diff --git a/mace/core/file_storage.cc b/mace/core/file_storage.cc
new file mode 100644
index 0000000000000000000000000000000000000000..e7916b512048cf87dfc8d0acafeb3e289d16e484
--- /dev/null
+++ b/mace/core/file_storage.cc
@@ -0,0 +1,108 @@
+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#include "mace/core/file_storage.h"
+
+#include <fstream>
+#include <memory>
+#include <utility>
+
+#include "mace/utils/logging.h"
+
+namespace mace {
+
+class FileStorageFactory::Impl {
+ public:
+  explicit Impl(const std::string &path);
+
+  std::unique_ptr<KVStorage> CreateStorage(const std::string &name);
+
+ private:
+  std::string path_;
+};
+
+FileStorageFactory::Impl::Impl(const std::string &path): path_(path) {}
+std::unique_ptr<KVStorage> FileStorageFactory::Impl::CreateStorage(
+    const std::string &name) {
+  return std::move(std::unique_ptr<KVStorage>(
+      new FileStorage(path_ + "/" + name)));
+}
+
+FileStorageFactory::FileStorageFactory(const std::string &path):
+    impl_(new FileStorageFactory::Impl(path)) {}
+
+FileStorageFactory::~FileStorageFactory() = default;
+
+std::unique_ptr<KVStorage> FileStorageFactory::CreateStorage(
+    const std::string &name) {
+  return impl_->CreateStorage(name);
+}
+
+FileStorage::FileStorage(const std::string &file_path):
+    file_path_(file_path) {}
+
+void FileStorage::Load() {
+  std::ifstream ifs(file_path_, std::ios::binary | std::ios::in);
+  if (ifs.is_open()) {
+    int64_t data_size = 0;
+    ifs.read(reinterpret_cast<char *>(&data_size), sizeof(data_size));
+    while (data_size--) {
+      int32_t key_size = 0;
+      ifs.read(reinterpret_cast<char *>(&key_size), sizeof(key_size));
+      std::string key(key_size, ' ');
+      ifs.read(&key[0], key_size);
+
+      int32_t value_size = 0;
+      ifs.read(reinterpret_cast<char *>(&value_size),
+               sizeof(value_size));
+
+      std::vector<unsigned char> value(value_size);
+      ifs.read(reinterpret_cast<char *>(value.data()),
+               value_size);
+      data_.emplace(key, value);
+    }
+    ifs.close();
+  } else {
+    LOG(INFO) << "No file to Read.";
+  }
+}
+
+bool FileStorage::Insert(const std::string &key,
+                         const std::vector<unsigned char> &value) {
+  data_.emplace(key, value);
+  return true;
+}
+
+std::vector<unsigned char> *FileStorage::Find(const std::string &key) {
+  auto iter = data_.find(key);
+  if (iter == data_.end()) return nullptr;
+
+  return &(iter->second);
+}
+
+void FileStorage::Flush() {
+  std::ofstream ofs(file_path_,
+                    std::ios::binary | std::ios::out);
+  if (ofs.is_open()) {
+    int64_t data_size = data_.size();
+    ofs.write(reinterpret_cast<const char *>(&data_size),
+              sizeof(data_size));
+    for (auto &kv : data_) {
+      int32_t key_size = static_cast<int32_t>(kv.first.size());
+      ofs.write(reinterpret_cast<const char *>(&key_size), sizeof(key_size));
+      ofs.write(kv.first.c_str(), key_size);
+
+      int32_t value_size = static_cast<int32_t>(kv.second.size());
+      ofs.write(reinterpret_cast<const char *>(&value_size),
+                sizeof(value_size));
+      ofs.write(reinterpret_cast<const char*>(kv.second.data()),
+                value_size);
+    }
+    ofs.close();
+  } else {
+    LOG(WARNING) << "Write failed, please check directory exists";
+  }
+}
+
+};  // namespace mace
diff --git a/mace/core/file_storage.h b/mace/core/file_storage.h
new file mode 100644
index 0000000000000000000000000000000000000000..b1c90b50351f559a0d0a8315256b39fa1eeb6bca
--- /dev/null
+++ b/mace/core/file_storage.h
@@ -0,0 +1,34 @@
+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#ifndef MACE_CORE_FILE_STORAGE_H_
+#define MACE_CORE_FILE_STORAGE_H_
+
+#include <map>
+#include <string>
+#include <vector>
+
+#include "mace/public/mace_runtime.h"
+
+namespace mace {
+
+class FileStorage : public KVStorage {
+ public:
+  explicit FileStorage(const std::string &file_path);
+
+ public:
+  void Load() override;
+  bool Insert(const std::string &key,
+              const std::vector<unsigned char> &value) override;
