Use ellpack for prediction only when sparsepage doesn't exist. (#5504)

dmlc · Apr 10, 2020 · 6671b42 · 6671b42
1 parent ad826e9
commit 6671b42
Show file tree

Hide file tree

Showing 35 changed files with 166 additions and 116 deletions.
diff --git a/include/xgboost/gbm.h b/include/xgboost/gbm.h
@@ -181,12 +181,6 @@ class GradientBooster : public Model, public Configurable {
       const std::string& name,
       GenericParameter const* generic_param,
       LearnerModelParam const* learner_model_param);
-
-  static void AssertGPUSupport() {
-#ifndef XGBOOST_USE_CUDA
-    LOG(FATAL) << "XGBoost version not compiled with GPU support.";
-#endif  // XGBOOST_USE_CUDA
-  }
 };
 
 /*!

diff --git a/src/c_api/c_api.cc b/src/c_api/c_api.cc
@@ -85,7 +85,7 @@ XGB_DLL int XGDMatrixCreateFromArrayInterfaceColumns(char const* c_json_strs,
                                                      int nthread,
                                                      DMatrixHandle* out) {
   API_BEGIN();
-  LOG(FATAL) << "XGBoost not compiled with CUDA";
+  common::AssertGPUSupport();
   API_END();
 }
 
@@ -94,7 +94,7 @@ XGB_DLL int XGDMatrixCreateFromArrayInterface(char const* c_json_strs,
                                               int nthread,
                                               DMatrixHandle* out) {
   API_BEGIN();
-  LOG(FATAL) << "XGBoost not compiled with CUDA";
+  common::AssertGPUSupport();
   API_END();
 }
 
@@ -521,7 +521,7 @@ XGB_DLL int XGBoosterPredictFromArrayInterfaceColumns(BoosterHandle handle,
                                                       float const** out_result) {
   API_BEGIN();
   CHECK_HANDLE();
-  LOG(FATAL) << "XGBoost not compiled with CUDA.";
+  common::AssertGPUSupport();
   API_END();
 }
 XGB_DLL int XGBoosterPredictFromArrayInterface(BoosterHandle handle,
@@ -535,7 +535,7 @@ XGB_DLL int XGBoosterPredictFromArrayInterface(BoosterHandle handle,
                                                const float **out_result) {
   API_BEGIN();
   CHECK_HANDLE();
-  LOG(FATAL) << "XGBoost not compiled with CUDA.";
+  common::AssertGPUSupport();
   API_END();
 }
 #endif  // !defined(XGBOOST_USE_CUDA)

diff --git a/src/common/common.h b/src/common/common.h
@@ -147,6 +147,13 @@ class Range {
 };
 
 int AllVisibleGPUs();
+
+inline void AssertGPUSupport() {
+#ifndef XGBOOST_USE_CUDA
+    LOG(FATAL) << "XGBoost version not compiled with GPU support.";
+#endif  // XGBOOST_USE_CUDA
+}
+
 }  // namespace common
 }  // namespace xgboost
 #endif  // XGBOOST_COMMON_COMMON_H_
diff --git a/src/data/data.cc b/src/data/data.cc
@@ -293,7 +293,7 @@ void MetaInfo::SetInfo(const char* key, const void* dptr, DataType dtype, size_t
 
 #if !defined(XGBOOST_USE_CUDA)
 void MetaInfo::SetInfo(const char * c_key, std::string const& interface_str) {
-  LOG(FATAL) << "XGBoost version is not compiled with GPU support";
+  common::AssertGPUSupport();
 }
 #endif  // !defined(XGBOOST_USE_CUDA)
 

diff --git a/src/gbm/gblinear.cc b/src/gbm/gblinear.cc
@@ -22,6 +22,7 @@
 
 #include "gblinear_model.h"
 #include "../common/timer.h"
+#include "../common/common.h"
 
 namespace xgboost {
 namespace gbm {
@@ -68,7 +69,7 @@ class GBLinear : public GradientBooster {
     updater_->Configure(cfg);
     monitor_.Init("GBLinear");
     if (param_.updater == "gpu_coord_descent") {
-      this->AssertGPUSupport();
+      common::AssertGPUSupport();
     }
   }
 

