Support column split in approx tree method (#8847)

include/xgboost/data.h

@@ -529,6 +529,11 @@ class DMatrix {
     return Info().num_nonzero_ == Info().num_row_ * Info().num_col_;
   }
 
+  /*! \brief Whether the data is split row-wise. */
+  bool IsRowSplit() const {
+    return Info().data_split_mode == DataSplitMode::kRow;
+  }
+
   /*! \brief Whether the data is split column-wise. */
   bool IsColumnSplit() const {
     return Info().data_split_mode == DataSplitMode::kCol;

src/data/data.cc

@@ -912,6 +912,7 @@ DMatrix* DMatrix::Load(const std::string& uri, bool silent, DataSplitMode data_s
     if (!cache_file.empty()) {
       LOG(FATAL) << "Column-wise data split is not support for external memory.";
     }
+    LOG(CONSOLE) << "Splitting data by column";
     auto* sliced = dmat->SliceCol(npart, partid);
     delete dmat;
     return sliced;

src/tree/hist/evaluate_splits.h

@@ -38,6 +38,7 @@ class HistEvaluator {
   TrainParam param_;
   std::shared_ptr<common::ColumnSampler> column_sampler_;
   TreeEvaluator tree_evaluator_;
+  bool is_col_split_{false};
   FeatureInteractionConstraintHost interaction_constraints_;
   std::vector<NodeEntry> snode_;
 
@@ -355,7 +356,24 @@ class HistEvaluator {
             tloc_candidates[n_threads * nidx_in_set + tidx].split);
       }
     }
+
+    if (is_col_split_) {
+      // With column-wise data split, we gather the best splits from all the workers and update the
+      // expand entries accordingly.
+      auto const world = collective::GetWorldSize();
+      auto const rank = collective::GetRank();
+      auto const num_entries = entries.size();
+      std::vector<ExpandEntry> buffer{num_entries * world};
+      std::copy_n(entries.cbegin(), num_entries, buffer.begin() + num_entries * rank);
+      collective::Allgather(buffer.data(), buffer.size() * sizeof(ExpandEntry));
+      for (auto worker = 0; worker < world; ++worker) {
+        for (auto nidx_in_set = 0; nidx_in_set < entries.size(); ++nidx_in_set) {
+          entries[nidx_in_set].split.Update(buffer[worker * num_entries + nidx_in_set].split);
+        }
+      }
+    }
   }
+
   // Add splits to tree, handles all statistic
   void ApplyTreeSplit(ExpandEntry const& candidate, RegTree *p_tree) {
     auto evaluator = tree_evaluator_.GetEvaluator();
@@ -429,7 +447,8 @@ class HistEvaluator {
                          std::shared_ptr<common::ColumnSampler> sampler)
       : ctx_{ctx}, param_{param},
         column_sampler_{std::move(sampler)},
-        tree_evaluator_{param, static_cast<bst_feature_t>(info.num_col_), Context::kCpuId} {
+        tree_evaluator_{param, static_cast<bst_feature_t>(info.num_col_), Context::kCpuId},
+        is_col_split_{info.data_split_mode == DataSplitMode::kCol} {
     interaction_constraints_.Configure(param, info.num_col_);
     column_sampler_->Init(ctx, info.num_col_, info.feature_weights.HostVector(),
                           param_.colsample_bynode, param_.colsample_bylevel,

src/tree/hist/histogram.h

@@ -98,7 +98,7 @@ class HistogramBuilder {
                    std::vector<ExpandEntry> const &nodes_for_explicit_hist_build,
                    std::vector<ExpandEntry> const &nodes_for_subtraction_trick,
                    RegTree const *p_tree) {
-    if (is_distributed_) {
+    if (is_distributed_ && !is_col_split_) {
       this->AddHistRowsDistributed(starting_index, sync_count, nodes_for_explicit_hist_build,
                                    nodes_for_subtraction_trick, p_tree);
     } else {

src/tree/updater_approx.cc

@@ -90,7 +90,9 @@ class GloablApproxBuilder {
     for (auto const &g : gpair) {
       root_sum.Add(g);
     }
-    collective::Allreduce<collective::Operation::kSum>(reinterpret_cast<double *>(&root_sum), 2);
+    if (p_fmat->IsRowSplit()) {
+      collective::Allreduce<collective::Operation::kSum>(reinterpret_cast<double *>(&root_sum), 2);
+    }
     std::vector<CPUExpandEntry> nodes{best};
     size_t i = 0;
     auto space = ConstructHistSpace(partitioner_, nodes);

tests/cpp/tree/test_histmaker.cc

@@ -1,5 +1,4 @@
 #include <gtest/gtest.h>
-
 #include <xgboost/tree_model.h>
 #include <xgboost/tree_updater.h>
 
@@ -8,26 +7,34 @@
 namespace xgboost {
 namespace tree {
 
-TEST(GrowHistMaker, InteractionConstraint) {
-  size_t constexpr kRows = 32;
-  size_t constexpr kCols = 16;
+std::shared_ptr<DMatrix> GenerateDMatrix(std::size_t rows, std::size_t cols){
+  return RandomDataGenerator{rows, cols, 0.6f}.Seed(3).GenerateDMatrix();
+}
 
