The flag of mkldnn is enabled iff it is necessary

test=develop

paddle/fluid/pybind/pybind.cc

@@ -86,6 +86,14 @@ bool IsCompiledWithCUDA() {
 #endif
 }
 
+bool IsCompiledWithMKLDNN() {
+#ifndef PADDLE_WITH_MKLDNN
+  return false;
+#else
+  return true;
+#endif
+}
+
 bool IsCompiledWithBrpc() {
 #ifndef PADDLE_WITH_DISTRIBUTE
   return false;
@@ -848,6 +856,7 @@ All parameter, weight, gradient are variables in Paddle.
         [](bool init_p2p) { framework::InitDevices(init_p2p); });
 
   m.def("is_compiled_with_cuda", IsCompiledWithCUDA);
+  m.def("is_compiled_with_mkldnn", IsCompiledWithMKLDNN);
   m.def("is_compiled_with_brpc", IsCompiledWithBrpc);
   m.def("is_compiled_with_dist", IsCompiledWithDIST);
 #ifdef PADDLE_WITH_CUDA

python/paddle/fluid/__init__.py

@@ -125,14 +125,13 @@ def __bootstrap__():
     os.environ['OMP_NUM_THREADS'] = str(num_threads)
     sysstr = platform.system()
     read_env_flags = [
-        'check_nan_inf', 'benchmark', 'eager_delete_scope', 'use_mkldnn',
-        'use_ngraph', 'initial_cpu_memory_in_mb', 'init_allocated_mem',
-        'free_idle_memory', 'paddle_num_threads', "dist_threadpool_size",
-        'eager_delete_tensor_gb', 'fast_eager_deletion_mode',
-        'allocator_strategy', 'reader_queue_speed_test_mode',
-        'print_sub_graph_dir', 'pe_profile_fname', 'warpctc_dir',
-        'inner_op_parallelism', 'enable_parallel_graph',
-        'multiple_of_cupti_buffer_size'
+        'check_nan_inf', 'benchmark', 'eager_delete_scope', 'use_ngraph',
+        'initial_cpu_memory_in_mb', 'init_allocated_mem', 'free_idle_memory',
+        'paddle_num_threads', "dist_threadpool_size", 'eager_delete_tensor_gb',
+        'fast_eager_deletion_mode', 'allocator_strategy',
+        'reader_queue_speed_test_mode', 'print_sub_graph_dir',
+        'pe_profile_fname', 'warpctc_dir', 'inner_op_parallelism',
+        'enable_parallel_graph', 'multiple_of_cupti_buffer_size'
     ]
     if 'Darwin' not in sysstr:
         read_env_flags.append('use_pinned_memory')
@@ -140,6 +139,9 @@ def __bootstrap__():
     if os.name != 'nt':
         read_env_flags.append('cpu_deterministic')
 
+    if core.is_compiled_with_mkldnn():
+        read_env_flags.append('use_mkldnn')
+
     if core.is_compiled_with_dist():
         read_env_flags.append('rpc_deadline')
         read_env_flags.append('rpc_server_profile_path')