make MKLDNNLayer input grad as a vector

paddle/gserver/layers/MKLDNNAddtoLayer.cpp

@@ -69,16 +69,15 @@ void MKLDNNAddtoLayer::resetFwd(std::vector<primitive>& pipeline,
 }
 
 void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
-                                MKLDNNMatrixPtr& in,
+                                std::vector<MKLDNNMatrixPtr>& inputs,
                                 MKLDNNMatrixPtr& out) {
-  resetBwdBuffers(inGrads_, biasGrad_, out);
-  in = inGrads_[0];
+  resetBwdBuffers(inputs, biasGrad_, out);
 
   // backward only need share output grad to input grad
-  for (size_t i = 0; i < inGrads_.size(); i++) {
-    if (inGrads_[i] != nullptr) {
-      inGrads_[i] = out;
-      inputLayers_[i]->getOutputGrad()->setData(inGrads_[i]->getData());
+  for (size_t i = 0; i < inputs.size(); i++) {
+    if (inputs[i] != nullptr) {
+      inputs[i] = out;
+      inputLayers_[i]->getOutputGrad()->setData(inputs[i]->getData());
     }
   }
 

paddle/gserver/layers/MKLDNNAddtoLayer.h

@@ -26,8 +26,6 @@ namespace paddle {
  */
 class MKLDNNAddtoLayer : public MKLDNNLayer {
 protected:
-  std::vector<MKLDNNMatrixPtr> inGrads_;
-
   // layer size == ic * ih * iw == oc * oh *ow, and can not be changed
   size_t layerSize_;
 
@@ -56,23 +54,11 @@ public:
                 MKLDNNMatrixPtr& out) override;
 
   void resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                MKLDNNMatrixPtr& in,
+                std::vector<MKLDNNMatrixPtr>& inputs,
                 MKLDNNMatrixPtr& out) override;
 
   void updateWeights(const UpdateCallback& callback) override;
 
-  void printGradFormat() override {
-    if (extOutGrad_) {
-      VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat();
-    }
-    if (outGrad_) {
-      VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< ";
-    }
-    for (size_t i = 0; i < inGrads_.size(); ++i) {
-      VLOG(MKLDNN_FMTS) << i << " input: " << inGrads_[i]->getFormat() << "<<<";
-    }
-  }
-
 protected:
   void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
                        MKLDNNMatrixPtr& bias,

paddle/gserver/layers/MKLDNNBatchNormLayer.cpp

@@ -146,15 +146,15 @@ void MKLDNNBatchNormLayer::resetFwd(std::vector<primitive>& pipeline,
 }
 
 void MKLDNNBatchNormLayer::resetBwd(std::vector<primitive>& pipeline,
-                                    MKLDNNMatrixPtr& in,
+                                    std::vector<MKLDNNMatrixPtr>& inputs,
                                     MKLDNNMatrixPtr& out) {
   std::shared_ptr<bn_bwd::primitive_desc> pd;
 
-  resetBwdBuffers(in, wgtGrad_, out);
+  resetBwdBuffers(inputs[0], wgtGrad_, out);
 
-  resetBwdPD(pd, in, wgtGrad_, out);
+  resetBwdPD(pd, inputs[0], wgtGrad_, out);
 
-  resetBwdPipeline(pipeline, pd, in, wgtGrad_, out);
+  resetBwdPipeline(pipeline, pd, inputs[0], wgtGrad_, out);
 }
 
 void MKLDNNBatchNormLayer::forward(PassType passType) {

paddle/gserver/layers/MKLDNNBatchNormLayer.h

@@ -80,7 +80,7 @@ public:
                 MKLDNNMatrixPtr& out) override;
 
   void resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                MKLDNNMatrixPtr& in,
+                std::vector<MKLDNNMatrixPtr>& inputs,
                 MKLDNNMatrixPtr& out) override;
 
   void updateWeights(const UpdateCallback& callback) override;

paddle/gserver/layers/MKLDNNConcatLayer.cpp

@@ -70,12 +70,11 @@ void MKLDNNConcatLayer::resetFwd(std::vector<primitive>& pipeline,
 }
 
 void MKLDNNConcatLayer::resetBwd(std::vector<primitive>& pipeline,
-                                 MKLDNNMatrixPtr& in,
+                                 std::vector<MKLDNNMatrixPtr>& inputs,
                                  MKLDNNMatrixPtr& out) {
-  resetBwdBuffers(inGrads_, out);
-  in = inGrads_[0];
+  resetBwdBuffers(inputs, out);
 
