diff --git a/mace/core/net.cc b/mace/core/net.cc
index 6c1533a2bff14e95c0f4c85f055c4df64ef29de1..55a1c830804795f8ab6e3c7e7507d3542a052c87 100644
--- a/mace/core/net.cc
+++ b/mace/core/net.cc
@@ -36,7 +36,7 @@ bool SimpleNet::Run(RunMetadata *run_metadata) {
     VLOG(1) << "Running operator " << op->debug_def().name() << "("
             << op->debug_def().type() << ").";
     OperatorStats *op_stats = nullptr;
-    if (device_type_ != DeviceType::OPENCL && run_metadata) {
+    if (run_metadata && device_type_ != DeviceType::OPENCL) {
       op_stats = run_metadata->add_op_stats();
       op_stats->set_operator_name(op->debug_def().name());
       op_stats->set_type(op->debug_def().type());
@@ -57,16 +57,16 @@ bool SimpleNet::Run(RunMetadata *run_metadata) {
         op_stats->set_type(op->debug_def().type());
 
         op_stats->set_all_start_micros(
-            OpenCLRuntime::GetEventProfilingStartInfo() / 1000);
+            OpenCLRuntime::Get()->GetEventProfilingStartInfo() / 1000);
         op_stats->set_op_start_rel_micros(
-            OpenCLRuntime::GetEventProfilingStartInfo() / 1000 -
+            OpenCLRuntime::Get()->GetEventProfilingStartInfo() / 1000 -
             op_stats->all_start_micros());
 
         op_stats->set_op_end_rel_micros(
-            OpenCLRuntime::GetEventProfilingEndInfo() / 1000 -
+            OpenCLRuntime::Get()->GetEventProfilingEndInfo() / 1000 -
             op_stats->all_start_micros());
         op_stats->set_all_end_rel_micros(
-            OpenCLRuntime::GetEventProfilingEndInfo() / 1000 -
+            OpenCLRuntime::Get()->GetEventProfilingEndInfo() / 1000 -
             op_stats->all_start_micros());
       } else {
         op_stats->set_op_end_rel_micros(NowInMicroSec() -
diff --git a/mace/core/runtime/opencl/opencl_runtime.cc b/mace/core/runtime/opencl/opencl_runtime.cc
index e925e9fb2a79e79802502758bfa33270a39d8a0a..ae67c09973ca897cd9c35f45f8645e3f468f533c 100644
--- a/mace/core/runtime/opencl/opencl_runtime.cc
+++ b/mace/core/runtime/opencl/opencl_runtime.cc
@@ -119,7 +119,10 @@ OpenCLRuntime::OpenCLRuntime(cl::Context context,
   kernel_path_ = std::string(kernel_path == nullptr ? "" : kernel_path) + "/";
 }
 
-OpenCLRuntime::~OpenCLRuntime() { delete profiling_ev_; }
+OpenCLRuntime::~OpenCLRuntime() {
+  if (profiling_ev_)
+    delete profiling_ev_;
+}
 
 cl::Context &OpenCLRuntime::context() { return context_; }
 
diff --git a/mace/core/runtime/opencl/opencl_runtime.h b/mace/core/runtime/opencl/opencl_runtime.h
index 647fe172faf0d3062f0027b270badb2ae84326df..88086998d0779d8c7688de28e77d908611ba36ba 100644
--- a/mace/core/runtime/opencl/opencl_runtime.h
+++ b/mace/core/runtime/opencl/opencl_runtime.h
@@ -19,10 +19,10 @@ class OpenCLRuntime {
   static OpenCLRuntime *Get();
 
   static void EnableProfiling();
-  static cl::Event *GetDefaultEvent();
+  cl::Event *GetDefaultEvent();
 