diff --git a/src/gbm/gbtree.cc b/src/gbm/gbtree.cc
@@ -172,7 +172,7 @@ void GBTree::ConfigureUpdaters() {
       tparam_.updater_seq = "grow_quantile_histmaker";
       break;
     case TreeMethod::kGPUHist: {
-      this->AssertGPUSupport();
+      common::AssertGPUSupport();
       tparam_.updater_seq = "grow_gpu_hist";
       break;
     }
@@ -391,17 +391,21 @@ GBTree::GetPredictor(HostDeviceVector<float> const *out_pred,
       CHECK(gpu_predictor_);
       return gpu_predictor_;
 #else
-      this->AssertGPUSupport();
+      common::AssertGPUSupport();
 #endif  // defined(XGBOOST_USE_CUDA)
     }
     CHECK(cpu_predictor_);
     return cpu_predictor_;
   }
 
-  auto on_device =
-      f_dmat &&
-      (f_dmat->PageExists<EllpackPage>() ||
-      (*(f_dmat->GetBatches<SparsePage>().begin())).data.DeviceCanRead());
+  // Data comes from Device DMatrix.
+  auto is_ellpack = f_dmat && f_dmat->PageExists<EllpackPage>() &&
+                    !f_dmat->PageExists<SparsePage>();
+  // Data comes from device memory, like CuDF or CuPy.
+  auto is_from_device =
+      f_dmat && f_dmat->PageExists<SparsePage>() &&
+      (*(f_dmat->GetBatches<SparsePage>().begin())).data.DeviceCanRead();
+  auto on_device = is_ellpack || is_from_device;
 
   // Use GPU Predictor if data is already on device and gpu_id is set.
   if (on_device && generic_param_->gpu_id >= 0) {
@@ -434,7 +438,7 @@ GBTree::GetPredictor(HostDeviceVector<float> const *out_pred,
     CHECK(gpu_predictor_);
     return gpu_predictor_;
 #else
-    this->AssertGPUSupport();
+    common::AssertGPUSupport();
     return cpu_predictor_;
 #endif  // defined(XGBOOST_USE_CUDA)
   }

diff --git a/src/predictor/gpu_predictor.cu b/src/predictor/gpu_predictor.cu
@@ -348,21 +348,21 @@ class GPUPredictor : public xgboost::Predictor {
     model_.Init(model, tree_begin, tree_end, generic_param_->gpu_id);
     out_preds->SetDevice(generic_param_->gpu_id);
 
-    if (dmat->PageExists<EllpackPage>()) {
+    if (dmat->PageExists<SparsePage>()) {
+      size_t batch_offset = 0;
+      for (auto &batch : dmat->GetBatches<SparsePage>()) {
+        this->PredictInternal(batch, model.learner_model_param->num_feature,
+                              out_preds, batch_offset);
+        batch_offset += batch.Size() * model.learner_model_param->num_output_group;
+      }
+    } else {
       size_t batch_offset = 0;
       for (auto const& page : dmat->GetBatches<EllpackPage>()) {
         this->PredictInternal(
             page.Impl()->GetDeviceAccessor(generic_param_->gpu_id), out_preds,
             batch_offset);
         batch_offset += page.Impl()->n_rows;
       }
-    } else {
-      size_t batch_offset = 0;
-      for (auto &batch : dmat->GetBatches<SparsePage>()) {
-        this->PredictInternal(batch, model.learner_model_param->num_feature,
-                              out_preds, batch_offset);
-        batch_offset += batch.Size() * model.learner_model_param->num_output_group;
-      }
     }
   }
 

diff --git a/tests/cpp/c_api/test_c_api.cc b/tests/cpp/c_api/test_c_api.cc
@@ -82,7 +82,7 @@ TEST(CAPI, Version) {
 
 TEST(CAPI, ConfigIO) {
   size_t constexpr kRows = 10;
-  auto p_dmat = RandomDataGenerator(kRows, 10, 0).GenerateDMatix();
+  auto p_dmat = RandomDataGenerator(kRows, 10, 0).GenerateDMatrix();
   std::vector<std::shared_ptr<DMatrix>> mat {p_dmat};
   std::vector<bst_float> labels(kRows);
   for (size_t i = 0; i < labels.size(); ++i) {
@@ -115,7 +115,7 @@ TEST(CAPI, JsonModelIO) {
   size_t constexpr kRows = 10;
   dmlc::TemporaryDirectory tempdir;
 
