Various small cleanups. (dmlc#11097)

- Add deduction guide for Span.
- Merge data split mode setter into the `MetaInfo` column sync method.
- Check CUDA warning for cross space function calls.

Author: trivialfis

cmake/Utils.cmake

@@ -81,7 +81,8 @@ function(xgboost_set_cuda_flags target)
     $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>
     $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=${OpenMP_CXX_FLAGS}>
     $<$<COMPILE_LANGUAGE:CUDA>:-Xfatbin=-compress-all>
-    $<$<COMPILE_LANGUAGE:CUDA>:--default-stream per-thread>)
+    $<$<COMPILE_LANGUAGE:CUDA>:--default-stream per-thread>
+  )
 
   if(FORCE_COLORED_OUTPUT)
     if(FORCE_COLORED_OUTPUT AND (CMAKE_GENERATOR STREQUAL "Ninja") AND
@@ -200,6 +201,12 @@ macro(xgboost_target_properties target)
   if(WIN32 AND MINGW)
     target_compile_options(${target} PUBLIC -static-libstdc++)
   endif()
+
+  if(NOT WIN32 AND ENABLE_ALL_WARNINGS)
+    target_compile_options(${target} PRIVATE
+      $<$<COMPILE_LANGUAGE:CUDA>:-Werror=cross-execution-space-call>
+    )
+  endif()
 endmacro()
 
 # Custom definitions used in xgboost.

include/xgboost/c_api.h

@@ -899,6 +899,8 @@ XGB_DLL int XGBoosterFree(BoosterHandle handle);
  * @brief Reset the booster object to release data caches used for training.
  *
  * @since 3.0.0
+ *
+ * @return 0 when success, -1 when failure happens
  */
 XGB_DLL int XGBoosterReset(BoosterHandle handle);
 

include/xgboost/data.h

@@ -172,31 +172,27 @@ class MetaInfo {
    *                     columns.
    */
   void Extend(MetaInfo const& that, bool accumulate_rows, bool check_column);
-
   /**
    * @brief Synchronize the number of columns across all workers.
    *
    * Normally we just need to find the maximum number of columns across all workers, but
    * in vertical federated learning, since each worker loads its own list of columns,
    * we need to sum them.
    */
-  void SynchronizeNumberOfColumns(Context const* ctx);
-
-  /*! \brief Whether the data is split row-wise. */
-  bool IsRowSplit() const {
-    return data_split_mode == DataSplitMode::kRow;
-  }
+  void SynchronizeNumberOfColumns(Context const* ctx, DataSplitMode split_mode);
 
+  /** @brief Whether the data is split row-wise. */
+  [[nodiscard]] bool IsRowSplit() const { return data_split_mode == DataSplitMode::kRow; }
   /** @brief Whether the data is split column-wise. */
-  bool IsColumnSplit() const { return data_split_mode == DataSplitMode::kCol; }
+  [[nodiscard]] bool IsColumnSplit() const { return data_split_mode == DataSplitMode::kCol; }
   /** @brief Whether this is a learning to rank data. */
-  bool IsRanking() const { return !group_ptr_.empty(); }
+  [[nodiscard]] bool IsRanking() const { return !group_ptr_.empty(); }
 
-  /*!
-   * \brief A convenient method to check if we are doing vertical federated learning, which requires
+  /**
+   * @brief A convenient method to check if we are doing vertical federated learning, which requires
    * some special processing.
    */
-  bool IsVerticalFederated() const;
+  [[nodiscard]] bool IsVerticalFederated() const;
 
   /*!
    * \brief A convenient method to check if the MetaInfo should contain labels.
@@ -330,10 +326,9 @@ struct HostSparsePageView {
   common::Span<bst_idx_t const> offset;
   common::Span<Entry const> data;
 
-  Inst operator[](size_t i) const {
+  [[nodiscard]] Inst operator[](std::size_t i) const {
     auto size = *(offset.data() + i + 1) - *(offset.data() + i);
-    return {data.data() + *(offset.data() + i),
-            static_cast<Inst::index_type>(size)};
+    return {data.data() + *(offset.data() + i), static_cast<Inst::index_type>(size)};
   }
 
