unittests/FuzzMutate/ReservoirSamplerTest.cpp

   1 //===- ReservoirSampler.cpp - Tests for the ReservoirSampler --------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9
  10 #include "llvm/FuzzMutate/Random.h"
  11 #include "gtest/gtest.h"
  12 #include <random>
  13
  14 using namespace llvm;
  15
  16 TEST(ReservoirSamplerTest, OneItem) {
  17   std::mt19937 Rand;
  18   auto Sampler = makeSampler(Rand, 7, 1);
  19   ASSERT_FALSE(Sampler.isEmpty());
  20   ASSERT_EQ(7, Sampler.getSelection());
  21 }
  22
  23 TEST(ReservoirSamplerTest, NoWeight) {
  24   std::mt19937 Rand;
  25   auto Sampler = makeSampler(Rand, 7, 0);
  26   ASSERT_TRUE(Sampler.isEmpty());
  27 }
  28
  29 TEST(ReservoirSamplerTest, Uniform) {
  30   std::mt19937 Rand;
  31
  32   // Run three chi-squared tests to check that the distribution is reasonably
  33   // uniform.
  34   std::vector<int> Items = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
  35
  36   int Failures = 0;
  37   for (int Run = 0; Run < 3; ++Run) {
  38     std::vector<int> Counts(Items.size(), 0);
  39
  40     // We need $np_s > 5$ at minimum, but we're better off going a couple of
  41     // orders of magnitude larger.
  42     int N = Items.size() * 5 * 100;
  43     for (int I = 0; I < N; ++I) {
  44       auto Sampler = makeSampler(Rand, Items);
  45       Counts[Sampler.getSelection()] += 1;
  46     }
  47
  48     // Knuth. TAOCP Vol. 2, 3.3.1 (8):
  49     // $V = \frac{1}{n} \sum_{s=1}^{k} \left(\frac{Y_s^2}{p_s}\right) - n$
  50     double Ps = 1.0 / Items.size();
  51     double Sum = 0.0;
  52     for (int Ys : Counts)
  53       Sum += Ys * Ys / Ps;
  54     double V = (Sum / N) - N;
  55
  56     assert(Items.size() == 10 && "Our chi-squared values assume 10 items");
  57     // Since we have 10 items, there are 9 degrees of freedom and the table of
  58     // chi-squared values is as follows:
  59     //
  60     //     | p=1%  |   5%  |  25%  |  50%  |  75%  |  95%  |  99%  |
  61     // v=9 | 2.088 | 3.325 | 5.899 | 8.343 | 11.39 | 16.92 | 21.67 |
  62     //
  63     // Check that we're in the likely range of results.
  64     //if (V < 2.088 || V > 21.67)
  65     if (V < 2.088 || V > 21.67)
  66       ++Failures;
  67   }
  68   EXPECT_LT(Failures, 3) << "Non-uniform distribution?";
  69 }