refine wrap to support GPU test=develop (#1892)

mobile/src/io/paddle_mobile_wrap.cpp

@@ -49,167 +49,274 @@ DDim make_ddim(const std::vector<int64_t> &dims) {
 }
 
 // tensor class
+
 Tensor::Tensor(float *data, DDim ddim) {
   this->data_ = data;
   this->ddim_ = ddim;
 }
 
-template <typename T>
-float *Tensor::data() const {
-  return this->data_;
-}
+float *Tensor::data() const { return this->data_; }
 
 DDim Tensor::dims() const { return this->ddim_; }
 
 // net class
-template <typename Device>
-void Net<Device>::SetThreadNum(int threads) {
-  auto engine =
-      (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-  if (engine != nullptr) {
-    engine->SetThreadNum(threads);
+
+void Net::SetThreadNum(int threads) {
+  if (this->device_ == kCPU) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+    if (engine != nullptr) {
+      engine->SetThreadNum(threads);
+    }
   }
 }
 
-template <typename Device>
-PMStatus Net<Device>::Load(const std::string &dirname, const bool optimize,
-                           const bool quantification, const int batch_size,
-                           const bool lod_mode) {
-  auto engine =
-      (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-  if (engine != nullptr) {
-    paddle_mobile::PMStatus status =
-        engine->Load(dirname, false, false, 1, true);
-    return status == paddle_mobile::PMSuccess ? PMSuccess : PMUnKownError;
+void Net::SetCLPath(std::string path) {
+  if (this->device_ == kGPU_CL) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+    engine->SetCLPath(path);
   }
-  return PMUnKownError;
-}
-
-template <typename Device>
-PMStatus Net<Device>::Load(const std::string &model_path,
-                           const std::string &para_path, const bool optimize,
-                           const bool quantification, const int batch_size,
-                           const bool lod_mode) {
-  auto engine =
-      (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-  if (engine != nullptr) {
-    paddle_mobile::PMStatus status =
-        engine->Load(model_path, para_path, false, false, 1, true);
-    return status == paddle_mobile::PMSuccess ? PMSuccess : PMUnKownError;
+}
+
+bool Net::Load(const std::string &dirname, const bool optimize,
+               const bool quantification, const int batch_size,
+               const bool lod_mode) {
+  if (this->device_ == kCPU) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+    if (engine != nullptr) {
+      paddle_mobile::PMStatus status =
+          engine->Load(dirname, optimize, quantification, batch_size, lod_mode);
+      return status == paddle_mobile::PMSuccess;
+    }
+  } else if (this->device_ == kGPU_CL) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+    if (engine != nullptr) {
+      paddle_mobile::PMStatus status =
+          engine->Load(dirname, optimize, quantification, batch_size, lod_mode);
+      return status == paddle_mobile::PMSuccess;
+    }
   }
-  return PMUnKownError;
-}
-
-template <typename Device>
-bool Net<Device>::LoadCombinedMemory(size_t model_len, const uint8_t *model_buf,
-                                     size_t combined_params_len,
-                                     uint8_t *combined_params_buf,
-                                     bool optimize, bool quantification,
-                                     int batch_size, bool lod_mode) {
-  auto engine =
-      (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-  if (engine != nullptr) {
-    bool status =
-        engine->LoadCombinedMemory(model_len, model_buf, combined_params_len,
-                                   combined_params_buf, false, false, 1, true);
-    return status;
+  return false;
+}
+
+bool Net::Load(const std::string &model_path, const std::string &para_path,
+               const bool optimize, const bool quantification,
+               const int batch_size, const bool lod_mode) {
+  if (this->device_ == kCPU) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+    if (engine != nullptr) {
+      paddle_mobile::PMStatus status =
+          engine->Load(model_path, para_path, optimize, quantification,
+                       batch_size, lod_mode);
+      return status == paddle_mobile::PMSuccess;
+    }
+  } else if (this->device_ == kGPU_CL) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+    if (engine != nullptr) {
+      paddle_mobile::PMStatus status =
+          engine->Load(model_path, para_path, optimize, quantification,
+                       batch_size, lod_mode);
+      return status == paddle_mobile::PMSuccess;
+    }
   }
   return false;
 }
 
-template <typename Device>
-PMStatus Net<Device>::Predict(const Tensor &input) {
-  auto engine =
-      (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-  if (engine != nullptr) {
-    auto input_data = input.data<float>();
-    auto input_dims = input.dims();
-    std::vector<int64_t> input_dims_as_vector = input_dims.dims;
-    paddle_mobile::framework::Tensor input_inner(
-        input_data, paddle_mobile::framework::make_ddim(input_dims_as_vector));
-    paddle_mobile::PMStatus status = engine->Predict(input_inner);
-    return status == paddle_mobile::PMSuccess ? PMSuccess : PMUnKownError;
+bool Net::LoadCombinedMemory(size_t model_len, const uint8_t *model_buf,
+                             size_t combined_params_len,
+                             uint8_t *combined_params_buf, bool optimize,
+                             bool quantification, int batch_size,
+                             bool lod_mode) {
+  if (this->device_ == kCPU) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+    if (engine != nullptr) {
+      bool status = engine->LoadCombinedMemory(
+          model_len, model_buf, combined_params_len, combined_params_buf,
+          optimize, quantification, batch_size, lod_mode);
+      return status;
+    }
+  } else if (this->device_ == kGPU_CL) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+    if (engine != nullptr) {
+      bool status = engine->LoadCombinedMemory(
+          model_len, model_buf, combined_params_len, combined_params_buf,
+          optimize, quantification, batch_size, lod_mode);
+      return status;
+    }
   }
-  return PMUnKownError;
+  return false;
 }
 
-template <typename Device>
-std::vector<float> Net<Device>::Predict(const std::vector<float> &input,
-                                        const std::vector<int64_t> &dims) {
-  auto engine =
-      (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-  if (engine != nullptr) {
-    auto result = engine->Predict(input, dims);
-    return result;
+std::vector<float> Net::Predict(const std::vector<float> &input,
+                                const std::vector<int64_t> &dims) {
+  if (this->device_ == kCPU) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+    if (engine != nullptr) {
+      auto result = engine->Predict(input, dims);
+      return result;
+    }
+  } else if (this->device_ == kGPU_CL) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+    if (engine != nullptr) {
+      auto result = engine->Predict(input, dims);
+      return result;
+    }
   }
   return std::vector<float>();
 }
 
