about summary refs log tree commit diff
path: root/absl/random/internal/fast_uniform_bits.h
diff options
context:
space:
mode:
authorAbseil Team <absl-team@google.com>2019-06-21T20·11-0700
committerGennadiy Rozental <rogeeff@google.com>2019-06-21T20·18-0400
commite9324d926a9189e222741fce6e676f0944661a72 (patch)
treea08568a709940c376454da34c9d8aac021378e5f /absl/random/internal/fast_uniform_bits.h
parent43ef2148c0936ebf7cb4be6b19927a9d9d145b8f (diff)
Export of internal Abseil changes.
--
7a6ff16a85beb730c172d5d25cf1b5e1be885c56 by Laramie Leavitt <lar@google.com>:

Internal change.

PiperOrigin-RevId: 254454546

--
ff8f9bafaefc26d451f576ea4a06d150aed63f6f by Andy Soffer <asoffer@google.com>:

Internal changes

PiperOrigin-RevId: 254451562

--
deefc5b651b479ce36f0b4ef203e119c0c8936f2 by CJ Johnson <johnsoncj@google.com>:

Account for subtracting unsigned values from the size of InlinedVector

PiperOrigin-RevId: 254450625

--
3c677316a27bcadc17e41957c809ca472d5fef14 by Andy Soffer <asoffer@google.com>:

Add C++17's std::make_from_tuple to absl/utility/utility.h

PiperOrigin-RevId: 254411573

--
4ee3536a918830eeec402a28fc31a62c7c90b940 by CJ Johnson <johnsoncj@google.com>:

Adds benchmark for the rest of the InlinedVector public API

PiperOrigin-RevId: 254408378

--
e5a21a00700ee83498ff1efbf649169756463ee4 by CJ Johnson <johnsoncj@google.com>:

Updates the definition of InlinedVector::shrink_to_fit() to be exception safe and adds exception safety tests for it.

PiperOrigin-RevId: 254401387

--
2ea82e72b86d82d78b4e4712a63a55981b53c64b by Laramie Leavitt <lar@google.com>:

Use absl::InsecureBitGen in place of std::mt19937
in tests absl/random/...distribution_test.cc

PiperOrigin-RevId: 254289444

--
fa099e02c413a7ffda732415e8105cad26a90337 by Andy Soffer <asoffer@google.com>:

Internal changes

PiperOrigin-RevId: 254286334

--
ce34b7f36933b30cfa35b9c9a5697a792b5666e4 by Andy Soffer <asoffer@google.com>:

Internal changes

PiperOrigin-RevId: 254273059

--
6f9c473da7c2090c2e85a37c5f00622e8a912a89 by Jorg Brown <jorg@google.com>:

Change absl::container_internal::CompressedTuple to instantiate its
internal Storage class with the name of the type it's holding, rather
than the name of the Tuple.  This is not an externally-visible change,
other than less compiler memory is used and less debug information is
generated.

PiperOrigin-RevId: 254269285

--
8bd3c186bf2fc0c55d8a2dd6f28a5327502c9fba by Andy Soffer <asoffer@google.com>:

Adding short-hand IntervalClosed for IntervalClosedClosed and IntervalOpen for
IntervalOpenOpen.

PiperOrigin-RevId: 254252419

--
ea957f99b6a04fccd42aa05605605f3b44b1ecfd by Abseil Team <absl-team@google.com>:

Do not directly use __SIZEOF_INT128__.

In order to avoid linker errors when building with clang-cl (__fixunsdfti, __udivti3 and __fixunssfti are undefined), this CL uses ABSL_HAVE_INTRINSIC_INT128 which is not defined for clang-cl.

