newlib/libm/machine/amdgcn/v64df_tanh.c

   1 /*
   2  * Copyright 2023 Siemens
   3  *
   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
  12  * they apply.
  13  */
  14
  15 /*
  16  * Copyright (c) 1994-2009  Red Hat, Inc. All rights reserved.
  17  *
  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.
  28  */
  29
  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
  34  * Hall, 1980.
  35  *****************************************************************/
  36
  37 /* Based on newlib/libm/mathfp/s_tanh.c in Newlib.  */
  38
  39 #include "amdgcnmach.h"
  40
  41 v64df v64df_exp_aux (v64df, v64di);
  42
  43 static const double LN3_OVER2 = 0.54930614433405484570;
  44 static const double p[] = { -0.16134119023996228053e+4,
  45                             -0.99225929672236083313e+2,
  46                             -0.96437492777225469787 };
  47 static const double q[] = { 0.48402357071988688686e+4,
  48                             0.22337720718962312926e+4,
  49                             0.11274474380534949335e+3 };
  50
  51 #if defined (__has_builtin) && __has_builtin (__builtin_gcn_fabsv)
  52
  53 DEF_VD_MATH_FUNC (v64df, tanh, v64df x)
  54 {
  55   FUNCTION_INIT (v64df);
  56
  57   v64df f = __builtin_gcn_fabsv (x);
  58   v64df res;
  59
  60   /* Check if the input is too big. */
  61   VECTOR_IF (f > BIGX, cond)
  62     VECTOR_COND_MOVE (res, VECTOR_INIT (1.0), cond);
  63
  64   VECTOR_ELSEIF (f > LN3_OVER2, cond)
  65     VECTOR_COND_MOVE (res, 1.0 - 2.0 / (v64df_exp_aux (2 * f, __mask) + 1.0),
  66                       cond);
  67
  68   /* Check if the input is too small. */
  69   VECTOR_ELSEIF (f < z_rooteps, cond)
  70     VECTOR_COND_MOVE (res, f, cond);
  71
  72   /* Calculate the Taylor series. */
  73   VECTOR_ELSE (cond)
  74     v64df g = f * f;
  75
  76     v64df P = (p[2] * g + p[1]) * g + p[0];
  77     v64df Q = ((g + q[2]) * g + q[1]) * g + q[0];
  78     v64df R = g * (P / Q);
  79
  80     VECTOR_COND_MOVE (res, f + f * R, cond);
  81   VECTOR_ENDIF
  82
  83   VECTOR_COND_MOVE (res, -res, x < 0.0);
  84
  85   VECTOR_RETURN (res, NO_COND);
  86
  87   FUNCTION_RETURN;
  88 }
  89
  90 DEF_VARIANTS (tanh, df, df)
  91
  92 #endif