   [[nodiscard]] size_t Size() const { return offset.size() == 0 ? 0 : offset.size() - 1; }

include/xgboost/span.h

@@ -37,6 +37,7 @@
 #include <limits>  // numeric_limits
 #include <type_traits>
 #include <utility>  // for move
+#include <vector>   // for vector
 
 #if defined(__CUDACC__)
 #include <cuda_runtime.h>
@@ -707,6 +708,12 @@ class IterSpan {
     return it_ + size();
   }
 };
+
+template <typename T>
+Span(std::vector<T> const&) -> Span<T const>;
+
+template <typename T>
+Span(std::vector<T>&) -> Span<T>;
 }  // namespace xgboost::common
 
 

include/xgboost/tree_model.h

@@ -186,7 +186,8 @@ class RegTree : public Model {
       return this->DefaultLeft() ? this->LeftChild() : this->RightChild();
     }
     /*! \brief feature index of split condition */
-    [[nodiscard]] XGBOOST_DEVICE unsigned SplitIndex() const {
+    [[nodiscard]] XGBOOST_DEVICE bst_feature_t SplitIndex() const {
+      static_assert(!std::is_signed_v<bst_feature_t>);
       return sindex_ & ((1U << 31) - 1U);
     }
     /*! \brief when feature is unknown, whether goes to left child */

src/common/host_device_vector.cu

@@ -413,13 +413,13 @@ template class HostDeviceVector<bst_float>;
 template class HostDeviceVector<double>;
 template class HostDeviceVector<GradientPair>;
 template class HostDeviceVector<GradientPairPrecise>;
-template class HostDeviceVector<int32_t>;   // bst_node_t
-template class HostDeviceVector<uint8_t>;
-template class HostDeviceVector<int8_t>;
+template class HostDeviceVector<std::int32_t>;   // bst_node_t
+template class HostDeviceVector<std::uint8_t>;
+template class HostDeviceVector<std::int8_t>;
 template class HostDeviceVector<FeatureType>;
 template class HostDeviceVector<Entry>;
 template class HostDeviceVector<bst_idx_t>;
-template class HostDeviceVector<uint32_t>;  // bst_feature_t
+template class HostDeviceVector<std::uint32_t>;  // bst_feature_t
 template class HostDeviceVector<RegTree::Node>;
 template class HostDeviceVector<RegTree::CategoricalSplitMatrix::Segment>;
 template class HostDeviceVector<RTreeNodeStat>;

src/common/transform_iterator.h

@@ -62,6 +62,15 @@ class IndexTransformIter {
     ++(*this);
     return ret;
   }
+  IndexTransformIter &operator--() {
+    iter_--;
+    return *this;
+  }
+  IndexTransformIter operator--(int) {
+    auto ret = *this;
+    --(*this);
+    return ret;
+  }
   IndexTransformIter &operator+=(difference_type n) {
     iter_ += n;
     return *this;

src/data/data.cc

@@ -770,7 +770,8 @@ void MetaInfo::Extend(MetaInfo const& that, bool accumulate_rows, bool check_col
   }
 }
 
-void MetaInfo::SynchronizeNumberOfColumns(Context const* ctx) {
+void MetaInfo::SynchronizeNumberOfColumns(Context const* ctx, DataSplitMode split_mode) {
+  this->data_split_mode = split_mode;
   auto op = IsColumnSplit() ? collective::Op::kSum : collective::Op::kMax;
   auto rc = collective::Allreduce(ctx, linalg::MakeVec(&num_col_, 1), op);
   collective::SafeColl(rc);

src/data/iterative_dmatrix.cu

@@ -100,7 +100,6 @@ void IterativeDMatrix::InitFromCUDA(Context const* ctx, BatchParam const& p,
   }
 
   iter.Reset();
-  // Synchronise worker columns
 }
 
 IterativeDMatrix::IterativeDMatrix(std::shared_ptr<EllpackPage> ellpack, MetaInfo const& info,

src/data/proxy_dmatrix.h

@@ -173,7 +173,7 @@ struct ExternalDataInfo {
     info.num_row_ = this->accumulated_rows;
     info.num_col_ = this->n_features;
     info.num_nonzero_ = this->nnz;
-    info.SynchronizeNumberOfColumns(ctx);
+    info.SynchronizeNumberOfColumns(ctx, DataSplitMode::kRow);
     this->Validate();
   }
 };