-  Context ctx;
-
-  auto p_dmat = RandomDataGenerator{kRows, kCols, 0.6f}.Seed(3).GenerateDMatrix();
-
-  HostDeviceVector<GradientPair> gradients (kRows);
-  std::vector<GradientPair>& h_gradients = gradients.HostVector();
+std::unique_ptr<HostDeviceVector<GradientPair>> GenerateGradients(std::size_t rows) {
+  auto p_gradients = std::make_unique<HostDeviceVector<GradientPair>>(rows);
+  auto& h_gradients = p_gradients->HostVector();
 
   xgboost::SimpleLCG gen;
   xgboost::SimpleRealUniformDistribution<bst_float> dist(0.0f, 1.0f);
 
-  for (size_t i = 0; i < kRows; ++i) {
-    bst_float grad = dist(&gen);
-    bst_float hess = dist(&gen);
-    h_gradients[i] = GradientPair(grad, hess);
+  for (auto i = 0; i < rows; ++i) {
+    auto grad = dist(&gen);
+    auto hess = dist(&gen);
+    h_gradients[i] = GradientPair{grad, hess};
   }
 
+  return p_gradients;
+}
+
+TEST(GrowHistMaker, InteractionConstraint)
+{
+  auto constexpr kRows = 32;
+  auto constexpr kCols = 16;
+  auto p_dmat = GenerateDMatrix(kRows, kCols);
+  auto p_gradients = GenerateGradients(kRows);
+
+  Context ctx;
   {
     // With constraints
     RegTree tree;
@@ -39,7 +46,7 @@ TEST(GrowHistMaker, InteractionConstraint) {
         {"interaction_constraints", "[[0, 1]]"},
         {"num_feature", std::to_string(kCols)}});
     std::vector<HostDeviceVector<bst_node_t>> position(1);
-    updater->Update(&gradients, p_dmat.get(), position, {&tree});
+    updater->Update(p_gradients.get(), p_dmat.get(), position, {&tree});
 
     ASSERT_EQ(tree.NumExtraNodes(), 4);
     ASSERT_EQ(tree[0].SplitIndex(), 1);
@@ -56,7 +63,7 @@ TEST(GrowHistMaker, InteractionConstraint) {
         TreeUpdater::Create("grow_histmaker", &ctx, ObjInfo{ObjInfo::kRegression})};
     updater->Configure(Args{{"num_feature", std::to_string(kCols)}});
     std::vector<HostDeviceVector<bst_node_t>> position(1);
-    updater->Update(&gradients, p_dmat.get(), position, {&tree});
+    updater->Update(p_gradients.get(), p_dmat.get(), position, {&tree});
 
     ASSERT_EQ(tree.NumExtraNodes(), 10);
     ASSERT_EQ(tree[0].SplitIndex(), 1);
@@ -66,5 +73,53 @@ TEST(GrowHistMaker, InteractionConstraint) {
   }
 }
 
+namespace {
+void TestColumnSplit(int32_t rows, int32_t cols, RegTree const& expected_tree) {
+  auto p_dmat = GenerateDMatrix(rows, cols);
+  auto p_gradients = GenerateGradients(rows);
+  Context ctx;
+  std::unique_ptr<TreeUpdater> updater{
+      TreeUpdater::Create("grow_histmaker", &ctx, ObjInfo{ObjInfo::kRegression})};
+  updater->Configure(Args{{"num_feature", std::to_string(cols)}});
+  std::vector<HostDeviceVector<bst_node_t>> position(1);
+
+  std::unique_ptr<DMatrix> sliced{
+      p_dmat->SliceCol(collective::GetWorldSize(), collective::GetRank())};
+
+  RegTree tree;
+  tree.param.num_feature = cols;
+  updater->Update(p_gradients.get(), sliced.get(), position, {&tree});
+
+  EXPECT_EQ(tree.NumExtraNodes(), 10);
+  EXPECT_EQ(tree[0].SplitIndex(), 1);
+
+  EXPECT_NE(tree[tree[0].LeftChild()].SplitIndex(), 0);
+  EXPECT_NE(tree[tree[0].RightChild()].SplitIndex(), 0);
+
+  EXPECT_EQ(tree, expected_tree);
+}
+}  // anonymous namespace
+
+TEST(GrowHistMaker, ColumnSplit) {
+  auto constexpr kRows = 32;
+  auto constexpr kCols = 16;
+
+  RegTree expected_tree;
+  expected_tree.param.num_feature = kCols;
+  {
+    auto p_dmat = GenerateDMatrix(kRows, kCols);
+    auto p_gradients = GenerateGradients(kRows);
+    Context ctx;
+    std::unique_ptr<TreeUpdater> updater{
+        TreeUpdater::Create("grow_histmaker", &ctx, ObjInfo{ObjInfo::kRegression})};
+    updater->Configure(Args{{"num_feature", std::to_string(kCols)}});
+    std::vector<HostDeviceVector<bst_node_t>> position(1);
+    updater->Update(p_gradients.get(), p_dmat.get(), position, {&expected_tree});
+  }
+
+  auto constexpr kWorldSize = 2;
+  RunWithInMemoryCommunicator(kWorldSize, TestColumnSplit, kRows, kCols, std::cref(expected_tree));
+}
+
 }  // namespace tree
 }  // namespace xgboost