diff --git a/mace/examples/BUILD b/mace/examples/BUILD
index 96843a2e5bd0886299ad4b0f84f9811baba6ead5..7b6a268d56a78bce9879d9707d71f9f8eb4f29c6 100644
--- a/mace/examples/BUILD
+++ b/mace/examples/BUILD
@@ -39,15 +39,3 @@ cc_binary(
         "//mace/utils:command_line_flags",
     ],
 )
-
-cc_binary(
-    name = "example",
-    srcs = ["example.cc"],
-    copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
-    linkopts = ["-fopenmp"],
-    linkstatic = 1,
-    deps = [
-        "//mace/codegen:generated_models_lib",
-        "//mace/utils:command_line_flags",
-    ],
-)
diff --git a/mace/examples/example.cc b/mace/examples/example.cc
deleted file mode 100644
index 5dbd2c4440a6e6f5b9133ad57165a77076d8df2a..0000000000000000000000000000000000000000
--- a/mace/examples/example.cc
+++ /dev/null
@@ -1,145 +0,0 @@
-//
-// Copyright (c) 2017 XiaoMi All rights reserved.
-//
-
-/**
- * Usage:
- * mace_run --model=mobi_mace.pb \
- *          --input=input_node  \
- *          --output=MobilenetV1/Logits/conv2d/convolution  \
- *          --input_shape=1,3,224,224   \
- *          --input_file=input_data \
- *          --output_file=mace.out  \
- *          --device=NEON
- */
-#include <sys/time.h>
-#include <fstream>
-#include <numeric>
-#include <iostream>
-#include <cstdlib>
-#include "mace/utils/command_line_flags.h"
-#include "mace/core/mace.h"
-#include "mace/utils/logging.h"
-
-using namespace std;
-using namespace mace;
-
-namespace mace {
-extern NetDef MACE_MODEL_FUNCTION();
-}
-void ParseShape(const string &str, vector<int64_t> *shape) {
-  string tmp = str;
-  while (!tmp.empty()) {
-    int dim = atoi(tmp.data());
-    shape->push_back(dim);
-    size_t next_offset = tmp.find(",");
-    if (next_offset == string::npos) {
-      break;
-    } else {
-      tmp = tmp.substr(next_offset + 1);
-    }
-  }
-}
-
-DeviceType ParseDeviceType(const string &device_str) {
-  if(device_str.compare("CPU") == 0) {
-    return DeviceType::CPU;
-  } else if (device_str.compare("NEON") == 0) {
-    return DeviceType::NEON;
-  } else if (device_str.compare("OPENCL") == 0) {
-    return DeviceType::OPENCL;
-  } else {
-    return DeviceType::CPU;
-  }
-}
-
-int main(int argc, char **argv) {
-  string model_file;
-  string input_node;
-  string output_node;
-  string input_shape;
-  string input_file;
-  string output_file;
-  string device;
-  int round = 1;
-
-  std::vector<Flag> flag_list = {
-      Flag("model", &model_file, "model file name"),
-      Flag("input", &input_node, "input node"),
-      Flag("output", &output_node, "output node"),
-      Flag("input_shape", &input_shape, "input shape, separated by comma"),
-      Flag("input_file", &input_file, "input file name"),
-      Flag("output_file", &output_file, "output file name"),
-      Flag("device", &device, "CPU/NEON"),
-      Flag("round", &round, "round"),
-  };
-
-  string usage = Flags::Usage(argv[0], flag_list);
-  const bool parse_result = Flags::Parse(&argc, argv, flag_list);
-
-  if (!parse_result) {
-    LOG(ERROR) << usage;
-    return -1;
-  }
-
-  VLOG(0) << "model: " << model_file << std::endl
-          << "input: " << input_node << std::endl
-          << "output: " << output_node << std::endl
-          << "input_shape: " << input_shape << std::endl
-          << "input_file: " << input_file << std::endl
-          << "output_file: " << output_file << std::endl
-          << "device: " << device << std::endl
-          << "round: " << round << std::endl;
-
-  vector<int64_t> shape;
-  ParseShape(input_shape, &shape);
-
-  // load model
-  NetDef net_def = mace::MACE_MODEL_FUNCTION();
-
-  DeviceType device_type = ParseDeviceType(device);