PiperOrigin-RevId: 254250739

--
89ab385cd26b34d64130bce856253aaba96d2345 by Andy Soffer <asoffer@google.com>:

Internal changes

PiperOrigin-RevId: 254242321

--
cffc793d93eca6d6bdf7de733847b6ab4a255ae9 by CJ Johnson <johnsoncj@google.com>:

Adds benchmark for InlinedVector::reserve(size_type)

PiperOrigin-RevId: 254199226

--
c90c7a9fa3c8f0c9d5114036979548b055ea2f2a by Gennadiy Rozental <rogeeff@google.com>:

Import of CCTZ from GitHub.

PiperOrigin-RevId: 254072387

--
c4c388beae016c9570ab54ffa1d52660e4a85b7b by Laramie Leavitt <lar@google.com>:

Internal cleanup.

PiperOrigin-RevId: 254062381

--
d3c992e221cc74e5372d0c8fa410170b6a43c062 by Tom Manshreck <shreck@google.com>:

Update distributions.h to Abseil standards

PiperOrigin-RevId: 254054946

--
d15ad0035c34ef11b14fadc5a4a2d3ec415f5518 by CJ Johnson <johnsoncj@google.com>:

Removes functions with only one caller from the implementation details of InlinedVector by manually inlining the definitions

PiperOrigin-RevId: 254005427

--
2f37e807efc3a8ef1f4b539bdd379917d4151520 by Andy Soffer <asoffer@google.com>:

Initial release of Abseil Random

PiperOrigin-RevId: 253999861

--
24ed1694b6430791d781ed533a8f8ccf6cac5856 by CJ Johnson <johnsoncj@google.com>:

Updates the definition of InlinedVector::assign(...)/InlinedVector::operator=(...) to new, exception-safe implementations with exception safety tests to boot

PiperOrigin-RevId: 253993691

--
5613d95f5a7e34a535cfaeadce801441e990843e by CJ Johnson <johnsoncj@google.com>:

Adds benchmarks for InlinedVector::shrink_to_fit()

PiperOrigin-RevId: 253989647

--
2a96ddfdac40bbb8cb6a7f1aeab90917067c6e63 by Abseil Team <absl-team@google.com>:

Initial release of Abseil Random

PiperOrigin-RevId: 253927497

--
bf1aff8fc9ffa921ad74643e9525ecf25b0d8dc1 by Andy Soffer <asoffer@google.com>:

Initial release of Abseil Random

PiperOrigin-RevId: 253920512

--
bfc03f4a3dcda3cf3a4b84bdb84cda24e3394f41 by Laramie Leavitt <lar@google.com>:

Internal change.

PiperOrigin-RevId: 253886486

--
05036cfcc078ca7c5f581a00dfb0daed568cbb69 by Eric Fiselier <ericwf@google.com>:

Don't include `winsock2.h` because it drags in `windows.h` and friends,
and they define awful macros like OPAQUE, ERROR, and more. This has the
potential to break abseil users.

Instead we only forward declare `timeval` and require Windows users
include `winsock2.h` themselves. This is both inconsistent and poor QoI, but so
including 'windows.h' is bad too.

