| 1 |
9 |
jlechner |
/* -------------------------------------------------------------- */
|
| 2 |
|
|
/* (C)Copyright 2007,2008, */
|
| 3 |
|
|
/* International Business Machines Corporation */
|
| 4 |
|
|
/* All Rights Reserved. */
|
| 5 |
|
|
/* */
|
| 6 |
|
|
/* Redistribution and use in source and binary forms, with or */
|
| 7 |
|
|
/* without modification, are permitted provided that the */
|
| 8 |
|
|
/* following conditions are met: */
|
| 9 |
|
|
/* */
|
| 10 |
|
|
/* - Redistributions of source code must retain the above copyright*/
|
| 11 |
|
|
/* notice, this list of conditions and the following disclaimer. */
|
| 12 |
|
|
/* */
|
| 13 |
|
|
/* - Redistributions in binary form must reproduce the above */
|
| 14 |
|
|
/* copyright notice, this list of conditions and the following */
|
| 15 |
|
|
/* disclaimer in the documentation and/or other materials */
|
| 16 |
|
|
/* provided with the distribution. */
|
| 17 |
|
|
/* */
|
| 18 |
|
|
/* - Neither the name of IBM Corporation nor the names of its */
|
| 19 |
|
|
/* contributors may be used to endorse or promote products */
|
| 20 |
|
|
/* derived from this software without specific prior written */
|
| 21 |
|
|
/* permission. */
|
| 22 |
|
|
/* */
|
| 23 |
|
|
/* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND */
|
| 24 |
|
|
/* CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, */
|
| 25 |
|
|
/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
|
| 26 |
|
|
/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
|
| 27 |
|
|
/* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR */
|
| 28 |
|
|
/* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, */
|
| 29 |
|
|
/* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT */
|
| 30 |
|
|
/* NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; */
|
| 31 |
|
|
/* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) */
|
| 32 |
|
|
/* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN */
|
| 33 |
|
|
/* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR */
|
| 34 |
|
|
/* OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, */
|
| 35 |
|
|
/* EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */
|
| 36 |
|
|
/* -------------------------------------------------------------- */
|
| 37 |
|
|
/* PROLOG END TAG zYx */
|
| 38 |
|
|
#ifdef __SPU__
|
| 39 |
|
|
#ifndef _TANHD2_H_
|
| 40 |
|
|
#define _TANHD2_H_ 1
|
| 41 |
|
|
|
| 42 |
|
|
#include <spu_intrinsics.h>
|
| 43 |
|
|
|
| 44 |
|
|
#include "expd2.h"
|
| 45 |
|
|
#include "divd2.h"
|
| 46 |
|
|
|
| 47 |
|
|
|
| 48 |
|
|
/*
|
| 49 |
|
|
* Taylor coefficients for tanh
|
| 50 |
|
|
*/
|
| 51 |
|
|
#define TANH_TAY01 1.000000000000000000000000000000E0
|
| 52 |
|
|
#define TANH_TAY02 -3.333333333333333333333333333333E-1
|
| 53 |
|
|
#define TANH_TAY03 1.333333333333333333333333333333E-1
|
| 54 |
|
|
#define TANH_TAY04 -5.396825396825396825396825396825E-2
|
| 55 |
|
|
#define TANH_TAY05 2.186948853615520282186948853616E-2
|
| 56 |
|
|
#define TANH_TAY06 -8.863235529902196568863235529902E-3
|
| 57 |
|
|
#define TANH_TAY07 3.592128036572481016925461369906E-3
|
| 58 |
|
|
#define TANH_TAY08 -1.455834387051318268249485180702E-3
|
| 59 |
|
|
#define TANH_TAY09 5.900274409455859813780759937000E-4
|
| 60 |
|
|
#define TANH_TAY10 -2.391291142435524814857314588851E-4
|
| 61 |
|
|
#define TANH_TAY11 9.691537956929450325595875000389E-5
|
| 62 |
|
|
#define TANH_TAY12 -3.927832388331683405337080809312E-5
|
| 63 |
|
|
#define TANH_TAY13 1.591890506932896474074427981657E-5
|
| 64 |
|
|
#define TANH_TAY14 -6.451689215655430763190842315303E-6
|
| 65 |
|
|
#define TANH_TAY15 2.614771151290754554263594256410E-6
|
| 66 |
|
|
#define TANH_TAY16 -1.059726832010465435091355394125E-6
|
| 67 |
|
|
#define TANH_TAY17 4.294911078273805854820351280397E-7
|
| 68 |
|
|
|
| 69 |
|
|
|
| 70 |
|
|
/*
|
| 71 |
|
|
* FUNCTION
|
| 72 |
|
|
* vector double _tanhd2(vector double x)
|
| 73 |
|
|
*
|
| 74 |
|
|
* DESCRIPTION
|
| 75 |
|
|
* The _tanhd2 function computes the hyperbolic tangent for each
|
| 76 |
|
|
* element of the input vector.