+  std::vector<unsigned char> *Find(const std::string &key) override;
+  void Flush() override;
+
+ private:
+  std::string file_path_;
+  std::map<std::string, std::vector<unsigned char>> data_;
+};
+
+}  // namespace mace
+
+#endif  // MACE_CORE_FILE_STORAGE_H_
diff --git a/mace/core/file_storage_engine.cc b/mace/core/file_storage_engine.cc
deleted file mode 100644
index a3cd12f180875d885a49f7057ebe3e195e19c545..0000000000000000000000000000000000000000
--- a/mace/core/file_storage_engine.cc
+++ /dev/null
@@ -1,74 +0,0 @@
-//
-// Copyright (c) 2017 XiaoMi All rights reserved.
-//
-
-#include "mace/core/file_storage_engine.h"
-
-#include <fstream>
-
-#include "mace/utils/logging.h"
-
-namespace mace {
-
-std::string FileStorageEngine::kStoragePath  // NOLINT(runtime/string)
-    = "/data/local/tmp";
-
-FileStorageEngine::FileStorageEngine(const std::string &file_name):
-    file_name_(file_name) {}
-
-void FileStorageEngine::Write(
-    const std::map<std::string, std::vector<unsigned char>> &data) {
-  const std::string file_path = kStoragePath + "/" + file_name_;
-
-  std::ofstream ofs(file_path,
-                    std::ios::binary | std::ios::out);
-  if (ofs.is_open()) {
-    int64_t data_size = data.size();
-    ofs.write(reinterpret_cast<const char *>(&data_size),
-              sizeof(data_size));
-    for (auto &kv : data) {
-      int32_t key_size = static_cast<int32_t>(kv.first.size());
-      ofs.write(reinterpret_cast<const char *>(&key_size), sizeof(key_size));
-      ofs.write(kv.first.c_str(), key_size);
-
-      int32_t value_size = static_cast<int32_t>(kv.second.size());
-      ofs.write(reinterpret_cast<const char *>(&value_size),
-                sizeof(value_size));
-      ofs.write(reinterpret_cast<const char*>(kv.second.data()),
-                value_size);
-    }
-    ofs.close();
-  } else {
-    LOG(WARNING) << "Write failed, please check directory exists";
-  }
-}
-
-void FileStorageEngine::Read(
-    std::map<std::string, std::vector<unsigned char>> *data) {
-  const std::string file_path = kStoragePath + "/" + file_name_;
-  std::ifstream ifs(file_path, std::ios::binary | std::ios::in);
-  if (ifs.is_open()) {
-    int64_t data_size = 0;
-    ifs.read(reinterpret_cast<char *>(&data_size), sizeof(data_size));
-    while (data_size--) {
-      int32_t key_size = 0;
-      ifs.read(reinterpret_cast<char *>(&key_size), sizeof(key_size));
-      std::string key(key_size, ' ');
-      ifs.read(&key[0], key_size);
-
-      int32_t value_size = 0;
-      ifs.read(reinterpret_cast<char *>(&value_size),
-               sizeof(value_size));
-
-      std::vector<unsigned char> program_binary(value_size);
-      ifs.read(reinterpret_cast<char *>(program_binary.data()),
-               value_size);
-      data->emplace(key, program_binary);
-    }
-    ifs.close();
-  } else {
-    LOG(INFO) << "No file to Read.";
-  }
-}
-
-};  // namespace mace
diff --git a/mace/core/file_storage_engine.h b/mace/core/file_storage_engine.h
deleted file mode 100644
index 9ce742a4d2cdeabb79c04549a3fd8a8969f1b715..0000000000000000000000000000000000000000
--- a/mace/core/file_storage_engine.h
+++ /dev/null
@@ -1,33 +0,0 @@
-//
-// Copyright (c) 2017 XiaoMi All rights reserved.
-//
-
-#ifndef MACE_CORE_FILE_STORAGE_ENGINE_H_
-#define MACE_CORE_FILE_STORAGE_ENGINE_H_
-
-#include <map>
-#include <string>
-#include <vector>
-
-#include "mace/public/mace_runtime.h"
-
-namespace mace {
-
-class FileStorageEngine : public KVStorageEngine {
- public:
-  explicit FileStorageEngine(const std::string &file_name);
- public:
-  void Write(
-      const std::map<std::string, std::vector<unsigned char>> &data) override;
-  void Read(
-      std::map<std::string, std::vector<unsigned char>> *data) override;
- private:
-  std::string file_name_;
-
- public:
-  static std::string kStoragePath;
-};
-
-}  // namespace mace
-
-#endif  // MACE_CORE_FILE_STORAGE_ENGINE_H_
diff --git a/mace/core/mace.cc b/mace/core/mace.cc
index f5b73a3933ac93de1161c683e5daf0985386e392..484c9472af8e9e12124c78b054d539584e4d2801 100644
--- a/mace/core/mace.cc
+++ b/mace/core/mace.cc
@@ -4,7 +4,7 @@
 
