Merge pull request #4953 from tensor-tang/merge_grad_gtest

refine the mkldnn logic

paddle/gserver/activations/MKLDNNActivation.cpp

@@ -126,7 +126,7 @@ void MKLDNNEltwiseActivation::resetFwd(Argument& act) {
   copyInVal_ = nullptr;
   if (act.grad && algo == algorithm::eltwise_tanh) {
     // tanh need save src input for backward
-    inVal_ = MKLDNNMatrix::create(nullptr, val_->getPrimitiveDesc());
+    inVal_ = MKLDNNMatrix::create(val_->getPrimitiveDesc());
     copyInVal_ = std::make_shared<mkldnn::reorder>(*val_, *inVal_);
     CHECK(copyInVal_) << "should not be emptry";
     pipelineFwd_.push_back(*copyInVal_);
@@ -145,7 +145,7 @@ void MKLDNNEltwiseActivation::resetBwd(Argument& act) {
   algorithm algo = getAlgo(this->getName());
   float alpha = getBwdAlpha();
   float beta = getBeta();
-  grad_ = MKLDNNMatrix::create(act.grad, val_->getPrimitiveDesc());
+  grad_ = MKLDNNMatrix::create(val_->getPrimitiveDesc(), act.grad);
   auto eng = CPUEngine::Instance().getEngine();
   auto bwdDesc = eltwise_bwd::desc(
       algo, grad_->getMemoryDesc(), val_->getMemoryDesc(), alpha, beta);
@@ -230,7 +230,7 @@ void MKLDNNActivation::resetFwd(Argument& act) {
     int ic = cnt_ / bs / ih / iw;
     CHECK_EQ(cnt_, (size_t)bs * ic * ih * iw);
     val_ = MKLDNNMatrix::create(
-        act.value, {bs, ic, ih, iw}, mkldnn::memory::format::nchw, *engine_);
+        {bs, ic, ih, iw}, mkldnn::memory::format::nchw, *engine_, act.value);
     CHECK(val_);
     val_->downSpatial();
   }

paddle/gserver/layers/MKLDNNBase.h

@@ -21,8 +21,8 @@ namespace paddle {
 typedef enum {
   MKLDNN_BASE = 1,   // basical info of MKLDNN
   MKLDNN_TESTS = 1,  // gtest info of MKLDNN
-  MKLDNN_SIZES = 2,  // size info of MKLDNN
+  MKLDNN_FMTS = 2,   // format info of MKLDNN
-  MKLDNN_FMTS = 3,   // format info of MKLDNN
+  MKLDNN_SIZES = 3,  // size info of MKLDNN
   MKLDNN_ALL = 4,    // show all info of MKLDNN
 } MKLDNN_LOG_LEVEL;
 

paddle/gserver/layers/MKLDNNConvLayer.h

@@ -48,17 +48,6 @@ protected:
   // save forward primitive_desc, which can be used backward
   std::shared_ptr<conv_fwd::primitive_desc> fwdPD_;
 
-  // MKLDNNMatrixPtr which should be created from CPU Device
-  MKLDNNMatrixPtr cpuInVal_;
-  MKLDNNMatrixPtr cpuInGrad_;
-  MKLDNNMatrixPtr cpuOutVal_;
-  MKLDNNMatrixPtr cpuOutGrad_;
-  // convert handle between CPU device and MKLDNN device
-  std::shared_ptr<mkldnn::reorder> cvtInVal_;
-  std::shared_ptr<mkldnn::reorder> cvtInGrad_;
-  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
-  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
-
   // whether the weight has been init
   bool hasInitedWgt_;
 
@@ -94,8 +83,6 @@ public:
                 MKLDNNMatrixPtr& bias,
                 MKLDNNMatrixPtr& out) override;
 
-  void updateInputData() override;
-
   void updateWeights(const UpdateCallback& callback) override;
 
   void convertWeightsFromPaddle() override;
@@ -109,26 +96,6 @@ public:
                        << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
   }
 
