[SYCL] Add dask support for distributed (#10812)

plugin/sycl/common/hist_util.cc

@@ -31,6 +31,33 @@ template void InitHist(::sycl::queue qu,
                        GHistRow<double, MemoryType::on_device>* hist,
                        size_t size, ::sycl::event* event);
 
+/*!
+ * \brief Copy histogram from src to dst
+ */
+template<typename GradientSumT>
+void CopyHist(::sycl::queue qu,
+              GHistRow<GradientSumT, MemoryType::on_device>* dst,
+              const GHistRow<GradientSumT, MemoryType::on_device>& src,
+              size_t size) {
+  GradientSumT* pdst = reinterpret_cast<GradientSumT*>(dst->Data());
+  const GradientSumT* psrc = reinterpret_cast<const GradientSumT*>(src.DataConst());
+
+  qu.submit([&](::sycl::handler& cgh) {
+    cgh.parallel_for<>(::sycl::range<1>(2 * size), [=](::sycl::item<1> pid) {
+      const size_t i = pid.get_id(0);
+      pdst[i] = psrc[i];
+    });
+  }).wait();
+}
+template void CopyHist(::sycl::queue qu,
+                       GHistRow<float, MemoryType::on_device>* dst,
+                       const GHistRow<float, MemoryType::on_device>& src,
+                       size_t size);
+template void CopyHist(::sycl::queue qu,
+                       GHistRow<double, MemoryType::on_device>* dst,
+                       const GHistRow<double, MemoryType::on_device>& src,
+                       size_t size);
+
 /*!
  * \brief Compute Subtraction: dst = src1 - src2
  */

plugin/sycl/common/hist_util.h

@@ -36,6 +36,15 @@ void InitHist(::sycl::queue qu,
               GHistRow<GradientSumT, MemoryType::on_device>* hist,
               size_t size, ::sycl::event* event);
 
+/*!
+ * \brief Copy histogram from src to dst
+ */
+template<typename GradientSumT>
+void CopyHist(::sycl::queue qu,
+              GHistRow<GradientSumT, MemoryType::on_device>* dst,
+              const GHistRow<GradientSumT, MemoryType::on_device>& src,
+              size_t size);
+
 /*!
  * \brief Compute subtraction: dst = src1 - src2
  */

plugin/sycl/tree/hist_row_adder.h

@@ -39,6 +39,42 @@ class BatchHistRowsAdder: public HistRowsAdder<GradientSumT> {
   }
 };
 
+
+template <typename GradientSumT>
+class DistributedHistRowsAdder: public HistRowsAdder<GradientSumT> {
+ public:
+  void AddHistRows(HistUpdater<GradientSumT>* builder,
+                   std::vector<int>* sync_ids, RegTree *p_tree) override {
+    builder->builder_monitor_.Start("AddHistRows");
+    const size_t explicit_size = builder->nodes_for_explicit_hist_build_.size();
+    const size_t subtaction_size = builder->nodes_for_subtraction_trick_.size();
+    std::vector<int> merged_node_ids(explicit_size + subtaction_size);
+    for (size_t i = 0; i < explicit_size; ++i) {
+      merged_node_ids[i] = builder->nodes_for_explicit_hist_build_[i].nid;
+    }
+    for (size_t i = 0; i < subtaction_size; ++i) {
+      merged_node_ids[explicit_size + i] =
+      builder->nodes_for_subtraction_trick_[i].nid;
+    }
+    std::sort(merged_node_ids.begin(), merged_node_ids.end());
+    sync_ids->clear();
+    for (auto const& nid : merged_node_ids) {
+      if ((*p_tree)[nid].IsLeftChild()) {
+        builder->hist_.AddHistRow(nid);
+        builder->hist_local_worker_.AddHistRow(nid);
+        sync_ids->push_back(nid);
+      }
+    }
+    for (auto const& nid : merged_node_ids) {
+      if (!((*p_tree)[nid].IsLeftChild())) {
+        builder->hist_.AddHistRow(nid);
+        builder->hist_local_worker_.AddHistRow(nid);
+      }
+    }
+    builder->builder_monitor_.Stop("AddHistRows");
+  }
+};
+
 }  // namespace tree
 }  // namespace sycl
 }  // namespace xgboost

plugin/sycl/tree/hist_synchronizer.h

@@ -61,6 +61,68 @@ class BatchHistSynchronizer: public HistSynchronizer<GradientSumT> {
   std::vector<::sycl::event> hist_sync_events_;
 };
 