 #include <memory>
 
-#include "mace/core/file_storage_engine.h"
+#include "mace/core/file_storage.h"
 #include "mace/core/net.h"
 #include "mace/core/runtime/hexagon/hexagon_control_wrapper.h"
 #include "mace/core/runtime/opencl/opencl_runtime.h"
@@ -69,8 +69,7 @@ class MaceEngine::Impl {
   explicit Impl(const NetDef *net_def,
                 DeviceType device_type,
                 const std::vector<std::string> &input_nodes,
-                const std::vector<std::string> &output_nodes,
-                const std::string &internal_storage_path);
+                const std::vector<std::string> &output_nodes);
   ~Impl();
 
   MaceStatus Run(const std::map<std::string, MaceTensor> &inputs,
@@ -90,8 +89,7 @@ class MaceEngine::Impl {
 MaceEngine::Impl::Impl(const NetDef *net_def,
                        DeviceType device_type,
                        const std::vector<std::string> &input_nodes,
-                       const std::vector<std::string> &output_nodes,
-                       const std::string &internal_storage_path)
+                       const std::vector<std::string> &output_nodes)
     : op_registry_(new OperatorRegistry()),
       device_type_(device_type),
       ws_(new Workspace()),
@@ -99,7 +97,6 @@ MaceEngine::Impl::Impl(const NetDef *net_def,
       hexagon_controller_(nullptr) {
   LOG(INFO) << "MACE version: " << MaceVersion();
   // Set storage path for internal usage
-  FileStorageEngine::kStoragePath = internal_storage_path;
   for (auto input_name : input_nodes) {
     ws_->CreateTensor(MakeString("mace_input_node_", input_name, ":0"),
                       GetDeviceAllocator(device_type_), DT_FLOAT);
@@ -206,11 +203,9 @@ MaceStatus MaceEngine::Impl::Run(
 MaceEngine::MaceEngine(const NetDef *net_def,
                        DeviceType device_type,
                        const std::vector<std::string> &input_nodes,
-                       const std::vector<std::string> &output_nodes,
-                       const std::string &internal_storage_path) {
+                       const std::vector<std::string> &output_nodes) {
   impl_ = std::unique_ptr<MaceEngine::Impl>(
-      new MaceEngine::Impl(net_def, device_type, input_nodes,
-                           output_nodes, internal_storage_path));
+      new MaceEngine::Impl(net_def, device_type, input_nodes, output_nodes));
 }
 
