[MLU]add compiler options and remove redundant code (#40705)

paddle/fluid/imperative/reducer.cc

@@ -1109,8 +1109,9 @@ void Reducer::FinalizeBackward() {
 
   if (find_unused_vars_each_step_) {
 // TODO(liuyuhui) support xpu about Tensorcopy/TensorFromVector/TensorToVector
-#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) || \
-    defined(PADDLE_WITH_GLOO) || defined(PADDLE_WITH_ASCEND_CL)
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) ||      \
+    defined(PADDLE_WITH_GLOO) || defined(PADDLE_WITH_ASCEND_CL) || \
+    defined(PADDLE_WITH_CNCL)
     ProcessUnusedDenseVars();
 #endif
     // Initialize local used vars

paddle/fluid/operators/math/concat_and_split.cc

@@ -243,8 +243,6 @@ class ConcatFunctor<platform::MLUDeviceContext, T> {
 
     const int axis_t = axis;
     const int ins_size_t = ins_size;
-    auto place = context.GetPlace();
-    output->mutable_data<T>(place);
 
     // mlu should do sth
     // init ins tensors
@@ -295,7 +293,6 @@ class SplitFunctor<platform::MLUDeviceContext, T> {
     std::vector<cnnlTensorDescriptor_t> desc_vector;
     for (size_t i = 0; i < out_size; i++) {
       (*outputs)[i]->Resize(outs_dims[i]);
-      (*outputs)[i]->mutable_data<T>(context.GetPlace());
       output_descs.emplace_back(
           MLUCnnlTensorDesc(*(*outputs)[i], CNNL_LAYOUT_ARRAY,
                             ToCnnlDataType((*outputs)[i]->dtype())));