update new feature of mlperf (#735) (#42706)

* update new feature of mlperf
Co-authored-by: Nyaozhixin <522190855@qq.com>

paddle/fluid/platform/device/ipu/ipu_executor.cc

@@ -215,7 +215,12 @@ void Executor::Run(const std::vector<const Tensor *> &inputs,
 
   popart::StepIO stepio(popart_inputs, popart_anchors);
   VLOG(10) << "Running...";
-  session_->run(stepio);
+  if (ipu_strategy_->popart_options.createImplicitPipeliningFwdOnlyProgram &&
+      ipu_strategy_->runtime_options.enable_eval) {
+    session_->run("implicitPipeliningFwdOnly", stepio);
+  } else {
+    session_->run(stepio);
+  }
   VLOG(10) << "Running...done";
 }
 

paddle/fluid/platform/device/ipu/ipu_strategy.cc

@@ -285,6 +285,8 @@ IpuStrategy::IpuStrategy() {
   ADD_POPART_BOOL_OPTION_ALIAS(
       schedule_non_weight_update_gradient_consumers_early,
       scheduleNonWeightUpdateGradientConsumersEarly);
+  ADD_POPART_BOOL_OPTION_ALIAS(create_implicit_pipelining_fwd_only_program,
+                               createImplicitPipeliningFwdOnlyProgram);
 
   ADD_POPART_DOUBLE_OPTION_ALIAS(outline_sequence_break_cost,
                                  outlineSequenceBreakCost);