Optimize move semantics of NodeEntry

apache/tvm#2576
Making copies of shared_ptr is very expensive, 100x more than moving. This PR reduces lock contention by using move semantics in NNVM nodes

.gitmodules

@@ -25,7 +25,7 @@
 	url = https://github.com/dmlc/cub
 [submodule "3rdparty/tvm"]
 	path = 3rdparty/tvm
-	url = https://github.com/dmlc/tvm
+	url = https://github.com/larroy/tvm
 [submodule "3rdparty/onnx-tensorrt"]
 	path = 3rdparty/onnx-tensorrt
 	url = https://github.com/onnx/onnx-tensorrt.git

src/c_api/c_api_function.cc

@@ -56,7 +56,7 @@ std::vector<nnvm::NodeEntry> Gradient(
 
   std::vector<nnvm::NodeEntry> ret;
   for (uint32_t i = 0; i < g->num_outputs(); ++i) {
-    ret.emplace_back(nnvm::NodeEntry{g, i, 0});
+    ret.emplace_back(std::move(g), i, 0);
   }
 
   return ret;

src/executor/graph_executor.cc

@@ -146,7 +146,7 @@ nnvm::NodeEntry AggregateGradient(std::vector<nnvm::NodeEntry>&& v) {
     ng->attrs.op = Op::Get("_zeros_without_dtype");
     ng->attrs.name = "zeros_without_dtype";
     ng->attrs.op->attr_parser(&(ng->attrs));
-    return nnvm::NodeEntry{ng, 0, 0};
+    return nnvm::NodeEntry(std::move(ng), 0, 0);
   }
 
   // remove zero in the sum. at least keep 1.
@@ -167,7 +167,7 @@ nnvm::NodeEntry AggregateGradient(std::vector<nnvm::NodeEntry>&& v) {
       sum_node->attrs.dict["num_args"] = std::to_string(v.size());
       sum_node->attrs.op->attr_parser(&(sum_node->attrs));
       sum_node->inputs = std::move(v);
-      return nnvm::NodeEntry{sum_node, 0, 0};
+      return nnvm::NodeEntry(std::move(sum_node), 0, 0);
     } else {
       // use a stream line of plus instead
       nnvm::NodeEntry ret = v[0];
@@ -197,7 +197,7 @@ nnvm::NodeEntry AggregateGradient(std::vector<nnvm::NodeEntry>&& v) {
         x->attrs.op = ewise_plus_op;
         x->attrs.name = os.str();
         x->inputs = {ret, v[i]};
-        ret = nnvm::NodeEntry{x, 0, 0};
+        ret = nnvm::NodeEntry(std::move(x), 0, 0);
       }
       // identity node is used to avoid exposure of dummy plus node
       // when its output get assigned to another space.
@@ -246,7 +246,7 @@ nnvm::Graph GraphExecutor::InitFullGraph(nnvm::Symbol symbol,
   }
   if (!need_grad_) return g;
   for (size_t i = 0; i < g.outputs.size(); ++i) {
-    NodeEntry ngrad{nnvm::Node::Create(), 0, 0};
+    NodeEntry ngrad;
     head_grad_entry_.emplace_back(AttrHint(ngrad, g.outputs[i]));
     head_grad_map_[ngrad.node.get()] = i;
   }

src/imperative/cached_op.cc

@@ -160,7 +160,7 @@ CachedOp::CachedOp(
   {
     ograd_entries_.reserve(fwd_graph_.outputs.size());
     for (size_t i = 0; i < fwd_graph_.outputs.size(); ++i) {
-      ograd_entries_.emplace_back(NodeEntry{Node::Create(), 0, 0});
+      ograd_entries_.emplace_back();
     }
 
     std::vector<NodeEntry> xs;
@@ -169,7 +169,7 @@ CachedOp::CachedOp(
       auto nid = idx.input_nodes()[i];
       if (idx.mutable_input_nodes().count(nid)) continue;
       fwd_input_to_grad_output_[i] = xs.size();
-      xs.emplace_back(NodeEntry{idx[nid].weak_ref.lock(), 0, 0});
+      xs.emplace_back(idx[nid].weak_ref.lock(), 0, 0);
     }
 
     CHECK_GT(xs.size(), 0)

src/imperative/imperative.cc

@@ -303,7 +303,7 @@ std::vector<NDArray*> Imperative::Backward(
   std::vector<NodeEntry> ograd_entries;
   ograd_entries.reserve(ograds.size());
   for (size_t i = 0; i < outputs.size(); ++i) {
-    ograd_entries.emplace_back(NodeEntry{Node::Create(), 0, 0});
+    ograd_entries.emplace_back();
     AGInfo& info = AGInfo::Create(ograd_entries.back().node);
     info.ctx = outputs[i]->ctx();
     if (ograds[i] != nullptr) {

