diff --git a/paddle/gserver/layers/Layer.h b/paddle/gserver/layers/Layer.h
index edef36194aabdb9c122ec3423deb036169a34d7c..4002a3d0747a86ab7b495ffe52247521831b71b8 100644
--- a/paddle/gserver/layers/Layer.h
+++ b/paddle/gserver/layers/Layer.h
@@ -49,6 +49,12 @@ struct LayerState {
 };
 typedef std::shared_ptr<LayerState> LayerStatePtr;
 
+/// Paddle device ID, MKLDNN is -2, CPU is -1
+enum PADDLE_DEVICE_ID {
+  MKLDNN_DEVICE = -2,
+  CPU_DEVICE = -1,
+};
+
 /**
  * @brief Base class for layer.
  * Define necessary variables and functions for every layer.
@@ -59,11 +65,6 @@ protected:
   LayerConfig config_;
   /// whether to use GPU
   bool useGpu_;
-  /// Paddle device ID, MKLDNN is -2, CPU is -1
-  enum PADDLE_DEVICE_ID {
-    MKLDNN_DEVICE = -2,
-    CPU_DEVICE = -1,
-  };
   /// Device Id. MKLDNN is -2, CPU is -1, and GPU is 0, 1, 2 ...
   int deviceId_;
   /// Input layers
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp
index 53433cef35a377a73f87b041fdcfadd848dd2ec9..f70343251ad4fbb99f9614618f6d1bff1174f15e 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp
@@ -14,7 +14,6 @@ limitations under the License. */
 
 #include "MKLDNNFcLayer.h"
 #include "paddle/utils/Logging.h"
-#include "paddle/utils/Stat.h"
 
 using namespace mkldnn;  // NOLINT
 typedef memory::format format;
@@ -40,6 +39,8 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap,
   oc_ = getSize();
   oh_ = 1;
   ow_ = 1;
+  ih_ = 1;
+  iw_ = 1;
 
   // input size can not change in FC
   iLayerSize_ = inputLayers_[0]->getSize();
@@ -77,67 +78,53 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
   wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
 }
 
-void MKLDNNFcLayer::reshape() {
-  const Argument& input = getInput(0, getPrev(0)->getDeviceId());
-  int batchSize = input.getBatchSize();
-  if (bs_ == batchSize) {
-    return;
-  }
-  bs_ = batchSize;
-  ih_ = input.getFrameHeight();
-  iw_ = input.getFrameWidth();
-  if (ih_ == 0) {
-    ih_ = 1;
-  }
-  if (iw_ == 0) {
-    iw_ = 1;
-  }
-  CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
-  ic_ = iLayerSize_ / (ih_ * iw_);
-  CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible";
-  CHECK_EQ(size_t(oc_), getSize());
-  printSizeInfo();
+void MKLDNNFcLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
 
-  // reset output
-  output_.setFrameHeight(oh_);
-  output_.setFrameWidth(ow_);
-  resetOutput(bs_, oc_);
+  CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
+  ic = iLayerSize_ / (ih * iw);
+  CHECK_EQ(size_t(ic * ih * iw), iLayerSize_) << "not divisible";
+  CHECK_EQ(size_t(oc), getSize());
 
-  // reset mkldnn forward
-  resetFwd();
-  needResetBwd_ = true;
+  reshapeOutput(oh, ow);
+  resizeOutput(bs, oc);
 
-  convertWeightsFromPaddle();
+  printSizeInfo();
 }
 
-void MKLDNNFcLayer::resetFwd() {
+void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  pipeline.clear();
   bool hasBias = biases_ && biases_->getW();
-  const MatrixPtr& wgt = weight_->getW();
-  const MatrixPtr& bias = hasBias ? biases_->getW() : nullptr;
-  const MatrixPtr& out = output_.value;
+  const MatrixPtr& wgtVal = weight_->getW();
+  const MatrixPtr& biasVal = hasBias ? biases_->getW() : nullptr;
+  const MatrixPtr& outVal = output_.value;
 
