External memory support for hist (#7531)

* Generate column matrix from gHistIndex.
* Avoid synchronization with the sparse page once the cache is written.
* Cleanups: Remove member variables/functions, change the update routine to look like approx and gpu_hist.
* Remove pruner.

amalgamation/xgboost-all0.cc

@@ -48,17 +48,18 @@
 #include "../src/predictor/cpu_predictor.cc"
 
 // trees
+#include "../src/tree/constraints.cc"
+#include "../src/tree/hist/param.cc"
 #include "../src/tree/param.cc"
 #include "../src/tree/tree_model.cc"
 #include "../src/tree/tree_updater.cc"
+#include "../src/tree/updater_approx.cc"
 #include "../src/tree/updater_colmaker.cc"
-#include "../src/tree/updater_quantile_hist.cc"
+#include "../src/tree/updater_histmaker.cc"
 #include "../src/tree/updater_prune.cc"
+#include "../src/tree/updater_quantile_hist.cc"
 #include "../src/tree/updater_refresh.cc"
 #include "../src/tree/updater_sync.cc"
-#include "../src/tree/updater_histmaker.cc"
-#include "../src/tree/updater_approx.cc"
-#include "../src/tree/constraints.cc"
 
 // linear
 #include "../src/linear/linear_updater.cc"

demo/guide-python/external_memory.py

@@ -7,6 +7,9 @@ instead of Quantile DMatrix.  The feature is not ready for production use yet.
 
     .. versionadded:: 1.5.0
 
+
+See :doc:`the tutorial </tutorials/external_memory>` for more details.
+
 """
 import os
 import xgboost
@@ -77,9 +80,14 @@ def main(tmpdir: str) -> xgboost.Booster:
     missing = np.NaN
     Xy = xgboost.DMatrix(it, missing=missing, enable_categorical=False)
 
-    # Other tree methods including ``hist`` and ``gpu_hist`` also work, but has some
+    # Other tree methods including ``hist`` and ``gpu_hist`` also work, see tutorial in
-    # caveats.  This is still an experimental feature.
+    # doc for details.
-    booster = xgboost.train({"tree_method": "approx"}, Xy, evals=[(Xy, "Train")])
+    booster = xgboost.train(
+        {"tree_method": "approx", "max_depth": 2},
+        Xy,
+        evals=[(Xy, "Train")],
+        num_boost_round=10,
+    )
     return booster
 
 

demo/guide-python/feature_weights.py

@@ -27,7 +27,7 @@ def main(args):
     dtrain.set_info(feature_weights=fw)
 
     bst = xgboost.train({'tree_method': 'hist',
-                         'colsample_bynode': 0.5},
+                         'colsample_bynode': 0.2},
                         dtrain, num_boost_round=10,
                         evals=[(dtrain, 'd')])
     feature_map = bst.get_fscore()

doc/tutorials/external_memory.rst

@@ -127,9 +127,12 @@ the tree method still concatenate all the chunks into 1 final histogram index du
 performance reason, but in compressed format.  So its scalability has an upper bound but
 still has lower memory cost in general.
 
-********
+***********
-CPU Hist
+CPU Version
-********
+***********
 
-It's limited by the same factor of GPU Hist, except that gradient based sampling is not
+For CPU histogram based tree methods (``approx``, ``hist``) it's recommended to use
-yet supported on CPU.
+``grow_policy=depthwise`` for performance reason.  Iterating over data batches is slow,
+with ``depthwise`` policy XGBoost can build a entire layer of tree nodes with a few
+iterations, while with ``lossguide`` XGBoost needs to iterate over the data set for each
+tree node.

src/common/column_matrix.h

@@ -1,5 +1,5 @@
 /*!
- * Copyright 2017 by Contributors
+ * Copyright 2017-2022 by Contributors
  * \file column_matrix.h
  * \brief Utility for fast column-wise access
  * \author Philip Cho
@@ -8,21 +8,22 @@
 #ifndef XGBOOST_COMMON_COLUMN_MATRIX_H_
 #define XGBOOST_COMMON_COLUMN_MATRIX_H_
 
+#include <dmlc/endian.h>
+
+#include <algorithm>
 #include <limits>
-#include <vector>
 #include <memory>
-#include "hist_util.h"
+#include <vector>
+
 #include "../data/gradient_index.h"
+#include "hist_util.h"
 
 namespace xgboost {
 namespace common {
 
 class ColumnMatrix;
 /*! \brief column type */
-enum ColumnType {
+enum ColumnType : uint8_t { kDenseColumn, kSparseColumn };
-  kDenseColumn,
-  kSparseColumn
-};
 
 /*! \brief a column storage, to be used with ApplySplit. Note that each
     bin id is stored as index[i] + index_base.
@@ -34,9 +35,7 @@ class Column {
   static constexpr int32_t kMissingId = -1;
 
   Column(ColumnType type, common::Span<const BinIdxType> index, const uint32_t index_base)
-      : type_(type),
+      : type_(type), index_(index), index_base_(index_base) {}
-        index_(index),
-        index_base_(index_base) {}
 
   virtual ~Column() = default;
 
@@ -67,10 +66,9 @@ class Column {
 template <typename BinIdxType>
 class SparseColumn : public Column<BinIdxType> {
  public:
-  SparseColumn(ColumnType type, common::Span<const BinIdxType> index,
+  SparseColumn(ColumnType type, common::Span<const BinIdxType> index, uint32_t index_base,
-              uint32_t index_base, common::Span<const size_t> row_ind)
+               common::Span<const size_t> row_ind)
-      : Column<BinIdxType>(type, index, index_base),
+      : Column<BinIdxType>(type, index, index_base), row_ind_(row_ind) {}
-        row_ind_(row_ind) {}
 
   const size_t* GetRowData() const { return row_ind_.data(); }
 
@@ -98,9 +96,7 @@ class SparseColumn: public Column<BinIdxType> {
     return p - row_data;
   }
 
-  size_t GetRowIdx(size_t idx) const {
+  size_t GetRowIdx(size_t idx) const { return row_ind_.data()[idx]; }
-    return row_ind_.data()[idx];
-  }
 
  private:
   /* indexes of rows */
@@ -110,9 +106,8 @@ class SparseColumn: public Column<BinIdxType> {
 template <typename BinIdxType, bool any_missing>
 class DenseColumn : public Column<BinIdxType> {
  public:
-  DenseColumn(ColumnType type, common::Span<const BinIdxType> index,
+  DenseColumn(ColumnType type, common::Span<const BinIdxType> index, uint32_t index_base,
-              uint32_t index_base, const std::vector<bool>& missing_flags,
+              const std::vector<bool>& missing_flags, size_t feature_offset)
-              size_t feature_offset)
       : Column<BinIdxType>(type, index, index_base),
         missing_flags_(missing_flags),
         feature_offset_(feature_offset) {}
@@ -126,9 +121,7 @@ class DenseColumn: public Column<BinIdxType> {
     }
   }
 
-  size_t GetInitialState(const size_t first_row_id) const {
+  size_t GetInitialState(const size_t first_row_id) const { return 0; }
-    return 0;
-  }
 
  private:
   /* flags for missing values in dense columns */
@@ -141,28 +134,26 @@ class DenseColumn: public Column<BinIdxType> {
 class ColumnMatrix {
  public:
   // get number of features
-  inline bst_uint GetNumFeature() const {
+  bst_feature_t GetNumFeature() const { return static_cast<bst_feature_t>(type_.size()); }
-    return static_cast<bst_uint>(type_.size());
-  }
 
   // construct column matrix from GHistIndexMatrix
-  inline void Init(const GHistIndexMatrix& gmat, double sparse_threshold, int32_t n_threads) {
+  inline void Init(SparsePage const& page, const GHistIndexMatrix& gmat, double sparse_threshold,
-    const int32_t nfeature = static_cast<int32_t>(gmat.cut.Ptrs().size() - 1);
+                   int32_t n_threads) {
+    auto const nfeature = static_cast<bst_feature_t>(gmat.cut.Ptrs().size() - 1);
     const size_t nrow = gmat.row_ptr.size() - 1;
     // identify type of each column
     feature_counts_.resize(nfeature);
     type_.resize(nfeature);
     std::fill(feature_counts_.begin(), feature_counts_.end(), 0);
     uint32_t max_val = std::numeric_limits<uint32_t>::max();
-    for (int32_t fid = 0; fid < nfeature; ++fid) {
+    for (bst_feature_t fid = 0; fid < nfeature; ++fid) {
       CHECK_LE(gmat.cut.Ptrs()[fid + 1] - gmat.cut.Ptrs()[fid], max_val);
     }
     bool all_dense = gmat.IsDense();
     gmat.GetFeatureCounts(&feature_counts_[0]);
     // classify features
-    for (int32_t fid = 0; fid < nfeature; ++fid) {
+    for (bst_feature_t fid = 0; fid < nfeature; ++fid) {
-      if (static_cast<double>(feature_counts_[fid])
+      if (static_cast<double>(feature_counts_[fid]) < sparse_threshold * nrow) {
-                 < sparse_threshold * nrow) {
         type_[fid] = kSparseColumn;
         all_dense = false;
       } else {
@@ -175,7 +166,7 @@ class ColumnMatrix {
     feature_offsets_.resize(nfeature + 1);
     size_t accum_index_ = 0;
     feature_offsets_[0] = accum_index_;
-    for (int32_t fid = 1; fid < nfeature + 1; ++fid) {
+    for (bst_feature_t fid = 1; fid < nfeature + 1; ++fid) {
       if (type_[fid - 1] == kDenseColumn) {
         accum_index_ += static_cast<size_t>(nrow);
       } else {
@@ -197,6 +188,7 @@ class ColumnMatrix {
     const bool noMissingValues = NoMissingValues(gmat.row_ptr[nrow], nrow, nfeature);
     any_missing_ = !noMissingValues;
 
+    missing_flags_.clear();
     if (noMissingValues) {
       missing_flags_.resize(feature_offsets_[nfeature], false);
     } else {
@@ -207,26 +199,26 @@ class ColumnMatrix {
     if (all_dense) {
       BinTypeSize gmat_bin_size = gmat.index.GetBinTypeSize();
       if (gmat_bin_size == kUint8BinsTypeSize) {
-        SetIndexAllDense(gmat.index.data<uint8_t>(), gmat, nrow, nfeature, noMissingValues,
+        SetIndexAllDense(page, gmat.index.data<uint8_t>(), gmat, nrow, nfeature, noMissingValues,
                          n_threads);
       } else if (gmat_bin_size == kUint16BinsTypeSize) {
-        SetIndexAllDense(gmat.index.data<uint16_t>(), gmat, nrow, nfeature, noMissingValues,
+        SetIndexAllDense(page, gmat.index.data<uint16_t>(), gmat, nrow, nfeature, noMissingValues,
                          n_threads);
       } else {
         CHECK_EQ(gmat_bin_size, kUint32BinsTypeSize);
-        SetIndexAllDense(gmat.index.data<uint32_t>(), gmat, nrow, nfeature, noMissingValues,
+        SetIndexAllDense(page, gmat.index.data<uint32_t>(), gmat, nrow, nfeature, noMissingValues,
                          n_threads);
       }
       /* For sparse DMatrix gmat.index.getBinTypeSize() returns always kUint32BinsTypeSize
          but for ColumnMatrix we still have a chance to reduce the memory consumption */
     } else {
       if (bins_type_size_ == kUint8BinsTypeSize) {
-          SetIndex<uint8_t>(gmat.index.data<uint32_t>(), gmat, nfeature);
+        SetIndex<uint8_t>(page, gmat.index.data<uint32_t>(), gmat, nfeature);
       } else if (bins_type_size_ == kUint16BinsTypeSize) {
-          SetIndex<uint16_t>(gmat.index.data<uint32_t>(), gmat, nfeature);
+        SetIndex<uint16_t>(page, gmat.index.data<uint32_t>(), gmat, nfeature);
       } else {
         CHECK_EQ(bins_type_size_, kUint32BinsTypeSize);
-          SetIndex<uint32_t>(gmat.index.data<uint32_t>(), gmat, nfeature);
+        SetIndex<uint32_t>(page, gmat.index.data<uint32_t>(), gmat, nfeature);
       }
     }
   }
@@ -250,8 +242,8 @@ class ColumnMatrix {
 
     const size_t feature_offset = feature_offsets_[fid];  // to get right place for certain feature
     const size_t column_size = feature_offsets_[fid + 1] - feature_offset;
-    common::Span<const BinIdxType> bin_index = { reinterpret_cast<const BinIdxType*>(
+    common::Span<const BinIdxType> bin_index = {
-                                                 &index_[feature_offset * bins_type_size_]),
+        reinterpret_cast<const BinIdxType*>(&index_[feature_offset * bins_type_size_]),
         column_size};
     std::unique_ptr<const Column<BinIdxType> > res;
     if (type_[fid] == ColumnType::kDenseColumn) {
@@ -266,8 +258,8 @@ class ColumnMatrix {
   }
 
   template <typename T>
-  inline void SetIndexAllDense(T const* index, const GHistIndexMatrix& gmat, const size_t nrow,
+  inline void SetIndexAllDense(SparsePage const& page, T const* index, const GHistIndexMatrix& gmat,
-                               const size_t nfeature, const bool noMissingValues,
+                               const size_t nrow, const size_t nfeature, const bool noMissingValues,
                                int32_t n_threads) {
     T* local_index = reinterpret_cast<T*>(&index_[0]);
 
