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

1 files changed, 260 insertions, 0 deletions

diff --git a/absl/random/gaussian_distribution.h b/absl/random/gaussian_distribution.h
new file mode 100644
index 000000000000..1d1347bce0a0
--- /dev/null
+++ b/absl/random/gaussian_distribution.h
@@ -0,0 +1,260 @@
+// 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_GAUSSIAN_DISTRIBUTION_H_
+#define ABSL_RANDOM_GAUSSIAN_DISTRIBUTION_H_
+
+// absl::gaussian_distribution implements the Ziggurat algorithm
+// for generating random gaussian numbers.
+//
+// Implementation based on "The Ziggurat Method for Generating Random Variables"
+// by George Marsaglia and Wai Wan Tsang: http://www.jstatsoft.org/v05/i08/
+//
+
+#include <cmath>
+#include <cstdint>
+#include <istream>
+#include <limits>
+#include <type_traits>
+
+#include "absl/random/internal/distribution_impl.h"
+#include "absl/random/internal/fast_uniform_bits.h"
+#include "absl/random/internal/iostream_state_saver.h"
+
+namespace absl {
+namespace random_internal {
+
+// absl::gaussian_distribution_base implements the underlying ziggurat algorithm
+// using the ziggurat tables generated by the gaussian_distribution_gentables
+// binary.
+//
+// The specific algorithm has some of the improvements suggested by the
+// 2005 paper, "An Improved Ziggurat Method to Generate Normal Random Samples",
+// Jurgen A Doornik.  (https://www.doornik.com/research/ziggurat.pdf)
+class gaussian_distribution_base {
+ public:
+  template <typename URBG>
+  inline double zignor(URBG& g);  // NOLINT(runtime/references)
+
+ private:
+  friend class TableGenerator;
+
+  template <typename URBG>
+  inline double zignor_fallback(URBG& g,  // NOLINT(runtime/references)
+                                bool neg);
+
+  // Constants used for the gaussian distribution.
+  static constexpr double kR = 3.442619855899;  // Start of the tail.
+  static constexpr double kRInv = 0.29047645161474317;  // ~= (1.0 / kR) .
+  static constexpr double kV = 9.91256303526217e-3;
+  static constexpr uint64_t kMask = 0x07f;
+
+  // The ziggurat tables store the pdf(f) and inverse-pdf(x) for equal-area
+  // points on one-half of the normal distribution, where the pdf function,
+  // pdf = e ^ (-1/2 *x^2), assumes that the mean = 0 & stddev = 1.
+  //
+  // These tables are just over 2kb in size; larger tables might improve the
+  // distributions, but also lead to more cache pollution.
+  //
+  // x = {3.71308, 3.44261, 3.22308, ..., 0}
+  // f = {0.00101, 0.00266, 0.00554, ..., 1}
+  struct Tables {
+    double x[kMask + 2];
+    double f[kMask + 2];
+  };
+  static const Tables zg_;
+  random_internal::FastUniformBits<uint64_t> fast_u64_;
+};
+
+}  // namespace random_internal
+
+// absl::gaussian_distribution:
+// Generates a number conforming to a Gaussian distribution.
+template <typename RealType = double>
+class gaussian_distribution : random_internal::gaussian_distribution_base {
+ public:
+  using result_type = RealType;
+
+  class param_type {
+   public:
+    using distribution_type = gaussian_distribution;
+
+    explicit param_type(result_type mean = 0, result_type stddev = 1)
+        : mean_(mean), stddev_(stddev) {}
+
+    // Returns the mean distribution parameter.  The mean specifies the location
+    // of the peak.  The default value is 0.0.
+    result_type mean() const { return mean_; }
+
+    // Returns the deviation distribution parameter.  The default value is 1.0.