+template <typename GradientSumT>
+class DistributedHistSynchronizer: public HistSynchronizer<GradientSumT> {
+ public:
+  void SyncHistograms(HistUpdater<GradientSumT>* builder,
+                      const std::vector<int>& sync_ids,
+                      RegTree *p_tree) override {
+    builder->builder_monitor_.Start("SyncHistograms");
+    const size_t nbins = builder->hist_builder_.GetNumBins();
+    for (int node = 0; node < builder->nodes_for_explicit_hist_build_.size(); node++) {
+      const auto entry = builder->nodes_for_explicit_hist_build_[node];
+      auto& this_hist = builder->hist_[entry.nid];
+      // // Store posible parent node
+      auto& this_local = builder->hist_local_worker_[entry.nid];
+      common::CopyHist(builder->qu_, &this_local, this_hist, nbins);
+
+      if (!(*p_tree)[entry.nid].IsRoot()) {
+        const size_t parent_id = (*p_tree)[entry.nid].Parent();
+        auto sibling_nid = entry.GetSiblingId(p_tree, parent_id);
+        auto& parent_hist = builder->hist_local_worker_[parent_id];
+
+        auto& sibling_hist = builder->hist_[sibling_nid];
+        common::SubtractionHist(builder->qu_, &sibling_hist, parent_hist,
+                                this_hist, nbins, ::sycl::event());
+        builder->qu_.wait_and_throw();
+        // Store posible parent node
+        auto& sibling_local = builder->hist_local_worker_[sibling_nid];
+        common::CopyHist(builder->qu_, &sibling_local, sibling_hist, nbins);
+      }
+    }
+    builder->ReduceHists(sync_ids, nbins);
+
+    ParallelSubtractionHist(builder, builder->nodes_for_explicit_hist_build_, p_tree);
+    ParallelSubtractionHist(builder, builder->nodes_for_subtraction_trick_, p_tree);
+
+    builder->builder_monitor_.Stop("SyncHistograms");
+  }
+
+  void ParallelSubtractionHist(HistUpdater<GradientSumT>* builder,
+                               const std::vector<ExpandEntry>& nodes,
+                               const RegTree * p_tree) {
+    const size_t nbins = builder->hist_builder_.GetNumBins();
+    for (int node = 0; node < nodes.size(); node++) {
+      const auto entry = nodes[node];
+      if (!((*p_tree)[entry.nid].IsLeftChild())) {
+        auto& this_hist = builder->hist_[entry.nid];
+
+        if (!(*p_tree)[entry.nid].IsRoot()) {
+          const size_t parent_id = (*p_tree)[entry.nid].Parent();
+          auto& parent_hist = builder->hist_[parent_id];
+          auto& sibling_hist = builder->hist_[entry.GetSiblingId(p_tree, parent_id)];
+          common::SubtractionHist(builder->qu_, &this_hist, parent_hist,
+                                  sibling_hist, nbins, ::sycl::event());
+          builder->qu_.wait_and_throw();
+        }
+      }
+    }
+  }
+
+ private:
+  std::vector<::sycl::event> hist_sync_events_;
+};
+
 }  // namespace tree
 }  // namespace sycl
 }  // namespace xgboost

plugin/sycl/tree/hist_updater.cc

@@ -22,6 +22,30 @@ using ::sycl::ext::oneapi::plus;
 using ::sycl::ext::oneapi::minimum;
 using ::sycl::ext::oneapi::maximum;
 
+template <typename GradientSumT>
+void HistUpdater<GradientSumT>::ReduceHists(const std::vector<int>& sync_ids,
+                                            size_t nbins) {
+  if (reduce_buffer_.size() < sync_ids.size() * nbins) {
+    reduce_buffer_.resize(sync_ids.size() * nbins);
+  }
+  for (size_t i = 0; i < sync_ids.size(); i++) {
+    auto& this_hist = hist_[sync_ids[i]];
+    const GradientPairT* psrc = reinterpret_cast<const GradientPairT*>(this_hist.DataConst());
+    qu_.memcpy(reduce_buffer_.data() + i * nbins, psrc, nbins*sizeof(GradientPairT)).wait();
+  }
+
+  auto buffer_vec = linalg::MakeVec(reinterpret_cast<GradientSumT*>(reduce_buffer_.data()),
+                                    2 * nbins * sync_ids.size());
+  auto rc = collective::Allreduce(ctx_, buffer_vec, collective::Op::kSum);
+  SafeColl(rc);
+
+  for (size_t i = 0; i < sync_ids.size(); i++) {
+    auto& this_hist = hist_[sync_ids[i]];
+    GradientPairT* psrc = reinterpret_cast<GradientPairT*>(this_hist.Data());
+    qu_.memcpy(psrc, reduce_buffer_.data() + i * nbins, nbins*sizeof(GradientPairT)).wait();
+  }
+}
+
 template <typename GradientSumT>
 void HistUpdater<GradientSumT>::SetHistSynchronizer(
     HistSynchronizer<GradientSumT> *sync) {
@@ -492,6 +516,7 @@ void HistUpdater<GradientSumT>::InitData(
     // initialize histogram collection
     uint32_t nbins = gmat.cut.Ptrs().back();
     hist_.Init(qu_, nbins);
+    hist_local_worker_.Init(qu_, nbins);
 
     hist_buffer_.Init(qu_, nbins);
     size_t buffer_size = kBufferSize;

plugin/sycl/tree/hist_updater.h

@@ -87,7 +87,10 @@ class HistUpdater {
 
  protected:
   friend class BatchHistSynchronizer<GradientSumT>;
+  friend class DistributedHistSynchronizer<GradientSumT>;
+
   friend class BatchHistRowsAdder<GradientSumT>;
+  friend class DistributedHistRowsAdder<GradientSumT>;
 
   struct SplitQuery {
     bst_node_t nid;
@@ -183,6 +186,8 @@ class HistUpdater {
                            RegTree* p_tree,
                            const USMVector<GradientPair, MemoryType::on_device>& gpair);
 