-  auto p_dmat = RandomDataGenerator(kRows, 10, 0).GenerateDMatix();
+  auto p_dmat = RandomDataGenerator(kRows, 10, 0).GenerateDMatrix();
   std::vector<std::shared_ptr<DMatrix>> mat {p_dmat};
   std::vector<bst_float> labels(kRows);
   for (size_t i = 0; i < labels.size(); ++i) {

diff --git a/tests/cpp/common/test_column_matrix.cc b/tests/cpp/common/test_column_matrix.cc
@@ -13,7 +13,7 @@ TEST(DenseColumn, Test) {
                           static_cast<uint64_t>(std::numeric_limits<uint16_t>::max()) + 1,
                           static_cast<uint64_t>(std::numeric_limits<uint16_t>::max()) + 2};
   for (size_t max_num_bin : max_num_bins) {
-    auto dmat = RandomDataGenerator(100, 10, 0.0).GenerateDMatix();
+    auto dmat = RandomDataGenerator(100, 10, 0.0).GenerateDMatrix();
     GHistIndexMatrix gmat;
     gmat.Init(dmat.get(), max_num_bin);
     ColumnMatrix column_matrix;
@@ -61,7 +61,7 @@ TEST(SparseColumn, Test) {
                           static_cast<uint64_t>(std::numeric_limits<uint16_t>::max()) + 1,
                           static_cast<uint64_t>(std::numeric_limits<uint16_t>::max()) + 2};
   for (size_t max_num_bin : max_num_bins) {
-    auto dmat = RandomDataGenerator(100, 1, 0.85).GenerateDMatix();
+    auto dmat = RandomDataGenerator(100, 1, 0.85).GenerateDMatrix();
     GHistIndexMatrix gmat;
     gmat.Init(dmat.get(), max_num_bin);
     ColumnMatrix column_matrix;
@@ -102,7 +102,7 @@ TEST(DenseColumnWithMissing, Test) {
                               static_cast<uint64_t>(std::numeric_limits<uint16_t>::max()) + 1,
                               static_cast<uint64_t>(std::numeric_limits<uint16_t>::max()) + 2 };
   for (size_t max_num_bin : max_num_bins) {
-    auto dmat = RandomDataGenerator(100, 1, 0.5).GenerateDMatix();
+    auto dmat = RandomDataGenerator(100, 1, 0.5).GenerateDMatrix();
     GHistIndexMatrix gmat;
     gmat.Init(dmat.get(), max_num_bin);
     ColumnMatrix column_matrix;

diff --git a/tests/cpp/common/test_hist_util.cc b/tests/cpp/common/test_hist_util.cc
@@ -128,7 +128,7 @@ TEST(CutsBuilder, SearchGroupInd) {
   size_t constexpr kRows = 17;
   size_t constexpr kCols = 15;
 
-  auto p_mat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatix();
+  auto p_mat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
 
   std::vector<bst_int> group(kNumGroups);
   group[0] = 2;
@@ -155,7 +155,7 @@ TEST(SparseCuts, SingleThreadedBuild) {
   size_t constexpr kCols = 31;
   size_t constexpr kBins = 256;
 
-  auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatix();
+  auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
 
   common::GHistIndexMatrix hmat;
   hmat.Init(p_fmat.get(), kBins);
@@ -206,12 +206,12 @@ TEST(SparseCuts, MultiThreadedBuild) {
       };
 
   {
-    auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatix();
+    auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
     Compare(p_fmat.get());
   }
 
   {
-    auto p_fmat = RandomDataGenerator(kRows, kCols, 0.0001).GenerateDMatix();
+    auto p_fmat = RandomDataGenerator(kRows, kCols, 0.0001).GenerateDMatrix();
     Compare(p_fmat.get());
   }
 
