diff --git a/src/common/dep_core.h b/src/common/dep_core.h
deleted file mode 100644
index a9fdca5b1de0307ed9bde99dcc65ca92fd5aee53..0000000000000000000000000000000000000000
--- a/src/common/dep_core.h
+++ /dev/null
@@ -1,67 +0,0 @@
-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#pragma once
-
-#ifdef PADDLE_EXECUTOR_MULTITHREAD
-#include <string>
-#include <unordered_map>
-#include <vector>
-#include "framework/operator.h"
-
-namespace paddle_mobile {
-
-class depCore {
- public:
-  template <typename Dtype>
-  void analysisDep(
-      const std::vector<std::shared_ptr<framework::OperatorBase<Dtype>>>& ops) {
-    std::unordered_map<std::string, int> vars;
-    size_t nop = ops.size();
-    deps.resize(nop);
-    next.resize(nop);
-    for (size_t i = 0; i < nop; i++) {
-      const auto& op = ops[i];
-      for (const auto& kv : op->Inputs()) {
-        for (const auto& v : kv.second) {
-          if (vars.find(v) == vars.end()) {
-            continue;
-          }
-          int di = vars[v];
-          if (di == i) {
-            continue;
-          }
-          if (std::find(deps[i].begin(), deps[i].end(), di) != deps[i].end()) {
-            continue;
-          }
-          deps[i].push_back(di);
-          next[di].push_back(i);
-        }
-      }
-      for (const auto& kv : op->Outputs()) {
-        for (const auto& v : kv.second) {
-          vars[v] = i;
-        }
-      }
-    }
-  }
-  const std::vector<int>& getNext(int i) { return next[i]; }
-  const std::vector<int>& getDeps(int i) { return deps[i]; }
-  std::vector<std::vector<int>> deps;
-  std::vector<std::vector<int>> next;
-};
-
-}  // namespace paddle_mobile
-
-#endif
diff --git a/src/common/macros.h b/src/common/macros.h
deleted file mode 100644
index ee38f19c9285b369e48c550b67f6c397695e73cf..0000000000000000000000000000000000000000
--- a/src/common/macros.h
+++ /dev/null
@@ -1,17 +0,0 @@
-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#pragma once
-
-#define EXPORT __attribute__((visibility("default")))
diff --git a/src/common/type_define.h b/src/common/type_define.h
index a25a19f11f444c1a92b033c0f721174bda44ed50..bedbd2a75e6de73b4f741cd0dd5ab7fcc36a998e 100644
--- a/src/common/type_define.h
+++ b/src/common/type_define.h
@@ -15,32 +15,173 @@ limitations under the License. */
 #pragma once
 
 #include <functional>
-#include <map>
 #include <string>
 #include <vector>
-#include "framework/attribute.h"
-#include "framework/scope.h"
 
 namespace paddle_mobile {
 
+typedef enum {
+  _void = 0,
+  _float,
+  _int,
+  _uint16_t,
+  _double,
+  _int64_t,
+  _size_t,
+  _int16_t,
+  _int8_t,
+  _uint8_t,
+  _bool,
+  _string,
+  _floats = 100,
+  _ints,
+  _int64_ts,
+  _size_ts,
+  _bools,
+  _strings,
+  _const_float = 200,
+  _const_int,
+  _block = 300,
+  _tensor,
+  _lod_tensor,
+  _blocks,
+  _tensors,
+  _lod_tensors,
+  _p_block = 400,
+  _p_tensor,
+  _p_lod_tensor,
+  _p_blocks,
+  _p_tensors,
+  _p_lod_tensors,
+  _scopes = 500,
+  _selected_rows,
+  _dim0 = 600,
+  _dim1,
+  _dim2,
+  _dim3,
+  _dim4,
+  _dim5,
+  _dim6,
+  _dim7,
+  _dim8,
+  _dim9,
+#ifdef PADDLE_MOBILE_CL
+  _cl_image,
+#endif
+} kTypeId_t;
+
+template <typename T>
+struct TypeIdWrapper {
+  inline std::string name();
+  inline kTypeId_t hash_code();
+};
+
+template <typename T>
+struct type_id {
+  const kTypeId_t hash_code() const { return TypeIdWrapper<T>().hash_code(); }
+  const std::string name() const { return TypeIdWrapper<T>().name(); }
+
+  template <typename OtherType>
+  bool operator==(const type_id<OtherType> &operand) const {
+    return this->hash_code() == operand.hash_code();
+  }
+};
+
+#define OVERIDE_TYPEID_OPERATOR(oprand)                                    \
+  template <typename T>                                                    \
+  inline bool operator oprand(const kTypeId_t &t0, const type_id<T> &t1) { \
+    return t0 oprand t1.hash_code();                                       \
+  }                                                                        \
+  template <typename T>                                                    \
+  inline bool operator oprand(const type_id<T> &t0, const kTypeId_t &t1) { \
+    return t1 oprand t0.hash_code();                                       \
+  }
+
+OVERIDE_TYPEID_OPERATOR(==)
+OVERIDE_TYPEID_OPERATOR(!=)
+
 namespace framework {
-template <typename Dtype>
-class OperatorBase;
-class OpDesc;
 class BlockDesc;
-class InferShapeContext;
+class Tensor;
+class LoDTensor;
+class SelectedRows;
+class Scope;
+#ifdef PADDLE_MOBILE_CL
+class CLImage;
+#endif
+
+template <int>
+struct Dim;
 }  // namespace framework
 
-using VariableNameMap = std::map<std::string, std::vector<std::string>>;
+#define REGISTER_TYPE_ID(Type, TypeName)                         \
+  template <>                                                    \
+  struct TypeIdWrapper<Type> {                                   \
+    inline std::string name() { return std::string(#TypeName); } \
+    inline kTypeId_t hash_code() { return kTypeId_t::TypeName; } \
+  };
+
+REGISTER_TYPE_ID(void, _void)
+REGISTER_TYPE_ID(float, _float)
+REGISTER_TYPE_ID(int, _int)
+REGISTER_TYPE_ID(uint16_t, _uint16_t)
+REGISTER_TYPE_ID(double, _double)
+REGISTER_TYPE_ID(int64_t, _int64_t)
+REGISTER_TYPE_ID(size_t, _size_t)
+REGISTER_TYPE_ID(int16_t, _int16_t)
+REGISTER_TYPE_ID(int8_t, _int8_t)
+REGISTER_TYPE_ID(uint8_t, _uint8_t)
+REGISTER_TYPE_ID(bool, _bool)
+REGISTER_TYPE_ID(std::string, _string)
+REGISTER_TYPE_ID(std::vector<float>, _floats)
+REGISTER_TYPE_ID(std::vector<int>, _ints)
+REGISTER_TYPE_ID(std::vector<int64_t>, _int64_ts)
+REGISTER_TYPE_ID(std::vector<size_t>, _size_ts)
+REGISTER_TYPE_ID(std::vector<bool>, _bools)
+REGISTER_TYPE_ID(std::vector<std::string>, _strings)
+
+REGISTER_TYPE_ID(float const, _const_float)
+REGISTER_TYPE_ID(int const, _const_int)
+
+REGISTER_TYPE_ID(framework::BlockDesc, _block)
+REGISTER_TYPE_ID(framework::Tensor, _tensor)
+REGISTER_TYPE_ID(framework::LoDTensor, _lod_tensor)
+REGISTER_TYPE_ID(std::vector<framework::BlockDesc>, _blocks)
+REGISTER_TYPE_ID(std::vector<framework::Tensor>, _tensors)
+REGISTER_TYPE_ID(std::vector<framework::LoDTensor>, _lod_tensors)
+
+REGISTER_TYPE_ID(framework::BlockDesc *, _p_block)
+REGISTER_TYPE_ID(framework::Tensor *, _p_tensor)
+REGISTER_TYPE_ID(framework::LoDTensor *, _p_lod_tensor)
+REGISTER_TYPE_ID(std::vector<framework::BlockDesc *>, _p_blocks)
+REGISTER_TYPE_ID(std::vector<framework::Tensor *>, _p_tensors)
+REGISTER_TYPE_ID(std::vector<framework::LoDTensor *>, _p_lod_tensors)
+
+REGISTER_TYPE_ID(std::vector<framework::Scope *>, _scopes);
+REGISTER_TYPE_ID(framework::SelectedRows, _selected_rows)
+REGISTER_TYPE_ID(framework::Dim<0>, _dim0)
+REGISTER_TYPE_ID(framework::Dim<1>, _dim1)
+REGISTER_TYPE_ID(framework::Dim<2>, _dim2)
+REGISTER_TYPE_ID(framework::Dim<3>, _dim3)
+REGISTER_TYPE_ID(framework::Dim<4>, _dim4)
+REGISTER_TYPE_ID(framework::Dim<5>, _dim5)
+REGISTER_TYPE_ID(framework::Dim<6>, _dim6)
+REGISTER_TYPE_ID(framework::Dim<7>, _dim7)
+REGISTER_TYPE_ID(framework::Dim<8>, _dim8)
+REGISTER_TYPE_ID(framework::Dim<9>, _dim9)
+
+#ifdef PADDLE_MOBILE_CL
+REGISTER_TYPE_ID(framework::CLImage, _cl_image)
+#endif
+}  // namespace paddle_mobile
 
-template <typename Dtype>
-using OpCreator = std::function<framework::OperatorBase<Dtype> *(
-    const std::string & /*type*/, const VariableNameMap & /*inputs*/,
-    const VariableNameMap & /*outputs*/,
-    const framework::AttributeMap & /*attrs*/, framework::Scope * /*scope*/)>;
+namespace std {
 
-using InferVarTypeFN = std::function<void(const framework::OpDesc & /*op_desc*/,
-                                          framework::BlockDesc * /*block*/)>;
+template <>
+struct hash<paddle_mobile::kTypeId_t> {
+  size_t operator()(const paddle_mobile::kTypeId_t &t) const {
+    return std::hash<int>{}(static_cast<int>(t));
+  }
+};
 
-using InferShapeFN = std::function<void(framework::InferShapeContext *)>;
-};  // namespace paddle_mobile
+}  // namespace std
diff --git a/src/common/types.h b/src/common/types.h
old mode 100755
new mode 100644
index 35c1659c5a246c49f6e0ccb31c3e63ce4fdd2e71..f8e1fd26ea34a1f6ab7cd5c6eeac75e901d0e02b
--- a/src/common/types.h
+++ b/src/common/types.h
@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <map>
 #include <string>
 #include <unordered_map>
 #include <utility>
@@ -211,4 +212,6 @@ extern std::unordered_map<
     std::string, std::pair<std::vector<std::string>, std::vector<std::string>>>
     op_input_output_key;
 
+typedef std::map<std::string, std::vector<std::string>> VariableNameMap;
+
 }  // namespace paddle_mobile
diff --git a/src/common/variant.h b/src/common/variant.h
index 4aa4f47c628caec438ecd00522d90ebf299da6a0..87fd822243dff30deb3fbcd3c61ffbe8807c3938 100644
--- a/src/common/variant.h
+++ b/src/common/variant.h
@@ -19,6 +19,7 @@ limitations under the License. */
 #include <string>
 #include "common/enforce.h"
 #include "common/log.h"
+#include "common/type_define.h"
 
 namespace paddle_mobile {
 
@@ -33,11 +34,11 @@ struct VariantHelper {
                                  ? sizeof(F)
                                  : VariantHelper<Ts...>::size;
 