-template <typename Device>
-PMStatus Net<Device>::Predict() {
-  auto engine =
-      (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-  if (engine != nullptr) {
-    paddle_mobile::PMStatus status = engine->Predict();
-    return status == paddle_mobile::PMSuccess ? PMSuccess : PMUnKownError;
+bool Net::Predict() {
+  if (this->device_ == kCPU) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+    if (engine != nullptr) {
+      paddle_mobile::PMStatus status = engine->Predict();
+      return status == paddle_mobile::PMSuccess;
+    }
+  } else if (this->device_ == kGPU_CL) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+    if (engine != nullptr) {
+      paddle_mobile::PMStatus status = engine->Predict();
+      return status == paddle_mobile::PMSuccess;
+    }
+  }
+  return false;
+}
+
+bool Net::Predict(const Tensor &input) {
+  if (this->device_ == kCPU) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+    if (engine != nullptr) {
+      auto input_data = input.data();
+      auto input_dims = input.dims();
+      std::vector<int64_t> input_dims_as_vector = input_dims.dims;
+      paddle_mobile::framework::Tensor input_inner(
+          input_data,
+          paddle_mobile::framework::make_ddim(input_dims_as_vector));
+      paddle_mobile::PMStatus status = engine->Predict(input_inner);
+      return status == paddle_mobile::PMSuccess;
+    }
+  } else if (this->device_ == kGPU_CL) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+    if (engine != nullptr) {
+      auto input_data = input.data();
+      auto input_dims = input.dims();
+      std::vector<int64_t> input_dims_as_vector = input_dims.dims;
+      paddle_mobile::framework::Tensor input_inner(
+          input_data,
+          paddle_mobile::framework::make_ddim(input_dims_as_vector));
+      paddle_mobile::PMStatus status = engine->Predict(input_inner);
+      return status == paddle_mobile::PMSuccess;
+    }
   }
-  return PMUnKownError;
-}
-
-template <typename Device>
-void Net<Device>::Feed(const std::string &var_name, const Tensor &input) {
-  auto engine =
-      (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-  if (engine != nullptr) {
-    auto input_data = input.data<float>();
-    auto input_dims = input.dims();
-    std::vector<int64_t> input_dims_as_vector = input_dims.dims;
-    paddle_mobile::framework::Tensor input_inner(
-        input_data, paddle_mobile::framework::make_ddim(input_dims_as_vector));
-    engine->Feed(var_name, input_inner);
+  return false;
+}
+
+void Net::Feed(const std::string &var_name, const Tensor &input) {
+  if (this->device_ == kCPU) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+    if (engine != nullptr) {
+      auto input_data = input.data();
+      auto input_dims = input.dims();
+      std::vector<int64_t> input_dims_as_vector = input_dims.dims;
+      paddle_mobile::framework::Tensor input_inner(
+          input_data,
+          paddle_mobile::framework::make_ddim(input_dims_as_vector));
+      engine->Feed(var_name, input_inner);
+    }
+  } else if (this->device_ == kGPU_CL) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+    if (engine != nullptr) {
+      auto input_data = input.data();
+      auto input_dims = input.dims();
+      std::vector<int64_t> input_dims_as_vector = input_dims.dims;
+      paddle_mobile::framework::Tensor input_inner(
+          input_data,
+          paddle_mobile::framework::make_ddim(input_dims_as_vector));
+      engine->Feed(var_name, input_inner);
+    }
   }
 }
 
-template <typename Device>
-std::shared_ptr<Tensor> Net<Device>::Fetch(const std::string &var_name) {
-  auto engine =
-      (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-  if (engine != nullptr) {
-    auto output_inner = engine->Fetch(var_name);
-    auto ddim_inner = output_inner->dims();
-    std::vector<int64_t> ddim_as_vector;
-    for (int i = 0; i < ddim_inner.size(); i++) {
-      ddim_as_vector.push_back(ddim_inner[i]);
-    }
-    auto ddim = make_ddim(ddim_as_vector);
-    auto output_data = output_inner->data<float>();
-    std::shared_ptr<Tensor> ptr(new Tensor(output_data, ddim));
-    return ptr;
+std::shared_ptr<Tensor> Net::Fetch(const std::string &var_name) {
+  if (this->device_ == kCPU) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+    if (engine != nullptr) {
+      auto output_inner = engine->Fetch(var_name);
+      auto ddim_inner = output_inner->dims();
+      std::vector<int64_t> ddim_as_vector;
+      for (int i = 0; i < ddim_inner.size(); i++) {
+        ddim_as_vector.push_back(ddim_inner[i]);
+      }
+      auto ddim = make_ddim(ddim_as_vector);
+      auto output_data = output_inner->data<float>();
+      std::shared_ptr<Tensor> ptr(new Tensor(output_data, ddim));
+      return ptr;
+    }
+  } else if (this->device_ == kGPU_CL) {
+    auto engine =
+        (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+    if (engine != nullptr) {
+      auto output_inner = engine->Fetch(var_name);
+      auto ddim_inner = output_inner->dims();
+      std::vector<int64_t> ddim_as_vector;
+      for (int i = 0; i < ddim_inner.size(); i++) {
+        ddim_as_vector.push_back(ddim_inner[i]);
+      }
+      auto ddim = make_ddim(ddim_as_vector);
+      auto output_data = output_inner->data<float>();
+      std::shared_ptr<Tensor> ptr(new Tensor(output_data, ddim));
+      return ptr;
+    }
   }
   return nullptr;
 }
 
-template <typename Device>
-Net<Device>::Net() {
+Net::Net(DeviceTypeEnum device) {
   if (this->engine_ == nullptr) {
     PaddleMobileConfigInternal config;
-    this->engine_ = new paddle_mobile::PaddleMobile<paddle_mobile::CPU>(config);
+    this->device_ = device;
+    if (this->device_ == kCPU) {
+      this->engine_ =
+          new paddle_mobile::PaddleMobile<paddle_mobile::CPU>(config);
+    } else if (this->device_ == kGPU_CL) {
+      this->engine_ =
+          new paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL>(config);
+    }
   }
 }
 
-template <typename Device>
-Net<Device>::~Net() {
+Net::~Net() {
   if (this->engine_ != nullptr) {
-    auto engine =
-        (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
-    delete engine;
-    this->engine_ = nullptr;
+    if (this->device_ == kCPU) {
+      auto engine =
+          (paddle_mobile::PaddleMobile<paddle_mobile::CPU> *)this->engine_;
+      delete engine;
+      this->engine_ = nullptr;
+    } else if (this->device_ == kGPU_CL) {
+      auto engine =
+          (paddle_mobile::PaddleMobile<paddle_mobile::GPU_CL> *)this->engine_;
+      delete engine;
+      this->engine_ = nullptr;
+    }
   }
 }
 
-template class Net<CPU>;
-template float *Tensor::data<float>() const;
-
 #endif
 
 }  // namespace wrap

mobile/src/io/paddle_mobile_wrap.h

@@ -28,84 +28,67 @@ namespace wrap {
 #ifndef PADDLE_MOBILE_FPGA
 
 // device type
-enum DeviceTypeEnum {
-  kINVALID = -1,
+__attribute__((__visibility__("default"))) enum DeviceTypeEnum {
   kCPU = 0,
-  kFPGA = 1,
-  kGPU_MALI = 2,
-  kGPU_CL = 3
+  kGPU_CL = 1
 };
 
