Optimized ApplySplit, BuildHist and UpdatePredictCache functions on C…

…PU (#5244) * Split up sparse and dense build hist kernels. * Add `PartitionBuilder`.
dmlc · Feb 29, 2020 · 1b97eaf · 1b97eaf
1 parent b81f8cb
commit 1b97eaf
Show file tree

Hide file tree

Showing 9 changed files with 683 additions and 376 deletions.
diff --git a/src/common/column_matrix.h b/src/common/column_matrix.h
@@ -37,6 +37,7 @@ class Column {
   size_t Size() const { return len_; }
   uint32_t GetGlobalBinIdx(size_t idx) const { return index_base_ + index_[idx]; }
   uint32_t GetFeatureBinIdx(size_t idx) const { return index_[idx]; }
+  common::Span<const uint32_t> GetFeatureBinIdxPtr() const { return { index_, len_ }; }
   // column.GetFeatureBinIdx(idx) + column.GetBaseIdx(idx) ==
   // column.GetGlobalBinIdx(idx)
   uint32_t GetBaseIdx() const { return index_base_; }
@@ -186,8 +187,8 @@ class ColumnMatrix {
 
   std::vector<size_t> feature_counts_;
   std::vector<ColumnType> type_;
-  SimpleArray<uint32_t> index_;  // index_: may store smaller integers; needs padding
-  SimpleArray<size_t> row_ind_;
+  std::vector<uint32_t> index_;  // index_: may store smaller integers; needs padding
+  std::vector<size_t> row_ind_;
   std::vector<ColumnBoundary> boundary_;
 
   // index_base_[fid]: least bin id for feature fid

diff --git a/src/common/hist_util.cc b/src/common/hist_util.cc
@@ -672,7 +672,7 @@ void GHistIndexBlockMatrix::Init(const GHistIndexMatrix& gmat,
 }
 
 /*!
- * \brief fill a histogram by zeroes
+ * \brief fill a histogram by zeros in range [begin, end)
  */
 void InitilizeHistByZeroes(GHistRow hist, size_t begin, size_t end) {
   memset(hist.data() + begin, '\0', (end-begin)*sizeof(tree::GradStats));
@@ -719,43 +719,144 @@ void SubtractionHist(GHistRow dst, const GHistRow src1, const GHistRow src2,
   }
 }
 
-
-void GHistBuilder::BuildHist(const std::vector<GradientPair>& gpair,
-                             const RowSetCollection::Elem row_indices,
-                             const GHistIndexMatrix& gmat,
-                             GHistRow hist) {
-  const size_t* rid =  row_indices.begin;
-  const size_t nrows = row_indices.Size();
-  const uint32_t* index = gmat.index.data();
-  const size_t* row_ptr =  gmat.row_ptr.data();
+struct Prefetch {
+ public:
+  static constexpr size_t kCacheLineSize = 64;
+  static constexpr size_t kPrefetchOffset = 10;
+  static constexpr size_t kPrefetchStep =
+      kCacheLineSize / sizeof(decltype(GHistIndexMatrix::index)::value_type);
+
+ private:
+  static constexpr size_t kNoPrefetchSize =
+      kPrefetchOffset + kCacheLineSize /
+      sizeof(decltype(GHistIndexMatrix::row_ptr)::value_type);
+
+ public:
+  static size_t NoPrefetchSize(size_t rows) {
+    return std::min(rows, kNoPrefetchSize);
+  }
+};
+
+constexpr size_t Prefetch::kNoPrefetchSize;
+
+template<typename FPType, bool do_prefetch>
+void BuildHistDenseKernel(const std::vector<GradientPair>& gpair,
+                          const RowSetCollection::Elem row_indices,
+                          const GHistIndexMatrix& gmat,
+                          const size_t n_features,
+                          GHistRow hist) {
+  const size_t size = row_indices.Size();
+  const size_t* rid = row_indices.begin;
   const float* pgh = reinterpret_cast<const float*>(gpair.data());
+  const uint32_t* gradient_index = gmat.index.data();
+  FPType* hist_data = reinterpret_cast<FPType*>(hist.data());
+
+  const uint32_t two {2};  // Each element from 'gpair' and 'hist' contains
+                           // 2 FP values: gradient and hessian.
+                           // So we need to multiply each row-index/bin-index by 2
+                           // to work with gradient pairs as a singe row FP array
+
+  for (size_t i = 0; i < size; ++i) {
+    const size_t icol_start = rid[i] * n_features;
+    const size_t idx_gh = two * rid[i];
+
+    if (do_prefetch) {
+      const size_t icol_start_prefetch = rid[i + Prefetch::kPrefetchOffset] * n_features;
 