-  inline static void Destroy(size_t id, void *data) {
-    if (id == typeid(F).hash_code()) {
+  inline static void Destroy(kTypeId_t type, void *data) {
+    if (type == type_id<F>()) {
       reinterpret_cast<F *>(data)->~F();
     } else {
-      VariantHelper<Ts...>::Destroy(id, data);
+      VariantHelper<Ts...>::Destroy(type, data);
     }
   }
 };
@@ -45,11 +46,11 @@ struct VariantHelper {
 template <typename F>
 struct VariantHelper<F> {
   static const size_t size = sizeof(F);
-  inline static void Destroy(size_t id, void *data) {
-    if (id == typeid(F).hash_code()) {
-      //              reinterpret_cast<F*>(data)->~F();
+  inline static void Destroy(kTypeId_t type, void *data) {
+    if (type == type_id<F>()) {
+      // reinterpret_cast<F*>(data)->~F();
     } else {
-      //              std::cout << "未匹配到 " << std::endl;
+      // std::cout << "未匹配到 " << std::endl;
     }
   }
 };
@@ -57,7 +58,7 @@ struct VariantHelper<F> {
 template <size_t size>
 class RawData {
  public:
-  char data[size];
+  char data[size];  // NOLINT
   RawData() {}
   RawData(const RawData &raw_data) { memcpy(data, raw_data.data, size); }
 
@@ -69,32 +70,33 @@ class RawData {
 
 template <typename... Ts>
 struct Variant {
+  Variant() : type_(invalid_type()) {}
+
   Variant(const Variant &variant) {
-    type_id = variant.type_id;
-    data = variant.data;
+    type_ = variant.type_;
+    data_ = variant.data_;
   }
 
-  Variant() : type_id(invalid_type()) {}
-  ~Variant() {
-    //        helper::Destroy(type_id, &data);
+  virtual ~Variant() {
+    // helper::Destroy(type_id, &data);
   }
 
   template <typename T, typename... Args>
   void Set(Args &&... args) {
-    helper::Destroy(type_id, data.data);
-    new (data.data) T(std::forward<Args>(args)...);
-    type_id = typeid(T).hash_code();
+    helper::Destroy(type_, data_.data);
+    new (data_.data) T(std::forward<Args>(args)...);
+    type_ = type_id<T>().hash_code();
   }
 
-  void SetString(std::string &string) {
-    helper::Destroy(type_id, data.data);
-    type_id = typeid(std::string).hash_code();
-    strcpy(data.data, string.c_str());
+  void SetString(const std::string &string) {
+    helper::Destroy(type_, data_.data);
+    type_ = type_id<std::string>().hash_code();
+    strcpy(data_.data, string.c_str());  // NOLINT
   }
 
   std::string GetString() const {
-    if (type_id == typeid(std::string).hash_code()) {
-      return std::string(data.data);
+    if (type_ == type_id<std::string>()) {
+      return std::string(data_.data);
     } else {
       PADDLE_MOBILE_THROW_EXCEPTION(
           " bad cast in variant data type not a string ");
@@ -104,28 +106,25 @@ struct Variant {
 
   template <typename T>
   T &Get() const {
-    if (type_id == typeid(std::string).hash_code()) {
+    if (type_ == type_id<std::string>()) {
       PADDLE_MOBILE_THROW_EXCEPTION(
           "Please use getString to get an string (to avoid of an issue with "
           "gcc "
           "stl lib with string copy)");
       exit(0);
-    } else if (type_id == typeid(T).hash_code()) {
-      return *const_cast<T *>(reinterpret_cast<const T *>(data.data));
     } else {
-      PADDLE_MOBILE_THROW_EXCEPTION(" bad cast in variant");
-      exit(0);
+      return *const_cast<T *>(reinterpret_cast<const T *>(data_.data));
     }
   }
 
-  size_t TypeId() const { return type_id; }
+  kTypeId_t TypeId() const { return type_; }
 
  private:
-  static inline size_t invalid_type() { return typeid(void).hash_code(); }
+  static inline kTypeId_t invalid_type() { return type_id<void>().hash_code(); }
   typedef VariantHelper<Ts...> helper;
-  size_t type_id;
+  kTypeId_t type_ = type_id<void>().hash_code();
   // todo use an anto size to suite this.
-  RawData<64> data;
+  RawData<64> data_;
 };
 
 template <typename T>
diff --git a/src/fpga/V1/api.cpp b/src/fpga/V1/api.cpp
index 4ad252d41420442844aee25dddd7dcbac22aef2d..2bcae0cb4745bb27c7a995446a6f7e2cfc54c694 100644
--- a/src/fpga/V1/api.cpp
+++ b/src/fpga/V1/api.cpp
@@ -28,8 +28,8 @@ namespace fpga {
 void format_image(framework::Tensor *image_tensor) {
   auto dims = image_tensor->dims();
   auto channel = dims[1], height = dims[2], width = dims[3];
-  std::type_index input_type = image_tensor->type();
-  if (input_type == typeid(float)) {
+  kTypeId_t input_type = image_tensor->type();
+  if (input_type == type_id<float>()) {
     auto data_ptr = image_tensor->data<float>();
     auto external_ptr = reinterpret_cast<float *>(image_tensor->external_data);
     float *p_data = external_ptr == nullptr ? data_ptr : external_ptr;
@@ -51,7 +51,7 @@ void format_image(framework::Tensor *image_tensor) {
 }
 
 void format_ofm(framework::Tensor *ofm_tensor) {
-  if (ofm_tensor->type() == typeid(float)) {
+  if (ofm_tensor->type() == type_id<float>()) {
     format_fp32_ofm(ofm_tensor);
   } else {
     format_fp16_ofm(ofm_tensor);
@@ -72,7 +72,7 @@ void format_fp16_ofm(framework::Tensor *ofm_tensor) {
   auto p = fpga_malloc(memory_size);
   // memset(p, 0, memory_size);
   ofm_tensor->reset_data_ptr(p);
-  ofm_tensor->set_type(typeid(half));
+  ofm_tensor->set_type(type_id<half>().hash_code());
   ofm_tensor->fpga_data_num = memory_size / sizeof(half);
   fpga::fpga_flush(p, memory_size);
 }
@@ -92,7 +92,7 @@ void format_fp16_ofm(framework::Tensor *ofm_tensor, framework::DDim dims) {
   auto p = fpga_malloc(memory_size);
   // memset(p, 0, memory_size);
   ofm_tensor->reset_data_ptr(p);
-  ofm_tensor->set_type(typeid(half));
+  ofm_tensor->set_type(type_id<half>().hash_code());
   ofm_tensor->fpga_data_num = memory_size / sizeof(half);
   fpga::fpga_flush(p, memory_size);
 }
@@ -112,7 +112,7 @@ void format_fp32_ofm(framework::Tensor *ofm_tensor) {
   auto p = fpga_malloc(memory_size);
   // memset(p, 0, memory_size);
   ofm_tensor->reset_data_ptr(p);
-  ofm_tensor->set_type(typeid(float));
+  ofm_tensor->set_type(type_id<float>().hash_code());
   ofm_tensor->fpga_data_num = memory_size / sizeof(float);
   fpga::fpga_flush(p, memory_size);
 }
@@ -171,7 +171,7 @@ void format_filter(framework::Tensor *filter_tensor, float max_value,
   filter::format_filter(&new_data, num, channel, height, width, group_num,
                         max_value);
   filter_tensor->reset_data_ptr(new_data);
-  filter_tensor->set_type(typeid(int8_t));
+  filter_tensor->set_type(type_id<int8_t>().hash_code());
 }
 void format_dwconv_filter(framework::Tensor *filter_tensor, float *scale_ptr) {
   auto dims = filter_tensor->dims();
@@ -182,7 +182,7 @@ void format_dwconv_filter(framework::Tensor *filter_tensor, float *scale_ptr) {
   fpga_copy(new_data, data_ptr, memory_size);
   filter::format_dwconv_filter(&new_data, num, height, width, scale_ptr);
   filter_tensor->reset_data_ptr(new_data);
-  filter_tensor->set_type(typeid(int16_t));
+  filter_tensor->set_type(type_id<int16_t>().hash_code());
 }
 
 void format_DWDconv_filter(framework::Tensor *filter_tensor, float *scale_ptr,
@@ -207,7 +207,7 @@ void format_DWDconv_filter(framework::Tensor *filter_tensor, float *scale_ptr,
   //      framework::make_ddim({num, 1, height, width});
   //  filter_tensor->Resize(dims_new);
   filter_tensor->reset_data_ptr(new_data);
-  filter_tensor->set_type(typeid(int16_t));
+  filter_tensor->set_type(type_id<int16_t>().hash_code());
 }
 
 void format_fc_filter(framework::Tensor *filter_tensor, float max_value) {
@@ -222,7 +222,7 @@ void format_fc_filter(framework::Tensor *filter_tensor, float max_value) {
   filter::format_fc_filter(&new_data, num, channel, height, width, 1,
                            max_value);
   filter_tensor->reset_data_ptr(new_data);
-  filter_tensor->set_type(typeid(int8_t));
+  filter_tensor->set_type(type_id<int8_t>().hash_code());
 }
 void format_deconv_filter(framework::Tensor *filter_tensor, float max_value,
                           int group_num, int stride) {
@@ -249,7 +249,7 @@ void format_deconv_filter(framework::Tensor *filter_tensor, float max_value,
       framework::make_ddim({num, channel, height, width});
   filter_tensor->Resize(dims_new);
   filter_tensor->reset_data_ptr(new_data);
-  filter_tensor->set_type(typeid(int8_t));
+  filter_tensor->set_type(type_id<int8_t>().hash_code());
 }
 
 void format_bias_scale_array(float **bias_scale_array,
@@ -273,7 +273,7 @@ void format_concat_output(framework::Tensor *out, int height, int width,
   auto ddim = framework::make_ddim({1, sum_channel, height, width});
   out->Resize(ddim);
   out->reset_data_ptr(data_ptr);
-  out->set_type(typeid(half));
+  out->set_type(type_id<half>().hash_code());
 }
 void format_conv_data(framework::Tensor *filter_tensor,
                       framework::Tensor *ofm_tensor, float **bs_ptr,
diff --git a/src/framework/attribute.h b/src/framework/attribute.h
index e00cee09b36a4372c938f356900faab88e610010..01c4a8d7d4b59b8c0a92913618b4407de769b044 100644
--- a/src/framework/attribute.h
+++ b/src/framework/attribute.h
@@ -128,31 +128,30 @@ class Attribute {
 
   template <typename Vistor>
   static typename Vistor::type_t ApplyVistor(Vistor vistor, Attribute attr) {
-    if (attr.variant_.TypeId() == typeid(int).hash_code()) {  // NOLINT
+    if (attr.variant_.TypeId() == type_id<int>()) {  // NOLINT
       return vistor(attr.variant_.Get<int>());
-    } else if (attr.variant_.TypeId() == typeid(float).hash_code()) {  // NOLINT
+    } else if (attr.variant_.TypeId() == type_id<float>()) {  // NOLINT
       return vistor(attr.variant_.Get<float>());
-    } else if (attr.variant_.TypeId() == typeid(string).hash_code()) {
+    } else if (attr.variant_.TypeId() == type_id<string>()) {
       return vistor(attr.variant_.GetString());
-    } else if (attr.variant_.TypeId() == typeid(vector<int>).hash_code()) {
+    } else if (attr.variant_.TypeId() == type_id<vector<int>>()) {
       return vistor(attr.variant_.Get<vector<int>>());
-    } else if (attr.variant_.TypeId() == typeid(vector<float>).hash_code()) {
+    } else if (attr.variant_.TypeId() == type_id<vector<float>>()) {
       return vistor(attr.variant_.Get<vector<float>>());
-    } else if (attr.variant_.TypeId() == typeid(vector<string>).hash_code()) {
+    } else if (attr.variant_.TypeId() == type_id<vector<string>>()) {
       return vistor(attr.variant_.Get<vector<string>>());
-    } else if (attr.variant_.TypeId() == typeid(bool).hash_code()) {  // NOLINT
+    } else if (attr.variant_.TypeId() == type_id<bool>()) {  // NOLINT
       return vistor(attr.variant_.Get<bool>());
-    } else if (attr.variant_.TypeId() == typeid(vector<bool>).hash_code()) {
+    } else if (attr.variant_.TypeId() == type_id<vector<bool>>()) {
       return vistor(attr.variant_.Get<vector<bool>>());
-    } else if (attr.variant_.TypeId() == typeid(int64_t).hash_code()) {
+    } else if (attr.variant_.TypeId() == type_id<int64_t>()) {
       return vistor(attr.variant_.Get<int64_t>());
-    } else if (attr.variant_.TypeId() ==
-               typeid(framework::BlockDesc *).hash_code()) {
+    } else if (attr.variant_.TypeId() == type_id<framework::BlockDesc *>()) {
       return vistor(attr.variant_.Get<framework::BlockDesc *>());
     } else if (attr.variant_.TypeId() ==
-               typeid(vector<framework::BlockDesc *>).hash_code()) {
+               type_id<vector<framework::BlockDesc *>>()) {
       return vistor(attr.variant_.Get<vector<framework::BlockDesc *>>());
-    } else if (attr.variant_.TypeId() == typeid(vector<int64_t>).hash_code()) {
+    } else if (attr.variant_.TypeId() == type_id<vector<int64_t>>()) {
       return vistor(attr.variant_.Get<vector<int64_t>>());
     } else {
       PADDLE_MOBILE_THROW_EXCEPTION("type not support");
diff --git a/src/framework/cl/cl_tensor.h b/src/framework/cl/cl_tensor.h
index 93a3648bfe8054ed0b54728bb7000423949f494a..a0ed438f9773dbacaaa7446d594719c0cf12b32e 100644
--- a/src/framework/cl/cl_tensor.h
+++ b/src/framework/cl/cl_tensor.h
@@ -53,12 +53,12 @@ class CLTensor : TensorBase {
     int64_t size = numel() * sizeof(T);
 
     holder_.reset(new PlaceholderImpl(
-        size, reinterpret_cast<void *>(const_cast<T *>(data)), typeid(T),
-        context_, command_queue_));
+        size, reinterpret_cast<void *>(const_cast<T *>(data)),
+        type_id<T>().hash_code(), context_, command_queue_));
     return reinterpret_cast<cl_mem>(holder_->ptr());
   }
 
