1 #include <gtest/gtest.h>
2 #include "test_quantile.h"
3 #include "../../../src/common/quantile.h"
4 #include "../../../src/common/hist_util.h"
5
6 namespace xgboost {
7 namespace common {
8
TEST(Quantile,LoadBalance)9 TEST(Quantile, LoadBalance) {
10 size_t constexpr kRows = 1000, kCols = 100;
11 auto m = RandomDataGenerator{kRows, kCols, 0}.GenerateDMatrix();
12 std::vector<bst_feature_t> cols_ptr;
13 for (auto const &page : m->GetBatches<SparsePage>()) {
14 cols_ptr = HostSketchContainer::LoadBalance(page, kCols, 13);
15 }
16 size_t n_cols = 0;
17 for (size_t i = 1; i < cols_ptr.size(); ++i) {
18 n_cols += cols_ptr[i] - cols_ptr[i - 1];
19 }
20 CHECK_EQ(n_cols, kCols);
21 }
22
TestDistributedQuantile(size_t rows,size_t cols)23 void TestDistributedQuantile(size_t rows, size_t cols) {
24 std::string msg {"Skipping AllReduce test"};
25 int32_t constexpr kWorkers = 4;
26 InitRabitContext(msg, kWorkers);
27 auto world = rabit::GetWorldSize();
28 if (world != 1) {
29 ASSERT_EQ(world, kWorkers);
30 } else {
31 return;
32 }
33
34 std::vector<MetaInfo> infos(2);
35 auto& h_weights = infos.front().weights_.HostVector();
36 h_weights.resize(rows);
37 SimpleLCG lcg;
38 SimpleRealUniformDistribution<float> dist(3, 1000);
39 std::generate(h_weights.begin(), h_weights.end(), [&]() { return dist(&lcg); });
40 std::vector<bst_row_t> column_size(cols, rows);
41 size_t n_bins = 64;
42
43 // Generate cuts for distributed environment.
44 auto sparsity = 0.5f;
45 auto rank = rabit::GetRank();
46 auto m = RandomDataGenerator{rows, cols, sparsity}
47 .Seed(rank)
48 .Lower(.0f)
49 .Upper(1.0f)
50 .GenerateDMatrix();
51 HostSketchContainer sketch_distributed(
52 column_size, n_bins, m->Info().feature_types.ConstHostSpan(), false,
53 OmpGetNumThreads(0));
54 for (auto const &page : m->GetBatches<SparsePage>()) {
55 sketch_distributed.PushRowPage(page, m->Info());
56 }
57 HistogramCuts distributed_cuts;
58 sketch_distributed.MakeCuts(&distributed_cuts);
59
60 // Generate cuts for single node environment
61 rabit::Finalize();
62 CHECK_EQ(rabit::GetWorldSize(), 1);
63 std::for_each(column_size.begin(), column_size.end(), [=](auto& size) { size *= world; });
64 HostSketchContainer sketch_on_single_node(
65 column_size, n_bins, m->Info().feature_types.ConstHostSpan(), false,
66 OmpGetNumThreads(0));
67 for (auto rank = 0; rank < world; ++rank) {
68 auto m = RandomDataGenerator{rows, cols, sparsity}
69 .Seed(rank)
70 .Lower(.0f)
71 .Upper(1.0f)
72 .GenerateDMatrix();
73 for (auto const &page : m->GetBatches<SparsePage>()) {
74 sketch_on_single_node.PushRowPage(page, m->Info());
75 }
76 }
77
78 HistogramCuts single_node_cuts;
79 sketch_on_single_node.MakeCuts(&single_node_cuts);
80
81 auto const& sptrs = single_node_cuts.Ptrs();
82 auto const& dptrs = distributed_cuts.Ptrs();
83 auto const& svals = single_node_cuts.Values();
84 auto const& dvals = distributed_cuts.Values();
85 auto const& smins = single_node_cuts.MinValues();
86 auto const& dmins = distributed_cuts.MinValues();
87
88 ASSERT_EQ(sptrs.size(), dptrs.size());
89 for (size_t i = 0; i < sptrs.size(); ++i) {
90 ASSERT_EQ(sptrs[i], dptrs[i]);
91 }
92
93 ASSERT_EQ(svals.size(), dvals.size());
94 for (size_t i = 0; i < svals.size(); ++i) {
95 ASSERT_NEAR(svals[i], dvals[i], 2e-2f);
96 }
97
