[performance] Batch DGLGraph in C++ end. (dmlc#2155)

* upd

* upd

* upd

* upd

* upd

* upd

* upd

* upd

* upd

* upd

* upd

* fix

* upd

* upd

* upd

* upd

* fix

* upd

Co-authored-by: VoVAllen <[email protected]>

python/dgl/backend/tensorflow/tensor.py

@@ -145,7 +145,8 @@ def to_backend_ctx(dglctx):
 
 
 def astype(input, ty):
-    return tf.cast(input, dtype=ty)
+    with tf.device(input.device):
+        return tf.cast(input, dtype=ty)
 
 
 def asnumpy(input):

python/dgl/batch.py

@@ -1,12 +1,14 @@
 """Utilities for batching/unbatching graphs."""
 from collections.abc import Mapping
-from collections import defaultdict
 
 from . import backend as F
 from .base import ALL, is_all, DGLError, dgl_warning
+from .heterograph_index import disjoint_union
+from .heterograph import DGLHeteroGraph
 from . import convert
 from . import utils
 
+
 __all__ = ['batch', 'unbatch', 'batch_hetero', 'unbatch_hetero']
 
 def batch(graphs, ndata=ALL, edata=ALL, *, node_attrs=None, edge_attrs=None):
@@ -156,61 +158,44 @@ def batch(graphs, ndata=ALL, edata=ALL, *, node_attrs=None, edge_attrs=None):
         dgl_warning('Arguments edge_attrs has been deprecated. Please use'
                     ' edata instead.')
         edata = edge_attrs
-    if not (is_all(ndata) or isinstance(ndata, list)):
+    if not (is_all(ndata) or isinstance(ndata, list) or ndata is None):
         raise DGLError('Invalid argument ndata: must be a string list but got {}.'.format(
             type(ndata)))
-    if not (is_all(edata) or isinstance(edata, list)):
+    if not (is_all(edata) or isinstance(edata, list) or edata is None):
         raise DGLError('Invalid argument edata: must be a string list but got {}.'.format(
             type(edata)))
     if any(g.is_block for g in graphs):
         raise DGLError("Batching a block is not supported.")
 
-    utils.check_all_same_device(graphs, 'graphs')
-    utils.check_all_same_idtype(graphs, 'graphs')
-    relations = graphs[0].canonical_etypes
-    ntypes = graphs[0].ntypes
-    idtype = graphs[0].idtype
-    device = graphs[0].device
-
-    # Batch graph structure for each relation graph
-    edge_dict = defaultdict(list)
-    num_nodes_dict = defaultdict(int)
-    for g in graphs:
-        for rel in relations:
-            srctype, etype, dsttype = rel
-            u, v = g.edges(order='eid', etype=rel)
-            src = u + num_nodes_dict[srctype]
-            dst = v + num_nodes_dict[dsttype]
-            edge_dict[rel].append((src, dst))
-        for ntype in ntypes:
-            num_nodes_dict[ntype] += g.number_of_nodes(ntype)
-    for rel in relations:
-        src, dst = zip(*edge_dict[rel])
-        edge_dict[rel] = (F.cat(src, 0), F.cat(dst, 0))
-    retg = convert.heterograph(edge_dict, num_nodes_dict, idtype=idtype, device=device)
+    relations = list(sorted(graphs[0].canonical_etypes))
+    ntypes = list(sorted(graphs[0].ntypes))
+    etypes = [etype for _, etype, _ in relations]
+
+    gidx = disjoint_union(graphs[0]._graph.metagraph, [g._graph for g in graphs])
+    retg = DGLHeteroGraph(gidx, ntypes, etypes)
 
     # Compute batch num nodes
     bnn = {}
-    for ntype in graphs[0].ntypes:
+    for ntype in ntypes:
         bnn[ntype] = F.cat([g.batch_num_nodes(ntype) for g in graphs], 0)
     retg.set_batch_num_nodes(bnn)
 
     # Compute batch num edges
     bne = {}
-    for etype in graphs[0].canonical_etypes:
+    for etype in relations:
         bne[etype] = F.cat([g.batch_num_edges(etype) for g in graphs], 0)
     retg.set_batch_num_edges(bne)
 
