Implement GPU predict leaf. (#6187)

include/xgboost/gbm.h

@@ -144,7 +144,7 @@ class GradientBooster : public Model, public Configurable {
    *    we do not limit number of trees, this parameter is only valid for gbtree, but not for gblinear
    */
   virtual void PredictLeaf(DMatrix* dmat,
-                           std::vector<bst_float>* out_preds,
+                           HostDeviceVector<bst_float>* out_preds,
                            unsigned ntree_limit = 0) = 0;
 
   /*!

include/xgboost/predictor.h

@@ -164,10 +164,6 @@ class Predictor {
                                unsigned ntree_limit = 0) = 0;
 
   /**
-   * \fn  virtual void Predictor::PredictLeaf(DMatrix* dmat,
-   * std::vector<bst_float>* out_preds, const gbm::GBTreeModel& model, unsigned
-   * ntree_limit = 0) = 0;
-   *
    * \brief predict the leaf index of each tree, the output will be nsample *
    * ntree vector this is only valid in gbtree predictor.
    *
@@ -177,7 +173,7 @@ class Predictor {
    * \param           ntree_limit (Optional) The ntree limit.
    */
 
-  virtual void PredictLeaf(DMatrix* dmat, std::vector<bst_float>* out_preds,
+  virtual void PredictLeaf(DMatrix* dmat, HostDeviceVector<bst_float>* out_preds,
                            const gbm::GBTreeModel& model,
                            unsigned ntree_limit = 0) = 0;
 

src/gbm/gblinear.cc

@@ -147,9 +147,7 @@ class GBLinear : public GradientBooster {
     }
   }
 
-  void PredictLeaf(DMatrix*,
+  void PredictLeaf(DMatrix *, HostDeviceVector<bst_float> *, unsigned) override {
-                   std::vector<bst_float>*,
-                   unsigned) override {
     LOG(FATAL) << "gblinear does not support prediction of leaf index";
   }
 

src/gbm/gbtree.h

@@ -278,10 +278,9 @@ class GBTree : public GradientBooster {
   }
 
   void PredictLeaf(DMatrix* p_fmat,
-                   std::vector<bst_float>* out_preds,
+                   HostDeviceVector<bst_float>* out_preds,
                    unsigned ntree_limit) override {
-    CHECK(configured_);
+    this->GetPredictor()->PredictLeaf(p_fmat, out_preds, model_, ntree_limit);
-    cpu_predictor_->PredictLeaf(p_fmat, out_preds, model_, ntree_limit);
   }
 
   void PredictContribution(DMatrix* p_fmat,

src/learner.cc

@@ -1106,7 +1106,7 @@ class LearnerImpl : public LearnerIO {
       gbm_->PredictInteractionContributions(data.get(), out_preds, ntree_limit,
                                             approx_contribs);
     } else if (pred_leaf) {
-      gbm_->PredictLeaf(data.get(), &out_preds->HostVector(), ntree_limit);
+      gbm_->PredictLeaf(data.get(), out_preds, ntree_limit);
     } else {
       auto local_cache = this->GetPredictionCache();
       auto& prediction = local_cache->Cache(data, generic_parameters_.gpu_id);

src/predictor/cpu_predictor.cc

@@ -345,7 +345,7 @@ class CPUPredictor : public Predictor {
     }
   }
 
-  void PredictLeaf(DMatrix* p_fmat, std::vector<bst_float>* out_preds,
+  void PredictLeaf(DMatrix* p_fmat, HostDeviceVector<bst_float>* out_preds,
                    const gbm::GBTreeModel& model, unsigned ntree_limit) override {
     const int nthread = omp_get_max_threads();
     InitThreadTemp(nthread, model.learner_model_param->num_feature, &this->thread_temp_);
@@ -355,7 +355,7 @@ class CPUPredictor : public Predictor {
     if (ntree_limit == 0 || ntree_limit > model.trees.size()) {
       ntree_limit = static_cast<unsigned>(model.trees.size());
     }
-    std::vector<bst_float>& preds = *out_preds;
+    std::vector<bst_float>& preds = out_preds->HostVector();
     preds.resize(info.num_row_ * ntree_limit);
     // start collecting the prediction
     for (const auto &batch : p_fmat->GetBatches<SparsePage>()) {

src/predictor/gpu_predictor.cu

@@ -169,7 +169,7 @@ struct DeviceAdapterLoader {
 };
 
 template <typename Loader>
-__device__ float GetLeafWeight(bst_uint ridx, const RegTree::Node* tree,
+__device__ float GetLeafWeight(bst_row_t ridx, const RegTree::Node* tree,
                                common::Span<FeatureType const> split_types,
                                common::Span<RegTree::Segment const> d_cat_ptrs,
                                common::Span<uint32_t const> d_categories,
@@ -201,6 +201,49 @@ __device__ float GetLeafWeight(bst_uint ridx, const RegTree::Node* tree,
   return tree[nidx].LeafValue();
 }
 
