Use realloc for histogram cache and expose the cache limit. (#9455)

doc/parameter.rst

@@ -226,6 +226,15 @@ Parameters for Tree Booster
     - ``one_output_per_tree``: One model for each target.
     - ``multi_output_tree``:  Use multi-target trees.
 
+* ``max_cached_hist_node``, [default = 65536]
+
+  Maximum number of cached nodes for CPU histogram.
+
+  .. versionadded:: 2.0.0
+
+  - For most of the cases this parameter should not be set except for growing deep trees
+    on CPU.
+
 .. _cat-param:
 
 Parameters for Categorical Feature

python-package/xgboost/testing/params.py

@@ -42,7 +42,7 @@ hist_parameter_strategy = strategies.fixed_dictionaries(
 )
 
 hist_cache_strategy = strategies.fixed_dictionaries(
-    {"internal_max_cached_hist_node": strategies.sampled_from([1, 4, 1024, 2**31])}
+    {"max_cached_hist_node": strategies.sampled_from([1, 4, 1024, 2**31])}
 )
 
 hist_multi_parameter_strategy = strategies.fixed_dictionaries(

src/common/ref_resource_view.h

@@ -35,6 +35,13 @@ class RefResourceView {
   size_type size_{0};
   std::shared_ptr<common::ResourceHandler> mem_{nullptr};
 
+ protected:
+  void Init(value_type* ptr, size_type size, std::shared_ptr<common::ResourceHandler> mem) {
+    ptr_ = ptr;
+    size_ = size;
+    mem_ = std::move(mem);
+  }
+
  public:
   RefResourceView(value_type* ptr, size_type n, std::shared_ptr<common::ResourceHandler> mem)
       : ptr_{ptr}, size_{n}, mem_{std::move(mem)} {
@@ -60,11 +67,11 @@ class RefResourceView {
 
   RefResourceView() = default;
   RefResourceView(RefResourceView const& that) = delete;
-  RefResourceView(RefResourceView&& that) = delete;
   RefResourceView& operator=(RefResourceView const& that) = delete;
   /**
    * @brief We allow move assignment for lazy initialization.
    */
+  RefResourceView(RefResourceView&& that) = default;
   RefResourceView& operator=(RefResourceView&& that) = default;
 
   [[nodiscard]] size_type size() const { return size_; }  // NOLINT
@@ -154,5 +161,33 @@ template <typename T>
   auto resource = std::make_shared<common::MallocResource>(n_elements * sizeof(T));
   return RefResourceView{resource->DataAs<T>(), n_elements, resource, init};
 }
+
+template <typename T>
+class ReallocVector : public RefResourceView<T> {
+  static_assert(!std::is_reference_v<T>);
+  static_assert(!std::is_const_v<T>);
+  static_assert(std::is_trivially_copyable_v<T>);
+
+  using Upper = RefResourceView<T>;
+  using size_type = typename Upper::size_type;    // NOLINT
+  using value_type = typename Upper::value_type;  // NOLINT
+
+ public:
+  ReallocVector() : RefResourceView<T>{MakeFixedVecWithMalloc(0, T{})} {}
+
+  ReallocVector(size_type n, value_type const& init)
+      : RefResourceView<T>{MakeFixedVecWithMalloc(n, init)} {}
+  ReallocVector(ReallocVector const& that) = delete;
+  ReallocVector(ReallocVector&& that) = delete;
+  ReallocVector& operator=(ReallocVector const& that) = delete;
+  ReallocVector& operator=(ReallocVector&& that) = delete;
+
+  void Resize(typename Upper::size_type new_size) {
+    auto resource = std::dynamic_pointer_cast<common::MallocResource>(this->Resource());
+    CHECK(resource);
+    resource->Resize(new_size * sizeof(T));
+    this->Init(resource->template DataAs<T>(), new_size, resource);
+  }
+};
 }  // namespace xgboost::common
 #endif  // XGBOOST_COMMON_REF_RESOURCE_VIEW_H_

src/tree/hist/hist_cache.h

@@ -5,12 +5,14 @@
 #define XGBOOST_TREE_HIST_HIST_CACHE_H_
 #include <cstddef>  // for size_t
 #include <map>      // for map
+#include <memory>   // for unique_ptr
 #include <vector>   // for vector
 
