log_uniform_int_distribution_test.cc (9772B)
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 #include "absl/random/log_uniform_int_distribution.h" 16 17 #include <cstddef> 18 #include <cstdint> 19 #include <iterator> 20 #include <random> 21 #include <sstream> 22 #include <string> 23 #include <vector> 24 25 #include "gmock/gmock.h" 26 #include "gtest/gtest.h" 27 #include "absl/log/log.h" 28 #include "absl/random/internal/chi_square.h" 29 #include "absl/random/internal/distribution_test_util.h" 30 #include "absl/random/internal/pcg_engine.h" 31 #include "absl/random/internal/sequence_urbg.h" 32 #include "absl/random/random.h" 33 #include "absl/strings/str_cat.h" 34 #include "absl/strings/str_format.h" 35 #include "absl/strings/str_replace.h" 36 #include "absl/strings/strip.h" 37 38 namespace { 39 40 template <typename IntType> 41 class LogUniformIntDistributionTypeTest : public ::testing::Test {}; 42 43 using IntTypes = ::testing::Types<int8_t, int16_t, int32_t, int64_t, // 44 uint8_t, uint16_t, uint32_t, uint64_t>; 45 TYPED_TEST_SUITE(LogUniformIntDistributionTypeTest, IntTypes); 46 47 TYPED_TEST(LogUniformIntDistributionTypeTest, SerializeTest) { 48 using param_type = 49 typename absl::log_uniform_int_distribution<TypeParam>::param_type; 50 using Limits = std::numeric_limits<TypeParam>; 51 52 constexpr int kCount = 1000; 53 absl::InsecureBitGen gen; 54 for (const auto& param : { 55 param_type(0, 1), // 56 param_type(0, 2), // 57 param_type(0, 2, 10), // 58 param_type(9, 32, 4), // 59 param_type(1, 101, 10), // 60 param_type(1, Limits::max() / 2), // 61 param_type(0, Limits::max() - 1), // 62 param_type(0, Limits::max(), 2), // 63 param_type(0, Limits::max(), 10), // 64 param_type(Limits::min(), 0), // 65 param_type(Limits::lowest(), Limits::max()), // 66 param_type(Limits::min(), Limits::max()), // 67 }) { 68 // Validate parameters. 69 const auto min = param.min(); 70 const auto max = param.max(); 71 const auto base = param.base(); 72 absl::log_uniform_int_distribution<TypeParam> before(min, max, base); 73 EXPECT_EQ(before.min(), param.min()); 74 EXPECT_EQ(before.max(), param.max()); 75 EXPECT_EQ(before.base(), param.base()); 76 77 { 78 absl::log_uniform_int_distribution<TypeParam> via_param(param); 79 EXPECT_EQ(via_param, before); 80 } 81 82 // Validate stream serialization. 83 std::stringstream ss; 84 ss << before; 85 86 absl::log_uniform_int_distribution<TypeParam> after(3, 6, 17); 87 88 EXPECT_NE(before.max(), after.max()); 89 EXPECT_NE(before.base(), after.base()); 90 EXPECT_NE(before.param(), after.param()); 91 EXPECT_NE(before, after); 92 93 ss >> after; 94 95 EXPECT_EQ(before.min(), after.min()); 96 EXPECT_EQ(before.max(), after.max()); 97 EXPECT_EQ(before.base(), after.base()); 98 EXPECT_EQ(before.param(), after.param()); 99 EXPECT_EQ(before, after); 100 101 // Smoke test. 102 auto sample_min = after.max(); 103 auto sample_max = after.min(); 104 for (int i = 0; i < kCount; i++) { 105 auto sample = after(gen); 106 EXPECT_GE(sample, after.min()); 107 EXPECT_LE(sample, after.max()); 108 if (sample > sample_max) sample_max = sample; 109 if (sample < sample_min) sample_min = sample; 110 } 111 LOG(INFO) << "Range: " << sample_min << ", " << sample_max; 112 } 113 } 114 115 using log_uniform_i32 = absl::log_uniform_int_distribution<int32_t>; 116 117 class LogUniformIntChiSquaredTest 118 : public testing::TestWithParam<log_uniform_i32::param_type> { 119 public: 120 // The ChiSquaredTestImpl provides a chi-squared goodness of fit test for 121 // data generated by the log-uniform-int distribution. 122 double ChiSquaredTestImpl(); 123 124 // We use a fixed bit generator for distribution accuracy tests. This allows 125 // these tests to be deterministic, while still testing the qualify of the 126 // implementation. 127 absl::random_internal::pcg64_2018_engine rng_{0x2B7E151628AED2A6}; 128 }; 129 130 double LogUniformIntChiSquaredTest::ChiSquaredTestImpl() { 131 using absl::random_internal::kChiSquared; 132 133 const auto& param = GetParam(); 134 135 // Check the distribution of L=log(log_uniform_int_distribution, base), 136 // expecting that L is roughly uniformly distributed, that is: 137 // 138 // P[L=0] ~= P[L=1] ~= ... ~= P[L=log(max)] 139 // 140 // For a total of X entries, each bucket should contain some number of samples 141 // in the interval [X/k - a, X/k + a]. 142 // 143 // Where `a` is approximately sqrt(X/k). This is validated by bucketing 144 // according to the log function and using a chi-squared test for uniformity. 