+template <typename Loader>
+__device__ bst_node_t GetLeafIndex(bst_row_t ridx, const RegTree::Node* tree,
+                                   Loader const& loader) {
+  bst_node_t nidx = 0;
+  RegTree::Node n = tree[nidx];
+  while (!n.IsLeaf()) {
+    float fvalue = loader.GetElement(ridx, n.SplitIndex());
+    // Missing value
+    if (isnan(fvalue)) {
+      nidx = n.DefaultChild();
+      n = tree[nidx];
+    } else {
+      if (fvalue < n.SplitCond()) {
+        nidx = n.LeftChild();
+        n = tree[nidx];
+      } else {
+        nidx = n.RightChild();
+        n = tree[nidx];
+      }
+    }
+  }
+  return nidx;
+}
+
+template <typename Loader, typename Data>
+__global__ void PredictLeafKernel(Data data,
+                                  common::Span<const RegTree::Node> d_nodes,
+                                  common::Span<float> d_out_predictions,
+                                  common::Span<size_t const> d_tree_segments,
+                                  size_t tree_begin, size_t tree_end, size_t num_features,
+                                  size_t num_rows, size_t entry_start, bool use_shared) {
+  bst_row_t ridx = blockDim.x * blockIdx.x + threadIdx.x;
+  if (ridx >= num_rows) {
+    return;
+  }
+  Loader loader(data, use_shared, num_features, num_rows, entry_start);
+  for (int tree_idx = tree_begin; tree_idx < tree_end; ++tree_idx) {
+    const RegTree::Node* d_tree = &d_nodes[d_tree_segments[tree_idx - tree_begin]];
+    auto leaf = GetLeafIndex(ridx, d_tree, loader);
+    d_out_predictions[ridx * (tree_end - tree_begin) + tree_idx] = leaf;
+  }
+}
+
 template <typename Loader, typename Data>
 __global__ void
 PredictKernel(Data data, common::Span<const RegTree::Node> d_nodes,
@@ -437,6 +480,19 @@ void ExtractPaths(dh::device_vector<gpu_treeshap::PathElement>* paths,
   });
 }
 
+namespace {
+template <size_t kBlockThreads>
+size_t SharedMemoryBytes(size_t cols, size_t max_shared_memory_bytes) {
+  // No way max_shared_memory_bytes that is equal to 0.
+  CHECK_GT(max_shared_memory_bytes, 0);
+  size_t shared_memory_bytes =
+      static_cast<size_t>(sizeof(float) * cols * kBlockThreads);
+  if (shared_memory_bytes > max_shared_memory_bytes) {
+    shared_memory_bytes = 0;
+  }
+  return shared_memory_bytes;
+}
+}  // anonymous namespace
 
 class GPUPredictor : public xgboost::Predictor {
  private:
@@ -450,13 +506,10 @@ class GPUPredictor : public xgboost::Predictor {
     size_t num_rows = batch.Size();
     auto GRID_SIZE = static_cast<uint32_t>(common::DivRoundUp(num_rows, BLOCK_THREADS));
 
-    auto shared_memory_bytes =
+    size_t shared_memory_bytes =
-        static_cast<size_t>(sizeof(float) * num_features * BLOCK_THREADS);
+        SharedMemoryBytes<BLOCK_THREADS>(num_features, max_shared_memory_bytes_);
-    bool use_shared = true;
+    bool use_shared = shared_memory_bytes != 0;
-    if (shared_memory_bytes > max_shared_memory_bytes_) {
+
-      shared_memory_bytes = 0;
-      use_shared = false;
-    }
     size_t entry_start = 0;
     SparsePageView data(batch.data.DeviceSpan(), batch.offset.DeviceSpan(),
                         num_features);
@@ -608,13 +661,9 @@ class GPUPredictor : public xgboost::Predictor {
     const uint32_t BLOCK_THREADS = 128;
     auto GRID_SIZE = static_cast<uint32_t>(common::DivRoundUp(info.num_row_, BLOCK_THREADS));
 