-  double* hist_data = reinterpret_cast<double*>(hist.data());
+      PREFETCH_READ_T0(pgh + two * rid[i + Prefetch::kPrefetchOffset]);
+      for (size_t j = icol_start_prefetch; j < icol_start_prefetch + n_features;
+           j += Prefetch::kPrefetchStep) {
+        PREFETCH_READ_T0(gradient_index + j);
+      }
+    }
+
+    for (size_t j = icol_start; j < icol_start + n_features; ++j) {
+      const uint32_t idx_bin = two * gradient_index[j];
+
+      hist_data[idx_bin]   += pgh[idx_gh];
+      hist_data[idx_bin+1] += pgh[idx_gh+1];
+    }
+  }
+}
+
+template<typename FPType, bool do_prefetch>
+void BuildHistSparseKernel(const std::vector<GradientPair>& gpair,
+                           const RowSetCollection::Elem row_indices,
+                           const GHistIndexMatrix& gmat,
+                           GHistRow hist) {
+  const size_t size = row_indices.Size();
+  const size_t* rid = row_indices.begin;
+  const float* pgh = reinterpret_cast<const float*>(gpair.data());
+  const uint32_t* gradient_index = gmat.index.data();
+  const size_t* row_ptr =  gmat.row_ptr.data();
+  FPType* hist_data = reinterpret_cast<FPType*>(hist.data());
 
-  const size_t cache_line_size = 64;
-  const size_t prefetch_offset = 10;
-  size_t no_prefetch_size = prefetch_offset + cache_line_size/sizeof(*rid);
-  no_prefetch_size = no_prefetch_size > nrows ? nrows : no_prefetch_size;
+  const uint32_t two {2};  // Each element from 'gpair' and 'hist' contains
+                           // 2 FP values: gradient and hessian.
+                           // So we need to multiply each row-index/bin-index by 2
+                           // to work with gradient pairs as a singe row FP array
 
-  for (size_t i = 0; i < nrows; ++i) {
+  for (size_t i = 0; i < size; ++i) {
     const size_t icol_start = row_ptr[rid[i]];
     const size_t icol_end = row_ptr[rid[i]+1];
+    const size_t idx_gh = two * rid[i];
+
+    if (do_prefetch) {
+      const size_t icol_start_prftch = row_ptr[rid[i+Prefetch::kPrefetchOffset]];
+      const size_t icol_end_prefect = row_ptr[rid[i+Prefetch::kPrefetchOffset]+1];
 
-    if (i < nrows - no_prefetch_size) {
-      PREFETCH_READ_T0(row_ptr + rid[i + prefetch_offset]);
-      PREFETCH_READ_T0(pgh + 2*rid[i + prefetch_offset]);
+      PREFETCH_READ_T0(pgh + two * rid[i + Prefetch::kPrefetchOffset]);
+      for (size_t j = icol_start_prftch; j < icol_end_prefect; j+=Prefetch::kPrefetchStep) {
+        PREFETCH_READ_T0(gradient_index + j);
+      }
     }
 
     for (size_t j = icol_start; j < icol_end; ++j) {
-      const uint32_t idx_bin = 2*index[j];
-      const size_t idx_gh = 2*rid[i];
-
-      hist_data[idx_bin] += pgh[idx_gh];
+      const uint32_t idx_bin = two * gradient_index[j];
+      hist_data[idx_bin]   += pgh[idx_gh];
       hist_data[idx_bin+1] += pgh[idx_gh+1];
     }
   }
 }
 
+template<typename FPType, bool do_prefetch>
+void BuildHistKernel(const std::vector<GradientPair>& gpair,
+                     const RowSetCollection::Elem row_indices,
+                     const GHistIndexMatrix& gmat, const bool isDense, GHistRow hist) {
+  if (row_indices.Size() && isDense) {
+    const size_t* row_ptr =  gmat.row_ptr.data();
+    const size_t n_features = row_ptr[row_indices.begin[0]+1] - row_ptr[row_indices.begin[0]];
+    BuildHistDenseKernel<FPType, do_prefetch>(gpair, row_indices, gmat, n_features, hist);
+  } else {
+    BuildHistSparseKernel<FPType, do_prefetch>(gpair, row_indices, gmat, hist);
+  }
+}
+
+void GHistBuilder::BuildHist(const std::vector<GradientPair>& gpair,
+                             const RowSetCollection::Elem row_indices,
+                             const GHistIndexMatrix& gmat,
+                             GHistRow hist,
+                             bool isDense) {
+  using FPType = decltype(tree::GradStats::sum_grad);
+  const size_t nrows = row_indices.Size();
+  const size_t no_prefetch_size = Prefetch::NoPrefetchSize(nrows);
+
+  // if need to work with all rows from bin-matrix (e.g. root node)
+  const bool contiguousBlock = (row_indices.begin[nrows - 1] - row_indices.begin[0]) == (nrows - 1);
+
+  if (contiguousBlock) {
+    // contiguous memory access, built-in HW prefetching is enough
+    BuildHistKernel<FPType, false>(gpair, row_indices, gmat, isDense, hist);
+  } else {
+    const RowSetCollection::Elem span1(row_indices.begin, row_indices.end - no_prefetch_size);
+    const RowSetCollection::Elem span2(row_indices.end - no_prefetch_size, row_indices.end);
+
+    BuildHistKernel<FPType, true>(gpair, span1, gmat, isDense, hist);
+    // no prefetching to avoid loading extra memory
+    BuildHistKernel<FPType, false>(gpair, span2, gmat, isDense, hist);
+  }
+}
+
 void GHistBuilder::BuildBlockHist(const std::vector<GradientPair>& gpair,
                                   const RowSetCollection::Elem row_indices,
                                   const GHistIndexBlockMatrix& gmatb,