@@ -285,51 +277,34 @@ class ColumnMatrix {
       });
     } else {
       /* to handle rows in all batches, sum of all batch sizes equal to gmat.row_ptr.size() - 1 */
-      size_t rbegin = 0;
+      auto get_bin_idx = [&](auto bin_id, auto rid, bst_feature_t fid) {
-      for (const auto &batch : gmat.p_fmat->GetBatches<SparsePage>()) {
+        // T* begin = &local_index[feature_offsets_[fid]];
-        const xgboost::Entry* data_ptr = batch.data.HostVector().data();
-        const std::vector<bst_row_t>& offset_vec = batch.offset.HostVector();
-        const size_t batch_size = batch.Size();
-        CHECK_LT(batch_size, offset_vec.size());
-        for (size_t rid = 0; rid < batch_size; ++rid) {
-          const size_t size = offset_vec[rid + 1] - offset_vec[rid];
-          SparsePage::Inst inst = {data_ptr + offset_vec[rid], size};
-          const size_t ibegin = gmat.row_ptr[rbegin + rid];
-          const size_t iend = gmat.row_ptr[rbegin + rid + 1];
-          CHECK_EQ(ibegin + inst.size(), iend);
-          size_t j = 0;
-          size_t fid = 0;
-          for (size_t i = ibegin; i < iend; ++i, ++j) {
-              fid = inst[j].index;
         const size_t idx = feature_offsets_[fid];
         /* rbegin allows to store indexes from specific SparsePage batch */
-              local_index[idx + rbegin + rid] = index[i];
+        local_index[idx + rid] = bin_id;
-              missing_flags_[idx + rbegin + rid] = false;
+
-          }
+        missing_flags_[idx + rid] = false;
-        }
+      };
-        rbegin += batch.Size();
+      this->SetIndexSparse(page, index, gmat, nfeature, get_bin_idx);
-      }
     }
   }
 
-  template<typename T>
+  // FIXME(jiamingy): In the future we might want to simply use binary search to simplify
-  inline void SetIndex(uint32_t const* index, const GHistIndexMatrix& gmat,
+  // this and remove the dependency on SparsePage.  This way we can have quantilized
-                       const size_t nfeature) {
+  // matrix for host similar to `DeviceQuantileDMatrix`.
-    std::vector<size_t> num_nonzeros;
+  template <typename T, typename BinFn>
-    num_nonzeros.resize(nfeature);
+  void SetIndexSparse(SparsePage const& batch, T* index, const GHistIndexMatrix& gmat,
-    std::fill(num_nonzeros.begin(), num_nonzeros.end(), 0);
+                      const size_t nfeature, BinFn&& assign_bin) {
-
+    std::vector<size_t> num_nonzeros(nfeature, 0ul);
-    T* local_index = reinterpret_cast<T*>(&index_[0]);
-    size_t rbegin = 0;
-    for (const auto &batch : gmat.p_fmat->GetBatches<SparsePage>()) {
     const xgboost::Entry* data_ptr = batch.data.HostVector().data();
     const std::vector<bst_row_t>& offset_vec = batch.offset.HostVector();
-      const size_t batch_size = batch.Size();
+    auto rbegin = 0;
+    const size_t batch_size = gmat.Size();
     CHECK_LT(batch_size, offset_vec.size());
+
     for (size_t rid = 0; rid < batch_size; ++rid) {
       const size_t ibegin = gmat.row_ptr[rbegin + rid];
       const size_t iend = gmat.row_ptr[rbegin + rid + 1];
-        size_t fid = 0;
       const size_t size = offset_vec[rid + 1] - offset_vec[rid];
       SparsePage::Inst inst = {data_ptr + offset_vec[rid], size};
 
@@ -337,36 +312,110 @@ class ColumnMatrix {
       size_t j = 0;
       for (size_t i = ibegin; i < iend; ++i, ++j) {
         const uint32_t bin_id = index[i];
+        auto fid = inst[j].index;
+        assign_bin(bin_id, rid, fid);
+      }
+    }
+  }
 
-          fid = inst[j].index;
+  template <typename T>
+  inline void SetIndex(SparsePage const& page, uint32_t const* index, const GHistIndexMatrix& gmat,
+                       const size_t nfeature) {
+    T* local_index = reinterpret_cast<T*>(&index_[0]);
+    std::vector<size_t> num_nonzeros;
+    num_nonzeros.resize(nfeature);
+    std::fill(num_nonzeros.begin(), num_nonzeros.end(), 0);
+
+    auto get_bin_idx = [&](auto bin_id, auto rid, bst_feature_t fid) {
       if (type_[fid] == kDenseColumn) {
         T* begin = &local_index[feature_offsets_[fid]];
-            begin[rid + rbegin] = bin_id - index_base_[fid];
+        begin[rid] = bin_id - index_base_[fid];
-            missing_flags_[feature_offsets_[fid] + rid + rbegin] = false;
+        missing_flags_[feature_offsets_[fid] + rid] = false;
       } else {
         T* begin = &local_index[feature_offsets_[fid]];
         begin[num_nonzeros[fid]] = bin_id - index_base_[fid];
-            row_ind_[feature_offsets_[fid] + num_nonzeros[fid]] = rid + rbegin;
+        row_ind_[feature_offsets_[fid] + num_nonzeros[fid]] = rid;
         ++num_nonzeros[fid];
       }
-        }
+    };
-      }
+    this->SetIndexSparse(page, index, gmat, nfeature, get_bin_idx);
-      rbegin += batch.Size();
-    }
-  }
-  BinTypeSize GetTypeSize() const {
-    return bins_type_size_;
   }
 
+  BinTypeSize GetTypeSize() const { return bins_type_size_; }
+
   // This is just an utility function
-  bool NoMissingValues(const size_t n_elements,
+  bool NoMissingValues(const size_t n_elements, const size_t n_row, const size_t n_features) {
-                             const size_t n_row, const size_t n_features) {
     return n_elements == n_features * n_row;
   }
 
   // And this returns part of state
-  bool AnyMissing() const {
+  bool AnyMissing() const { return any_missing_; }
-    return any_missing_;
+
+  // IO procedures for external memory.
+  bool Read(dmlc::SeekStream* fi, uint32_t const* index_base) {
+    fi->Read(&index_);
+    fi->Read(&feature_counts_);
+#if !DMLC_LITTLE_ENDIAN
+    // s390x
+    std::vector<std::underlying_type<ColumnType>::type> int_types;
+    fi->Read(&int_types);
+    type_.resize(int_types.size());
+    std::transform(
+        int_types.begin(), int_types.end(), type_.begin(),
+        [](std::underlying_type<ColumnType>::type i) { return static_cast<ColumnType>(i); });
+#else
+    fi->Read(&type_);
+#endif  // !DMLC_LITTLE_ENDIAN
+
+    fi->Read(&row_ind_);
+    fi->Read(&feature_offsets_);
+    index_base_ = index_base;
+#if !DMLC_LITTLE_ENDIAN
+    std::underlying_type<BinTypeSize>::type v;
+    fi->Read(&v);
+    bins_type_size_ = static_cast<BinTypeSize>(v);
+#else
+    fi->Read(&bins_type_size_);
+#endif
+
+    fi->Read(&any_missing_);
+    return true;
+  }
+
+  size_t Write(dmlc::Stream* fo) const {
+    size_t bytes{0};
+
+    auto write_vec = [&](auto const& vec) {
+      fo->Write(vec);
+      bytes += vec.size() * sizeof(typename std::remove_reference_t<decltype(vec)>::value_type) +
+               sizeof(uint64_t);
+    };
+    write_vec(index_);
+    write_vec(feature_counts_);
+#if !DMLC_LITTLE_ENDIAN
+    // s390x
+    std::vector<std::underlying_type<ColumnType>::type> int_types(type_.size());
+    std::transform(type_.begin(), type_.end(), int_types.begin(), [](ColumnType t) {
+      return static_cast<std::underlying_type<ColumnType>::type>(t);
+    });
+    write_vec(int_types);
+#else
+    write_vec(type_);
+#endif  // !DMLC_LITTLE_ENDIAN
+    write_vec(row_ind_);
+    write_vec(feature_offsets_);
+
+#if !DMLC_LITTLE_ENDIAN
+    auto v = static_cast<std::underlying_type<BinTypeSize>::type>(bins_type_size_);
+    fo->Write(v);
+#else
+    fo->Write(bins_type_size_);
+#endif  // DMLC_LITTLE_ENDIAN
+    bytes += sizeof(bins_type_size_);
+    fo->Write(any_missing_);
+    bytes += sizeof(any_missing_);
+
+    return bytes;
   }
 
  private:

src/data/ellpack_page_source.cu

@@ -1,5 +1,5 @@
 /*!
- * Copyright 2019-2021 XGBoost contributors
+ * Copyright 2019-2022 XGBoost contributors
  */
 #include <memory>
 #include <utility>
@@ -12,6 +12,13 @@ namespace data {
 void EllpackPageSource::Fetch() {
   dh::safe_cuda(cudaSetDevice(param_.gpu_id));
   if (!this->ReadCache()) {
+    if (count_ != 0 && !sync_) {
+      // source is initialized to be the 0th page during construction, so when count_ is 0
+      // there's no need to increment the source.
+      ++(*source_);
+    }
+    // This is not read from cache so we still need it to be synced with sparse page source.
+    CHECK_EQ(count_, source_->Iter());
     auto const &csr = source_->Page();
     this->page_.reset(new EllpackPage{});
     auto *impl = this->page_->Impl();

src/data/ellpack_page_source.h

@@ -1,5 +1,5 @@
 /*!
- * Copyright 2019-2021 by XGBoost Contributors
+ * Copyright 2019-2022 by XGBoost Contributors
  */
 
 #ifndef XGBOOST_DATA_ELLPACK_PAGE_SOURCE_H_
@@ -25,15 +25,17 @@ class EllpackPageSource : public PageSourceIncMixIn<EllpackPage> {
   std::unique_ptr<common::HistogramCuts> cuts_;
 
  public:
-  EllpackPageSource(
+  EllpackPageSource(float missing, int nthreads, bst_feature_t n_features, size_t n_batches,
-      float missing, int nthreads, bst_feature_t n_features, size_t n_batches,
                     std::shared_ptr<Cache> cache, BatchParam param,
-      std::unique_ptr<common::HistogramCuts> cuts, bool is_dense,
+                    std::unique_ptr<common::HistogramCuts> cuts, bool is_dense, size_t row_stride,
-      size_t row_stride, common::Span<FeatureType const> feature_types,
+                    common::Span<FeatureType const> feature_types,
                     std::shared_ptr<SparsePageSource> source)
-      : PageSourceIncMixIn(missing, nthreads, n_features, n_batches, cache),
+      : PageSourceIncMixIn(missing, nthreads, n_features, n_batches, cache, false),
-        is_dense_{is_dense}, row_stride_{row_stride}, param_{std::move(param)},
+        is_dense_{is_dense},
-        feature_types_{feature_types}, cuts_{std::move(cuts)} {
+        row_stride_{row_stride},
+        param_{std::move(param)},
+        feature_types_{feature_types},
+        cuts_{std::move(cuts)} {
     this->source_ = source;
     this->Fetch();
   }

src/data/gradient_index.cc

@@ -144,7 +144,6 @@ void GHistIndexMatrix::Init(DMatrix *p_fmat, int max_bins, double sparse_thresh,
   hit_count.resize(nbins, 0);
   hit_count_tloc_.resize(n_threads * nbins, 0);
 
