diff --git a/mace/kernels/opencl/batch_norm_opencl.cc b/mace/kernels/opencl/batch_norm_opencl.cc
index 4f24249d448409feabda8f002ed3beae491146ed..23bedc6eb30a2753ba67c966c8102385dd293dd0 100644
--- a/mace/kernels/opencl/batch_norm_opencl.cc
+++ b/mace/kernels/opencl/batch_norm_opencl.cc
@@ -44,42 +44,25 @@ void BatchNormFunctor<DeviceType::OPENCL, float>::operator()(
   bm_kernel.setArg(idx++, lws[1] * sizeof(float), nullptr);
   bm_kernel.setArg(idx++, lws[1] * sizeof(float), nullptr);
 
-  std::function<std::vector<std::vector<uint32_t>>()> params_generator = nullptr;
-  std::function<cl_int(const std::vector<uint32_t>& params)> func;
-  if (Tuning()) {
-    params_generator = [&kwg_size]()->std::vector<std::vector<uint32_t>> {
-      return {{1, 1, 64},
-              {1, 1, 128},
-              {1, kwg_size/32, 32},
-              {1, kwg_size/64, 64},
-              {1, kwg_size/128, 128},
-              {1, 1, kwg_size},
-              {1, kwg_size, 1}};
-    };
-    func = [&](const std::vector<uint32_t>& params)->cl_int {
-      cl::Event event;
-      cl_int error = runtime->command_queue().enqueueNDRangeKernel(
-          bm_kernel, cl::NullRange,
-          cl::NDRange(gws[0], gws[1], gws[2]),
-          cl::NDRange(params[0], params[1], params[2]),
-          nullptr,
-          &event);
+  auto params_generator = [&kwg_size]()->std::vector<std::vector<uint32_t>> {
+    return {{1, 1, 64},
+            {1, 1, 128},
+            {1, kwg_size/16, 16},
+            {1, kwg_size/32, 32},
+            {1, kwg_size/64, 64},
+            {1, kwg_size/128, 128},
+            {1, 1, kwg_size},
+            {1, kwg_size, 1}};
+  };
+  auto func = [&](const std::vector<uint32_t>& params)->cl_int {
+    cl_int error = runtime->command_queue().enqueueNDRangeKernel(
+        bm_kernel, cl::NullRange,
+        cl::NDRange(gws[0], gws[1], gws[2]),
+        cl::NDRange(params[0], params[1], params[2]));
 
-      MACE_CHECK(error == CL_SUCCESS);
-      event.wait();
-      return error;
-    };
-  } else {
-    func = [&](const std::vector<uint32_t>& params)->cl_int {
-      cl_int error = runtime->command_queue().enqueueNDRangeKernel(
-          bm_kernel, cl::NullRange,
-          cl::NDRange(gws[0], gws[1], gws[2]),
-          cl::NDRange(params[0], params[1], params[2]));
-
-      MACE_CHECK(error == CL_SUCCESS);
-      return error;
-    };
-  }
+    MACE_CHECK(error == CL_SUCCESS);
+    return error;
+  };
   std::stringstream ss;
   ss << "batch_norm_opencl_kernel_"
       << input->dim(0) << "_"
diff --git a/mace/utils/tuner.h b/mace/utils/tuner.h
index 79a3cf3e49ca56030173b97398c5909e08cd633f..de96f87e27402ba757d3cc713a7b0146c7d5bb49 100644
--- a/mace/utils/tuner.h
+++ b/mace/utils/tuner.h
@@ -10,18 +10,14 @@
 #include <string>
 #include <unordered_map>
 #include <fstream>
-#include <thread>
 #include <limits>
 
