[Bug Fix]Fix gradient accumulator (#39577)

* merge legacy to fluid

* Remove legacy code

* Remove legacy code

* Remove DataType test

* Using Tensor directly instead of using EagerTensor

* support gradient_accumulation

* make test_imperative_lod_tensor_to_selected_rows longer

* make test_imperative_lod_tensor_to_selected_rows longer

* refine code

* Rename all EagerTensor to Tensor

* Rename some EagerTensor to Tensor

* rename EagerTensor to EagerVariable

* add more test

* fix different device gradient_accmulator bug

* merge develop

* remove useless tests

paddle/fluid/imperative/gradient_accumulator.cc

@@ -300,13 +300,10 @@ void TensorAdd(const VarType& src, VarType* dst) {
                         "should be equal, Otherwise, the calculation results "
                         "will be incorrect."));
 
-#ifdef PADDLE_WITH_XPU
   // if src and dst are in different place, copy dst to src's place
   if (dst_tensor->place() != place) {
     paddle::framework::TensorCopySync(*dst_tensor, place, dst_tensor);
   }
-#endif
-
 #define PADDLE_TENSOR_ADD(cpp_type)                                  \
   if (data_type == framework::DataTypeTrait<cpp_type>::DataType()) { \
     TensorAddFunctor<cpp_type> func(                                 \

paddle/fluid/imperative/tests/test_gradient_accmulator.cc

@@ -163,10 +163,30 @@ TEST(test_add_functor, add_functor) {
   gpu_res = TensorddTest(gpu_place, gpu_place, static_cast<double>(1.0),
                          static_cast<double>(2.0));
   EXPECT_EQ(gpu_res, 0);
+
+  // normal
+  gpu_res = TensorddTest(gpu_place, gpu_place, static_cast<float>(1.0),
+                         static_cast<float>(2.0));
+  EXPECT_EQ(gpu_res, 0);
   gpu_res =
       TensorddTest(gpu_place, gpu_place, static_cast<platform::float16>(1.0),
                    static_cast<platform::float16>(2.0));
   EXPECT_EQ(gpu_res, 0);
+  // different places
+  gpu_res = TensorddTest(cpu_place, gpu_place, static_cast<float>(1.0),
+                         static_cast<float>(2.0));
+  EXPECT_EQ(gpu_res, 0);
+  gpu_res = TensorddTest(gpu_place, cpu_place, static_cast<float>(1.0),
+                         static_cast<float>(2.0));
+  EXPECT_EQ(gpu_res, 0);
+  gpu_res =
+      TensorddTest(cpu_place, gpu_place, static_cast<platform::float16>(1.0),
+                   static_cast<platform::float16>(2.0));
+  EXPECT_EQ(gpu_res, 0);
+  gpu_res =
+      TensorddTest(gpu_place, cpu_place, static_cast<platform::float16>(1.0),
+                   static_cast<platform::float16>(2.0));
+  EXPECT_EQ(gpu_res, 0);
 #endif
 
 #ifdef PADDLE_WITH_XPU