98 ASSERT_EQ(smins.size(), dmins.size());
99 for (size_t i = 0; i < smins.size(); ++i) {
100 ASSERT_FLOAT_EQ(smins[i], dmins[i]);
101 }
102 }
103
TEST(Quantile,DistributedBasic)104 TEST(Quantile, DistributedBasic) {
105 #if defined(__unix__)
106 constexpr size_t kRows = 10, kCols = 10;
107 TestDistributedQuantile(kRows, kCols);
108 #endif
109 }
110
TEST(Quantile,Distributed)111 TEST(Quantile, Distributed) {
112 #if defined(__unix__)
113 constexpr size_t kRows = 1000, kCols = 200;
114 TestDistributedQuantile(kRows, kCols);
115 #endif
116 }
117
TEST(Quantile,SameOnAllWorkers)118 TEST(Quantile, SameOnAllWorkers) {
119 #if defined(__unix__)
120 std::string msg{"Skipping Quantile AllreduceBasic test"};
121 int32_t constexpr kWorkers = 4;
122 InitRabitContext(msg, kWorkers);
123 auto world = rabit::GetWorldSize();
124 if (world != 1) {
125 CHECK_EQ(world, kWorkers);
126 } else {
127 LOG(WARNING) << msg;
128 return;
129 }
130
131 constexpr size_t kRows = 1000, kCols = 100;
132 RunWithSeedsAndBins(
133 kRows, [=](int32_t seed, size_t n_bins, MetaInfo const &info) {
134 auto rank = rabit::GetRank();
135 HostDeviceVector<float> storage;
136 auto m = RandomDataGenerator{kRows, kCols, 0}
137 .Device(0)
138 .Seed(rank + seed)
139 .GenerateDMatrix();
140 auto cuts = SketchOnDMatrix(m.get(), n_bins);
141 std::vector<float> cut_values(cuts.Values().size() * world, 0);
142 std::vector<
143 typename std::remove_reference_t<decltype(cuts.Ptrs())>::value_type>
144 cut_ptrs(cuts.Ptrs().size() * world, 0);
145 std::vector<float> cut_min_values(cuts.MinValues().size() * world, 0);
146
147 size_t value_size = cuts.Values().size();
148 rabit::Allreduce<rabit::op::Max>(&value_size, 1);
149 size_t ptr_size = cuts.Ptrs().size();
150 rabit::Allreduce<rabit::op::Max>(&ptr_size, 1);
151 CHECK_EQ(ptr_size, kCols + 1);
152 size_t min_value_size = cuts.MinValues().size();
153 rabit::Allreduce<rabit::op::Max>(&min_value_size, 1);
154 CHECK_EQ(min_value_size, kCols);
155
156 size_t value_offset = value_size * rank;
157 std::copy(cuts.Values().begin(), cuts.Values().end(),
158 cut_values.begin() + value_offset);
159 size_t ptr_offset = ptr_size * rank;
160 std::copy(cuts.Ptrs().cbegin(), cuts.Ptrs().cend(),
161 cut_ptrs.begin() + ptr_offset);
162 size_t min_values_offset = min_value_size * rank;
163 std::copy(cuts.MinValues().cbegin(), cuts.MinValues().cend(),
164 cut_min_values.begin() + min_values_offset);
165
166 rabit::Allreduce<rabit::op::Sum>(cut_values.data(), cut_values.size());
167 rabit::Allreduce<rabit::op::Sum>(cut_ptrs.data(), cut_ptrs.size());
168 rabit::Allreduce<rabit::op::Sum>(cut_min_values.data(), cut_min_values.size());
169
170 for (int32_t i = 0; i < world; i++) {
171 for (size_t j = 0; j < value_size; ++j) {
172 size_t idx = i * value_size + j;
173 ASSERT_NEAR(cuts.Values().at(j), cut_values.at(idx), kRtEps);
174 }
175
176 for (size_t j = 0; j < ptr_size; ++j) {
177 size_t idx = i * ptr_size + j;
178 ASSERT_EQ(cuts.Ptrs().at(j), cut_ptrs.at(idx));
179 }
180
181 for (size_t j = 0; j < min_value_size; ++j) {
182 size_t idx = i * min_value_size + j;
183 ASSERT_EQ(cuts.MinValues().at(j), cut_min_values.at(idx));
184 }
185 }
186 });
187 rabit::Finalize();
188 #endif // defined(__unix__)
189 }
190 } // namespace common
191 } // namespace xgboost
192