PiperOrigin-RevId: 253852615
GitOrigin-RevId: 7a6ff16a85beb730c172d5d25cf1b5e1be885c56
Change-Id: Icd6aff87da26f29ec8915da856f051129987cef6
Diffstat (limited to 'absl/random/internal/fast_uniform_bits.h')
-rw-r--r--absl/random/internal/fast_uniform_bits.h299
1 files changed, 299 insertions, 0 deletions
diff --git a/absl/random/internal/fast_uniform_bits.h b/absl/random/internal/fast_uniform_bits.h
new file mode 100644
index 000000000000..23eabbc8444c
--- /dev/null
+++ b/absl/random/internal/fast_uniform_bits.h
@@ -0,0 +1,299 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef ABSL_RANDOM_INTERNAL_FAST_UNIFORM_BITS_H_
+#define ABSL_RANDOM_INTERNAL_FAST_UNIFORM_BITS_H_
+
+#include <cstddef>
+#include <cstdint>
+#include <limits>
+#include <type_traits>
+
+namespace absl {
+namespace random_internal {
+// Computes the length of the range of values producible by the URBG, or returns
+// zero if that would encompass the entire range of representable values in
+// URBG::result_type.
+template <typename URBG>
+constexpr typename URBG::result_type constexpr_range() {
+  using result_type = typename URBG::result_type;
+  return ((URBG::max)() == (std::numeric_limits<result_type>::max)() &&
+          (URBG::min)() == std::numeric_limits<result_type>::lowest())
+             ? result_type{0}
+             : (URBG::max)() - (URBG::min)() + result_type{1};
+}
+
+// FastUniformBits implements a fast path to acquire uniform independent bits
+// from a type which conforms to the [rand.req.urbg] concept.
+// Parameterized by:
+//  `UIntType`: the result (output) type
+//  `Width`: binary output width
+//
+// The std::independent_bits_engine [rand.adapt.ibits] adaptor can be
+// instantiated from an existing generator through a copy or a move. It does
+// not, however, facilitate the production of pseudorandom bits from an un-owned
+// generator that will outlive the std::independent_bits_engine instance.
+template <typename UIntType = uint64_t,
+          size_t Width = std::numeric_limits<UIntType>::digits>
+class FastUniformBits {
+  static_assert(std::is_unsigned<UIntType>::value,
+                "Class-template FastUniformBits<> must be parameterized using "
+                "an unsigned type.");
+
+  // `kWidth` is the width, in binary digits, of the output. By default it is
+  // the number of binary digits in the `result_type`.
+  static constexpr size_t kWidth = Width;
+  static_assert(kWidth > 0,
+                "Class-template FastUniformBits<> Width argument must be > 0");
+
+  static_assert(kWidth <= std::numeric_limits<UIntType>::digits,
+                "Class-template FastUniformBits<> Width argument must be <= "
+                "width of UIntType.");
+
+  static constexpr bool kIsMaxWidth =
+      (kWidth >= std::numeric_limits<UIntType>::digits);
+
+  // Computes a mask of `n` bits for the `UIntType`.
+  static constexpr UIntType constexpr_mask(size_t n) {
+    return (UIntType(1) << n) - 1;
+  }
+
+ public:
+  using result_type = UIntType;
+
+  static constexpr result_type(min)() { return 0; }
+  static constexpr result_type(max)() {
+    return kIsMaxWidth ? (std::numeric_limits<result_type>::max)()
+                       : constexpr_mask(kWidth);
+  }
+
+  template <typename URBG>
+  result_type operator()(URBG& g);  // NOLINT(runtime/references)
+
+ private:
+  // Variate() generates a single random variate, always returning a value
+  // in the closed interval [0 ... FastUniformBitsURBGConstants::kRangeMask]
+  // (kRangeMask+1 is a power of 2).
+  template <typename URBG>
+  typename URBG::result_type Variate(URBG& g);  // NOLINT(runtime/references)
+
+  // generate() generates a random value, dispatched on whether
+  // the underlying URNG must loop over multiple calls or not.
+  template <typename URBG>
+  result_type Generate(URBG& g,  // NOLINT(runtime/references)
+                       std::true_type /* avoid_looping */);
+
+  template <typename URBG>
+  result_type Generate(URBG& g,  // NOLINT(runtime/references)