@@ -360,7 +360,7 @@ TEST(HistUtil, IndexBinBound) {
 
   size_t bin_id = 0;
   for (auto max_bin : bin_sizes) {
-    auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatix();
+    auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
 
     common::GHistIndexMatrix hmat;
     hmat.Init(p_fmat.get(), max_bin);
@@ -381,7 +381,7 @@ TEST(HistUtil, SparseIndexBinBound) {
 
   size_t bin_id = 0;
   for (auto max_bin : bin_sizes) {
-    auto p_fmat = RandomDataGenerator(kRows, kCols, 0.2).GenerateDMatix();
+    auto p_fmat = RandomDataGenerator(kRows, kCols, 0.2).GenerateDMatrix();
     common::GHistIndexMatrix hmat;
     hmat.Init(p_fmat.get(), max_bin);
     EXPECT_EQ(expected_bin_type_sizes[bin_id++], hmat.index.GetBinTypeSize());
@@ -404,7 +404,7 @@ TEST(HistUtil, IndexBinData) {
   size_t constexpr kCols = 10;
 
   for (auto max_bin : kBinSizes) {
-    auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatix();
+    auto p_fmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
     common::GHistIndexMatrix hmat;
     hmat.Init(p_fmat.get(), max_bin);
     uint32_t* offsets = hmat.index.Offset();
@@ -434,7 +434,7 @@ TEST(HistUtil, SparseIndexBinData) {
   size_t constexpr kCols = 10;
 
   for (auto max_bin : bin_sizes) {
-    auto p_fmat = RandomDataGenerator(kRows, kCols, 0.2).GenerateDMatix();
+    auto p_fmat = RandomDataGenerator(kRows, kCols, 0.2).GenerateDMatrix();
     common::GHistIndexMatrix hmat;
     hmat.Init(p_fmat.get(), max_bin);
     EXPECT_EQ(hmat.index.Offset(), nullptr);

diff --git a/tests/cpp/data/test_adapter.cc b/tests/cpp/data/test_adapter.cc
@@ -68,7 +68,7 @@ TEST(Adapter, CSCAdapterColsMoreThanRows) {
 }
 
 TEST(CAPI, DMatrixSliceAdapterFromSimpleDMatrix) {
-  auto p_dmat = RandomDataGenerator(6, 2, 1.0).GenerateDMatix();
+  auto p_dmat = RandomDataGenerator(6, 2, 1.0).GenerateDMatrix();
 
   std::vector<int> ridx_set = {1, 3, 5};
   data::DMatrixSliceAdapter adapter(p_dmat.get(),

diff --git a/tests/cpp/data/test_device_dmatrix.cu b/tests/cpp/data/test_device_dmatrix.cu
@@ -23,7 +23,7 @@ TEST(DeviceDMatrix, RowMajor) {
   auto adapter = common::AdapterFromData(x_device, num_rows, num_columns);
 
   data::DeviceDMatrix dmat(&adapter,
-                                  std::numeric_limits<float>::quiet_NaN(), 1, 256);
+                           std::numeric_limits<float>::quiet_NaN(), 1, 256);
 
   auto &batch = *dmat.GetBatches<EllpackPage>({0, 256, 0}).begin();
   auto impl = batch.Impl();
@@ -60,7 +60,7 @@ TEST(DeviceDMatrix, RowMajorMissing) {
   EXPECT_EQ(iterator[1], impl->GetDeviceAccessor(0).NullValue());
   EXPECT_EQ(iterator[5], impl->GetDeviceAccessor(0).NullValue());
   // null values get placed after valid values in a row
-  EXPECT_EQ(iterator[7], impl->GetDeviceAccessor(0).NullValue()); 
+  EXPECT_EQ(iterator[7], impl->GetDeviceAccessor(0).NullValue());
   EXPECT_EQ(dmat.Info().num_col_, num_columns);
   EXPECT_EQ(dmat.Info().num_row_, num_rows);
   EXPECT_EQ(dmat.Info().num_nonzero_, num_rows*num_columns-3);

diff --git a/tests/cpp/data/test_ellpack_page.cu b/tests/cpp/data/test_ellpack_page.cu
@@ -17,7 +17,7 @@ namespace xgboost {
 TEST(EllpackPage, EmptyDMatrix) {
   constexpr int kNRows = 0, kNCols = 0, kMaxBin = 256;
   constexpr float kSparsity = 0;
-  auto dmat = RandomDataGenerator(kNRows, kNCols, kSparsity).GenerateDMatix();
+  auto dmat = RandomDataGenerator(kNRows, kNCols, kSparsity).GenerateDMatrix();
   auto& page = *dmat->GetBatches<EllpackPage>({0, kMaxBin}).begin();
   auto impl = page.Impl();
   ASSERT_EQ(impl->row_stride, 0);