-  inline cl_mem mutable_data(std::type_index type) {
+  inline cl_mem mutable_data(kTypeId_t type) {
     if (holder_ != nullptr) {
       holder_->set_type(type);
     }
@@ -77,7 +77,7 @@ class CLTensor : TensorBase {
    */
   template <typename T>
   inline cl_mem mutable_data() {
-    return reinterpret_cast<cl_mem>(mutable_data(typeid(T)));
+    return reinterpret_cast<cl_mem>(mutable_data(type_id<T>().hash_code()));
   }
 
   /**
@@ -132,7 +132,7 @@ class CLTensor : TensorBase {
   void *host_ptr_ = nullptr;
 
   struct PlaceholderImpl : public Placeholder {
-    PlaceholderImpl(size_t size, void *input, std::type_index type,
+    PlaceholderImpl(size_t size, void *input, kTypeId_t type,
                     cl_context context, cl_command_queue command_queue)
         : ptr_(clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
                               size, reinterpret_cast<void *>(input), NULL)),
@@ -142,7 +142,7 @@ class CLTensor : TensorBase {
           context_(context),
           command_queue_(command_queue) {}
 
-    PlaceholderImpl(size_t size, std::type_index type, cl_context context,
+    PlaceholderImpl(size_t size, kTypeId_t type, cl_context context,
                     cl_command_queue command_queue)
         : ptr_(clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, NULL)),
           size_(size),
@@ -155,9 +155,9 @@ class CLTensor : TensorBase {
 
     virtual void *ptr() const { return static_cast<void *>(ptr_.get()); }
 
-    virtual std::type_index type() const { return type_; }
+    virtual kTypeId_t type() const { return type_; }
 
-    virtual void set_type(std::type_index type) { type_ = type; }
+    virtual void set_type(kTypeId_t type) { type_ = type; }
 
     virtual void resize(size_t size) {
       if (size > capatity_) {
@@ -175,7 +175,7 @@ class CLTensor : TensorBase {
     size_t capatity_;
 
     /* the current type of memory */
-    std::type_index type_;
+    kTypeId_t type_;
 
     cl_context context_;
     cl_command_queue command_queue_;
diff --git a/src/framework/data_type.cpp b/src/framework/data_type.cpp
index 0bcf7d9f67dae28db5a316476778b4132b39b274..5eaf3ecaf599d5ed34c5396bbba2c9e0e68736fd 100644
--- a/src/framework/data_type.cpp
+++ b/src/framework/data_type.cpp
@@ -16,17 +16,17 @@ limitations under the License. */
 #include <stdint.h>
 #include <string>
 #include <unordered_map>
+#include "common/type_define.h"
 
 namespace paddle_mobile {
 namespace framework {
 
 struct DataTypeMap {
-  std::unordered_map<std::type_index,
-                     _PaddleMobile__Framework__Proto__VarType__Type>
+  std::unordered_map<kTypeId_t, _PaddleMobile__Framework__Proto__VarType__Type>
       cpp_to_proto_;
-  std::unordered_map<int, std::type_index> proto_to_cpp_;
+  std::unordered_map<int, kTypeId_t> proto_to_cpp_;
   std::unordered_map<int, std::string> proto_to_str_;
-  std::unordered_map<std::type_index, size_t> cpp_to_size_;
+  std::unordered_map<kTypeId_t, size_t> cpp_to_size_;
 };
 
 static DataTypeMap* InitDataTypeMap();
@@ -42,10 +42,11 @@ template <typename T>
 static inline void RegisterType(
     DataTypeMap* map, _PaddleMobile__Framework__Proto__VarType__Type proto_type,
     const std::string& name) {
-  map->proto_to_cpp_.emplace(static_cast<int>(proto_type), typeid(T));
-  map->cpp_to_proto_.emplace(typeid(T), proto_type);
+  map->proto_to_cpp_.emplace(static_cast<int>(proto_type),
+                             type_id<T>().hash_code());
+  map->cpp_to_proto_.emplace(type_id<T>().hash_code(), proto_type);
   map->proto_to_str_.emplace(static_cast<int>(proto_type), name);
-  map->cpp_to_size_.emplace(typeid(T), sizeof(T));
+  map->cpp_to_size_.emplace(type_id<T>().hash_code(), sizeof(T));
 }
 
 static DataTypeMap* InitDataTypeMap() {
@@ -70,17 +71,15 @@ static DataTypeMap* InitDataTypeMap() {
   return retv;
 }
 
-_PaddleMobile__Framework__Proto__VarType__Type ToDataType(
-    std::type_index type) {
+_PaddleMobile__Framework__Proto__VarType__Type ToDataType(kTypeId_t type) {
   auto it = gDataTypeMap().cpp_to_proto_.find(type);
   if (it != gDataTypeMap().cpp_to_proto_.end()) {
     return it->second;
   }
-  PADDLE_MOBILE_THROW_EXCEPTION("Not support %s as tensor type", type.name());
+  PADDLE_MOBILE_THROW_EXCEPTION("Not support %d as tensor type", type);
 }
 
-std::type_index ToTypeIndex(
-    _PaddleMobile__Framework__Proto__VarType__Type type) {
+kTypeId_t ToTypeIndex(_PaddleMobile__Framework__Proto__VarType__Type type) {
   auto it = gDataTypeMap().proto_to_cpp_.find(static_cast<int>(type));
   if (it != gDataTypeMap().proto_to_cpp_.end()) {
     return it->second;
diff --git a/src/framework/data_type.h b/src/framework/data_type.h
index 231bb62e1f7a24bdf65b417d18679ca31bf13d83..bda823ada49b108e1f3279173f87daefc3403b5b 100644
--- a/src/framework/data_type.h
+++ b/src/framework/data_type.h
@@ -15,18 +15,17 @@ limitations under the License. */
 #pragma once
 
 #include <string>
-#include <typeindex>
 #include "common/enforce.h"
+#include "common/type_define.h"
 #include "framework/framework.pb-c.h"
 
 namespace paddle_mobile {
 
 namespace framework {
 
-extern _PaddleMobile__Framework__Proto__VarType__Type ToDataType(
-    std::type_index type);
-extern std::type_index ToTypeIndex(
-    _PaddleMobile__Framework__Proto__VarType__Type type);
+_PaddleMobile__Framework__Proto__VarType__Type ToDataType(kTypeId_t type);
+
+kTypeId_t ToTypeIndex(_PaddleMobile__Framework__Proto__VarType__Type type);
 
 inline _PaddleMobile__Framework__Proto__VarType__Type ToDataType(int type) {
   return static_cast<_PaddleMobile__Framework__Proto__VarType__Type>(type);
diff --git a/src/framework/ddim.h b/src/framework/ddim.h
index 74dd288ba88108af3de895c3ef535dedaa5edfc2..8ba0756a6b6774a5641242c03ebeb86454b38ad2 100644
--- a/src/framework/ddim.h
+++ b/src/framework/ddim.h
@@ -22,7 +22,7 @@ limitations under the License. */
 
 #include "common/enforce.h"
 #include "common/variant.h"
-#include "dim.h"
+#include "framework/dim.h"
 
 namespace paddle_mobile {
 namespace framework {
@@ -40,25 +40,25 @@ struct DDim {
 
   template <typename Vistor>
   static typename Vistor::type_t ApplyVistor(Vistor vistor, const DDim &d) {
-    if (d.var.TypeId() == typeid(Dim<0>).hash_code()) {
+    if (d.var.TypeId() == type_id<Dim<0>>()) {
       return vistor(d.var.Get<Dim<0>>());
-    } else if (d.var.TypeId() == typeid(Dim<1>).hash_code()) {
+    } else if (d.var.TypeId() == type_id<Dim<1>>()) {
       return vistor(d.var.Get<Dim<1>>());
-    } else if (d.var.TypeId() == typeid(Dim<2>).hash_code()) {
+    } else if (d.var.TypeId() == type_id<Dim<2>>()) {
       return vistor(d.var.Get<Dim<2>>());
-    } else if (d.var.TypeId() == typeid(Dim<3>).hash_code()) {
+    } else if (d.var.TypeId() == type_id<Dim<3>>()) {
       return vistor(d.var.Get<Dim<3>>());
-    } else if (d.var.TypeId() == typeid(Dim<4>).hash_code()) {
+    } else if (d.var.TypeId() == type_id<Dim<4>>()) {
       return vistor(d.var.Get<Dim<4>>());
-    } else if (d.var.TypeId() == typeid(Dim<5>).hash_code()) {
+    } else if (d.var.TypeId() == type_id<Dim<5>>()) {
       return vistor(d.var.Get<Dim<5>>());
-    } else if (d.var.TypeId() == typeid(Dim<6>).hash_code()) {
+    } else if (d.var.TypeId() == type_id<Dim<6>>()) {
       return vistor(d.var.Get<Dim<6>>());
-    } else if (d.var.TypeId() == typeid(Dim<7>).hash_code()) {
+    } else if (d.var.TypeId() == type_id<Dim<7>>()) {
       return vistor(d.var.Get<Dim<7>>());
-    } else if (d.var.TypeId() == typeid(Dim<8>).hash_code()) {
+    } else if (d.var.TypeId() == type_id<Dim<8>>()) {
       return vistor(d.var.Get<Dim<8>>());
-    } else if (d.var.TypeId() == typeid(Dim<9>).hash_code()) {
+    } else if (d.var.TypeId() == type_id<Dim<9>>()) {
       return vistor(d.var.Get<Dim<9>>());
     } else {
       PADDLE_MOBILE_ENFORCE(false, " dim not support");
diff --git a/src/framework/executor.cpp b/src/framework/executor.cpp
index b9f455bd2708c82bd2165dc34e0a22150de40fac..f337d1d184d0d4880a28df7d066301ca0d3e1460 100644
--- a/src/framework/executor.cpp
+++ b/src/framework/executor.cpp
@@ -62,7 +62,7 @@ Executor<Device, T>::Executor(const Program<Device> &program,
       use_optimize_ ? program_.optimizeProgram : program_.originProgram;
   PADDLE_MOBILE_ENFORCE(program_desc_ != nullptr,
                         "program_desc_ should not be nullptr");
-#ifndef PADDLE_MOBILE_FPGA
+#if !defined(PADDLE_MOBILE_FPGA) && !defined(PADDLE_MOBILE_CL)
   pass::MemoryOptPass()(program_desc_.get(), program_.scope.get());
 #endif
   // resize feed and fetch list
@@ -302,25 +302,9 @@ bool Executor<Device, T>::varInputMemory(
     const std::shared_ptr<VarDesc> &var_desc, Variable *var) const {
 #ifdef PADDLE_MOBILE_FPGA
   framework::LoDTensor *tensor = var->template GetMutable<LoDTensor>();
-  tensor->init(typeid(float));
+  tensor->init(type_id<float>().hash_code());
   return true;
 #endif
-  auto TypeId = [](const VarType_Type &type) -> std::type_index {
-    switch (type) {
-      case VARTYPE_TYPE_BOOL:
-        return typeid(bool);
-      case VARTYPE_TYPE_FP32:
-        return typeid(float);
-      case VARTYPE_TYPE_INT8:
-        return typeid(int8_t);
-      case VARTYPE_TYPE_INT32:
-        return typeid(int);
-      case VARTYPE_TYPE_INT64:
-        return typeid(int64_t);
-      default:
-        PADDLE_MOBILE_THROW_EXCEPTION("got unhandled var type `%d`", type);
-    }
-  };
 
   auto type = var_desc->Type();
   if (type == VARTYPE_TYPE_LOD_TENSOR) {
@@ -390,13 +374,6 @@ void Executor<Device, T>::SetInput(const Tensor &input,
   framework::LoDTensor &target =
       feed_var->template GetMutable<framework::LoDTensorArray>()->at(index);
 