 MaceEngine::~MaceEngine() = default;
diff --git a/mace/core/mace_runtime.cc b/mace/core/mace_runtime.cc
index fb5110bce17472fc21aadf357b8257f64dc8807f..5f94195370990480c0003836cbde8d2a529c70c3 100644
--- a/mace/core/mace_runtime.cc
+++ b/mace/core/mace_runtime.cc
@@ -8,6 +8,8 @@
 
 namespace mace {
 
+std::shared_ptr<KVStorageFactory> kStorageFactory = nullptr;
+
 void ConfigOpenCLRuntime(GPUPerfHint gpu_perf_hint,
                          GPUPriorityHint gpu_priority_hint) {
   VLOG(1) << "Set GPU configurations, gpu_perf_hint: " << gpu_perf_hint
@@ -15,9 +17,9 @@ void ConfigOpenCLRuntime(GPUPerfHint gpu_perf_hint,
   OpenCLRuntime::Configure(gpu_perf_hint, gpu_priority_hint);
 }
 
-void ConfigKVStorageEngine(std::shared_ptr<KVStorageEngine> storage_engine) {
+void ConfigKVStorageFactory(std::shared_ptr<KVStorageFactory> storage_factory) {
   VLOG(1) << "Set internal KV Storage Engine";
-  OpenCLRuntime::Configure(storage_engine);
+  kStorageFactory = storage_factory;
 }
 
 void ConfigOmpThreads(int omp_num_threads) {
diff --git a/mace/core/runtime/opencl/opencl_runtime.cc b/mace/core/runtime/opencl/opencl_runtime.cc
index 89db2e9b03141518c76f0703df231da4049ac30c..9d7045f2ca4728f4a76cc4ced522dab4151392ac 100644
--- a/mace/core/runtime/opencl/opencl_runtime.cc
+++ b/mace/core/runtime/opencl/opencl_runtime.cc
@@ -12,7 +12,7 @@
 #include <vector>
 #include <utility>
 
-#include "mace/core/file_storage_engine.h"
+#include "mace/core/file_storage.h"
 #include "mace/core/runtime/opencl/opencl_extension.h"
 #include "mace/public/mace.h"
 #include "mace/utils/tuner.h"
@@ -182,7 +182,6 @@ void OpenCLProfilingTimer::ClearTiming() {
 GPUPerfHint OpenCLRuntime::kGPUPerfHint = GPUPerfHint::PERF_DEFAULT;
 GPUPriorityHint OpenCLRuntime::kGPUPriorityHint =
     GPUPriorityHint::PRIORITY_DEFAULT;
-std::shared_ptr<KVStorageEngine> OpenCLRuntime::kStorageEngine(nullptr);
 
 OpenCLRuntime *OpenCLRuntime::Global() {
   static OpenCLRuntime runtime(kGPUPerfHint, kGPUPriorityHint);
@@ -194,9 +193,6 @@ void OpenCLRuntime::Configure(GPUPerfHint gpu_perf_hint,
   OpenCLRuntime::kGPUPerfHint = gpu_perf_hint;
   OpenCLRuntime::kGPUPriorityHint = gpu_priority_hint;
 }
-void OpenCLRuntime::Configure(std::shared_ptr<KVStorageEngine> storage_engine) {
-  OpenCLRuntime::kStorageEngine = std::move(storage_engine);
-}
 
 void GetAdrenoContextProperties(std::vector<cl_context_properties> *properties,
                                 GPUPerfHint gpu_perf_hint,
@@ -239,7 +235,8 @@ void GetAdrenoContextProperties(std::vector<cl_context_properties> *properties,
 }
 
 OpenCLRuntime::OpenCLRuntime(GPUPerfHint gpu_perf_hint,
-                             GPUPriorityHint gpu_priority_hint) {
+                             GPUPriorityHint gpu_priority_hint):
+    storage_(nullptr) {
   LoadOpenCLLibrary();
 
   std::vector<cl::Platform> all_platforms;
@@ -312,17 +309,16 @@ OpenCLRuntime::OpenCLRuntime(GPUPerfHint gpu_perf_hint,
                                                       &err);
   MACE_CHECK_CL_SUCCESS(err);
 
-  this->program_map_changed = false;
+  this->program_map_changed_ = false;
 
