diff --git a/paddle/gserver/layers/CrossChannelNormLayer.cpp b/paddle/gserver/layers/CrossChannelNormLayer.cpp
index 3fbccc11032ca..d72503217f1c9 100644
--- a/paddle/gserver/layers/CrossChannelNormLayer.cpp
+++ b/paddle/gserver/layers/CrossChannelNormLayer.cpp
@@ -36,6 +36,16 @@ MatrixPtr CrossChannelNormLayer::createSpatialMatrix(MatrixPtr data,
       data->getData() + iter * spatialDim, 1, spatialDim, false, useGpu_);
 }
 
+bool CrossChannelNormLayer::init(const LayerMap& layerMap,
+                                 const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK(parameters_[0]);
+  const NormConfig& conf = config_.inputs(0).norm_conf();
+  channels_ = conf.channels();
+  scale_.reset(new Weight(channels_, 1, parameters_[0]));
+  return true;
+}
+
 void CrossChannelNormLayer::forward(PassType passType) {
   Layer::forward(passType);
   MatrixPtr inV = getInputValue(0);
@@ -51,9 +61,7 @@ void CrossChannelNormLayer::forward(PassType passType) {
   Matrix::resizeOrCreate(dataBuffer_, batchSize, dataDim, false, useGpu_);
   Matrix::resizeOrCreate(spatialBuffer_, 1, spatialDim, false, useGpu_);
   Matrix::resizeOrCreate(normBuffer_, batchSize, spatialDim, false, useGpu_);
-  normBuffer_->zeroMem();
-  // add eps to avoid overflow
-  normBuffer_->addScalar(*normBuffer_, 1e-6);
+
   inV->square2(*dataBuffer_);
   for (size_t i = 0; i < batchSize; i++) {
     const MatrixPtr inVTmp = createSampleMatrix(inV, i, spatialDim);
@@ -63,6 +71,8 @@ void CrossChannelNormLayer::forward(PassType passType) {
 
     // compute norm.
     spatialBuffer_->sumCols(*dataTmp, 1, 0);
+    // add eps to avoid overflow
+    spatialBuffer_->add(1e-6);
     spatialBuffer_->sqrt2(*spatialBuffer_);
     normTmp->copyFrom(*spatialBuffer_);
     outVTmp->copyFrom(*inVTmp);
@@ -82,6 +92,9 @@ void CrossChannelNormLayer::backward(const UpdateCallback& callback) {
   size_t dataDim = inG->getWidth();
   size_t spatialDim = dataDim / channels_;
 
+  MatrixPtr inGBuffer;
+  Matrix::resizeOrCreate(inGBuffer, channels_, spatialDim, false, useGpu_);
+
   dataBuffer_->dotMul(*outG, *outV);
   Matrix::resizeOrCreate(scaleDiff_, channels_, 1, false, useGpu_);
   Matrix::resizeOrCreate(channelBuffer_, channels_, 1, false, useGpu_);
@@ -100,22 +113,24 @@ void CrossChannelNormLayer::backward(const UpdateCallback& callback) {
     scaleDiff_->add(*channelBuffer_, 1.);
 
     sampleBuffer_->dotMul(*inVTmp, *outGTmp);
-    spatialBuffer_->sumCols(*sampleBuffer_, 1., 1.);
+    spatialBuffer_->sumCols(*sampleBuffer_, 1., 0.);
     // scale the grad
-    inGTmp->copyFrom(*inVTmp);
-    inGTmp->mulRowVector(*spatialBuffer_);
+    inGBuffer->copyFrom(*inVTmp);
+    inGBuffer->mulRowVector(*spatialBuffer_);
     // divide by square of norm
     spatialBuffer_->dotMul(*normTmp, *normTmp);
-    inGTmp->divRowVector(*spatialBuffer_);
+    inGBuffer->divRowVector(*spatialBuffer_);
     // subtract
-    inGTmp->add(*outGTmp, -1, 1);
+    inGBuffer->add(*outGTmp, -1, 1);
     // divide by norm
-    inGTmp->divRowVector(*normTmp);
+    inGBuffer->divRowVector(*normTmp);
     // scale the diff
-    inGTmp->mulColVector(*scale_->getW());
+    inGBuffer->mulColVector(*scale_->getW());
+
+    inGTmp->add(*inGBuffer);
   }
   // updata scale
-  if (scale_->getWGrad()) scale_->getWGrad()->copyFrom(*scaleDiff_);
+  if (scale_->getWGrad()) scale_->getWGrad()->add(*scaleDiff_);
   scale_->getParameterPtr()->incUpdate(callback);
 }
 
