fix unnecessary tensor sync

mindspore/ccsrc/backend/session/executor.cc

@@ -42,6 +42,7 @@ void UpdateOutputTensors(const VectorRef *outputs,
       if (tensor->NeedSyncDeviceToHostImmediately()) {
         tensor->data_sync();
         tensor->set_device_address(nullptr);
+        tensor->set_sync_status(kNeedSyncHostToDevice);
       }
     }
   }

mindspore/ccsrc/runtime/device/cpu/cpu_kernel_runtime.cc

@@ -248,7 +248,6 @@ void CPUKernelRuntime::BindInputOutput(session::KernelGraph *kernel_graph, const
                                        tensor->data_c())) {
           MS_LOG(EXCEPTION) << "Parameter node sync host to device failed!";
         }
-        tensor->set_sync_status(kNeedSyncHostToDevice);
       }
       address->ref_count_ = INIT_NODE_REF;
       tensor->set_device_address(address);

mindspore/core/ir/tensor.cc

@@ -557,7 +557,6 @@ void Tensor::data_sync() const {
   if (!device_sync_->SyncDeviceToHost(shape(), static_cast<size_t>(data().nbytes()), data_type(), data_c())) {
     MS_LOG(EXCEPTION) << "SyncDeviceToHost failed.";
   }
-  sync_status_ = kNeedSyncHostToDevice;
 }
 
 TypeId Tensor::set_data_type(const TypeId data_type) {

mindspore/core/ir/tensor.h

@@ -289,7 +289,7 @@ class Tensor : public MetaTensor {
     if (event_ != nullptr) {
       event_->Wait();
     }
-    event_ == nullptr;
+    event_ = nullptr;
   }
 
   void set_sync_status(TensorSyncStatus sync_status) { sync_status_ = sync_status; }
@@ -306,7 +306,7 @@ class Tensor : public MetaTensor {
   bool init_flag_{false};
   TensorDataPtr data_{nullptr};
   std::string id_{""};
-  std::shared_ptr<WaitEvent> event_{nullptr};
+  mutable std::shared_ptr<WaitEvent> event_{nullptr};
   mutable TensorSyncStatus sync_status_{kNeedSyncHostToDevice};
   DeviceSyncPtr device_sync_{nullptr};
   std::vector<Axis> padding_type_;