-template <DeviceTypeEnum T>
-struct DeviceType {};
-
-typedef DeviceType<kCPU> CPU;
-typedef DeviceType<kFPGA> FPGA;
-typedef DeviceType<kGPU_MALI> GPU_MALI;
-typedef DeviceType<kGPU_CL> GPU_CL;
-
 // ddim class
 class DDim {
  public:
-  int size();
-  int64_t &operator[](int idx);
-  int64_t operator[](int idx) const;
+  __attribute__((__visibility__("default"))) int size();
+  __attribute__((__visibility__("default"))) int64_t &operator[](int idx);
+  __attribute__((__visibility__("default"))) int64_t operator[](int idx) const;
 
-  std::vector<int64_t> dims;
+  __attribute__((__visibility__("default"))) std::vector<int64_t> dims;
 };
-DDim make_ddim(const std::vector<int64_t> &dims);
+__attribute__((__visibility__("default"))) DDim make_ddim(
+    const std::vector<int64_t> &dims);
 
 // tensor class
 class Tensor {
  public:
-  Tensor(float *data, DDim ddim);
+  __attribute__((__visibility__("default"))) Tensor(float *data, DDim ddim);
 
-  template <typename T>
-  float *data() const;
-  DDim dims() const;
+  __attribute__((__visibility__("default"))) float *data() const;
+  __attribute__((__visibility__("default"))) DDim dims() const;
 
+ private:
   float *data_;
   DDim ddim_;
 };
 
-// pm status
-enum PMStatus {
-  PMSuccess = 0xFF,        /*!< No errors */
-  PMNotInitialized = 0x01, /*!< Data not initialized. */
-  PMInvalidValue = 0x02,   /*!< Incorrect variable value. */
-  PMMemAllocFailed = 0x03, /*!< Memory allocation error. */
-  PMUnKownError = 0x04,    /*!< Unknown error. */
-  PMOutOfAuthority = 0x05, /*!< Try to modified data not your own*/
-  PMOutOfMem = 0x06,       /*!< OOM error*/
-  PMUnImplError = 0x07,    /*!< Unimplement error. */
-  PMWrongDevice = 0x08     /*!< un-correct device. */
-};
-
 // net class
-template <typename Device>
 class Net {
  public:
-  Net();
-  ~Net();
-  void SetThreadNum(int thread_num);
-  PMStatus Load(const std::string &dirname, const bool optimize = false,
-                const bool quantification = false, const int batch_size = 1,
-                const bool lod_mode = false);
-  PMStatus Load(const std::string &model_path, const std::string &para_path,
-                const bool optimize = false, const bool quantification = false,
-                const int batch_size = 1, const bool lod_mode = false);
-  bool LoadCombinedMemory(size_t model_len, const uint8_t *model_buf,
-                          size_t combined_params_len,
-                          uint8_t *combined_params_buf, bool optimize = false,
-                          bool quantification = false, int batch_size = 1,
-                          bool lod_mode = false);
-  PMStatus Predict(const Tensor &input);
-  std::vector<float> Predict(const std::vector<float> &input,
-                             const std::vector<int64_t> &dims);
-  PMStatus Predict();
-  void Feed(const std::string &var_name, const Tensor &input);
-  std::shared_ptr<Tensor> Fetch(const std::string &var_name);
+  __attribute__((__visibility__("default"))) Net(DeviceTypeEnum device);
+  __attribute__((__visibility__("default"))) ~Net();
+  __attribute__((__visibility__("default"))) void SetThreadNum(int thread_num);
+  __attribute__((__visibility__("default"))) void SetCLPath(std::string path);
+  __attribute__((__visibility__("default"))) bool Load(
+      const std::string &dirname, const bool optimize = false,
+      const bool quantification = false, const int batch_size = 1,
+      const bool lod_mode = false);
+  __attribute__((__visibility__("default"))) bool Load(
+      const std::string &model_path, const std::string &para_path,
+      const bool optimize = false, const bool quantification = false,
+      const int batch_size = 1, const bool lod_mode = false);
+  __attribute__((__visibility__("default"))) bool LoadCombinedMemory(
+      size_t model_len, const uint8_t *model_buf, size_t combined_params_len,
+      uint8_t *combined_params_buf, bool optimize = false,
+      bool quantification = false, int batch_size = 1, bool lod_mode = false);
+  __attribute__((__visibility__("default"))) std::vector<float> Predict(
+      const std::vector<float> &input, const std::vector<int64_t> &dims);
+  __attribute__((__visibility__("default"))) bool Predict();
+  __attribute__((__visibility__("default"))) bool Predict(const Tensor &input);
+  __attribute__((__visibility__("default"))) void Feed(
+      const std::string &var_name, const Tensor &input);
+  __attribute__((__visibility__("default"))) std::shared_ptr<Tensor> Fetch(
+      const std::string &var_name);
+
+ private:
   void *engine_ = nullptr;
+  DeviceTypeEnum device_;
 };
 
 #endif

mobile/src/protobuf-c/protobuf-c.c → mobile/src/protobuf-c/protobuf-c.cpp

@@ -191,7 +191,8 @@ void PaddleMobile__Framework__protobuf_c_buffer_simple_append(
 
     if (allocator == NULL) allocator = &protobuf_c__allocator;
     while (new_alloced < new_len) new_alloced += new_alloced;
-    new_data = PaddleMobile__Framework__do_alloc(allocator, new_alloced);
+    new_data =
+        (uint8_t *)PaddleMobile__Framework__do_alloc(allocator, new_alloced);
     if (!new_data) return;
     memcpy(new_data, simp->data, simp->len);
     if (simp->must_free_data)
@@ -905,7 +906,7 @@ static size_t PaddleMobile__Framework__parse_tag_and_wiretype(
   unsigned shift = 4;
   unsigned rv;
 
-  *wiretype_out = data[0] & 7;
+  *wiretype_out = (PaddleMobile__Framework__ProtobufCWireType)(data[0] & 7);
   if ((data[0] & 0x80) == 0) {
     *tag_out = tag;
     return 1;
@@ -1013,7 +1014,7 @@ static protobuf_c_boolean PaddleMobile__Framework__merge_messages(
                   fields[i].type);
           uint8_t *new_field;
 
-          new_field = PaddleMobile__Framework__do_alloc(
+          new_field = (uint8_t *)PaddleMobile__Framework__do_alloc(
               allocator, (*n_earlier + *n_latter) * el_size);
           if (!new_field) return FALSE;
 
