diff --git a/mace/ops/ops_test_util.cc b/mace/ops/ops_test_util.cc
index 5233ccde1e6a7043f864e558045b823bd63c9507..1cb835f60fd3a16e3c8bd12cfae38f91a8c57e99 100644
--- a/mace/ops/ops_test_util.cc
+++ b/mace/ops/ops_test_util.cc
@@ -100,6 +100,14 @@ void OpDefBuilder::Finalize(OperatorDef *op_def) const {
   *op_def = op_def_;
 }
 
+namespace {
+std::string GetStoragePathFromEnv() {
+  char *storage_path_str = getenv("MACE_INTERNAL_STORAGE_PATH");
+  if (storage_path_str == nullptr) return "";
+  return storage_path_str;
+}
+}  // namespace
+
 OpTestContext *OpTestContext::Get(int num_threads,
                                   CPUAffinityPolicy cpu_affinity_policy,
                                   bool use_gemmlowp) {
@@ -112,7 +120,7 @@ OpTestContext *OpTestContext::Get(int num_threads,
 OpTestContext::OpTestContext(int num_threads,
                              CPUAffinityPolicy cpu_affinity_policy,
                              bool use_gemmlowp)
-    : gpu_context_(new GPUContext()),
+    : gpu_context_(new GPUContext(GetStoragePathFromEnv())),
       opencl_mem_types_({MemoryType::GPU_IMAGE}) {
   device_map_[DeviceType::CPU] = std::unique_ptr<Device>(
       new CPUDevice(num_threads,
diff --git a/mace/public/mace.h b/mace/public/mace.h
index 01818ef5719b48298bd501967bb91cb99521336f..90b1c16fe6260ce46d67e049e609534d33ef45de 100644
--- a/mace/public/mace.h
+++ b/mace/public/mace.h
@@ -146,6 +146,9 @@ class MaceStatus {
 ///
 /// There are some data in common between different MaceEngines using GPU,
 /// use one GPUContext could avoid duplication.
+///
+/// Thread-safe.
+/// You could use one GPUContext for multiple parallel MaceEngines.
 class GPUContext;
 
 /// \brief GPUContext builder.
diff --git a/mace/test/mace_api_mt_test.cc b/mace/test/mace_api_mt_test.cc
index 3a55a4dce485869d998e7350a94a8a629c2c13a0..65264cf02121f8834db763d66f1229c491196795 100644
--- a/mace/test/mace_api_mt_test.cc
+++ b/mace/test/mace_api_mt_test.cc
@@ -39,7 +39,6 @@ void MaceRunFunc(const int in_out_size) {
 
   std::shared_ptr<NetDef> net_def(new NetDef());
 
-
   std::vector<half> data;
   ops::test::GenerateRandomRealTypeData<half>(filter_shape, &data);
   AddTensor<half>(
@@ -63,6 +62,7 @@ void MaceRunFunc(const int in_out_size) {
   }
 
   MaceEngineConfig config(DeviceType::GPU);
+  config.SetGPUContext(mace::ops::test::OpTestContext::Get()->gpu_context());
 
   MaceEngine engine(config);
   MaceStatus status = engine.Init(net_def.get(), input_names, output_names,
diff --git a/tools/device.py b/tools/device.py
index fccbf4428bf1c0376e922fb65a3ac93640b8ed89..90683eb87afed561a96e35c1402ec58536d46478 100644
--- a/tools/device.py
+++ b/tools/device.py
@@ -901,6 +901,8 @@ class DeviceWrapper:
             stdout_buf = []
             process_output = sh_commands.make_output_processor(stdout_buf)
 
+            internal_storage_dir = self.create_internal_storage_dir()
+
             if simpleperf and self.system == SystemType.android:
                 self.push(sh_commands.find_simpleperf_library(abi),
                           self.data_dir)
@@ -909,6 +911,7 @@ class DeviceWrapper:
                     ld_preload,
                     'MACE_OUT_OF_RANGE_CHECK=%s' % out_of_range_check,
                     'MACE_OPENCL_PROFILING=%d' % opencl_profiling,
+                    'MACE_INTERNAL_STORAGE_PATH=%s' % internal_storage_dir,
                     'MACE_CPP_MIN_VLOG_LEVEL=%d' % vlog_level,
                     simpleperf_cmd,
                     'stat',
@@ -928,6 +931,7 @@ class DeviceWrapper:
                     ld_preload,
                     'MACE_OUT_OF_RANGE_CHECK=%d' % out_of_range_check,
                     'MACE_OPENCL_PROFILNG=%d' % opencl_profiling,
+                    'MACE_INTERNAL_STORAGE_PATH=%s' % internal_storage_dir,
                     'MACE_CPP_MIN_VLOG_LEVEL=%d' % vlog_level,
                     device_bin_full_path,
                     args