diff --git a/tests/cpp/data/test_simple_dmatrix.cc b/tests/cpp/data/test_simple_dmatrix.cc
@@ -220,7 +220,7 @@ TEST(SimpleDMatrix, FromFile) {
 TEST(SimpleDMatrix, Slice) {
   const int kRows = 6;
   const int kCols = 2;
-  auto p_dmat = RandomDataGenerator(kRows, kCols, 1.0).GenerateDMatix();
+  auto p_dmat = RandomDataGenerator(kRows, kCols, 1.0).GenerateDMatrix();
   auto &labels = p_dmat->Info().labels_.HostVector();
   auto &weights = p_dmat->Info().weights_.HostVector();
   auto &base_margin = p_dmat->Info().base_margin_.HostVector();

diff --git a/tests/cpp/gbm/test_gbtree.cc b/tests/cpp/gbm/test_gbtree.cc
@@ -55,7 +55,7 @@ TEST(GBTree, WrongUpdater) {
   size_t constexpr kRows = 17;
   size_t constexpr kCols = 15;
 
-  auto p_dmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatix();
+  auto p_dmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
 
   p_dmat->Info().labels_.Resize(kRows);
 
@@ -67,10 +67,11 @@ TEST(GBTree, WrongUpdater) {
 
 #ifdef XGBOOST_USE_CUDA
 TEST(GBTree, ChoosePredictor) {
+  // The test ensures data don't get pulled into device.
   size_t constexpr kRows = 17;
   size_t constexpr kCols = 15;
 
-  auto p_dmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatix();
+  auto p_dmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
 
   auto& data = (*(p_dmat->GetBatches<SparsePage>().begin())).data;
   p_dmat->Info().labels_.Resize(kRows);
@@ -195,7 +196,7 @@ TEST(Dart, JsonIO) {
 TEST(Dart, Prediction) {
   size_t constexpr kRows = 16, kCols = 10;
 
-  auto p_mat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatix();
+  auto p_mat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
 
   std::vector<bst_float> labels (kRows);
   for (size_t i = 0; i < kRows; ++i) {

diff --git a/tests/cpp/helpers.cc b/tests/cpp/helpers.cc
@@ -260,8 +260,8 @@ void RandomDataGenerator::GenerateCSR(
 }
 
 std::shared_ptr<DMatrix>
-RandomDataGenerator::GenerateDMatix(bool with_label, bool float_label,
-                                    size_t classes) const {
+RandomDataGenerator::GenerateDMatrix(bool with_label, bool float_label,
+                                     size_t classes) const {
   HostDeviceVector<float> data;
   HostDeviceVector<bst_row_t> rptrs;
   HostDeviceVector<bst_feature_t> columns;
@@ -399,7 +399,7 @@ std::unique_ptr<GradientBooster> CreateTrainedGBM(
   std::unique_ptr<GradientBooster> gbm {
     GradientBooster::Create(name, generic_param, learner_model_param)};
   gbm->Configure(kwargs);
-  auto p_dmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatix();
+  auto p_dmat = RandomDataGenerator(kRows, kCols, 0).GenerateDMatrix();
 
   std::vector<float> labels(kRows);
   for (size_t i = 0; i < kRows; ++i) {

diff --git a/tests/cpp/helpers.cu b/tests/cpp/helpers.cu
@@ -0,0 +1,17 @@
+#include "helpers.h"
+#include "../../src/data/device_adapter.cuh"
+#include "../../src/data/device_dmatrix.h"
+
+namespace xgboost {
+std::shared_ptr<DMatrix> RandomDataGenerator::GenerateDeviceDMatrix(bool with_label,
+                                                                    bool float_label,
+                                                                    size_t classes) {
+  std::vector<HostDeviceVector<float>> storage(cols_);
+  std::string arr = this->GenerateColumnarArrayInterface(&storage);
+  auto adapter = data::CudfAdapter(arr);
+  std::shared_ptr<DMatrix> m {
+    new data::DeviceDMatrix{&adapter,
+          std::numeric_limits<float>::quiet_NaN(), 1, 256}};
+  return m;
+}
+}  // namespace xgboost