use static variable to do cache instead of thread local in thread frequent switching case (#18428)

paddle/fluid/framework/transfer_scope_cache.cc

@@ -17,12 +17,61 @@
 namespace paddle {
 namespace framework {
 
+#ifdef PADDLE_WITH_MKLDNN
+using transfer_data_cache_map = std::unordered_map<size_t, Scope*>;
+using transfer_scope_cache_map = std::unordered_set<Scope*>;
+static std::unordered_map<size_t, transfer_data_cache_map*>
+    static_transfer_data_caches;
+static std::unordered_map<size_t, transfer_scope_cache_map*>
+    static_transfer_scope_caches;
+#endif
+
 std::unordered_map<size_t, Scope*>& global_transfer_data_cache() {
+#ifdef PADDLE_WITH_MKLDNN
+  size_t sid = platform::get_cur_mkldnn_session_id();
+
+  // if there is specific mkldnn tid setting from user.
+  if (sid != platform::kMKLDNNSessionID_Default) {
+    sid = std::hash<std::thread::id>()(std::this_thread::get_id());
+
+    static std::mutex acquire_barrier;
+    std::lock_guard<std::mutex> block_until_finish_this_job(acquire_barrier);
+
+    auto map_it = static_transfer_data_caches.find(sid);
+    if (map_it == static_transfer_data_caches.end()) {
+      auto* x = new transfer_data_cache_map;
+      static_transfer_data_caches[sid] = x;
+      return *x;
+    } else {
+      return *static_transfer_data_caches[sid];
+    }
+  }
+#endif
   thread_local auto* x = new std::unordered_map<size_t, Scope*>;
   return *x;
 }
 
 std::unordered_set<Scope*>& global_transfer_scope_cache() {
+#ifdef PADDLE_WITH_MKLDNN
+  size_t sid = platform::get_cur_mkldnn_session_id();
+
+  // if there is specific mkldnn session id setting from user.
+  if (sid != platform::kMKLDNNSessionID_Default) {
+    sid = std::hash<std::thread::id>()(std::this_thread::get_id());
+
+    static std::mutex acquire_barrier;
+    std::lock_guard<std::mutex> block_until_finish_this_job(acquire_barrier);
+
+    auto map_it = static_transfer_scope_caches.find(sid);
+    if (map_it == static_transfer_scope_caches.end()) {
+      auto* x = new transfer_scope_cache_map;
+      static_transfer_scope_caches[sid] = x;
+      return *x;
+    } else {
+      return *static_transfer_scope_caches[sid];
+    }
+  }
+#endif
   thread_local auto* x = new std::unordered_set<Scope*>;
   return *x;
 }

paddle/fluid/inference/tests/api/analyzer_bert_tester.cc

@@ -230,7 +230,7 @@ TEST(Analyzer_bert, compare_determine) {
                        inputs);
 }
 
-TEST(Analyzer_bert, transfer_scope_cache) {
+void verify_transfer_scope_cache(bool is_static = false) {
   AnalysisConfig config;
   SetConfig(&config);
 
@@ -251,6 +251,11 @@ TEST(Analyzer_bert, transfer_scope_cache) {
     threads.emplace_back([&, i]() {
       std::getline(fin, line);
       ParseLine(line, &input);
+#ifdef PADDLE_WITH_MKLDNN
+      // Use static method to handle transfer_scope_cache()
+      // TODO(intel) explicit session id setting will be deprecated.
+      if (is_static) platform::set_cur_mkldnn_session_id(1);
+#endif
       predictor->Run(input, &output, FLAGS_batch_size);
       global_transfer_scope_cache.insert(
           &paddle::framework::global_transfer_scope_cache());
@@ -261,12 +266,31 @@ TEST(Analyzer_bert, transfer_scope_cache) {
     threads.clear();
     std::vector<PaddleTensor>().swap(input);
   }
-  // Since paddle::framework::global_transfer_scope_cache() and
-  // paddle::framework::global_transfer_data_cache() are thread_local,
-  // their pointer should be different among different thread id.
-  PADDLE_ENFORCE(global_transfer_scope_cache.size(), threads_num);
-  PADDLE_ENFORCE(global_transfer_data_cache.size(), threads_num);
+#ifdef PADDLE_WITH_MKLDNN
+  if (is_static) {
+    // Use static method to do transfer_scope_cache() instead of thread_local
+    // so paddle::framework::global_transfer_data_cache() should be 1
+    PADDLE_ENFORCE(global_transfer_scope_cache.size(), 1);
+    PADDLE_ENFORCE(global_transfer_data_cache.size(), 1);
+  } else {
+#endif
+    // Since paddle::framework::global_transfer_scope_cache() and
+    // paddle::framework::global_transfer_data_cache() are thread_local,
+    // their pointer should be different among different thread id.
+    PADDLE_ENFORCE(global_transfer_scope_cache.size(), threads_num);
+    PADDLE_ENFORCE(global_transfer_data_cache.size(), threads_num);
+#ifdef PADDLE_WITH_MKLDNN
+  }
+#endif
 }
 
+TEST(Analyzer_bert, threadlocal_transfer_scope_cache) {
+  verify_transfer_scope_cache();
+}
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_bert, static_transfer_scope_cache) {
+  verify_transfer_scope_cache(true);
+}
+#endif
 }  // namespace inference
 }  // namespace paddle