-  resetBwdPipeline(pipeline, bwds_, inGrads_, out);
+  resetBwdPipeline(pipeline, bwds_, inputs, out);
 }
 
 void MKLDNNConcatLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,

paddle/gserver/layers/MKLDNNConcatLayer.h

@@ -26,7 +26,6 @@ namespace paddle {
  */
 class MKLDNNConcatLayer : public MKLDNNLayer {
 protected:
-  std::vector<MKLDNNMatrixPtr> inGrads_;
   std::vector<std::shared_ptr<mkldnn::primitive>> bwds_;
   // input channel numbers
   std::vector<int> channels_;
@@ -53,7 +52,7 @@ public:
                 MKLDNNMatrixPtr& out) override;
 
   void resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                MKLDNNMatrixPtr& in,
+                std::vector<MKLDNNMatrixPtr>& inputs,
                 MKLDNNMatrixPtr& out) override;
 
   void printSizeInfo() override {
@@ -67,19 +66,6 @@ public:
                        << ", " << ow_;
   }
 
-  void printGradFormat() override {
-    if (extOutGrad_) {
-      VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat();
-    }
-    if (outGrad_) {
-      VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< ";
-    }
-    for (size_t i = 0; i < inGrads_.size(); ++i) {
-      VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName()
-                        << ": " << inGrads_[i]->getFormat() << "<<<";
-    }
-  }
-
 protected:
   void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
                        MKLDNNMatrixPtr& out);

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -115,7 +115,7 @@ void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,
 }
 
 void MKLDNNConvLayer::resetBwd(std::vector<primitive>& pipeline,
-                               MKLDNNMatrixPtr& in,
+                               std::vector<MKLDNNMatrixPtr>& inputs,
                                MKLDNNMatrixPtr& out) {
   std::shared_ptr<conv_bwdWgt::primitive_desc> bwdWgtPD;
   std::shared_ptr<conv_bwdData::primitive_desc> bwdDataPD;
@@ -124,9 +124,10 @@ void MKLDNNConvLayer::resetBwd(std::vector<primitive>& pipeline,
 
   resetBwdDataPD(bwdDataPD);
 
-  resetBwdBuffers(bwdWgtPD, bwdDataPD, in, wgtGrad_, biasGrad_, out);
+  resetBwdBuffers(bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out);
 
-  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgtGrad_, biasGrad_, out);
+  resetBwdPipeline(
+      pipeline, bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out);
 }
 
 void MKLDNNConvLayer::updateWeights(const UpdateCallback& callback) {

paddle/gserver/layers/MKLDNNConvLayer.h

@@ -76,7 +76,7 @@ public:
                 MKLDNNMatrixPtr& out) override;
 
   void resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                MKLDNNMatrixPtr& in,
+                std::vector<MKLDNNMatrixPtr>& inputs,
                 MKLDNNMatrixPtr& out) override;
 
   void updateWeights(const UpdateCallback& callback) override;

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -97,18 +97,19 @@ void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
 }
 
 void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
-                             MKLDNNMatrixPtr& in,
+                             std::vector<MKLDNNMatrixPtr>& inputs,
                              MKLDNNMatrixPtr& out) {
   std::shared_ptr<fc_bwdWgt::primitive_desc> bwdWgtPD;
   std::shared_ptr<fc_bwdData::primitive_desc> bwdDataPD;
 
-  resetBwdBuffers(in, wgtGrad_, biasGrad_, out);
+  resetBwdBuffers(inputs[0], wgtGrad_, biasGrad_, out);
 
   resetBwdWgtPD(bwdWgtPD, wgtGrad_, biasGrad_, out);
 
-  resetBwdDataPD(bwdDataPD, in, out);
+  resetBwdDataPD(bwdDataPD, inputs[0], out);
 