-  if (config_.load_when_predict) {
-    if (input_dim_last_ != input.dims()) {
-      InitNoPersistableMemory(input);
-      input_dim_last_ = input.dims();
-    }
-  }
-
   target.Resize(input.dims());
   target.ShareDataWith(input);
 }
@@ -412,13 +389,6 @@ void Executor<Device, T>::SetInput(const LoDTensor &input,
   framework::LoDTensor &target =
       feed_var->template GetMutable<framework::LoDTensorArray>()->at(index);
 
-  if (config_.load_when_predict) {
-    if (input_dim_last_ != input.dims()) {
-      InitNoPersistableMemory(input);
-      input_dim_last_ = input.dims();
-    }
-  }
-
   target.Resize(input.dims());
   target.ShareDataWith(input);
   target.set_lod(input.lod());
diff --git a/src/framework/load_ops.h b/src/framework/load_ops.h
index a969a73b9edfe1e7a475ea7191bfcd552d09357c..983a544cda8faa756562914d26f36c4ab8608be2 100644
--- a/src/framework/load_ops.h
+++ b/src/framework/load_ops.h
@@ -268,6 +268,9 @@ LOAD_OP1(sequence_expand, CPU);
 #ifdef SEQUENCE_POOL_OP
 LOAD_OP1(sequence_pool, CPU);
 #endif
+#ifdef SEQUENCE_SOFTMAX_OP
+LOAD_OP1(sequence_softmax, CPU);
+#endif
 #ifdef LOG_OP
 LOAD_OP1(log, CPU);
 #endif
diff --git a/src/framework/lod_tensor.cpp b/src/framework/lod_tensor.cpp
index e165e55507ed04a9b63e4ad5eb002f206c71d96c..0a1a6f881d0aaf20cc9ca6a109e2915464025aa4 100644
--- a/src/framework/lod_tensor.cpp
+++ b/src/framework/lod_tensor.cpp
@@ -12,52 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "lod_tensor.h"
+#include "framework/lod_tensor.h"
 #include <algorithm>
 
 namespace paddle_mobile {
 namespace framework {
 
-// std::ostream &operator<<(std::ostream &os, const LoD &lod) {
-//  os << "{";
-//  for (auto &v : lod) {
-//    os << "{";
-//    bool is_first = true;
-//    for (auto &i : v) {
-//      if (is_first) {
-//        os << i;
-//        is_first = false;
-//      } else {
-//        os << ", " << i;
-//      }
-//    }
-//    os << "}";
-//  }
-//  os << "}";
-//
-//  return os;
-//}
-//
-// std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
-//  PADDLE_MOBILE_ENFORCE(t.type().hash_code() == typeid(float).hash_code(),
-//                        "t.type() is not float");
-//  os << "dim: " << t.dims() << "\n";
-//  os << "lod: " << t.lod() << "\n";
-//  // only print first ten elements
-//  int64_t size = t.numel() < 10 ? t.numel() : 10;
-//  for (int64_t i = 0; i < size; ++i) {
-//    os << t.data<float>()[i] << " ";
-//  }
-//
-//  return os;
-//}
-
-// std::string LoDToString(const LoD &lod) {
-//  std::ostringstream stream;
-//  stream << lod;
-//  return stream.str();
-//}
-
 LoD SliceInLevel(const LoD &in, size_t level, size_t elem_begin,
                  size_t elem_end) {
   PADDLE_MOBILE_ENFORCE(level < in.size(), "level should >= in.size()");
diff --git a/src/framework/lod_tensor.h b/src/framework/lod_tensor.h
index e96fe0e501f06814914677d0fb6248a11fa1dde5..6d67b517ff38a3bb4b2ae6978de1290079596aa6 100644
--- a/src/framework/lod_tensor.h
+++ b/src/framework/lod_tensor.h
@@ -211,17 +211,17 @@ inline Print &operator<<(Print &printer, const LoDTensor &tensor) {
   stride = stride > 0 ? stride : 1;
 #ifndef PADDLE_MOBILE_FPGA
   for (int i = 0; i < tensor.numel(); i += stride) {
-    if (tensor.type() == typeid(float)) {
+    if (tensor.type() == type_id<float>()) {
       printer << tensor.data<float>()[i] << " ";
-    } else if (tensor.type() == typeid(int32_t)) {
+    } else if (tensor.type() == type_id<int32_t>()) {
       printer << tensor.data<int32_t>()[i] << " ";
-    } else if (tensor.type() == typeid(int64_t)) {
+    } else if (tensor.type() == type_id<int64_t>()) {
       printer << tensor.data<int64_t>()[i] << " ";
-    } else if (tensor.type() == typeid(int8_t)) {
+    } else if (tensor.type() == type_id<int8_t>()) {
       printer << static_cast<int>(tensor.data<int8_t>()[i]) << " ";
-    } else if (tensor.type() == typeid(int32_t)) {
+    } else if (tensor.type() == type_id<int32_t>()) {
       printer << tensor.data<int32_t>()[i] << " ";
-    } else if (tensor.type() == typeid(bool)) {
+    } else if (tensor.type() == type_id<bool>()) {
       printer << tensor.data<bool>()[i] << " ";
     }
   }
diff --git a/src/framework/mixed_vector.h b/src/framework/mixed_vector.h
index 031d73179c991229ec99ebdde927b0ad1532d82b..6e46164fb791fdc28a9dddff4b38c3d3b346b5c7 100644
--- a/src/framework/mixed_vector.h
+++ b/src/framework/mixed_vector.h
@@ -17,7 +17,6 @@
 #include <algorithm>
 #include <initializer_list>
 #include <vector>
-
 #include "framework/tensor.h"
 #include "framework/tensor_util.h"
 
@@ -198,7 +197,7 @@ class Vector {
   }
 
   size_t capacity() const {
-    return cpu_vec_.memory_size() / SizeOfType(typeid(T));
+    return cpu_vec_.memory_size() / SizeOfType(type_id<T>().hash_code());
   }
 
   // reserve data
diff --git a/src/framework/op_info.h b/src/framework/op_info.h
index 16b3487955ce05721e6e3f3e79b6d8ebd180e020..c250f61664108b57cb5eba7411638fd4d5b8dc9b 100644
--- a/src/framework/op_info.h
+++ b/src/framework/op_info.h
@@ -14,13 +14,24 @@ limitations under the License. */
 
 #pragma once
 
+#include <functional>
 #include <string>
 #include "common/log.h"
 #include "common/type_define.h"
+#include "framework/scope.h"
 
 namespace paddle_mobile {
 namespace framework {
 
+template <typename Dtype>
+class OperatorBase;
+
+template <typename Dtype>
+using OpCreator = std::function<framework::OperatorBase<Dtype> *(
+    const std::string & /*type*/, const VariableNameMap & /*inputs*/,
+    const VariableNameMap & /*outputs*/,
+    const framework::AttributeMap & /*attrs*/, framework::Scope * /*scope*/)>;
+
 template <typename Dtype>
 struct OpInfo {
   OpCreator<Dtype> creator_;
@@ -79,8 +90,6 @@ class OpInfoMap {
  private:
   OpInfoMap() = default;
   std::unordered_map<std::string, OpInfo<Dtype>> map_;
-
-  //  DISABLE_COPY_AND_ASSIGN(OpInfoMap);
 };
 
 }  // namespace framework
diff --git a/src/framework/operator.h b/src/framework/operator.h
index 93f23e9d1067392a0bb43e8f764dd6695249d996..c8b3a5ccf796c5cce33efb697cd6d6aef8b9d21c 100644
--- a/src/framework/operator.h
+++ b/src/framework/operator.h
@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <functional>
 #include <map>
 #include <string>
 #include <utility>
diff --git a/src/framework/program/block_desc.h b/src/framework/program/block_desc.h
index dd33a274266cb503cea0b960c026276d90cea57a..86dd832d1b108878313d5a34ada3ce5a2b9c321e 100644
--- a/src/framework/program/block_desc.h
+++ b/src/framework/program/block_desc.h
@@ -14,6 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include <memory>
+#include <vector>
 #include "framework/framework.pb-c.h"
 #include "framework/program/op_desc.h"
 #include "framework/program/var_desc.h"
@@ -26,8 +28,8 @@ class BlockDesc {
   friend class Node;
   friend class ProgramOptimize;
   BlockDesc() {}
-  BlockDesc(PaddleMobile__Framework__Proto__BlockDesc *desc);
-  BlockDesc(const BlockDesc &block_desc)
+  explicit BlockDesc(PaddleMobile__Framework__Proto__BlockDesc *desc);
+  explicit BlockDesc(const BlockDesc &block_desc)
       : index_(block_desc.index_), parent_index_(block_desc.parent_index_) {
     for (auto &op_desc : block_desc.ops_) {
       std::shared_ptr<OpDesc> copy_op_desc = std::make_shared<OpDesc>(*op_desc);
diff --git a/src/framework/program/op_desc.h b/src/framework/program/op_desc.h
index f9579df034ccfb3e68863b27eef1df935ba266d5..89c877ba120db630f30dc21d48c0347b0668cb22 100644
--- a/src/framework/program/op_desc.h
+++ b/src/framework/program/op_desc.h
@@ -18,7 +18,8 @@ limitations under the License. */
 #include <vector>
 
 #include "common/log.h"
-#include "common/type_define.h"
+#include "common/types.h"
+#include "framework/attribute.h"
 #include "framework/framework.pb-c.h"
 
 namespace paddle_mobile {
diff --git a/src/framework/program/program-optimize/node.h b/src/framework/program/program-optimize/node.h
index b86fc96a67e290540c94487497fced55abf09041..5b5ae7796f3c99f991100d3313563900fcf59250 100644
--- a/src/framework/program/program-optimize/node.h
+++ b/src/framework/program/program-optimize/node.h
@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <cinttypes>
 #include <map>
+#include <memory>
 #include <string>
 #include <utility>
 #include <vector>
diff --git a/src/framework/program/program_desc.cpp b/src/framework/program/program_desc.cpp
index 23781fe77962608b515ebb7c0479f284ebfc4277..88cac11d283cd0c5b72a6809deb34df8206c3e25 100644
--- a/src/framework/program/program_desc.cpp
+++ b/src/framework/program/program_desc.cpp
@@ -72,7 +72,7 @@ void ProgramDesc::Description(std::string header) const {
         }
       }
       for (auto &attr : op->GetAttrMap()) {
-        if (attr.first == "op_callstack") continue;
+        if (attr.first == "op_callstack" || attr.first == "sub_block") continue;
         LOG(kLOG_DEBUG2) << "attr name: " << attr.first;
         LOG(kLOG_DEBUG3) << "argument - " << attr.second;
       }
diff --git a/src/framework/tensor.h b/src/framework/tensor.h
index 4fb06c654983b9e1b8441b074d5a30220fb960a2..fe2f266137b98df778e21818a42ab7f9e13e617d 100644
--- a/src/framework/tensor.h
+++ b/src/framework/tensor.h
@@ -19,8 +19,6 @@ limitations under the License. */
 #include <fstream>
 #include <memory>
 #include <string>
-#include <type_traits>
-#include <typeindex>
 #include <vector>
 