-  if (kStorageEngine == nullptr) {
+  extern std::shared_ptr<KVStorageFactory> kStorageFactory;
+  if (kStorageFactory != nullptr) {
     const std::string cl_compiled_file_name = "mace_cl_compiled_program.bin";
-    kStorageEngine = std::move(
-        std::unique_ptr<FileStorageEngine>(
-            new FileStorageEngine(cl_compiled_file_name)));
-  }
+    storage_ = kStorageFactory->CreateStorage(cl_compiled_file_name);
 
-  if (platform_info_ != kCompiledProgramPlatform) {
-    kStorageEngine->Read(&program_content_map_);
+    if (platform_info_ != kCompiledProgramPlatform) {
+      storage_->Load();
+    }
   }
 }
 
@@ -377,12 +373,13 @@ bool OpenCLRuntime::BuildProgramFromCache(
     const std::string &build_options_str,
     cl::Program *program) {
   // Find from binary
-  auto it_content = this->program_content_map_.find(built_program_key);
-  if (it_content == this->program_content_map_.end()) {
+  if (this->storage_ == nullptr) return false;
+  auto content = this->storage_->Find(built_program_key);
+  if (content == nullptr) {
     return false;
   }
 
-  *program = cl::Program(context(), {device()}, {it_content->second});
+  *program = cl::Program(context(), {device()}, {*content});
   cl_int ret = program->build({device()}, build_options_str.c_str());
   if (ret != CL_SUCCESS) {
     if (program->getBuildInfo<CL_PROGRAM_BUILD_STATUS>(device()) ==
@@ -451,9 +448,11 @@ void OpenCLRuntime::BuildProgramFromSource(
         reinterpret_cast<unsigned char const *>(program_binaries[0].get()) +
             program_binary_sizes[0]);
 
-    this->program_content_map_.emplace(built_program_key,
-                                       content);
-    this->program_map_changed = true;
+    if (this->storage_ != nullptr) {
+      this->storage_->Insert(built_program_key, content);
+      this->program_map_changed_ = true;
+    }
+
     VLOG(3) << "Program from source: " << built_program_key;
   }
 }
@@ -504,9 +503,9 @@ cl::Kernel OpenCLRuntime::BuildKernel(
 }
 
 void OpenCLRuntime::SaveBuiltCLProgram() {
-  if (program_map_changed) {
-    kStorageEngine->Write(program_content_map_);
-    program_map_changed = false;
+  if (program_map_changed_ && storage_ != nullptr) {
+    storage_->Flush();
+    program_map_changed_ = false;
   }
 }
 
diff --git a/mace/core/runtime/opencl/opencl_runtime.h b/mace/core/runtime/opencl/opencl_runtime.h
index d0a509328e149a174c9176122e1259a8a8e82b22..f6cf4815ce7f2264fe1859f157258b55538095bc 100644
--- a/mace/core/runtime/opencl/opencl_runtime.h
+++ b/mace/core/runtime/opencl/opencl_runtime.h
@@ -56,7 +56,7 @@ class OpenCLRuntime {
  public:
   static OpenCLRuntime *Global();
   static void Configure(GPUPerfHint, GPUPriorityHint);
-  static void Configure(std::shared_ptr<KVStorageEngine> storage_engine);
+  static void Configure(std::shared_ptr<KVStorage> storage_engine);
 
   cl::Context &context();
   cl::Device &device();
@@ -108,16 +108,15 @@ class OpenCLRuntime {
   std::shared_ptr<cl::Device> device_;
   std::shared_ptr<cl::CommandQueue> command_queue_;
   std::map<std::string, cl::Program> built_program_map_;
-  std::map<std::string, std::vector<unsigned char>> program_content_map_;
   std::mutex program_build_mutex_;
   GPUType gpu_type_;
   std::string opencl_version_;
   std::string platform_info_;
-  bool program_map_changed;
+  bool program_map_changed_;
+  std::unique_ptr<KVStorage> storage_;
 
   static GPUPerfHint kGPUPerfHint;
   static GPUPriorityHint kGPUPriorityHint;
-  static std::shared_ptr<KVStorageEngine> kStorageEngine;
 };
 