-  VLOG(0) << device_type;
-  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";
-  mace::MaceEngine engine(&net_def, device_type);
-
-  VLOG(0) << "Run model";
-  std::vector<int64_t> output_shape;
-  timeval tv1, tv2;
-  gettimeofday(&tv1, NULL);
-  for (int i = 0; i < round; ++i) {
-    engine.Run(input_data.get(), shape, output_shape);
-  }
-  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) {
-    ofstream out_file(output_file, ios::binary);
-    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 (auto i : output_shape) {
-      ss << i << ", ";
-    }
-    ss << "]";
-    VLOG(0) << ss.str();
-  }
-}
\ No newline at end of file
diff --git a/mace/examples/mace_run.cc b/mace/examples/mace_run.cc
index d62fc1229a670f6920c3025ffc4b1352f04b1ea4..d6fb321e70c4d36c6d8cf57fc30d44a29eb7c9ef 100644
--- a/mace/examples/mace_run.cc
+++ b/mace/examples/mace_run.cc
@@ -12,11 +12,14 @@
  *          --output_file=mace.out  \
  *          --device=NEON
  */
+#include <sys/time.h>
 #include <fstream>
-#include "mace/core/net.h"
-#include "mace/core/workspace.h"
+#include <numeric>
+#include <iostream>
+#include <cstdlib>
 #include "mace/utils/command_line_flags.h"
-#include "mace/utils/env_time.h"
+#include "mace/core/mace.h"
+#include "mace/utils/logging.h"
 
 using namespace std;
 using namespace mace;
@@ -24,7 +27,7 @@ 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());
@@ -88,86 +91,61 @@ 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);
 
-  int64_t t0 = utils::NowMicros();
+  // load model
   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";
+  VLOG(0) << device_type;
+  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();
-  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();
-  }
+  mace::MaceEngine engine(&net_def, device_type);
 
+  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";
-
-  VLOG(0) << "Run";
-  t0 = utils::NowMicros();
-  for (int i = 0; i < round; ++i) {
-    net->Run();
+  for (int i = 0; i < 1; ++i) {
+    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());
+  VLOG(0) << "Run model";
+  timeval tv1, tv2;
+  gettimeofday(&tv1, NULL);
+  for (int i = 0; i < round; ++i) {
+    engine.Run(input_data.get(), shape, output_shape);
+  }
+  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_example.sh b/tools/validate_example.sh
deleted file mode 100644
index 4957b45651752001197d6e2c2e6d058359069075..0000000000000000000000000000000000000000
--- a/tools/validate_example.sh
+++ /dev/null
@@ -1,128 +0,0 @@
-#!/bin/bash
-# Must run at root dir of mace project.
-set +x
-Usage() {
-  echo 'Usage: bash tools/validate_gcn.sh tf_model_path image_size [tuning]'
-}
-
-if [ $# -lt 2 ];then
-  Usage
-  exit -1
-fi
-
-TF_MODEL_FILE_PATH=$1
-MODEL_DIR=$(dirname ${TF_MODEL_FILE_PATH})
-MACE_SOURCE_DIR=`/bin/pwd`
-MACE_MODEL_NAME='mace_model.pb'
-INPUT_FILE_NAME='model_input'
-OUTPUT_FILE_NAME='gcn.out'
-OUTPUT_LIST_FILE='gcn.list'
-PHONE_DATA_DIR="/data/local/tmp/${MACE_MODEL_NAME}"
-KERNEL_DIR="${PHONE_DATA_DIR}/cl/"
-IMAGE_SIZE=$2
-MODEL_TAG=GCN${IMAGE_SIZE}
-CODEGEN_DIR=${MACE_SOURCE_DIR}/mace/codegen
-MODEL_CODEGEN_DIR=${CODEGEN_DIR}/models/gcn-$IMAGE_SIZE
-CL_CODEGEN_DIR=${CODEGEN_DIR}/opencl
-CL_BIN_DIR=${CODEGEN_DIR}/opencl_bin
-TUNING_CODEGEN_DIR=${CODEGEN_DIR}/tuning
-TUNING_OR_NOT=${3:-0}
-
-build_and_run()
-{
-  EMBED_OPENCL_BINARY=$1
-  if [ "$EMBED_OPENCL_BINARY" = true ]; then
-    EMBED_OPENCL_BINARY_BUILD_FLAGS="--define embed_binary_program=true"
-  fi
-
-  bazel build -c opt --strip always mace/examples:example \
-    --crosstool_top=//external:android/crosstool \
-    --host_crosstool_top=@bazel_tools//tools/cpp:toolchain \