     # Batch node feature
     if ndata is not None:
-        for ntype in graphs[0].ntypes:
+        for ntype in ntypes:
             feat_dicts = [g.nodes[ntype].data for g in graphs if g.number_of_nodes(ntype) > 0]
             ret_feat = _batch_feat_dicts(feat_dicts, ndata, 'nodes["{}"].data'.format(ntype))
             retg.nodes[ntype].data.update(ret_feat)
 
     # Batch edge feature
     if edata is not None:
-        for etype in graphs[0].canonical_etypes:
+        for etype in relations:
             feat_dicts = [g.edges[etype].data for g in graphs if g.number_of_edges(etype) > 0]
             ret_feat = _batch_feat_dicts(feat_dicts, edata, 'edges[{}].data'.format(etype))
             retg.edges[etype].data.update(ret_feat)

src/array/array.cc

@@ -202,7 +202,6 @@ IdArray Relabel_(const std::vector<IdArray>& arrays) {
 }
 
 NDArray Concat(const std::vector<IdArray>& arrays) {
-  CHECK(arrays.size() > 1) << "Number of arrays should larger than 1";
   IdArray ret;
 
   int64_t len = 0, offset = 0;

src/array/cuda/utils.h

@@ -23,7 +23,9 @@ namespace cuda {
  * and is also power of two.
  */
 inline int FindNumThreads(int dim, int max_nthrs = CUDA_MAX_NUM_THREADS) {
-  CHECK_NE(dim, 0);
+  CHECK_GE(dim, 0);
+  if (dim == 0)
+    return 1;
   int ret = max_nthrs;
   while (ret > dim) {
     ret = ret >> 1;

src/array/union_partition.cc

@@ -10,8 +10,6 @@ namespace dgl {
 namespace aten {
 ///////////////////////// COO Based Operations/////////////////////////
 COOMatrix DisjointUnionCoo(const std::vector<COOMatrix>& coos) {
-  CHECK(coos.size() > 1) <<
-    "The length of input COOMatrix vector should be larger than 1";
   uint64_t src_offset = 0, dst_offset = 0;
   int64_t edge_data_offset = 0;
   bool has_data = false;
@@ -114,8 +112,6 @@ std::vector<COOMatrix> DisjointPartitionCooBySizes(
 
 ///////////////////////// CSR Based Operations/////////////////////////
 CSRMatrix DisjointUnionCsr(const std::vector<CSRMatrix>& csrs) {
-  CHECK(csrs.size() > 1) <<
-    "The length of input CSRMatrix vector should be larger than 1";
   uint64_t src_offset = 0, dst_offset = 0;
   int64_t indices_offset = 0;
   bool has_data = false;

src/graph/heterograph.cc

@@ -124,6 +124,10 @@ void HeteroGraphSanityCheck(GraphPtr meta_graph, const std::vector<HeteroGraphPt
   for (const auto &rg : rel_graphs) {
     CHECK_EQ(rg->NumEdgeTypes(), 1) << "Each relation graph must have only one edge type.";
   }
+  auto ctx = rel_graphs[0]->Context();
+  for (const auto &rg : rel_graphs) {
+    CHECK_EQ(rg->Context(), ctx) << "Each relation graph must have the same context.";
+  }
 }
 
 std::vector<int64_t>

src/graph/heterograph_capi.cc

@@ -559,28 +559,6 @@ DGL_REGISTER_GLOBAL("heterograph_index._CAPI_DGLHeteroDisjointPartitionBySizes_v
     *rv = ret_list;
 });
 
-DGL_REGISTER_GLOBAL("heterograph_index._CAPI_DGLHeteroDisjointUnion")
-.set_body([] (DGLArgs args, DGLRetValue* rv) {
-    GraphRef meta_graph = args[0];
-    List<HeteroGraphRef> component_graphs = args[1];
-    CHECK(component_graphs.size() > 0)
-      << "Expect graph list has at least one graph";
-    std::vector<HeteroGraphPtr> component_ptrs;
-    component_ptrs.reserve(component_graphs.size());
-    const int64_t bits = component_graphs[0]->NumBits();
-    for (const auto& component : component_graphs) {
-      component_ptrs.push_back(component.sptr());
-      CHECK_EQ(component->NumBits(), bits)
-        << "Expect graphs to batch have the same index dtype(int" << bits
-        << "), but got int" << component->NumBits();
-    }
-    ATEN_ID_BITS_SWITCH(bits, IdType, {
-      auto hgptr =
-        DisjointUnionHeteroGraph<IdType>(meta_graph.sptr(), component_ptrs);
-      *rv = HeteroGraphRef(hgptr);
-    });
-});
-
 DGL_REGISTER_GLOBAL("heterograph_index._CAPI_DGLHeteroDisjointPartitionBySizes")
 .set_body([] (DGLArgs args, DGLRetValue* rv) {
     HeteroGraphRef hg = args[0];