-  void printValueFormatFlow() override {
-    if (cpuInVal_) {
-      VLOG(MKLDNN_FMTS) << cpuInVal_->getFormat() << " >>>";
-    }
-    MKLDNNLayer::printValueFormatFlow();
-    if (cpuOutVal_) {
-      VLOG(MKLDNN_FMTS) << " >>> " << cpuOutVal_->getFormat();
-    }
-  }
-
-  void printGradFormatFlow() override {
-    if (cpuInGrad_) {
-      VLOG(MKLDNN_FMTS) << cpuInGrad_->getFormat() << " <<<";
-    }
-    MKLDNNLayer::printGradFormatFlow();
-    if (cpuOutGrad_) {
-      VLOG(MKLDNN_FMTS) << " <<< " << cpuOutGrad_->getFormat();
-    }
-  }
-
 protected:
   /**
    * load the dims settings of this conv
@@ -162,23 +129,6 @@ protected:
                         MKLDNNMatrixPtr& bias,
                         MKLDNNMatrixPtr& out);
 
-  /**
-   * reset MKLDNNMatrix of input value
-   */
-  void resetInValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
-                    MKLDNNMatrixPtr& in);
-  /**
-   * reset MKLDNNMatrix of weight and bias value
-   */
-  void resetWgtBiasValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
-                         MKLDNNMatrixPtr& wgt,
-                         MKLDNNMatrixPtr& bias);
-  /**
-   * reset MKLDNNMatrix of output value
-   */
-  void resetOutValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
-                     MKLDNNMatrixPtr& out);
-
   /**
    * reset the backward weight primitive descriptor.
    */
@@ -207,22 +157,6 @@ protected:
                         MKLDNNMatrixPtr& bias,
                         MKLDNNMatrixPtr& out);
 
-  /**
-   * reset MKLDNNMatrix of output grad
-   */
-  void resetOutGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
-                    MKLDNNMatrixPtr& out);
-  /**
-   * reset MKLDNNMatrix of weight and bias grad
-   */
-  void resetWgtBiasGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
-                        MKLDNNMatrixPtr& wgt,
-                        MKLDNNMatrixPtr& bias);
-  /**
-   * reset MKLDNNMatrix of input grad
-   */
-  void resetInGrad(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
-                   MKLDNNMatrixPtr& in);
   /**
    * reset MKLDNNMatrix of weight value for backward data
    * since the primitive_desc would be different with wgtVal_

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -62,7 +62,7 @@ void MKLDNNFcLayer::convertWeightsFromPaddle() {
   CHECK(wgtVal_) << "should have been initialized";
   bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
   auto targetDim = wgtVal_->getDims();
-  auto srcFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
+  auto srcFmt = hasNoSpatial_ ? format::io : format::ihwo;
   wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
   hasInitedWgt_ = true;
 }
@@ -71,7 +71,7 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
   CHECK(wgtVal_) << "should have been initialized";
   bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
   auto targetDim = wgtVal_->getDims();
-  auto dstFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
+  auto dstFmt = hasNoSpatial_ ? format::io : format::ihwo;
   wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
 }
 
@@ -100,8 +100,6 @@ void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
   resetFwdPD(fwdPD_, in, wgt, bias, out);
 
   resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
-
-  printValueFormatFlow();
 }
 
 void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
@@ -119,12 +117,6 @@ void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
   resetBwdDataPD(bwdDataPD, in, out);
 
   resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
-
-  printGradFormatFlow();
-}
-
-void MKLDNNFcLayer::updateInputData() {
-  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
 }
 
 void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
@@ -139,51 +131,30 @@ void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
                                     MKLDNNMatrixPtr& bias,
                                     MKLDNNMatrixPtr& out) {
   resetInValue(in);
-
+  CHECK(in);
-  resetWgtBiasValue(wgt, bias);
-
-  resetOutValue(out);
-}
-
-void MKLDNNFcLayer::resetInValue(MKLDNNMatrixPtr& in) {
-  if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& dnnIn = getInputValue(0);
-    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
-    CHECK(in) << "Input should be MKLDNNMatrix";
-  } else {
-    CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
-    in = MKLDNNMatrix::create(
-        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
-  }
   in->downSpatial();
-}
 