+                       std::false_type /* avoid_looping */);
+};
+
+// FastUniformBitsURBGConstants computes the URBG-derived constants used
+// by FastUniformBits::Generate and FastUniformBits::Variate.
+// Parameterized by the FastUniformBits parameter:
+//   `URBG`: The underlying UniformRandomNumberGenerator.
+//
+// The values here indicate the URBG range as well as providing an indicator
+// whether the URBG output is a power of 2, and kRangeMask, which allows masking
+// the generated output to kRangeBits.
+template <typename URBG>
+class FastUniformBitsURBGConstants {
+  // Computes the floor of the log. (i.e., std::floor(std::log2(N));
+  static constexpr size_t constexpr_log2(size_t n) {
+    return (n <= 1) ? 0 : 1 + constexpr_log2(n / 2);
+  }
+
+  // Computes a mask of n bits for the URBG::result_type.
+  static constexpr typename URBG::result_type constexpr_mask(size_t n) {
+    return (typename URBG::result_type(1) << n) - 1;
+  }
+
+ public:
+  using result_type = typename URBG::result_type;
+
+  // The range of the URNG, max - min + 1, or zero if that result would cause
+  // overflow.
+  static constexpr result_type kRange = constexpr_range<URBG>();
+
+  static constexpr bool kPowerOfTwo =
+      (kRange == 0) || ((kRange & (kRange - 1)) == 0);
+
+  // kRangeBits describes the number number of bits suitable to mask off of URNG
+  // variate, which is:
+  // kRangeBits = floor(log2(kRange))
+  static constexpr size_t kRangeBits =
+      kRange == 0 ? std::numeric_limits<result_type>::digits
+                  : constexpr_log2(kRange);
+
+  // kRangeMask is the mask used when sampling variates from the URNG when the
+  // width of the URNG range is not a power of 2.
+  // Y = (2 ^ kRange) - 1
+  static constexpr result_type kRangeMask =
+      kRange == 0 ? (std::numeric_limits<result_type>::max)()
+                  : constexpr_mask(kRangeBits);
+
+  static_assert((URBG::max)() != (URBG::min)(),
+                "Class-template FastUniformBitsURBGConstants<> "
+                "URBG::max and URBG::min may not be equal.");
+
+  static_assert(std::is_unsigned<result_type>::value,
+                "Class-template FastUniformBitsURBGConstants<> "
+                "URBG::result_type must be unsigned.");
+
+  static_assert(kRangeMask > 0,
+                "Class-template FastUniformBitsURBGConstants<> "
+                "URBG does not generate sufficient random bits.");
+
+  static_assert(kRange == 0 ||
+                    kRangeBits < std::numeric_limits<result_type>::digits,
+                "Class-template FastUniformBitsURBGConstants<> "
+                "URBG range computation error.");
+};
+
+// FastUniformBitsLoopingConstants computes the looping constants used
+// by FastUniformBits::Generate. These constants indicate how multiple
+// URBG::result_type values are combined into an output_value.
+// Parameterized by the FastUniformBits parameters:
+//  `UIntType`: output type.
+//  `Width`: binary output width,
+//  `URNG`: The underlying UniformRandomNumberGenerator.
+//
+// The looping constants describe the sets of loop counters and mask values
+// which control how individual variates are combined the final output.  The
+// algorithm ensures that the number of bits used by any individual call differs
+// by at-most one bit from any other call. This is simplified into constants
+// which describe two loops, with the second loop parameters providing one extra
+// bit per variate.
+//
+// See [rand.adapt.ibits] for more details on the use of these constants.
+template <typename UIntType, size_t Width, typename URBG>
+class FastUniformBitsLoopingConstants {
+ private:
+  static constexpr size_t kWidth = Width;
+  using urbg_result_type = typename URBG::result_type;
+  using uint_result_type = UIntType;
+
+ public:
+  using result_type =
+      typename std::conditional<(sizeof(urbg_result_type) <=
+                                 sizeof(uint_result_type)),
+                                uint_result_type, urbg_result_type>::type;
+
+ private:
+  // Estimate N as ceil(width / urng width), and W0 as (width / N).
+  static constexpr size_t kRangeBits =
+      FastUniformBitsURBGConstants<URBG>::kRangeBits;
+
+  // The range of the URNG, max - min + 1, or zero if that result would cause
+  // overflow.
+  static constexpr result_type kRange = constexpr_range<URBG>();
+  static constexpr size_t kEstimateN =
+      kWidth / kRangeBits + (kWidth % kRangeBits != 0);
+  static constexpr size_t kEstimateW0 = kWidth / kEstimateN;
+  static constexpr result_type kEstimateY0 = (kRange >> kEstimateW0)
+                                             << kEstimateW0;
+
+ public:
+  // Parameters for the two loops:
+  // kN0, kN1 are the number of underlying calls required for each loop.
+  // KW0, kW1 are shift widths for each loop.
+  //
+  static constexpr size_t kN1 = (kRange - kEstimateY0) >
+                                        (kEstimateY0 / kEstimateN)
+                                    ? kEstimateN + 1
+                                    : kEstimateN;
+  static constexpr size_t kN0 = kN1 - (kWidth % kN1);
+  static constexpr size_t kW0 = kWidth / kN1;
+  static constexpr size_t kW1 = kW0 + 1;
+
+  static constexpr result_type kM0 = (result_type(1) << kW0) - 1;
+  static constexpr result_type kM1 = (result_type(1) << kW1) - 1;
+
+  static_assert(
+      kW0 <= kRangeBits,
+      "Class-template FastUniformBitsLoopingConstants::kW0 too large.");
+
+  static_assert(
+      kW0 > 0,
+      "Class-template FastUniformBitsLoopingConstants::kW0 too small.");
+};
+
+template <typename UIntType, size_t Width>
+template <typename URBG>
+typename FastUniformBits<UIntType, Width>::result_type
+FastUniformBits<UIntType, Width>::operator()(
+    URBG& g) {  // NOLINT(runtime/references)
+  using constants = FastUniformBitsURBGConstants<URBG>;
+  return Generate(
+      g, std::integral_constant<bool, constants::kRangeMask >= (max)()>{});
+}
+
+template <typename UIntType, size_t Width>
+template <typename URBG>
+typename URBG::result_type FastUniformBits<UIntType, Width>::Variate(
+    URBG& g) {  // NOLINT(runtime/references)
+  using constants = FastUniformBitsURBGConstants<URBG>;
+  if (constants::kPowerOfTwo) {
+    return g() - (URBG::min)();
+  }
+
+  // Use rejection sampling to ensure uniformity across the range.
+  typename URBG::result_type u;
+  do {
+    u = g() - (URBG::min)();
+  } while (u > constants::kRangeMask);
+  return u;
+}
+
+template <typename UIntType, size_t Width>
+template <typename URBG>
+typename FastUniformBits<UIntType, Width>::result_type
+FastUniformBits<UIntType, Width>::Generate(
+    URBG& g,  // NOLINT(runtime/references)
+    std::true_type /* avoid_looping */) {
+  // The width of the result_type is less than than the width of the random bits
+  // provided by URNG.  Thus, generate a single value and then simply mask off
+  // the required bits.
+  return Variate(g) & (max)();
+}
+
+template <typename UIntType, size_t Width>
+template <typename URBG>
+typename FastUniformBits<UIntType, Width>::result_type
+FastUniformBits<UIntType, Width>::Generate(
+    URBG& g,  // NOLINT(runtime/references)
+    std::false_type /* avoid_looping */) {
+  // The width of the result_type is wider than the number of random bits
+  // provided by URNG. Thus we merge several variates of URNG into the result
+  // using a shift and mask.  The constants type generates the parameters used
+  // ensure that the bits are distributed across all the invocations of the
+  // underlying URNG.
+  using constants = FastUniformBitsLoopingConstants<UIntType, Width, URBG>;
+
+  result_type s = 0;
+  for (size_t n = 0; n < constants::kN0; ++n) {
+    auto u = Variate(g);
+    s = (s << constants::kW0) + (u & constants::kM0);
+  }
+  for (size_t n = constants::kN0; n < constants::kN1; ++n) {
+    auto u = Variate(g);
+    s = (s << constants::kW1) + (u & constants::kM1);
+  }
+  return s;
+}
+
+}  // namespace random_internal
+}  // namespace absl
+
+#endif  // ABSL_RANDOM_INTERNAL_FAST_UNIFORM_BITS_H_