+    result_type stddev() const { return stddev_; }
+
+    friend bool operator==(const param_type& a, const param_type& b) {
+      return a.mean_ == b.mean_ && a.stddev_ == b.stddev_;
+    }
+
+    friend bool operator!=(const param_type& a, const param_type& b) {
+      return !(a == b);
+    }
+
+   private:
+    result_type mean_;
+    result_type stddev_;
+
+    static_assert(
+        std::is_floating_point<RealType>::value,
+        "Class-template absl::gaussian_distribution<> must be parameterized "
+        "using a floating-point type.");
+  };
+
+  gaussian_distribution() : gaussian_distribution(0) {}
+
+  explicit gaussian_distribution(result_type mean, result_type stddev = 1)
+      : param_(mean, stddev) {}
+
+  explicit gaussian_distribution(const param_type& p) : param_(p) {}
+
+  void reset() {}
+
+  // Generating functions
+  template <typename URBG>
+  result_type operator()(URBG& g) {  // NOLINT(runtime/references)
+    return (*this)(g, param_);
+  }
+
+  template <typename URBG>
+  result_type operator()(URBG& g,  // NOLINT(runtime/references)
+                         const param_type& p);
+
+  param_type param() const { return param_; }
+  void param(const param_type& p) { param_ = p; }
+
+  result_type(min)() const {
+    return -std::numeric_limits<result_type>::infinity();
+  }
+  result_type(max)() const {
+    return std::numeric_limits<result_type>::infinity();
+  }
+
+  result_type mean() const { return param_.mean(); }
+  result_type stddev() const { return param_.stddev(); }
+
+  friend bool operator==(const gaussian_distribution& a,
+                         const gaussian_distribution& b) {
+    return a.param_ == b.param_;
+  }
+  friend bool operator!=(const gaussian_distribution& a,
+                         const gaussian_distribution& b) {
+    return a.param_ != b.param_;
+  }
+
+ private:
+  param_type param_;
+};
+
+// --------------------------------------------------------------------------
+// Implementation details only below
+// --------------------------------------------------------------------------
+
+template <typename RealType>
+template <typename URBG>
+typename gaussian_distribution<RealType>::result_type
+gaussian_distribution<RealType>::operator()(
+    URBG& g,  // NOLINT(runtime/references)
+    const param_type& p) {
+  return p.mean() + p.stddev() * static_cast<result_type>(zignor(g));
+}
+
+template <typename CharT, typename Traits, typename RealType>
+std::basic_ostream<CharT, Traits>& operator<<(
+    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
+    const gaussian_distribution<RealType>& x) {
+  auto saver = random_internal::make_ostream_state_saver(os);
+  os.precision(random_internal::stream_precision_helper<RealType>::kPrecision);
+  os << x.mean() << os.fill() << x.stddev();
+  return os;
+}
+
+template <typename CharT, typename Traits, typename RealType>
+std::basic_istream<CharT, Traits>& operator>>(
+    std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
+    gaussian_distribution<RealType>& x) {   // NOLINT(runtime/references)
+  using result_type = typename gaussian_distribution<RealType>::result_type;
+  using param_type = typename gaussian_distribution<RealType>::param_type;
+
+  auto saver = random_internal::make_istream_state_saver(is);
+  auto mean = random_internal::read_floating_point<result_type>(is);
+  if (is.fail()) return is;
+  auto stddev = random_internal::read_floating_point<result_type>(is);
+  if (!is.fail()) {
+    x.param(param_type(mean, stddev));
+  }
+  return is;
+}
+
+namespace random_internal {
+
+template <typename URBG>
+inline double gaussian_distribution_base::zignor_fallback(URBG& g, bool neg) {
+  // This fallback path happens approximately 0.05% of the time.
+  double x, y;
+  do {
+    // kRInv = 1/r, U(0, 1)
+    x = kRInv * std::log(RandU64ToDouble<PositiveValueT, false>(fast_u64_(g)));
+    y = -std::log(RandU64ToDouble<PositiveValueT, false>(fast_u64_(g)));
+  } while ((y + y) < (x * x));
+  return neg ? (x - kR) : (kR - x);
+}
+
+template <typename URBG>
+inline double gaussian_distribution_base::zignor(
+    URBG& g) {  // NOLINT(runtime/references)
+  while (true) {
+    // We use a single uint64_t to generate both a double and a strip.
+    // These bits are unused when the generated double is > 1/2^5.
+    // This may introduce some bias from the duplicated low bits of small
+    // values (those smaller than 1/2^5, which all end up on the left tail).
+    uint64_t bits = fast_u64_(g);
+    int i = static_cast<int>(bits & kMask);  // pick a random strip
+    double j = RandU64ToDouble<SignedValueT, false>(bits);  // U(-1, 1)
+    const double x = j * zg_.x[i];
+
+    // Retangular box. Handles >97% of all cases.
+    // For any given box, this handles between 75% and 99% of values.
+    // Equivalent to U(01) < (x[i+1] / x[i]), and when i == 0, ~93.5%
+    if (std::abs(x) < zg_.x[i + 1]) {
+      return x;
+    }
+
+    // i == 0: Base box. Sample using a ratio of uniforms.
+    if (i == 0) {
+      // This path happens about 0.05% of the time.
+      return zignor_fallback(g, j < 0);
+    }
+
+    // i > 0: Wedge samples using precomputed values.
+    double v = RandU64ToDouble<PositiveValueT, false>(fast_u64_(g));  // U(0, 1)
+    if ((zg_.f[i + 1] + v * (zg_.f[i] - zg_.f[i + 1])) <
+        std::exp(-0.5 * x * x)) {
+      return x;
+    }
+
+    // The wedge was missed; reject the value and try again.
+  }
+}
+
+}  // namespace random_internal
+}  // namespace absl
+
+#endif  // ABSL_RANDOM_GAUSSIAN_DISTRIBUTION_H_