diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc
index 31516a884baefc2e50b7774a7993740f6930212f..f98011e896f4033ef210e0eb69f93ce7800a3cd6 100644
--- a/paddle/fluid/framework/tensor_util.cc
+++ b/paddle/fluid/framework/tensor_util.cc
@@ -73,18 +73,12 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
       memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                    stream);
     } else {
-      // NOTE(zcd): Because TensorCopy is an async operation, when the src_place
-      // and dst_place are two different GPU, to ensure that the operation can
-      // be carried out correctly, we should make ctx wait.
-      // If ctx_place and src_place are the same, we should add ctx.Wait()
-      // after memory::Copy; if ctx_place and dst_place are the same, we should
-      // add ctx.Wait() before memory::Copy.
       if (platform::is_same_place(ctx_place, src_place)) {
         memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                      stream);
-        ctx.Wait();
+        platform::DeviceContextPool::Instance().Get(src.place())->Wait();
       } else if (platform::is_same_place(ctx_place, dst_place)) {
-        ctx.Wait();
+        platform::DeviceContextPool::Instance().Get(src.place())->Wait();
         memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                      stream);
       } else {
@@ -97,13 +91,6 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
 
 void TensorCopy(const Tensor& src, const platform::Place& dst_place,
                 Tensor* dst) {
-  // NOTE(zcd): If the src.place() and dst_place are two different GPU,
-  // the copy operation is carried out on the dst_place's stream. This is
-  // very important, because TensorCopy is an async operator, and in most
-  // case, once this copy operator returns, dst is to be used in dst_place's
-  // stream, if this copy operation is carried out on the src_place's stream,
-  // when dst is used in dst_place's stream the copy operation may be
-  // not completed.
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   const platform::DeviceContext* dev_ctx;
   if (platform::is_gpu_place(dst_place)) {
diff --git a/paddle/fluid/framework/tensor_util.h b/paddle/fluid/framework/tensor_util.h
index dca279b69382b80e055f661cefe84b81326704b5..4457382ade37a12f5f3613fc4113fbf1f6f91124 100644
--- a/paddle/fluid/framework/tensor_util.h
+++ b/paddle/fluid/framework/tensor_util.h
@@ -23,10 +23,25 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+// NOTE(zcd): Because TensorCopy is an async operation, when the src_place
+// and dst_place are two different GPU, to ensure that the operation can
+// be carried out correctly, there is a src_ctx wait operation in TensorCopy.
+// If ctx_place and src_place are the same, src_ctx.Wait() is added
+// after memory::Copy; if ctx_place and dst_place are the same,
+// src_ctx.Wait() is added before memory::Copy.
 void TensorCopy(const Tensor& src, const platform::Place& dst_place,
                 const platform::DeviceContext& ctx, Tensor* dst);
+
+// NOTE(zcd): If the src.place() and dst_place are two different GPU,
+// the copy operation is carried out on the dst_place's stream. This is
+// very important, because TensorCopy is an async operator, and in most
+// case, once this copy operator returns, dst is to be used in dst_place's
+// stream, if this copy operation is carried out on the src_place's stream,
+// when dst is used in dst_place's stream the copy operation may be
+// not completed.
 void TensorCopy(const Tensor& src, const platform::Place& dst_place,
                 Tensor* dst);
+
 void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
                     Tensor* dst);