-  static cl_ulong GetEventProfilingStartInfo();
-  static cl_ulong GetEventProfilingEndInfo();
+  cl_ulong GetEventProfilingStartInfo();
+  cl_ulong GetEventProfilingEndInfo();
 
 
   cl::Context &context();
diff --git a/mace/kernels/opencl/addn.cc b/mace/kernels/opencl/addn.cc
index c6e2101042f67ca72d697ec087585848bebd0ebd..d1ee123d9e7b6cffc5ebea5ec605c104232d2eb0 100644
--- a/mace/kernels/opencl/addn.cc
+++ b/mace/kernels/opencl/addn.cc
@@ -31,7 +31,7 @@ static void Add2(const Tensor *input0, const Tensor *input1, Tensor *output) {
       addn_kernel, cl::NullRange,
       cl::NDRange(gws),
       cl::NDRange(lws),
-      NULL, OpenCLRuntime::GetDefaultEvent());
+      NULL, OpenCLRuntime::Get()->GetDefaultEvent());
   MACE_CHECK(error == CL_SUCCESS);
 }
 
diff --git a/mace/kernels/opencl/batch_norm_opencl.cc b/mace/kernels/opencl/batch_norm_opencl.cc
index 06d2e196fdb446f44e9642a2c0fdf30482562490..c7997d4e13a571c7831c153c979d257b52b29788 100644
--- a/mace/kernels/opencl/batch_norm_opencl.cc
+++ b/mace/kernels/opencl/batch_norm_opencl.cc
@@ -62,7 +62,7 @@ void BatchNormFunctor<DeviceType::OPENCL, float>::operator()(
         bm_kernel, cl::NullRange,
         cl::NDRange(gws[0], gws[1], gws[2]),
         cl::NDRange(params[0], params[1], params[2]),
-        NULL, OpenCLRuntime::GetDefaultEvent());
+        NULL, OpenCLRuntime::Get()->GetDefaultEvent());
 
     MACE_CHECK(error == CL_SUCCESS) << "Error code: " << error;
     return error;
diff --git a/mace/kernels/opencl/conv_2d_opencl_1x1.cc b/mace/kernels/opencl/conv_2d_opencl_1x1.cc
index 8f234be1b98757440b1e1fb0a40211e61891739f..8f019207d2d570b7222acf08b2d1cd0d1e960eda 100644
--- a/mace/kernels/opencl/conv_2d_opencl_1x1.cc
+++ b/mace/kernels/opencl/conv_2d_opencl_1x1.cc
@@ -91,7 +91,7 @@ void Conv1x1V2(const Tensor *input,
       cl::NDRange(static_cast<int>(batch), static_cast<int>(channel_blocks),
                   static_cast<int>(pixel_blocks)),
       cl::NDRange(1, 2, kwg_size / 2),
-      NULL, OpenCLRuntime::GetDefaultEvent());
+      NULL, OpenCLRuntime::Get()->GetDefaultEvent());
   MACE_CHECK(error == CL_SUCCESS, error);
 }
 
@@ -178,7 +178,7 @@ void Conv1x1V3(const Tensor *input,
       cl::NDRange(static_cast<int>(channel_blocks), static_cast<int>(height),
                   static_cast<int>(width)),
       cl::NDRange(1, 2, kwg_size / 2),
-      NULL, OpenCLRuntime::GetDefaultEvent());
+      NULL, OpenCLRuntime::Get()->GetDefaultEvent());
   MACE_CHECK(error == CL_SUCCESS, error);
 }
 
diff --git a/mace/kernels/opencl/conv_2d_opencl_3x3.cc b/mace/kernels/opencl/conv_2d_opencl_3x3.cc
index 3078e4b8eb1e51687bbe23eb2102c226480170b8..c2f6ba7f20309de689d43a724185440e7fae0917 100644
--- a/mace/kernels/opencl/conv_2d_opencl_3x3.cc
+++ b/mace/kernels/opencl/conv_2d_opencl_3x3.cc
@@ -52,7 +52,7 @@ static void InnerConv2dK3x3S12(const Tensor *input, const Tensor *filter,
       conv_kernel, cl::NullRange,
       cl::NDRange(gws[0], gws[1], gws[2]),
       cl::NDRange(lws[0], lws[1], lws[2]),
-      NULL, OpenCLRuntime::GetDefaultEvent());
+      NULL, OpenCLRuntime::Get()->GetDefaultEvent());
   MACE_CHECK(error == CL_SUCCESS);
 }
 