-#include "../../common/hist_util.h"  // for GHistRow, ConstGHistRow
+#include "../../common/hist_util.h"          // for GHistRow, ConstGHistRow
-#include "xgboost/base.h"            // for bst_node_t, bst_bin_t
+#include "../../common/ref_resource_view.h"  // for ReallocVector
-#include "xgboost/logging.h"         // for CHECK_GT
+#include "xgboost/base.h"                    // for bst_node_t, bst_bin_t
-#include "xgboost/span.h"            // for Span
+#include "xgboost/logging.h"                 // for CHECK_GT
+#include "xgboost/span.h"                    // for Span
 
 namespace xgboost::tree {
 /**
@@ -32,7 +34,8 @@ class BoundedHistCollection {
   std::size_t current_size_{0};
 
   // stores the histograms in a contiguous buffer
-  std::vector<GradientPairPrecise> data_;
+  using Vec = common::ReallocVector<GradientPairPrecise>;
+  std::unique_ptr<Vec> data_{new Vec{}};  // nvcc 12.1 trips over std::make_unique
 
   // number of histogram bins across all features
   bst_bin_t n_total_bins_{0};
@@ -42,13 +45,14 @@ class BoundedHistCollection {
   bool has_exceeded_{false};
 
  public:
+  BoundedHistCollection() = default;
   common::GHistRow operator[](std::size_t idx) {
     auto offset = node_map_.at(idx);
-    return common::Span{data_.data(), data_.size()}.subspan(offset, n_total_bins_);
+    return common::Span{data_->data(), data_->size()}.subspan(offset, n_total_bins_);
   }
   common::ConstGHistRow operator[](std::size_t idx) const {
     auto offset = node_map_.at(idx);
-    return common::Span{data_.data(), data_.size()}.subspan(offset, n_total_bins_);
+    return common::Span{data_->data(), data_->size()}.subspan(offset, n_total_bins_);
   }
   void Reset(bst_bin_t n_total_bins, std::size_t n_cached_nodes) {
     n_total_bins_ = n_total_bins;
@@ -81,8 +85,8 @@ class BoundedHistCollection {
     auto n_new_nodes = nodes_to_build.size() + nodes_to_sub.size();
     auto alloc_size = n_new_nodes * n_total_bins_;
     auto new_size = alloc_size + current_size_;
-    if (new_size > data_.size()) {
+    if (new_size > data_->size()) {
-      data_.resize(new_size);
+      data_->Resize(new_size);
     }
     for (auto nidx : nodes_to_build) {
       node_map_[nidx] = current_size_;

src/tree/hist/histogram.h

@@ -63,7 +63,7 @@ class HistogramBuilder {
              bool is_col_split, HistMakerTrainParam const *param) {
     n_threads_ = ctx->Threads();
     param_ = p;
-    hist_.Reset(total_bins, param->internal_max_cached_hist_node);
+    hist_.Reset(total_bins, param->max_cached_hist_node);
     buffer_.Init(total_bins);
     is_distributed_ = is_distributed;
     is_col_split_ = is_col_split;

src/tree/hist/param.h

@@ -13,7 +13,7 @@ struct HistMakerTrainParam : public XGBoostParameter<HistMakerTrainParam> {
   constexpr static std::size_t DefaultNodes() { return static_cast<std::size_t>(1) << 16; }
 
   bool debug_synchronize{false};
-  std::size_t internal_max_cached_hist_node{DefaultNodes()};
+  std::size_t max_cached_hist_node{DefaultNodes()};
 
   void CheckTreesSynchronized(RegTree const* local_tree) const;
 
