diff --git a/mace/core/mace.cc b/mace/core/mace.cc
index bb181d200965577e2b2cf77302d94509b683e57d..f693d6b55fa9576bb77afddbb10adbf526f91407 100644
--- a/mace/core/mace.cc
+++ b/mace/core/mace.cc
@@ -4,6 +4,9 @@
 
 #include "mace/core/mace.h"
 #include "mace/core/types.h"
+#include "mace/core/net.h"
+#include "mace/core/workspace.h"
+#include "mace/utils/logging.h"
 
 namespace mace {
 
@@ -149,6 +152,17 @@ void Argument::set_strings(const std::vector<std::string> &value) {
   std::copy(value.begin(), value.end(), strings_.begin());
 }
 
+// Node Input
+void NodeInput::CopyFrom(const NodeInput &from) {
+  node_id_ = from.node_id();
+  output_port_ = from.output_port();
+}
+int NodeInput::node_id() const {
+  return node_id_;
+}
+int NodeInput::output_port() const {
+  return output_port_;
+}
 
 // OutputShape
 OutputShape::OutputShape() {}
@@ -338,6 +352,51 @@ uint32_t MemoryBlock::y() const {
   return y_;
 }
 
+// MemoryArena
+const std::vector<MemoryBlock> &MemoryArena::mem_block() const {
+  return mem_block_;
+}
+std::vector<MemoryBlock> &MemoryArena::mutable_mem_block() {
+  return mem_block_;
+}
+int MemoryArena::mem_block_size() const {
+  return mem_block_.size();
+}
+
+// InputInfo
+const std::string &InputInfo::name() const {
+  return name_;
+}
+int32_t InputInfo::node_id() const {
+  return node_id_;
+}
+int32_t InputInfo::max_byte_size() const {
+  return max_byte_size_;
+}
+DataType InputInfo::data_type() const {
+  return data_type_;
+}
+const std::vector<int32_t> &InputInfo::dims() const {
+  return dims_;
+}
+
+// OutputInfo
+const std::string &OutputInfo::name() const {
+  return name_;
+}
+int32_t OutputInfo::node_id() const {
+  return node_id_;
+}
+int32_t OutputInfo::max_byte_size() const {
+  return max_byte_size_;
+}
+DataType OutputInfo::data_type() const {
+  return data_type_;
+}
+const std::vector<int32_t> &OutputInfo::dims() const {
+  return dims_;
+}
+
 // NetDef
 NetDef::NetDef() : has_bits_(0) {}
 
@@ -421,4 +480,49 @@ const OperatorDef &NetDef::op(const int idx) const {
   MACE_CHECK(0 <= idx && idx < op_size());
   return op_[idx];
 }
+
+// Mace Engine
+MaceEngine::MaceEngine(const NetDef *net_def, DeviceType device_type):
+    device_type_(device_type), ws_(new Workspace()), net_(nullptr) {
+
+  ws_->LoadModelTensor(*net_def, device_type);
+
+  // Init model
+  auto net = CreateNet(*net_def, ws_.get(), device_type, NetMode::INIT);
+  if(!net->Run()) {
+    LOG(FATAL) << "Net init run failed";
+  }
+  ws_->CreateTensor("mace_input_node:0", GetDeviceAllocator(device_type_), DT_FLOAT);
+  net_ = std::move(CreateNet(*net_def, ws_.get(), device_type));
+}
+MaceEngine::~MaceEngine() {
+}
+const float *MaceEngine::Run(const float *input,
+                             const std::vector<index_t> &input_shape,
+                             std::vector<int64_t> &output_shape) {
+  Tensor *input_tensor =
+      ws_->CreateTensor("mace_input_node:0", GetDeviceAllocator(device_type_), DT_FLOAT);
+  input_tensor->Resize(input_shape);
+  {
+    Tensor::MappingGuard input_guard(input_tensor);
+    float *input_data = input_tensor->mutable_data<float>();
+    memcpy(input_data, input, input_tensor->size() * sizeof(float));
+  }
+  if(!net_->Run()) {
+    LOG(FATAL) << "Net run failed";
+  }
+  // save output
+  const Tensor *output = ws_->GetTensor("mace_output_node:0");
+
+  if (output != nullptr) {
+    Tensor::MappingGuard output_guard(output);
+    auto shape = output->shape();
+    output_shape.resize(shape.size());
+    std::copy(shape.begin(), shape.end(), output_shape.begin());
+    return output->data<float>();
+  } else {
+    return nullptr;
+  }
+}
+
 } //  namespace mace
diff --git a/mace/core/mace.h b/mace/core/mace.h
index e1607af9df16f8f5eec3d24f780e23145807b64a..c56a53b5f3dc62305310b7d4981e137236627225 100644
--- a/mace/core/mace.h
+++ b/mace/core/mace.h
@@ -7,7 +7,7 @@
 #include <cstdint>
 #include <vector>
 #include <string>
-#include "mace/utils/logging.h"
+#include <memory>
 
