fix precision error with fp16 input on PyNative mode

mindspore/ccsrc/backend/optimizer/ascend/ascend_helper.cc

@@ -167,7 +167,8 @@ AnfNodePtr InsertTransOpForMultipleOutput(const FuncGraphPtr &func_graph, const
 }
 }  // namespace
 void RefreshKernelBuildInfo(const std::string &input_format, const std::string &output_format,
-                            const AnfNodePtr &trans_data, const std::vector<kernel::Axis> &reshape_type) {
+                            const AnfNodePtr &trans_data, const std::vector<kernel::Axis> &reshape_type,
+                            const TypeId &type_id) {
   MS_EXCEPTION_IF_NULL(trans_data);
   auto ori_build_info = AnfAlgo::GetSelectKernelBuildInfo(trans_data);
   MS_EXCEPTION_IF_NULL(ori_build_info);
@@ -176,6 +177,10 @@ void RefreshKernelBuildInfo(const std::string &input_format, const std::string &
   builder->SetInputReshapeType({reshape_type});
   builder->SetOutputReshapeType({reshape_type});
   builder->SetOutputsFormat({output_format});
+  if (type_id != kTypeUnknown) {
+    builder->SetOutputsDeviceType({type_id});
+    builder->SetInputsDeviceType({type_id});
+  }
   AnfAlgo::SetSelectKernelBuildInfo(builder->Build(), trans_data.get());
 }
 

mindspore/ccsrc/backend/optimizer/ascend/ascend_helper.h

@@ -86,7 +86,8 @@ class OpFinder {
 using OpFinderPtr = std::shared_ptr<OpFinder>;
 
 void RefreshKernelBuildInfo(const std::string &input_format, const std::string &output_format,
-                            const AnfNodePtr &trans_data, const std::vector<kernel::Axis> &reshape_type = {});
+                            const AnfNodePtr &trans_data, const std::vector<kernel::Axis> &reshape_type = {},
+                            const TypeId &type_id = kTypeUnknown);
 
 CNodePtr NewTransOpNode(const FuncGraphPtr &func_graph, const AnfNodePtr &input, const KernelSelectPtr &kernel_select,
                         const bool need_padding, const std::string &op_name);

mindspore/ccsrc/backend/optimizer/ascend/format_type/deal_ref_trans_and_cast.cc

@@ -107,7 +107,7 @@ AnfNodePtr AddAdditionalToRefOutput(const FuncGraphPtr &func_graph, const CNodeP
   if (origin_format != cur_format && cur_shape.size() > 1) {
     auto kernel_select = std::make_shared<KernelSelect>();
     final_node = NewTransOpNode(func_graph, final_node, kernel_select, false, prim::KPrimTransData->name());
-    RefreshKernelBuildInfo(cur_format, origin_format, final_node);
+    RefreshKernelBuildInfo(cur_format, origin_format, final_node, {}, cur_type);
     final_index = 0;
     MS_EXCEPTION_IF_NULL(final_node);
     MS_LOG(INFO) << "DealRefTransAndCast add trans op, op debug info is " << final_node->DebugString();