-void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
+  auto outPD =
-                                      MKLDNNMatrixPtr& bias) {
+      MKLDNNMatrix::createPrimitiveDesc({bs_, oc_}, format::nc, engine_);
+  resetOutValue(out, outPD);
+
   format wgtFmt = format::oihw;
-  if (inVal_->getFormat() == format::nChw8c) {
+  if (in->getFormat() == format::nChw8c) {
     wgtFmt = format::oIhw8i;
-  } else if (inVal_->getFormat() == format::nChw16c) {
+  } else if (in->getFormat() == format::nChw16c) {
     wgtFmt = format::oIhw16i;
   }
-  wgt = MKLDNNMatrix::create(
+  auto wgtPD =
-      weight_->getW(), {oc_, ic_, ih_, iw_}, wgtFmt, engine_);
+      MKLDNNMatrix::createPrimitiveDesc({oc_, ic_, ih_, iw_}, wgtFmt, engine_);
+  resetWithMatrix(wgt, weight_->getW(), wgtPD);
   wgt->downSpatial();
-  VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
-
-  bias = (biases_ && biases_->getW())
-             ? MKLDNNMatrix::create(biases_->getW(), {oc_}, format::x, engine_)
-             : nullptr;
-}
 
-void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
+  if (biases_ && biases_->getW()) {
-  out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
+    auto biasPD = MKLDNNMatrix::createPrimitiveDesc({oc_}, format::x, engine_);
-  if (!outputIsOnlyMKLDNN()) {
+    resetWithMatrix(bias, biases_->getW(), biasPD);
-    // fc cpu output value do not need create convert, just share data
+  } else {
-    getOutput(CPU_DEVICE).value->setData(out->getData());
+    bias = nullptr;
   }
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
 }
 
 void MKLDNNFcLayer::resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
@@ -219,7 +190,6 @@ void MKLDNNFcLayer::resetFwdPipeline(
   } else {
     fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out));
   }
-
   pipeline.push_back(*fwd_);
 }
 
@@ -227,44 +197,18 @@ void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
                                     MKLDNNMatrixPtr& wgt,
                                     MKLDNNMatrixPtr& bias,
                                     MKLDNNMatrixPtr& out) {
-  resetOutGrad(out);
+  CHECK(inVal_ && outVal_);
-
+  resetOutGrad(out, outVal_->getPrimitiveDesc());
-  resetWgtBiasGrad(wgt, bias);
+  resetInGrad(in, inVal_->getPrimitiveDesc());
-
-  resetInGrad(in);
-}
-
-void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
-  CHECK(outVal_);
-  if (outputIsOnlyMKLDNN()) {
-    MKLDNNLayer::resetOutGrad(out, outVal_->getPrimitiveDesc());
-  } else {
-    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
-    output_.grad->setData(cpuOut->getData());
-    out = MKLDNNMatrix::create(cpuOut, outVal_->getPrimitiveDesc());
-  }
-}
 
-void MKLDNNFcLayer::resetWgtBiasGrad(MKLDNNMatrixPtr& wgt,
-                                     MKLDNNMatrixPtr& bias) {
   CHECK(wgtVal_);
-  wgt = MKLDNNMatrix::create(weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
+  resetWithMatrix(wgt, weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
 
-  bias = nullptr;
+  if (biasVal_) {
-  if (biasVal_ == nullptr) {
+    resetWithMatrix(bias, biases_->getWGrad(), biasVal_->getPrimitiveDesc());
-    return;
+  } else {
-  }
+    bias = nullptr;
-  bias =
-      MKLDNNMatrix::create(biases_->getWGrad(), biasVal_->getPrimitiveDesc());
-}
-
-void MKLDNNFcLayer::resetInGrad(MKLDNNMatrixPtr& in) {
-  in = nullptr;
-  if (inputLayers_[0]->getOutput().grad == nullptr) {
-    return;
   }
-  CHECK(inVal_);
-  MKLDNNLayer::resetInGrad(in, inVal_->getPrimitiveDesc());
 }
 
 void MKLDNNFcLayer::resetBwdWgtPD(

paddle/gserver/layers/MKLDNNFcLayer.h

@@ -66,8 +66,6 @@ public:
                 MKLDNNMatrixPtr& bias,
                 MKLDNNMatrixPtr& out) override;
 
-  void updateInputData() override;
-
   void updateWeights(const UpdateCallback& callback) override;
 
   void convertWeightsFromPaddle() override;
@@ -84,9 +82,6 @@ protected:
                        MKLDNNMatrixPtr& wgt,
                        MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);
-  void resetInValue(MKLDNNMatrixPtr& in);
-  void resetWgtBiasValue(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
-  void resetOutValue(MKLDNNMatrixPtr& out);
   void resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
                   MKLDNNMatrixPtr in,
                   MKLDNNMatrixPtr wgt,
@@ -109,9 +104,6 @@ protected:
                        MKLDNNMatrixPtr& wgt,
                        MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);
-  void resetOutGrad(MKLDNNMatrixPtr& out);
-  void resetWgtBiasGrad(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
-  void resetInGrad(MKLDNNMatrixPtr& in);
   void resetBwdWgtPD(std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
                      MKLDNNMatrixPtr& wgt,
                      MKLDNNMatrixPtr& bias,