src/graph/transform/union_partition.cc

@@ -91,33 +91,34 @@ HeteroGraphPtr DisjointUnionHeteroGraph2(
   std::vector<HeteroGraphPtr> rel_graphs(meta_graph->NumEdges());
   std::vector<int64_t> num_nodes_per_type(meta_graph->NumVertices(), 0);
 
+  // Loop over all ntypes
+  for (dgl_type_t vtype = 0; vtype < meta_graph->NumVertices(); ++vtype) {
+    uint64_t offset = 0;
+    for (const auto &cg : component_graphs)
+      offset += cg->NumVertices(vtype);
+    num_nodes_per_type[vtype] = offset;
+  }
+
   // Loop over all canonical etypes
   for (dgl_type_t etype = 0; etype < meta_graph->NumEdges(); ++etype) {
     auto pair = meta_graph->FindEdge(etype);
     const dgl_type_t src_vtype = pair.first;
     const dgl_type_t dst_vtype = pair.second;
-    uint64_t src_offset = 0, dst_offset = 0;
     HeteroGraphPtr rgptr = nullptr;
 
     const dgl_format_code_t code =\
       component_graphs[0]->GetRelationGraph(etype)->GetAllowedFormats();
     // do some preprocess
-    for (size_t i = 0; i < component_graphs.size(); ++i) {
-      const auto& cg = component_graphs[i];
+    for (const auto &cg : component_graphs) {
       const dgl_format_code_t cur_code = cg->GetRelationGraph(etype)->GetAllowedFormats();
       if (cur_code != code)
         LOG(FATAL) << "All components should have the same formats";
-
-      // Update offsets
-      src_offset += cg->NumVertices(src_vtype);
-      dst_offset += cg->NumVertices(dst_vtype);
     }
 
     // prefer COO
     if (FORMAT_HAS_COO(code)) {
       std::vector<aten::COOMatrix> coos;
-      for (size_t i = 0; i < component_graphs.size(); ++i) {
-        const auto& cg = component_graphs[i];
+      for (const auto &cg : component_graphs) {
         aten::COOMatrix coo = cg->GetCOOMatrix(etype);
         coos.push_back(coo);
       }
@@ -128,8 +129,7 @@ HeteroGraphPtr DisjointUnionHeteroGraph2(
           (src_vtype == dst_vtype) ? 1 : 2, res, code);
     } else if (FORMAT_HAS_CSR(code)) {
       std::vector<aten::CSRMatrix> csrs;
-      for (size_t i = 0; i < component_graphs.size(); ++i) {
-        const auto& cg = component_graphs[i];
+      for (const auto &cg : component_graphs) {
         aten::CSRMatrix csr = cg->GetCSRMatrix(etype);
         csrs.push_back(csr);
       }
@@ -141,8 +141,7 @@ HeteroGraphPtr DisjointUnionHeteroGraph2(
     } else if (FORMAT_HAS_CSC(code)) {
       // CSR and CSC have the same storage format, i.e. CSRMatrix
       std::vector<aten::CSRMatrix> cscs;
-      for (size_t i = 0; i < component_graphs.size(); ++i) {
-        const auto& cg = component_graphs[i];
+      for (const auto &cg : component_graphs) {
         aten::CSRMatrix csc = cg->GetCSCMatrix(etype);
         cscs.push_back(csc);
       }
@@ -152,8 +151,6 @@ HeteroGraphPtr DisjointUnionHeteroGraph2(
         (src_vtype == dst_vtype) ? 1 : 2, res, code);
     }
     rel_graphs[etype] = rgptr;
-    num_nodes_per_type[src_vtype] = src_offset;
-    num_nodes_per_type[dst_vtype] = dst_offset;
   }
 
   return CreateHeteroGraph(meta_graph, rel_graphs, std::move(num_nodes_per_type));
@@ -272,56 +269,6 @@ std::vector<HeteroGraphPtr> DisjointPartitionHeteroBySizes2(
   return rst;
 }
 