-  this->p_fmat = p_fmat;
   size_t new_size = 1;
   for (const auto &batch : p_fmat->GetBatches<SparsePage>()) {
     new_size += batch.Size();
@@ -164,6 +163,16 @@ void GHistIndexMatrix::Init(DMatrix *p_fmat, int max_bins, double sparse_thresh,
     prev_sum = row_ptr[rbegin + batch.Size()];
     rbegin += batch.Size();
   }
+  this->columns_ = std::make_unique<common::ColumnMatrix>();
+
+  // hessian is empty when hist tree method is used or when dataset is empty
+  if (hess.empty() && !std::isnan(sparse_thresh)) {
+    // hist
+    CHECK(!sorted_sketch);
+    for (auto const &page : p_fmat->GetBatches<SparsePage>()) {
+      this->columns_->Init(page, *this, sparse_thresh, n_threads);
+    }
+  }
 }
 
 void GHistIndexMatrix::Init(SparsePage const &batch, common::Span<FeatureType const> ft,
@@ -187,6 +196,10 @@ void GHistIndexMatrix::Init(SparsePage const &batch, common::Span<FeatureType co
   size_t prev_sum = 0;
 
   this->PushBatch(batch, ft, rbegin, prev_sum, nbins, n_threads);
+  this->columns_ = std::make_unique<common::ColumnMatrix>();
+  if (!std::isnan(sparse_thresh)) {
+    this->columns_->Init(batch, *this, sparse_thresh, n_threads);
+  }
 }
 
 void GHistIndexMatrix::ResizeIndex(const size_t n_index, const bool isDense) {
@@ -205,4 +218,17 @@ void GHistIndexMatrix::ResizeIndex(const size_t n_index, const bool isDense) {
     index.Resize((sizeof(uint32_t)) * n_index);
   }
 }
+
+common::ColumnMatrix const &GHistIndexMatrix::Transpose() const {
+  CHECK(columns_);
+  return *columns_;
+}
+
+bool GHistIndexMatrix::ReadColumnPage(dmlc::SeekStream *fi) {
+  return this->columns_->Read(fi, this->cut.Ptrs().data());
+}
+
+size_t GHistIndexMatrix::WriteColumnPage(dmlc::Stream *fo) const {
+  return this->columns_->Write(fo);
+}
 }  // namespace xgboost

src/data/gradient_index.h

@@ -40,7 +40,6 @@ class GHistIndexMatrix {
   std::vector<size_t> hit_count;
   /*! \brief The corresponding cuts */
   common::HistogramCuts cut;
-  DMatrix* p_fmat;
   /*! \brief max_bin for each feature. */
   size_t max_num_bins;
   /*! \brief base row index for current page (used by external memory) */
@@ -119,8 +118,12 @@ class GHistIndexMatrix {
     return row_ptr.empty() ? 0 : row_ptr.size() - 1;
   }
 
+  bool ReadColumnPage(dmlc::SeekStream* fi);
+  size_t WriteColumnPage(dmlc::Stream* fo) const;
+
+  common::ColumnMatrix const& Transpose() const;
+
  private:
-  // unused at the moment: https://github.com/dmlc/xgboost/pull/7531
   std::unique_ptr<common::ColumnMatrix> columns_;
   std::vector<size_t> hit_count_tloc_;
   bool isDense_;

src/data/gradient_index_format.cc

@@ -1,5 +1,5 @@
 /*!
- * Copyright 2021 XGBoost contributors
+ * Copyright 2021-2022 XGBoost contributors
  */
 #include "sparse_page_writer.h"
 #include "gradient_index.h"
@@ -7,7 +7,6 @@
 
 namespace xgboost {
 namespace data {
-
 class GHistIndexRawFormat : public SparsePageFormat<GHistIndexMatrix> {
  public:
   bool Read(GHistIndexMatrix* page, dmlc::SeekStream* fi) override {
@@ -50,6 +49,8 @@ class GHistIndexRawFormat : public SparsePageFormat<GHistIndexMatrix> {
     if (is_dense) {
       page->index.SetBinOffset(page->cut.Ptrs());
     }
+
+    page->ReadColumnPage(fi);
     return true;
   }
 
@@ -81,6 +82,8 @@ class GHistIndexRawFormat : public SparsePageFormat<GHistIndexMatrix> {
     bytes += sizeof(page.base_rowid);
     fo->Write(page.IsDense());
     bytes += sizeof(page.IsDense());
+
+    bytes += page.WriteColumnPage(fo);
     return bytes;
   }
 };

src/data/gradient_index_page_source.cc

@@ -7,11 +7,18 @@ namespace xgboost {
 namespace data {
 void GradientIndexPageSource::Fetch() {
   if (!this->ReadCache()) {
+    if (count_ != 0 && !sync_) {
+      // source is initialized to be the 0th page during construction, so when count_ is 0
+      // there's no need to increment the source.
+      ++(*source_);
+    }
+    // This is not read from cache so we still need it to be synced with sparse page source.
+    CHECK_EQ(count_, source_->Iter());
     auto const& csr = source_->Page();
     this->page_.reset(new GHistIndexMatrix());
     CHECK_NE(cuts_.Values().size(), 0);
-    this->page_->Init(*csr, feature_types_, cuts_, max_bin_per_feat_, is_dense_,
+    this->page_->Init(*csr, feature_types_, cuts_, max_bin_per_feat_, is_dense_, sparse_thresh_,
-                      sparse_thresh_, nthreads_);
+                      nthreads_);
     this->WriteCache();
   }
 }

src/data/gradient_index_page_source.h

@@ -22,13 +22,14 @@ class GradientIndexPageSource : public PageSourceIncMixIn<GHistIndexMatrix> {
  public:
   GradientIndexPageSource(float missing, int nthreads, bst_feature_t n_features, size_t n_batches,
                           std::shared_ptr<Cache> cache, BatchParam param,
-                          common::HistogramCuts cuts, bool is_dense, int32_t max_bin_per_feat,
+                          common::HistogramCuts cuts, bool is_dense,
                           common::Span<FeatureType const> feature_types,
                           std::shared_ptr<SparsePageSource> source)
-      : PageSourceIncMixIn(missing, nthreads, n_features, n_batches, cache),
+      : PageSourceIncMixIn(missing, nthreads, n_features, n_batches, cache,
+                           std::isnan(param.sparse_thresh)),
         cuts_{std::move(cuts)},
         is_dense_{is_dense},
-        max_bin_per_feat_{max_bin_per_feat},
+        max_bin_per_feat_{param.max_bin},
         feature_types_{feature_types},
         sparse_thresh_{param.sparse_thresh} {
     this->source_ = source;

src/data/sparse_page_dmatrix.cc

@@ -159,21 +159,6 @@ BatchSet<SortedCSCPage> SparsePageDMatrix::GetSortedColumnBatches() {
 
 BatchSet<GHistIndexMatrix> SparsePageDMatrix::GetGradientIndex(const BatchParam &param) {
   CHECK_GE(param.max_bin, 2);
-  if (param.hess.empty() && !param.regen) {
-    // hist method doesn't support full external memory implementation, so we concatenate
-    // all index here.
-    if (!ghist_index_page_ || (param != batch_param_ && param != BatchParam{})) {
-      this->InitializeSparsePage();
-      ghist_index_page_.reset(new GHistIndexMatrix{this, param.max_bin, param.sparse_thresh,
-                                                   param.regen, ctx_.Threads()});
-      this->InitializeSparsePage();
-      batch_param_ = param;
-    }
-    auto begin_iter = BatchIterator<GHistIndexMatrix>(
-        new SimpleBatchIteratorImpl<GHistIndexMatrix>(ghist_index_page_));
-    return BatchSet<GHistIndexMatrix>(begin_iter);
-  }
-
   auto id = MakeCache(this, ".gradient_index.page", cache_prefix_, &cache_info_);
   this->InitializeSparsePage();
   if (!cache_info_.at(id)->written || RegenGHist(batch_param_, param)) {
@@ -190,10 +175,9 @@ BatchSet<GHistIndexMatrix> SparsePageDMatrix::GetGradientIndex(const BatchParam
     ghist_index_source_.reset();
     CHECK_NE(cuts.Values().size(), 0);
     auto ft = this->info_.feature_types.ConstHostSpan();
-    ghist_index_source_.reset(
+    ghist_index_source_.reset(new GradientIndexPageSource(
-        new GradientIndexPageSource(this->missing_, this->ctx_.Threads(), this->Info().num_col_,
+        this->missing_, this->ctx_.Threads(), this->Info().num_col_, this->n_batches_,
-                                    this->n_batches_, cache_info_.at(id), param, std::move(cuts),
+        cache_info_.at(id), param, std::move(cuts), this->IsDense(), ft, sparse_page_source_));
-                                    this->IsDense(), param.max_bin, ft, sparse_page_source_));
   } else {
     CHECK(ghist_index_source_);
     ghist_index_source_->Reset();

src/data/sparse_page_dmatrix.cu

@@ -11,6 +11,9 @@ namespace data {
 BatchSet<EllpackPage> SparsePageDMatrix::GetEllpackBatches(const BatchParam& param) {
   CHECK_GE(param.gpu_id, 0);
   CHECK_GE(param.max_bin, 2);
+  if (!(batch_param_ != BatchParam{})) {
+    CHECK(param != BatchParam{}) << "Batch parameter is not initialized.";
+  }
   auto id = MakeCache(this, ".ellpack.page", cache_prefix_, &cache_info_);
   size_t row_stride = 0;
   this->InitializeSparsePage();

src/data/sparse_page_source.h

@@ -23,6 +23,7 @@
 #include "proxy_dmatrix.h"
 
 #include "../common/common.h"
+#include "../common/timer.h"
 
 namespace xgboost {
 namespace data {
@@ -118,26 +119,30 @@ class SparsePageSourceImpl : public BatchIteratorImpl<S> {
     size_t n_prefetch_batches = std::min(kPreFetch, n_batches_);
     CHECK_GT(n_prefetch_batches, 0) << "total batches:" << n_batches_;
     size_t fetch_it = count_;
+
     for (size_t i = 0; i < n_prefetch_batches; ++i, ++fetch_it) {
       fetch_it %= n_batches_;  // ring
-      if (ring_->at(fetch_it).valid()) { continue; }
+      if (ring_->at(fetch_it).valid()) {
+        continue;
+      }
       auto const *self = this;  // make sure it's const
       CHECK_LT(fetch_it, cache_info_->offset.size());
       ring_->at(fetch_it) = std::async(std::launch::async, [fetch_it, self]() {
+        common::Timer timer;
+        timer.Start();
         std::unique_ptr<SparsePageFormat<S>> fmt{CreatePageFormat<S>("raw")};
         auto n = self->cache_info_->ShardName();
         size_t offset = self->cache_info_->offset.at(fetch_it);
-        std::unique_ptr<dmlc::SeekStream> fi{
+        std::unique_ptr<dmlc::SeekStream> fi{dmlc::SeekStream::CreateForRead(n.c_str())};
-            dmlc::SeekStream::CreateForRead(n.c_str())};
         fi->Seek(offset);
         CHECK_EQ(fi->Tell(), offset);
         auto page = std::make_shared<S>();
         CHECK(fmt->Read(page.get(), fi.get()));
+        LOG(INFO) << "Read a page in " << timer.ElapsedSeconds() << " seconds.";
         return page;
       });
     }
-    CHECK_EQ(std::count_if(ring_->cbegin(), ring_->cend(),
+    CHECK_EQ(std::count_if(ring_->cbegin(), ring_->cend(), [](auto const& f) { return f.valid(); }),
-                           [](auto const &f) { return f.valid(); }),
              n_prefetch_batches)
         << "Sparse DMatrix assumes forward iteration.";
     page_ = (*ring_)[count_].get();
@@ -146,12 +151,18 @@ class SparsePageSourceImpl : public BatchIteratorImpl<S> {
 
   void WriteCache() {
     CHECK(!cache_info_->written);
+    common::Timer timer;
+    timer.Start();
     std::unique_ptr<SparsePageFormat<S>> fmt{CreatePageFormat<S>("raw")};
     if (!fo_) {
       auto n = cache_info_->ShardName();
       fo_.reset(dmlc::Stream::Create(n.c_str(), "w"));
     }
     auto bytes = fmt->Write(*page_, fo_.get());
+    timer.Stop();
+
+    LOG(INFO) << static_cast<double>(bytes) / 1024.0 / 1024.0 << " MB written in "
+              << timer.ElapsedSeconds() << " seconds.";
     cache_info_->offset.push_back(bytes);
   }
 