paddle/gserver/layers/MKLDNNPoolLayer.cpp

@@ -85,8 +85,6 @@ void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,
   resetFwdPD(fwdPD_, in, out);
 
   resetFwdPipeline(pipeline, fwdPD_, in, out);
-
-  printValueFormatFlow();
 }
 
 void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
@@ -101,65 +99,22 @@ void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
   resetBwdPD(pd, in, out);
 
   resetBwdPipeline(pipeline, pd, in, out);
-
-  printGradFormatFlow();
-}
-
-void MKLDNNPoolLayer::updateInputData() {
-  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
 }
 
 void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
                                       MKLDNNMatrixPtr& out) {
   resetInValue(in);
 
-  resetOutValue(out);
-}
-
-void MKLDNNPoolLayer::resetInValue(MKLDNNMatrixPtr& in) {
-  if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& dnnIn = getInputValue(0);
-    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
-    CHECK(in) << "Input should be MKLDNNMatrix";
-  } else {
-    CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
-    in = MKLDNNMatrix::create(
-        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
-  }
-}
-
-void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
-  CHECK(inVal_) << "Should reset input value first";
   memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
-  out = MKLDNNMatrix::create(
+  CHECK(in);
-      output_.value, outDims, inVal_->getFormat(), engine_);
+  auto outPD =
-
+      MKLDNNMatrix::createPrimitiveDesc(outDims, in->getFormat(), engine_);
-  // create reorder if output value has cpu device and pd do not match
+  resetOutValue(out, outPD);
-  cpuOutVal_ = nullptr;
-  cvtOutVal_ = nullptr;
-  if (!outputIsOnlyMKLDNN()) {
-    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
-    cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
-    if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
-      out = MKLDNNMatrix::create(nullptr, out->getPrimitiveDesc());
-      cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
-      CHECK(cvtOutVal_) << "should not be emptry";
-    } else {
-      cpuOut->setData(output_.value->getData());
-      cpuOutVal_ = out;
-    }
-    output_.value = std::dynamic_pointer_cast<Matrix>(cpuOutVal_);
-    return;
-  }
-  output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
 }
 
 void MKLDNNPoolLayer::resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
                                  MKLDNNMatrixPtr in,
                                  MKLDNNMatrixPtr out) {
-  memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
-  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
   memory::dims kernels = memory::dims{fh_, fw_};
   memory::dims strides = memory::dims{sh_, sw_};
   memory::dims padL = memory::dims{ph_, pw_};
@@ -194,58 +149,26 @@ void MKLDNNPoolLayer::resetFwdPipeline(
              ? std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out, *workspace_))
              : std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out));
   pipeline.push_back(*fwd_);
-
-  if (cvtOutVal_) {
-    pipeline.push_back(*cvtOutVal_);
-  }
 }
 
 void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
                                       MKLDNNMatrixPtr& out) {
-  resetOutGrad(out);
+  CHECK(inVal_ && outVal_);
-
+  resetOutGrad(out, outVal_->getPrimitiveDesc());
-  resetInGrad(in);
+  resetInGrad(in, inVal_->getPrimitiveDesc());
-}
-void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
-  cpuOutGrad_ = nullptr;
-  cvtOutGrad_ = nullptr;
-  CHECK(outVal_);
-  if (outputIsOnlyMKLDNN()) {
-    MKLDNNLayer::resetOutGrad(out, outVal_->getPrimitiveDesc());
-  } else {
-    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
-    // always share the same grad data of CPU output
-    // then the activation can get the right grad from output_.grad
-    output_.grad->setData(cpuOut->getData());
-    cpuOutGrad_ = MKLDNNMatrix::create(
-        cpuOut, memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_);
-    if (cpuOutGrad_->getPrimitiveDesc() != outVal_->getPrimitiveDesc()) {
-      out = MKLDNNMatrix::create(nullptr, outVal_->getPrimitiveDesc());
-      cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
-      CHECK(cvtOutGrad_) << "should not be emptry";
-    } else {
-      out = cpuOutGrad_;
-    }
-  }
-}
-
-void MKLDNNPoolLayer::resetInGrad(MKLDNNMatrixPtr& in) {
-  in = nullptr;
-  if (inputLayers_[0]->getOutput().grad == nullptr) {
-    return;
-  }
-  CHECK(inVal_);
-  MKLDNNLayer::resetInGrad(in, inVal_->getPrimitiveDesc());
 }
 