-    auto shared_memory_bytes =
+    size_t shared_memory_bytes =
-        static_cast<size_t>(sizeof(float) * m->NumColumns() * BLOCK_THREADS);
+        SharedMemoryBytes<BLOCK_THREADS>(info.num_col_, max_shared_memory_bytes);
-    bool use_shared = true;
+    bool use_shared = shared_memory_bytes != 0;
-    if (shared_memory_bytes > max_shared_memory_bytes) {
-      shared_memory_bytes = 0;
-      use_shared = false;
-    }
     size_t entry_start = 0;
 
     dh::LaunchKernel {GRID_SIZE, BLOCK_THREADS, shared_memory_bytes} (
@@ -780,11 +829,65 @@ class GPUPredictor : public xgboost::Predictor {
                << " is not implemented in GPU Predictor.";
   }
 
-  void PredictLeaf(DMatrix*, std::vector<bst_float>*,
+  void PredictLeaf(DMatrix* p_fmat, HostDeviceVector<bst_float>* predictions,
-                   const gbm::GBTreeModel&,
+                   const gbm::GBTreeModel& model,
-                   unsigned) override {
+                   unsigned ntree_limit) override {
-    LOG(FATAL) << "[Internal error]: " << __func__
+    dh::safe_cuda(cudaSetDevice(generic_param_->gpu_id));
-               << " is not implemented in GPU Predictor.";
+    ConfigureDevice(generic_param_->gpu_id);
+
+    const MetaInfo& info = p_fmat->Info();
+    constexpr uint32_t kBlockThreads = 128;
+    size_t shared_memory_bytes =
+        SharedMemoryBytes<kBlockThreads>(info.num_col_, max_shared_memory_bytes_);
+    bool use_shared = shared_memory_bytes != 0;
+    bst_feature_t num_features = info.num_col_;
+    bst_row_t num_rows = info.num_row_;
+    size_t entry_start = 0;
+
+    uint32_t real_ntree_limit = ntree_limit * model.learner_model_param->num_output_group;
+    if (real_ntree_limit == 0 || real_ntree_limit > model.trees.size()) {
+      real_ntree_limit = static_cast<uint32_t>(model.trees.size());
+    }
+    predictions->SetDevice(generic_param_->gpu_id);
+    predictions->Resize(num_rows * real_ntree_limit);
+    model_.Init(model, 0, real_ntree_limit, generic_param_->gpu_id);
+
+    if (p_fmat->PageExists<SparsePage>()) {
+      for (auto const& batch : p_fmat->GetBatches<SparsePage>()) {
+        batch.data.SetDevice(generic_param_->gpu_id);
+        batch.offset.SetDevice(generic_param_->gpu_id);
+        bst_row_t batch_offset = 0;
+        SparsePageView data{batch.data.DeviceSpan(), batch.offset.DeviceSpan(),
+                            model.learner_model_param->num_feature};
+        size_t num_rows = batch.Size();
+        auto grid =
+            static_cast<uint32_t>(common::DivRoundUp(num_rows, kBlockThreads));
+        dh::LaunchKernel {grid, kBlockThreads, shared_memory_bytes} (
+            PredictLeafKernel<SparsePageLoader, SparsePageView>, data,
+            model_.nodes.ConstDeviceSpan(),
+            predictions->DeviceSpan().subspan(batch_offset),
+            model_.tree_segments.ConstDeviceSpan(),
+            model_.tree_beg_, model_.tree_end_, num_features, num_rows,
+            entry_start, use_shared);
+        batch_offset += batch.Size();
+      }
+    } else {
+      for (auto const& batch : p_fmat->GetBatches<EllpackPage>()) {
+        bst_row_t batch_offset = 0;
+        EllpackDeviceAccessor data{batch.Impl()->GetDeviceAccessor(generic_param_->gpu_id)};
+        size_t num_rows = batch.Size();
+        auto grid =
+            static_cast<uint32_t>(common::DivRoundUp(num_rows, kBlockThreads));
+        dh::LaunchKernel {grid, kBlockThreads, shared_memory_bytes} (
+            PredictLeafKernel<EllpackLoader, EllpackDeviceAccessor>, data,
+            model_.nodes.ConstDeviceSpan(),
+            predictions->DeviceSpan().subspan(batch_offset),
+            model_.tree_segments.ConstDeviceSpan(),
+            model_.tree_beg_, model_.tree_end_, num_features, num_rows,
+            entry_start, use_shared);
+        batch_offset += batch.Size();
+      }
+    }
   }
 
   void Configure(const std::vector<std::pair<std::string, std::string>>& cfg) override {
@@ -801,7 +904,7 @@ class GPUPredictor : public xgboost::Predictor {
 
   std::mutex lock_;
   DeviceModel model_;
-  size_t max_shared_memory_bytes_;
+  size_t max_shared_memory_bytes_ { 0 };
 };
 