@@ -22,7 +22,7 @@ struct HistMakerTrainParam : public XGBoostParameter<HistMakerTrainParam> {
     DMLC_DECLARE_FIELD(debug_synchronize)
         .set_default(false)
         .describe("Check if all distributed tree are identical after tree construction.");
-    DMLC_DECLARE_FIELD(internal_max_cached_hist_node)
+    DMLC_DECLARE_FIELD(max_cached_hist_node)
         .set_default(DefaultNodes())
         .set_lower_bound(1)
         .describe("Maximum number of nodes in CPU histogram cache. Only for internal usage.");

src/tree/updater_gpu_hist.cu

@@ -866,6 +866,9 @@ class GPUGlobalApproxMaker : public TreeUpdater {
     // Used in test to count how many configurations are performed
     LOG(DEBUG) << "[GPU Approx]: Configure";
     hist_maker_param_.UpdateAllowUnknown(args);
+    if (hist_maker_param_.max_cached_hist_node != HistMakerTrainParam::DefaultNodes()) {
+      LOG(WARNING) << "The `max_cached_hist_node` is ignored in GPU.";
+    }
     dh::CheckComputeCapability();
     initialised_ = false;
 

tests/cpp/tree/hist/test_evaluate_splits.cc

@@ -51,7 +51,7 @@ void TestEvaluateSplits(bool force_read_by_column) {
   row_set_collection.Init();
 
   HistMakerTrainParam hist_param;
-  hist.Reset(gmat.cut.Ptrs().back(), hist_param.internal_max_cached_hist_node);
+  hist.Reset(gmat.cut.Ptrs().back(), hist_param.max_cached_hist_node);
   hist.AllocateHistograms({0});
   common::BuildHist<false>(row_gpairs, row_set_collection[0], gmat, hist[0], force_read_by_column);
 
@@ -118,7 +118,7 @@ TEST(HistMultiEvaluator, Evaluate) {
   linalg::Vector<GradientPairPrecise> root_sum({2}, Context::kCpuId);
   for (bst_target_t t{0}; t < n_targets; ++t) {
     auto &hist = histogram[t];
-    hist.Reset(n_bins * n_features, hist_param.internal_max_cached_hist_node);
+    hist.Reset(n_bins * n_features, hist_param.max_cached_hist_node);
     hist.AllocateHistograms({0});
     auto node_hist = hist[0];
     node_hist[0] = {-0.5, 0.5};
@@ -235,7 +235,7 @@ auto CompareOneHotAndPartition(bool onehot) {
     entries.front().nid = 0;
     entries.front().depth = 0;
 
-    hist.Reset(gmat.cut.TotalBins(), hist_param.internal_max_cached_hist_node);
+    hist.Reset(gmat.cut.TotalBins(), hist_param.max_cached_hist_node);
     hist.AllocateHistograms({0});
     auto node_hist = hist[0];
 
@@ -265,7 +265,7 @@ TEST(HistEvaluator, Categorical) {
 TEST_F(TestCategoricalSplitWithMissing, HistEvaluator) {
   BoundedHistCollection hist;
   HistMakerTrainParam hist_param;
-  hist.Reset(cuts_.TotalBins(), hist_param.internal_max_cached_hist_node);
+  hist.Reset(cuts_.TotalBins(), hist_param.max_cached_hist_node);
   hist.AllocateHistograms({0});
   auto node_hist = hist[0];
   ASSERT_EQ(node_hist.size(), feature_histogram_.size());

tests/cpp/tree/hist/test_histogram.cc

@@ -516,7 +516,7 @@ class OverflowTest : public ::testing::TestWithParam<std::tuple<bool, bool>> {
     Context ctx;
     HistMakerTrainParam hist_param;
     if (limit) {
-      hist_param.Init(Args{{"internal_max_cached_hist_node", "1"}});
+      hist_param.Init(Args{{"max_cached_hist_node", "1"}});
     }
 
     std::shared_ptr<DMatrix> Xy =

tests/cpp/tree/test_evaluate_splits.h

@@ -59,7 +59,7 @@ class TestPartitionBasedSplit : public ::testing::Test {
     cuts_.min_vals_.Resize(1);
 
     HistMakerTrainParam hist_param;
-    hist_.Reset(cuts_.TotalBins(), hist_param.internal_max_cached_hist_node);
+    hist_.Reset(cuts_.TotalBins(), hist_param.max_cached_hist_node);
     hist_.AllocateHistograms({0});
     auto node_hist = hist_[0];