[1.x] Backport Faster pointwise fusion graph pass (#19269)

src/executor/exec_pass.h

@@ -222,22 +222,14 @@ Graph DetectInplaceAddTo(Graph g);
 Graph EliminateCommonExpr(Graph && g);
 
 /*!
- * \brief Fuse pointwise operations in the forward pass.
+ * \brief Fuse pointwise operations in the graph.
  *
  * \param g input graph (needs to be entire graph, not just forward part)
+ * \param num_forward_outputs number of outputs in the graph produced by the forward pass
  *
- * \return graph with fused pointwise operations in the forward pass
+ * \return copy of the graph with fused pointwise operations
  */
-Graph FusePointwiseForward(Graph&& g);
-
-/*!
- * \brief Fuse pointwise operations in the backward pass.
- *
- * \param g input graph (needs to be entire graph, not just forward part)
- *
- * \return graph with fused pointwise operations in the backward pass
- */
-Graph FusePointwiseBackward(Graph&& g);
+Graph FusePointwise(const Graph& g, const size_t num_forward_outputs);
 
 /*!
  * \brief Issue a one-time warning that fusion is not possible for this platform or build.

src/executor/graph_executor.cc

@@ -1009,10 +1009,7 @@ Graph GraphExecutor::InitGraph(nnvm::Symbol symbol,
     common::CopyGraph(&unoptimized_graph, g, false);
 
     if (common::CheckForInputNameDuplicates(unoptimized_graph.indexed_graph())) {
-      g.attrs["num_forward_outputs"] = std::make_shared<nnvm::any>(num_forward_outputs_);
-      g = FusePointwiseForward(std::move(g));
-      g.attrs["num_forward_outputs"] = std::make_shared<nnvm::any>(num_forward_outputs_);
-      g = FusePointwiseBackward(std::move(g));
+      g = exec::FusePointwise(std::move(g), num_forward_outputs_);
       // Check the topological order of inputs
       const auto &original_inputs = unoptimized_graph.indexed_graph().input_nodes();
       const auto &new_inputs = g.indexed_graph().input_nodes();