From d4c0734840420314298e4a330ddd4f10d957e8e7 Mon Sep 17 00:00:00 2001
From: tensor-tang <jian.j.tang@intel.com>
Date: Mon, 11 Sep 2017 14:08:13 +0800
Subject: [PATCH] remove convertOutputToOtherDevice

---
 paddle/gserver/layers/MKLDNNFcLayer.cpp | 23 ++++-------------------
 paddle/gserver/layers/MKLDNNFcLayer.h   |  2 --
 paddle/gserver/layers/MKLDNNLayer.h     | 13 +++++++------
 3 files changed, 11 insertions(+), 27 deletions(-)

diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp
index 8318c8c51..f4deb351f 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp
@@ -77,24 +77,6 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
   wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
 }
 
-void MKLDNNFcLayer::convertOutputToOtherDevice() {
-  copyOutputInfoToOtherDevice();
-  // find other cpu device and reorder output to cpu device
-  int cnt = 0;
-  for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
-    if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
-      // fc cpu output value do not need convert
-      // just share point
-      outputOtherDevice_[i].value = output_.value;
-      ++cnt;
-    }
-  }
-
-  if (cnt > 1) {
-    LOG(WARNING) << "should not have more than one CPU devie";
-  }
-}
-
 void MKLDNNFcLayer::reshape() {
   const Argument& input = getInput(0, getPrev(0)->getDeviceId());
   int batchSize = input.getBatchSize();
@@ -155,7 +137,10 @@ void MKLDNNFcLayer::resetFwd() {
   // change original output value to mkldnn output value
   output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
   if (!outputIsOnlyMKLDNN()) {
-    convertOutputToOtherDevice();
+    copyOutputInfoToOtherDevice();
+    // fc cpu output value do not need create convert
+    // just share point
+    getOutput(CPU_DEVICE).value->setData(output_.value->getData());
   }
 
   // create forward handle
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h
index e138a6faf..e2657a8d5 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.h
+++ b/paddle/gserver/layers/MKLDNNFcLayer.h
@@ -72,8 +72,6 @@ protected:
    * only would be called when needed
    */
   void resetBwd();
-
-  void convertOutputToOtherDevice() override;
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h
index ed1ad7c0b..1a3e949fb 100644
--- a/paddle/gserver/layers/MKLDNNLayer.h
+++ b/paddle/gserver/layers/MKLDNNLayer.h
@@ -113,12 +113,6 @@ public:
    */
   virtual void convertWeightsToPaddle() {}
 
-  /**
-   * convert MKLDNN output to other device.
-   * only support CPU device yet
-   */
-  virtual void convertOutputToOtherDevice() {}
-
   /**
    * print info about sizes
    */
@@ -155,6 +149,7 @@ protected:
    *        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());
@@ -163,6 +158,12 @@ protected:
       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";
     }
   }
 
-- 
GitLab