 namespace mace {
 
@@ -111,17 +111,10 @@ class Argument {
 
 class NodeInput {
  public:
-  void CopyFrom(const NodeInput &from) {
-    node_id_ = from.node_id();
-    output_port_ = from.output_port();
-  }
+  void CopyFrom(const NodeInput &from);
  public:
-  int node_id() const {
-    return node_id_;
-  }
-  int output_port() const {
-    return output_port_;
-  }
+  int node_id() const;
+  int output_port() const;
  private:
   int node_id_;
   int output_port_;
@@ -218,15 +211,9 @@ class MemoryBlock {
 
 class MemoryArena {
  public:
-  inline const std::vector<MemoryBlock> &mem_block() const {
-    return mem_block_;
-  }
-  inline std::vector<MemoryBlock> &mutable_mem_block() {
-    return mem_block_;
-  }
-  inline int mem_block_size() const {
-    return mem_block_.size();
-  }
+  const std::vector<MemoryBlock> &mem_block() const;
+  std::vector<MemoryBlock> &mutable_mem_block();
+  int mem_block_size() const;
  private:
   std::vector<MemoryBlock> mem_block_;
 
@@ -235,21 +222,11 @@ class MemoryArena {
 // for hexagon mace-nnlib
 class InputInfo {
  public:
-  const std::string &name() const {
-    return name_;
-  }
-  int32_t node_id() const {
-    return node_id_;
-  }
-  int32_t max_byte_size() const {
-    return max_byte_size_;
-  }
-  DataType data_type() const {
-    return data_type_;
-  }
-  const std::vector<int32_t> &dims() const {
-    return dims_;
-  }
+  const std::string &name() const;
+  int32_t node_id() const;
+  int32_t max_byte_size() const;
+  DataType data_type() const;
+  const std::vector<int32_t> &dims() const;
  private:
   std::string name_;
   int32_t node_id_;
@@ -260,21 +237,11 @@ class InputInfo {
 
 class OutputInfo {
  public:
-  const std::string &name() const {
-    return name_;
-  }
-  int32_t node_id() const {
-    return node_id_;
-  }
-  int32_t max_byte_size() const {
-    return max_byte_size_;
-  }
-  DataType data_type() const {
-    return data_type_;
-  }
-  const std::vector<int32_t> &dims() const {
-    return dims_;
-  }
+  const std::string &name() const;
+  int32_t node_id() const;
+  int32_t max_byte_size() const;
+  DataType data_type() const;
+  const std::vector<int32_t> &dims() const;
  private:
   std::string name_;
   int32_t node_id_;
@@ -333,5 +300,21 @@ class NetDef {
   uint32_t has_bits_;
 };
 
