diff --git a/paddle/gserver/layers/MKLDNNConvLayer.cpp b/paddle/gserver/layers/MKLDNNConvLayer.cpp
index 9088744beebd25ac105737fe3b012de143c66a7c..2647cb600653b4f43322016afb231a55f4db5642 100644
--- a/paddle/gserver/layers/MKLDNNConvLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNConvLayer.cpp
@@ -294,12 +294,9 @@ void MKLDNNConvLayer::resetOutValue(
     std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& out) {
   out = MKLDNNMatrix::create(output_.value, pd->dst_primitive_desc());
 
-  // change original output value from cpu matrix to mkldnn matrix
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
-
   // create reorder if output value has cpu device and pd do not match
   cpuOutVal_ = nullptr;
-  cpuOutVal_ = nullptr;
+  cvtOutVal_ = nullptr;
   if (!outputIsOnlyMKLDNN()) {
     const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
     memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp
index f60e221a6ec2ff513789a24e9f59bb25aef437b5..66b358bcea53f61ddcc15323704fa9f154fb2a73 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp
@@ -172,12 +172,10 @@ void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
 
 void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
   out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
-  // change original output value to mkldnn output value
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
   if (!outputIsOnlyMKLDNN()) {
     // fc cpu output value do not need create convert
     // just share point
-    getOutput(CPU_DEVICE).value->setData(output_.value->getData());
+    getOutput(CPU_DEVICE).value->setData(out->getData());
   }
 }
 
diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h
index 169679c8297542cac4a43f5a8e1af311ad9282df..c4e4a6874e6fdb491c344c70dfea422dc0924cd9 100644
--- a/paddle/gserver/layers/MKLDNNLayer.h
+++ b/paddle/gserver/layers/MKLDNNLayer.h
@@ -119,6 +119,10 @@ public:
         inputElemenCnt_ = elemenCnt;
         reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
         resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
+        if (outVal_) {
+          // change original output value to mkldnn output value
+          output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
+        }
         convertWeightsFromPaddle();
         needResetBwd_ = true;
       }
diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.cpp b/paddle/gserver/layers/MKLDNNPoolLayer.cpp
index 48b2f5a4cb37f6a9c4b1fdc6178c914b46c76e63..b62dfb7c54258a593aa50d5b30096423f375c69d 100644
--- a/paddle/gserver/layers/MKLDNNPoolLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNPoolLayer.cpp
@@ -134,7 +134,6 @@ void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
   memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
   out = MKLDNNMatrix::create(
       output_.value, outDims, inVal_->getFormat(), engine_);
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
 
   // create reorder if output value has cpu device and pd do not match
   cpuOutVal_ = nullptr;
diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp
index b593f65fe49ef2271ad7cd0f609c9b828be03037..7620365efa345dbd78605223994a425b6582c259 100644
--- a/paddle/gserver/tests/test_MKLDNN.cpp
+++ b/paddle/gserver/tests/test_MKLDNN.cpp
@@ -162,7 +162,6 @@ void testPoolLayer(const testPoolDesc& pm) {
        0});
   LayerInputConfig* input = cfg.layerConfig.add_inputs();
   PoolConfig* pool = input->mutable_pool_conf();
-  // pool->set_pool_type(poolType);
   pool->set_channels(pm.ch);
   pool->set_img_size(pm.iw);
   pool->set_img_size_y(pm.ih);