@@ -280,15 +291,24 @@ template <typename S>
 class PageSourceIncMixIn : public SparsePageSourceImpl<S> {
  protected:
   std::shared_ptr<SparsePageSource> source_;
+  using Super = SparsePageSourceImpl<S>;
+  // synchronize the row page, `hist` and `gpu_hist` don't need the original sparse page
+  // so we avoid fetching it.
+  bool sync_{true};
 
  public:
-  using SparsePageSourceImpl<S>::SparsePageSourceImpl;
+  PageSourceIncMixIn(float missing, int nthreads, bst_feature_t n_features, uint32_t n_batches,
+                     std::shared_ptr<Cache> cache, bool sync)
+      : Super::SparsePageSourceImpl{missing, nthreads, n_features, n_batches, cache}, sync_{sync} {}
+
   PageSourceIncMixIn& operator++() final {
     TryLockGuard guard{this->single_threaded_};
+    if (sync_) {
       ++(*source_);
+    }
 
     ++this->count_;
-    this->at_end_ = source_->AtEnd();
+    this->at_end_ = this->count_ == this->n_batches_;
 
     if (this->at_end_) {
       this->cache_info_->Commit();
@@ -299,7 +319,10 @@ class PageSourceIncMixIn : public SparsePageSourceImpl<S> {
     } else {
       this->Fetch();
     }
+
+    if (sync_) {
       CHECK_EQ(source_->Iter(), this->count_);
+    }
     return *this;
   }
 };
@@ -318,12 +341,9 @@ class CSCPageSource : public PageSourceIncMixIn<CSCPage> {
   }
 
  public:
-  CSCPageSource(
+  CSCPageSource(float missing, int nthreads, bst_feature_t n_features, uint32_t n_batches,
-      float missing, int nthreads, bst_feature_t n_features, uint32_t n_batches,
+                std::shared_ptr<Cache> cache, std::shared_ptr<SparsePageSource> source)
-      std::shared_ptr<Cache> cache,
+      : PageSourceIncMixIn(missing, nthreads, n_features, n_batches, cache, true) {
-      std::shared_ptr<SparsePageSource> source)
-      : PageSourceIncMixIn(missing, nthreads, n_features,
-                                     n_batches, cache) {
     this->source_ = source;
     this->Fetch();
   }
@@ -349,7 +369,7 @@ class SortedCSCPageSource : public PageSourceIncMixIn<SortedCSCPage> {
   SortedCSCPageSource(float missing, int nthreads, bst_feature_t n_features,
                       uint32_t n_batches, std::shared_ptr<Cache> cache,
                       std::shared_ptr<SparsePageSource> source)
-      : PageSourceIncMixIn(missing, nthreads, n_features, n_batches, cache) {
+      : PageSourceIncMixIn(missing, nthreads, n_features, n_batches, cache, true) {
     this->source_ = source;
     this->Fetch();
   }

src/tree/hist/histogram.h

@@ -1,5 +1,5 @@
 /*!
- * Copyright 2021 by XGBoost Contributors
+ * Copyright 2021-2022 by XGBoost Contributors
  */
 #ifndef XGBOOST_TREE_HIST_HISTOGRAM_H_
 #define XGBOOST_TREE_HIST_HISTOGRAM_H_
@@ -8,10 +8,11 @@
 #include <limits>
 #include <vector>
 
-#include "rabit/rabit.h"
-#include "xgboost/tree_model.h"
 #include "../../common/hist_util.h"
 #include "../../data/gradient_index.h"
+#include "expand_entry.h"
+#include "rabit/rabit.h"
+#include "xgboost/tree_model.h"
 
 namespace xgboost {
 namespace tree {
@@ -323,6 +324,25 @@ template <typename GradientSumT, typename ExpandEntry> class HistogramBuilder {
     (*sync_count) = std::max(1, n_left);
   }
 };
+
+// Construct a work space for building histogram.  Eventually we should move this
+// function into histogram builder once hist tree method supports external memory.
+template <typename Partitioner>
+common::BlockedSpace2d ConstructHistSpace(Partitioner const &partitioners,
+                                          std::vector<CPUExpandEntry> const &nodes_to_build) {
+  std::vector<size_t> partition_size(nodes_to_build.size(), 0);
+  for (auto const &partition : partitioners) {
+    size_t k = 0;
+    for (auto node : nodes_to_build) {
+      auto n_rows_in_node = partition.Partitions()[node.nid].Size();
+      partition_size[k] = std::max(partition_size[k], n_rows_in_node);
+      k++;
+    }
+  }
+  common::BlockedSpace2d space{
+      nodes_to_build.size(), [&](size_t nidx_in_set) { return partition_size[nidx_in_set]; }, 256};
+  return space;
+}
 }      // namespace tree
 }      // namespace xgboost
 #endif  // XGBOOST_TREE_HIST_HISTOGRAM_H_

src/tree/hist/param.cc

@@ -0,0 +1,10 @@
+/*!
+ * Copyright 2022 XGBoost contributors
+ */
+#include "param.h"
+
+namespace xgboost {
+namespace tree {
+DMLC_REGISTER_PARAMETER(CPUHistMakerTrainParam);
+}  // namespace tree
+}  // namespace xgboost

src/tree/updater_approx.cc

@@ -94,7 +94,7 @@ class GloablApproxBuilder {
     rabit::Allreduce<rabit::op::Sum, double>(reinterpret_cast<double *>(&root_sum), 2);
     std::vector<CPUExpandEntry> nodes{best};
     size_t i = 0;
-    auto space = this->ConstructHistSpace(nodes);
+    auto space = ConstructHistSpace(partitioner_, nodes);
     for (auto const &page : p_fmat->GetBatches<GHistIndexMatrix>(BatchSpec(param_, hess))) {
       histogram_builder_.BuildHist(i, space, page, p_tree, partitioner_.at(i).Partitions(), nodes,
                                    {}, gpair);
@@ -123,25 +123,6 @@ class GloablApproxBuilder {
     monitor_->Stop(__func__);
   }
 
-  // Construct a work space for building histogram.  Eventually we should move this
-  // function into histogram builder once hist tree method supports external memory.
-  common::BlockedSpace2d ConstructHistSpace(
-      std::vector<CPUExpandEntry> const &nodes_to_build) const {
-    std::vector<size_t> partition_size(nodes_to_build.size(), 0);
-    for (auto const &partition : partitioner_) {
-      size_t k = 0;
-      for (auto node : nodes_to_build) {
-        auto n_rows_in_node = partition.Partitions()[node.nid].Size();
-        partition_size[k] = std::max(partition_size[k], n_rows_in_node);
-        k++;
-      }
-    }
-    common::BlockedSpace2d space{nodes_to_build.size(),
-                                 [&](size_t nidx_in_set) { return partition_size[nidx_in_set]; },
-                                 256};
-    return space;
-  }
-
   void BuildHistogram(DMatrix *p_fmat, RegTree *p_tree,
                       std::vector<CPUExpandEntry> const &valid_candidates,
                       std::vector<GradientPair> const &gpair, common::Span<float> hess) {
@@ -164,7 +145,7 @@ class GloablApproxBuilder {
     }
 
     size_t i = 0;
-    auto space = this->ConstructHistSpace(nodes_to_build);
+    auto space = ConstructHistSpace(partitioner_, nodes_to_build);
     for (auto const &page : p_fmat->GetBatches<GHistIndexMatrix>(BatchSpec(param_, hess))) {
       histogram_builder_.BuildHist(i, space, page, p_tree, partitioner_.at(i).Partitions(),
                                    nodes_to_build, nodes_to_sub, gpair);
@@ -191,7 +172,7 @@ class GloablApproxBuilder {
     Driver<CPUExpandEntry> driver(static_cast<TrainParam::TreeGrowPolicy>(param_.grow_policy));
     auto &tree = *p_tree;
     driver.Push({this->InitRoot(p_fmat, gpair, hess, p_tree)});
-    bst_node_t num_leaves = 1;
+    bst_node_t num_leaves{1};
     auto expand_set = driver.Pop();
 
     /**
@@ -223,10 +204,10 @@ class GloablApproxBuilder {
       }
 
       monitor_->Start("UpdatePosition");
-      size_t i = 0;
+      size_t page_id = 0;
       for (auto const &page : p_fmat->GetBatches<GHistIndexMatrix>(BatchSpec(param_, hess))) {
-        partitioner_.at(i).UpdatePosition(ctx_, page, applied, p_tree);
+        partitioner_.at(page_id).UpdatePosition(ctx_, page, applied, p_tree);
-        i++;
+        page_id++;
       }
       monitor_->Stop("UpdatePosition");
 
@@ -288,9 +269,9 @@ class GlobalApproxUpdater : public TreeUpdater {
     out["hist_param"] = ToJson(hist_param_);
   }
 
-  void InitData(TrainParam const &param, HostDeviceVector<GradientPair> *gpair,
+  void InitData(TrainParam const &param, HostDeviceVector<GradientPair> const *gpair,
                 std::vector<GradientPair> *sampled) {
-    auto const &h_gpair = gpair->HostVector();
+    auto const &h_gpair = gpair->ConstHostVector();
     sampled->resize(h_gpair.size());
     std::copy(h_gpair.cbegin(), h_gpair.cend(), sampled->begin());
     auto &rnd = common::GlobalRandom();

src/tree/updater_quantile_hist.cc

@@ -4,80 +4,39 @@
  * \brief use quantized feature values to construct a tree
  * \author Philip Cho, Tianqi Checn, Egor Smirnov
  */
-#include <dmlc/timer.h>
+#include "./updater_quantile_hist.h"
+
 #include <rabit/rabit.h>
 
 #include <algorithm>
-#include <cmath>
-#include <iomanip>
 #include <memory>
 #include <numeric>
-#include <queue>
 #include <string>
 #include <utility>
 #include <vector>
 
+#include "../common/column_matrix.h"
+#include "../common/hist_util.h"
+#include "../common/random.h"
+#include "../common/threading_utils.h"
+#include "constraints.h"
+#include "hist/evaluate_splits.h"
+#include "param.h"
 #include "xgboost/logging.h"
 #include "xgboost/tree_updater.h"
 
-#include "constraints.h"
-#include "param.h"
-#include "./updater_quantile_hist.h"
-#include "./split_evaluator.h"
-#include "../common/random.h"
-#include "../common/hist_util.h"
-#include "../common/row_set.h"
-#include "../common/column_matrix.h"
-#include "../common/threading_utils.h"
-
 namespace xgboost {
 namespace tree {
 
 DMLC_REGISTRY_FILE_TAG(updater_quantile_hist);
 
-DMLC_REGISTER_PARAMETER(CPUHistMakerTrainParam);
-
 void QuantileHistMaker::Configure(const Args &args) {
-  // initialize pruner
-  if (!pruner_) {
-    pruner_.reset(TreeUpdater::Create("prune", ctx_, task_));
-  }
-  pruner_->Configure(args);
   param_.UpdateAllowUnknown(args);
   hist_maker_param_.UpdateAllowUnknown(args);
 }
 