-template <class IdType>
-HeteroGraphPtr DisjointUnionHeteroGraph(
-    GraphPtr meta_graph, const std::vector<HeteroGraphPtr>& component_graphs) {
-  CHECK_GT(component_graphs.size(), 0) << "Input graph list is empty";
-  std::vector<HeteroGraphPtr> rel_graphs(meta_graph->NumEdges());
-  std::vector<int64_t> num_nodes_per_type(meta_graph->NumVertices(), 0);
-
-  // Loop over all canonical etypes
-  for (dgl_type_t etype = 0; etype < meta_graph->NumEdges(); ++etype) {
-    auto pair = meta_graph->FindEdge(etype);
-    const dgl_type_t src_vtype = pair.first;
-    const dgl_type_t dst_vtype = pair.second;
-    IdType src_offset = 0, dst_offset = 0;
-    std::vector<IdType> result_src, result_dst;
-
-    // Loop over all graphs
-    for (size_t i = 0; i < component_graphs.size(); ++i) {
-      const auto& cg = component_graphs[i];
-      EdgeArray edges = cg->Edges(etype);
-      size_t num_edges = cg->NumEdges(etype);
-      const IdType* edges_src_data = static_cast<const IdType*>(edges.src->data);
-      const IdType* edges_dst_data = static_cast<const IdType*>(edges.dst->data);
-
-      // Loop over all edges
-      for (size_t j = 0; j < num_edges; ++j) {
-        // TODO(mufei): Should use array operations to implement this.
-        result_src.push_back(edges_src_data[j] + src_offset);
-        result_dst.push_back(edges_dst_data[j] + dst_offset);
-      }
-      // Update offsets
-      src_offset += cg->NumVertices(src_vtype);
-      dst_offset += cg->NumVertices(dst_vtype);
-    }
-    HeteroGraphPtr rgptr = UnitGraph::CreateFromCOO(
-        (src_vtype == dst_vtype) ? 1 : 2, src_offset, dst_offset,
-        aten::VecToIdArray(result_src, sizeof(IdType) * 8),
-        aten::VecToIdArray(result_dst, sizeof(IdType) * 8));
-    rel_graphs[etype] = rgptr;
-    num_nodes_per_type[src_vtype] = src_offset;
-    num_nodes_per_type[dst_vtype] = dst_offset;
-  }
-  return CreateHeteroGraph(meta_graph, rel_graphs, std::move(num_nodes_per_type));
-}
-
-template HeteroGraphPtr DisjointUnionHeteroGraph<int32_t>(
-    GraphPtr meta_graph, const std::vector<HeteroGraphPtr>& component_graphs);
-
-template HeteroGraphPtr DisjointUnionHeteroGraph<int64_t>(
-    GraphPtr meta_graph, const std::vector<HeteroGraphPtr>& component_graphs);
-
 template <class IdType>
 std::vector<HeteroGraphPtr> DisjointPartitionHeteroBySizes(
     GraphPtr meta_graph, HeteroGraphPtr batched_graph, IdArray vertex_sizes, IdArray edge_sizes) {

src/runtime/ndarray.cc

@@ -204,9 +204,10 @@ NDArray NDArray::Empty(std::vector<int64_t> shape,
   // setup memory content
   size_t size = GetDataSize(ret.data_->dl_tensor);
   size_t alignment = GetDataAlignment(ret.data_->dl_tensor);
-  ret.data_->dl_tensor.data =
-      DeviceAPI::Get(ret->ctx)->AllocDataSpace(
-          ret->ctx, size, alignment, ret->dtype);
+  if (size > 0)
+    ret.data_->dl_tensor.data =
+        DeviceAPI::Get(ret->ctx)->AllocDataSpace(
+            ret->ctx, size, alignment, ret->dtype);
   return ret;
 }
 

tests/compute/test_batched_heterograph.py

@@ -325,6 +325,6 @@ def test_unbatch2(idtype):
     #test_topology('int32')
     #test_batching_batched('int32')
     #test_batched_features('int32')
-    #test_empty_relation('int32')
+    # test_empty_relation('int64')
     #test_to_device('int32')
     pass