 XGBOOST_REGISTER_PREDICTOR(GPUPredictor, "gpu_predictor")

tests/cpp/helpers.cc

@@ -348,6 +348,13 @@ RandomDataGenerator::GenerateDMatrix(bool with_label, bool float_label,
       gen.GenerateDense(&out->Info().labels_);
     }
   }
+  if (device_ >= 0) {
+    out->Info().labels_.SetDevice(device_);
+    for (auto const& page : out->GetBatches<SparsePage>()) {
+      page.data.SetDevice(device_);
+      page.offset.SetDevice(device_);
+    }
+  }
   return out;
 }
 

tests/cpp/predictor/test_cpu_predictor.cc

@@ -46,9 +46,10 @@ TEST(CpuPredictor, Basic) {
   }
 
   // Test predict leaf
-  std::vector<float> leaf_out_predictions;
+  HostDeviceVector<float> leaf_out_predictions;
   cpu_predictor->PredictLeaf(dmat.get(), &leaf_out_predictions, model);
-  for (auto v : leaf_out_predictions) {
+  auto const& h_leaf_out_predictions = leaf_out_predictions.ConstHostVector();
+  for (auto v : h_leaf_out_predictions) {
     ASSERT_EQ(v, 0);
   }
 
@@ -112,10 +113,11 @@ TEST(CpuPredictor, ExternalMemory) {
   }
 
   // Test predict leaf
-  std::vector<float> leaf_out_predictions;
+  HostDeviceVector<float> leaf_out_predictions;
   cpu_predictor->PredictLeaf(dmat.get(), &leaf_out_predictions, model);
-  ASSERT_EQ(leaf_out_predictions.size(), dmat->Info().num_row_);
+  auto const& h_leaf_out_predictions = leaf_out_predictions.ConstHostVector();
-  for (const auto& v : leaf_out_predictions) {
+  ASSERT_EQ(h_leaf_out_predictions.size(), dmat->Info().num_row_);
+  for (const auto& v : h_leaf_out_predictions) {
     ASSERT_EQ(v, 0);
   }
 

tests/cpp/predictor/test_gpu_predictor.cu

@@ -190,6 +190,7 @@ TEST(GPUPredictor, ShapStump) {
   EXPECT_EQ(phis[4], 0.0);
   EXPECT_EQ(phis[5], param.base_score);
 }
+
 TEST(GPUPredictor, Shap) {
   LearnerModelParam param;
   param.num_feature = 1;
@@ -224,5 +225,28 @@ TEST(GPUPredictor, Shap) {
 TEST(GPUPredictor, CategoricalPrediction) {
   TestCategoricalPrediction("gpu_predictor");
 }
+
+TEST(GPUPredictor, PredictLeafBasic) {
+  size_t constexpr kRows = 5, kCols = 5;
+  auto dmat = RandomDataGenerator(kRows, kCols, 0).Device(0).GenerateDMatrix();
+  auto lparam = CreateEmptyGenericParam(GPUIDX);
+  std::unique_ptr<Predictor> gpu_predictor =
+      std::unique_ptr<Predictor>(Predictor::Create("gpu_predictor", &lparam));
+  gpu_predictor->Configure({});
+
+  LearnerModelParam param;
+  param.num_feature = kCols;
+  param.base_score = 0.0;
+  param.num_output_group = 1;
+
+  gbm::GBTreeModel model = CreateTestModel(&param);
+
+  HostDeviceVector<float> leaf_out_predictions;
+  gpu_predictor->PredictLeaf(dmat.get(), &leaf_out_predictions, model);
+  auto const& h_leaf_out_predictions = leaf_out_predictions.ConstHostVector();
+  for (auto v : h_leaf_out_predictions) {
+    ASSERT_EQ(v, 0);
+  }
+}
 }  // namespace predictor
 }  // namespace xgboost

tests/python-gpu/test_gpu_prediction.py

@@ -1,4 +1,5 @@
 import sys
+import json
 import unittest
 import pytest
 
@@ -9,6 +10,7 @@ from hypothesis import given, strategies, assume, settings, note
 sys.path.append("tests/python")
 import testing as tm
 from test_predict import run_threaded_predict  # noqa
+from test_predict import run_predict_leaf      # noqa
 
 rng = np.random.RandomState(1994)
 