-template <typename GradientSumT>
+void QuantileHistMaker::Update(HostDeviceVector<GradientPair> *gpair, DMatrix *dmat,
-void QuantileHistMaker::SetBuilder(const size_t n_trees,
-                                   std::unique_ptr<Builder<GradientSumT>>* builder, DMatrix* dmat) {
-  builder->reset(
-      new Builder<GradientSumT>(n_trees, param_, std::move(pruner_), dmat, task_, ctx_));
-}
-
-template<typename GradientSumT>
-void QuantileHistMaker::CallBuilderUpdate(const std::unique_ptr<Builder<GradientSumT>>& builder,
-                                          HostDeviceVector<GradientPair> *gpair,
-                                          DMatrix *dmat,
-                                          GHistIndexMatrix const& gmat,
                                const std::vector<RegTree *> &trees) {
-  for (auto tree : trees) {
-    builder->Update(gmat, column_matrix_, gpair, dmat, tree);
-  }
-}
-
-void QuantileHistMaker::Update(HostDeviceVector<GradientPair> *gpair,
-                               DMatrix *dmat,
-                               const std::vector<RegTree *> &trees) {
-  auto it = dmat->GetBatches<GHistIndexMatrix>(HistBatch(param_)).begin();
-  auto p_gmat = it.Page();
-  if (dmat != p_last_dmat_ || is_gmat_initialized_ == false) {
-    updater_monitor_.Start("GmatInitialization");
-    column_matrix_.Init(*p_gmat, param_.sparse_threshold, ctx_->Threads());
-    updater_monitor_.Stop("GmatInitialization");
-    // A proper solution is puting cut matrix in DMatrix, see:
-    // https://github.com/dmlc/xgboost/issues/5143
-    is_gmat_initialized_ = true;
-  }
   // rescale learning rate according to size of trees
   float lr = param_.learning_rate;
   param_.learning_rate = lr / trees.size();
@@ -86,19 +45,23 @@ void QuantileHistMaker::Update(HostDeviceVector<GradientPair> *gpair,
   const size_t n_trees = trees.size();
   if (hist_maker_param_.single_precision_histogram) {
     if (!float_builder_) {
-      this->SetBuilder(n_trees, &float_builder_, dmat);
+      float_builder_.reset(new Builder<float>(n_trees, param_, dmat, task_, ctx_));
     }
-    CallBuilderUpdate(float_builder_, gpair, dmat, *p_gmat, trees);
   } else {
     if (!double_builder_) {
-      SetBuilder(n_trees, &double_builder_, dmat);
+      double_builder_.reset(new Builder<double>(n_trees, param_, dmat, task_, ctx_));
+    }
+  }
+
+  for (auto p_tree : trees) {
+    if (hist_maker_param_.single_precision_histogram) {
+      this->float_builder_->UpdateTree(gpair, dmat, p_tree);
+    } else {
+      this->double_builder_->UpdateTree(gpair, dmat, p_tree);
     }
-    CallBuilderUpdate(double_builder_, gpair, dmat, *p_gmat, trees);
   }
 
   param_.learning_rate = lr;
-
-  p_last_dmat_ = dmat;
 }
 
 bool QuantileHistMaker::UpdatePredictionCache(const DMatrix *data,
@@ -113,23 +76,18 @@ bool QuantileHistMaker::UpdatePredictionCache(const DMatrix *data,
 }
 
 template <typename GradientSumT>
-template <bool any_missing>
+CPUExpandEntry QuantileHistMaker::Builder<GradientSumT>::InitRoot(
-void QuantileHistMaker::Builder<GradientSumT>::InitRoot(
+    DMatrix *p_fmat, RegTree *p_tree, const std::vector<GradientPair> &gpair_h) {
-    DMatrix *p_fmat, RegTree *p_tree, const std::vector<GradientPair> &gpair_h,
-    int *num_leaves, std::vector<CPUExpandEntry> *expand) {
   CPUExpandEntry node(RegTree::kRoot, p_tree->GetDepth(0), 0.0f);
 
-  nodes_for_explicit_hist_build_.clear();
-  nodes_for_subtraction_trick_.clear();
-  nodes_for_explicit_hist_build_.push_back(node);
-
-  auto const& row_set_collection = partitioner_.front().Partitions();
   size_t page_id = 0;
-  for (auto const& gidx :
+  auto space = ConstructHistSpace(partitioner_, {node});
-       p_fmat->GetBatches<GHistIndexMatrix>(HistBatch(param_))) {
+  for (auto const &gidx : p_fmat->GetBatches<GHistIndexMatrix>(HistBatch(param_))) {
-    this->histogram_builder_->BuildHist(
+    std::vector<CPUExpandEntry> nodes_to_build{node};
-        page_id, gidx, p_tree, row_set_collection,
+    std::vector<CPUExpandEntry> nodes_to_sub;
-        nodes_for_explicit_hist_build_, nodes_for_subtraction_trick_, gpair_h);
+    this->histogram_builder_->BuildHist(page_id, space, gidx, p_tree,
+                                        partitioner_.at(page_id).Partitions(), nodes_to_build,
+                                        nodes_to_sub, gpair_h);
     ++page_id;
   }
 
@@ -165,148 +123,118 @@ void QuantileHistMaker::Builder<GradientSumT>::InitRoot(
     (*p_tree)[RegTree::kRoot].SetLeaf(param_.learning_rate * weight);
 
     std::vector<CPUExpandEntry> entries{node};
-    builder_monitor_->Start("EvaluateSplits");
+    monitor_->Start("EvaluateSplits");
     auto ft = p_fmat->Info().feature_types.ConstHostSpan();
     for (auto const &gmat : p_fmat->GetBatches<GHistIndexMatrix>(HistBatch(param_))) {
-      evaluator_->EvaluateSplits(histogram_builder_->Histogram(), gmat.cut, ft,
+      evaluator_->EvaluateSplits(histogram_builder_->Histogram(), gmat.cut, ft, *p_tree, &entries);
-                                 *p_tree, &entries);
       break;
     }
-    builder_monitor_->Stop("EvaluateSplits");
+    monitor_->Stop("EvaluateSplits");
     node = entries.front();
   }
 
-  expand->push_back(node);
+  return node;
-  ++(*num_leaves);
 }
 
 template <typename GradientSumT>
-void QuantileHistMaker::Builder<GradientSumT>::AddSplitsToTree(
+void QuantileHistMaker::Builder<GradientSumT>::BuildHistogram(
-          const std::vector<CPUExpandEntry>& expand,
+    DMatrix *p_fmat, RegTree *p_tree, std::vector<CPUExpandEntry> const &valid_candidates,
-          RegTree *p_tree,
+    std::vector<GradientPair> const &gpair) {
-          int *num_leaves,
+  std::vector<CPUExpandEntry> nodes_to_build(valid_candidates.size());
-          std::vector<CPUExpandEntry>* nodes_for_apply_split) {
+  std::vector<CPUExpandEntry> nodes_to_sub(valid_candidates.size());
-  for (auto const& entry : expand) {
+
-    if (entry.IsValid(param_, *num_leaves)) {
+  size_t n_idx = 0;
-      nodes_for_apply_split->push_back(entry);
+  for (auto const &c : valid_candidates) {
-      evaluator_->ApplyTreeSplit(entry, p_tree);
+    auto left_nidx = (*p_tree)[c.nid].LeftChild();
-      (*num_leaves)++;
+    auto right_nidx = (*p_tree)[c.nid].RightChild();
-    }
+    auto fewer_right = c.split.right_sum.GetHess() < c.split.left_sum.GetHess();
+
+    auto build_nidx = left_nidx;
+    auto subtract_nidx = right_nidx;
+    if (fewer_right) {
+      std::swap(build_nidx, subtract_nidx);
     }
+    nodes_to_build[n_idx] = CPUExpandEntry{build_nidx, p_tree->GetDepth(build_nidx), {}};
+    nodes_to_sub[n_idx] = CPUExpandEntry{subtract_nidx, p_tree->GetDepth(subtract_nidx), {}};
+    n_idx++;
   }
 
-// Split nodes to 2 sets depending on amount of rows in each node
+  size_t page_id{0};
-// Histograms for small nodes will be built explicitly
+  auto space = ConstructHistSpace(partitioner_, nodes_to_build);
-// Histograms for big nodes will be built by 'Subtraction Trick'
+  for (auto const &gidx : p_fmat->GetBatches<GHistIndexMatrix>(HistBatch(param_))) {
-// Exception: in distributed setting, we always build the histogram for the left child node
+    histogram_builder_->BuildHist(page_id, space, gidx, p_tree,
-//    and use 'Subtraction Trick' to built the histogram for the right child node.
+                                  partitioner_.at(page_id).Partitions(), nodes_to_build,
-//    This ensures that the workers operate on the same set of tree nodes.
+                                  nodes_to_sub, gpair);
-template <typename GradientSumT>
+    ++page_id;
-void QuantileHistMaker::Builder<GradientSumT>::SplitSiblings(
-    const std::vector<CPUExpandEntry> &nodes_for_apply_split,
-    std::vector<CPUExpandEntry> *nodes_to_evaluate, RegTree *p_tree) {
-  builder_monitor_->Start("SplitSiblings");
-  auto const& row_set_collection = this->partitioner_.front().Partitions();
-  for (auto const& entry : nodes_for_apply_split) {
-    int nid = entry.nid;
-
-    const int cleft = (*p_tree)[nid].LeftChild();
-    const int cright = (*p_tree)[nid].RightChild();
-    const CPUExpandEntry left_node = CPUExpandEntry(cleft, p_tree->GetDepth(cleft), 0.0);
-    const CPUExpandEntry right_node = CPUExpandEntry(cright, p_tree->GetDepth(cright), 0.0);
-    nodes_to_evaluate->push_back(left_node);
-    nodes_to_evaluate->push_back(right_node);
-    if (row_set_collection[cleft].Size() < row_set_collection[cright].Size()) {
-      nodes_for_explicit_hist_build_.push_back(left_node);
-      nodes_for_subtraction_trick_.push_back(right_node);
-    } else {
-      nodes_for_explicit_hist_build_.push_back(right_node);
-      nodes_for_subtraction_trick_.push_back(left_node);
   }
 }
-  CHECK_EQ(nodes_for_subtraction_trick_.size(), nodes_for_explicit_hist_build_.size());
-  builder_monitor_->Stop("SplitSiblings");
-}
 
 template <typename GradientSumT>
-template <bool any_missing>
 void QuantileHistMaker::Builder<GradientSumT>::ExpandTree(
-    const GHistIndexMatrix& gmat,
+    DMatrix *p_fmat, RegTree *p_tree, const std::vector<GradientPair> &gpair_h) {
-    const common::ColumnMatrix& column_matrix,
+  monitor_->Start(__func__);
-    DMatrix* p_fmat,
-    RegTree* p_tree,
-    const std::vector<GradientPair>& gpair_h) {
-  builder_monitor_->Start("ExpandTree");
-  int num_leaves = 0;
 
   Driver<CPUExpandEntry> driver(static_cast<TrainParam::TreeGrowPolicy>(param_.grow_policy));
-  std::vector<CPUExpandEntry> expand;
+  driver.Push(this->InitRoot(p_fmat, p_tree, gpair_h));
-  InitRoot<any_missing>(p_fmat, p_tree, gpair_h, &num_leaves, &expand);
+  bst_node_t num_leaves{1};
-  driver.Push(expand[0]);
+  auto expand_set = driver.Pop();
 
-  int32_t depth = 0;
+  while (!expand_set.empty()) {
-  while (!driver.IsEmpty()) {
+    // candidates that can be further splited.
-    expand = driver.Pop();
+    std::vector<CPUExpandEntry> valid_candidates;
-    depth = expand[0].depth + 1;
+    // candidaates that can be applied.
-    std::vector<CPUExpandEntry> nodes_for_apply_split;
+    std::vector<CPUExpandEntry> applied;
-    std::vector<CPUExpandEntry> nodes_to_evaluate;
+    int32_t depth = expand_set.front().depth + 1;
-    nodes_for_explicit_hist_build_.clear();
+    for (auto const& candidate : expand_set) {
-    nodes_for_subtraction_trick_.clear();
+      if (!candidate.IsValid(param_, num_leaves)) {
-
+        continue;
-    AddSplitsToTree(expand, p_tree, &num_leaves, &nodes_for_apply_split);
+      }
-
+      evaluator_->ApplyTreeSplit(candidate, p_tree);
-    if (nodes_for_apply_split.size() != 0) {
+      applied.push_back(candidate);
-      HistRowPartitioner &partitioner = this->partitioner_.front();
+      num_leaves++;
-      if (gmat.cut.HasCategorical()) {
+      if (CPUExpandEntry::ChildIsValid(param_, depth, num_leaves)) {
-        partitioner.UpdatePosition<any_missing, true>(this->ctx_, gmat, column_matrix,
+        valid_candidates.emplace_back(candidate);
-                                                      nodes_for_apply_split, p_tree);
+      }
-      } else {
-        partitioner.UpdatePosition<any_missing, false>(this->ctx_, gmat, column_matrix,
-                                                       nodes_for_apply_split, p_tree);
     }
 