|
| 77 |
|
|
*
|
| 78 |
|
|
* We use the following to approximate tanh:
|
| 79 |
|
|
*
|
| 80 |
|
|
* |x| <= .25: Taylor Series
|
| 81 |
|
|
* |x| > .25: tanh(x) = (exp(2x) - 1)/(exp(2x) + 1)
|
| 82 |
|
|
*
|
| 83 |
|
|
*
|
| 84 |
|
|
* SPECIAL CASES:
|
| 85 |
|
|
* - tanh(+/- 0) = +/-0
|
| 86 |
|
|
* - tanh(+/- infinity) = +/- 1
|
| 87 |
|
|
* - tanh(NaN) = NaN
|
| 88 |
|
|
*
|
| 89 |
|
|
*/
|
| 90 |
|
|
|
| 91 |
|
|
static __inline vector double _tanhd2(vector double x)
|
| 92 |
|
|
{
|
| 93 |
|
|
vector double signbit = spu_splats(-0.0);
|
| 94 |
|
|
vector double oned = spu_splats(1.0);
|
| 95 |
|
|
vector double twod = spu_splats(2.0);
|
| 96 |
|
|
vector double infd = (vector double)spu_splats(0x7FF0000000000000ull);
|
| 97 |
|
|
vector double xabs;
|
| 98 |
|
|
vector double x2;
|
| 99 |
|
|
vector unsigned long long gttaylor;
|
| 100 |
|
|
vector double e;
|
| 101 |
|
|
vector double tresult;
|
| 102 |
|
|
vector double eresult;
|
| 103 |
|
|
vector double result;
|
| 104 |
|
|
|
| 105 |
|
|
xabs = spu_andc(x, signbit);
|
| 106 |
|
|
|
| 107 |
|
|
/*
|
| 108 |
|
|
* This is where we switch from Taylor Series
|
| 109 |
|
|
* to exponential formula.
|
| 110 |
|
|
*/
|
| 111 |
|
|
gttaylor = spu_cmpgt(xabs, spu_splats(0.25));
|
| 112 |
|
|
|
| 113 |
|
|
|
| 114 |
|
|
/*
|
| 115 |
|
|
* Taylor Series Approximation
|
| 116 |
|
|
*/
|
| 117 |
|
|
x2 = spu_mul(x,x);
|
| 118 |
|
|
tresult = spu_madd(x2, spu_splats(TANH_TAY11), spu_splats(TANH_TAY10));
|
| 119 |
|
|
tresult = spu_madd(x2, tresult, spu_splats(TANH_TAY09));
|
| 120 |
|
|
tresult = spu_madd(x2, tresult, spu_splats(TANH_TAY08));
|
| 121 |
|
|
tresult = spu_madd(x2, tresult, spu_splats(TANH_TAY07));
|
| 122 |
|
|
tresult = spu_madd(x2, tresult, spu_splats(TANH_TAY06));
|
| 123 |
|
|
tresult = spu_madd(x2, tresult, spu_splats(TANH_TAY05));
|
| 124 |
|
|
tresult = spu_madd(x2, tresult, spu_splats(TANH_TAY04));
|
| 125 |
|
|
tresult = spu_madd(x2, tresult, spu_splats(TANH_TAY03));
|
| 126 |
|
|
tresult = spu_madd(x2, tresult, spu_splats(TANH_TAY02));
|
| 127 |
|
|
tresult = spu_madd(x2, tresult, spu_splats(TANH_TAY01));
|
| 128 |
|
|
tresult = spu_mul(xabs, tresult);
|
| 129 |
|
|
|
| 130 |
|
|
|
| 131 |
|
|
/*
|
| 132 |
|
|
* Exponential Formula
|
| 133 |
|
|
* Our expd2 function gives a more accurate result in general
|
| 134 |
|
|
* with xabs instead of x for x<0. We correct for sign later.
|
| 135 |
|
|
*/
|
| 136 |
|
|
e = _expd2(spu_mul(xabs, twod));
|
| 137 |
|
|
eresult = _divd2(spu_sub(e, oned), spu_add(e, oned));
|
| 138 |
|
|
|
| 139 |
|
|
|
| 140 |
|
|
/*
|
| 141 |
|
|
* Select Taylor or exp result.
|
| 142 |
|
|
*/
|
| 143 |
|
|
result = spu_sel(tresult, eresult, gttaylor);
|
| 144 |
|
|
|
| 145 |
|
|
/*
|
| 146 |
|
|
* Inf and NaN special cases. NaN is already in result
|
| 147 |
|
|
* for x = NaN.
|
| 148 |
|
|
*/
|
| 149 |
|
|
result = spu_sel(result, oned, spu_cmpeq(xabs, infd));
|
| 150 |
|
|
|
| 151 |
|
|
/*
|
| 152 |
|
|
* Antisymmetric function - preserve sign bit of x
|
| 153 |
|
|
* in the result.
|
| 154 |
|
|
*/
|
| 155 |
|
|
result = spu_sel(result, x, (vec_ullong2)signbit);
|
| 156 |
|
|
|
| 157 |
|
|
return result;
|
| 158 |
|
|
}
|
| 159 |
|
|
|
| 160 |
|
|
#endif /* _TANHD2_H_ */
|
| 161 |
|
|
#endif /* __SPU__ */
|