+  void ReduceHists(const std::vector<int>& sync_ids, size_t nbins);
+
   inline static bool LossGuide(ExpandEntry lhs, ExpandEntry rhs) {
     if (lhs.GetLossChange() == rhs.GetLossChange()) {
       return lhs.GetNodeId() > rhs.GetNodeId();  // favor small timestamp
@@ -230,6 +235,8 @@ class HistUpdater {
   common::ParallelGHistBuilder<GradientSumT> hist_buffer_;
   /*! \brief culmulative histogram of gradients. */
   common::HistCollection<GradientSumT, MemoryType::on_device> hist_;
+  /*! \brief culmulative local parent histogram of gradients. */
+  common::HistCollection<GradientSumT, MemoryType::on_device> hist_local_worker_;
 
   /*! \brief TreeNode Data: statistics for each constructed node */
   std::vector<NodeEntry<GradientSumT>> snode_host_;
@@ -258,6 +265,8 @@ class HistUpdater {
   USMVector<bst_float, MemoryType::on_device> out_preds_buf_;
   bst_float* out_pred_ptr = nullptr;
 
+  std::vector<GradientPairT> reduce_buffer_;
+
   ::sycl::queue qu_;
 };
 

plugin/sycl/tree/updater_quantile_hist.cc

@@ -51,7 +51,8 @@ void QuantileHistMaker::SetPimpl(std::unique_ptr<HistUpdater<GradientSumT>>* pim
                 param_,
                 int_constraint_, dmat));
   if (collective::IsDistributed()) {
-    LOG(FATAL) << "Distributed mode is not yet upstreamed for sycl";
+    (*pimpl)->SetHistSynchronizer(new DistributedHistSynchronizer<GradientSumT>());
+    (*pimpl)->SetHistRowsAdder(new DistributedHistRowsAdder<GradientSumT>());
   } else {
     (*pimpl)->SetHistSynchronizer(new BatchHistSynchronizer<GradientSumT>());
     (*pimpl)->SetHistRowsAdder(new BatchHistRowsAdder<GradientSumT>());

python-package/xgboost/core.py

@@ -306,6 +306,7 @@ def _check_distributed_params(kwargs: Dict[str, Any]) -> None:
         raise TypeError(msg)
 
     if device and device.find(":") != -1:
+        if device != "sycl:gpu":
             raise ValueError(
                 "Distributed training doesn't support selecting device ordinal as GPUs are"
                 " managed by the distributed frameworks. use `device=cuda` or `device=gpu`"

tests/ci_build/conda_env/linux_sycl_test.yml

@@ -17,5 +17,6 @@ dependencies:
 - pytest
 - pytest-timeout
 - pytest-cov
+- dask
 - dpcpp_linux-64
 - onedpl-devel

tests/python-sycl/test_sycl_simple_dask.py

@@ -0,0 +1,42 @@
+from xgboost import dask as dxgb
+from xgboost import testing as tm
+
+from hypothesis import given, strategies, assume, settings, note
+
+import dask.array as da
+import dask.distributed
+
+
+def train_result(client, param, dtrain, num_rounds):
+    result = dxgb.train(
+        client,
+        param,
+        dtrain,
+        num_rounds,
+        verbose_eval=False,
+        evals=[(dtrain, "train")],
+    )
+    return result
+
+
+class TestSYCLDask:
+    # The simplest test verify only one node training.
+    def test_simple(self):
+        cluster = dask.distributed.LocalCluster(n_workers=1)
+        client = dask.distributed.Client(cluster)
+
+        param = {}
+        param["tree_method"] = "hist"
+        param["device"] = "sycl"
+        param["verbosity"] = 0
+        param["objective"] = "reg:squarederror"
+
+        # X and y must be Dask dataframes or arrays
+        num_obs = 1e4
+        num_features = 20
+        X = da.random.random(size=(num_obs, num_features), chunks=(1000, num_features))
+        y = da.random.random(size=(num_obs, 1), chunks=(1000, 1))
+        dtrain = dxgb.DaskDMatrix(client, X, y)
+
+        result = train_result(client, param, dtrain, 10)
+        assert tm.non_increasing(result["history"]["train"]["rmse"])