diff --git a/lite/api/test_classify_lite_bm.cc b/lite/api/test_classify_lite_bm.cc
index b2507e28adbe050e4715e0c28a433a259607e7a9..e3d6f4b84472b9655e6d310a79ae67df3048e871 100644
--- a/lite/api/test_classify_lite_bm.cc
+++ b/lite/api/test_classify_lite_bm.cc
@@ -14,6 +14,7 @@
 
 #include <gflags/gflags.h>
 #include <gtest/gtest.h>
+#include <pthread.h>
 #include <fstream>
 #include <vector>
 #include "lite/api/cxx_api.h"
@@ -30,14 +31,18 @@ DEFINE_string(input_img_txt_path,
 namespace paddle {
 namespace lite {
 
-void TestModel(const std::vector<Place>& valid_places) {
+const int g_batch_size = 4;
+const int g_thread_num = 10;
+
+void instance_run() {
   lite::Predictor predictor;
   std::vector<std::string> passes;
+  std::vector<Place> valid_places({Place{TARGET(kBM), PRECISION(kFloat)},
+                                   Place{TARGET(kX86), PRECISION(kFloat)}});
   predictor.Build(FLAGS_model_dir, "", "", valid_places, passes);
-
   auto* input_tensor = predictor.GetInput(0);
-  input_tensor->Resize(DDim(
-      std::vector<DDim::value_type>({1, 3, FLAGS_im_height, FLAGS_im_width})));
+  input_tensor->Resize(DDim(std::vector<DDim::value_type>(
+      {g_batch_size, 3, FLAGS_im_height, FLAGS_im_width})));
   auto* data = input_tensor->mutable_data<float>();
   auto item_size = input_tensor->dims().production();
   if (FLAGS_input_img_txt_path.empty()) {
@@ -45,12 +50,15 @@ void TestModel(const std::vector<Place>& valid_places) {
       data[i] = 1;
     }
   } else {
-    std::fstream fs(FLAGS_input_img_txt_path, std::ios::in);
-    if (!fs.is_open()) {
-      LOG(FATAL) << "open input_img_txt error.";
-    }
-    for (int i = 0; i < item_size; i++) {
-      fs >> data[i];
+    for (int j = 0; j < g_batch_size; j++) {
+      std::fstream fs(FLAGS_input_img_txt_path, std::ios::in);
+      if (!fs.is_open()) {
+        LOG(FATAL) << "open input_img_txt error.";
+      }
+      for (int i = 0; i < item_size / g_batch_size; i++) {
+        fs >> data[i];
+      }
+      data += j * item_size / g_batch_size;
     }
   }
   for (int i = 0; i < FLAGS_warmup; ++i) {
@@ -72,6 +80,7 @@ void TestModel(const std::vector<Place>& valid_places) {
   FILE* fp = fopen("result.txt", "wb");
   for (int i = 0; i < out.size(); i++) {
     auto* out_data = out[i]->data<float>();
+    LOG(INFO) << out[i]->numel();
     for (int j = 0; j < out[i]->numel(); j++) {
       fprintf(fp, "%f\n", out_data[j]);
     }
@@ -79,6 +88,16 @@ void TestModel(const std::vector<Place>& valid_places) {
   fclose(fp);
 }
 
+void TestModel(const std::vector<Place>& valid_places) {
+  std::vector<std::unique_ptr<std::thread>> instances_vec;
+  for (int i = 0; i < g_thread_num; ++i) {
+    instances_vec.emplace_back(new std::thread(&instance_run));
+  }
+  for (int i = 0; i < g_thread_num; ++i) {
+    instances_vec[i]->join();
+  }
+}
+
 TEST(Classify, test_bm) {
   std::vector<Place> valid_places({Place{TARGET(kBM), PRECISION(kFloat)},
                                    Place{TARGET(kX86), PRECISION(kFloat)}});
diff --git a/lite/kernels/bm/bridges/graph.cc b/lite/kernels/bm/bridges/graph.cc
index 32b10f5020605f8f55a2dee3ae2861e591f2f6ed..aeb810f028b69f1da7373956319f9b93e41c72cf 100644
--- a/lite/kernels/bm/bridges/graph.cc
+++ b/lite/kernels/bm/bridges/graph.cc
@@ -20,11 +20,14 @@ namespace lite {
 namespace subgraph {
 namespace bm {
 
+pthread_mutex_t Graph::mutex_compiler_ = PTHREAD_MUTEX_INITIALIZER;
+
 void Graph::AddNode(const std::string& name) {
   nodes_.insert(std::make_pair(name, name));
 }
 
 void Graph::CreateCompilerHandle() {
+  pthread_mutex_lock(&mutex_compiler_);
 #ifdef BM1682
   compiler_handle_ = create_bmcompiler("BM1682");
 #else
@@ -33,6 +36,8 @@ void Graph::CreateCompilerHandle() {
   CHECK(compiler_handle_ != nullptr);
 }
 
+void Graph::UnlockCompilerMutex() { pthread_mutex_unlock(&mutex_compiler_); }
+
 }  // namespace bm
 }  // namespace subgraph
 }  // namespace lite
diff --git a/lite/kernels/bm/bridges/graph.h b/lite/kernels/bm/bridges/graph.h
index 40dadcc92d44e8f3cf73dea63d4c7cf2899cda1f..c54f4d7ad00fa58fe2a30365abc53c589ce4e253 100644
--- a/lite/kernels/bm/bridges/graph.h
+++ b/lite/kernels/bm/bridges/graph.h
@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <pthread.h>
 #include <memory>
 #include <string>
 #include <unordered_map>
@@ -36,10 +37,12 @@ class Graph {
   }
   void CreateCompilerHandle();
   void* GetCompilerHandle() { return compiler_handle_; }
+  void UnlockCompilerMutex();
 
  private:
   std::unordered_map<std::string, std::string> nodes_;
   void* compiler_handle_;
+  static pthread_mutex_t mutex_compiler_;
 };
 
 }  // namespace bm
diff --git a/lite/kernels/bm/subgraph_compute.cc b/lite/kernels/bm/subgraph_compute.cc
index c6059461d1e790064407009cfc0aa3cfcdec8935..d7640e1ac7326d9764380469dc97a7806b044437 100644
--- a/lite/kernels/bm/subgraph_compute.cc
+++ b/lite/kernels/bm/subgraph_compute.cc
@@ -40,6 +40,7 @@ int SubgraphEngine::BuildDeviceProgram() {
     op->CheckShape();
     op->InferShape();
     std::string op_type = op->op_info()->Type();
+    LOG(INFO) << op_type;
     if (!bridges.Exists(op_type, TARGET(kBM))) {
       return subgraph::FAILED;
     }
@@ -59,6 +60,7 @@ int SubgraphEngine::BuildDeviceProgram() {
   unsigned int data_size = 0;
   bm_hd_ = static_cast<bm_handle_t>(ctx.GetHandle());
   finish_bmcompiler_data(graph.GetCompilerHandle(), &bmodel_data, &data_size);
+  graph.UnlockCompilerMutex();
   bmrt_hd_ = bmrt_create(bm_hd_);
   if (false == bmrt_load_bmodel_data(bmrt_hd_, bmodel_data, data_size)) {
     return subgraph::FAILED;