145 146 const bool is_2 = (param.base() == 2); 147 const double base_log = 1.0 / std::log(param.base()); 148 const auto bucket_index = [base_log, is_2, ¶m](int32_t x) { 149 uint64_t y = static_cast<uint64_t>(x) - param.min(); 150 return (y == 0) ? 0 151 : is_2 ? static_cast<int>(1 + std::log2(y)) 152 : static_cast<int>(1 + std::log(y) * base_log); 153 }; 154 const int max_bucket = bucket_index(param.max()); // inclusive 155 const size_t trials = 15 + (max_bucket + 1) * 10; 156 157 log_uniform_i32 dist(param); 158 159 std::vector<int64_t> buckets(max_bucket + 1); 160 for (size_t i = 0; i < trials; ++i) { 161 const auto sample = dist(rng_); 162 // Check the bounds. 163 ABSL_ASSERT(sample <= dist.max()); 164 ABSL_ASSERT(sample >= dist.min()); 165 // Convert the output of the generator to one of num_bucket buckets. 166 int bucket = bucket_index(sample); 167 ABSL_ASSERT(bucket <= max_bucket); 168 ++buckets[bucket]; 169 } 170 171 // The null-hypothesis is that the distribution is uniform with respect to 172 // log-uniform-int bucketization. 173 const int dof = buckets.size() - 1; 174 const double expected = trials / static_cast<double>(buckets.size()); 175 176 const double threshold = absl::random_internal::ChiSquareValue(dof, 0.98); 177 178 double chi_square = absl::random_internal::ChiSquareWithExpected( 179 std::begin(buckets), std::end(buckets), expected); 180 181 const double p = absl::random_internal::ChiSquarePValue(chi_square, dof); 182 183 if (chi_square > threshold) { 184 LOG(INFO) << "values"; 185 for (size_t i = 0; i < buckets.size(); i++) { 186 LOG(INFO) << i << ": " << buckets[i]; 187 } 188 LOG(INFO) << "trials=" << trials << "\n" 189 << kChiSquared << "(data, " << dof << ") = " << chi_square << " (" 190 << p << ")\n" 191 << kChiSquared << " @ 0.98 = " << threshold; 192 } 193 return p; 194 } 195 196 TEST_P(LogUniformIntChiSquaredTest, MultiTest) { 197 const int kTrials = 5; 198 int failures = 0; 199 for (int i = 0; i < kTrials; i++) { 200 double p_value = ChiSquaredTestImpl(); 201 if (p_value < 0.005) { 202 failures++; 203 } 204 } 205 206 // There is a 0.10% chance of producing at least one failure, so raise the 207 // failure threshold high enough to allow for a flake rate < 10,000. 208 EXPECT_LE(failures, 4); 209 } 210 211 // Generate the parameters for the test. 212 std::vector<log_uniform_i32::param_type> GenParams() { 213 using Param = log_uniform_i32::param_type; 214 using Limits = std::numeric_limits<int32_t>; 215 216 return std::vector<Param>{ 217 Param{0, 1, 2}, 218 Param{1, 1, 2}, 219 Param{0, 2, 2}, 220 Param{0, 3, 2}, 221 Param{0, 4, 2}, 222 Param{0, 9, 10}, 223 Param{0, 10, 10}, 224 Param{0, 11, 10}, 225 Param{1, 10, 10}, 226 Param{0, (1 << 8) - 1, 2}, 227 Param{0, (1 << 8), 2}, 228 Param{0, (1 << 30) - 1, 2}, 229 Param{-1000, 1000, 10}, 230 Param{0, Limits::max(), 2}, 231 Param{0, Limits::max(), 3}, 232 Param{0, Limits::max(), 10}, 233 Param{Limits::min(), 0}, 234 Param{Limits::min(), Limits::max(), 2}, 235 }; 236 } 237 238 std::string ParamName( 239 const ::testing::TestParamInfo<log_uniform_i32::param_type>& info) { 240 const auto& p = info.param; 241 std::string name = 242 absl::StrCat("min_", p.min(), "__max_", p.max(), "__base_", p.base()); 243 return absl::StrReplaceAll(name, {{"+", "_"}, {"-", "_"}, {".", "_"}}); 244 } 245 246 INSTANTIATE_TEST_SUITE_P(All, LogUniformIntChiSquaredTest, 247 ::testing::ValuesIn(GenParams()), ParamName); 248 249 // NOTE: absl::log_uniform_int_distribution is not guaranteed to be stable. 250 TEST(LogUniformIntDistributionTest, StabilityTest) { 251 using testing::ElementsAre; 252 // absl::uniform_int_distribution stability relies on 253 // absl::random_internal::LeadingSetBit, std::log, std::pow. 254 absl::random_internal::sequence_urbg urbg( 255 {0x0003eb76f6f7f755ull, 0xFFCEA50FDB2F953Bull, 0xC332DDEFBE6C5AA5ull, 256 0x6558218568AB9702ull, 0x2AEF7DAD5B6E2F84ull, 0x1521B62829076170ull, 257 0xECDD4775619F1510ull, 0x13CCA830EB61BD96ull, 0x0334FE1EAA0363CFull, 258 0xB5735C904C70A239ull, 0xD59E9E0BCBAADE14ull, 0xEECC86BC60622CA7ull}); 259 260 std::vector<int> output(6); 261 262 { 263 absl::log_uniform_int_distribution<int32_t> dist(0, 256); 264 std::generate(std::begin(output), std::end(output), 265 [&] { return dist(urbg); }); 266 EXPECT_THAT(output, ElementsAre(256, 66, 4, 6, 57, 103)); 267 } 268 urbg.reset(); 269 { 270 absl::log_uniform_int_distribution<int32_t> dist(0, 256, 10); 271 std::generate(std::begin(output), std::end(output), 272 [&] { return dist(urbg); }); 273 EXPECT_THAT(output, ElementsAre(8, 4, 0, 0, 0, 69)); 274 } 275 } 276 277 } // namespace