support num_iteration_per_run

paddle/fluid/framework/details/async_ssa_graph_executor.cc

@@ -56,6 +56,9 @@ FeedFetchList AsyncSSAGraphExecutor::Run(
   for (size_t i = 0; i < places_.size(); ++i) {
     auto call = [this, i, &fetch_tensors]() -> FeedFetchList {
       try {
+        for (size_t j = 0; j < strategy_.num_iteration_per_run_ - 1; ++j) {
+          executors_[i]->Run(fetch_tensors);
+        }
         return executors_[i]->Run(fetch_tensors);
       } catch (...) {
         exception_holder_.Catch(std::current_exception());

paddle/fluid/framework/details/execution_strategy.h

@@ -28,6 +28,8 @@ struct ExecutionStrategy {
   size_t num_iteration_per_drop_scope_{1};
   ExecutorType type_{kDefault};
   bool dry_run_{false};
+  size_t num_iteration_per_run_{1};  // only use with async_ssa_graph_executor
+                                     // and pyreader with data queue
 };
 
 }  //  namespace details

paddle/fluid/pybind/pybind.cc

@@ -892,6 +892,17 @@ All parameter, weight, gradient are variables in Paddle.
                     2. In some NLP model, it may cause the GPU memory is insufficient,
                        in this case, you should reduce `num_iteration_per_drop_scope`.
               )DOC")
+      .def_property(
+          "num_iteration_per_run",
+          [](const ExecutionStrategy &self) {
+            return self.num_iteration_per_run_;
+          },
+          [](ExecutionStrategy &self, size_t num_iteration_per_run) {
+            self.num_iteration_per_run_ = num_iteration_per_run;
+          },
+          R"DOC(This config that how many iteration the executor will run when
+                user call pe.run() in python
+              )DOC")
       .def_property("_dry_run",
                     [](const ExecutionStrategy &self) { return self.dry_run_; },
                     [](ExecutionStrategy &self, bool dry_run) {