diff --git a/paddle/gserver/layers/NormLayer.cpp b/paddle/gserver/layers/NormLayer.cpp
index e094078bfe86e..caef7100929c7 100644
--- a/paddle/gserver/layers/NormLayer.cpp
+++ b/paddle/gserver/layers/NormLayer.cpp
@@ -56,14 +56,4 @@ bool ResponseNormLayer::init(const LayerMap& layerMap,
   return true;
 }
 
-bool CrossChannelNormLayer::init(const LayerMap& layerMap,
-                                 const ParameterMap& parameterMap) {
-  Layer::init(layerMap, parameterMap);
-  CHECK(parameters_[0]);
-  const NormConfig& conf = config_.inputs(0).norm_conf();
-  channels_ = conf.channels();
-  scale_.reset(new Weight(channels_, 1, parameters_[0]));
-  return true;
-}
-
 }  // namespace paddle
diff --git a/paddle/gserver/tests/LayerGradUtil.cpp b/paddle/gserver/tests/LayerGradUtil.cpp
index e3591ba4df88f..15b8cedeb8316 100644
--- a/paddle/gserver/tests/LayerGradUtil.cpp
+++ b/paddle/gserver/tests/LayerGradUtil.cpp
@@ -465,7 +465,6 @@ void initTestLayer(TestConfig testConf,
                            ParameterConfig paraConfig) {
     paraConfig.set_name(paraName);
     paraConfig.set_size(paraSize);
-    paraConfig.set_initial_std(1);
     paraConfig.set_is_static(isStatic);
     auto para =
         std::make_shared<Parameter>(paraConfig, FLAGS_use_gpu, initialize);
@@ -499,6 +498,9 @@ void initTestLayer(TestConfig testConf,
         paraConfig.add_dims((*layerMap)[input.input_layer_name()]->getSize());
         paraConfig.add_dims(testConf.layerConfig.size());
       }
+      CHECK_GE(testConf.paramInitialStd, 0);
+      paraConfig.set_initial_mean(testConf.paramInitialMean);
+      paraConfig.set_initial_std(testConf.paramInitialStd);
       initParameter(paraName, paraSize, inputDef.isStatic, false, paraConfig);
     }
   }
diff --git a/paddle/gserver/tests/LayerGradUtil.h b/paddle/gserver/tests/LayerGradUtil.h
index 18a6525a145fb..d299b4dd09418 100644
--- a/paddle/gserver/tests/LayerGradUtil.h
+++ b/paddle/gserver/tests/LayerGradUtil.h
@@ -125,12 +125,16 @@ struct TestConfig {
   LayerConfig layerConfig;
   std::vector<InputDef> inputDefs;
   size_t biasSize;
+  real paramInitialMean;
+  real paramInitialStd;
   bool testAccumulate;
   bool testState;
   bool staticBias;
   bool testBatchState;
   TestConfig()
       : biasSize(0),
+        paramInitialMean(0.0),
+        paramInitialStd(1.0),
         testAccumulate(true),
         testState(false),
         staticBias(false),
diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp
index c041f1380cabd..67251f08e34fa 100644
--- a/paddle/gserver/tests/test_LayerGrad.cpp
+++ b/paddle/gserver/tests/test_LayerGrad.cpp
@@ -1669,6 +1669,8 @@ TEST(Layer, PadLayer) {
 
 TEST(Layer, CrossChannelNormLayer) {
   TestConfig config;
+  config.paramInitialMean = 1.;
+  config.paramInitialStd = 0.;
   config.layerConfig.set_type("norm");
   config.layerConfig.set_size(100);
   LayerInputConfig* input = config.layerConfig.add_inputs();
@@ -1682,7 +1684,7 @@ TEST(Layer, CrossChannelNormLayer) {
   config.inputDefs.push_back({INPUT_DATA, "layer_0", 100, 10});
 
   for (auto useGpu : {false, true}) {
-    testLayerGrad(config, "cross-channel-norm", 10, false, useGpu, false, 5);
+    testLayerGrad(config, "cross-channel-norm", 10, false, useGpu, false);
   }
 }