 #include "common/enforce.h"
@@ -83,7 +81,7 @@ class Tensor : public TensorBase {
     return *this;
   }
 
-  inline void *mutable_data(std::type_index type) {
+  inline void *mutable_data(const kTypeId_t type) {
     if (holder_ != nullptr) {
       holder_->set_type(type);
     }
@@ -108,7 +106,7 @@ class Tensor : public TensorBase {
   template <typename T>
   inline T *mutable_data() {
     static_assert(std::is_pod<T>::value, "T must be POD");
-    return reinterpret_cast<T *>(mutable_data(typeid(T)));
+    return reinterpret_cast<T *>(mutable_data(type_id<T>().hash_code()));
   }
 
   /**
@@ -165,9 +163,9 @@ class Tensor : public TensorBase {
     check_memory_size();
     PADDLE_MOBILE_ENFORCE(
         (std::is_same<T, void>::value ||
-         holder_->type().hash_code() == typeid(T).hash_code()),
-        "Tensor holds the wrong type, it holds %s, requested %s",
-        this->holder_->type().name(), typeid(T).name());
+         holder_->type() == type_id<T>().hash_code()),
+        "Tensor holds the wrong type, it holds %d, requested %d",
+        this->holder_->type(), type_id<T>().hash_code());
 
     return reinterpret_cast<T *>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                                  offset_);
@@ -179,9 +177,9 @@ class Tensor : public TensorBase {
     check_memory_size();
     PADDLE_MOBILE_ENFORCE(
         (std::is_same<T, void>::value ||
-         holder_->type().hash_code() == typeid(T).hash_code()),
-        "Tensor holds the wrong type, it holds %s, requested %s",
-        this->holder_->type().name(), typeid(T).name());
+         holder_->type() == type_id<T>().hash_code()),
+        "Tensor holds the wrong type, it holds %d, requested %d",
+        this->holder_->type(), type_id<T>().hash_code());
 
     return reinterpret_cast<const T *>(
         reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
@@ -189,7 +187,7 @@ class Tensor : public TensorBase {
 
  private:
   struct PlaceholderImpl : public Placeholder {
-    PlaceholderImpl(size_t size, std::type_index type)
+    PlaceholderImpl(size_t size, const kTypeId_t type)
         : ptr_(static_cast<uint8_t *>(memory::Alloc(size)),
                memory::PODDeleter<uint8_t>()),
           size_(size),
@@ -203,9 +201,9 @@ class Tensor : public TensorBase {
 
     virtual void *ptr() const { return static_cast<void *>(ptr_.get()); }
 
-    virtual std::type_index type() const { return type_; }
+    virtual kTypeId_t type() const { return type_; }
 
-    virtual void set_type(std::type_index type) { type_ = type; }
+    virtual void set_type(const kTypeId_t type) { type_ = type; }
 
     virtual void resize(size_t size) {
       if (size > capatity_) {
@@ -223,7 +221,7 @@ class Tensor : public TensorBase {
     size_t capatity_;
 
     /* the current type of memory */
-    std::type_index type_;
+    kTypeId_t type_;
   };
 
 #ifdef PADDLE_MOBILE_FPGA
@@ -231,13 +229,13 @@ class Tensor : public TensorBase {
   inline void reset_data_ptr(void *p) {
     ((PlaceholderImpl *)(holder_.get()))->ptr_.reset((uint8_t *)p);  // NOLINT
   }
-  inline void set_type(std::type_index type) { holder_->set_type(type); }
+  inline void set_type(const kTypeId_t type) { holder_->set_type(type); }
   inline void *get_data() {
     return (
         void *)(((PlaceholderImpl *)(holder_.get()))->ptr_.get());  // NOLINT
   }
 
-  inline void *init(std::type_index type) {
+  inline void *init(const kTypeId_t type) {
     if (holder_ != nullptr) {
       holder_->set_type(type);
     }
@@ -265,15 +263,15 @@ inline Print &operator<<(Print &printer, const Tensor &tensor) {
   stride = stride > 0 ? stride : 1;
 #ifndef PADDLE_MOBILE_FPGA
   for (int i = 0; i < tensor.numel(); i += stride) {
-    if (tensor.type() == typeid(float)) {
+    if (tensor.type() == type_id<float>()) {
       printer << tensor.data<float>()[i] << " ";
-    } else if (tensor.type() == typeid(int32_t)) {
+    } else if (tensor.type() == type_id<int32_t>()) {
       printer << tensor.data<int32_t>()[i] << " ";
-    } else if (tensor.type() == typeid(int64_t)) {
+    } else if (tensor.type() == type_id<int64_t>()) {
       printer << tensor.data<int64_t>()[i] << " ";
-    } else if (tensor.type() == typeid(int8_t)) {
+    } else if (tensor.type() == type_id<int8_t>()) {
       printer << static_cast<int>(tensor.data<int8_t>()[i]) << " ";
-    } else if (tensor.type() == typeid(int32_t)) {
+    } else if (tensor.type() == type_id<int32_t>()) {
       printer << tensor.data<int32_t>()[i] << " ";
     }
   }
diff --git a/src/framework/tensor_base.h b/src/framework/tensor_base.h
index e5ab7793c0eb392126359b050b9f0b8bc25c3515..027f1165a08509431fd1281f7b05174a7c64b7cc 100644
--- a/src/framework/tensor_base.h
+++ b/src/framework/tensor_base.h
@@ -14,10 +14,8 @@ limitations under the License. */
 
 #pragma once
 
-#include <type_traits>
-#include <typeindex>
-
 #include "common/enforce.h"
+#include "common/type_define.h"
 #include "common/types.h"
 #include "framework/ddim.h"
 
@@ -29,8 +27,8 @@ struct SizeOfTypeFunctor;
 
 template <typename T>
 struct SizeOfTypeFunctor<T> {
-  size_t operator()(std::type_index type) const {
-    if (typeid(T).hash_code() == type.hash_code()) {
+  size_t operator()(const kTypeId_t type) const {
+    if (type_id<T>().hash_code() == type) {
       return sizeof(T);
     } else {
       return 0UL;
@@ -40,12 +38,12 @@ struct SizeOfTypeFunctor<T> {
 
 template <>
 struct SizeOfTypeFunctor<> {
-  size_t operator()(std::type_index type) const { return 0UL; }
+  size_t operator()(const kTypeId_t type) const { return 0UL; }
 };
 
 template <typename HEAD, typename... TAIL>
 struct SizeOfTypeFunctor<HEAD, TAIL...> {
-  size_t operator()(std::type_index type) const {
+  size_t operator()(const kTypeId_t type) const {
     SizeOfTypeFunctor<HEAD> head;
     size_t head_size = head(type);
     if (head_size != 0) {
@@ -56,13 +54,13 @@ struct SizeOfTypeFunctor<HEAD, TAIL...> {
   }
 };
 
-static inline size_t SizeOfType(std::type_index type) {
+static inline size_t SizeOfType(const kTypeId_t type) {
   SizeOfTypeFunctor<int8_t, int, half, float, double, int16_t, int64_t, bool,
                     size_t>
       functor;
   size_t size = functor(type);
 
-  PADDLE_MOBILE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name());
+  PADDLE_MOBILE_ENFORCE(size != 0UL, "Cannot get size of type %d", type);
   return size;
 }
 
@@ -78,7 +76,7 @@ class TensorBase {
   /*! Return the numel of the memory block. */
   inline int64_t numel() const { return product(dims_); }
 
-  std::type_index type() const {
+  kTypeId_t type() const {
     PADDLE_MOBILE_ENFORCE(
         holder_ != nullptr,
         "Tensor not initialized yet when Tensor::type() is called.")
@@ -114,9 +112,9 @@ class TensorBase {
 
     virtual size_t size() const = 0;
 
-    virtual std::type_index type() const = 0;
+    virtual kTypeId_t type() const = 0;
 
-    virtual void set_type(std::type_index type) = 0;
+    virtual void set_type(kTypeId_t type) = 0;
 
     virtual void resize(size_t size) = 0;
   };
diff --git a/src/framework/variable.h b/src/framework/variable.h
index 5bff63f068ca13fa6252006c4618a7f8a9d3b2f7..30486cb34721a06ef5babf5cc796302b23834617 100644
--- a/src/framework/variable.h
+++ b/src/framework/variable.h
@@ -16,13 +16,10 @@ limitations under the License. */
 
 #include <memory>
 #include <string>
-#include <typeindex>
-#include <typeinfo>
-#include "../common/variant.h"
+#include "common/variant.h"
 
 namespace paddle_mobile {
 namespace framework {
-using std::string;
 
 class Variable {
  public:
@@ -33,7 +30,7 @@ class Variable {
 
   template <typename T>
   const T GetValue() const {
-    if (typeid(T) == typeid(std::string)) {
+    if (type_id<T>().hash_code() == type_id<std::string>().hash_code()) {
       PADDLE_MOBILE_THROW_EXCEPTION(
           "Please use getString to get an string (to avoid of an issue with "
           "gcc "
@@ -60,38 +57,40 @@ class Variable {
 
   template <typename T>
   bool IsType() const {
-    return holder_ != nullptr && holder_->Type() == typeid(T);
+    return holder_ != nullptr && holder_->Type() == type_id<T>().hash_code();
   }
 
   void Clear() { holder_.reset(); }
 
-  std::type_index Type() const { return holder_->Type(); }
+  kTypeId_t Type() const { return holder_->Type(); }
 
  private:
   struct Placeholder {
     Placeholder() = default;
     virtual ~Placeholder() = default;
 
-    virtual const std::type_info &Type() const = 0;
+    virtual kTypeId_t Type() const = 0;
     virtual void *Ptr() const = 0;
   };
 
   template <typename T>
   struct PlaceholderImp : public Placeholder {
-    explicit PlaceholderImp(T *ptr) : ptr_(ptr), type_(typeid(T)) {}
+    explicit PlaceholderImp(T *ptr)
+        : ptr_(ptr), type_(type_id<T>().hash_code()) {}
 
