newlib/libm/machine/amdgcn/v64sf_sine.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/sf_sine.c in Newlib.  */
  38
  39 #include "amdgcnmach.h"
  40
  41 v64si v64sf_numtestf (v64sf);
  42
  43 static const float HALF_PI = 1.570796326;
  44 static const float ONE_OVER_PI = 0.318309886;
  45 static const float r[] = { -0.1666665668,
  46                             0.8333025139e-02,
  47                            -0.1980741872e-03,
  48                             0.2601903036e-5 };
  49
  50 #if defined (__has_builtin) && __has_builtin (__builtin_gcn_fabsvf)
  51
  52 DEF_VS_MATH_FUNC (v64sf, sinef, v64sf x, int cosine)
  53 {
  54   const float YMAX = 210828714.0;
  55
  56   FUNCTION_INIT (v64sf);
  57
  58   v64si num_type = v64sf_numtestf (x);
  59   VECTOR_IF (num_type == NAN, cond)
  60     errno = EDOM;
  61     VECTOR_RETURN (x, cond);
  62   VECTOR_ELSEIF (num_type == INF, cond)
  63     errno = EDOM;
  64     VECTOR_RETURN (VECTOR_INIT (z_notanum_f.f), cond);
  65   VECTOR_ENDIF
  66
  67   /* Use sin and cos properties to ease computations. */
  68   v64si sgn;
  69   v64sf y;
  70
  71   if (cosine)
  72     {
  73       sgn = VECTOR_INIT (0);
  74       y = __builtin_gcn_fabsvf (x) + HALF_PI;
  75     }
  76   else
  77     {
  78       sgn = x < 0.0f;
  79       y = VECTOR_MERGE (-x, x, x < 0.0f);
  80     }
  81
  82   /* Check for values of y that will overflow here. */
  83   VECTOR_IF (y > YMAX, cond)
  84     errno = ERANGE;
  85     VECTOR_RETURN (x, cond);
  86   VECTOR_ENDIF
  87
  88   /* Calculate the exponent. */
  89   v64si Nneg = __builtin_convertvector (y * ONE_OVER_PI - 0.5f, v64si);
  90   v64si Npos = __builtin_convertvector (y * ONE_OVER_PI + 0.5f, v64si);
  91   v64si N = VECTOR_MERGE (Nneg, Npos, y < 0.0f);
  92   v64sf XN = __builtin_convertvector (N, v64sf);
  93
  94   VECTOR_COND_MOVE (sgn, ~sgn, (N & 1) != 0);
  95
  96   if (cosine)
  97     XN -= 0.5;
  98
  99   y = __builtin_gcn_fabsvf (x) - XN * (float) __PI;
 100
 101   v64sf res;
 102
 103   VECTOR_IF ((-z_rooteps_f < y) & (y < z_rooteps_f), cond)
 104     VECTOR_COND_MOVE (res, y, cond);
 105   VECTOR_ELSE (cond)
 106     v64sf g = y * y;
 107
 108     /* Calculate the Taylor series. */
 109     v64sf R = (((r[3] * g + r[2]) * g + r[1]) * g + r[0]) * g;
 110
 111     /* Finally, compute the result. */
 112     VECTOR_COND_MOVE (res, y + y * R, cond);
 113   VECTOR_ENDIF
 114
 115   VECTOR_COND_MOVE (res, -res, sgn);
 116
 117   VECTOR_RETURN (res, NO_COND);
 118
 119   FUNCTION_RETURN;
 120 }
 121
 122 #endif