@@ -1102,7 +1103,7 @@ static protobuf_c_boolean PaddleMobile__Framework__merge_messages(
         case PROTOBUF_C_TYPE_STRING: {
           char *e_str = *(char **)earlier_elem;
           char *l_str = *(char **)latter_elem;
-          const char *d_str = def_val;
+          const char *d_str = (const char *)def_val;
 
           need_to_merge = e_str != d_str && l_str == d_str;
           break;
@@ -1286,7 +1287,7 @@ static protobuf_c_boolean PaddleMobile__Framework__parse_required_member(
   unsigned len = scanned_member->len;
   const uint8_t *data = scanned_member->data;
   PaddleMobile__Framework__ProtobufCWireType wire_type =
-      scanned_member->wire_type;
+      (PaddleMobile__Framework__ProtobufCWireType)scanned_member->wire_type;
 
   switch (scanned_member->field->type) {
     case PROTOBUF_C_TYPE_ENUM:
@@ -1330,36 +1331,40 @@ static protobuf_c_boolean PaddleMobile__Framework__parse_required_member(
           PaddleMobile__Framework__parse_boolean(len, data);
       return TRUE;
     case PROTOBUF_C_TYPE_STRING: {
-      char **pstr = member;
+      char **pstr = (char **)member;
       unsigned pref_len = scanned_member->length_prefix_len;
 
       if (wire_type != PROTOBUF_C_WIRE_TYPE_LENGTH_PREFIXED) return FALSE;
 
       if (maybe_clear && *pstr != NULL) {
-        const char *def = scanned_member->field->default_value;
+        const char *def = (const char *)scanned_member->field->default_value;
         if (*pstr != NULL && *pstr != def)
           PaddleMobile__Framework__do_free(allocator, *pstr);
       }
-      *pstr = PaddleMobile__Framework__do_alloc(allocator, len - pref_len + 1);
+      *pstr = (char *)PaddleMobile__Framework__do_alloc(allocator,
+                                                        len - pref_len + 1);
       if (*pstr == NULL) return FALSE;
       memcpy(*pstr, data + pref_len, len - pref_len);
       (*pstr)[len - pref_len] = 0;
       return TRUE;
     }
     case PROTOBUF_C_TYPE_BYTES: {
-      PaddleMobile__Framework__ProtobufCBinaryData *bd = member;
+      PaddleMobile__Framework__ProtobufCBinaryData *bd =
+          (PaddleMobile__Framework__ProtobufCBinaryData *)member;
       const PaddleMobile__Framework__ProtobufCBinaryData *def_bd;
       unsigned pref_len = scanned_member->length_prefix_len;
 
       if (wire_type != PROTOBUF_C_WIRE_TYPE_LENGTH_PREFIXED) return FALSE;
 
-      def_bd = scanned_member->field->default_value;
+      def_bd = (const PaddleMobile__Framework__ProtobufCBinaryData *)
+                   scanned_member->field->default_value;
       if (maybe_clear && bd->data != NULL &&
           (def_bd == NULL || bd->data != def_bd->data)) {
         PaddleMobile__Framework__do_free(allocator, bd->data);
       }
       if (len - pref_len > 0) {
-        bd->data = PaddleMobile__Framework__do_alloc(allocator, len - pref_len);
+        bd->data = (uint8_t *)PaddleMobile__Framework__do_alloc(allocator,
+                                                                len - pref_len);
         if (bd->data == NULL) return FALSE;
         memcpy(bd->data, data + pref_len, len - pref_len);
       } else {
@@ -1369,7 +1374,8 @@ static protobuf_c_boolean PaddleMobile__Framework__parse_required_member(
       return TRUE;
     }
     case PROTOBUF_C_TYPE_MESSAGE: {
-      PaddleMobile__Framework__ProtobufCMessage **pmessage = member;
+      PaddleMobile__Framework__ProtobufCMessage **pmessage =
+          (PaddleMobile__Framework__ProtobufCMessage **)member;
       PaddleMobile__Framework__ProtobufCMessage *subm;
       const PaddleMobile__Framework__ProtobufCMessage *def_mess;
       protobuf_c_boolean merge_successful = TRUE;
@@ -1377,10 +1383,12 @@ static protobuf_c_boolean PaddleMobile__Framework__parse_required_member(
 
       if (wire_type != PROTOBUF_C_WIRE_TYPE_LENGTH_PREFIXED) return FALSE;
 
-      def_mess = scanned_member->field->default_value;
+      def_mess = (const PaddleMobile__Framework__ProtobufCMessage *)
+                     scanned_member->field->default_value;
       subm = PaddleMobile__Framework__protobuf_c_message_unpack(
-          scanned_member->field->descriptor, allocator, len - pref_len,
-          data + pref_len);
+          (const PaddleMobile__Framework__ProtobufCMessageDescriptor *)
+              scanned_member->field->descriptor,
+          allocator, len - pref_len, data + pref_len);
 
       if (maybe_clear && *pmessage != NULL && *pmessage != def_mess) {
         if (subm != NULL)
@@ -1418,25 +1426,29 @@ static protobuf_c_boolean PaddleMobile__Framework__parse_oneof_member(
 
     switch (old_field->type) {
       case PROTOBUF_C_TYPE_STRING: {
-        char **pstr = member;
-        const char *def = old_field->default_value;
+        char **pstr = (char **)member;
+        const char *def = (const char *)old_field->default_value;
         if (*pstr != NULL && *pstr != def)
           PaddleMobile__Framework__do_free(allocator, *pstr);
         break;
       }
       case PROTOBUF_C_TYPE_BYTES: {
-        PaddleMobile__Framework__ProtobufCBinaryData *bd = member;
+        PaddleMobile__Framework__ProtobufCBinaryData *bd =
+            (PaddleMobile__Framework__ProtobufCBinaryData *)member;
         const PaddleMobile__Framework__ProtobufCBinaryData *def_bd =
-            old_field->default_value;
+            (const PaddleMobile__Framework__ProtobufCBinaryData *)
+                old_field->default_value;
         if (bd->data != NULL && (def_bd == NULL || bd->data != def_bd->data)) {
           PaddleMobile__Framework__do_free(allocator, bd->data);
         }
         break;
       }
       case PROTOBUF_C_TYPE_MESSAGE: {
-        PaddleMobile__Framework__ProtobufCMessage **pmessage = member;
+        PaddleMobile__Framework__ProtobufCMessage **pmessage =
+            (PaddleMobile__Framework__ProtobufCMessage **)member;
         const PaddleMobile__Framework__ProtobufCMessage *def_mess =
-            old_field->default_value;
+            (const PaddleMobile__Framework__ProtobufCMessage *)
+                old_field->default_value;
         if (*pmessage != NULL && *pmessage != def_mess)
           PaddleMobile__Framework__protobuf_c_message_free_unpacked(*pmessage,
                                                                     allocator);
@@ -1651,10 +1663,11 @@ static protobuf_c_boolean PaddleMobile__Framework__parse_member(
     PaddleMobile__Framework__ProtobufCMessageUnknownField *ufield =
         message->unknown_fields + (message->n_unknown_fields++);
     ufield->tag = scanned_member->tag;
-    ufield->wire_type = scanned_member->wire_type;
+    ufield->wire_type =
+        (PaddleMobile__Framework__ProtobufCWireType)scanned_member->wire_type;
     ufield->len = scanned_member->len;
-    ufield->data =
-        PaddleMobile__Framework__do_alloc(allocator, scanned_member->len);
+    ufield->data = (uint8_t *)PaddleMobile__Framework__do_alloc(
+        allocator, scanned_member->len);
     if (ufield->data == NULL) return FALSE;
     memcpy(ufield->data, scanned_member->data, ufield->len);
     return TRUE;
@@ -1810,13 +1823,14 @@ PaddleMobile__Framework__protobuf_c_message_unpack(
 
   if (allocator == NULL) allocator = &protobuf_c__allocator;
 