diff --git a/mace/kernels/opencl/depthwise_conv_opencl_3x3.cc b/mace/kernels/opencl/depthwise_conv_opencl_3x3.cc
index 01a2fa1b0d1b8d8e2d1fb7cee598f8f958e85952..da581da895f4c9f64a34d7179337bcf101fd122a 100644
--- a/mace/kernels/opencl/depthwise_conv_opencl_3x3.cc
+++ b/mace/kernels/opencl/depthwise_conv_opencl_3x3.cc
@@ -60,7 +60,7 @@ static void InnerDepthwiseConvOpenclK3x3S12(const Tensor *input,
       conv_kernel, cl::NullRange,
       cl::NDRange(gws[0], gws[1], gws[2]),
       cl::NDRange(lws[0], lws[1], lws[2]),
-      NULL, OpenCLRuntime::GetDefaultEvent());
+      NULL, OpenCLRuntime::Get()->GetDefaultEvent());
   MACE_CHECK(error == CL_SUCCESS);
 }
 
diff --git a/mace/kernels/opencl/pooling_opencl.cc b/mace/kernels/opencl/pooling_opencl.cc
index 8c85b78ad5a808e748045c5c2de40c4ead70b1ec..c805078264acc004b2fa1591257b565876dc5a5c 100644
--- a/mace/kernels/opencl/pooling_opencl.cc
+++ b/mace/kernels/opencl/pooling_opencl.cc
@@ -52,7 +52,7 @@ static void Pooling3(const Tensor *input,
       pooling_kernel, cl::NullRange,
       cl::NDRange(gws[0], gws[1], gws[2]),
       cl::NDRange(lws[0], lws[1], lws[2]),
-      NULL, OpenCLRuntime::GetDefaultEvent());
+      NULL, OpenCLRuntime::Get()->GetDefaultEvent());
   MACE_CHECK(error == CL_SUCCESS);
 }
 
@@ -101,7 +101,7 @@ static void PoolingN(const Tensor *input,
       pooling_kernel, cl::NullRange,
       cl::NDRange(gws[0], gws[1], gws[2]),
       cl::NDRange(lws[0], lws[1], lws[2]),
-      NULL, OpenCLRuntime::GetDefaultEvent());
+      NULL, OpenCLRuntime::Get()->GetDefaultEvent());
   MACE_CHECK(error == CL_SUCCESS);
 }
 
diff --git a/mace/kernels/opencl/relu_opencl.cc b/mace/kernels/opencl/relu_opencl.cc
index 086a653fb68bfbb18750781f9722b46b1e2f3432..60281afa65dd2fd041dd13afc03a3bc43a5f1699 100644
--- a/mace/kernels/opencl/relu_opencl.cc
+++ b/mace/kernels/opencl/relu_opencl.cc
@@ -36,7 +36,7 @@ void ReluFunctor<DeviceType::OPENCL, float>::operator()(const Tensor *input,
         relu_kernel, cl::NullRange,
         cl::NDRange(gws),
         cl::NDRange(lws),
-        NULL, OpenCLRuntime::GetDefaultEvent());
+        NULL, OpenCLRuntime::Get()->GetDefaultEvent());
     MACE_CHECK(error == CL_SUCCESS);
   } else {
     auto relu_kernel  = runtime->BuildKernel("relu", "relux", built_options);
@@ -53,7 +53,7 @@ void ReluFunctor<DeviceType::OPENCL, float>::operator()(const Tensor *input,
         relu_kernel, cl::NullRange,
         cl::NDRange(gws),
         cl::NDRange(lws),
-        NULL, OpenCLRuntime::GetDefaultEvent());
+        NULL, OpenCLRuntime::Get()->GetDefaultEvent());
     MACE_CHECK(error == CL_SUCCESS);
   }
 }
diff --git a/mace/kernels/opencl/resize_bilinear_opencl.cc b/mace/kernels/opencl/resize_bilinear_opencl.cc
index 11b6ee016456f0c4a440b3c344e0d1ba948f7a6a..1e4c2c5461fb4aa2067cf99584f4a0f0b606669f 100644
--- a/mace/kernels/opencl/resize_bilinear_opencl.cc
+++ b/mace/kernels/opencl/resize_bilinear_opencl.cc
@@ -50,7 +50,7 @@ void ResizeBilinearFunctor<DeviceType::OPENCL, float>::operator()(
                   static_cast<int>(out_height), static_cast<int>(out_width)),
       // TODO (heliangliang) tuning and fix when kwg_size < devisor
       cl::NDRange(1, 16, kwg_size / 16),
-      NULL, OpenCLRuntime::GetDefaultEvent());
+      NULL, OpenCLRuntime::Get()->GetDefaultEvent());
   MACE_CHECK(error == CL_SUCCESS, error);
 }
 
