Merge pull request #7183 from tensor-tang/use_mkl_packed

Add flag use_mkl_packed

paddle/trainer/TrainerConfigHelper.cpp

@@ -29,6 +29,7 @@ DECLARE_bool(with_gpu);
 DECLARE_bool(parallel_nn);
 DECLARE_string(config_args);
 DECLARE_bool(use_mkldnn);
+DECLARE_bool(use_mkl_packed);
 
 const char *kConfigParserModuleName = "paddle.trainer.config_parser";
 const char *kConfigParserFuncName = "parse_config_and_serialize";
@@ -46,6 +47,7 @@ TrainerConfigHelper::TrainerConfigHelper(const std::string &configFilePath)
              << ",with_cost=" << FLAGS_with_cost << ",use_gpu=" << FLAGS_use_gpu
              << ",parallel_nn=" << FLAGS_parallel_nn
              << ",use_mkldnn=" << FLAGS_use_mkldnn
+             << ",use_mkl_packed=" << FLAGS_use_mkl_packed
              << ",cudnn_version=" << hl_get_cudnn_lib_version();
   if (!FLAGS_config_args.empty()) {
     configArgs << "," << FLAGS_config_args;

paddle/utils/Flags.cpp

@@ -27,6 +27,13 @@ DEFINE_bool(use_mkldnn, false, "Default still keep use CPU training");
 DEFINE_bool(use_mkldnn, false, "Only support CPU training");
 #endif
 
+#ifdef PADDLE_WITH_MKLML
+// TODO(TJ): change to true when fully confirmed
+DEFINE_bool(use_mkl_packed, false, "Whether to use MKL Packed Optimization");
+#else
+DEFINE_bool(use_mkl_packed, false, "Not to use MKL Packed Optimization");
+#endif
+
 DEFINE_bool(parallel_nn,
             false,
             "Whether to use multi-threads to calculate one neural network."

paddle/utils/Flags.h

@@ -41,3 +41,4 @@ DECLARE_string(predict_file);
 DECLARE_bool(prev_batch_state);
 DECLARE_string(init_model_path);
 DECLARE_bool(use_mkldnn);
+DECLARE_bool(use_mkl_packed);

python/paddle/trainer/config_parser.py

@@ -3622,8 +3622,13 @@ class ConcatenateLayer2(LayerBase):
 
 @config_layer('recurrent')
 class RecurrentLayer(LayerBase):
+    layer_type = 'recurrent'
+
     def __init__(self, name, inputs, reversed=False, bias=True, **xargs):
-        super(RecurrentLayer, self).__init__(name, 'recurrent', 0, inputs,
+        use_mkl_packed = bool(
+            int(g_command_config_args.get("use_mkl_packed", 0)))
+        self.layer_type = 'mkl_packed_recurrent' if use_mkl_packed else 'recurrent'
+        super(RecurrentLayer, self).__init__(name, self.layer_type, 0, inputs,
                                              **xargs)
         config_assert(len(self.inputs) == 1, 'RecurrentLayer must have 1 input')
         input_layer = self.get_input_layer(0)

python/paddle/v2/__init__.py

@@ -135,6 +135,8 @@ def init(**kwargs):
         cp.g_command_config_args['use_gpu'] = kwargs['use_gpu']
     if 'use_mkldnn' in kwargs:
         cp.g_command_config_args['use_mkldnn'] = kwargs['use_mkldnn']
+    if 'use_mkl_packed' in kwargs:
+        cp.g_command_config_args['use_mkl_packed'] = kwargs['use_mkl_packed']
     assert 'parallel_nn' not in kwargs, ("currently 'parallel_nn' is not "
                                          "supported in v2 APIs.")