-  rv = PaddleMobile__Framework__do_alloc(allocator, desc->sizeof_message);
+  rv = (PaddleMobile__Framework__ProtobufCMessage *)
+      PaddleMobile__Framework__do_alloc(allocator, desc->sizeof_message);
   if (!rv) return (NULL);
   scanned_member_slabs[0] = first_member_slab;
 
   required_fields_bitmap_len = (desc->n_fields + 7) / 8;
   if (required_fields_bitmap_len > sizeof(required_fields_bitmap_stack)) {
-    required_fields_bitmap = PaddleMobile__Framework__do_alloc(
+    required_fields_bitmap = (unsigned char *)PaddleMobile__Framework__do_alloc(
         allocator, required_fields_bitmap_len);
     if (!required_fields_bitmap) {
       PaddleMobile__Framework__do_free(allocator, rv);
@@ -1944,7 +1958,7 @@ PaddleMobile__Framework__protobuf_c_message_unpack(
       size = sizeof(ScannedMember)
              << (which_slab + FIRST_SCANNED_MEMBER_SLAB_SIZE_LOG2);
       scanned_member_slabs[which_slab] =
-          PaddleMobile__Framework__do_alloc(allocator, size);
+          (ScannedMember *)PaddleMobile__Framework__do_alloc(allocator, size);
       if (scanned_member_slabs[which_slab] == NULL)
         goto error_cleanup_during_scan;
     }
@@ -2012,10 +2026,13 @@ PaddleMobile__Framework__protobuf_c_message_unpack(
 
   /* allocate space for unknown fields */
   if (n_unknown) {
-    rv->unknown_fields = PaddleMobile__Framework__do_alloc(
-        allocator,
-        n_unknown *
-            sizeof(PaddleMobile__Framework__ProtobufCMessageUnknownField));
+    rv->unknown_fields =
+        (PaddleMobile__Framework__ProtobufCMessageUnknownField *)
+            PaddleMobile__Framework__do_alloc(
+                allocator,
+                n_unknown *
+                    sizeof(
+                        PaddleMobile__Framework__ProtobufCMessageUnknownField));
     if (rv->unknown_fields == NULL) goto error_cleanup;
   }
 
@@ -2118,7 +2135,9 @@ void PaddleMobile__Framework__protobuf_c_message_free_unpacked(
                        .data;
       const PaddleMobile__Framework__ProtobufCBinaryData *default_bd;
 
-      default_bd = desc->fields[f].default_value;
+      default_bd =
+          (const PaddleMobile__Framework__ProtobufCBinaryData *)desc->fields[f]
+              .default_value;
       if (data != NULL && (default_bd == NULL || default_bd->data != data)) {
         PaddleMobile__Framework__do_free(allocator, data);
       }
@@ -2166,7 +2185,8 @@ protobuf_c_boolean PaddleMobile__Framework__protobuf_c_message_check(
     void *field = STRUCT_MEMBER_P(message, f->offset);
 
     if (label == PROTOBUF_C_LABEL_REPEATED) {
-      size_t *quantity = STRUCT_MEMBER_P(message, f->quantifier_offset);
+      size_t *quantity =
+          (size_t *)STRUCT_MEMBER_P(message, f->quantifier_offset);
 
       if (*quantity > 0 && *(void **)field == NULL) {
         return FALSE;
@@ -2208,9 +2228,10 @@ protobuf_c_boolean PaddleMobile__Framework__protobuf_c_message_check(
         char *string = *(char **)field;
         if (label == PROTOBUF_C_LABEL_REQUIRED && string == NULL) return FALSE;
       } else if (type == PROTOBUF_C_TYPE_BYTES) {
-        protobuf_c_boolean *has =
-            STRUCT_MEMBER_P(message, f->quantifier_offset);
-        PaddleMobile__Framework__ProtobufCBinaryData *bd = field;
+        protobuf_c_boolean *has = (protobuf_c_boolean *)STRUCT_MEMBER_P(
+            message, f->quantifier_offset);
+        PaddleMobile__Framework__ProtobufCBinaryData *bd =
+            (PaddleMobile__Framework__ProtobufCBinaryData *)field;
         if (label == PROTOBUF_C_LABEL_REQUIRED || *has == TRUE) {
           if (bd->len > 0 && bd->data == NULL) return FALSE;
         }

mobile/test/net/test_wrap.cpp

@@ -12,27 +12,41 @@ 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 <fstream>
 #include <iostream>
 #include <sstream>
-#include "../test_helper.h"
-#include "../test_include.h"
+#include <vector>
 #include "io/paddle_mobile_wrap.h"
 
 int main(int argc, char *argv[]) {
 #ifndef PADDLE_MOBILE_FPGA
-  paddle_mobile::wrap::Net<paddle_mobile::wrap::CPU> *net =
-      new paddle_mobile::wrap::Net<paddle_mobile::wrap::CPU>();
+  paddle_mobile::wrap::Net *net =
+      new paddle_mobile::wrap::Net(paddle_mobile::wrap::kGPU_CL);
+  net->SetCLPath("/data/local/tmp/bin");
   net->Load("./checked_model/model", "./checked_model/params", false, false, 1,
             true);
-  int size = 1 * 3 * 64 * 64;
+  int size = 1 * 3 * 416 * 416;
+  std::vector<int64_t> shape{1, 3, 416, 416};
   float *data = new float[size];
   for (int i = 0; i < size; i++) {
     data[i] = 0.0;
   }
-  std::vector<int64_t> shape{1, 3, 64, 64};
+  std::ifstream infile;
+  infile.open("input.txt");
+  for (int i = 0; i < size; i++) {
+    infile >> data[i];
+  }
+  infile.close();
+  // input as vector
+  // std::vector<float> data_as_vector(data, data + size);
+  // auto output = net->Predict(data_as_vector, shape);
+  // for (auto item : output) {
+  //     std::cout << item << std::endl;
+  // }
+  // input as float pointer
   paddle_mobile::wrap::Tensor input(data,
                                     paddle_mobile::wrap::make_ddim(shape));
-  net->Feed("data", input);
+  net->Feed("image", input);
   net->Predict();
   auto output = net->Fetch("save_infer_model/scale_0");
   int output_size = 1;
@@ -44,7 +58,8 @@ int main(int argc, char *argv[]) {
   std::cout << std::endl;
   std::cout << "output data: ";
   for (int i = 0; i < output_size; i++) {
-    std::cout << output->data<float>()[i] << std::endl;
+    std::cout << output->data()[i] << std::endl;
   }
 #endif
+  return 0;
 }

