• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2017 The Abseil Authors.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //      https://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 
15 #ifndef ABSL_RANDOM_INTERNAL_FAST_UNIFORM_BITS_H_
16 #define ABSL_RANDOM_INTERNAL_FAST_UNIFORM_BITS_H_
17 
18 #include <cstddef>
19 #include <cstdint>
20 #include <limits>
21 #include <type_traits>
22 
23 #include "absl/base/config.h"
24 #include "absl/meta/type_traits.h"
25 
26 namespace absl {
27 ABSL_NAMESPACE_BEGIN
28 namespace random_internal {
29 // Returns true if the input value is zero or a power of two. Useful for
30 // determining if the range of output values in a URBG
31 template <typename UIntType>
IsPowerOfTwoOrZero(UIntType n)32 constexpr bool IsPowerOfTwoOrZero(UIntType n) {
33   return (n == 0) || ((n & (n - 1)) == 0);
34 }
35 
36 // Computes the length of the range of values producible by the URBG, or returns
37 // zero if that would encompass the entire range of representable values in
38 // URBG::result_type.
39 template <typename URBG>
RangeSize()40 constexpr typename URBG::result_type RangeSize() {
41   using result_type = typename URBG::result_type;
42   static_assert((URBG::max)() != (URBG::min)(), "URBG range cannot be 0.");
43   return ((URBG::max)() == (std::numeric_limits<result_type>::max)() &&
44           (URBG::min)() == std::numeric_limits<result_type>::lowest())
45              ? result_type{0}
46              : ((URBG::max)() - (URBG::min)() + result_type{1});
47 }
48 
49 // Computes the floor of the log. (i.e., std::floor(std::log2(N));
50 template <typename UIntType>
IntegerLog2(UIntType n)51 constexpr UIntType IntegerLog2(UIntType n) {
52   return (n <= 1) ? 0 : 1 + IntegerLog2(n >> 1);
53 }
54 
55 // Returns the number of bits of randomness returned through
56 // `PowerOfTwoVariate(urbg)`.
57 template <typename URBG>
NumBits()58 constexpr size_t NumBits() {
59   return RangeSize<URBG>() == 0
60              ? std::numeric_limits<typename URBG::result_type>::digits
61              : IntegerLog2(RangeSize<URBG>());
62 }
63 
64 // Given a shift value `n`, constructs a mask with exactly the low `n` bits set.
65 // If `n == 0`, all bits are set.
66 template <typename UIntType>
MaskFromShift(size_t n)67 constexpr UIntType MaskFromShift(size_t n) {
68   return ((n % std::numeric_limits<UIntType>::digits) == 0)
69              ? ~UIntType{0}
70              : (UIntType{1} << n) - UIntType{1};
71 }
72 
73 // Tags used to dispatch FastUniformBits::generate to the simple or more complex
74 // entropy extraction algorithm.
75 struct SimplifiedLoopTag {};
76 struct RejectionLoopTag {};
77 
78 // FastUniformBits implements a fast path to acquire uniform independent bits
79 // from a type which conforms to the [rand.req.urbg] concept.
80 // Parameterized by:
81 //  `UIntType`: the result (output) type
82 //
83 // The std::independent_bits_engine [rand.adapt.ibits] adaptor can be
84 // instantiated from an existing generator through a copy or a move. It does
85 // not, however, facilitate the production of pseudorandom bits from an un-owned
86 // generator that will outlive the std::independent_bits_engine instance.
87 template <typename UIntType = uint64_t>
88 class FastUniformBits {
89  public:
90   using result_type = UIntType;
91 
result_type(min)92   static constexpr result_type(min)() { return 0; }
result_type(max)93   static constexpr result_type(max)() {
94     return (std::numeric_limits<result_type>::max)();
95   }
96 
97   template <typename URBG>
98   result_type operator()(URBG& g);  // NOLINT(runtime/references)
99 
100  private:
101   static_assert(std::is_unsigned<UIntType>::value,
102                 "Class-template FastUniformBits<> must be parameterized using "
103                 "an unsigned type.");
104 
105   // Generate() generates a random value, dispatched on whether
106   // the underlying URBG must use rejection sampling to generate a value,
107   // or whether a simplified loop will suffice.