-  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgtGrad_, biasGrad_, out);
+  resetBwdPipeline(
+      pipeline, bwdWgtPD, bwdDataPD, inputs[0], wgtGrad_, biasGrad_, out);
 }
 
 void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {

paddle/gserver/layers/MKLDNNFcLayer.h

@@ -59,7 +59,7 @@ public:
                 MKLDNNMatrixPtr& out) override;
 
   void resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                MKLDNNMatrixPtr& in,
+                std::vector<MKLDNNMatrixPtr>& inputs,
                 MKLDNNMatrixPtr& out) override;
 
   void updateWeights(const UpdateCallback& callback) override;

paddle/gserver/layers/MKLDNNLayer.cpp

@@ -91,22 +91,13 @@ void MKLDNNLayer::backward(const UpdateCallback& callback) {
   if (needResetBwd_) {
     VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward";
     pipelineBwd_.clear();
+    inGrads_.resize(inputLayers_.size(), nullptr);
+    extInGrads_.resize(inputLayers_.size(), nullptr);
+    cvtInGrads_.resize(inputLayers_.size(), nullptr);
     pipelineMergeGrad_.clear();
     mergeGrad_ = nullptr;
-    resetBwd(pipelineBwd_, inGrad_, outGrad_);
-    // external output grad is not necessary
-    // since output may be mkldnn internal buffer or merge them directly.
-    CHECK(outGrad_) << "internal output grad is necessary";
-    if (extOutGrad_) {
-      CHECK_EQ(extOutGrad_->getData(), output_.grad->getData())
-          << "the external buffer should share the same data with output_.grad";
-    }
-    if (cvtOutGrad_) {
-      pipelineBwd_.insert(pipelineBwd_.begin(), *cvtOutGrad_);
-    }
-    if (cvtInGrad_) {
-      pipelineBwd_.push_back(*cvtInGrad_);
-    }
+    resetBwd(pipelineBwd_, inGrads_, outGrad_);
+    prepareGradConversions(pipelineBwd_);
     printGradFormat();
     needResetBwd_ = false;
   }
@@ -214,8 +205,8 @@ void MKLDNNLayer::resetOutValue(MKLDNNMatrixPtr& out,
 void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in,
                               memory::primitive_desc intPD,
                               size_t idx) {
-  cvtInGrad_ = nullptr;
-  extInGrad_ = nullptr;
+  cvtInGrads_[idx] = nullptr;
+  extInGrads_[idx] = nullptr;
   in = nullptr;
   LayerPtr& input = inputLayers_[idx];
   if (input->getOutputGrad() == nullptr) {
@@ -237,19 +228,20 @@ void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in,
     return;
   }
 
-  extInGrad_ = in;
-  if (isPaddleFormat(extInGrad_->getFormat())) {
+  extInGrads_[idx] = in;
+  if (isPaddleFormat(extInGrads_[idx]->getFormat())) {
     return;
   }
   // need create reorder
   CHECK(extInVals_[idx] != nullptr &&
         isPaddleFormat(extInVals_[idx]->getFormat()))
       << "should have external input value and the format must be nchw(nc)";
-  extInGrad_ = MKLDNNMatrix::create(extInVals_[idx]->getPrimitiveDesc(), inMat);
+  extInGrads_[idx] =
+      MKLDNNMatrix::create(extInVals_[idx]->getPrimitiveDesc(), inMat);
   CHECK_PRIMITIVE_DESC_EQ(inVals_[idx], intPD);
   in = MKLDNNMatrix::create(intPD);
-  cvtInGrad_ = MKLDNNMatrix::createReorder(in, extInGrad_);
-  CHECK(cvtInGrad_);
+  cvtInGrads_[idx] = MKLDNNMatrix::createReorder(in, extInGrads_[idx]);
+  CHECK(cvtInGrads_[idx]);
 }
 
