should run resetBwd before bwdAct

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -449,13 +449,14 @@ void MKLDNNConvLayer::resetOutGrad(
   cvtOutGrad_ = nullptr;
   if (!outputIsOnlyMKLDNN()) {
     const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
+    outMat->setData(cpuOut->getData());
     // same PrimitiveDesc with cpuInVal_
     CHECK(cpuOutVal_);
     cpuOutGrad_ = MKLDNNMatrix::create(cpuOut, cpuOutVal_->getPrimitiveDesc());
     if (cpuOutGrad_->getPrimitiveDesc() == out->getPrimitiveDesc()) {
-      outMat->setData(cpuOut->getData());
       out = cpuOutGrad_;
     } else {
+      out = MKLDNNMatrix::create(nullptr, wgtPD->diff_dst_primitive_desc());
       cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
       CHECK(cvtOutGrad_);
     }

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -232,6 +232,7 @@ void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
 void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
   // TODO(TJ): merge outgrad
   int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
+  output_.grad->setData(getOutput(device).grad->getData());
   // for MKLDNN device:
   // can not directly cast outputgrad to mkldnnmatrix,
   // since each layer can not write the inputgrad to mkldnn inputgrad.

paddle/gserver/layers/MKLDNNLayer.h

@@ -141,18 +141,16 @@ public:
   }
 
   void backward(const UpdateCallback& callback) override {
-    /* Do derivation */ {
+    if (needResetBwd_) {
+      resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
+      needResetBwd_ = false;
+    }
+    {
       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_);
     }