   if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& in = getInputValue(0);
-    inVal_ = std::dynamic_pointer_cast<MKLDNNMatrix>(in);
-    CHECK(inVal_) << "Input should be MKLDNNMatrix";
+    const MatrixPtr& inVal = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(inVal);
+    CHECK(in) << "Input should be MKLDNNMatrix";
   } else {
     CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& in = getInputValue(0, CPU_DEVICE);
-    inVal_ = MKLDNNMatrix::create(
-        in, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
-  }
-  inVal_->downSpatial();
-  wgtVal_ = MKLDNNMatrix::create(
-      wgt, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
-  wgtVal_->downSpatial();
-  biasVal_ =
-      hasBias ? MKLDNNMatrix::create(bias, {oc_}, format::x, engine_) : nullptr;
-  outVal_ = MKLDNNMatrix::create(out, {bs_, oc_}, format::nc, engine_);
+    const MatrixPtr& inVal = getInputValue(0, CPU_DEVICE);
+    in = MKLDNNMatrix::create(
+        inVal, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
+  }
+  in->downSpatial();
+  wgt = MKLDNNMatrix::create(
+      wgtVal, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
+  wgt->downSpatial();
+  bias = hasBias ? MKLDNNMatrix::create(biasVal, {oc_}, format::x, engine_)
+                 : nullptr;
+  out = MKLDNNMatrix::create(outVal, {bs_, oc_}, format::nc, engine_);
 
   // change original output value to mkldnn output value
-  output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
   if (!outputIsOnlyMKLDNN()) {
-    copyOutputInfoToOtherDevice();
     // fc cpu output value do not need create convert
     // just share point
     getOutput(CPU_DEVICE).value->setData(output_.value->getData());
@@ -146,27 +133,31 @@ void MKLDNNFcLayer::resetFwd() {
   // create forward handle
   prop_kind pk = prop_kind::forward;
   fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk,
-                                                inVal_->getMemoryDesc(),
-                                                wgtVal_->getMemoryDesc(),
-                                                biasVal_->getMemoryDesc(),
-                                                outVal_->getMemoryDesc())
+                                                in->getMemoryDesc(),
+                                                wgt->getMemoryDesc(),
+                                                bias->getMemoryDesc(),
+                                                out->getMemoryDesc())
                                  : fc_fwd::desc(pk,
-                                                inVal_->getMemoryDesc(),
-                                                wgtVal_->getMemoryDesc(),
-                                                outVal_->getMemoryDesc());
+                                                in->getMemoryDesc(),
+                                                wgt->getMemoryDesc(),
+                                                out->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
   if (hasBias) {
-    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_));
+    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *bias, *out));
   } else {
-    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_));
+    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *out));
   }
   printValueFormatFlow();
 
-  pipelineFwd_.clear();
-  pipelineFwd_.push_back(*fwd_);
+  pipeline.push_back(*fwd_);
 }
 
-void MKLDNNFcLayer::resetBwd() {
+void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  pipeline.clear();
   if (!needResetBwd_) {
     return;
   }
@@ -175,8 +166,8 @@ void MKLDNNFcLayer::resetBwd() {
 
   /// backward weight
   CHECK(inVal_) << "Should have input value";
-  const MatrixPtr& wgt = weight_->getWGrad();
-  const MatrixPtr& bias = hasBias ? biases_->getWGrad() : nullptr;
+  const MatrixPtr& wgtGrad = weight_->getWGrad();
+  const MatrixPtr& biasGrad = hasBias ? biases_->getWGrad() : nullptr;
 
   // TODO(TJ): merge outgrad
   int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
@@ -187,107 +178,66 @@ void MKLDNNFcLayer::resetBwd() {
   // for CPU device:
   // fc do not need to convert from cpu device since output is always nc format
   // only need create from cpu device
-  const MatrixPtr& out = getOutput(device).grad;
-  outGrad_ = MKLDNNMatrix::create(out, outVal_->getPrimitiveDesc());
-  wgtGrad_ = MKLDNNMatrix::create(wgt, wgtVal_->getPrimitiveDesc());
-  biasGrad_ = hasBias ? MKLDNNMatrix::create(bias, biasVal_->getPrimitiveDesc())
-                      : nullptr;
+  const MatrixPtr& outGrad = getOutput(device).grad;
+  out = MKLDNNMatrix::create(outGrad, outVal_->getPrimitiveDesc());
+  wgt = MKLDNNMatrix::create(wgtGrad, wgtVal_->getPrimitiveDesc());
+  bias = hasBias ? MKLDNNMatrix::create(biasGrad, biasVal_->getPrimitiveDesc())
+                 : nullptr;
 
