[PIR slice]Change deep_copy to share_buffer in eager backward (#72966)

* change deep_copy to share_buffer in eager backward

* share grad_out_tensor only for shared_buffer io

Author: changeyoung98

paddle/fluid/eager/backward.cc

@@ -192,9 +192,30 @@ std::vector<paddle::Tensor> RunBackward(
       // Feed given tensor if it's provided
       VLOG(3) << "Fill grad input tensor " << i << "with give grad tensor";
 
-      // Deep copy
-      node_input_buffers_dict[grad_node]->CopyValueFromTensor(
-          input_info.first, input_info.second, grad_tensors[i]);
+      bool use_shared_buffer = false;
+      // Check if inputs and outputs are equal in size and share the same buffer
+      if (tensors.size() == inputs.size() &&
+          tensors[i].numel() == inputs[i].numel()) {
+        auto output_tensor =
+            std::dynamic_pointer_cast<phi::DenseTensor>(tensors[i].impl());
+        auto input_tensor =
+            std::dynamic_pointer_cast<phi::DenseTensor>(inputs[i].impl());
+        use_shared_buffer = output_tensor->IsSharedBufferWith(*input_tensor);
+      }
+
+      if (use_shared_buffer) {
+        // Share buffer with given grad_tensor
+        paddle::small_vector<std::vector<paddle::Tensor>, kSlotSmallVectorSize>
+            inputs_grad_tensors;
+        inputs_grad_tensors.push_back({grad_tensors[i]});
+        auto grad_holder = GradTensorHolder(std::move(inputs_grad_tensors));
+        node_input_buffers_dict[grad_node] =
+            std::make_unique<GradTensorHolder>(grad_holder);
+      } else {
+        // Deep copy
+        node_input_buffers_dict[grad_node]->CopyValueFromTensor(
+            input_info.first, input_info.second, grad_tensors[i]);
+      }
     } else {
       VLOG(3) << "Fill grad input tensor " << i << " with 1.0";
       // Initialize tensor with 1.0