diff --git a/src/common/hist_util.h b/src/common/hist_util.h
@@ -1,5 +1,5 @@
 /*!
- * Copyright 2017 by Contributors
+ * Copyright 2017-2020 by Contributors
  * \file hist_util.h
  * \brief Utility for fast histogram aggregation
  * \author Philip Cho, Tianqi Chen
@@ -25,75 +25,6 @@
 
 namespace xgboost {
 namespace common {
-
-/*
- * \brief A thin wrapper around dynamically allocated C-style array.
- * Make sure to call resize() before use.
- */
-template<typename T>
-struct SimpleArray {
-  ~SimpleArray() {
-    std::free(ptr_);
-    ptr_ = nullptr;
-  }
-
-  void resize(size_t n) {
-    T* ptr = static_cast<T*>(std::malloc(n * sizeof(T)));
-    CHECK(ptr) << "Failed to allocate memory";
-    if (ptr_) {
-      std::memcpy(ptr, ptr_, n_ * sizeof(T));
-      std::free(ptr_);
-    }
-    ptr_ = ptr;
-    n_ = n;
-  }
-
-  T& operator[](size_t idx) {
-    return ptr_[idx];
-  }
-
-  T& operator[](size_t idx) const {
-    return ptr_[idx];
-  }
-
-  size_t size() const {
-    return n_;
-  }
-
-  T back() const {
-    return ptr_[n_-1];
-  }
-
-  T* data() {
-    return ptr_;
-  }
-
-  const T* data() const {
-    return ptr_;
-  }
-
-
-  T* begin() {
-    return ptr_;
-  }
-
-  const T* begin() const {
-    return ptr_;
-  }
-
-  T* end() {
-    return ptr_ + n_;
-  }
-
-  const T* end() const {
-    return ptr_ + n_;
-  }
-
- private:
-  T* ptr_ = nullptr;
-  size_t n_ = 0;
-};
-
 /*!
  * \brief A single row in global histogram index.
  *  Directly represent the global index in the histogram entry.
@@ -161,7 +92,7 @@ class HistogramCuts {
     return idx;
   }
 
-  BinIdx SearchBin(Entry const& e) {
+  BinIdx SearchBin(Entry const& e) const {
     return SearchBin(e.fvalue, e.index);
   }
 };
@@ -261,8 +192,9 @@ size_t DeviceSketch(int device,
 
 /*!
  * \brief preprocessed global index matrix, in CSR format
- *  Transform floating values to integer index in histogram
- *  This is a global histogram index.
+ *
+ *  Transform floating values to integer index in histogram This is a global histogram
+ *  index for CPU histogram.  On GPU ellpack page is used.
  */
 struct GHistIndexMatrix {
   /*! \brief row pointer to rows by element position */
@@ -606,17 +538,15 @@ class ParallelGHistBuilder {
  */
 class GHistBuilder {
  public:
-  // initialize builder
-  inline void Init(size_t nthread, uint32_t nbins) {
-    nthread_ = nthread;
-    nbins_ = nbins;
-  }
+  GHistBuilder() : nthread_{0}, nbins_{0} {}
+  GHistBuilder(size_t nthread, uint32_t nbins) : nthread_{nthread}, nbins_{nbins} {}
 
   // construct a histogram via histogram aggregation
   void BuildHist(const std::vector<GradientPair>& gpair,
                  const RowSetCollection::Elem row_indices,
                  const GHistIndexMatrix& gmat,
-                 GHistRow hist);
+                 GHistRow hist,
+                 bool isDense);
   // same, with feature grouping
   void BuildBlockHist(const std::vector<GradientPair>& gpair,
                       const RowSetCollection::Elem row_indices,
@@ -625,7 +555,7 @@ class GHistBuilder {
   // construct a histogram via subtraction trick
   void SubtractionTrick(GHistRow self, GHistRow sibling, GHistRow parent);
 
-  uint32_t GetNumBins() {
+  uint32_t GetNumBins() const {
       return nbins_;
   }