   // create memory primitive desc
   fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
                                       inVal_->getMemoryDesc(),
-                                      wgtGrad_->getMemoryDesc(),
-                                      outGrad_->getMemoryDesc());
+                                      wgt->getMemoryDesc(),
+                                      out->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
   fc_bwdWgt::desc bwdWgtDesc = hasBias
                                    ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgtGrad_->getMemoryDesc(),
-                                                     biasGrad_->getMemoryDesc(),
-                                                     outGrad_->getMemoryDesc())
+                                                     wgt->getMemoryDesc(),
+                                                     bias->getMemoryDesc(),
+                                                     out->getMemoryDesc())
                                    : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgtGrad_->getMemoryDesc(),
-                                                     outGrad_->getMemoryDesc());
+                                                     wgt->getMemoryDesc(),
+                                                     out->getMemoryDesc());
   fc_bwdWgt::primitive_desc bwdWgtPD =
       fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
 
   if (hasBias) {
-    bwdWgt_.reset(
-        new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_, *biasGrad_));
+    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt, *bias));
   } else {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_));
+    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt));
   }
-  pipelineBwd_.clear();
-  pipelineBwd_.push_back(*bwdWgt_);
+  pipeline.push_back(*bwdWgt_);
 
   /// backward data
-  const MatrixPtr& in = inputLayers_[0]->getOutput().grad;
-  if (in == nullptr) {
+  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inGrad == nullptr) {
     return;
   }
   if (getInput(0, MKLDNN_DEVICE).getAllCount() > 1) {
     // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
   } else {
-    inGrad_ = MKLDNNMatrix::create(in, inVal_->getPrimitiveDesc());
+    in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
   }
 
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(inVal_->getMemoryDesc(),
-                                                  wgtGrad_->getMemoryDesc(),
-                                                  outGrad_->getMemoryDesc());
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
+      inVal_->getMemoryDesc(), wgt->getMemoryDesc(), out->getMemoryDesc());
   fc_bwdData::primitive_desc bwdDataPD =
       fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
 
   CHECK(wgtVal_) << "Should have weight memory";
-  bwdData_.reset(new fc_bwdData(bwdDataPD, *outGrad_, *wgtVal_, *inGrad_));
+  bwdData_.reset(new fc_bwdData(bwdDataPD, *out, *wgtVal_, *in));
   printGradFormatFlow();
-  pipelineBwd_.push_back(*bwdData_);
+  pipeline.push_back(*bwdData_);
 }
 
 void MKLDNNFcLayer::updateInputData() {
-  if (inputLayers_[0]->getType() != "data") {
-    return;
-  }
-  real* iData = getInputValue(0, CPU_DEVICE)->getData();
-  inVal_->setData(iData);
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
 }
 
-void MKLDNNFcLayer::forward(PassType passType) {
-  Layer::forward(passType);
-  reshape();
-
-  {
-    REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
-    updateInputData();
-
-    // just submit forward pipeline
-    stream_->submit(pipelineFwd_);
-  }
-
-  /* activation */ {
-    REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
-    forwardActivation();
-  }
-}
-
-void MKLDNNFcLayer::backward(const UpdateCallback& callback) {
-  /* Do derivation */ {
-    REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
-    backwardActivation();
-  }
-
-  {
-    REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
-    resetBwd();
-
-    // just sumbmit backward pipeline
-    stream_->submit(pipelineBwd_);
-  }
-
-  {
-    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
-    weight_->getParameterPtr()->incUpdate(callback);
-    if (biases_ && biases_->getWGrad()) {
-      biases_->getParameterPtr()->incUpdate(callback);
-    }
+void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
+  weight_->getParameterPtr()->incUpdate(callback);
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
   }
 }
 }  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h
index 4ad67a16e056a718c45a28babcf22a7cd571b15c..3119f863496df092da13c08bf733f13c42e53780 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.h
+++ b/paddle/gserver/layers/MKLDNNFcLayer.h
@@ -45,35 +45,28 @@ public:
   bool init(const LayerMap& layerMap,
             const ParameterMap& parameterMap) override;
 
-  void convertWeightsFromPaddle() override;
-
-  void convertWeightsToPaddle() override;
+  void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
 
-  void forward(PassType passType) override;
+  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
 