108   template <typename URBG>
109   result_type Generate(URBG& g,  // NOLINT(runtime/references)
110                        SimplifiedLoopTag);
111 
112   template <typename URBG>
113   result_type Generate(URBG& g,  // NOLINT(runtime/references)
114                        RejectionLoopTag);
115 };
116 
117 template <typename UIntType>
118 template <typename URBG>
119 typename FastUniformBits<UIntType>::result_type
operator()120 FastUniformBits<UIntType>::operator()(URBG& g) {  // NOLINT(runtime/references)
121   // kRangeMask is the mask used when sampling variates from the URBG when the
122   // width of the URBG range is not a power of 2.
123   // Y = (2 ^ kRange) - 1
124   static_assert((URBG::max)() > (URBG::min)(),
125                 "URBG::max and URBG::min may not be equal.");
126 
127   using tag = absl::conditional_t<IsPowerOfTwoOrZero(RangeSize<URBG>()),
128                                   SimplifiedLoopTag, RejectionLoopTag>;
129   return Generate(g, tag{});
130 }
131 
132 template <typename UIntType>
133 template <typename URBG>
134 typename FastUniformBits<UIntType>::result_type
Generate(URBG & g,SimplifiedLoopTag)135 FastUniformBits<UIntType>::Generate(URBG& g,  // NOLINT(runtime/references)
136                                     SimplifiedLoopTag) {
137   // The simplified version of FastUniformBits works only on URBGs that have
138   // a range that is a power of 2. In this case we simply loop and shift without
139   // attempting to balance the bits across calls.
140   static_assert(IsPowerOfTwoOrZero(RangeSize<URBG>()),
141                 "incorrect Generate tag for URBG instance");
142 
143   static constexpr size_t kResultBits =
144       std::numeric_limits<result_type>::digits;
145   static constexpr size_t kUrbgBits = NumBits<URBG>();
146   static constexpr size_t kIters =
147       (kResultBits / kUrbgBits) + (kResultBits % kUrbgBits != 0);
148   static constexpr size_t kShift = (kIters == 1) ? 0 : kUrbgBits;
149   static constexpr auto kMin = (URBG::min)();
150 
151   result_type r = static_cast<result_type>(g() - kMin);
152   for (size_t n = 1; n < kIters; ++n) {
153     r = (r << kShift) + static_cast<result_type>(g() - kMin);
154   }
155   return r;
156 }
157 
158 template <typename UIntType>
159 template <typename URBG>
160 typename FastUniformBits<UIntType>::result_type
Generate(URBG & g,RejectionLoopTag)161 FastUniformBits<UIntType>::Generate(URBG& g,  // NOLINT(runtime/references)
162                                     RejectionLoopTag) {
163   static_assert(!IsPowerOfTwoOrZero(RangeSize<URBG>()),
164                 "incorrect Generate tag for URBG instance");
165   using urbg_result_type = typename URBG::result_type;
166 
167   // See [rand.adapt.ibits] for more details on the constants calculated below.
168   //
169   // It is preferable to use roughly the same number of bits from each generator
170   // call, however this is only possible when the number of bits provided by the
171   // URBG is a divisor of the number of bits in `result_type`. In all other
172   // cases, the number of bits used cannot always be the same, but it can be
173   // guaranteed to be off by at most 1. Thus we run two loops, one with a
174   // smaller bit-width size (`kSmallWidth`) and one with a larger width size
175   // (satisfying `kLargeWidth == kSmallWidth + 1`). The loops are run
176   // `kSmallIters` and `kLargeIters` times respectively such
177   // that
178   //
179   //    `kResultBits == kSmallIters * kSmallBits
180   //                    + kLargeIters * kLargeBits`
181   //
182   // where `kResultBits` is the total number of bits in `result_type`.