 }  // namespace mace
diff --git a/mace/examples/example.cc b/mace/examples/example.cc
index 9a50c0aac0d469a886922fb234f166372c461c9e..e70025d6c171087fad4393bc3620f728517ac564 100644
--- a/mace/examples/example.cc
+++ b/mace/examples/example.cc
@@ -157,9 +157,13 @@ bool RunModel(const std::vector<std::string> &input_names,
   const std::string kernel_file_path =
                   "/data/local/tmp/mace_run/cl";
 
+  // Config internal kv storage factory.
+  std::shared_ptr<KVStorageFactory> storage_factory(
+      new FileStorageFactory(kernel_file_path));
+  ConfigKVStorageFactory(storage_factory);
   // Init model
   mace::MaceEngine engine(&net_def, device_type, input_names,
-                          output_names, kernel_file_path);
+                          output_names);
   if (device_type == DeviceType::OPENCL || device_type == DeviceType::HEXAGON) {
     mace::MACE_MODEL_TAG::UnloadModelData(model_data);
   }
diff --git a/mace/public/mace.h b/mace/public/mace.h
index 13469db278f2d71f7766f9e07cc5e94a31830c73..db57fbcb9eaf27c3f8ea2f16ba062c6ca73a09fd 100644
--- a/mace/public/mace.h
+++ b/mace/public/mace.h
@@ -54,8 +54,7 @@ class MaceEngine {
   explicit MaceEngine(const NetDef *net_def,
                       DeviceType device_type,
                       const std::vector<std::string> &input_nodes,
-                      const std::vector<std::string> &output_nodes,
-                      const std::string &internal_storage_path);
+                      const std::vector<std::string> &output_nodes);
   ~MaceEngine();
 
   MaceStatus Run(const std::map<std::string, MaceTensor> &inputs,
diff --git a/mace/public/mace_runtime.h b/mace/public/mace_runtime.h
index 0c8eef59724d58641cc2944870a4e7fe022c03e9..55397701397647318bfefbbb0b2a3a443c502ae7 100644
--- a/mace/public/mace_runtime.h
+++ b/mace/public/mace_runtime.h
@@ -31,16 +31,36 @@ enum GPUPriorityHint {
 
 enum CPUPowerOption { DEFAULT = 0, HIGH_PERFORMANCE = 1, BATTERY_SAVE = 2 };
 
-class KVStorageEngine {
+class KVStorage {
  public:
-  virtual void Write(
-      const std::map<std::string, std::vector<unsigned char>> &data) = 0;
-  virtual void Read(
-      std::map<std::string, std::vector<unsigned char>> *data) = 0;
+  virtual void Load() = 0;
+  virtual bool Insert(const std::string &key,
+                      const std::vector<unsigned char> &value) = 0;
+  virtual std::vector<unsigned char> *Find(const std::string &key) = 0;
+  virtual void Flush() = 0;
 };
 
+class KVStorageFactory {
+ public:
+  virtual std::unique_ptr<KVStorage> CreateStorage(const std::string &name) = 0;
+};
+
+class FileStorageFactory : public KVStorageFactory {
+ public:
+  explicit FileStorageFactory(const std::string &path);
+
+  ~FileStorageFactory();
+
+  std::unique_ptr<KVStorage> CreateStorage(const std::string &name) override;
+
+ private:
+  class Impl;
+  std::unique_ptr<Impl> impl_;
+};
+
+void ConfigKVStorageFactory(std::shared_ptr<KVStorageFactory> storage_factory);
+
 void ConfigOpenCLRuntime(GPUPerfHint, GPUPriorityHint);
-void ConfigKVStorageEngine(std::shared_ptr<KVStorageEngine> storage_engine);
 void ConfigOmpThreads(int omp_num_threads);
 void ConfigCPUPowerOption(CPUPowerOption power_option);
 
diff --git a/mace/python/tools/binary_codegen.py b/mace/python/tools/binary_codegen.py
index 4f03a5f6995fa99415f300c70e88026a8963ed2a..21c6a84929e3527dd2b1e97feb4e9473d7a03ffa 100644
--- a/mace/python/tools/binary_codegen.py
+++ b/mace/python/tools/binary_codegen.py
@@ -25,6 +25,7 @@ def generate_cpp_source():
     with open(binary_path, "rb") as f:
       binary_array = np.fromfile(f, dtype=np.uint8)
 