+class Workspace;
+class NetBase;
+
+class MaceEngine {
+ public:
+  explicit MaceEngine(const NetDef *net_def, DeviceType device_type);
+  ~MaceEngine();
+  const float *Run(const float *input,
+                   const std::vector<int64_t> &input_shape,
+                   std::vector<int64_t> &output_shape);
+ private:
+  DeviceType device_type_;
+  std::unique_ptr<Workspace> ws_;
+  std::unique_ptr<NetBase> net_;
+};
+
 } //  namespace mace
 #endif //  MACE_CORE_MACE_H_
diff --git a/mace/core/net.cc b/mace/core/net.cc
index c24e22d5a3e8a4057535b8b861bbafbf33271450..3a65de85a2715297c01f6701469f888852cada9c 100644
--- a/mace/core/net.cc
+++ b/mace/core/net.cc
@@ -3,6 +3,8 @@
 //
 
 #include "mace/core/net.h"
+#include "mace/core/operator.h"
+#include "mace/core/workspace.h"
 #include "mace/utils/utils.h"
 
 namespace mace {
diff --git a/mace/core/net.h b/mace/core/net.h
index 7c8b510786f494d60df98ba1395b983a05bc4c78..89744712eba97520d38e55a97d6e11645e3ca325 100644
--- a/mace/core/net.h
+++ b/mace/core/net.h
@@ -6,12 +6,14 @@
 #define MACE_CORE_NET_H_
 
 #include "mace/core/common.h"
-#include "mace/core/operator.h"
-#include "mace/core/workspace.h"
 #include "mace/core/mace.h"
 
 namespace mace {
 
+class RunMetadata;
+class OperatorBase;
+class Workspace;
+
 class NetBase {
  public:
   NetBase(const std::shared_ptr<const NetDef> &net_def,
diff --git a/mace/examples/mace_run.cc b/mace/examples/mace_run.cc
index 86b40632577eb5a594c6a29c1d3316b6cc693d95..1e3834f9c585b7fe9cc1c8d03574b22c04535e3c 100644
--- a/mace/examples/mace_run.cc
+++ b/mace/examples/mace_run.cc
@@ -12,9 +12,14 @@
  *          --output_file=mace.out  \
  *          --device=NEON
  */
+#include <sys/time.h>
 #include <fstream>
-#include "mace/core/net.h"
+#include <numeric>
+#include <iostream>
+#include <cstdlib>
 #include "mace/utils/command_line_flags.h"
+#include "mace/core/mace.h"
+#include "mace/utils/logging.h"
 #include "mace/utils/env_time.h"
 
 using namespace std;
@@ -23,7 +28,7 @@ using namespace mace;
 namespace mace {
 extern NetDef MACE_MODEL_FUNCTION();
 }
-void ParseShape(const string &str, vector<index_t> *shape) {
+void ParseShape(const string &str, vector<int64_t> *shape) {
   string tmp = str;
   while (!tmp.empty()) {
     int dim = atoi(tmp.data());
@@ -87,86 +92,71 @@ int main(int argc, char **argv) {
           << "device: " << device << std::endl
           << "round: " << round << std::endl;
 
-  vector<index_t> shape;
+  vector<int64_t> shape;
   ParseShape(input_shape, &shape);
 
+  // load model
   int64_t t0 = utils::NowMicros();
   NetDef net_def = mace::MACE_MODEL_FUNCTION();
   int64_t t1 = utils::NowMicros();
   LOG(INFO) << "CreateNetDef duration: " << t1 - t0 << "us";
   int64_t init_micros = t1 - t0;
 
-  t0 = utils::NowMicros();
   DeviceType device_type = ParseDeviceType(device);
   VLOG(1) << "Device Type" << device_type;
-  Workspace ws;
-  ws.LoadModelTensor(net_def, device_type);
-  Tensor *input_tensor =
-      ws.CreateTensor(input_node + ":0", GetDeviceAllocator(device_type), DT_FLOAT);
-  input_tensor->Resize(shape);
-  t1 = utils::NowMicros();
-  init_micros += t1 - t0;
-  LOG(INFO) << "CreateWorkspaceTensor duration: " << t1 - t0 << "us";
+  int64_t input_size = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int64_t>());
+  std::unique_ptr<float[]> input_data(new float[input_size]);
+
+  // load input
+  ifstream in_file(input_file, ios::in | ios::binary);
+  in_file.read(reinterpret_cast<char *>(input_data.get()),
+               input_size * sizeof(float));
+  in_file.close();
 
