Copy data from Ellpack to GHist. (#8215)

src/data/ellpack_page.cu

@@ -547,4 +547,15 @@ EllpackDeviceAccessor EllpackPageImpl::GetDeviceAccessor(
                                                NumSymbols()),
           feature_types};
 }
+EllpackDeviceAccessor EllpackPageImpl::GetHostAccessor(
+    common::Span<FeatureType const> feature_types) const {
+  return {Context::kCpuId,
+          cuts_,
+          is_dense,
+          row_stride,
+          base_rowid,
+          n_rows,
+          common::CompressedIterator<uint32_t>(gidx_buffer.ConstHostPointer(), NumSymbols()),
+          feature_types};
+}
 }  // namespace xgboost

src/data/ellpack_page.cuh

@@ -43,12 +43,18 @@ struct EllpackDeviceAccessor {
         base_rowid(base_rowid),
         n_rows(n_rows) ,gidx_iter(gidx_iter),
         feature_types{feature_types} {
-    cuts.cut_values_.SetDevice(device);
-    cuts.cut_ptrs_.SetDevice(device);
-    cuts.min_vals_.SetDevice(device);
-    gidx_fvalue_map = cuts.cut_values_.ConstDeviceSpan();
-    feature_segments = cuts.cut_ptrs_.ConstDeviceSpan();
-    min_fvalue = cuts.min_vals_.ConstDeviceSpan();
+    if (device == Context::kCpuId) {
+      gidx_fvalue_map = cuts.cut_values_.ConstHostSpan();
+      feature_segments = cuts.cut_ptrs_.ConstHostSpan();
+      min_fvalue = cuts.min_vals_.ConstHostSpan();
+    } else {
+      cuts.cut_values_.SetDevice(device);
+      cuts.cut_ptrs_.SetDevice(device);
+      cuts.min_vals_.SetDevice(device);
+      gidx_fvalue_map = cuts.cut_values_.ConstDeviceSpan();
+      feature_segments = cuts.cut_ptrs_.ConstDeviceSpan();
+      min_fvalue = cuts.min_vals_.ConstDeviceSpan();
+    }
   }
   // Get a matrix element, uses binary search for look up Return NaN if missing
   // Given a row index and a feature index, returns the corresponding cut value
@@ -202,6 +208,7 @@ class EllpackPageImpl {
   EllpackDeviceAccessor
   GetDeviceAccessor(int device,
                     common::Span<FeatureType const> feature_types = {}) const;
+  EllpackDeviceAccessor GetHostAccessor(common::Span<FeatureType const> feature_types = {}) const;
 
  private:
   /*!

src/data/gradient_index.cc

@@ -53,7 +53,7 @@ GHistIndexMatrix::GHistIndexMatrix(DMatrix *p_fmat, bst_bin_t max_bins_per_feat,
     // hist
     CHECK(!sorted_sketch);
     for (auto const &page : p_fmat->GetBatches<SparsePage>()) {
-      this->columns_->Init(page, *this, sparse_thresh, n_threads);
+      this->columns_->InitFromSparse(page, *this, sparse_thresh, n_threads);
     }
   }
 }
@@ -66,6 +66,12 @@ GHistIndexMatrix::GHistIndexMatrix(MetaInfo const &info, common::HistogramCuts &
       max_num_bins(max_bin_per_feat),
       isDense_{info.num_col_ * info.num_row_ == info.num_nonzero_} {}
 
+#if !defined(XGBOOST_USE_CUDA)
+GHistIndexMatrix::GHistIndexMatrix(Context const *, MetaInfo const &, EllpackPage const &,
+                                   BatchParam const &) {
+  common::AssertGPUSupport();
+}
+#endif  // defined(XGBOOST_USE_CUDA)
 
 GHistIndexMatrix::~GHistIndexMatrix() = default;
 