-  void backward(const UpdateCallback& callback) override;
+  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
 
   void updateInputData() override;
 
-protected:
-  /**
-   * reshape the input image sizes
-   * and reset output buffer size
-   * and reset mkldnn forward
-   */
-  void reshape();
-
-  /**
-   * reset the forward primitve and memory
-   * only would be called when input size changes
-   */
-  void resetFwd();
-
-  /**
-   * reset the backward primitve and memory for mkldnn fc
-   * only would be called when needed
-   */
-  void resetBwd();
+  void updateWeights(const UpdateCallback& callback) override;
+
+  void convertWeightsFromPaddle() override;
+
+  void convertWeightsToPaddle() override;
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h
index 543364edceff684bdcd002a8f4f10e7ce5e6953b..169679c8297542cac4a43f5a8e1af311ad9282df 100644
--- a/paddle/gserver/layers/MKLDNNLayer.h
+++ b/paddle/gserver/layers/MKLDNNLayer.h
@@ -19,6 +19,7 @@ limitations under the License. */
 #include "MKLDNNBase.h"
 #include "mkldnn.hpp"
 #include "paddle/math/MKLDNNMatrix.h"
+#include "paddle/utils/Stat.h"
 
 DECLARE_bool(use_mkldnn);
 
@@ -33,6 +34,8 @@ typedef std::shared_ptr<MKLDNNLayer> MKLDNNLayerPtr;
  */
 class MKLDNNLayer : public Layer {
 protected:
+  // input value element count
+  size_t inputElemenCnt_;
   // batch size
   int bs_;
   // input image channel, height and width
@@ -52,7 +55,7 @@ protected:
   std::vector<mkldnn::primitive> pipelineFwd_;
   std::vector<mkldnn::primitive> pipelineBwd_;
 
-  // MKLDNNMatrixPtr
+  // MKLDNNMatrixPtr with internal format
   MKLDNNMatrixPtr inVal_;
   MKLDNNMatrixPtr inGrad_;
   MKLDNNMatrixPtr outVal_;
@@ -65,6 +68,7 @@ protected:
 public:
   explicit MKLDNNLayer(const LayerConfig& config)
       : Layer(config),
+        inputElemenCnt_(0),
         bs_(0),
         ic_(0),
         ih_(0),
@@ -95,12 +99,104 @@ public:
     if (!Layer::init(layerMap, parameterMap)) {
       return false;
     }
+    checkCPUOutputsNumber();
 
     stream_.reset(new MKLDNNStream());
     engine_ = CPUEngine::Instance().getEngine();
     return true;
   }
 
+  void forward(PassType passType) override {
+    passType_ = passType;
+
+    {
+      REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
+      CHECK(!inputLayers_.empty());
+      copySeqInfoToOutputs();
+      size_t elemenCnt = inputLayers_[0]->getOutput().value->getElementCnt();
+      if (inputElemenCnt_ != elemenCnt) {
+        // reset when input total sizes changed, not only the batchsize
+        inputElemenCnt_ = elemenCnt;
+        reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
+        resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
+        convertWeightsFromPaddle();
+        needResetBwd_ = true;
+      }
+
+      if (inputLayers_[0]->getType() == "data") {
+        updateInputData();
+      }
+
+      stream_->submit(pipelineFwd_);
+    }
+
+    /* activation */ {
+      REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
+      forwardActivation();
+    }
+  }
+
+  void backward(const UpdateCallback& callback) override {
+    /* Do derivation */ {
+      REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
+      backwardActivation();
+    }
+
+    {
+      REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
+      if (needResetBwd_) {
+        resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
+        needResetBwd_ = false;
+      }
+
+      stream_->submit(pipelineBwd_);
+    }
+
+    {
+      REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+      updateWeights(callback);
+    }
+  }
+
+  /**
+   * reshape the input image sizes
+   * and reset output image and buffer size
+   * output channel can not be changed
+   */
+  virtual void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) = 0;
+
+  /**
+   * reset the mkldnn forward primitve and memory
+   * only would be called when input size changes
+   */
+  virtual void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out) = 0;
+
+  /**
+   * reset the mkldnn backward primitve and memory for mkldnn fc
+   * only would be called when needed
+   */
+  virtual void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out) = 0;
+
+  /**
+   * Update input value data when input layer is "data" type.
+   * Since the input value data address might be changed.
+   */
+  virtual void updateInputData() {}
+
+  /**
+   * Update weights and biases if necessary.
+   */
+  virtual void updateWeights(const UpdateCallback& callback) {}
+
   /**
    * convert weight from paddle format to mkldnn format
    * weight_ will be override
@@ -114,10 +210,38 @@ public:
   virtual void convertWeightsToPaddle() {}
 