-      SplitSiblings(nodes_for_apply_split, &nodes_to_evaluate, p_tree);
+    monitor_->Start("UpdatePosition");
-
+    size_t page_id{0};
-      if (param_.max_depth == 0 || depth < param_.max_depth) {
+    for (auto const &page : p_fmat->GetBatches<GHistIndexMatrix>(HistBatch(param_))) {
-        size_t i = 0;
+      partitioner_.at(page_id).UpdatePosition(ctx_, page, applied, p_tree);
-        for (auto const &gidx : p_fmat->GetBatches<GHistIndexMatrix>(HistBatch(param_))) {
+      ++page_id;
-          this->histogram_builder_->BuildHist(i, gidx, p_tree, partitioner_.front().Partitions(),
-                                              nodes_for_explicit_hist_build_,
-                                              nodes_for_subtraction_trick_, gpair_h);
-          ++i;
-        }
-      } else {
-        int starting_index = std::numeric_limits<int>::max();
-        int sync_count = 0;
-        this->histogram_builder_->AddHistRows(
-            &starting_index, &sync_count, nodes_for_explicit_hist_build_,
-            nodes_for_subtraction_trick_, p_tree);
     }
+    monitor_->Stop("UpdatePosition");
 
-      builder_monitor_->Start("EvaluateSplits");
+    std::vector<CPUExpandEntry> best_splits;
+    if (!valid_candidates.empty()) {
+      this->BuildHistogram(p_fmat, p_tree, valid_candidates, gpair_h);
+      auto const &tree = *p_tree;
+      for (auto const &candidate : valid_candidates) {
+        int left_child_nidx = tree[candidate.nid].LeftChild();
+        int right_child_nidx = tree[candidate.nid].RightChild();
+        CPUExpandEntry l_best{left_child_nidx, depth, 0.0};
+        CPUExpandEntry r_best{right_child_nidx, depth, 0.0};
+        best_splits.push_back(l_best);
+        best_splits.push_back(r_best);
+      }
+      auto const &histograms = histogram_builder_->Histogram();
       auto ft = p_fmat->Info().feature_types.ConstHostSpan();
-      evaluator_->EvaluateSplits(this->histogram_builder_->Histogram(),
+      for (auto const &gmat : p_fmat->GetBatches<GHistIndexMatrix>(HistBatch(param_))) {
-                                 gmat.cut, ft, *p_tree, &nodes_to_evaluate);
+        evaluator_->EvaluateSplits(histograms, gmat.cut, ft, *p_tree, &best_splits);
-      builder_monitor_->Stop("EvaluateSplits");
+        break;
+      }
+    }
+    driver.Push(best_splits.begin(), best_splits.end());
+    expand_set = driver.Pop();
+  }
 
-      for (size_t i = 0; i < nodes_for_apply_split.size(); ++i) {
+  monitor_->Stop(__func__);
-        CPUExpandEntry left_node = nodes_to_evaluate.at(i * 2 + 0);
-        CPUExpandEntry right_node = nodes_to_evaluate.at(i * 2 + 1);
-        driver.Push(left_node);
-        driver.Push(right_node);
-      }
-    }
-  }
-  builder_monitor_->Stop("ExpandTree");
 }
 
 template <typename GradientSumT>
-void QuantileHistMaker::Builder<GradientSumT>::Update(
+void QuantileHistMaker::Builder<GradientSumT>::UpdateTree(HostDeviceVector<GradientPair> *gpair,
-    const GHistIndexMatrix &gmat,
-    const common::ColumnMatrix &column_matrix,
-    HostDeviceVector<GradientPair> *gpair,
                                                           DMatrix *p_fmat, RegTree *p_tree) {
-  builder_monitor_->Start("Update");
+  monitor_->Start(__func__);
 
   std::vector<GradientPair> *gpair_ptr = &(gpair->HostVector());
   // in case 'num_parallel_trees != 1' no posibility to change initial gpair
@@ -315,18 +243,12 @@ void QuantileHistMaker::Builder<GradientSumT>::Update(
     gpair_local_ = *gpair_ptr;
     gpair_ptr = &gpair_local_;
   }
-  p_last_fmat_mutable_ = p_fmat;
 
-  this->InitData(gmat, p_fmat, *p_tree, gpair_ptr);
+  this->InitData(p_fmat, *p_tree, gpair_ptr);
 
-  if (column_matrix.AnyMissing()) {
+  ExpandTree(p_fmat, p_tree, *gpair_ptr);
-    ExpandTree<true>(gmat, column_matrix, p_fmat, p_tree, *gpair_ptr);
-  } else {
-    ExpandTree<false>(gmat, column_matrix, p_fmat, p_tree, *gpair_ptr);
-  }
-  pruner_->Update(gpair, p_fmat, std::vector<RegTree*>{p_tree});
 
-  builder_monitor_->Stop("Update");
+  monitor_->Stop(__func__);
 }
 
 template <typename GradientSumT>
@@ -334,20 +256,20 @@ bool QuantileHistMaker::Builder<GradientSumT>::UpdatePredictionCache(
     DMatrix const *data, linalg::VectorView<float> out_preds) const {
   // p_last_fmat_ is a valid pointer as long as UpdatePredictionCache() is called in
   // conjunction with Update().
-  if (!p_last_fmat_ || !p_last_tree_ || data != p_last_fmat_ ||
+  if (!p_last_fmat_ || !p_last_tree_ || data != p_last_fmat_) {
-      p_last_fmat_ != p_last_fmat_mutable_) {
     return false;
   }
-  builder_monitor_->Start(__func__);
+  monitor_->Start(__func__);
   CHECK_EQ(out_preds.Size(), data->Info().num_row_);
   UpdatePredictionCacheImpl(ctx_, p_last_tree_, partitioner_, *evaluator_, param_, out_preds);
-  builder_monitor_->Stop(__func__);
+  monitor_->Stop(__func__);
   return true;
 }
 
 template <typename GradientSumT>
 void QuantileHistMaker::Builder<GradientSumT>::InitSampling(const DMatrix &fmat,
                                                             std::vector<GradientPair> *gpair) {
+  monitor_->Start(__func__);
   const auto &info = fmat.Info();
   auto& rnd = common::GlobalRandom();
   std::vector<GradientPair>& gpair_ref = *gpair;
@@ -380,6 +302,7 @@ void QuantileHistMaker::Builder<GradientSumT>::InitSampling(const DMatrix &fmat,
   }
   exc.Rethrow();
 #endif  // XGBOOST_CUSTOMIZE_GLOBAL_PRNG
+  monitor_->Stop(__func__);
 }
 template<typename GradientSumT>
 size_t QuantileHistMaker::Builder<GradientSumT>::GetNumberOfTrees() {
@@ -387,10 +310,9 @@ size_t QuantileHistMaker::Builder<GradientSumT>::GetNumberOfTrees() {
 }
 
 template <typename GradientSumT>
-void QuantileHistMaker::Builder<GradientSumT>::InitData(const GHistIndexMatrix &gmat, DMatrix *fmat,
+void QuantileHistMaker::Builder<GradientSumT>::InitData(DMatrix *fmat, const RegTree &tree,
-                                                        const RegTree &tree,
                                                         std::vector<GradientPair> *gpair) {
-  builder_monitor_->Start("InitData");
+  monitor_->Start(__func__);
   const auto& info = fmat->Info();
 
   {
@@ -406,18 +328,14 @@ void QuantileHistMaker::Builder<GradientSumT>::InitData(const GHistIndexMatrix &
       partitioner_.emplace_back(page.Size(), page.base_rowid, this->ctx_->Threads());
       ++page_id;
     }
-    histogram_builder_->Reset(n_total_bins, BatchParam{param_.max_bin, param_.sparse_threshold},
+    histogram_builder_->Reset(n_total_bins, HistBatch(param_), ctx_->Threads(), page_id,
-                              ctx_->Threads(), page_id, rabit::IsDistributed());
+                              rabit::IsDistributed());
 
     if (param_.subsample < 1.0f) {
       CHECK_EQ(param_.sampling_method, TrainParam::kUniform)
           << "Only uniform sampling is supported, "
           << "gradient-based sampling is only support by GPU Hist.";
-      builder_monitor_->Start("InitSampling");
       InitSampling(*fmat, gpair);
-      builder_monitor_->Stop("InitSampling");
-      // We should check that the partitioning was done correctly
-      // and each row of the dataset fell into exactly one of the categories
     }
   }
 
@@ -426,7 +344,7 @@ void QuantileHistMaker::Builder<GradientSumT>::InitData(const GHistIndexMatrix &
   evaluator_.reset(new HistEvaluator<GradientSumT, CPUExpandEntry>{
       param_, info, this->ctx_->Threads(), column_sampler_, task_});
 
-  builder_monitor_->Stop("InitData");
+  monitor_->Stop(__func__);
 }
 
 void HistRowPartitioner::FindSplitConditions(const std::vector<CPUExpandEntry> &nodes,
@@ -470,21 +388,8 @@ void HistRowPartitioner::AddSplitsToRowSet(const std::vector<CPUExpandEntry> &no
 template struct QuantileHistMaker::Builder<float>;
 template struct QuantileHistMaker::Builder<double>;
 
-XGBOOST_REGISTER_TREE_UPDATER(FastHistMaker, "grow_fast_histmaker")
-.describe("(Deprecated, use grow_quantile_histmaker instead.)"
-          " Grow tree using quantized histogram.")
-.set_body(
-    [](ObjInfo task) {
-      LOG(WARNING) << "grow_fast_histmaker is deprecated, "
-                   << "use grow_quantile_histmaker instead.";
-      return new QuantileHistMaker(task);
-    });
-
 XGBOOST_REGISTER_TREE_UPDATER(QuantileHistMaker, "grow_quantile_histmaker")
     .describe("Grow tree using quantized histogram.")
-.set_body(
+    .set_body([](ObjInfo task) { return new QuantileHistMaker(task); });
-    [](ObjInfo task) {
-      return new QuantileHistMaker(task);
-    });
 }  // namespace tree
 }  // namespace xgboost

src/tree/updater_quantile_hist.h

@@ -7,7 +7,6 @@
 #ifndef XGBOOST_TREE_UPDATER_QUANTILE_HIST_H_
 #define XGBOOST_TREE_UPDATER_QUANTILE_HIST_H_
 
-#include <dmlc/timer.h>
 #include <rabit/rabit.h>
 #include <xgboost/tree_updater.h>
 
@@ -29,7 +28,6 @@
 #include "constraints.h"
 #include "./param.h"
 #include "./driver.h"
-#include "./split_evaluator.h"
 #include "../common/random.h"
 #include "../common/timer.h"
 #include "../common/hist_util.h"
@@ -194,6 +192,24 @@ class HistRowPartitioner {
     AddSplitsToRowSet(nodes, p_tree);
   }
 
+  void UpdatePosition(GenericParameter const* ctx, GHistIndexMatrix const& page,
+                      std::vector<CPUExpandEntry> const& applied, RegTree const* p_tree) {
+    auto const& column_matrix = page.Transpose();
+    if (page.cut.HasCategorical()) {
+      if (column_matrix.AnyMissing()) {
+        this->template UpdatePosition<true, true>(ctx, page, column_matrix, applied, p_tree);
+      } else {
+        this->template UpdatePosition<false, true>(ctx, page, column_matrix, applied, p_tree);
+      }
+    } else {
+      if (column_matrix.AnyMissing()) {
+        this->template UpdatePosition<true, false>(ctx, page, column_matrix, applied, p_tree);
+      } else {
+        this->template UpdatePosition<false, false>(ctx, page, column_matrix, applied, p_tree);
+      }
+    }
+  }
+
   auto const& Partitions() const { return row_set_collection_; }
   size_t Size() const {
     return std::distance(row_set_collection_.begin(), row_set_collection_.end());
@@ -209,9 +225,7 @@ inline BatchParam HistBatch(TrainParam const& param) {
 /*! \brief construct a tree using quantized feature values */
 class QuantileHistMaker: public TreeUpdater {
  public:
-  explicit QuantileHistMaker(ObjInfo task) : task_{task} {
+  explicit QuantileHistMaker(ObjInfo task) : task_{task} {}
-    updater_monitor_.Init("QuantileHistMaker");
-  }
   void Configure(const Args& args) override;
 
   void Update(HostDeviceVector<GradientPair>* gpair,
@@ -256,10 +270,6 @@ class QuantileHistMaker: public TreeUpdater {
   CPUHistMakerTrainParam hist_maker_param_;
   // training parameter
   TrainParam param_;
-  // column accessor
-  common::ColumnMatrix column_matrix_;
-  DMatrix const* p_last_dmat_ {nullptr};
-  bool is_gmat_initialized_ {false};
 
