Fix external memory race in colmaker. (#4980)

* Move `GetColDensity` out of omp parallel block.

src/common/io.cc

@@ -10,6 +10,7 @@
 #include <string>
 
 #include "xgboost/logging.h"
+#include "io.h"
 
 namespace xgboost {
 namespace common {

src/common/io.h

@@ -19,8 +19,8 @@ using MemoryFixSizeBuffer = rabit::utils::MemoryFixSizeBuffer;
 using MemoryBufferStream = rabit::utils::MemoryBufferStream;
 
 /*!
- * \brief Input stream that support additional PeekRead
- *  operation, besides read.
+ * \brief Input stream that support additional PeekRead operation,
+ *  besides read.
  */
 class PeekableInStream : public dmlc::Stream {
  public:

src/data/ellpack_page.cu

@@ -121,7 +121,7 @@ void EllpackPageImpl::InitInfo(int device,
 
 // Initialize the buffer to stored compressed features.
 void EllpackPageImpl::InitCompressedData(int device, size_t num_rows) {
-  int num_symbols = matrix.info.n_bins + 1;
+  size_t num_symbols = matrix.info.n_bins + 1;
 
   // Required buffer size for storing data matrix in ELLPack format.
   size_t compressed_size_bytes = common::CompressedBufferWriter::CalculateBufferSize(

src/data/ellpack_page.cuh

@@ -140,7 +140,7 @@ struct RowStateOnDevice {
 // to begin processing on each device
 class DeviceHistogramBuilderState {
  public:
-  explicit DeviceHistogramBuilderState(int n_rows) : device_row_state_(n_rows) {}
+  explicit DeviceHistogramBuilderState(size_t n_rows) : device_row_state_(n_rows) {}
 
   const RowStateOnDevice& GetRowStateOnDevice() const {
     return device_row_state_;

src/data/sparse_page_dmatrix.cc

@@ -41,6 +41,7 @@ class SparseBatchIteratorImpl : public BatchIteratorImpl<T> {
 
 BatchSet<SparsePage> SparsePageDMatrix::GetRowBatches() {
   auto cast = dynamic_cast<SparsePageSource<SparsePage>*>(row_source_.get());
+  CHECK(cast);
   cast->BeforeFirst();
   cast->Next();
   auto begin_iter = BatchIterator<SparsePage>(

src/data/sparse_page_source.h

@@ -1,5 +1,5 @@
 /*!
- *  Copyright (c) 2014 by Contributors
+ *  Copyright (c) 2014-2019 by Contributors
  * \file page_csr_source.h
  *  External memory data source, saved with sparse_batch_page binary format.
  * \author Tianqi Chen
@@ -7,8 +7,6 @@
 #ifndef XGBOOST_DATA_SPARSE_PAGE_SOURCE_H_
 #define XGBOOST_DATA_SPARSE_PAGE_SOURCE_H_
 
-#include <xgboost/base.h>
-#include <xgboost/data.h>
 #include <dmlc/threadediter.h>
 #include <dmlc/timer.h>
 
@@ -20,6 +18,9 @@
 #include <utility>
 #include <vector>
 
+#include "xgboost/base.h"
+#include "xgboost/data.h"
+
 #include "sparse_page_writer.h"
 #include "../common/common.h"
 

src/data/sparse_page_writer.h

@@ -1,5 +1,5 @@
 /*!
- * Copyright (c) 2014 by Contributors
+ * Copyright (c) 2014-2019 by Contributors
  * \file sparse_page_writer.h
  * \author Tianqi Chen
  */

src/objective/rank_obj.cu

@@ -34,7 +34,7 @@ DMLC_REGISTRY_FILE_TAG(rank_obj_gpu);
 #endif  // defined(XGBOOST_USE_CUDA)
 
 struct LambdaRankParam : public XGBoostParameter<LambdaRankParam> {
-  int num_pairsample;
+  size_t num_pairsample;
   float fix_list_weight;
   // declare parameters
   DMLC_DECLARE_PARAMETER(LambdaRankParam) {
@@ -337,7 +337,7 @@ class SortedLabelList {
     // Launch a kernel that populates the segment information for the different groups
     uint32_t *gsegs = group_segments_.data().get();
     const unsigned *dgroups = dgroups_.data().get();
-    int ngroups = dgroups_.size();
+    size_t ngroups = dgroups_.size();
     dh::LaunchN(device_id_, num_elems, nullptr, [=] __device__(unsigned idx){
       // Find the group first
       int group_idx = dh::UpperBound(dgroups, ngroups, idx);
@@ -405,10 +405,10 @@ class SortedLabelList {
                         float weight_normalization_factor) {
     // Group info on device
     const unsigned *dgroups = dgroups_.data().get();
-    int ngroups = dgroups_.size();
+    size_t ngroups = dgroups_.size();
 
-    uint32_t total_items = group_segments_.size();
-    int niter = param_.num_pairsample * total_items;
+    auto total_items = group_segments_.size();
+    size_t niter = param_.num_pairsample * total_items;
 
     float fix_list_weight = param_.fix_list_weight;
 

src/tree/updater_colmaker.cc

@@ -608,17 +608,25 @@ class ColMaker: public TreeUpdater {
         poption = static_cast<int>(num_features) * 2 < this->nthread_ ? 1 : 0;
       }
       if (poption == 0) {
+        std::vector<float> densities(num_features);
+        CHECK_EQ(feat_set.size(), num_features);
+        for (bst_omp_uint i = 0; i < num_features; ++i) {
+          int32_t const fid = feat_set[i];
+          densities.at(i) = p_fmat->GetColDensity(fid);
+        }
+
 #pragma omp parallel for schedule(dynamic, batch_size)
         for (bst_omp_uint i = 0; i < num_features; ++i) {
-          int fid = feat_set[i];
-          const int tid = omp_get_thread_num();
+          int32_t const fid = feat_set[i];
+          int32_t const tid = omp_get_thread_num();
           auto c = batch[fid];
           const bool ind = c.size() != 0 && c[0].fvalue == c[c.size() - 1].fvalue;
-          if (param_.NeedForwardSearch(p_fmat->GetColDensity(fid), ind)) {
+          auto const density = densities[i];
+          if (param_.NeedForwardSearch(density, ind)) {
             this->EnumerateSplit(c.data(), c.data() + c.size(), +1,
                                  fid, gpair, info, stemp_[tid]);
           }
-          if (param_.NeedBackwardSearch(p_fmat->GetColDensity(fid), ind)) {
+          if (param_.NeedBackwardSearch(density, ind)) {
             this->EnumerateSplit(c.data() + c.size() - 1, c.data() - 1, -1,
                                  fid, gpair, info, stemp_[tid]);
           }