Avoid thread block with sparse data. (#7255)

Author: trivialfis

src/predictor/cpu_predictor.cc

@@ -253,17 +253,32 @@ class CPUPredictor : public Predictor {
                       gbm::GBTreeModel const &model, int32_t tree_begin,
                       int32_t tree_end) const {
     const int threads = omp_get_max_threads();
+    constexpr double kDensityThresh = .5;
+    size_t total = std::max(p_fmat->Info().num_row_ * p_fmat->Info().num_col_,
+                            static_cast<uint64_t>(1));
+    double density = static_cast<double>(p_fmat->Info().num_nonzero_) /
+                     static_cast<double>(total);
+    bool blocked = density > kDensityThresh;
+
     std::vector<RegTree::FVec> feat_vecs;
-    InitThreadTemp(threads * kBlockOfRowsSize,
+    InitThreadTemp(threads * (blocked ? kBlockOfRowsSize : 1),
                    model.learner_model_param->num_feature, &feat_vecs);
-    for (auto const& batch : p_fmat->GetBatches<SparsePage>()) {
+    for (auto const &batch : p_fmat->GetBatches<SparsePage>()) {
       CHECK_EQ(out_preds->size(),
-               p_fmat->Info().num_row_ * model.learner_model_param->num_output_group);
+               p_fmat->Info().num_row_ *
+                   model.learner_model_param->num_output_group);
       size_t constexpr kUnroll = 8;
-      PredictBatchByBlockOfRowsKernel<SparsePageView<kUnroll>,
-                          kBlockOfRowsSize>(SparsePageView<kUnroll>{&batch},
-                                              out_preds, model, tree_begin,
-                                              tree_end, &feat_vecs);
+      if (blocked) {
+        PredictBatchByBlockOfRowsKernel<SparsePageView<kUnroll>,
+                                        kBlockOfRowsSize>(
+            SparsePageView<kUnroll>{&batch}, out_preds, model, tree_begin,
+            tree_end, &feat_vecs);
+
+      } else {
+        PredictBatchByBlockOfRowsKernel<SparsePageView<kUnroll>, 1>(
+            SparsePageView<kUnroll>{&batch}, out_preds, model, tree_begin,
+            tree_end, &feat_vecs);
+      }
     }
   }
 
@@ -316,7 +331,7 @@ class CPUPredictor : public Predictor {
                          tree_end);
   }
 
-  template <typename Adapter>
+  template <typename Adapter, size_t kBlockSize>
   void DispatchedInplacePredict(dmlc::any const &x, std::shared_ptr<DMatrix> p_m,
                                 const gbm::GBTreeModel &model, float missing,
                                 PredictionCacheEntry *out_preds,
@@ -336,9 +351,9 @@ class CPUPredictor : public Predictor {
     std::vector<Entry> workspace(m->NumColumns() * 8 * threads);
     auto &predictions = out_preds->predictions.HostVector();
     std::vector<RegTree::FVec> thread_temp;
-    InitThreadTemp(threads * kBlockOfRowsSize,
-                   model.learner_model_param->num_feature, &thread_temp);
-    PredictBatchByBlockOfRowsKernel<AdapterView<Adapter>, kBlockOfRowsSize>(
+    InitThreadTemp(threads * kBlockSize, model.learner_model_param->num_feature,
+                   &thread_temp);
+    PredictBatchByBlockOfRowsKernel<AdapterView<Adapter>, kBlockSize>(
         AdapterView<Adapter>(m.get(), missing, common::Span<Entry>{workspace}),
         &predictions, model, tree_begin, tree_end, &thread_temp);
   }
@@ -348,16 +363,16 @@ class CPUPredictor : public Predictor {
                       PredictionCacheEntry *out_preds, uint32_t tree_begin,
                       unsigned tree_end) const override {
     if (x.type() == typeid(std::shared_ptr<data::DenseAdapter>)) {
-      this->DispatchedInplacePredict<data::DenseAdapter>(
+      this->DispatchedInplacePredict<data::DenseAdapter, kBlockOfRowsSize>(
           x, p_m, model, missing, out_preds, tree_begin, tree_end);
     } else if (x.type() == typeid(std::shared_ptr<data::CSRAdapter>)) {
-      this->DispatchedInplacePredict<data::CSRAdapter>(
+      this->DispatchedInplacePredict<data::CSRAdapter, 1>(
           x, p_m, model, missing, out_preds, tree_begin, tree_end);
     } else if (x.type() == typeid(std::shared_ptr<data::ArrayAdapter>)) {
-      this->DispatchedInplacePredict<data::ArrayAdapter> (
+      this->DispatchedInplacePredict<data::ArrayAdapter, kBlockOfRowsSize> (
           x, p_m, model, missing, out_preds, tree_begin, tree_end);
     } else if (x.type() == typeid(std::shared_ptr<data::CSRArrayAdapter>)) {
-      this->DispatchedInplacePredict<data::CSRArrayAdapter> (
+      this->DispatchedInplacePredict<data::CSRArrayAdapter, 1> (
           x, p_m, model, missing, out_preds, tree_begin, tree_end);
     } else {
       return false;

