add python interface of mkldnn_batch_norm

Author: tensor-tang

python/paddle/trainer/config_parser.py

@@ -2420,6 +2420,7 @@ def __init__(self,
         # If not use is_static, even set learning_rate = 0, decay_rate = 0,
         # these paras will change if set average_window in configure.
         use_gpu = bool(int(g_command_config_args.get("use_gpu", 0)))
+        use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0)))
         is_shared = True if not use_gpu else False
         for i in xrange(2):
             inputs.append(
@@ -2433,11 +2434,17 @@ def __init__(self,
 
         parallel_nn = bool(int(g_command_config_args.get("parallel_nn", 0)))
         cudnn_version = int(g_command_config_args.get("cudnn_version", 0))
-        # Automatically select cudnn_batch_norm for GPU and batch_norm for CPU.
-        # Also based on cudnn version.
+        # Automatically select cudnn_batch_norm for GPU, batch_norm for CPU
+        # and mkldnn_batch_norm for MKLDNN. Also based on cudnn version.
+        if batch_norm_type == "mkldnn_batch_norm":
+            config_assert(use_mkldnn, "mkldnn_batch_norm only support MKLDNN")
         use_cudnn = use_gpu and batch_norm_type != "batch_norm" and \
+                not use_mkldnn and batch_norm_type != "mkldnn_batch_norm" and \
                 ((not parallel_nn) or self.config.device > -1)
-        self.layer_type = "cudnn_batch_norm" if use_cudnn else "batch_norm"
+        if use_cudnn:
+            self.layer_type = "cudnn_batch_norm"
+        else:
+            self.layer_type = "mkldnn_batch_norm" if use_mkldnn else "batch_norm"
         super(BatchNormLayer, self).__init__(
             name, self.layer_type, 0, inputs=inputs, **xargs)
 

python/paddle/trainer_config_helpers/layers.py

@@ -3014,16 +3014,19 @@ def batch_norm_layer(input,
     :param input: batch normalization input. Better be linear activation.
                 Because there is an activation inside batch_normalization.
     :type input: LayerOutput
-    :param batch_norm_type: We have batch_norm and cudnn_batch_norm. batch_norm
-                            supports both CPU and GPU. cudnn_batch_norm requires
-                            cuDNN version greater or equal to v4 (>=v4). But
-                            cudnn_batch_norm is faster and needs less memory
-                            than batch_norm. By default (None), we will
-                            automaticly select cudnn_batch_norm for GPU and
-                            batch_norm for CPU. Otherwise, select batch norm
-                            type based on the specified type. If you use cudnn_batch_norm,
+    :param batch_norm_type: We have batch_norm, mkldnn_batch_norm and cudnn_batch_norm.
+                            batch_norm supports CPU, MKLDNN and GPU. cudnn_batch_norm
+                            requires cuDNN version greater or equal to v4 (>=v4).
+                            But cudnn_batch_norm is faster and needs less
+                            memory than batch_norm. mkldnn_batch_norm requires
+                            enable use_mkldnn. By default (None), we will
+                            automaticly select cudnn_batch_norm for GPU,
+                            mkldnn_batch_norm for MKLDNN and batch_norm for CPU.
+                            Otherwise, select batch norm type based on the
+                            specified type. If you use cudnn_batch_norm,
                             we suggested you use latest version, such as v5.1.
     :type batch_norm_type: None | string, None or "batch_norm" or "cudnn_batch_norm"
+                           or "mkldnn_batch_norm"
     :param act: Activation Type. Better be relu. Because batch
                      normalization will normalize input near zero.
     :type act: BaseActivation
@@ -3063,6 +3066,7 @@ def batch_norm_layer(input,
         else:
             num_channels = input.size
     assert (batch_norm_type is None) or (batch_norm_type == "batch_norm") or \
+           (batch_norm_type == "mkldnn_batch_norm") or \
            (batch_norm_type == "cudnn_batch_norm")
     l = Layer(
         name=name,