+    print "Generate binary", binary_path
     idx = 0
     size, = struct.unpack("Q", binary_array[idx:idx+8])
     idx += 8
diff --git a/mace/python/tools/opencl_codegen.py b/mace/python/tools/opencl_codegen.py
index 220b8a18f54be3bf807e75c522c57350eb5e9ff7..18829635bce120a5dc9d912d4f66f76ef8e08239 100644
--- a/mace/python/tools/opencl_codegen.py
+++ b/mace/python/tools/opencl_codegen.py
@@ -14,15 +14,16 @@ FLAGS = None
 
 
 def generate_cpp_source():
-  cl_built_kernel_file_name = 'mace_cl_compiled_program.bin'
-  cl_platform_info_file_name = 'mace_cl_platform_info.txt'
   maps = {}
   platform_info = ''
-  for binary_dir in FLAGS.cl_binary_dirs.split(","):
-    binary_path = os.path.join(binary_dir, cl_built_kernel_file_name)
+  binary_dirs = FLAGS.cl_binary_dirs.strip().split(",")
+  for binary_dir in binary_dirs:
+    binary_path = os.path.join(binary_dir, FLAGS.built_kernel_file_name)
+    print 'Before generate opencl code:', binary_path
     if not os.path.exists(binary_path):
       continue
 
+    print 'generate opencl code:', binary_path
     with open(binary_path, "rb") as f:
       binary_array = np.fromfile(f, dtype=np.uint8)
 
@@ -43,7 +44,7 @@ def generate_cpp_source():
       for ele in value:
         maps[key].append(hex(ele))
 
-    cl_platform_info_path = os.path.join(binary_dir, cl_platform_info_file_name)
+    cl_platform_info_path = os.path.join(binary_dir, FLAGS.platform_info_file_name)
     with open(cl_platform_info_path, 'r') as f:
       curr_platform_info = f.read()
     if platform_info != "":
@@ -76,6 +77,16 @@ def parse_args():
       type=str,
       default="",
       help="The cl binaries directories.")
+  parser.add_argument(
+      "--built_kernel_file_name",
+      type=str,
+      default="",
+      help="The cl binaries directories.")
+  parser.add_argument(
+      "--platform_info_file_name",
+      type=str,
+      default="",
+      help="The cl binaries directories.")
   parser.add_argument(
       "--output_path",
       type=str,
diff --git a/mace/tools/validation/mace_run.cc b/mace/tools/validation/mace_run.cc
index a57acd0dc58f10332fd91325c86ae81230cdd91d..d4fa7c29a00073b9aff74d0a107e620540fc45cb 100644
--- a/mace/tools/validation/mace_run.cc
+++ b/mace/tools/validation/mace_run.cc
@@ -228,6 +228,9 @@ bool RunModel(const std::vector<std::string> &input_names,
 