   /**
-   * Update input value data when input layer is "data" type.
-   * Since the input value data address might be changed.
+   * add this interface as public for unit test
    */
-  virtual void updateInputData() {}
+  void addOutputArgument(int deviceId) { Layer::addOutputArgument(deviceId); }
+
+protected:
+  /**
+   * reshape the input image sizes and input batchsize
+   */
+  virtual void reshapeInput(int& batchsize, int& height, int& width) {
+    const Argument& input = inputLayers_[0]->getOutput();
+    batchsize = input.getBatchSize();
+    int h = input.getFrameHeight();
+    int w = input.getFrameWidth();
+    if (h != 0) {
+      height = h;
+    }
+    if (w != 0) {
+      width = w;
+    }
+  }
+
+  /**
+   * reshape output image sizes
+   */
+  virtual void reshapeOutput(size_t height, size_t width) {
+    output_.setFrameHeight(height);
+    output_.setFrameWidth(width);
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].setFrameHeight(height);
+      outputOtherDevice_[i].setFrameWidth(width);
+    }
+  }
 
   /**
    * print info about sizes
@@ -133,8 +257,8 @@ public:
    */
   virtual void printValueFormatFlow() {
     if (inVal_ && outVal_) {
-      VLOG(MKLDNN_FMTS) << "value format flow --- " << inVal_->getFormat()
-                        << " >>> " << outVal_->getFormat();
+      VLOG(MKLDNN_FMTS) << inVal_->getFormat() << " >>> "
+                        << outVal_->getFormat();
     }
   }
 
@@ -143,36 +267,12 @@ public:
    */
   virtual void printGradFormatFlow() {
     if (inGrad_ && outGrad_) {
-      VLOG(MKLDNN_FMTS) << "grad format flow --- " << inGrad_->getFormat()
-                        << " <<< " << outGrad_->getFormat();
+      VLOG(MKLDNN_FMTS) << inGrad_->getFormat() << " <<< "
+                        << outGrad_->getFormat();
     }
   }
 
 protected:
-  /**
-   * copy image size and sequence info to other device
-   * @note: can not directly use Layer::copyOutputToOtherDevice since here only
-   *        copy base info and do not copy data value
-   */
-  void copyOutputInfoToOtherDevice() {
-    int cnt = 0;
-    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
-      outputOtherDevice_[i].setFrameHeight(output_.getFrameHeight());
-      outputOtherDevice_[i].setFrameWidth(output_.getFrameWidth());
-      outputOtherDevice_[i].sequenceStartPositions =
-          output_.sequenceStartPositions;
-      outputOtherDevice_[i].subSequenceStartPositions =
-          output_.subSequenceStartPositions;
-      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
-      if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
-        ++cnt;
-      }
-    }
-    if (cnt > 1) {
-      LOG(WARNING) << "should not have more than one CPU devie";
-    }
-  }
-
   /**
    * If input only has MKLDNN device.
    * Otherwise, only support the previous layer using CPU device.
@@ -205,6 +305,7 @@ protected:
    */
   void setDevice(int id) { deviceId_ = id; }
 
