diff --git a/paddle/cuda/include/hl_batch_norm.h b/paddle/cuda/include/hl_batch_norm.h
index e1fea1316349c5ec9a1f69d4c34508184efbb59e..afc5e0b2deacc4aadf98b3f7ce115e534bbc5124 100644
--- a/paddle/cuda/include/hl_batch_norm.h
+++ b/paddle/cuda/include/hl_batch_norm.h
@@ -20,20 +20,18 @@ limitations under the License. */
 /**
  * @brief   batch norm inferece.
  *
- * @param[in]   input        input data.
- * @param[out]  output       output data.
- * @param[in]   scale        batch normalization scale parameter (in original
- *                           paper scale is referred to as gamma).
- * @param[in]   bias         batch normalization bias parameter (in original
- *                           paper scale is referred to as beta).
+ * @param[in]   input         input data.
+ * @param[out]  output        output data.
+ * @param[in]   scale         batch normalization scale parameter (in original
+ *                            paper scale is referred to as gamma).
+ * @param[in]   bias          batch normalization bias parameter (in original
+ *                            paper scale is referred to as beta).
  * @param[in]   estimatedMean
- * @param[in]   estimatedVar It is suggested that resultRunningMean,
- *                           resultRunningVariance from the
- *                           cudnnBatchNormalizationForwardTraining call
- *                           accumulated during the training phase are passed
- *                           as inputs here.
- * @param[in]   epsilon      Epsilon value used in the batch
- *                           normalization formula.
+ * @param[in]   estimatedVar  The moving mean and variance
+ *                            accumulated during the training phase are passed
+ *                            as inputs here.
+ * @param[in]   epsilon       Epsilon value used in the batch
+ *                            normalization formula.
  */
 extern void hl_batch_norm_cuda_inference(const real* input,
                                          real* output,
diff --git a/paddle/gserver/layers/CudnnBatchNormLayer.cpp b/paddle/gserver/layers/CudnnBatchNormLayer.cpp
index d99b50385e670190a3228ce9145828297a32f723..cc2cc21cdfd20e26f1b4f708535c3bd863f7871d 100644
--- a/paddle/gserver/layers/CudnnBatchNormLayer.cpp
+++ b/paddle/gserver/layers/CudnnBatchNormLayer.cpp
@@ -81,8 +81,8 @@ void CudnnBatchNormLayer::forward(PassType passType) {
   } else {
     // used movingMean and movingVar in testing
     if (batchSize > 1024) {
-      // when batchSize is larger than 1024, there is a bug
-      // in cudnn library.
+      // there is a bug in cudnn library when the batch size
+      // is larger than 1024.
       hl_batch_norm_cuda_inference(input,
                                    output,
                                    gamma,