mobile/tools/python/fluidtools/.gitignore

@@ -2,3 +2,4 @@
 !run.py
 !.gitignore
 !/model-encrypt-tool
+!test_wrap.py

mobile/tools/python/fluidtools/test_wrap.py

+# -*- coding: utf-8 -*
+import os
+import sys
+import math
+import subprocess
+import numpy as np
+import paddle.fluid as fluid
+
+model_path = "yolov2"
+checked_model_path = "checked_model"
+feed_path = "feeds"
+output_path = "outputs"
+diff_threshold = 0.05
+is_lod = False
+mobile_model_path = ""
+fast_check = False
+is_sample_step = False
+sample_step = 1
+sample_num = 20
+need_encrypt = False
+checked_encrypt_model_path = "checked_encrypt_model"
+output_var_filter = []
+output_key_filter = {}
+check_shape = False
+
+np.set_printoptions(linewidth=150)
+
+mobile_exec_root = "/data/local/tmp/bin"
+mobile_src_root = os.path.abspath("../../../")
+if mobile_src_root.endswith("/"):
+    mobile_src_root = mobile_src_root[:-1]
+
+dot = "•"
+black = lambda x: "\033[30m" + str(x) + "\033[0m"
+red = lambda x: "\033[31m" + str(x) + "\033[0m"
+green = lambda x: "\033[32m" + str(x) + "\033[0m"
+yellow = lambda x: "\033[33m" + str(x) + "\033[0m"
+reset = lambda x: "\033[0m" + str(x)
+
+def pp_tab(x, level=0):
+    header = ""
+    for i in range(0, level):
+        header += "\t"
+    print(header + str(x))
+def pp_black(x, level=0):
+    pp_tab(black(x) + reset(""), level)
+def pp_red(x, level=0):
+    pp_tab(red(x) + reset(""), level)
+def pp_green(x, level=0):
+    pp_tab(green(x) + reset(""), level)
+def pp_yellow(x, level=0):
+    pp_tab(yellow(x) + reset(""), level)
+
+def sh(command):
+    pipe = subprocess.Popen(command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
+    return pipe.stdout.read().decode("utf-8")
+def push(src, dest=""):
+    sh("adb push {} {}".format(src, mobile_exec_root + "/" + dest))
+
+pp_yellow(dot + " start inspecting fluid model")
+
+exe = fluid.Executor(fluid.CPUPlace())
+exe.run(fluid.default_startup_program())
+
+# 加载模型
+def load_model(model_path):
+    prog, feeds, fetches = fluid.io.load_inference_model(dirname=model_path, executor=exe, model_filename="model", params_filename="params")
+    return (prog, feeds, fetches)
+
+prog, feeds, fetches = load_model(model_path)
+
+# 强制要求所有张量的形状，在model和params中一致，并重新保存模型
+def resave_model(feed_kv):
+    if len(mobile_model_path) > 0:
+        pp_green("has set mobile_model_path, stop checking model & params", 1)
+        sh("cp {}/* {}".format(mobile_model_path, checked_model_path))
+        return
+    ops = prog.current_block().ops
+    vars = prog.current_block().vars
+    # 强制所有var为可持久化
+    p_names = []
+    for name in vars:
+        name = str(name)
+        v = fluid.framework._get_var(name, prog)
+        if not v.persistable:
+            v.persistable = True
+            p_names.append(name)
+    outputs = run_model(feed_kv=feed_kv)
+    has_found_wrong_shape = False
+    # 修正每个var的形状
+    for name in vars:
+        name = str(name)
+        v = vars[name]
+        if v.persistable:
+            v1 = fluid.global_scope().find_var(name)
+            try:
+                t1 = v1.get_tensor()
+                shape = t1.shape()
+            except:
+                continue
+            if v.desc.shape() != shape:
+                has_found_wrong_shape = True
+            v.desc.set_shape(shape)
+    # 恢复var的可持久化属性
+    for name in p_names:
+        v = fluid.framework._get_var(name, prog)
+        v.persistable = False
+    fluid.io.save_inference_model(dirname=checked_model_path, feeded_var_names=feeds, target_vars=fetches, executor=exe, main_program=prog, model_filename="model", params_filename="params")
+    if has_found_wrong_shape:
+        pp_red("has found wrong shape", 1)
+    else:
+        pp_green("has not found wrong shape", 1)
+    pp_green("new model is saved into directory 【{}】".format(checked_model_path), 1)
+
+# 分别加密model和params，加密key使用同一个
+def encrypt_model():
+    if not need_encrypt:
+        return
+    pp_yellow(dot + dot + " encrypting model")
+    if not os.path.exists(checked_encrypt_model_path):
+        os.mkdir(checked_encrypt_model_path)
+    res = sh("model-encrypt-tool/enc_key_gen -l 20 -c 232")
+    lines = res.split("\n")
+
+    for line in lines:
+        if line.startswith("key:"):
+            line = line.replace('key:','')
+            sh("model-encrypt-tool/enc_model_gen -k '{}' -c 2 -i checked_model/model -o "
+               "checked_model/model.ml".format(line))
+            sh("model-encrypt-tool/enc_model_gen -k '{}' -c 2 -i checked_model/params  -o checked_model/params.ml".format(line))
+            pp_green("model has been encrypted, key is : {}".format(line), 1)
+            sh("mv {} {}".format(checked_model_path + "/*.ml", checked_encrypt_model_path))
+            return
+    pp_red("model encrypt error", 1)
+
+# 生成feed的key-value对
+def gen_feed_kv():
+    feed_kv = {}
+    for feed_name in feeds:
+        feed_shape = get_feed_var_shape(feed_name)
+        data = np.random.random(feed_shape).astype("float32")
+        feed_kv[feed_name] = data
+    return feed_kv
+
+# 保存feed的key-value对
+def save_feed_kv(feed_kv):
+    for feed_name in feed_kv:
+        feed_data = feed_kv[feed_name]
+        feed_list = feed_data.flatten().tolist()
+        if not os.path.exists(feed_path):
+            os.mkdir(feed_path)
+        file_name = feed_name.replace("/", "_")
+        out_file = open(feed_path + "/" + file_name, "w")
+        for feed_item in feed_list:
+            out_file.write("{}\n".format(feed_item))
+        out_file.close()
+
+last_feed_var_name = None
+last_feed_file_name = None
+last_feed_var_lod = None
+# 加载feed的key-value对
+def load_feed_kv():
+    if not os.path.exists(feed_path):
+        return None
+    global last_feed_var_name
+    global last_feed_file_name
+    global last_feed_var_lod
+    feed_kv = {}
+    pp_yellow(dot + dot + " checking feed info")
+    pp_green("feed data is saved into directory 【{}】".format(feed_path), 1)
+    for feed_name in feeds:
+        feed_shape = get_feed_var_shape(feed_name)
+        pp_tab("feed var name : {}; feed var shape : {}".format(feed_name, feed_shape), 1)
+        file_name = feed_name.replace("/", "_")
+        last_feed_var_name = feed_name
+        last_feed_file_name = file_name
+        feed_file_path = feed_path + "/" + file_name
+        if not os.path.exists(feed_file_path):
+            return None
+        data = np.loadtxt(feed_file_path)
+        expected_len = 1
+        for dim in feed_shape:
+            expected_len *= dim
+        if len(np.atleast_1d(data)) != expected_len:
+            return None
