Merge pull request #5815 from tensor-tang/googlenet

remove the tmp buffer when merge grad

paddle/gserver/layers/MKLDNNLayer.cpp

@@ -294,22 +294,8 @@ void MKLDNNLayer::resetMergeGrad(MKLDNNMatrixPtr& out) {
     srcs.push_back(*src);
   }
 
-  // TODO(TJ): remove me when mkldnn sum support different formats
-  for (size_t i = 1; i < srcPDs.size(); ++i) {
-    CHECK(srcPDs[0] == srcPDs[i]);
-  }
-  tmpOutGrad_ = out;
-  tmpCvt_ = nullptr;
-  if (out->getPrimitiveDesc() != srcPDs[0]) {
-    tmpOutGrad_ = MKLDNNMatrix::create(srcPDs[0]);
-    tmpCvt_ = MKLDNNMatrix::createReorder(tmpOutGrad_, out);
-    CHECK(tmpCvt_);
-    pipelineMergeGrad_.push_back(*tmpCvt_);
-  }
-
-  auto sumPD =
-      sum::primitive_desc(tmpOutGrad_->getMemoryDesc(), scales, srcPDs);
-  mergeGrad_.reset(new sum(sumPD, srcs, *tmpOutGrad_));
+  auto sumPD = sum::primitive_desc(out->getMemoryDesc(), scales, srcPDs);
+  mergeGrad_.reset(new sum(sumPD, srcs, *out));
   pipelineMergeGrad_.insert(pipelineMergeGrad_.begin(), *mergeGrad_);
 }
 

paddle/gserver/layers/MKLDNNLayer.h

@@ -94,11 +94,6 @@ protected:
   std::vector<mkldnn::primitive> pipelineMergeGrad_;
   // tmp input argument to save input grad, only used to merge grad
   Argument tmpInArg_;
-  // since mkldnn sum do not support different formats:
-  // can refer to https://github.com/01org/mkl-dnn/issues/134
-  // so need create reorder manually and save tmp MKLDNNMatrix
-  MKLDNNMatrixPtr tmpOutGrad_;
-  std::shared_ptr<mkldnn::primitive> tmpCvt_;
 
 public:
   explicit MKLDNNLayer(const LayerConfig& config)