diff --git a/mace/kernels/opencl/space_to_batch_opecl.cc b/mace/kernels/opencl/space_to_batch_opecl.cc
index 52e6de377ae07f393ec4e803d56ee7c0437341e2..dc7058f981b99643fc38f3601ebb65441b2adf91 100644
--- a/mace/kernels/opencl/space_to_batch_opecl.cc
+++ b/mace/kernels/opencl/space_to_batch_opecl.cc
@@ -45,7 +45,7 @@ void SpaceToBatchFunctor<DeviceType::OPENCL, float>::operator()(Tensor *space_te
       s2b_kernel, cl::NullRange,
       cl::NDRange(gws[0], gws[1], gws[2]),
       cl::NDRange(lws[0], lws[1], lws[2]),
-      NULL, OpenCLRuntime::GetDefaultEvent());
+      NULL, OpenCLRuntime::Get()->GetDefaultEvent());
   MACE_CHECK(error == CL_SUCCESS);
 }
 
diff --git a/mace/ops/batch_norm_benchmark.cc b/mace/ops/batch_norm_benchmark.cc
index 499af6f29c5f1918f8233ef1e11ba155e35cc869..e0d56173d20e89799e7c2f1a9df33a90dbca47bd 100644
--- a/mace/ops/batch_norm_benchmark.cc
+++ b/mace/ops/batch_norm_benchmark.cc
@@ -3,6 +3,7 @@
 //
 
 #include "mace/core/operator.h"
+#include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/core/testing/test_benchmark.h"
 #include "mace/ops/ops_test_util.h"
 
@@ -12,6 +13,9 @@ static void BatchNorm(
     int iters, int batch, int channels, int height, int width) {
   mace::testing::StopTiming();
 
+  if ( D == OPENCL )
+    OpenCLRuntime::EnableProfiling();
+
   OpsTestNet net;
   OpDefBuilder("BatchNorm", "BatchNormBM")
       .Input("Input")
@@ -77,4 +81,4 @@ BM_BATCH_NORM(1, 512, 14, 14, float);
 BM_BATCH_NORM(1, 1024, 7, 7, float);
 BM_BATCH_NORM(32, 1, 256, 256, float);
 BM_BATCH_NORM(32, 3, 256, 256, float);
-}  //  namespace mace
\ No newline at end of file
+}  //  namespace mace
diff --git a/mace/utils/tuner.h b/mace/utils/tuner.h
index 1d36f7f5b170fc109bc7596bb556b0e8e3ed6959..38c29a8fe7e81a4ffc72bf048780d306ed1dd578 100644
--- a/mace/utils/tuner.h
+++ b/mace/utils/tuner.h
@@ -131,13 +131,14 @@ class Tuner {
                      double &time_us) {
     RetType res;
     int64_t total_time_us = 0;
-    const int64_t start_time = NowInMicroSec();
     for (int i = 0; i < num_runs; ++i) {
       res = func(params);
+      OpenCLRuntime::Get()->command_queue().finish();
+
+      double start_time = OpenCLRuntime::Get()->GetEventProfilingStartInfo() / 1000.0;
+      double end_time = OpenCLRuntime::Get()->GetEventProfilingEndInfo() / 1000.0;
+      total_time_us += end_time - start_time;
     }
-    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;