@@ -18,6 +20,11 @@ shap_parameter_strategy = strategies.fixed_dictionaries({
     'num_parallel_tree': strategies.sampled_from([1, 10]),
 }).filter(lambda x: x['max_depth'] > 0 or x['max_leaves'] > 0)
 
+predict_parameter_strategy = strategies.fixed_dictionaries({
+    'max_depth': strategies.integers(1, 8),
+    'num_parallel_tree': strategies.sampled_from([1, 4]),
+})
+
 
 class TestGPUPredict(unittest.TestCase):
     def test_predict(self):
@@ -223,3 +230,34 @@ class TestGPUPredict(unittest.TestCase):
         assert np.allclose(np.sum(shap, axis=(len(shap.shape) - 1, len(shap.shape) - 2)),
                            margin,
                            1e-3, 1e-3)
+
+    def test_predict_leaf_basic(self):
+        gpu_leaf = run_predict_leaf('gpu_predictor')
+        cpu_leaf = run_predict_leaf('cpu_predictor')
+        np.testing.assert_equal(gpu_leaf, cpu_leaf)
+
+    def run_predict_leaf_booster(self, param, num_rounds, dataset):
+        param = dataset.set_params(param)
+        m = dataset.get_dmat()
+        booster = xgb.train(param, dtrain=dataset.get_dmat(), num_boost_round=num_rounds)
+        booster.set_param({'predictor': 'cpu_predictor'})
+        cpu_leaf = booster.predict(m, pred_leaf=True)
+
+        booster.set_param({'predictor': 'gpu_predictor'})
+        gpu_leaf = booster.predict(m, pred_leaf=True)
+
+        np.testing.assert_equal(cpu_leaf, gpu_leaf)
+
+    @given(predict_parameter_strategy, tm.dataset_strategy)
+    @settings(deadline=None)
+    def test_predict_leaf_gbtree(self, param, dataset):
+        param['booster'] = 'gbtree'
+        param['tree_method'] = 'gpu_hist'
+        self.run_predict_leaf_booster(param, 10, dataset)
+
+    @given(predict_parameter_strategy, tm.dataset_strategy)
+    @settings(deadline=None)
+    def test_predict_leaf_dart(self, param, dataset):
+        param['booster'] = 'dart'
+        param['tree_method'] = 'gpu_hist'
+        self.run_predict_leaf_booster(param, 10, dataset)

tests/python/test_predict.py

@@ -23,6 +23,49 @@ def run_threaded_predict(X, rows, predict_func):
         assert f.result()
 
 
+def run_predict_leaf(predictor):
+    rows = 100
+    cols = 4
+    classes = 5
+    num_parallel_tree = 4
+    num_boost_round = 10
+    rng = np.random.RandomState(1994)
+    X = rng.randn(rows, cols)
+    y = rng.randint(low=0, high=classes, size=rows)
+    m = xgb.DMatrix(X, y)
+    booster = xgb.train(
+        {'num_parallel_tree': num_parallel_tree, 'num_class': classes,
+         'predictor': predictor, 'tree_method': 'hist'}, m,
+        num_boost_round=num_boost_round)
+
+    empty = xgb.DMatrix(np.ones(shape=(0, cols)))
+    empty_leaf = booster.predict(empty, pred_leaf=True)
+    assert empty_leaf.shape[0] == 0
+
+    leaf = booster.predict(m, pred_leaf=True)
+    assert leaf.shape[0] == rows
+    assert leaf.shape[1] == classes * num_parallel_tree * num_boost_round
+
+    for i in range(rows):
+        row = leaf[i, ...]
+        for j in range(num_boost_round):
+            start = classes * num_parallel_tree * j
+            end = classes * num_parallel_tree * (j + 1)
+            layer = row[start: end]
+            for c in range(classes):
+                tree_group = layer[c * num_parallel_tree:
+                                   (c+1) * num_parallel_tree]
+                assert tree_group.shape[0] == num_parallel_tree
+                # no subsampling so tree in same forest should output same
+                # leaf.
+                assert np.all(tree_group == tree_group[0])
+    return leaf
+
+
+def test_predict_leaf():
+    run_predict_leaf('cpu_predictor')
+
+
 class TestInplacePredict(unittest.TestCase):
     '''Tests for running inplace prediction'''
     def test_predict(self):