+        data = data.reshape(feed_shape).astype("float32")
+        
+        if is_lod:
+            data_shape = [1]
+            for dim in feed_shape:
+                data_shape.append(dim)
+            data = data.reshape(data_shape).astype("float32")
+            tensor = fluid.LoDTensor()
+            seq_lens = [len(seq) for seq in data]
+            cur_len = 0
+            lod = [cur_len]
+            for l in seq_lens:
+                cur_len += l
+                lod.append(cur_len)
+            data = data.reshape(feed_shape)
+            tensor.set(data, fluid.CPUPlace())
+            tensor.set_lod([lod])
+            last_feed_var_lod = lod
+            feed_kv[feed_name] = tensor
+        else:
+            feed_kv[feed_name] = data
+    return feed_kv
+
+# 运行模型
+def run_model(feed_kv=None):
+    if feed_kv is None:
+        feed_kv = gen_feed_kv()
+    outputs = exe.run(prog, feed=feed_kv, fetch_list=fetches, return_numpy=False)
+    results = []
+    for output in outputs:
+        results.append(np.array(output))
+    return results
+
+# 获取变量形状
+def get_var_shape(var_name):
+    vars = prog.current_block().vars
+    shape = vars[var_name].desc.shape()
+    for i in range(len(shape)):
+        dim = shape[i]
+        if dim == -1:
+            shape[i] = 1
+    return shape
+
+# 获取输入变量形状
+def get_feed_var_shape(var_name):
+    # 如果想写死输入形状，放开以下语句
+    # return [1, 3, 224, 224]
+    return get_var_shape(var_name)
+
+persistable_cache = []
+# 所有var，全部变成持久化
+def force_all_vars_to_persistable():
+    global persistable_cache
+    for var_name in vars.keys():
+        var_name = str(var_name)
+        v = fluid.framework._get_var(var_name, prog)
+        persistable = v.persistable
+        if not persistable:
+            persistable_cache.append(var_name)
+            v.persistable = True
+
+# 恢复持久化属性
+def restore_all_vars_persistable():
+    global persistable_cache
+    for var_name in vars.keys():
+        var_name = str(var_name)
+        v = fluid.framework._get_var(var_name, prog)
+        persistable = v.persistable
+        if var_name in persistable_cache:
+            v.persistable = False
+    persistable_cache = []
+
+# 获取var的数据
+def get_var_data(var_name, feed_kv=None):
+    output = np.array(fluid.global_scope().var(var_name).get_tensor())
+    return output
+
+output_var_cache = {}
+def tensor_sample(tensor):
+    if is_sample_step:
+        step = sample_step
+    else:
+        step = math.floor(len(tensor) / sample_num)
+    step = max(step, 1)
+    step = int(step)
+    sample = []
+    for i in range(0, len(tensor), step):
+        sample.append(tensor[i])
+    return sample
+
+op_cache = {}
+# 获取每层输出的数据
+def save_all_op_output(feed_kv=None):
+    force_all_vars_to_persistable()
+    outputs = run_model(feed_kv=feed_kv)
+    if not os.path.exists(output_path):
+        os.mkdir(output_path)
+    ops = prog.current_block().ops
+    fetch_names = []
+    for fetch in fetches:
+        fetch_names.append(fetch.name)
+    feed_names = feeds
+    for fetch_name in fetch_names:
+        output_var_filter.append(fetch_name)
+    for i in range(len(ops)):
+        op = ops[i]
+        var_name = None
+        var_name_index = -1
+        for index in range(len(op.output_names)):
+            if op.output_names[index] in ["Y", "Out", "Output"]:
+                var_name_index = index
+                break
+        if var_name_index != -1:
+            var_name = op.output_arg_names[var_name_index]
+        else:
+            for name in op.output_arg_names:
+                var_name = name
+                if "tmp" in name:
+                    break
+        if len(output_var_filter) > 0:
+            if var_name not in output_var_filter:
+                continue
+        # real_var_name = None
+        # if op.type == "fetch":
+        #     for name in op.input_arg_names:
+        #         real_var_name = name
+        #         if "tmp" in name:
+        #             break
+        # else:
+        #     real_var_name = var_name
+        if fast_check:
+            if var_name not in fetch_names and var_name not in feed_names:
+                continue
+        try:
+            data = get_var_data(var_name, feed_kv=feed_kv).flatten().tolist()
+            sample = tensor_sample(data)
+            output_var_cache[var_name] = (sample)
+            op_cache[i] = (var_name, op)
+            file_name = var_name.replace("/", "_")
+            out_file = open(output_path + "/" + file_name, "w")
+            if var_name in feed_names:
+                for item in data:
+                    out_file.write("{}\n".format(item))
+            else:
+                for item in sample:
+                    out_file.write("{}\n".format(item))
+            out_file.close()
+        except:
+            pass
+    for i in range(len(ops)):
+        op = ops[i]
+        if op.type not in output_key_filter:
+            continue
+        var_name = None
+        var_name_index = -1
+        for index in range(len(op.output_names)):
+            if op.output_names[index] in output_key_filter[op.type]:
+                var_name_index = index
+                break
+        if var_name_index != -1:
+            var_name = op.output_arg_names[var_name_index]
+        else:
+            continue
+        if len(output_var_filter) > 0:
+            if var_name not in output_var_filter:
+                continue
+        # real_var_name = None
+        # if op.type == "fetch":
+        #     for name in op.input_arg_names:
+        #         real_var_name = name
+        #         if "tmp" in name:
+        #             break
+        # else:
+        #     real_var_name = var_name
+        if fast_check:
+            if var_name not in fetch_names and var_name not in feed_names:
+                continue
+        try:
+            data = get_var_data(var_name, feed_kv=feed_kv).flatten().tolist()
+            sample = tensor_sample(data)
+            output_var_cache[var_name] = (sample)
+            op_cache[i] = (var_name, op)
+            file_name = var_name.replace("/", "_")
+            out_file = open(output_path + "/" + file_name, "w")
+            if var_name in feed_names:
+                for item in data:
+                    out_file.write("{}\n".format(item))
+            else:
+                for item in sample:
+                    out_file.write("{}\n".format(item))
+            out_file.close()
+        except:
+            pass
+    pp_green("all the op outputs are saved into directory 【{}】".format(output_path), 1)
+    restore_all_vars_persistable()
+
