvector (N
, v64df
);
94 VECTOR_COND_MOVE (z
, (Nf
* C2
+ z
) + Nf
* C1
, N
!= 0);
96 VECTOR_RETURN (z
, NO_COND
);
101 DEF_VARIANTS (log
, df
, df
)
103 DEF_VD_MATH_FUNC (v64df
, log1p
, v64df x
)
105 /* TODO: Implement algorithm with better precision. */
106 return v64df_log_aux (1 + x
, __mask
);
109 DEF_VARIANTS (log1p
, df
, df
)