@@ -99,7 +105,7 @@ GHistIndexMatrix::GHistIndexMatrix(SparsePage const &batch, common::Span<Feature
   this->PushBatch(batch, ft, n_threads);
   this->columns_ = std::make_unique<common::ColumnMatrix>();
   if (!std::isnan(sparse_thresh)) {
-    this->columns_->Init(batch, *this, sparse_thresh, n_threads);
+    this->columns_->InitFromSparse(batch, *this, sparse_thresh, n_threads);
   }
 }
 

src/data/gradient_index.cu

@@ -0,0 +1,111 @@
+/*!
+ * Copyright 2022 by XGBoost Contributors
+ */
+#include <memory>  // std::unique_ptr
+
+#include "../common/column_matrix.h"
+#include "../common/hist_util.h"  // Index
+#include "ellpack_page.cuh"
+#include "gradient_index.h"
+#include "xgboost/data.h"
+
+namespace xgboost {
+// Similar to GHistIndexMatrix::SetIndexData, but without the need for adaptor or bin
+// searching. Is there a way to unify the code?
+template <typename BinT, typename CompressOffset>
+void SetIndexData(Context const* ctx, EllpackPageImpl const* page,
+                  std::vector<size_t>* p_hit_count_tloc, CompressOffset&& get_offset,
+                  GHistIndexMatrix* out) {
+  auto accessor = page->GetHostAccessor();
+  auto const kNull = static_cast<bst_bin_t>(accessor.NullValue());
+
+  common::Span<BinT> index_data_span = {out->index.data<BinT>(), out->index.Size()};
+  auto n_bins_total = page->Cuts().TotalBins();
+
+  auto& hit_count_tloc = *p_hit_count_tloc;
+  hit_count_tloc.clear();
+  hit_count_tloc.resize(ctx->Threads() * n_bins_total, 0);
+
+  common::ParallelFor(page->Size(), ctx->Threads(), [&](auto i) {
+    auto tid = omp_get_thread_num();
+    size_t in_rbegin = page->row_stride * i;
+    size_t out_rbegin = out->row_ptr[i];
+    auto r_size = out->row_ptr[i + 1] - out->row_ptr[i];
+    for (size_t j = 0; j < r_size; ++j) {
+      auto bin_idx = accessor.gidx_iter[in_rbegin + j];
+      assert(bin_idx != kNull);
+      index_data_span[out_rbegin + j] = get_offset(bin_idx, j);
+      ++hit_count_tloc[tid * n_bins_total + bin_idx];
+    }
+  });
+}
+
+void GetRowPtrFromEllpack(Context const* ctx, EllpackPageImpl const* page,
+                          std::vector<size_t>* p_out) {
+  auto& row_ptr = *p_out;
+  row_ptr.resize(page->Size() + 1, 0);
+  if (page->is_dense) {
+    std::fill(row_ptr.begin() + 1, row_ptr.end(), page->row_stride);
+  } else {
+    auto accessor = page->GetHostAccessor();
+    auto const kNull = static_cast<bst_bin_t>(accessor.NullValue());
+
+    common::ParallelFor(page->Size(), ctx->Threads(), [&](auto i) {
+      size_t ibegin = page->row_stride * i;
+      for (size_t j = 0; j < page->row_stride; ++j) {
+        bst_bin_t bin_idx = accessor.gidx_iter[ibegin + j];
+        if (bin_idx != kNull) {
+          row_ptr[i + 1]++;
+        }
+      }
+    });
+  }
+  std::partial_sum(row_ptr.begin(), row_ptr.end(), row_ptr.begin());
+}
+
+GHistIndexMatrix::GHistIndexMatrix(Context const* ctx, MetaInfo const& info,
+                                   EllpackPage const& in_page, BatchParam const& p)
+    : max_num_bins{p.max_bin} {
+  auto page = in_page.Impl();
+  isDense_ = page->is_dense;
+
+  CHECK_EQ(info.num_row_, in_page.Size());
+
+  this->cut = page->Cuts();
+  // pull to host early, prevent race condition
+  this->cut.Ptrs();
+  this->cut.Values();
+  this->cut.MinValues();
+
+  this->ResizeIndex(info.num_nonzero_, page->is_dense);
+  if (page->is_dense) {
+    this->index.SetBinOffset(page->Cuts().Ptrs());
+  }
+
+  auto n_bins_total = page->Cuts().TotalBins();
+  GetRowPtrFromEllpack(ctx, page, &this->row_ptr);
+  if (page->is_dense) {
+    common::DispatchBinType(this->index.GetBinTypeSize(), [&](auto dtype) {
+      using T = decltype(dtype);
+      ::xgboost::SetIndexData<T>(ctx, page, &hit_count_tloc_, index.MakeCompressor<T>(), this);
+    });
+  } else {
+    // no compression
+    ::xgboost::SetIndexData<uint32_t>(
+        ctx, page, &hit_count_tloc_, [&](auto bin_idx, auto) { return bin_idx; }, this);
+  }
+
+  this->hit_count.resize(n_bins_total, 0);
+  this->GatherHitCount(ctx->Threads(), n_bins_total);
+
+  // sanity checks
+  CHECK_EQ(this->Features(), info.num_col_);
+  CHECK_EQ(this->Size(), info.num_row_);
+  CHECK(this->cut.cut_ptrs_.HostCanRead());
+  CHECK(this->cut.cut_values_.HostCanRead());
+  CHECK(this->cut.min_vals_.HostCanRead());
+
+  this->columns_ = std::make_unique<common::ColumnMatrix>(*this, p.sparse_thresh);
+  this->columns_->InitFromGHist(ctx, *this);
+}
+}  // namespace xgboost