 #include "mace/core/logging.h"
 #include "mace/utils/utils.h"
+#include "mace/core/runtime/opencl/opencl_runtime.h"
 
 namespace mace {
 
-bool Tuning() {
-  const char *tuning = getenv("MACE_TUNING");
-  return tuning != nullptr && tuning[0] == '1';
-}
 
 template<typename param_type>
 class Tuner {
@@ -34,22 +30,22 @@ class Tuner {
   template <typename RetType>
   RetType TuneOrRun(const std::string param_key,
               const std::vector<param_type> &default_param,
-              const std::function<std::vector<std::vector<param_type>>()> param_generator,
-              const std::function<RetType(const std::vector<param_type> &)>& func) {
+              const std::function<std::vector<std::vector<param_type>>()> &param_generator,
+              const std::function<RetType(const std::vector<param_type> &)> &func) {
 
-    if (param_generator == nullptr) {
+    if (IsTuning()) {
+      // tune
+      std::vector<param_type> opt_param = default_param;
+      RetType res = Tune<RetType>(param_generator, func, opt_param);
+      param_table_[param_key] = opt_param;
+      return res;
+    } else {
       // run
       if (param_table_.find(param_key) != param_table_.end()) {
         return func(param_table_[param_key]);
       } else {
         return func(default_param);
       }
-    } else {
-      // tune
-      std::vector<param_type> opt_param = default_param;
-      RetType res = Tune<RetType>(param_generator, func, opt_param);
-      param_table_[param_key] = opt_param;
-      return res;
     }
   }
 
@@ -66,6 +62,11 @@ class Tuner {
   Tuner(const Tuner&) = delete;
   Tuner& operator=(const Tuner&) = delete;
 
+  inline bool IsTuning() {
+    const char *tuning = getenv("MACE_TUNING");
+    return tuning != nullptr && strlen(tuning) == 1 && tuning[0] == '1';
+  }
+
   inline void WriteRunParameters() {
     VLOG(0) << path_;
     if (path_ != nullptr) {
@@ -127,24 +128,18 @@ class Tuner {
   inline RetType Run(const std::function<RetType(const std::vector<param_type> &)> &func,
                      const std::vector<param_type> &params,
                      int num_runs,
-                     int64_t sleep_millisecond,
                      double &time_us) {
     RetType res;
     int64_t total_time_us = 0;
-    int64_t actual_num_runs = 0;
-    bool util_max_time = (num_runs <= 0);
-    for (int i = 0; util_max_time || i < num_runs; ++i) {
-      const int64_t start_time = NowInMicroSec();
+    const int64_t start_time = NowInMicroSec();
+    for (int i = 0; i < num_runs; ++i) {
       res = func(params);
-      const int64_t end_time = NowInMicroSec();
-      total_time_us += end_time - start_time;
-      ++(actual_num_runs);
-
-      if (sleep_millisecond > 0) {
-        std::this_thread::sleep_for(std::chrono::milliseconds(sleep_millisecond));
-      }
     }
-    time_us = total_time_us * 1.0 / actual_num_runs;
+    OpenCLRuntime::Get()->command_queue().finish();
+    const int64_t end_time = NowInMicroSec();
+    total_time_us += end_time - start_time;
+
+    time_us = total_time_us * 1.0 / num_runs;
     return res;
   }
 
@@ -158,10 +153,10 @@ class Tuner {
     for (const auto &param: params) {
       double tmp_time = 0.0;
       // warm up
-      Run<RetType>(func, param, 2, 10, tmp_time);
+      Run<RetType>(func, param, 2, tmp_time);
 
       // run
-      RetType tmp_res = Run<RetType>(func, param, 10, 10, tmp_time);
+      RetType tmp_res = Run<RetType>(func, param, 10, tmp_time);
 
       // Check the execution time
       if (tmp_time < opt_time) {
diff --git a/mace/utils/tuner_test.cc b/mace/utils/tuner_test.cc
index 10982415fddbc620105b7291babd93d6454ebb5c..bcb5c620f3b553d3d6f8572fd88573d159d0a6cd 100644
--- a/mace/utils/tuner_test.cc
+++ b/mace/utils/tuner_test.cc
@@ -1,6 +1,7 @@
 //
 // Copyright (c) 2017 XiaoMi All rights reserved.
 //
+#include <thread>
 
 #include "gtest/gtest.h"