+private:
   /**
    * Set deviceId of the params used in this layer.
    */
@@ -228,6 +329,42 @@ protected:
       parameter->setDevice(id);
     }
   }
+
+  /**
+   * Check the cpu device number of outputOtherDevice_.
+   * should have only one at most.
+   */
+  void checkCPUOutputsNumber(int max = 1) {
+    int cnt = 0;
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
+        ++cnt;
+      }
+    }
+    CHECK_LE(cnt, max) << "too much CPU devies";
+  }
+
+  /**
+   * copy SeqInfo from input layer to this output and other output devices.
+   * @note: do not use getInput(0) since it used this deviceId_,
+   *        use "inputLayers_[0]->getOutput()" instead.
+   */
+  void copySeqInfoToOutputs() {
+    if (inputLayers_.empty() || !needSequenceInfo_) {
+      return;
+    }
+    const Argument& input = inputLayers_[0]->getOutput();
+    output_.sequenceStartPositions = input.sequenceStartPositions;
+    output_.subSequenceStartPositions = input.subSequenceStartPositions;
+    output_.cpuSequenceDims = input.cpuSequenceDims;
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].sequenceStartPositions =
+          output_.sequenceStartPositions;
+      outputOtherDevice_[i].subSequenceStartPositions =
+          output_.subSequenceStartPositions;
+      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
+    }
+  }
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/tests/MKLDNNTester.cpp b/paddle/gserver/tests/MKLDNNTester.cpp
index de1635be2af37cd0ba49010199a417090865b0e4..2f48e5b2d3ffc9337ed1314f6db6549e56263fdd 100644
--- a/paddle/gserver/tests/MKLDNNTester.cpp
+++ b/paddle/gserver/tests/MKLDNNTester.cpp
@@ -63,8 +63,12 @@ void MKLDNNTester::reset(const TestConfig& dnn,
     initTestLayer(
         configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i]));
   }
-  dnnLayer_ = testLayers_[DNN];
   refLayer_ = testLayers_[REF];
+  dnnLayer_ = std::dynamic_pointer_cast<MKLDNNLayer>(testLayers_[DNN]);
+  CHECK(dnnLayer_);
+  // for comparison with Paddle reference results,
+  // need manually add cpu device output for test
+  dnnLayer_->addOutputArgument(CPU_DEVICE);
   EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size());
   EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size());
 
@@ -109,20 +113,22 @@ void MKLDNNTester::randomBotDatas() {
 
 void MKLDNNTester::randomTopDiffs() {
   refLayer_->getOutputGrad()->randomizeUniform();
-  dnnLayer_->getOutputGrad()->copyFrom(*(refLayer_->getOutputGrad()));
-  VLOG(lvl_) << "Random dom Backward Input, TopDiff: ";
+  dnnLayer_->getOutput(CPU_DEVICE)
+      .grad->copyFrom(*(refLayer_->getOutputGrad()));
+  VLOG(lvl_) << "Random Backward Input, TopDiff: ";
   printMatrix(refLayer_->getOutputGrad());
 }
 
 void MKLDNNTester::checkForward() {
-  printTopDatas();
-  double delta = compareMatrix(testLayers_[DNN]->getOutputValue(),
-                               testLayers_[REF]->getOutputValue());
   VLOG(MKLDNN_ALL) << "Check Forward";
+  printTopDatas();
+  double delta = compareMatrix(dnnLayer_->getOutput(-1).value,
+                               refLayer_->getOutputValue());
   EXPECT_LE(fabs(delta), eps_);
 }
 
 void MKLDNNTester::checkBackwardData() {
+  VLOG(MKLDNN_ALL) << "Check Backward Data";
   // TODO(TJ): uncomment me when batch norm ready
   // const bool isBN = dnnLayer_->getType() == "mkldnn_batch_norm";
   for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) {
@@ -144,14 +150,12 @@ void MKLDNNTester::checkBackwardData() {
 }
 
 void MKLDNNTester::checkBackwardWgts() {
+  VLOG(MKLDNN_ALL) << "Check Backward Weight";
   CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
   vector<VectorPtr> dnnWgts;  // used to temply save mkldnn weights
   saveWgt(parameters_[DNN], dnnWgts);
 
-  const MKLDNNLayerPtr dnnlayer =
-      std::dynamic_pointer_cast<MKLDNNLayer>(dnnLayer_);
-  CHECK(dnnlayer);
-  dnnlayer->convertWeightsToPaddle();
+  dnnLayer_->convertWeightsToPaddle();
   for (size_t i = 0; i < parameters_[DNN].size(); ++i) {
     const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE);
     const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE);
@@ -189,38 +193,38 @@ void MKLDNNTester::restoreWgt(const vector<VectorPtr>& from,
 }
 
 // clear parameters grad
-void MKLDNNTester::clearWgtDiffs() {
+void MKLDNNTester::clearWgtDiffs(size_t id) {
+  CHECK_LE(id, parameters_.size());
   for (size_t n = 0; n < parameters_.size(); ++n) {
-    for (size_t i = 0; i < parameters_[n].size(); ++i) {
-      const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT);
-      if (grad) {
-        grad->zeroMem();
+    if (id == n || id == parameters_.size()) {
+      for (size_t i = 0; i < parameters_[n].size(); ++i) {
+        const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT);
+        if (grad) {
+          grad->zeroMem();
+        }
       }
     }
   }
 }
 