tests/cpp/common/test_hist_util.h

@@ -247,7 +247,7 @@ void TestCategoricalSketch(size_t n, size_t num_categories, int32_t num_bins,
   ASSERT_TRUE(is_unique);
 
   x.resize(n_uniques);
-  for (size_t i = 0; i < n_uniques; ++i) {
+  for (decltype(n_uniques) i = 0; i < n_uniques; ++i) {
     ASSERT_EQ(x[i], values[i]);
   }
 }

tests/cpp/predictor/test_cpu_predictor.cc

@@ -247,4 +247,9 @@ TEST(CpuPredictor, UpdatePredictionCache) {
 TEST(CpuPredictor, LesserFeatures) {
   TestPredictionWithLesserFeatures("cpu_predictor");
 }
+
+TEST(CpuPredictor, Sparse) {
+  TestSparsePrediction(0.2, "cpu_predictor");
+  TestSparsePrediction(0.8, "cpu_predictor");
+}
 }  // namespace xgboost

tests/cpp/predictor/test_gpu_predictor.cu

@@ -256,5 +256,10 @@ TEST(GPUPredictor, PredictLeafBasic) {
     ASSERT_EQ(v, 0);
   }
 }
+
+TEST(GPUPredictor, Sparse) {
+  TestSparsePrediction(0.2, "gpu_predictor");
+  TestSparsePrediction(0.8, "gpu_predictor");
+}
 }  // namespace predictor
 }  // namespace xgboost

tests/cpp/predictor/test_predictor.cc

@@ -11,6 +11,7 @@
 #include "test_predictor.h"
 
 #include "../helpers.h"
+#include "../../../src/data/adapter.h"
 #include "../../../src/common/io.h"
 #include "../../../src/common/categorical.h"
 #include "../../../src/common/bitfield.h"
@@ -355,4 +356,57 @@ void TestIterationRange(std::string name) {
     ASSERT_EQ(h_sliced, h_range);
   }
 }
+
+void TestSparsePrediction(float sparsity, std::string predictor) {
+  size_t constexpr kRows = 512, kCols = 128;
+  auto Xy = RandomDataGenerator(kRows, kCols, sparsity).GenerateDMatrix(true);
+  std::unique_ptr<Learner> learner{Learner::Create({Xy})};
+  learner->Configure();
+  for (size_t i = 0; i < 4; ++i) {
+    learner->UpdateOneIter(i, Xy);
+  }
+
+  HostDeviceVector<float> sparse_predt;
+
+  Json model{Object{}};
+  learner->SaveModel(&model);
+
+  learner.reset(Learner::Create({Xy}));
+  learner->LoadModel(model);
+
+  learner->SetParam("predictor", predictor);
+  learner->Predict(Xy, false, &sparse_predt, 0, 0);
+
+  std::vector<float> with_nan(kRows * kCols, std::numeric_limits<float>::quiet_NaN());
+  for (auto const& page : Xy->GetBatches<SparsePage>()) {
+    auto batch = page.GetView();
+    for (size_t i = 0; i < batch.Size(); ++i) {
+      auto row = batch[i];
+      for (auto e : row) {
+        with_nan[i * kCols + e.index] = e.fvalue;
+      }
+    }
+  }
+
+  learner->SetParam("predictor", "cpu_predictor");
+  // Xcode_12.4 doesn't compile with `std::make_shared`.
+  auto dense = std::shared_ptr<data::DenseAdapter>(
+      new data::DenseAdapter(with_nan.data(), kRows, kCols));
+  HostDeviceVector<float> *p_dense_predt;
+  learner->InplacePredict(dmlc::any(dense), nullptr, PredictionType::kValue,
+                          std::numeric_limits<float>::quiet_NaN(), &p_dense_predt,
+                          0, 0);
+
+  auto const& dense_predt = *p_dense_predt;
+  if (predictor == "cpu_predictor") {
+    ASSERT_EQ(dense_predt.HostVector(), sparse_predt.HostVector());
+  } else {
+    auto const &h_dense = dense_predt.HostVector();
+    auto const &h_sparse = sparse_predt.HostVector();
+    ASSERT_EQ(h_dense.size(), h_sparse.size());
+    for (size_t i = 0; i < h_dense.size(); ++i) {
+      ASSERT_FLOAT_EQ(h_dense[i], h_sparse[i]);
+    }
+  }
+}
 }  // namespace xgboost

tests/cpp/predictor/test_predictor.h

@@ -70,6 +70,8 @@ void TestCategoricalPrediction(std::string name);
 void TestCategoricalPredictLeaf(StringView name);
 
 void TestIterationRange(std::string name);
+
+void TestSparsePrediction(float sparsity, std::string predictor);
 }  // namespace xgboost
 
 #endif  // XGBOOST_TEST_PREDICTOR_H_