183   //
184   static constexpr size_t kResultBits =
185       std::numeric_limits<result_type>::digits;                      // w
186   static constexpr urbg_result_type kUrbgRange = RangeSize<URBG>();  // R
187   static constexpr size_t kUrbgBits = NumBits<URBG>();               // m
188 
189   // compute the initial estimate of the bits used.
190   // [rand.adapt.ibits] 2 (c)
191   static constexpr size_t kA =  // ceil(w/m)
192       (kResultBits / kUrbgBits) + ((kResultBits % kUrbgBits) != 0);  // n'
193 
194   static constexpr size_t kABits = kResultBits / kA;  // w0'
195   static constexpr urbg_result_type kARejection =
196       ((kUrbgRange >> kABits) << kABits);  // y0'
197 
198   // refine the selection to reduce the rejection frequency.
199   static constexpr size_t kTotalIters =
200       ((kUrbgRange - kARejection) <= (kARejection / kA)) ? kA : (kA + 1);  // n
201 
202   // [rand.adapt.ibits] 2 (b)
203   static constexpr size_t kSmallIters =
204       kTotalIters - (kResultBits % kTotalIters);                   // n0
205   static constexpr size_t kSmallBits = kResultBits / kTotalIters;  // w0
206   static constexpr urbg_result_type kSmallRejection =
207       ((kUrbgRange >> kSmallBits) << kSmallBits);  // y0
208 
209   static constexpr size_t kLargeBits = kSmallBits + 1;  // w0+1
210   static constexpr urbg_result_type kLargeRejection =
211       ((kUrbgRange >> kLargeBits) << kLargeBits);  // y1
212 
213   //
214   // Because `kLargeBits == kSmallBits + 1`, it follows that
215   //
216   //     `kResultBits == kSmallIters * kSmallBits + kLargeIters`
217   //
218   // and therefore
219   //
220   //     `kLargeIters == kTotalWidth % kSmallWidth`
221   //
222   // Intuitively, each iteration with the large width accounts for one unit
223   // of the remainder when `kTotalWidth` is divided by `kSmallWidth`. As
224   // mentioned above, if the URBG width is a divisor of `kTotalWidth`, then
225   // there would be no need for any large iterations (i.e., one loop would
226   // suffice), and indeed, in this case, `kLargeIters` would be zero.
227   static_assert(kResultBits == kSmallIters * kSmallBits +
228                                    (kTotalIters - kSmallIters) * kLargeBits,
229                 "Error in looping constant calculations.");
230 
231   // The small shift is essentially small bits, but due to the potential
232   // of generating a smaller result_type from a larger urbg type, the actual
233   // shift might be 0.
234   static constexpr size_t kSmallShift = kSmallBits % kResultBits;
235   static constexpr auto kSmallMask =
236       MaskFromShift<urbg_result_type>(kSmallShift);
237   static constexpr size_t kLargeShift = kLargeBits % kResultBits;
238   static constexpr auto kLargeMask =
239       MaskFromShift<urbg_result_type>(kLargeShift);
240 
241   static constexpr auto kMin = (URBG::min)();
242 
243   result_type s = 0;
244   for (size_t n = 0; n < kSmallIters; ++n) {
245     urbg_result_type v;
246     do {
247       v = g() - kMin;
248     } while (v >= kSmallRejection);
249 
250     s = (s << kSmallShift) + static_cast<result_type>(v & kSmallMask);
251   }
252 
253   for (size_t n = kSmallIters; n < kTotalIters; ++n) {
254     urbg_result_type v;
255     do {
256       v = g() - kMin;
257     } while (v >= kLargeRejection);
258 
259     s = (s << kLargeShift) + static_cast<result_type>(v & kLargeMask);
260   }
261   return s;
262 }
263 
264 }  // namespace random_internal
265 ABSL_NAMESPACE_END
266 }  // namespace absl
267 
268 #endif  // ABSL_RANDOM_INTERNAL_FAST_UNIFORM_BITS_H_
269