+ops = prog.current_block().ops
+vars = prog.current_block().vars
+
+pp_yellow(dot + dot + " checking op list")
+op_types = set()
+for op in ops:
+    op_types.add(op.type)
+pp_tab("op types : {}".format(op_types), 1)
+
+def check_mobile_results(args, fuse, mem_opt):
+    args = "{} {} {}".format("1" if fuse else "0", "1" if mem_opt else "0", args)
+    res = sh("adb shell \"cd {} && export LD_LIBRARY_PATH=. && ./test-net {}\"".format(mobile_exec_root, args))
+    lines = res.split("\n")
+    for line in lines:
+        print(line)
+    for line in lines:
+        if line.startswith("auto-test-debug"):
+            print(line)
+    pp_yellow(dot + dot + " checking paddle mobile results for {} -- {} ".format(green("【fusion】" if fuse else "【non fusion】"), green("【memory-optimization】" if mem_opt else "【non-memory-optimization】")))
+    mobile_var_cache = {}
+    for line in lines:
+        parts = line.split(" ")
+        if len(parts) < 2:
+            continue
+        if "auto-test" != parts[0]:
+            continue
+        if parts[1] == "load-time-cost":
+            pp_green("load time cost : {}".format(parts[2]), 1) 
+        elif parts[1] == "predict-time-cost":
+            pp_green("predict time cost : {}".format(parts[2]), 1) 
+        elif parts[1] == "preprocess-time-cost":
+            pp_green("preprocess time cost : {}".format(parts[2]), 1)
+        elif parts[1] == "var":
+            var_name = parts[2]
+            values = list(map(lambda x: float(x), parts[3:]))
+            mobile_var_cache[var_name] = values
+    error_index = None
+    error_values1 = None
+    error_values2 = None
+    checked_names = []
+    fetch_names = []
+    for fetch in fetches:
+        fetch_names.append(fetch.name)
+    for index in op_cache:
+        op_output_var_name, op = op_cache[index]
+        if mem_opt:
+            found_in_fetch = False
+            for fetch in fetches:
+                if op_output_var_name == fetch.name:
+                    found_in_fetch = True
+                    break
+            if not found_in_fetch:
+                continue
+        if not op_output_var_name in output_var_cache:
+            continue
+        if not op_output_var_name in mobile_var_cache:
+            continue
+        values1 = output_var_cache[op_output_var_name]
+        values2 = mobile_var_cache[op_output_var_name]
+        shape = get_var_shape(op_output_var_name) if check_shape else []
+        if len(values1) + len(shape) != len(values2):
+            error_index = index
+        for i in range(len(shape)):
+            v1 = shape[i]
+            v2 = values2[i]
+            if v1 != v2:
+                error_index = index
+                break
+        if error_index == None:
+            for i in range(len(values1)):
+                v1 = values1[i]
+                v2 = values2[len(shape) + i]
+                if abs(v1 - v2) > diff_threshold:
+                    error_index = index
+                    break
+        checked_names.append(op_output_var_name)
+        if error_index != None:
+            error_values1 = values1
+            error_values2 = values2
+            break
+    if error_index == None:
+        for name in fetch_names:
+            if name not in checked_names:
+                error_index = -1
+                break
+    if error_index == None:
+        pp_green("outputs are all correct", 1)
+    elif error_index == -1:
+        pp_red("outputs are missing")
+    else:
+        error_values1 = np.array(error_values1)
+        error_values2 = np.array(error_values2)
+        # pp_red("mobile op is not correct, error occurs at {}th op, op's type is {}")
+        pp_red("corresponding fluid op is {}th op, op's type is {}, wrong var name is {}".format(
+            error_index,op_cache[error_index][1].type,op_output_var_name), 1)
+        pp_red("fluid results are : ", 1)
+        pp_red(str(error_values1).replace("\n", "\n" + "\t" * 1), 1)
+        pp_yellow("paddle mobile results are : ", 1)
+        pp_red(str(error_values2).replace("\n", "\n" + "\t" * 1), 1)
+    # print(output_var_cache)
+    # print(mobile_var_cache)
+
+def main():
+    # 加载kv
+    feed_kv = load_feed_kv()
+    if feed_kv == None:
+        feed_kv = gen_feed_kv()
+        save_feed_kv(feed_kv)
+        feed_kv = load_feed_kv()
+    # 预测
+    pp_yellow(dot + dot + " checking inference")
+    outputs = run_model(feed_kv=feed_kv)
+    pp_tab("fluid output : {}".format(outputs), 1)
+    # 重新保存模型
+    pp_yellow(dot + dot + " checking model correctness")
+    resave_model(feed_kv=feed_kv)
+    # 输出加密模型
+    encrypt_model()
+    # 输出所有中间结果
+    pp_yellow(dot + dot + " checking output result of every op")
+    save_all_op_output(feed_kv=feed_kv)
+    pp_yellow(dot + dot + " checking fetch info")
+    for fetch in fetches:
+        fetch_name = fetch.name
+        fetch_shape = get_var_shape(fetch_name)
+        pp_tab("fetch var name : {}; fetch var shape : {}".format(fetch_name, fetch_shape), 1)
+    # 输出所有op、var信息
+    info_file = open("info.txt", "w")
+    for i in range(len(ops)):
+        op = ops[i]
+        info_file.write("{}th op: type - {}\n".format(i, op.type))
+        info_file.write("inputs:\n")
+        for var_name in op.input_arg_names:
+            try:
+                shape = get_var_shape(var_name)
+                shape_str = ", ".join(list(map(lambda x: str(x), shape)))
+                info_file.write("var {} : {}\n".format(var_name, shape_str))
+            except:
+                pass
+        info_file.write("outputs:\n")
+        for var_name in op.output_arg_names:
+            try:
+                shape = get_var_shape(var_name)
+                shape_str = ", ".join(list(map(lambda x: str(x), shape)))
+                info_file.write("var {} : {}\n".format(var_name, shape_str))
+            except:
+                pass
+    info_file.close()
+    # 开始检查mobile的正确性
+    print("")
+    print("==================================================")
+    print("")
+    pp_yellow(dot + " start inspecting paddle mobile correctness & performance")
+    push(checked_model_path)
+    push(feed_path + "/" + last_feed_file_name, "input.txt")
+    push(mobile_src_root + "/build/release/arm-v7a/build/libpaddle-mobile.so")
+    push(mobile_src_root + "/build/release/arm-v7a/build/cl_kernel")
+    push(mobile_src_root + "/test/build/test-wrap")
+    res = sh("adb shell 'cd {} && export LD_LIBRARY_PATH=. && ./test-wrap'".format(mobile_exec_root))
+    lines = res.split("\n")
+    for line in lines:
+        print(line)
+
+if __name__ == "__main__":
+    main()