   // Init model
   LOG(INFO) << "Run init";
+  std::shared_ptr<KVStorageFactory> storage_factory(
+      new FileStorageFactory(kernel_file_path));
+  ConfigKVStorageFactory(storage_factory);
   mace::MaceEngine engine(&net_def, device_type, input_names, output_names);
   if (device_type == DeviceType::OPENCL || device_type == DeviceType::HEXAGON) {
     mace::MACE_MODEL_TAG::UnloadModelData(model_data);
diff --git a/tools/env.sh b/tools/env.sh
index 4308e272e23d9446060cda4495aa3d50edbad64c..634fb93c6d5bf6726c461066d60446601ca8b62a 100644
--- a/tools/env.sh
+++ b/tools/env.sh
@@ -9,6 +9,8 @@ MODEL_CODEGEN_DIR=${CODEGEN_DIR}/models/${MODEL_TAG}
 CL_CODEGEN_DIR=${CODEGEN_DIR}/opencl
 TUNING_CODEGEN_DIR=${CODEGEN_DIR}/tuning
 VERSION_SOURCE_PATH=${CODEGEN_DIR}/version
+CL_BUILT_KERNEL_FILE_NAME=mace_cl_compiled_program.bin
+CL_PLATFORM_INFO_FILE_NAME=mace_cl_platform_info.txt
 if [ -z ${EMBED_MODEL_DATA} ]; then
   EMBED_MODEL_DATA=1
 fi
diff --git a/tools/generate_opencl_code.sh b/tools/generate_opencl_code.sh
index 381ce19df27a77c017adc03818cd887f68c6696d..37d1fb10c9e3f2474e7000c29d58cfb5aeed5a9e 100644
--- a/tools/generate_opencl_code.sh
+++ b/tools/generate_opencl_code.sh
@@ -26,6 +26,8 @@ if [ x"$TYPE" == x"source" ];then
 elif [ x"$#" == x"1" ];then
 
   python mace/python/tools/opencl_codegen.py \
+    --built_kernel_file_name=${CL_BUILT_KERNEL_FILE_NAME} \
+    --platform_info_file_name=${CL_PLATFORM_INFO_FILE_NAME} \
     --output_path=${CL_CODEGEN_DIR}/opencl_compiled_program.cc || exit 1
 
 else
@@ -39,14 +41,17 @@ else
 
   if [ "$PULL_OR_NOT" = 1 ]; then
     CL_BIN_DIR=${CL_BIN_DIRS}
-    rm -rf ${CL_BIN_DIR}
     mkdir -p ${CL_BIN_DIR}
+    rm -rf ${CL_BIN_DIR}/${CL_BUILT_KERNEL_FILE_NAME}
+    rm -rf ${CL_BIN_DIR}/${CL_PLATFORM_INFO_FILE_NAME}
     if [ x"$TARGET_ABI" != x"host" ]; then
       adb -s $DEVICE_ID pull ${COMPILED_PROGRAM_DIR}/. ${CL_BIN_DIR} > /dev/null
     fi
   fi
 
   python mace/python/tools/opencl_codegen.py \
+    --built_kernel_file_name=${CL_BUILT_KERNEL_FILE_NAME} \
+    --platform_info_file_name=${CL_PLATFORM_INFO_FILE_NAME} \
     --cl_binary_dirs=${CL_BIN_DIRS} \
     --output_path=${CL_CODEGEN_DIR}/opencl_compiled_program.cc || exit 1
 fi
diff --git a/tools/generate_tuning_param_code.sh b/tools/generate_tuning_param_code.sh
index 11d250e40cc5ce394b19f9749ab54c962b23f963..4d116c9fec7c140a2078e0d6f28e1c5181468e51 100644
--- a/tools/generate_tuning_param_code.sh
+++ b/tools/generate_tuning_param_code.sh
@@ -24,8 +24,8 @@ else
   fi
 
   if [ "$PULL_OR_NOT" = 1 ]; then
-    rm -rf ${BIN_DIRS}
     mkdir -p ${BIN_DIRS}
+    rm -rf ${BIN_DIRS}/mace_run.config
     if [ x"$TARGET_ABI" != x"host" ]; then
       adb -s $DEVICE_ID pull ${PHONE_DATA_DIR}/mace_run.config ${BIN_DIRS} > /dev/null
     fi
diff --git a/tools/mace_tools.py b/tools/mace_tools.py
index bfe73bafb1576c676e84ce71b72fa682800ea33d..d739bc41f873646adb3f8cb7a572134a763f9cb1 100644
--- a/tools/mace_tools.py
+++ b/tools/mace_tools.py
@@ -251,7 +251,6 @@ def merge_libs_and_tuning_results(target_soc, output_dir, model_output_dirs):
   build_production_code()
 
   model_output_dirs_str = ",".join(model_output_dirs)
-  print output_dir, model_output_dirs
   command = "bash tools/merge_libs.sh {} {} {}".format(target_soc, output_dir,
                                                        model_output_dirs_str)
   run_command(command)