src/nnvm/legacy_op_util.cc

@@ -321,17 +321,18 @@ inline std::vector<NodeEntry> OpPropGradient(
     const NodePtr& ptr,
     const std::vector<NodeEntry>& out_grads) {
   auto& prop = nnvm::get<ParsedOpProp>(ptr->attrs.parsed);
-  std::vector<NodeEntry> out_data(prop.outputs.size());
-  for (uint32_t i = 0; i < out_data.size(); ++i) {
-    out_data[i] = NodeEntry{ptr, i, 0};
+  std::vector<NodeEntry> out_data;
+  out_data.reserve(prop.outputs.size());
+  for (size_t i = 0; i < out_data.size(); ++i) {
+    out_data.emplace_back(ptr, i, 0);
   }
   std::vector<NodeEntry> in_data(
       ptr->inputs.begin(), ptr->inputs.begin() + prop.arguments.size());
   std::vector<NodeEntry> ograd(
       out_grads.begin(), out_grads.begin() + prop.ptr->NumVisibleOutputs());
   auto inputs = prop.ptr->BackwardInputs(ograd, in_data, out_data);
   // add all the auxiliary data
-  for (uint32_t i = 0; i < prop.aux_states.size(); ++i) {
+  for (size_t i = 0; i < prop.aux_states.size(); ++i) {
     inputs.emplace_back(ptr->inputs[i + prop.arguments.size()]);
   }
   NodePtr gnode = Node::Create();
@@ -340,17 +341,15 @@ inline std::vector<NodeEntry> OpPropGradient(
   gnode->attrs = ptr->attrs;
   gnode->attrs.op = back_op;
   gnode->attrs.name = ptr->attrs.name + "_backward";
-  std::vector<NodeEntry> in_grad(prop.arguments.size());
-  for (uint32_t i = 0; i < prop.arguments.size(); ++i) {
-    in_grad[i] = NodeEntry{gnode, i, 0};
+  std::vector<NodeEntry> in_grad;
+  in_grad.reserve(prop.arguments.size() + prop.aux_states.size());
+  for (size_t i = 0; i < prop.arguments.size(); ++i) {
+    in_grad.emplace_back(gnode, i, 0);
   }
   // attach no gradient node to forbid gradient on aux_state
   if (prop.aux_states.size() != 0) {
-    NodePtr ng = Node::Create();
-    ng->attrs.op = Op::Get("_NoGradient");
-    ng->attrs.name = "NoGradient";
-    for (uint32_t i = 0; i < prop.aux_states.size(); ++i) {
-      in_grad.emplace_back(NodeEntry{ng, 0, 0});
+    for (size_t i = 0; i < prop.aux_states.size(); ++i) {
+      in_grad.emplace_back(Node::Create(Op::Get("_NoGradient"), "NoGradient"), 0, 0);
     }
   }
   return in_grad;

src/operator/custom/custom.cc

@@ -225,7 +225,7 @@ std::vector<nnvm::NodeEntry> Gradient(
     size_t i = static_cast<size_t>(t);
     if (i >= params.num_outs + params.num_args) {
       uint32_t idx = static_cast<uint32_t>(i-params.num_outs-params.num_args);
-      g->inputs.push_back(nnvm::NodeEntry{n, idx, 0});
+      g->inputs.push_back(n, idx, 0);
     } else if (i >= params.num_outs) {
       g->inputs.push_back(n->inputs[i-params.num_outs]);
     } else {

src/operator/nn/lrn.cc

@@ -77,7 +77,7 @@ struct LRNGrad {
     std::vector<nnvm::NodeEntry> heads;
     heads.push_back(ograds[0]);  // out_grad
     heads.push_back(n->inputs[lrn_enum::kData]);
-    heads.emplace_back(nnvm::NodeEntry{n, lrn_enum::kTmpNorm, 0});
+    heads.emplace_back(n, lrn_enum::kTmpNorm, 0);
     return MakeGradNode(op_name, n, heads, n->attrs.dict);
   }
 };

src/operator/operator_common.h

@@ -446,7 +446,7 @@ inline std::vector<nnvm::NodeEntry> MakeNonlossGradNode(
   p->inputs.insert(p->inputs.end(), inputs.begin(), inputs.end());
   std::vector<nnvm::NodeEntry> ret;
   for (uint32_t i = 0; i < p->num_outputs(); ++i) {
-    ret.emplace_back(nnvm::NodeEntry{p, i, 0});
+    ret.emplace_back(std::move(p), i, 0);
   }
   return ret;
 }

src/operator/tensor/elemwise_sum.cc

@@ -49,12 +49,12 @@ std::vector<nnvm::NodeEntry> ElementWiseSumGrad(
       nnvm::Op::Get("identity");
   CHECK_EQ(ograds.size(), 1);
   std::vector<nnvm::NodeEntry> ret;
-  nnvm::NodeEntry n_out{n, 0, 0};
-  for (size_t i = 0; i < n->inputs.size(); i++) {
+  nnvm::NodeEntry n_out(n, 0, 0);
+  for (size_t i = 0; i < n->inputs.size(); ++i) {
     nnvm::NodePtr id_node = nnvm::Node::Create();
     id_node->attrs.op = copy_op;
     id_node->inputs = {ograds[0]};
-    ret.push_back(nnvm::NodeEntry{id_node, 0, 0});
+    ret.emplace_back(id_node, 0, 0);
   }
   return ret;
 }