-    virtual const std::type_info &Type() const { return type_; }
-    virtual void *Ptr() const override {
-      return static_cast<void *>(ptr_.get());
-    }
+    kTypeId_t Type() const override { return type_; }
+    void *Ptr() const override { return static_cast<void *>(ptr_.get()); }
 
     std::unique_ptr<T> ptr_;
-    const std::type_info &type_;
+    kTypeId_t type_;
   };
-  Variant<int, bool, string, float, double> variant;
-  std::unique_ptr<Placeholder> holder_;
+
   friend class Scope;
-  string name_;
+
+  Variant<int, bool, std::string, float, double> variant;
+  std::unique_ptr<Placeholder> holder_;
+  std::string name_;
 };
+
 }  // namespace framework
 }  // namespace paddle_mobile
diff --git a/src/io/api_paddle_mobile.cc b/src/io/api_paddle_mobile.cc
index 1f4769b282385207a5b53d6d678364393d7da6cc..e7782de201b9306ea3a65708ef12a00ca8f981c2 100644
--- a/src/io/api_paddle_mobile.cc
+++ b/src/io/api_paddle_mobile.cc
@@ -128,7 +128,7 @@ void ConvertTensors(const framework::Tensor &src, PaddleTensor *des) {
   des->layout = src.layout == framework::LAYOUT_HWC ? LAYOUT_HWC : LAYOUT_CHW;
 
   auto num = src.numel();
-  if (src.type() == typeid(float)) {
+  if (src.type() == type_id<float>()) {
     des->data.Reset(const_cast<float *>(src.data<float>()),
                     num * sizeof(float));
   } else {
@@ -143,7 +143,7 @@ void PaddleMobilePredictor<Device, T>::FeedPaddleTensors(
   auto num = inputs.size();
   std::vector<framework::Tensor> tensors(num, framework::Tensor());
   for (int i = 0; i < num; i++) {
-    tensors[i].init(typeid(float));
+    tensors[i].init(type_id<float>().hash_code());
     ConvertPaddleTensors(inputs[i], &tensors[i]);
   }
   paddle_mobile_->FeedTensorData(tensors);
diff --git a/src/io/paddle_inference_api.h b/src/io/paddle_inference_api.h
index e8c2c1daaa7bb41ccb7da0a969d941e0131249ab..b1e787c6496ee3c01918c1a9a143f35d7119f7a1 100644
--- a/src/io/paddle_inference_api.h
+++ b/src/io/paddle_inference_api.h
@@ -24,8 +24,8 @@ limitations under the License. */
 #include <cassert>
 #include <memory>
 #include <string>
-#include <typeindex>
 #include <vector>
+#include "common/type_define.h"
 
 namespace paddle_mobile {
 
@@ -88,7 +88,7 @@ struct PaddleTensor {
   // TODO(Superjomn) for LoD support, add a vector<vector<int>> field if needed.
   PaddleBuf data;  // blob of data.
   PaddleDType dtype;
-  std::type_index dtypeid = typeid(float);
+  kTypeId_t dtypeid;
   LayoutType layout;
 };
 
diff --git a/src/operators/beam_search_decode_op.cpp b/src/operators/beam_search_decode_op.cpp
index 9b01d2e17f363d3b729102a9747f6dc6682ea8aa..1038234fe8ecf1b68f5fdcd52c1f1b4491658ae7 100644
--- a/src/operators/beam_search_decode_op.cpp
+++ b/src/operators/beam_search_decode_op.cpp
@@ -14,8 +14,6 @@ limitations under the License. */
 
 #ifdef BEAM_SEARCH_DECODE_OP
 
-#pragma once
-
 #include "operators/beam_search_decode_op.h"
 
 namespace paddle_mobile {
diff --git a/src/operators/beam_search_op.cpp b/src/operators/beam_search_op.cpp
index 502510ebeefd29336531fac24d279e009f6b8d6d..5f83e536672daf6ae83f07209c40e6f900a9514d 100644
--- a/src/operators/beam_search_op.cpp
+++ b/src/operators/beam_search_op.cpp
@@ -14,8 +14,6 @@ limitations under the License. */
 
 #ifdef BEAM_SEARCH_OP
 
-#pragma once
-
 #include "operators/beam_search_op.h"
 
 namespace paddle_mobile {
diff --git a/src/operators/kernel/arm/compare_kernel.cpp b/src/operators/kernel/arm/compare_kernel.cpp
index d83fae1748bc9b942425219ea169b63d5c84b867..35bb13363cea5a8de173cc7491afa5cc421e5746 100644
--- a/src/operators/kernel/arm/compare_kernel.cpp
+++ b/src/operators/kernel/arm/compare_kernel.cpp
@@ -192,10 +192,10 @@ bool LessThanKernel<CPU, float>::Init(CompareParam<CPU> *param) {
 
 template <>
 void LessThanKernel<CPU, float>::Compute(const CompareParam<CPU> &param) {
-  if (param.input_x_->type() == typeid(int64_t)) {
+  if (param.input_x_->type() == type_id<int64_t>().hash_code()) {
     CompareCompute<int64_t, LESS_THAN>()(param.input_x_, param.input_y_,
                                          param.axis_, param.output_);
-  } else if (param.input_x_->type() == typeid(float)) {
+  } else if (param.input_x_->type() == type_id<float>().hash_code()) {
     CompareCompute<float, LESS_THAN>()(param.input_x_, param.input_y_,
                                        param.axis_, param.output_);
   } else {
diff --git a/src/operators/kernel/arm/concat_kernel.cpp b/src/operators/kernel/arm/concat_kernel.cpp
index 3c6a6f151f9b05ad0b69b40298ee5a47797d70af..3e585ec7214239ab5752d197e207fd55c6dffa68 100644
--- a/src/operators/kernel/arm/concat_kernel.cpp
+++ b/src/operators/kernel/arm/concat_kernel.cpp
@@ -27,7 +27,7 @@ bool ConcatKernel<CPU, float>::Init(ConcatParam<CPU> *param) {
 
 template <>
 void ConcatKernel<CPU, float>::Compute(const ConcatParam<CPU> &param) {
-  if (param.Inputs()[0]->type() == typeid(int8_t)) {
+  if (param.Inputs()[0]->type() == type_id<int8_t>().hash_code()) {
     ConcatCompute<int8_t>(param);
   } else {
     ConcatCompute<float>(param);
diff --git a/src/operators/kernel/arm/convolution/conv_common.cpp b/src/operators/kernel/arm/convolution/conv_common.cpp
index 86c6e8d3373f743f485e4a69599e8c8323ba0083..5a5c04c656c120c93e4b14b1242fa9accb3c2244 100644
--- a/src/operators/kernel/arm/convolution/conv_common.cpp
+++ b/src/operators/kernel/arm/convolution/conv_common.cpp
@@ -28,7 +28,7 @@ void InitBaseConvKernel(ConvParam<CPU> *param) {
 
   bool depth5x5 = conv5x5 && param->Groups() == param->Input()->dims()[1] &&
                   param->Input()->dims()[1] == param->Output()->dims()[1];
-  if (param->Filter()->type() == typeid(int8_t)) {
+  if (param->Filter()->type() == type_id<int8_t>().hash_code()) {
 #ifndef __aarch64__
     if (depth3x3 && param->Strides()[0] < 3 &&
         param->Strides()[0] == param->Strides()[1]) {
diff --git a/src/operators/kernel/arm/dequantize_kernel.cpp b/src/operators/kernel/arm/dequantize_kernel.cpp
index c9dfd6f93667cf283e83e21c0a1f3725bb847956..7c0d1cea18c90145c5dbc06de7cb97fa5ed289b6 100644
--- a/src/operators/kernel/arm/dequantize_kernel.cpp
+++ b/src/operators/kernel/arm/dequantize_kernel.cpp
@@ -34,7 +34,7 @@ void DequantizeKernel<CPU, float>::Compute(const DequantizeParam<CPU> &param) {
   LoDTensor *output = param.output_;
   float activation_scale = param.activation_scale_->data<float>()[0];
   float weight_scale = param.weight_scale_;
-  const int32_t *x = input->data<const int32_t>();
+  const int32_t *x = input->data<int32_t>();
   float *y = output->mutable_data<float>();
   size_t size = output->numel();
   // float scale = 1.f / (activation_scale * weight_scale);
diff --git a/src/operators/kernel/arm/quantize_kernel.cpp b/src/operators/kernel/arm/quantize_kernel.cpp
index 79186be79e5b4207e9f6b2b221bc1a4160e4e67e..515e9cf40dad4eedd307bc59309177910330a3bf 100644
--- a/src/operators/kernel/arm/quantize_kernel.cpp
+++ b/src/operators/kernel/arm/quantize_kernel.cpp
@@ -36,7 +36,7 @@ inline float32_t vmaxvq_f32(float32x4_t r) {
 template <RoundType R>
 inline void QuantizeOffline(const Tensor *input, const float scale,
                             const float max_abs, Tensor *output) {
-  const float *x = input->data<const float>();
+  const float *x = input->data<float>();
   int8_t *y = output->mutable_data<int8_t>();
   size_t remain = input->numel();
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
@@ -88,7 +88,7 @@ inline void QuantizeOffline(const Tensor *input, const float scale,
 template <RoundType R>
 inline void QuantizeOnline(const Tensor *input, const float scale,
                            Tensor *output) {
-  const float *x = input->data<const float>();
+  const float *x = input->data<float>();
   int8_t *y = output->mutable_data<int8_t>();
   size_t remain = input->numel();
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
@@ -143,7 +143,7 @@ static void Quantize(const Tensor *input, const float max_abs,
 
 float find_abs_max(const Tensor *input) {
   float max_abs = 0.f;
-  const float *x = input->data<const float>();
+  const float *x = input->data<float>();
   size_t remain = input->numel();
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
   size_t loop = remain >> 4;
diff --git a/src/operators/kernel/arm/tensor_array_read_write_kernel.cpp b/src/operators/kernel/arm/tensor_array_read_write_kernel.cpp
index 72fbb4cadbac4ef824692cfcaf55dd2d5c5a1166..bdf10574a82c920edfdfc870efe2f175ad7d7d07 100644
--- a/src/operators/kernel/arm/tensor_array_read_write_kernel.cpp
+++ b/src/operators/kernel/arm/tensor_array_read_write_kernel.cpp
@@ -28,7 +28,9 @@ void WriteToArrayKernel<CPU, float>::Compute(
     const WriteToArrayParam<CPU> &param) {
   int64_t offset = param.index_->data<int64_t>()[0];
   if (offset >= param.output_->size()) {
-    param.output_->resize(offset + 1);
+    while (param.output_->size() <= offset) {
+      param.output_->emplace_back();
+    }
   }
 
   framework::LoDTensor *out_tensor = &(param.output_->at(offset));
diff --git a/src/operators/kernel/arm/transpose2_kernel.cpp b/src/operators/kernel/arm/transpose2_kernel.cpp
index 6928df71e65dfe0aa28ceddda8f379f60fb65908..54c88015cb06728544258ddb47d6ce751130bc19 100644
--- a/src/operators/kernel/arm/transpose2_kernel.cpp
+++ b/src/operators/kernel/arm/transpose2_kernel.cpp
@@ -126,13 +126,13 @@ void Transpose2Kernel<CPU, float>::Compute(const Transpose2Param<CPU> &param) {
   const std::vector<int> &axis = param.Axis();
   bool shuffle_channel = IsShuffleChannel(axis);
   if (shuffle_channel) {
-    if (param.InputX()->type() == typeid(int8_t)) {
+    if (param.InputX()->type() == type_id<int8_t>().hash_code()) {
       ShuffleChannelCompute<int8_t>(param);
     } else {
       ShuffleChannelCompute<float>(param);
     }
   } else {
-    if (param.InputX()->type() == typeid(int8_t)) {
+    if (param.InputX()->type() == type_id<int8_t>().hash_code()) {
       Transpose2Compute<int8_t>(param);
     } else {
       Transpose2Compute<float>(param);
diff --git a/src/operators/kernel/arm/while_kernel.cpp b/src/operators/kernel/arm/while_kernel.cpp
index 63cd150ec977a5acf1fc24b1dd9bae75bf57b580..4794fd122cce141487598a19261fd3e0f66cfb27 100644
--- a/src/operators/kernel/arm/while_kernel.cpp
+++ b/src/operators/kernel/arm/while_kernel.cpp
@@ -35,6 +35,7 @@ class StepExecutor {
       auto op_handler = framework::OpRegistry<CPU>::CreateOp(
           op_desc->Type(), op_desc->GetInputs(), op_desc->GetOutputs(),
           op_desc->GetAttrMap(), scope_);
+      op_handler->Init();
       ops_of_block_[i] = op_handler;
     }
   }
diff --git a/src/operators/kernel/central-arm-func/mul_arm_func.h b/src/operators/kernel/central-arm-func/mul_arm_func.h
index 1f22ab98989644430e2484ca8d57fe2c4047e2f7..01d668021ba1affa4d827fdacbadbac7a66c794a 100644
--- a/src/operators/kernel/central-arm-func/mul_arm_func.h
+++ b/src/operators/kernel/central-arm-func/mul_arm_func.h
@@ -37,7 +37,7 @@ void MulCompute(const MulParam<CPU> &param) {
   if (out_dim.size() != 2) {
     out->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
   }
-  if (param.InputX()->type() == typeid(int8_t)) {
+  if (param.InputX()->type() == type_id<int8_t>().hash_code()) {
     out->mutable_data<int32_t>();
     math::MatMul<int8_t, int32_t>(x_matrix, false, y_matrix, false,
                                   static_cast<float>(1), out,
diff --git a/src/operators/kernel/central-arm-func/sum_arm_func.h b/src/operators/kernel/central-arm-func/sum_arm_func.h
index 36c7ac9694bde85fbf702ad8adf5ffda8744da1d..7d41c898db991c098c91e882fdb330f5ffb2b9bf 100644
--- a/src/operators/kernel/central-arm-func/sum_arm_func.h
+++ b/src/operators/kernel/central-arm-func/sum_arm_func.h
@@ -144,7 +144,7 @@ void SumCompute(const SumParam<CPU> &param) {
     }
   } else {
     PADDLE_MOBILE_THROW_EXCEPTION(
-        "Unexpected branch, output variable type is %s", outvar->Type().name());
+        "Unexpected branch, output variable type is %d", outvar->Type());
   }
 }
 }  // namespace operators
diff --git a/src/operators/kernel/fpga/V1/elementwise_add_kernel.cpp b/src/operators/kernel/fpga/V1/elementwise_add_kernel.cpp
index c549e5a6eee98f38f1806367054b925440e3ebf1..db4d2afbc1735ac7f733a9dea673c748bc6edc29 100644
--- a/src/operators/kernel/fpga/V1/elementwise_add_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/elementwise_add_kernel.cpp
@@ -25,7 +25,7 @@ template <>
 bool ElementwiseAddKernel<FPGA, float>::Init(ElementwiseAddParam<FPGA> *param) {
   auto *input_y = const_cast<LoDTensor *>(param->InputY());
   auto *out = param->Out();
-  if (input_y->type() != typeid(float)) {
+  if (input_y->type() != type_id<float>()) {
     paddle_mobile::fpga::ActivationType activation_enable =
         paddle_mobile::fpga::NONE;
     int16_t leaky_relu_negative_slope = 0;
@@ -62,11 +62,10 @@ bool ElementwiseAddKernel<FPGA, float>::Init(ElementwiseAddParam<FPGA> *param) {
     param->SetFpgaArgs(ewaddArgs);
   } else {
     param->float_input_x.Resize(param->InputX()->dims());
-    param->float_input_x.init(typeid(float));
+    param->float_input_x.init(type_id<float>().hash_code());
     fpga::format_fp32_ofm(&(param->float_input_x));
 