 void MKLDNNLayer::resetOutGrad(MKLDNNMatrixPtr& out,

paddle/gserver/layers/MKLDNNLayer.h

@@ -69,17 +69,17 @@ protected:
    */
   // below MKLDNNMatrix buffers are all internal buffers
   std::vector<MKLDNNMatrixPtr> inVals_;
-  MKLDNNMatrixPtr inGrad_;
+  std::vector<MKLDNNMatrixPtr> inGrads_;
   MKLDNNMatrixPtr outVal_;
   MKLDNNMatrixPtr outGrad_;
   // below are external value and grad
   std::vector<MKLDNNMatrixPtr> extInVals_;
-  MKLDNNMatrixPtr extInGrad_;
+  std::vector<MKLDNNMatrixPtr> extInGrads_;
   MKLDNNMatrixPtr extOutVal_;
   MKLDNNMatrixPtr extOutGrad_;
   // convert handle between external and internal buffers
   std::vector<std::shared_ptr<mkldnn::reorder>> cvtInVals_;
-  std::shared_ptr<mkldnn::reorder> cvtInGrad_;
+  std::vector<std::shared_ptr<mkldnn::reorder>> cvtInGrads_;
   std::shared_ptr<mkldnn::reorder> cvtOutVal_;
   std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
 
@@ -147,7 +147,7 @@ public:
    * weight and bias buffers should be coverd by child class itself
    */
   virtual void resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                        MKLDNNMatrixPtr& in,
+                        std::vector<MKLDNNMatrixPtr>& inputs,
                         MKLDNNMatrixPtr& out) = 0;
 
   /**
@@ -319,17 +319,19 @@ protected:
    * print the mkldnn memory format of grad
    */
   virtual void printGradFormat() {
-    if (extOutGrad_) {
-      VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat();
-    }
     if (outGrad_) {
-      VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< ";
+      VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< "
+                        << (extOutGrad_ ? extOutGrad_->getFormat()
+                                        : outGrad_->getFormat());
     }
-    if (inGrad_) {
-      VLOG(MKLDNN_FMTS) << inGrad_->getFormat() << " <<<";
+    for (size_t i = 0; i < inGrads_.size(); ++i) {
+      if (!inGrads_[i]) {
+        continue;
       }
-    if (extInGrad_) {
-      VLOG(MKLDNN_FMTS) << extInGrad_->getFormat() << " <<< ";
+      VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName()
+                        << ": " << (extInGrads_[i] ? extInGrads_[i]->getFormat()
+                                                   : inGrads_[i]->getFormat())
+                        << " <<< " << inGrads_[i]->getFormat() << " <<<";
     }
     if (wgtGrad_) {
       VLOG(MKLDNN_FMTS) << "Weight grad format: " << wgtGrad_->getFormat();
@@ -454,6 +456,23 @@ private:
       pipeline.push_back(*cvtOutVal_);
     }
   }
+  void prepareGradConversions(std::vector<mkldnn::primitive>& pipeline) {
+    // external output grad is not necessary
+    // since output may be mkldnn internal buffer or merge them directly.
+    CHECK(outGrad_) << "internal output grad is necessary";
+    if (extOutGrad_) {
+      CHECK_EQ(extOutGrad_->getData(), output_.grad->getData())
+          << "the external buffer should share the same data with output_.grad";
+    }
+    if (cvtOutGrad_) {
+      pipeline.insert(pipeline.begin(), *cvtOutGrad_);
+    }
+    for (size_t i = 0; i < cvtInGrads_.size(); ++i) {
+      if (cvtInGrads_[i]) {
+        pipeline.push_back(*cvtInGrads_[i]);
+      }
+    }
+  }
 };
 
 }  // namespace paddle

paddle/gserver/layers/MKLDNNPoolLayer.cpp

@@ -84,15 +84,15 @@ void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,
 }
 
 void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
-                               MKLDNNMatrixPtr& in,
+                               std::vector<MKLDNNMatrixPtr>& inputs,
                                MKLDNNMatrixPtr& out) {
   std::shared_ptr<pool_bwd::primitive_desc> pd;
 
-  resetBwdBuffers(in, out);
+  resetBwdBuffers(inputs[0], out);
 
-  resetBwdPD(pd, in, out);
+  resetBwdPD(pd, inputs[0], out);
 
-  resetBwdPipeline(pipeline, pd, in, out);
+  resetBwdPipeline(pipeline, pd, inputs[0], out);
 }
 
 void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,

paddle/gserver/layers/MKLDNNPoolLayer.h

@@ -60,7 +60,7 @@ public:
                 MKLDNNMatrixPtr& out) override;
 
   void resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                MKLDNNMatrixPtr& in,
+                std::vector<MKLDNNMatrixPtr>& inputs,
                 MKLDNNMatrixPtr& out) override;
 
   void printSizeInfo() override {