   // actual builder that runs the algorithm
   template<typename GradientSumT>
@@ -267,60 +277,40 @@ class QuantileHistMaker: public TreeUpdater {
    public:
     using GradientPairT = xgboost::detail::GradientPairInternal<GradientSumT>;
     // constructor
-    explicit Builder(const size_t n_trees, const TrainParam& param,
+    explicit Builder(const size_t n_trees, const TrainParam& param, DMatrix const* fmat,
-                     std::unique_ptr<TreeUpdater> pruner, DMatrix const* fmat, ObjInfo task,
+                     ObjInfo task, GenericParameter const* ctx)
-                     GenericParameter const* ctx)
         : n_trees_(n_trees),
           param_(param),
-          pruner_(std::move(pruner)),
           p_last_fmat_(fmat),
           histogram_builder_{new HistogramBuilder<GradientSumT, CPUExpandEntry>},
           task_{task},
           ctx_{ctx},
-          builder_monitor_{std::make_unique<common::Monitor>()} {
+          monitor_{std::make_unique<common::Monitor>()} {
-      builder_monitor_->Init("Quantile::Builder");
+      monitor_->Init("Quantile::Builder");
     }
     // update one tree, growing
-    void Update(const GHistIndexMatrix& gmat, const common::ColumnMatrix& column_matrix,
+    void UpdateTree(HostDeviceVector<GradientPair>* gpair, DMatrix* p_fmat, RegTree* p_tree);
-                HostDeviceVector<GradientPair>* gpair, DMatrix* p_fmat, RegTree* p_tree);
 
     bool UpdatePredictionCache(DMatrix const* data, linalg::VectorView<float> out_preds) const;
 
-   protected:
+   private:
     // initialize temp data structure
-    void InitData(const GHistIndexMatrix& gmat, DMatrix* fmat, const RegTree& tree,
+    void InitData(DMatrix* fmat, const RegTree& tree, std::vector<GradientPair>* gpair);
-                  std::vector<GradientPair>* gpair);
 
     size_t GetNumberOfTrees();
 
     void InitSampling(const DMatrix& fmat, std::vector<GradientPair>* gpair);
 
-    template <bool any_missing>
+    CPUExpandEntry InitRoot(DMatrix* p_fmat, RegTree* p_tree,
-    void InitRoot(DMatrix* p_fmat,
-                  RegTree *p_tree,
-                  const std::vector<GradientPair> &gpair_h,
-                  int *num_leaves, std::vector<CPUExpandEntry> *expand);
-
-    // Split nodes to 2 sets depending on amount of rows in each node
-    // Histograms for small nodes will be built explicitly
-    // Histograms for big nodes will be built by 'Subtraction Trick'
-    void SplitSiblings(const std::vector<CPUExpandEntry>& nodes,
-                       std::vector<CPUExpandEntry>* nodes_to_evaluate,
-                       RegTree *p_tree);
-
-    void AddSplitsToTree(const std::vector<CPUExpandEntry>& expand,
-                         RegTree *p_tree,
-                         int *num_leaves,
-                         std::vector<CPUExpandEntry>* nodes_for_apply_split);
-
-    template <bool any_missing>
-    void ExpandTree(const GHistIndexMatrix& gmat,
-                    const common::ColumnMatrix& column_matrix,
-                    DMatrix* p_fmat,
-                    RegTree* p_tree,
                             const std::vector<GradientPair>& gpair_h);
 
-    //  --data fields--
+    void BuildHistogram(DMatrix* p_fmat, RegTree* p_tree,
+                        std::vector<CPUExpandEntry> const& valid_candidates,
+                        std::vector<GradientPair> const& gpair);
+
+    void ExpandTree(DMatrix* p_fmat, RegTree* p_tree, const std::vector<GradientPair>& gpair_h);
+
+   private:
     const size_t n_trees_;
     const TrainParam& param_;
     std::shared_ptr<common::ColumnSampler> column_sampler_{
@@ -328,48 +318,24 @@ class QuantileHistMaker: public TreeUpdater {
 
     std::vector<GradientPair> gpair_local_;
 
-    std::unique_ptr<TreeUpdater> pruner_;
     std::unique_ptr<HistEvaluator<GradientSumT, CPUExpandEntry>> evaluator_;
-    // Right now there's only 1 partitioner in this vector, when external memory is fully
-    // supported we will have number of partitioners equal to number of pages.
     std::vector<HistRowPartitioner> partitioner_;
 
     // back pointers to tree and data matrix
     const RegTree* p_last_tree_{nullptr};
     DMatrix const* const p_last_fmat_;
-    DMatrix* p_last_fmat_mutable_;
 
-    // key is the node id which should be calculated by Subtraction Trick, value is the node which
-    // provides the evidence for subtraction
-    std::vector<CPUExpandEntry> nodes_for_subtraction_trick_;
-    // list of nodes whose histograms would be built explicitly.
-    std::vector<CPUExpandEntry> nodes_for_explicit_hist_build_;
-
-    enum class DataLayout { kDenseDataZeroBased, kDenseDataOneBased, kSparseData };
     std::unique_ptr<HistogramBuilder<GradientSumT, CPUExpandEntry>> histogram_builder_;
     ObjInfo task_;
     // Context for number of threads
     GenericParameter const* ctx_;
 
-    std::unique_ptr<common::Monitor> builder_monitor_;
+    std::unique_ptr<common::Monitor> monitor_;
   };
-  common::Monitor updater_monitor_;
-
-  template<typename GradientSumT>
-  void SetBuilder(const size_t n_trees, std::unique_ptr<Builder<GradientSumT>>*, DMatrix *dmat);
-
-  template<typename GradientSumT>
-  void CallBuilderUpdate(const std::unique_ptr<Builder<GradientSumT>>& builder,
-                         HostDeviceVector<GradientPair> *gpair,
-                         DMatrix *dmat,
-                         GHistIndexMatrix const& gmat,
-                         const std::vector<RegTree *> &trees);
 
  protected:
   std::unique_ptr<Builder<float>> float_builder_;
   std::unique_ptr<Builder<double>> double_builder_;
-
-  std::unique_ptr<TreeUpdater> pruner_;
   ObjInfo task_;
 };
 }  // namespace tree

tests/cpp/common/test_column_matrix.cc

@@ -21,7 +21,9 @@ TEST(DenseColumn, Test) {
     GHistIndexMatrix gmat{dmat.get(), max_num_bin, sparse_thresh, false,
                           common::OmpGetNumThreads(0)};
     ColumnMatrix column_matrix;
-    column_matrix.Init(gmat, 0.2, common::OmpGetNumThreads(0));
+    for (auto const& page : dmat->GetBatches<SparsePage>()) {
+      column_matrix.Init(page, gmat, sparse_thresh, common::OmpGetNumThreads(0));
+    }
 
     for (auto i = 0ull; i < dmat->Info().num_row_; i++) {
       for (auto j = 0ull; j < dmat->Info().num_col_; j++) {
@@ -68,7 +70,9 @@ TEST(SparseColumn, Test) {
     auto dmat = RandomDataGenerator(100, 1, 0.85).GenerateDMatrix();
     GHistIndexMatrix gmat{dmat.get(), max_num_bin, 0.5f, false, common::OmpGetNumThreads(0)};
     ColumnMatrix column_matrix;
-    column_matrix.Init(gmat, 0.5, common::OmpGetNumThreads(0));
+    for (auto const& page : dmat->GetBatches<SparsePage>()) {
+      column_matrix.Init(page, gmat, 1.0, common::OmpGetNumThreads(0));
+    }
     switch (column_matrix.GetTypeSize()) {
       case kUint8BinsTypeSize: {
           auto col = column_matrix.GetColumn<uint8_t, true>(0);
@@ -106,9 +110,11 @@ TEST(DenseColumnWithMissing, Test) {
                             static_cast<int32_t>(std::numeric_limits<uint16_t>::max()) + 2};
   for (int32_t max_num_bin : max_num_bins) {
     auto dmat = RandomDataGenerator(100, 1, 0.5).GenerateDMatrix();
-    GHistIndexMatrix gmat{dmat.get(), max_num_bin, 0.2, false, common::OmpGetNumThreads(0)};
+    GHistIndexMatrix gmat(dmat.get(), max_num_bin, 0.2, false, common::OmpGetNumThreads(0));
     ColumnMatrix column_matrix;
-    column_matrix.Init(gmat, 0.2, common::OmpGetNumThreads(0));
+    for (auto const& page : dmat->GetBatches<SparsePage>()) {
+      column_matrix.Init(page, gmat, 0.2, common::OmpGetNumThreads(0));
+    }
     switch (column_matrix.GetTypeSize()) {
       case kUint8BinsTypeSize: {
           auto col = column_matrix.GetColumn<uint8_t, true>(0);

tests/cpp/data/test_gradient_index_page_raw_format.cc

@@ -3,6 +3,7 @@
  */
 #include <gtest/gtest.h>
 
+#include "../../../src/common/column_matrix.h"
 #include "../../../src/data/gradient_index.h"
 #include "../../../src/data/sparse_page_source.h"
 #include "../helpers.h"
@@ -15,33 +16,31 @@ TEST(GHistIndexPageRawFormat, IO) {
   auto m = RandomDataGenerator{100, 14, 0.5}.GenerateDMatrix();
   dmlc::TemporaryDirectory tmpdir;
   std::string path = tmpdir.path + "/ghistindex.page";
+  auto batch = BatchParam{256, 0.5};
 
   {
     std::unique_ptr<dmlc::Stream> fo{dmlc::Stream::Create(path.c_str(), "w")};
-    for (auto const &index :
+    for (auto const &index : m->GetBatches<GHistIndexMatrix>(batch)) {
-         m->GetBatches<GHistIndexMatrix>({GenericParameter::kCpuId, 256})) {
       format->Write(index, fo.get());
     }
   }
 
   GHistIndexMatrix page;
-  std::unique_ptr<dmlc::SeekStream> fi{
+  std::unique_ptr<dmlc::SeekStream> fi{dmlc::SeekStream::CreateForRead(path.c_str())};
-      dmlc::SeekStream::CreateForRead(path.c_str())};
   format->Read(&page, fi.get());
 
