1 /*
2  *  Copyright 2008-2013 NVIDIA Corporation
3  *  Copyright 2013 Filipe RNC Maia
4  *
5  *  Licensed under the Apache License, Version 2.0 (the "License");
6  *  you may not use this file except in compliance with the License.
7  *  You may obtain a copy of the License at
8  *
9  *      http://www.apache.org/licenses/LICENSE-2.0
10  *
11  *  Unless required by applicable law or agreed to in writing, software
12  *  distributed under the License is distributed on an "AS IS" BASIS,
13  *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14  *  See the License for the specific language governing permissions and
15  *  limitations under the License.
16  */
17 
18 /*-
19  * Copyright (c) 2011 David Schultz <das@FreeBSD.ORG>
20  * All rights reserved.
21  *
22  * Redistribution and use in source and binary forms, with or without
23  * modification, are permitted provided that the following conditions
24  * are met:
25  * 1. Redistributions of source code must retain the above copyright
26  *    notice, this list of conditions and the following disclaimer.
27  * 2. Redistributions in binary form must reproduce the above copyright
28  *    notice, this list of conditions and the following disclaimer in the
29  *    documentation and/or other materials provided with the distribution.
30  *
31  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
32  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
35  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41  * SUCH DAMAGE.
42  */
43 
44 /* adapted from FreeBSD:
45  *    lib/msun/src/s_cexpf.c
46  *    lib/msun/src/k_exp.c
47  *
48  */
49 
50 #pragma once
51 
52 #include <thrust/complex.h>
53 #include <thrust/detail/complex/math_private.h>
54 
55 namespace thrust{
56 namespace detail{
57 namespace complex{
58 
59 __host__ __device__ inline
frexp_expf(float x,int * expt)60 float frexp_expf(float x, int *expt){
61   const uint32_t k = 235;                 /* constant for reduction */
62   const float kln2 =  162.88958740F;       /* k * ln2 */
63 
64   // should this be a double instead?
65   float exp_x;
66   uint32_t hx;
67 
68   exp_x = expf(x - kln2);
69   get_float_word(hx, exp_x);
70   *expt = (hx >> 23) - (0x7f + 127) + k;
71   set_float_word(exp_x, (hx & 0x7fffff) | ((0x7f + 127) << 23));
72   return (exp_x);
73 }
74 
75 __host__ __device__ inline
76 complex<float>
ldexp_cexpf(complex<float> z,int expt)77 ldexp_cexpf(complex<float> z, int expt)
78 {
79   float x, y, exp_x, scale1, scale2;
80   int ex_expt, half_expt;
81 
82   x = z.real();
83   y = z.imag();
84   exp_x = frexp_expf(x, &ex_expt);
85   expt += ex_expt;
86 
87   half_expt = expt / 2;
88   set_float_word(scale1, (0x7f + half_expt) << 23);
89   half_expt = expt - half_expt;
90   set_float_word(scale2, (0x7f + half_expt) << 23);
91 
92   return (complex<float>(std::cos(y) * exp_x * scale1 * scale2,
93 			 std::sin(y) * exp_x * scale1 * scale2));
94 }
95 
96 __host__ __device__ inline
cexpf(const complex<float> & z)97 complex<float> cexpf(const complex<float>& z){
98   float x, y, exp_x;
99   uint32_t hx, hy;
100 
101   const uint32_t
102     exp_ovfl  = 0x42b17218,		/* MAX_EXP * ln2 ~= 88.722839355 */
103     cexp_ovfl = 0x43400074;		/* (MAX_EXP - MIN_DENORM_EXP) * ln2 */
104 
105   x = z.real();
106   y = z.imag();
107 
108   get_float_word(hy, y);
109   hy &= 0x7fffffff;
110 
111   /* cexp(x + I 0) = exp(x) + I 0 */
112   if (hy == 0)
113     return (complex<float>(std::exp(x), y));
114   get_float_word(hx, x);
115   /* cexp(0 + I y) = cos(y) + I sin(y) */
116   if ((hx & 0x7fffffff) == 0){
117     return (complex<float>(std::cos(y), std::sin(y)));
118   }
119   if (hy >= 0x7f800000) {
120     if ((hx & 0x7fffffff) != 0x7f800000) {
121       /* cexp(finite|NaN +- I Inf|NaN) = NaN + I NaN */
122       return (complex<float>(y - y, y - y));
123     } else if (hx & 0x80000000) {
124       /* cexp(-Inf +- I Inf|NaN) = 0 + I 0 */
125       return (complex<float>(0.0, 0.0));
126     } else {
127       /* cexp(+Inf +- I Inf|NaN) = Inf + I NaN */
128       return (complex<float>(x, y - y));
129     }
130   }
131 
132   if (hx >= exp_ovfl && hx <= cexp_ovfl) {
133     /*
134      * x is between 88.7 and 192, so we must scale to avoid
135      * overflow in expf(x).
136      */
137     return (ldexp_cexpf(z, 0));
138   } else {
139     /*
140      * Cases covered here:
141      *  -  x < exp_ovfl and exp(x) won't overflow (common case)
142      *  -  x > cexp_ovfl, so exp(x) * s overflows for all s > 0
143      *  -  x = +-Inf (generated by exp())
144      *  -  x = NaN (spurious inexact exception from y)
145      */
146     exp_x = std::exp(x);
147     return (complex<float>(exp_x * std::cos(y), exp_x * std::sin(y)));
148   }
149 }
150 
151 } // namespace complex
152 
153 } // namespace detail
154 
155 template <>
156 __host__ __device__
exp(const complex<float> & z)157 inline complex<float> exp(const complex<float>& z){
158   return detail::complex::cexpf(z);
159 }
160 
161 } // namespace thrust
162