src/data/gradient_index.h

@@ -69,7 +69,7 @@ class GHistIndexMatrix {
         if (is_valid(elem)) {
           bst_bin_t bin_idx{-1};
           if (common::IsCat(ft, elem.column_idx)) {
-            bin_idx = cut.SearchCatBin(elem.value, elem.column_idx);
+            bin_idx = cut.SearchCatBin(elem.value, elem.column_idx, ptrs, values);
           } else {
             bin_idx = cut.SearchBin(elem.value, elem.column_idx, ptrs, values);
           }
@@ -81,6 +81,17 @@ class GHistIndexMatrix {
     });
   }
 
+  // Gather hit_count from all threads
+  void GatherHitCount(int32_t n_threads, bst_bin_t n_bins_total) {
+    CHECK_EQ(hit_count.size(), n_bins_total);
+    common::ParallelFor(n_bins_total, n_threads, [&](bst_omp_uint idx) {
+      for (int32_t tid = 0; tid < n_threads; ++tid) {
+        hit_count[idx] += hit_count_tloc_[tid * n_bins_total + idx];
+        hit_count_tloc_[tid * n_bins_total + idx] = 0;  // reset for next batch
+      }
+    });
+  }
+
   template <typename Batch, typename IsValid>
   void PushBatchImpl(int32_t n_threads, Batch const& batch, size_t rbegin, IsValid&& is_valid,
                      common::Span<FeatureType const> ft) {
@@ -95,33 +106,20 @@ class GHistIndexMatrix {
     if (isDense_) {
       index.SetBinOffset(cut.Ptrs());
     }
-    uint32_t const* offsets = index.Offset();
     if (isDense_) {
-      // Inside the lambda functions, bin_idx is the index for cut value across all
-      // features. By subtracting it with starting pointer of each feature, we can reduce
-      // it to smaller value and compress it to smaller types.
       common::DispatchBinType(index.GetBinTypeSize(), [&](auto dtype) {
         using T = decltype(dtype);
         common::Span<T> index_data_span = {index.data<T>(), index.Size()};
-        SetIndexData(
-            index_data_span, rbegin, ft, batch_threads, batch, is_valid, n_bins_total,
-            [offsets](auto bin_idx, auto fidx) { return static_cast<T>(bin_idx - offsets[fidx]); });
+        SetIndexData(index_data_span, rbegin, ft, batch_threads, batch, is_valid, n_bins_total,
+                     index.MakeCompressor<T>());
       });
     } else {
-      /* For sparse DMatrix we have to store index of feature for each bin
-         in index field to chose right offset. So offset is nullptr and index is
-         not reduced */
       common::Span<uint32_t> index_data_span = {index.data<uint32_t>(), n_index};
+      // no compression
       SetIndexData(index_data_span, rbegin, ft, batch_threads, batch, is_valid, n_bins_total,
                    [](auto idx, auto) { return idx; });
     }
-
-    common::ParallelFor(n_bins_total, n_threads, [&](bst_omp_uint idx) {
-      for (int32_t tid = 0; tid < n_threads; ++tid) {
-        hit_count[idx] += hit_count_tloc_[tid * n_bins_total + idx];
-        hit_count_tloc_[tid * n_bins_total + idx] = 0;  // reset for next batch
-      }
-    });
+    this->GatherHitCount(n_threads, n_bins_total);
   }
 