-void MKLDNNTester::clearBotDiffs() {
-  // dnn and ref
+void MKLDNNTester::clearBotDiffs(size_t id) {
+  CHECK_LE(id, dataLayers_.size());
   for (size_t n = 0; n < dataLayers_.size(); ++n) {
-    // all inputs layers
-    for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
-      dataLayers_[n][i]->getOutputGrad()->zeroMem();
+    if (id == n || id == dataLayers_.size()) {
+      // clear inputs layers of this specific layer
+      for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
+        dataLayers_[n][i]->getOutputGrad()->zeroMem();
+      }
     }
   }
 }
 
-void MKLDNNTester::clearBotDiffs(int n) {
-  CHECK_LT(n, NUM);
-  // all inputs layers
-  for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
-    dataLayers_[n][i]->getOutputGrad()->zeroMem();
-  }
-}
-
-void MKLDNNTester::clearTopDatas() {
+void MKLDNNTester::clearTopDatas(size_t id) {
+  CHECK_LE(id, testLayers_.size());
   for (size_t i = 0; i < testLayers_.size(); ++i) {
-    testLayers_[i]->getOutputValue()->zeroMem();
+    if (id == i || id == testLayers_.size()) {
+      testLayers_[i]->getOutputValue()->zeroMem();
+    }
   }
 }
 
@@ -300,16 +304,24 @@ void MKLDNNTester::runOnce() {
   checkForward();
 
   // test backward
+  // simple updater
+  UpdateCallback updateCallback = [](Parameter* para) {
+    auto& grad = para->getBuf(PARAMETER_GRADIENT);
+    auto& value = para->getBuf(PARAMETER_VALUE);
+    real lr = 1e-3;
+    value->add(*grad, lr);
+  };
   randomTopDiffs();
-  dnnLayer_->backward(nullptr);
-  refLayer_->backward(nullptr);
+  dnnLayer_->backward(updateCallback);
+  refLayer_->backward(updateCallback);
   checkBackwardData();
   checkBackwardWgts();
 
   // clear buffers
   // ref code will addto the diff, dnn code will writeto it
-  // and clearTopDatas() and clearWgtDiffs() should be coverd by test layers
+  // and clearTopDatas(REF) should be coverd by ref layers
   clearBotDiffs(REF);
+  clearWgtDiffs(REF);
 }
 
 void MKLDNNTester::run(const TestConfig& dnn,
diff --git a/paddle/gserver/tests/MKLDNNTester.h b/paddle/gserver/tests/MKLDNNTester.h
index e55e4493ffdfe45b8cfdee423febd1878b8b3d8a..5ac885638cde7693a0c847733e7a6149c1b7e6c2 100644
--- a/paddle/gserver/tests/MKLDNNTester.h
+++ b/paddle/gserver/tests/MKLDNNTester.h
@@ -18,6 +18,7 @@ limitations under the License. */
 #include <vector>
 #include "LayerGradUtil.h"
 #include "paddle/gserver/layers/MKLDNNBase.h"
+#include "paddle/gserver/layers/MKLDNNLayer.h"
 
 namespace paddle {
 
@@ -40,7 +41,8 @@ protected:
   vector<LayerMap> layerMaps_;
   vector<vector<ParameterPtr>> parameters_;
   vector<LayerPtr> testLayers_;
-  LayerPtr dnnLayer_, refLayer_;
+  LayerPtr refLayer_;
+  MKLDNNLayerPtr dnnLayer_;
 
   /// run some iterations, all the result should pass
   size_t iter_;
@@ -88,10 +90,10 @@ private:
   void checkBackwardData();
   void checkBackwardWgts();
 
-  void clearWgtDiffs();
-  void clearBotDiffs();
-  void clearBotDiffs(int n);  // clear specific layer
-  void clearTopDatas();
+  // clear specific layer, clear all when id equals NUM
+  void clearWgtDiffs(size_t id = NUM);
+  void clearBotDiffs(size_t id = NUM);
+  void clearTopDatas(size_t id = NUM);
 
   void printTopDatas();
   void printMatrix(const MatrixPtr& m);