     param->float_out.Resize(param->InputX()->dims());
-    // param->float_out.init(typeid(float));
     param->float_out.mutable_data<float>(param->InputX()->dims());
     fpga::format_fp32_ofm(&(param->float_out));
 
@@ -118,7 +117,7 @@ template <>
 void ElementwiseAddKernel<FPGA, float>::Compute(
     const ElementwiseAddParam<FPGA> &param) {
   auto input_y = const_cast<LoDTensor *>(param.InputY());
-  if (input_y->type() != typeid(float)) {
+  if (input_y->type() != type_id<float>()) {
     fpga::ComputeFpgaEWAdd(param.FpgaArgs());
   } else {
     auto input_x = const_cast<LoDTensor *>(param.InputX());
diff --git a/src/operators/kernel/fpga/V1/elementwise_mul_kernel.cpp b/src/operators/kernel/fpga/V1/elementwise_mul_kernel.cpp
index e421ddb78ff4f1a0f0c51c985db9c26666001d03..d744ae2c07810ae89418641799a37ea978d14139 100644
--- a/src/operators/kernel/fpga/V1/elementwise_mul_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/elementwise_mul_kernel.cpp
@@ -27,11 +27,11 @@ struct MulFunctor {
 template <>
 bool ElementwiseMulKernel<FPGA, float>::Init(ElementwiseMulParam<FPGA> *param) {
   param->float_input_x.Resize(param->InputX()->dims());
-  param->float_input_x.init(typeid(float));
+  param->float_input_x.init(type_id<float>().hash_code());
   fpga::format_fp32_ofm(&(param->float_input_x));
 
   param->float_out.Resize(param->InputX()->dims());
-  param->float_out.init(typeid(float));
+  param->float_out.init(type_id<float>().hash_code());
   fpga::format_fp32_ofm(&(param->float_out));
 
   auto *out = param->Out();
diff --git a/src/operators/kernel/fpga/V1/feed_kernel.cpp b/src/operators/kernel/fpga/V1/feed_kernel.cpp
index f57c517bb00b8d676beaabf24c662efcbe752aeb..28559b2b4bb96404febf5cf65a75e264166df20f 100644
--- a/src/operators/kernel/fpga/V1/feed_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/feed_kernel.cpp
@@ -23,7 +23,7 @@ bool FeedKernel<FPGA, float>::Init(FeedParam<FPGA> *param) {
   int col = param->Col();
   DLOG << "col = " << col;
   auto input = const_cast<LoDTensor *>(&param->InputX()->at(col));
-  input->init(typeid(float));
+  input->init(type_id<float>().hash_code());
   input->Resize(output->dims());
 
   if (output->dims().size() != 4) {
@@ -39,12 +39,12 @@ void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
   auto output = param.Out();
   int col = param.Col();
   auto input = const_cast<LoDTensor *>(&param.InputX()->at(col));
-  std::type_index input_type = input->type();
+  kTypeId_t input_type = input->type();
 
-  if (input_type == typeid(float)) {
-    input->init(typeid(float));
-  } else {  // input_type == typeid(int8_t)
-    input->init(typeid(int8_t));
+  if (input_type == type_id<float>()) {
+    input->init(type_id<float>().hash_code());
+  } else {
+    input->init(type_id<int8_t>().hash_code());
   }
   input->Resize(output->dims());
 
@@ -62,7 +62,7 @@ void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
   fpga::format_image(input);
   auto output_ptr = output->data<half>();
   fpga::BypassArgs args = {fpga::DATA_TYPE_FP32};
-  if (input_type == typeid(float)) {
+  if (input_type == type_id<float>()) {
     auto input_ptr = input->data<float>();
     auto external_ptr = reinterpret_cast<float *>(input->external_data);
     float *p_data = external_ptr == nullptr ? input_ptr : external_ptr;
@@ -81,7 +81,7 @@ void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
     args.output.scale_address = output->scale;
     fpga::PerformBypass(args);
     input->external_data = nullptr;
-  } else {  // input_type == typeid(int8_t)
+  } else {
     auto input_ptr = input->data<int8_t>();
     auto external_ptr = reinterpret_cast<int8_t *>(input->external_data);
     int8_t *p_data = external_ptr == nullptr ? input_ptr : external_ptr;
diff --git a/src/operators/kernel/fpga/V1/fetch_kernel.cpp b/src/operators/kernel/fpga/V1/fetch_kernel.cpp
index fad1e77643a017659bea3c27d4475aea2c00787d..87ede2af1ab2fa3225c0cd3e75c3fe0c8c8fb509 100644
--- a/src/operators/kernel/fpga/V1/fetch_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/fetch_kernel.cpp
@@ -21,10 +21,10 @@ bool FetchKernel<FPGA, float>::Init(FetchParam<FPGA> *param) {
   int col = param->Col();
   DLOG << "col = " << col;
   auto output = &(param->Out()->at(col));
-  if (input->type() == typeid(float)) {
+  if (input->type() == type_id<float>()) {
     return true;
   }
-  output->init(typeid(float));
+  output->init(type_id<float>().hash_code());
   output->Resize(input->dims());
   fpga::format_fp32_ofm(output);
   int outC = 1;
@@ -78,7 +78,7 @@ void FetchKernel<FPGA, float>::Compute(const FetchParam<FPGA> &param) {
   auto input = const_cast<LoDTensor *>(param.InputX());
   int col = param.Col();
   auto output = &param.Out()->at(col);
-  if (input->type() == typeid(float)) {
+  if (input->type() == type_id<float>()) {
     output->ShareDataWith(*input);
     return;
   }
diff --git a/src/operators/kernel/fpga/V1/pool_kernel.cpp b/src/operators/kernel/fpga/V1/pool_kernel.cpp
index 994fa151621956aa791d36cc0f4cd829dc88f3d1..7c8dba1696ecc15ba9748aabf1973445d23de95c 100644
--- a/src/operators/kernel/fpga/V1/pool_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/pool_kernel.cpp
@@ -28,7 +28,7 @@ bool PoolKernel<FPGA, float>::Init(PoolParam<FPGA> *param) {
   vector<int> paddings = param->Paddings();
   std::string pooling_type = param->PoolingType();
 
-  if (input->type() == typeid(float)) {
+  if (input->type() == type_id<float>()) {
     int channels = input->dims()[1];
     int height = input->dims()[2];
     int width = input->dims()[3];
@@ -70,7 +70,7 @@ template <>
 void PoolKernel<FPGA, float>::Compute(const PoolParam<FPGA> &param) {
   auto *input = const_cast<LoDTensor *>(param.Input());
 
-  if (input->type() == typeid(float)) {
+  if (input->type() == type_id<float>()) {
     auto *output = param.Output();
     auto in = input->data<float>();
     auto N = input->dims()[0];
diff --git a/src/operators/kernel/fpga/V1/proposal_kernel.cpp b/src/operators/kernel/fpga/V1/proposal_kernel.cpp
index 09c135ff5cdff0755dc41d96f90d4a3e3b345c27..bd6703bb81f1a4b70f2a3406b312160116ad38f5 100644
--- a/src/operators/kernel/fpga/V1/proposal_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/proposal_kernel.cpp
@@ -37,11 +37,11 @@ bool ProposalKernel<FPGA, float>::Init(ProposalParam<FPGA> *param) {
 
   param->float_bbox = std::make_shared<Tensor>();
   param->float_bbox->Resize(param->bbox_deltas_->dims());
-  param->float_bbox->init(typeid(float));
+  param->float_bbox->init(type_id<float>().hash_code());
   fpga::format_fp32_ofm(param->float_bbox.get());
   param->float_score = std::make_shared<Tensor>();
   param->float_score->Resize(param->scores_->dims());
-  param->float_score->init(typeid(float));
+  param->float_score->init(type_id<float>().hash_code());
   fpga::format_fp32_ofm(param->float_score.get());
 
   auto input = param->bbox_deltas_;
@@ -437,7 +437,6 @@ void ProposalKernel<FPGA, float>::Compute(const ProposalParam<FPGA> &param) {
   bbox_height = (uint32_t)(input_bbox->dims()[2]);
   bbox_width = (uint32_t)(input_bbox->dims()[3]);
 
-  // score_tmp->init(typeid(half));
   std::shared_ptr<Tensor> score_tmp = std::make_shared<Tensor>();
   score_tmp->Resize(param.scores_->dims());
   score_tmp->mutable_data<half>();
diff --git a/src/operators/kernel/fpga/V1/slice_kernel.cpp b/src/operators/kernel/fpga/V1/slice_kernel.cpp
index 39e5c64b34c2a6b0629a7f2ab07a8683e9c45edd..2fd6ef542e72ab3aceedee22f79bd591a00b7712 100644
--- a/src/operators/kernel/fpga/V1/slice_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/slice_kernel.cpp
@@ -25,7 +25,7 @@ bool SliceKernel<FPGA, float>::Init(SliceParam<FPGA>* param) {
   fpga::format_fp16_ofm(output);
   DLOG << "input: " << param->input_;
   DLOG << "output: " << param->output_;
-  if (param->input_->type() != typeid(half)) {
+  if (param->input_->type() != type_id<half>()) {
     DLOG << "wrong type";
   }
   return true;
diff --git a/src/operators/kernel/fpga/V1/softmax_kernel.cpp b/src/operators/kernel/fpga/V1/softmax_kernel.cpp
index 5537565bc2a4dc7563148617daf47eaa9a50ba91..ba86787c646c3fc67992c76f5ce34efdcb5bbe4a 100644
--- a/src/operators/kernel/fpga/V1/softmax_kernel.cpp
+++ b/src/operators/kernel/fpga/V1/softmax_kernel.cpp
@@ -26,7 +26,7 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   auto dims = framework::vectorize(input->dims());
   half *input_ptr;
   auto out = param->Out();
-  if (input->type() == typeid(float)) {
+  if (input->type() == type_id<float>()) {
     out->Resize(framework::make_ddim(dims));
     out->mutable_data<float>(framework::make_ddim(dims));
   } else {
@@ -50,7 +50,7 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   if (channel != 2) {  // Use CPU
     out->Resize(framework::make_ddim(dims));
     out->mutable_data<float>(framework::make_ddim(dims));
-    float_input->init(typeid(float));
+    float_input->init(type_id<float>().hash_code());
     float_input->mutable_data<float>(framework::make_ddim(dims));
     //  fpga::format_fp32_ofm(float_input);
     // fpga::format_fp32_ofm(out);
@@ -91,7 +91,7 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
 template <>
 void SoftmaxKernel<FPGA, float>::Compute(const SoftmaxParam<FPGA> &param) {
   auto *in_x = (param.InputX());
-  if (in_x->type() == typeid(half)) {
+  if (in_x->type() == type_id<half>()) {
     fpga::PerformBypass(param.FpgaArgs());
     if (param.FpgaArgs().output.activation.activation_type != fpga::SOFTMAX) {
       Tensor *out = param.Out();
diff --git a/src/operators/math/gemm/cblas.cc b/src/operators/math/gemm/cblas.cc
index 058b61f1114c664e8e5ac3ae31f85e4186b9fd8a..4428826552a6a05fc0ee64f173e81d60405d1249 100644
--- a/src/operators/math/gemm/cblas.cc
+++ b/src/operators/math/gemm/cblas.cc
@@ -14,8 +14,6 @@ limitations under the License. */
 