 void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
                                  MKLDNNMatrixPtr& in,
                                  MKLDNNMatrixPtr& out) {
+  pd = nullptr;
+  if (in == nullptr) {
+    return;
+  }
   memory::dims kernels = memory::dims{fh_, fw_};
   memory::dims strides = memory::dims{sh_, sw_};
   memory::dims padL = memory::dims{ph_, pw_};
   memory::dims padR = getPaddingR();
-  CHECK(in);
   CHECK(out);
   auto bwdDesc = pool_bwd::desc(poolAlgo_,
                                 in->getMemoryDesc(),
@@ -263,8 +186,8 @@ void MKLDNNPoolLayer::resetBwdPipeline(
     std::shared_ptr<pool_bwd::primitive_desc>& pd,
     MKLDNNMatrixPtr& in,
     MKLDNNMatrixPtr& out) {
-  if (cvtOutGrad_) {
+  if (pd == nullptr) {
-    pipeline.push_back(*cvtOutGrad_);
+    return;
   }
 
   bwdData_ =

paddle/gserver/layers/MKLDNNPoolLayer.h

@@ -38,13 +38,6 @@ protected:
   // pooling_avg or pooling_max
   mkldnn::algorithm poolAlgo_;
 
-  // MKLDNNMatrixPtr which should be created from CPU Device
-  MKLDNNMatrixPtr cpuOutVal_;
-  MKLDNNMatrixPtr cpuOutGrad_;
-  // convert handle between CPU device and MKLDNN device
-  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
-  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
-
   // save forward primitive_desc, which can be used backward
   std::shared_ptr<pool_fwd::primitive_desc> fwdPD_;
   // according to https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
@@ -74,8 +67,6 @@ public:
                 MKLDNNMatrixPtr& bias,
                 MKLDNNMatrixPtr& out) override;
 
-  void updateInputData() override;
-
   void printSizeInfo() override {
     MKLDNNLayer::printSizeInfo();
     VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
@@ -90,8 +81,6 @@ protected:
    *                    reset pipeline.
    */
   void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
-  void resetInValue(MKLDNNMatrixPtr& in);
-  void resetOutValue(MKLDNNMatrixPtr& out);
   void resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
                   MKLDNNMatrixPtr in,
                   MKLDNNMatrixPtr out);
@@ -106,8 +95,6 @@ protected:
    *                     reset pipeline.
    */
   void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
-  void resetOutGrad(MKLDNNMatrixPtr& out);
-  void resetInGrad(MKLDNNMatrixPtr& in);
   void resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
                   MKLDNNMatrixPtr& in,
                   MKLDNNMatrixPtr& out);

paddle/gserver/tests/MKLDNNTester.cpp

@@ -97,7 +97,7 @@ void MKLDNNTester::randomWgtDatas() {
     parameters_[REF][i]->randomize();
     dnnValue->copyFrom(*refValue);
 
-    VLOG(lvl_) << "Random weight data " << parameters_[DNN][i]->getName();
+    VLOG(MKLDNN_TESTS) << "Random weight " << parameters_[DNN][i]->getName();
     printVector(dnnValue);
   }
 }
@@ -109,7 +109,7 @@ void MKLDNNTester::randomBotDatas() {
     dataLayers_[REF][i]->getOutputValue()->randomizeUniform();
     dataLayers_[DNN][i]->getOutputValue()->copyFrom(
         *(dataLayers_[REF][i]->getOutputValue()));
-    VLOG(lvl_) << "Input " << i << " data:";
+    VLOG(MKLDNN_TESTS) << "Random Foward, InputValue " << i;
     printMatrix(dataLayers_[REF][i]->getOutputValue());
   }
 }
@@ -118,12 +118,12 @@ void MKLDNNTester::randomTopDiffs() {
   refLayer_->getOutputGrad()->randomizeUniform();
   dnnLayer_->getOutput(CPU_DEVICE)
       .grad->copyFrom(*(refLayer_->getOutputGrad()));
-  VLOG(lvl_) << "Random Backward Input, TopDiff: ";
+  VLOG(MKLDNN_TESTS) << "Random Backward, OutputGrad";
   printMatrix(refLayer_->getOutputGrad());
 }
 