  public:
@@ -129,12 +127,12 @@ class GHistIndexMatrix {
   std::vector<size_t> row_ptr;
   /*! \brief The index data */
   common::Index index;
-  /*! \brief hit count of each index */
+  /*! \brief hit count of each index, used for constructing the ColumnMatrix */
   std::vector<size_t> hit_count;
   /*! \brief The corresponding cuts */
   common::HistogramCuts cut;
   /*! \brief max_bin for each feature. */
-  size_t max_num_bins;
+  bst_bin_t max_num_bins;
   /*! \brief base row index for current page (used by external memory) */
   size_t base_rowid{0};
 
@@ -149,6 +147,13 @@ class GHistIndexMatrix {
    *        for push batch.
    */
   GHistIndexMatrix(MetaInfo const& info, common::HistogramCuts&& cuts, bst_bin_t max_bin_per_feat);
+  /**
+   * \brief Constructor fro Iterative DMatrix where we might copy an existing ellpack page
+   *        to host gradient index.
+   */
+  GHistIndexMatrix(Context const* ctx, MetaInfo const& info, EllpackPage const& page,
+                   BatchParam const& p);
+
   /**
    * \brief Constructor for external memory.
    */

src/data/iterative_dmatrix.cc

@@ -205,12 +205,11 @@ void IterativeDMatrix::InitFromCPU(DataIterHandle iter_handle, float missing,
 
 BatchSet<GHistIndexMatrix> IterativeDMatrix::GetGradientIndex(BatchParam const& param) {
   CheckParam(param);
-  CHECK(ghist_) << R"(`QuantileDMatrix` is not initialized with CPU data but used for CPU training.
-Possible solutions:
-- Use `DMatrix` instead.
-- Use CPU input for `QuantileDMatrix`.
-- Run training on GPU.
-)";
+  if (!ghist_) {
+    CHECK(ellpack_);
+    ghist_ = std::make_shared<GHistIndexMatrix>(&ctx_, Info(), *ellpack_, param);
+  }
+
   auto begin_iter =
       BatchIterator<GHistIndexMatrix>(new SimpleBatchIteratorImpl<GHistIndexMatrix>(ghist_));
   return BatchSet<GHistIndexMatrix>(begin_iter);

src/data/iterative_dmatrix.h

@@ -29,20 +29,17 @@ namespace data {
  * `QuantileDMatrix` is an intermediate storage for quantilization results including
  * quantile cuts and histogram index. Quantilization is designed to be performed on stream
  * of data (or batches of it). As a result, the `QuantileDMatrix` is also designed to work
- * with batches of data. During initializaion, it will walk through the data multiple
- * times iteratively in order to perform quantilization. This design can help us reduce
- * memory usage significantly by avoiding data concatenation along with removing the CSR
- * matrix `SparsePage`. However, it has its limitation (can be fixed if needed):
+ * with batches of data. During initializaion, it walks through the data multiple times
+ * iteratively in order to perform quantilization. This design helps us reduce memory
+ * usage significantly by avoiding data concatenation along with removing the CSR matrix
+ * `SparsePage`. However, it has its limitation (can be fixed if needed):
  *
  * - It's only supported by hist tree method (both CPU and GPU) since approx requires a
  *   re-calculation of quantiles for each iteration. We can fix this by retaining a
  *   reference to the callback if there are feature requests.
  *
  * - The CPU format and the GPU format are different, the former uses a CSR + CSC for
- *   histogram index while the latter uses only Ellpack. This results into a design that
- *   we can obtain the GPU format from CPU but the other way around is not yet
- *   supported. We can search the bin value from ellpack to recover the feature index when
- *   we support copying data from GPU to CPU.
+ *   histogram index while the latter uses only Ellpack.
  */
 class IterativeDMatrix : public DMatrix {
   MetaInfo info_;