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
 
-#pragma once
-
 #include "operators/math/gemm/cblas.h"
 #include "operators/math/gemm/executor.h"
 #include "operators/math/gemm/strategy.h"
diff --git a/src/operators/one_hot_op.cpp b/src/operators/one_hot_op.cpp
index 396f55318a80dc3177a0ae5f4b151eaec7806a6d..4367c2161eed977f2805a25b79a9ce126f0d5953 100644
--- a/src/operators/one_hot_op.cpp
+++ b/src/operators/one_hot_op.cpp
@@ -14,8 +14,6 @@ limitations under the License. */
 
 #ifdef ONE_HOT_OP
 
-#pragma once
-
 #include "operators/one_hot_op.h"
 
 namespace paddle_mobile {
diff --git a/src/operators/op_param.h b/src/operators/op_param.h
index 8679174e4cd5e3efe3abb0d8a7eff0a0c6290516..e144bf623ff5b28803bb454dbb11fcc810e9ca94 100644
--- a/src/operators/op_param.h
+++ b/src/operators/op_param.h
@@ -19,6 +19,7 @@ limitations under the License. */
 #include "common/log.h"
 #include "common/type_define.h"
 #include "common/types.h"
+#include "framework/attribute.h"
 #include "framework/lod_tensor.h"
 #include "framework/scope.h"
 #include "framework/tensor.h"
diff --git a/test/fpga/test_marker_api.cpp b/test/fpga/test_marker_api.cpp
index 48038e7b8fa9213e983f3ab25b17c6e8b83b84db..29cf6561df1afae1fae5494bcb8fa6606d3999df 100644
--- a/test/fpga/test_marker_api.cpp
+++ b/test/fpga/test_marker_api.cpp
@@ -104,7 +104,7 @@ void dump_stride_float(std::string filename,
 
 void dump_stride(std::string filename,
                  paddle_mobile::PaddleTensor input_tensor) {
-  if (input_tensor.dtypeid == typeid(float)) {
+  if (input_tensor.dtypeid == type_id<float>().hash_code()) {
     dump_stride_float(filename, input_tensor);
   } else {
     std::cout << "only support dumping float data" << std::endl;
@@ -156,13 +156,13 @@ int main() {
 
   std::cout << "Finishing initializing data" << std::endl;
   struct PaddleTensor t_img_info, t_img;
-  t_img_info.dtypeid = typeid(float);
+  t_img_info.dtypeid = type_id<float>().hash_code();
   t_img_info.layout = LAYOUT_HWC;
   t_img_info.shape = std::vector<int>({1, 3});
   t_img_info.name = "Image information";
   t_img_info.data.Reset(img_info, 3 * sizeof(float));
 
-  t_img.dtypeid = typeid(float);
+  t_img.dtypeid = type_id<float>().hash_code();
   // quantize(&img, img_length);
   // t_img.dtypeid = typeid(int8_t);
   t_img.layout = LAYOUT_HWC;
@@ -209,7 +209,7 @@ int main() {
     std::cout << "Finishing initializing data" << std::endl;
     struct PaddleTensor t_img1;
 
-    t_img1.dtypeid = typeid(float);
+    t_img1.dtypeid = type_id<float>().hash_code();
     t_img1.layout = LAYOUT_HWC;
     t_img1.shape = std::vector<int>({1, 14, 14, 144});
     t_img1.name = "Image information";
diff --git a/test/fpga/test_mobilenet_api.cpp b/test/fpga/test_mobilenet_api.cpp
index 4b372773937722942b70c584dda1eeb22339841f..09392e9d38f8a0312715da2aae74e1e039c2452e 100644
--- a/test/fpga/test_mobilenet_api.cpp
+++ b/test/fpga/test_mobilenet_api.cpp
@@ -96,7 +96,7 @@ void dump_stride_float(std::string filename, PaddleTensor input_tensor) {
 }
 
 void dump_stride(std::string filename, PaddleTensor input_tensor) {
-  if (input_tensor.dtypeid == typeid(float)) {
+  if (input_tensor.dtypeid == type_id<float>().hash_code()) {
     dump_stride_float(filename, input_tensor);
   } else {
     std::cout << "only support dumping float data" << std::endl;
@@ -131,7 +131,7 @@ int main() {
   std::cout << "Finishing initializing data" << std::endl;
   struct PaddleTensor t_img;
   t_img.dtype = FLOAT32;
-  t_img.dtypeid = typeid(float);
+  t_img.dtypeid = type_id<float>().hash_code();
   // quantize(&img, img_length);
   // t_img.dtype = INT8;
   // t_img.dtypeid = typeid(int8_t);
diff --git a/test/fpga/test_rfcn_api.cpp b/test/fpga/test_rfcn_api.cpp
index c62eeda9929ecaf1a3aca624301fd274872ea919..e86743cc7e7b8fdb48a034aebd5a53224674b6f5 100644
--- a/test/fpga/test_rfcn_api.cpp
+++ b/test/fpga/test_rfcn_api.cpp
@@ -20,8 +20,8 @@ limitations under the License. */
 #include <iostream>
 #include "../../src/io/paddle_inference_api.h"
 
-using namespace paddle_mobile;
-using namespace paddle_mobile::fpga;
+using namespace paddle_mobile;        // NOLINT
+using namespace paddle_mobile::fpga;  // NOLINT
 
 static const char *g_image = "../models/rfcn/data.bin";
 static const char *g_model = "../models/rfcn/model";
@@ -86,7 +86,7 @@ int main() {
 
   struct PaddleTensor t_img1;
 
-  t_img1.dtypeid = typeid(float);
+  t_img1.dtypeid = type_id<float>().hash_code();
   t_img1.layout = LAYOUT_HWC;
   t_img1.shape = std::vector<int>({1, 224, 224, 3});
   t_img1.name = "Image information";
@@ -117,13 +117,13 @@ int main() {
 
   std::cout << "Finishing initializing data" << std::endl;
   struct PaddleTensor t_img_info, t_img;
-  t_img.dtypeid = typeid(float);
+  t_img.dtypeid = type_id<float>().hash_code();
   t_img_info.layout = LAYOUT_HWC;
   t_img_info.shape = std::vector<int>({1, 3});
   t_img_info.name = "Image information";
   t_img_info.data.Reset(img_info, 3 * sizeof(float));
 
-  t_img.dtypeid = typeid(float);
+  t_img.dtypeid = type_id<float>().hash_code();
   t_img.layout = LAYOUT_HWC;
   t_img.shape = std::vector<int>({1, 432, 1280, 3});
   t_img.name = "Image information";
diff --git a/test/fpga/test_yolo_api.cpp b/test/fpga/test_yolo_api.cpp
index 4ef890506eb1c40638242b9767267756a64da787..f8f1a48abca74b76f77e7f8d59371b7672a8b453 100644
--- a/test/fpga/test_yolo_api.cpp
+++ b/test/fpga/test_yolo_api.cpp
@@ -95,7 +95,7 @@ void dump_stride_float(std::string filename, PaddleTensor input_tensor) {
 }
 
 void dump_stride(std::string filename, PaddleTensor input_tensor) {
-  if (input_tensor.dtypeid == typeid(float)) {
+  if (input_tensor.dtypeid == type_id<float>().hash_code()) {
     dump_stride_float(filename, input_tensor);
   } else {
     std::cout << "only support dumping float data" << std::endl;
@@ -131,10 +131,10 @@ int main() {
   std::cout << "Finishing initializing data" << std::endl;
   struct PaddleTensor t_img;
   // t_img.dtype = FLOAT32;
-  // t_img.dtypeid = typeid(float);
+  // t_img.dtypeid = type_id<float>().hash_code();
   quantize(&img, img_length);
   t_img.dtype = INT8;
-  t_img.dtypeid = typeid(int8_t);
+  t_img.dtypeid = type_id<int8_t>().hash_code();
   t_img.layout = LAYOUT_HWC;
   t_img.shape = std::vector<int>({1, 256, 416, 3});
   t_img.name = "Image information";
diff --git a/test/operators/test_dequantize_op.cpp b/test/operators/test_dequantize_op.cpp
index 41390d74a92d9cab615244600079373e84fb31d2..981439c66fdf9902b4451a322099f2e885a1b396 100644
--- a/test/operators/test_dequantize_op.cpp
+++ b/test/operators/test_dequantize_op.cpp
@@ -19,7 +19,7 @@ limitations under the License. */
 namespace paddle_mobile {
 
 void dequantize(const Tensor* input, const float scale, Tensor* output) {
-  const int32_t* x = input->data<const int32_t>();
+  const int32_t* x = input->data<int32_t>();
   float* y = output->mutable_data<float>();
   size_t size = output->numel();
   for (size_t i = 0; i < size; ++i) {
diff --git a/test/operators/test_gru_op.cpp b/test/operators/test_gru_op.cpp
index 704a4df0294e46cd661d5b010398baa6fa40a740..b14c4642fd7638aed0473790ac9ca3bc18476160 100644
--- a/test/operators/test_gru_op.cpp
+++ b/test/operators/test_gru_op.cpp
@@ -20,8 +20,8 @@ namespace paddle_mobile {
 
 template <typename Itype, typename Otype>
 int TestGruOp(int in_channels, int out_channels, std::string opname) {
-  int input_c = in_channels;
-  int output_c = out_channels;
+  size_t input_c = in_channels;
+  size_t output_c = out_channels;
   paddle_mobile::framework::LoD lod{{0, input_c}};
   int batch_size = lod.size();
   framework::DDim input_shape = framework::make_ddim({input_c, output_c * 3});
diff --git a/test/operators/test_quantize_op.cpp b/test/operators/test_quantize_op.cpp
index 2f8e08806be426e3cb23804514a01b0ca44a0fe1..d8e72e9b1472d0de48143b37ee2a7fe48ad4e174 100644
--- a/test/operators/test_quantize_op.cpp
+++ b/test/operators/test_quantize_op.cpp
@@ -65,7 +65,7 @@ static void quantize(const Tensor *input, const float scale, Tensor *output) {
   int output_w = output->dims()[3];
   size_t input_spatial = input_h * input_w;
   size_t output_spatial = output_h * output_w;
-  const float *x = input->data<const float>();
+  const float *x = input->data<float>();
   int8_t *y = output->mutable_data<int8_t>();
 
   for (int nc = 0; nc < batch_size * channels; ++nc) {
@@ -81,7 +81,7 @@ static void quantize(const Tensor *input, const float scale, Tensor *output) {
 
 static float find_abs_max(const Tensor *input) {
   float max_abs = 0.f;
-  const float *x = input->data<const float>();
+  const float *x = input->data<float>();
   size_t size = input->numel();
   for (size_t i = 0; i < size; ++i) {
     float value = std::abs(x[i]);