 void MKLDNNTester::checkForward() {
-  VLOG(MKLDNN_ALL) << "Check Forward";
+  VLOG(MKLDNN_TESTS) << "Check Forward";
   printTopDatas();
   double delta =
       compareMatrix(dnnLayer_->getOutputValue(), refLayer_->getOutputValue());
@@ -131,15 +131,15 @@ void MKLDNNTester::checkForward() {
 }
 
 void MKLDNNTester::checkBackwardData() {
-  VLOG(MKLDNN_ALL) << "Check Backward Data";
+  VLOG(MKLDNN_TESTS) << "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) {
     const MatrixPtr& dnnDiff = dataLayers_[DNN][i]->getOutputGrad();
     const MatrixPtr& refDiff = dataLayers_[REF][i]->getOutputGrad();
-    VLOG(lvl_) << "Mkldnn Backward Output BotDiff " << i;
+    VLOG(MKLDNN_ALL) << "MKLDNN Backward Result: InputGrad " << i;
     printMatrix(dnnDiff);
-    VLOG(lvl_) << "Reference Backward Output BotDiff " << i;
+    VLOG(MKLDNN_ALL) << "Reference Backward Result: InputGrad " << i;
     printMatrix(refDiff);
 
     double delta = compareMatrix(dnnDiff, refDiff);
@@ -153,7 +153,7 @@ void MKLDNNTester::checkBackwardData() {
 }
 
 void MKLDNNTester::checkBackwardWgts() {
-  VLOG(MKLDNN_ALL) << "Check Backward Weight";
+  VLOG(MKLDNN_TESTS) << "Check Backward Weight";
   CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
   vector<VectorPtr> dnnWgts;  // used to temply save mkldnn weights
   saveWgt(parameters_[DNN], dnnWgts);
@@ -165,9 +165,11 @@ void MKLDNNTester::checkBackwardWgts() {
   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);
-    VLOG(lvl_) << "Mkldnn Output weight " << parameters_[DNN][i]->getName();
+    VLOG(MKLDNN_ALL) << "MKLDNN Result: weight value"
+                     << parameters_[DNN][i]->getName();
     printVector(dnn);
-    VLOG(lvl_) << "Reference Output weight " << parameters_[REF][i]->getName();
+    VLOG(MKLDNN_ALL) << "Reference Result: weight value "
+                     << parameters_[REF][i]->getName();
     printVector(ref);
 
     double delta = compareVector(dnn, ref);
@@ -240,7 +242,8 @@ void MKLDNNTester::printTopDatas() {
   }
 
   for (int n = 0; n < NUM; ++n) {
-    VLOG(lvl_) << testLayers_[n]->getType() << " forward output TopData: ";
+    VLOG(MKLDNN_ALL) << testLayers_[n]->getType()
+                     << " Forward Result: OutputValue";
     printMatrix(testLayers_[n]->getOutputValue());
   }
 }
@@ -252,7 +255,7 @@ void MKLDNNTester::printMatrix(const MatrixPtr& m) {
 
   std::ostringstream ostr;
   m->print(ostr);
-  VLOG(lvl_) << std::endl << ostr.str();
+  VLOG(MKLDNN_ALL) << std::endl << ostr.str();
 }
 
 void MKLDNNTester::printVector(const VectorPtr& v) {
@@ -262,7 +265,7 @@ void MKLDNNTester::printVector(const VectorPtr& v) {
 
   std::ostringstream ostr;
   v->print(ostr, v->getSize());
-  VLOG(lvl_) << std::endl << ostr.str();
+  VLOG(MKLDNN_ALL) << std::endl << ostr.str();
 }
 
 double MKLDNNTester::getDelta(const real* d1,
@@ -314,7 +317,7 @@ void MKLDNNTester::runOnce() {
   UpdateCallback updateCallback = [](Parameter* para) {
     auto& grad = para->getBuf(PARAMETER_GRADIENT);
     auto& value = para->getBuf(PARAMETER_VALUE);
-    real lr = 1e-3;
+    real lr = 1e-2;
     value->add(*grad, lr);
     grad->zeroMem();
   };
@@ -340,10 +343,9 @@ void MKLDNNTester::run(const TestConfig& dnn,
                        size_t batchSize,
                        size_t inputImgH,
                        size_t inputImgW,
+                       bool printDetails,
                        size_t iter,
-                       float epsilon,
+                       float epsilon) {
-                       bool log,
-                       int level) {
   CHECK(dnn.layerConfig.type().compare(0, 7, "mkldnn_") == 0 ||
         dnn.layerConfig.active_type().compare(0, 7, "mkldnn_") == 0)
       << "should be MKLDNN layer or MKLDNN activation";
@@ -359,10 +361,9 @@ void MKLDNNTester::run(const TestConfig& dnn,
 