-    --cpu=arm64-v8a \
-    $EMBED_OPENCL_BINARY_BUILD_FLAGS \
-    --copt=-DMACE_MODEL_FUNCTION=Create${MODEL_TAG}
-
-  adb shell "mkdir -p ${PHONE_DATA_DIR}"
-  if [ "$EMBED_OPENCL_BINARY" = false ]; then
-    adb shell "rm -rf ${KERNEL_DIR}"
-    adb shell "mkdir -p ${KERNEL_DIR}"
-    adb push mace/kernels/opencl/cl/. ${KERNEL_DIR}
-  fi
-  adb push ${MODEL_DIR}/${INPUT_FILE_NAME} ${PHONE_DATA_DIR}
-  adb push bazel-bin/mace/examples/example ${PHONE_DATA_DIR}
-
-  num_threads=${1:-4}
-  if [[ "${TUNING_OR_NOT}" != "0" && "$EMBED_OPENCL_BINARY" != true ]];then
-    tuning_flag=1
-  else
-    tuning_flag=0
-  fi
-
-  adb </dev/null shell MACE_TUNING=${tuning_flag} \
-    MACE_CPP_MIN_VLOG_LEVEL=0 \
-    MACE_RUN_PARAMETER_PATH=${PHONE_DATA_DIR}/mace_run.config \
-    MACE_KERNEL_PATH=$KERNEL_DIR \
-    OMP_NUM_THREADS=$num_threads \
-    ${PHONE_DATA_DIR}/example \
-    --model=${PHONE_DATA_DIR}/${MACE_MODEL_NAME} \
-    --input=mace_input_node \
-    --output=mace_output_node \
-    --input_shape="1,${IMAGE_SIZE},${IMAGE_SIZE},3"\
-    --input_file=${PHONE_DATA_DIR}/${INPUT_FILE_NAME} \
-    --output_file=${PHONE_DATA_DIR}/${OUTPUT_FILE_NAME} \
-    --device=OPENCL   \
-    --round=2
-}
-
-echo "Step 1: Generate input data"
-python tools/validate.py --generate_data true --random_seed 1 \
- --input_file=${MODEL_DIR}/${INPUT_FILE_NAME} \
- --input_shape="${IMAGE_SIZE},${IMAGE_SIZE},3"
-
-echo "Step 2: Convert tf model to mace model and optimize memory"
-bazel build //mace/python/tools:tf_converter
-rm -rf ${CODEGEN_DIR}/models
-mkdir -p ${MODEL_CODEGEN_DIR}
-bazel-bin/mace/python/tools/tf_converter --input=${TF_MODEL_FILE_PATH} \
-                                         --output=${MODEL_CODEGEN_DIR}/mace_gcn${IMAGE_SIZE}.cc \
-                                         --input_node=input \
-                                         --output_node=GCN/br_result_2/fcn_br \
-                                         --data_type=DT_HALF \
-                                         --runtime=gpu \
-                                         --output_type=source \
-                                         --template=${MACE_SOURCE_DIR}/mace/python/tools/model.template \
-                                         --model_tag=${MODEL_TAG} \
-                                         --confuse=True
-
-echo "Step 3: Run model on the phone with files"
-build_and_run false
-
-echo "Step 4: Generate OpenCL binary program and config code"
-rm -rf ${CL_BIN_DIR}
-adb pull ${KERNEL_DIR} ${CL_BIN_DIR}
-rm -rf ${CL_CODEGEN_DIR}
-mkdir -p ${CL_CODEGEN_DIR}
-python mace/python/tools/opencl_codegen.py \
-  --cl_binary_dir=${CL_BIN_DIR} --output_path=${CL_CODEGEN_DIR}/opencl_compiled_program.cc
-
-echo "Step 5: Generate tuning source file"
-adb pull ${PHONE_DATA_DIR}/mace_run.config ${CL_BIN_DIR}
-mkdir -p ${TUNING_CODEGEN_DIR}
-python mace/python/tools/binary_codegen.py \
-  --binary_file=${CL_BIN_DIR}/mace_run.config --output_path=${TUNING_CODEGEN_DIR}/tuning_params.cc
-
-echo "Step 6: Run model on the phone using binary"
-build_and_run true
-
-echo "Step 7: Pull the mace run result."
-rm -rf ${MODEL_DIR}/${OUTPUT_FILE_NAME}
-adb </dev/null pull ${PHONE_DATA_DIR}/${OUTPUT_FILE_NAME} ${MODEL_DIR}
-
-echo "Step 8: Validate the result"
-python tools/validate.py --model_file ${TF_MODEL_FILE_PATH} \
-    --input_file ${MODEL_DIR}/${INPUT_FILE_NAME} \
-    --mace_out_file ${MODEL_DIR}/${OUTPUT_FILE_NAME} \
-    --input_node input \
-    --output_node GCN/br_result_2/fcn_br\
-    --input_shape "${IMAGE_SIZE},${IMAGE_SIZE},3" \
-    --output_shape "1,${IMAGE_SIZE},${IMAGE_SIZE},2"
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"