   // Init model
   VLOG(0) << "Run init";
   t0 = utils::NowMicros();
-  auto net = CreateNet(net_def, &ws, device_type, NetMode::INIT);
-  net->Run();
+  mace::MaceEngine engine(&net_def, device_type);
   t1 = utils::NowMicros();
   init_micros += t1 - t0;
   LOG(INFO) << "Net init duration: " << t1 - t0 << "us";
 
-  // run model
-  t0 = utils::NowMicros();
-  net = CreateNet(net_def, &ws, device_type);
-  t1 = utils::NowMicros();
-  init_micros += t1 - t0;
   LOG(INFO) << "Total init duration: " << init_micros << "us";
 
-  {
-    Tensor::MappingGuard input_guard(input_tensor);
-    float *input_data = input_tensor->mutable_data<float>();
-
-    // load input
-    ifstream in_file(input_file, ios::in | ios::binary);
-    in_file.read(reinterpret_cast<char *>(input_data),
-                 input_tensor->size() * sizeof(float));
-    in_file.close();
-  }
-
+  std::vector<int64_t> output_shape;
+  VLOG(0) << "warm up";
   // warm up
-  VLOG(0) << "Warm up";
-  t0 = utils::NowMicros();
-  net->Run();
-  t1 = utils::NowMicros();
-  LOG(INFO) << "1st run duration: " << t1 - t0 << "us";
+  for (int i = 0; i < 1; ++i) {
+    engine.Run(input_data.get(), shape, output_shape);
+  }
 
-  VLOG(0) << "Run";
-  t0 = utils::NowMicros();
+  VLOG(0) << "Run model";
+  timeval tv1, tv2;
+  gettimeofday(&tv1, NULL);
   for (int i = 0; i < round; ++i) {
-    net->Run();
+    engine.Run(input_data.get(), shape, output_shape);
   }
-  t1 = utils::NowMicros();
-  LOG(INFO) << "Average duration: " << (t1 - t0) / round << "us";
-
-  // save output
-  const Tensor *output = ws.GetTensor(output_node + ":0");
-
-  std::remove(output_file.c_str());
+  gettimeofday(&tv2, NULL);
+  std::cout << "avg duration: "
+       << ((tv2.tv_sec - tv1.tv_sec) * 1000 +
+           (tv2.tv_usec - tv1.tv_usec) / 1000) /
+              round
+       << endl;
+
+  const float *output = engine.Run(input_data.get(), shape, output_shape);
   if (output != nullptr) {
-    Tensor::MappingGuard output_guard(output);
     ofstream out_file(output_file, ios::binary);
-    out_file.write((const char *)(output->data<float>()),
-                   output->size() * sizeof(float));
+    int64_t output_size = std::accumulate(output_shape.begin(), output_shape.end(), 1, std::multiplies<int64_t>());
+    out_file.write((const char *) (output),
+                   output_size * sizeof(float));
     out_file.flush();
     out_file.close();
     stringstream ss;
     ss << "Output shape: [";
-    for (int i = 0; i < output->dim_size(); ++i) {
-      ss << output->dim(i) << ", ";
+    for (auto i : output_shape) {
+      ss << i << ", ";
     }
     ss << "]";
     VLOG(0) << ss.str();
   }
-}
+}
\ No newline at end of file
diff --git a/tools/validate_gcn.sh b/tools/validate_gcn.sh
index 393e8b1146cb7cd85f174a9136f643c3cf0dcf52..3ba8cfc816609899296cba3ace66066bba05695a 100644
--- a/tools/validate_gcn.sh
+++ b/tools/validate_gcn.sh
@@ -71,7 +71,7 @@ build_and_run()
     --input_file=${PHONE_DATA_DIR}/${INPUT_FILE_NAME} \
     --output_file=${PHONE_DATA_DIR}/${OUTPUT_FILE_NAME} \
     --device=OPENCL   \
-    --round=1
+    --round=2
 }
 
 echo "Step 1: Generate input data"