-  for (auto const &gidx :
+  for (auto const &gidx : m->GetBatches<GHistIndexMatrix>(batch)) {
-       m->GetBatches<GHistIndexMatrix>({GenericParameter::kCpuId, 256})) {
     auto const &loaded = gidx;
     ASSERT_EQ(loaded.cut.Ptrs(), page.cut.Ptrs());
     ASSERT_EQ(loaded.cut.MinValues(), page.cut.MinValues());
     ASSERT_EQ(loaded.cut.Values(), page.cut.Values());
     ASSERT_EQ(loaded.base_rowid, page.base_rowid);
     ASSERT_EQ(loaded.IsDense(), page.IsDense());
-    ASSERT_TRUE(std::equal(loaded.index.begin(), loaded.index.end(),
+    ASSERT_TRUE(std::equal(loaded.index.begin(), loaded.index.end(), page.index.begin()));
-                           page.index.begin()));
+    ASSERT_TRUE(std::equal(loaded.index.Offset(), loaded.index.Offset() + loaded.index.OffsetSize(),
-    ASSERT_TRUE(std::equal(loaded.index.Offset(),
-                           loaded.index.Offset() + loaded.index.OffsetSize(),
                            page.index.Offset()));
+
+    ASSERT_EQ(loaded.Transpose().GetTypeSize(), loaded.Transpose().GetTypeSize());
   }
 }
 } // namespace data

tests/cpp/tree/hist/test_histogram.cc

@@ -446,6 +446,12 @@ void TestHistogramExternalMemory(BatchParam batch_param, bool is_approx) {
 TEST(CPUHistogram, ExternalMemory) {
   int32_t constexpr kBins = 256;
   TestHistogramExternalMemory(BatchParam{kBins, common::Span<float>{}, false}, true);
+
+  float sparse_thresh{0.5};
+  TestHistogramExternalMemory({kBins, sparse_thresh}, false);
+  sparse_thresh = std::numeric_limits<float>::quiet_NaN();
+  TestHistogramExternalMemory({kBins, sparse_thresh}, false);
+
 }
 }  // namespace tree
 }  // namespace xgboost

tests/cpp/tree/test_quantile_hist.cc

@@ -18,138 +18,6 @@
 
 namespace xgboost {
 namespace tree {
-
-class QuantileHistMock : public QuantileHistMaker {
-  static double constexpr kEps = 1e-6;
-
-  template <typename GradientSumT>
-  struct BuilderMock : public QuantileHistMaker::Builder<GradientSumT> {
-    using RealImpl = QuantileHistMaker::Builder<GradientSumT>;
-
-    BuilderMock(const TrainParam &param, std::unique_ptr<TreeUpdater> pruner,
-                DMatrix const *fmat, GenericParameter const* ctx)
-        : RealImpl(1, param, std::move(pruner), fmat, ObjInfo{ObjInfo::kRegression}, ctx) {}
-
-   public:
-    void TestInitData(const GHistIndexMatrix& gmat,
-                      std::vector<GradientPair>* gpair,
-                      DMatrix* p_fmat,
-                      const RegTree& tree) {
-      RealImpl::InitData(gmat, p_fmat, tree, gpair);
-
-      /* The creation of HistCutMatrix and GHistIndexMatrix are not technically
-       * part of QuantileHist updater logic, but we include it here because
-       * QuantileHist updater object currently stores GHistIndexMatrix
-       * internally. According to https://github.com/dmlc/xgboost/pull/3803,
-       * we should eventually move GHistIndexMatrix out of the QuantileHist
-       * updater. */
-
-      const size_t num_row = p_fmat->Info().num_row_;
-      const size_t num_col = p_fmat->Info().num_col_;
-      /* Validate HistCutMatrix */
-      ASSERT_EQ(gmat.cut.Ptrs().size(), num_col + 1);
-      for (size_t fid = 0; fid < num_col; ++fid) {
-        const size_t ibegin = gmat.cut.Ptrs()[fid];
-        const size_t iend = gmat.cut.Ptrs()[fid + 1];
-        // Ordered,  but empty feature is allowed.
-        ASSERT_LE(ibegin, iend);
-        for (size_t i = ibegin; i < iend - 1; ++i) {
-          // Quantile points must be sorted in ascending order
-          // No duplicates allowed
-          ASSERT_LT(gmat.cut.Values()[i], gmat.cut.Values()[i + 1])
-              << "ibegin: " << ibegin << ", "
-              << "iend: " << iend;
-        }
-      }
-
-      /* Validate GHistIndexMatrix */
-      ASSERT_EQ(gmat.row_ptr.size(), num_row + 1);
-      ASSERT_LT(*std::max_element(gmat.index.begin(), gmat.index.end()),
-                gmat.cut.Ptrs().back());
-      for (const auto& batch : p_fmat->GetBatches<xgboost::SparsePage>()) {
-        auto page = batch.GetView();
-        for (size_t i = 0; i < batch.Size(); ++i) {
-          const size_t rid = batch.base_rowid + i;
-          ASSERT_LT(rid, num_row);
-          const size_t gmat_row_offset = gmat.row_ptr[rid];
-          ASSERT_LT(gmat_row_offset, gmat.index.Size());
-          SparsePage::Inst inst = page[i];
-          ASSERT_EQ(gmat.row_ptr[rid] + inst.size(), gmat.row_ptr[rid + 1]);
-          for (size_t j = 0; j < inst.size(); ++j) {
-            // Each entry of GHistIndexMatrix represents a bin ID
-            const size_t bin_id = gmat.index[gmat_row_offset + j];
-            const size_t fid = inst[j].index;
-            // The bin ID must correspond to correct feature
-            ASSERT_GE(bin_id, gmat.cut.Ptrs()[fid]);
-            ASSERT_LT(bin_id, gmat.cut.Ptrs()[fid + 1]);
-            // The bin ID must correspond to a region between two
-            // suitable quantile points
-            ASSERT_LT(inst[j].fvalue, gmat.cut.Values()[bin_id]);
-            if (bin_id > gmat.cut.Ptrs()[fid]) {
-              ASSERT_GE(inst[j].fvalue, gmat.cut.Values()[bin_id - 1]);
-            } else {
-              ASSERT_GE(inst[j].fvalue, gmat.cut.MinValues()[fid]);
-            }
-          }
-        }
-      }
-    }
-  };
-
-  int static constexpr kNRows = 8, kNCols = 16;
-  std::shared_ptr<xgboost::DMatrix> dmat_;
-  GenericParameter ctx_;
-  const std::vector<std::pair<std::string, std::string> > cfg_;
-  std::shared_ptr<BuilderMock<float> > float_builder_;
-  std::shared_ptr<BuilderMock<double> > double_builder_;
-
- public:
-  explicit QuantileHistMock(
-      const std::vector<std::pair<std::string, std::string> >& args,
-      const bool single_precision_histogram = false, bool batch = true) :
-      QuantileHistMaker{ObjInfo{ObjInfo::kRegression}}, cfg_{args} {
-    QuantileHistMaker::Configure(args);
-    dmat_ = RandomDataGenerator(kNRows, kNCols, 0.8).Seed(3).GenerateDMatrix();
-    ctx_.UpdateAllowUnknown(Args{});
-    if (single_precision_histogram) {
-      float_builder_.reset(new BuilderMock<float>(param_, std::move(pruner_), dmat_.get(), &ctx_));
-    } else {
-      double_builder_.reset(
-          new BuilderMock<double>(param_, std::move(pruner_), dmat_.get(), &ctx_));
-    }
-  }
-  ~QuantileHistMock() override = default;
-
-  static size_t GetNumColumns() { return kNCols; }
-
-  void TestInitData() {
-    int32_t constexpr kMaxBins = 4;
-    GHistIndexMatrix gmat{dmat_.get(), kMaxBins, 0.0f, false, common::OmpGetNumThreads(0)};
-
-    RegTree tree = RegTree();
-    tree.param.UpdateAllowUnknown(cfg_);
-
-    std::vector<GradientPair> gpair =
-        { {0.23f, 0.24f}, {0.23f, 0.24f}, {0.23f, 0.24f}, {0.23f, 0.24f},
-          {0.27f, 0.29f}, {0.27f, 0.29f}, {0.27f, 0.29f}, {0.27f, 0.29f} };
-    if (double_builder_) {
-      double_builder_->TestInitData(gmat, &gpair, dmat_.get(), tree);
-    } else {
-      float_builder_->TestInitData(gmat, &gpair, dmat_.get(), tree);
-    }
-  }
-};
-
-TEST(QuantileHist, InitData) {
-  std::vector<std::pair<std::string, std::string>> cfg
-      {{"num_feature", std::to_string(QuantileHistMock::GetNumColumns())}};
-  QuantileHistMock maker(cfg);
-  maker.TestInitData();
-  const bool single_precision_histogram = true;
-  QuantileHistMock maker_float(cfg, single_precision_histogram);
-  maker_float.TestInitData();
-}
-
 TEST(QuantileHist, Partitioner) {
   size_t n_samples = 1024, n_features = 1, base_rowid = 0;
   GenericParameter ctx;
@@ -163,45 +31,44 @@ TEST(QuantileHist, Partitioner) {
   auto Xy = RandomDataGenerator{n_samples, n_features, 0}.GenerateDMatrix(true);
   std::vector<CPUExpandEntry> candidates{{0, 0, 0.4}};
 
-  auto grad = GenerateRandomGradients(n_samples);
+  auto cuts = common::SketchOnDMatrix(Xy.get(), 64, ctx.Threads());
-  std::vector<float> hess(grad.Size());
-  std::transform(grad.HostVector().cbegin(), grad.HostVector().cend(), hess.begin(),
-                 [](auto gpair) { return gpair.GetHess(); });
 
-  for (auto const& page : Xy->GetBatches<GHistIndexMatrix>({64, 0.5})) {
+  for (auto const& page : Xy->GetBatches<SparsePage>()) {
+    GHistIndexMatrix gmat;
+    gmat.Init(page, {}, cuts, 64, false, 0.5, ctx.Threads());
     bst_feature_t const split_ind = 0;
     common::ColumnMatrix column_indices;
-    column_indices.Init(page, 0.5, ctx.Threads());
+    column_indices.Init(page, gmat, 0.5, ctx.Threads());
     {
-      auto min_value = page.cut.MinValues()[split_ind];
+      auto min_value = gmat.cut.MinValues()[split_ind];
       RegTree tree;
       HistRowPartitioner partitioner{n_samples, base_rowid, ctx.Threads()};
       GetSplit(&tree, min_value, &candidates);
-      partitioner.UpdatePosition<false, true>(&ctx, page, column_indices, candidates, &tree);
+      partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
       ASSERT_EQ(partitioner.Size(), 3);
       ASSERT_EQ(partitioner[1].Size(), 0);
       ASSERT_EQ(partitioner[2].Size(), n_samples);
     }
     {
       HistRowPartitioner partitioner{n_samples, base_rowid, ctx.Threads()};
-      auto ptr = page.cut.Ptrs()[split_ind + 1];
+      auto ptr = gmat.cut.Ptrs()[split_ind + 1];
-      float split_value = page.cut.Values().at(ptr / 2);
+      float split_value = gmat.cut.Values().at(ptr / 2);
       RegTree tree;
       GetSplit(&tree, split_value, &candidates);
       auto left_nidx = tree[RegTree::kRoot].LeftChild();
-      partitioner.UpdatePosition<false, true>(&ctx, page, column_indices, candidates, &tree);
+      partitioner.UpdatePosition<false, true>(&ctx, gmat, column_indices, candidates, &tree);
 
       auto elem = partitioner[left_nidx];
       ASSERT_LT(elem.Size(), n_samples);
       ASSERT_GT(elem.Size(), 1);
       for (auto it = elem.begin; it != elem.end; ++it) {
-        auto value = page.cut.Values().at(page.index[*it]);
+        auto value = gmat.cut.Values().at(gmat.index[*it]);
         ASSERT_LE(value, split_value);
       }
       auto right_nidx = tree[RegTree::kRoot].RightChild();
       elem = partitioner[right_nidx];
       for (auto it = elem.begin; it != elem.end; ++it) {
-        auto value = page.cut.Values().at(page.index[*it]);
+        auto value = gmat.cut.Values().at(gmat.index[*it]);
         ASSERT_GT(value, split_value) << *it;
       }
     }

tests/python/test_data_iterator.py

@@ -1,7 +1,7 @@
 import xgboost as xgb
 from xgboost.data import SingleBatchInternalIter as SingleBatch
 import numpy as np
-from testing import IteratorForTest
+from testing import IteratorForTest, non_increasing
 from typing import Tuple, List
 import pytest
 from hypothesis import given, strategies, settings
@@ -108,7 +108,7 @@ def run_data_iterator(
         evals_result=results_from_it,
         verbose_eval=False,
     )
-    it_predt = from_it.predict(Xy)
+    assert non_increasing(results_from_it["Train"]["rmse"])
 
     X, y = it.as_arrays()
     Xy = xgb.DMatrix(X, y)
@@ -125,13 +125,13 @@ def run_data_iterator(
         verbose_eval=False,
     )
     arr_predt = from_arrays.predict(Xy)
+    assert non_increasing(results_from_arrays["Train"]["rmse"])
 
-    if tree_method != "gpu_hist":
+    rtol = 1e-2
-        rtol = 1e-1  # flaky
+    # CPU sketching is more memory efficient but less consistent due to small chunks
-    else:
+    it_predt = from_it.predict(Xy)
-        # Model can be sensitive to quantiles, use 1e-2 to relax the test.
+    arr_predt = from_arrays.predict(Xy)
-        np.testing.assert_allclose(it_predt, arr_predt, rtol=1e-2)
+    np.testing.assert_allclose(it_predt, arr_predt, rtol=rtol)
-        rtol = 1e-6
 
     np.testing.assert_allclose(
         results_from_it["Train"]["rmse"],