   ih_ = inputImgH;
   iw_ = inputImgW;
+  log_ = printDetails;
   iter_ = iter;
   eps_ = epsilon;
-  log_ = log;
-  lvl_ = level;
 
   // Firstly test mkldnn init from PARAM_FORMAT_ORIGINAL weight
   reset(dnn, ref, batchSize);
@@ -531,9 +532,11 @@ void MKLDNNTester::getOutResult(const std::string& configPath,
 void MKLDNNTester::compareResult(DataOut& ref, DataOut& dnn, float eps) {
   CHECK_EQ(ref.outValues.size(), dnn.outValues.size());
   CHECK_EQ(ref.paraValues.size(), dnn.paraValues.size());
+  VLOG(MKLDNN_TESTS) << "compare value size: " << ref.outValues.size();
   for (size_t i = 0; i < ref.outValues.size(); i++) {
     EXPECT_LE(fabs(compareMatrix(ref.outValues[i], dnn.outValues[i])), eps);
   }
+  VLOG(MKLDNN_TESTS) << "compare param size: " << ref.outValues.size();
   for (size_t i = 0; i < ref.paraValues.size(); i++) {
     EXPECT_LE(fabs(compareVector(ref.paraValues[i], dnn.paraValues[i])), eps);
   }
@@ -544,9 +547,10 @@ void MKLDNNTester::runBranchesTest(const std::string& configPath,
                                    float eps) {
   DataIn in;
   initArgument(in, configPath, iter);
-
   DataOut outCpu, outDnn;
+  VLOG(MKLDNN_TESTS) << "runing cpu network";
   getOutResult(configPath, in, outCpu, false, iter);
+  VLOG(MKLDNN_TESTS) << "runing mkldnn network";
   getOutResult(configPath, in, outDnn, true, iter);
 
   compareResult(outCpu, outDnn, eps);

paddle/gserver/tests/MKLDNNTester.h

@@ -58,8 +58,6 @@ protected:
   size_t iter_;
   /// whether to print out the details
   bool log_;
-  /// vlog level to print the matrix details datas
-  int lvl_;
   /// epsilon
   float eps_;
   /// input image size, default 1
@@ -70,7 +68,6 @@ public:
     iter_ = iter;
     eps_ = epsilon;
     log_ = false;
-    lvl_ = MKLDNN_ALL;
   }
 
   ~MKLDNNTester() {}
@@ -81,10 +78,9 @@ public:
            size_t batchSize,
            size_t inputImgH = 1,
            size_t inputImgW = 1,
+           bool printDetails = false,
            size_t iter = 3,
-           float epsilon = 1e-4,
+           float epsilon = 1e-4);
-           bool log = false,
-           int level = MKLDNN_ALL);
   static void runBranchesTest(const std::string& configPath,
                               size_t iter = 3,
                               float eps = 1e-4);

paddle/math/MKLDNNMatrix.cpp

@@ -18,7 +18,7 @@ using namespace mkldnn;  // NOLINT
 
 namespace paddle {
 
-MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
+MKLDNNMatrixPtr MKLDNNMatrix::create(memory::primitive_desc pd, MatrixPtr m) {
   memory::desc md = pd.desc();
   size_t ndims = md.data.ndims;
   int* dims = md.data.dims;
@@ -41,12 +41,12 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
   return std::make_shared<MKLDNNMatrix>(cpuMatrix, pd);
 }
 
-MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
+MKLDNNMatrixPtr MKLDNNMatrix::create(memory::dims dims,
-                                     memory::dims dims,
                                      memory::format fmt,
                                      engine& eg,
+                                     MatrixPtr m,
                                      mkldnn::memory::data_type dtype) {
-  return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg));
+  return create(createPrimitiveDesc(dims, fmt, eg, dtype), m);
 }
 
 std::shared_ptr<reorder> MKLDNNMatrix::createReorder(const MKLDNNMatrixPtr& src,

paddle/math/MKLDNNMatrix.h

@@ -40,24 +40,37 @@ public:
   /**
    * Create MKLDNNMatrix from a MatrixPtr and memory primitive_desc
    */
-  static MKLDNNMatrixPtr create(MatrixPtr m, mkldnn::memory::primitive_desc pd);
+  static MKLDNNMatrixPtr create(mkldnn::memory::primitive_desc pd,
+                                MatrixPtr m = nullptr);
 
