check the generate_op is null or not and add DEPS of broadcast_op_handle and gather_op_handle

paddle/fluid/framework/details/CMakeLists.txt

@@ -21,11 +21,10 @@ cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph framewor
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
         simple_threadpool device_context)
 
-cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory)
-cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory)
-
 cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
 
+cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base variable_visitor scope ddim memory)
+cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope variable_visitor ddim memory)
 
 cc_test(broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
         device_context broadcast_op_handle)

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -61,8 +61,9 @@ void BroadcastOpHandle::RunImpl() {
                       "Places must be all on CPU or all on CUDA.");
 
     VariableVisitor::ShareDimsAndLoD(*in_var, out_var);
-    VariableVisitor::GetMutableTensor(out_var).mutable_data(out_p,
-                                                            in_tensor.type());
+    VariableVisitor::GetMutableTensor(out_var)
+        .Resize(in_tensor.dims())
+        .mutable_data(out_p, in_tensor.type());
 
     auto dev_ctx = dev_ctxes_[out_p];
     RunAndRecordEvent(out_p, [in_tensor, out_var, dev_ctx, out_p] {
@@ -74,9 +75,11 @@ void BroadcastOpHandle::RunImpl() {
 }
 
 void BroadcastOpHandle::WaitInputVarGenerated(const VarHandle &in_var) {
+  if (in_var.generated_op_) {
     for (auto &pair : dev_ctxes_) {
       in_var.generated_op_->Wait(pair.second);
     }
+  }
 }
 
 std::string BroadcastOpHandle::Name() const { return "broadcast"; }