optimize vlog before and after op run, move into op.run

paddle/fluid/framework/executor.cc

@@ -330,12 +330,7 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
   }
 
   for (auto& op : ctx->ops_) {
-    VLOG(4) << place_ << " " << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);
-    // NOTE! Please do not delete this line, it's usefull because the debug
-    // string before and after op.run are different, after run the output
-    // will have right shape which is usefull for debug.
-    VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
 
     if (FLAGS_benchmark) {
       VLOG(2) << "Memory used after operator " + op->Type() + " running: "

paddle/fluid/framework/operator.cc

@@ -127,7 +127,7 @@ static LoD GetLoD(const Scope& scope, const std::string& name) {
 }
 
 void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
-  VLOG(10) << "- " << DebugStringEx(&scope);
+  VLOG(4) << place << " " << DebugStringEx(&scope);
   if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
     PADDLE_THROW("Cannot run operator on place %s", place);
@@ -139,7 +139,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   platform::RecordEvent record_event(Type(), pool.Get(place));
   RunImpl(scope, place);
-  VLOG(10) << "+ " << DebugStringEx(&scope);
+  VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 
 bool OperatorBase::HasInputs(const std::string& name) const {