   /**
    * Create MKLDNNMatrix from a MatrixPtr and memory details info
    */
   static MKLDNNMatrixPtr create(
-      MatrixPtr m,
       mkldnn::memory::dims dims,
       mkldnn::memory::format fmt,
       mkldnn::engine& eg,
+      MatrixPtr m = nullptr,
       mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32);
 
+  /**
+   * Create primitive descriptor.
+   * default with f32 dtype
+   */
+  static mkldnn::memory::primitive_desc createPrimitiveDesc(
+      const mkldnn::memory::dims dims,
+      const mkldnn::memory::format& fmt,
+      const mkldnn::engine& eg,
+      const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) {
+    return mkldnn::memory::primitive_desc(memory::desc(dims, dtype, fmt), eg);
+  }
+
   /**
    * Create Memory descriptor.
    * default with any format and f32 dtype
    */
   static mkldnn::memory::desc createMemoryDesc(
-      const mkldnn::memory::dims& dims,
+      const mkldnn::memory::dims dims,
       const mkldnn::memory::format& fmt = mkldnn::memory::format::any,
       const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) {
     return mkldnn::memory::desc(dims, dtype, fmt);

paddle/trainer/tests/sample_trainer_config_branch_net.conf

@@ -17,7 +17,7 @@ from paddle.trainer_config_helpers import *
 ################################### Data Configuration ###################################
 TrainData(ProtoData(files = "trainer/tests/mnist.list"))
 ################################### Algorithm Configuration ###################################
-settings(batch_size = 256,
+settings(batch_size = 128,
          learning_method = MomentumOptimizer(momentum=0.5, sparse=False))
 ################################### Network Configuration ###################################
 data = data_layer(name ="input", size=784)
@@ -44,10 +44,11 @@ a2 = img_conv_layer(input=tmp,
             shared_biases=True,
             act=ReluActivation())
 
-tmp = concat_layer(input=[a1, a2])
+tmp = addto_layer(input=[a1, a2],
+            act=ReluActivation(),
+            bias_attr=False)
 
 tmp = img_pool_layer(input=tmp,
-            num_channels=64,
             pool_size=3,
             stride=2,
             padding=1,
@@ -55,35 +56,34 @@ tmp = img_pool_layer(input=tmp,
 
 b1 = img_conv_layer(input=tmp,
             filter_size=3,
-            num_filters=64,
+            num_filters=32,
             padding=1,
             shared_biases=True,
             act=ReluActivation())
 
 b1 = img_pool_layer(input=b1,
             pool_size=3,
-            stride=1,
+            stride=2,
-            padding=1,
+            padding=0,
             pool_type=MaxPooling())
 
 b2 = img_conv_layer(input=tmp,
-            filter_size=5,
+            filter_size=3,
             num_filters=64,
-            padding=2,
+            padding=1,
             shared_biases=True,
             act=ReluActivation())
 
 b2 = img_pool_layer(input=b2,
             pool_size=5,
-            stride=1,
+            stride=2,
-            padding=2,
+            padding=1,
             pool_type=MaxPooling())
 
-tmp = addto_layer(input=[b1, b2],
+tmp = concat_layer(input=[b1, b2])
-            act=ReluActivation(),
-            bias_attr=False)
 
 tmp = img_pool_layer(input=tmp,
+            num_channels=96,
             pool_size=3,
             stride=2,
             padding=1,

paddle/trainer/tests/sample_trainer_config_simple_net.conf

@@ -17,7 +17,7 @@ from paddle.trainer_config_helpers import *
 ################################### Data Configuration ###################################
 TrainData(ProtoData(files = "trainer/tests/mnist.list"))
 ################################### Algorithm Configuration ###################################
-settings(batch_size = 1000,
+settings(batch_size = 128,
          learning_method = MomentumOptimizer(momentum=0.5, sparse=False))
 ################################### Network Configuration ###################################
 data = data_layer(name ="input", size=784)