1// Optimizations for random number functions, x86 version -*- C++ -*-
2
3// Copyright (C) 2012-2020 Free Software Foundation, Inc.
4//
5// This file is part of the GNU ISO C++ Library.  This library is free
6// software; you can redistribute it and/or modify it under the
7// terms of the GNU General Public License as published by the
8// Free Software Foundation; either version 3, or (at your option)
9// any later version.
10
11// This library is distributed in the hope that it will be useful,
12// but WITHOUT ANY WARRANTY; without even the implied warranty of
13// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14// GNU General Public License for more details.
15
16// Under Section 7 of GPL version 3, you are granted additional
17// permissions described in the GCC Runtime Library Exception, version
18// 3.1, as published by the Free Software Foundation.
19
20// You should have received a copy of the GNU General Public License and
21// a copy of the GCC Runtime Library Exception along with this program;
22// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
23// <http://www.gnu.org/licenses/>.
24
25/** @file bits/opt_random.h
26 *  This is an internal header file, included by other library headers.
27 *  Do not attempt to use it directly. @headername{random}
28 */
29
30#ifndef _BITS_OPT_RANDOM_H
31#define _BITS_OPT_RANDOM_H 1
32
33#ifdef __SSE3__
34#include <pmmintrin.h>
35#endif
36
37
38#pragma GCC system_header
39
40
41namespace std _GLIBCXX_VISIBILITY(default)
42{
43_GLIBCXX_BEGIN_NAMESPACE_VERSION
44
45#ifdef __SSE3__
46  template<>
47    template<typename _UniformRandomNumberGenerator>
48      void
49      normal_distribution<double>::
50      __generate(typename normal_distribution<double>::result_type* __f,
51		 typename normal_distribution<double>::result_type* __t,
52		 _UniformRandomNumberGenerator& __urng,
53		 const param_type& __param)
54      {
55	typedef uint64_t __uctype;
56
57	if (__f == __t)
58	  return;
59
60	if (_M_saved_available)
61	  {
62	    _M_saved_available = false;
63	    *__f++ = _M_saved * __param.stddev() + __param.mean();
64
65	    if (__f == __t)
66	      return;
67	  }
68
69	constexpr uint64_t __maskval = 0xfffffffffffffull;
70	static const __m128i __mask = _mm_set1_epi64x(__maskval);
71	static const __m128i __two = _mm_set1_epi64x(0x4000000000000000ull);
72	static const __m128d __three = _mm_set1_pd(3.0);
73	const __m128d __av = _mm_set1_pd(__param.mean());
74
75	const __uctype __urngmin = __urng.min();
76	const __uctype __urngmax = __urng.max();
77	const __uctype __urngrange = __urngmax - __urngmin;
78	const __uctype __uerngrange = __urngrange + 1;
79
80	while (__f + 1 < __t)
81	  {
82	    double __le;
83	    __m128d __x;
84	    do
85	      {
86                union
87                {
88                  __m128i __i;
89                  __m128d __d;
90		} __v;
91
92		if (__urngrange > __maskval)
93		  {
94		    if (__detail::_Power_of_2(__uerngrange))
95		      __v.__i = _mm_and_si128(_mm_set_epi64x(__urng(),
96							     __urng()),
97					      __mask);
98		    else
99		      {
100			const __uctype __uerange = __maskval + 1;
101			const __uctype __scaling = __urngrange / __uerange;
102			const __uctype __past = __uerange * __scaling;
103			uint64_t __v1;
104			do
105			  __v1 = __uctype(__urng()) - __urngmin;
106			while (__v1 >= __past);
107			__v1 /= __scaling;
108			uint64_t __v2;
109			do
110			  __v2 = __uctype(__urng()) - __urngmin;
111			while (__v2 >= __past);
112			__v2 /= __scaling;
113
114			__v.__i = _mm_set_epi64x(__v1, __v2);
115		      }
116		  }
117		else if (__urngrange == __maskval)
118		  __v.__i = _mm_set_epi64x(__urng(), __urng());
119		else if ((__urngrange + 2) * __urngrange >= __maskval
120			 && __detail::_Power_of_2(__uerngrange))
121		  {
122		    uint64_t __v1 = __urng() * __uerngrange + __urng();
123		    uint64_t __v2 = __urng() * __uerngrange + __urng();
124
125		    __v.__i = _mm_and_si128(_mm_set_epi64x(__v1, __v2),
126					    __mask);
127		  }
128		else
129		  {
130		    size_t __nrng = 2;
131		    __uctype __high = __maskval / __uerngrange / __uerngrange;
132		    while (__high > __uerngrange)
133		      {
134			++__nrng;
135			__high /= __uerngrange;
136		      }
137		    const __uctype __highrange = __high + 1;
138		    const __uctype __scaling = __urngrange / __highrange;
139		    const __uctype __past = __highrange * __scaling;
140		    __uctype __tmp;
141
142		    uint64_t __v1;
143		    do
144		      {
145			do
146			  __tmp = __uctype(__urng()) - __urngmin;
147			while (__tmp >= __past);
148			__v1 = __tmp / __scaling;
149			for (size_t __cnt = 0; __cnt < __nrng; ++__cnt)
150			  {
151			    __tmp = __v1;
152			    __v1 *= __uerngrange;
153			    __v1 += __uctype(__urng()) - __urngmin;
154			  }
155		      }
156		    while (__v1 > __maskval || __v1 < __tmp);
157
158		    uint64_t __v2;
159		    do
160		      {
161			do
162			  __tmp = __uctype(__urng()) - __urngmin;
163			while (__tmp >= __past);
164			__v2 = __tmp / __scaling;
165			for (size_t __cnt = 0; __cnt < __nrng; ++__cnt)
166			  {
167			    __tmp = __v2;
168			    __v2 *= __uerngrange;
169			    __v2 += __uctype(__urng()) - __urngmin;
170			  }
171		      }
172		    while (__v2 > __maskval || __v2 < __tmp);
173
174		    __v.__i = _mm_set_epi64x(__v1, __v2);
175		  }
176
177		__v.__i = _mm_or_si128(__v.__i, __two);
178		__x = _mm_sub_pd(__v.__d, __three);
179		__m128d __m = _mm_mul_pd(__x, __x);
180		__le = _mm_cvtsd_f64(_mm_hadd_pd (__m, __m));
181              }
182            while (__le == 0.0 || __le >= 1.0);
183
184            double __mult = (std::sqrt(-2.0 * std::log(__le) / __le)
185                             * __param.stddev());
186
187            __x = _mm_add_pd(_mm_mul_pd(__x, _mm_set1_pd(__mult)), __av);
188
189            _mm_storeu_pd(__f, __x);
190            __f += 2;
191          }
192
193        if (__f != __t)
194          {
195            result_type __x, __y, __r2;
196
197            __detail::_Adaptor<_UniformRandomNumberGenerator, result_type>
198              __aurng(__urng);
199
200            do
201              {
202                __x = result_type(2.0) * __aurng() - 1.0;
203                __y = result_type(2.0) * __aurng() - 1.0;
204                __r2 = __x * __x + __y * __y;
205              }
206            while (__r2 > 1.0 || __r2 == 0.0);
207
208            const result_type __mult = std::sqrt(-2 * std::log(__r2) / __r2);
209            _M_saved = __x * __mult;
210            _M_saved_available = true;
211            *__f = __y * __mult * __param.stddev() + __param.mean();
212          }
213      }
214#endif
215
216
217_GLIBCXX_END_NAMESPACE_VERSION
218} // namespace
219
220
221#endif // _BITS_OPT_RANDOM_H
222