add mkldnn fc forward

paddle/gserver/layers/MkldnnFcLayer.cpp

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "MkldnnFcLayer.h"
+#include "paddle/utils/Stat.h"
 
 namespace paddle {
 
@@ -20,11 +21,82 @@ REGISTER_LAYER(mkldnn_fc, MkldnnFcLayer);
 
 bool MkldnnFcLayer::init(const LayerMap& layerMap,
                          const ParameterMap& parameterMap) {
-  return MkldnnLayer::init(layerMap, parameterMap);
+  if (!MkldnnLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+
+  CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet!";
+  CHECK_EQ(inputLayers_.size(), parameters_.size());
+  CHECK(!parameters_[0]->isSparse()) << "Do not support sparse yet";
+
+  // output size, cat not be changed
+  oc_ = getSize();
+  oh_ = 1;
+  ow_ = 1;
+
+  // input size can not change in FC
+  iLayerSize_ = inputLayers_[0]->getSize();
+  CHECK_EQ(parameters_[0]->getSize(), iLayerSize_ * oc_);
+
+  // create weight
+  weight_ =
+      std::unique_ptr<Weight>(new Weight(oc_, iLayerSize_, parameters_[0], 0));
+
+  // create biases
+  if (biasParameter_.get() != NULL) {
+    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_));
+  }
+  return true;
+}
+
+void MkldnnFcLayer::reshape() {
+  const Argument& input = getInput(0);
+  int batchSize = input.getBatchSize();
+  if (bs_ == batchSize) {
+    return;
+  }
+  bs_ = batchSize;
+  ih_ = input.getFrameHeight();
+  iw_ = input.getFrameWidth();
+  if (ih_ == 0) {
+    ih_ = 1;
+  }
+  if (iw_ == 0) {
+    iw_ = 1;
+  }
+  CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
+  ic_ = iLayerSize_ / (ih_ * iw_);
+  CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible";
+  CHECK_EQ(size_t(oc_), getSize());
+
+  // reset output
+  output_.setFrameHeight(oh_);
+  output_.setFrameWidth(ow_);
+  resetOutput(bs_, oc_);
 }
 
-void MkldnnFcLayer::forward(PassType passType) {}
+void MkldnnFcLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  reshape();
 
-void MkldnnFcLayer::backward(const UpdateCallback& callback) {}
+  {
+    REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
+    real* input = getInputValue(0)->getData();
+    real* output = getOutputValue()->getData();
+    real* wgt = weight_->getW()->getData();
+    bool hasBias = biases_ && biases_->getW();
+    real* bias = hasBias ? biases_->getW()->getData() : NULL;
+    mkldnnForwardFC(bs_, ic_, ih_, iw_, input, oc_, output, wgt, bias);
+  }
 
+  /* activation */ {
+    REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
+    forwardActivation();
+  }
+}
+
+void MkldnnFcLayer::backward(const UpdateCallback& callback) {
+  ;  // bool hasBias = biases_ && biases_->getWGrad();
+}
 }  // namespace paddle

paddle/gserver/layers/MkldnnFcLayer.h

@@ -26,6 +26,13 @@ namespace paddle {
  */
 class MkldnnFcLayer : public MkldnnLayer {
 protected:
+  // input layer size, can not be change after init
+  size_t iLayerSize_;  // == ic * ih * iw
+
+  // fc weight and bias
+  std::unique_ptr<Weight> weight_;
+  std::unique_ptr<Weight> biases_;
+
 public:
   explicit MkldnnFcLayer(const LayerConfig& config) : MkldnnLayer(config) {}
 
@@ -34,6 +41,8 @@ public:
   bool init(const LayerMap& layerMap,
             const ParameterMap& parameterMap) override;
 
+  void reshape();
+
   void forward(PassType passType) override;
 
   void backward(const UpdateCallback& callback) override;

paddle/gserver/layers/MkldnnLayer.cpp

@@ -0,0 +1,98 @@
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "MkldnnLayer.h"
+
+// using namespace mkldnn;  // NOLINT
+using mem = mkldnn::memory;  // NOLINT
+typedef mem::format format;
+typedef mkldnn::inner_product_forward fc_fwd;
+typedef mkldnn::inner_product_backward_weights fc_bwdWgt;
+typedef mkldnn::inner_product_backward_data fc_bwdData;
+
+namespace paddle {
+
+bool MkldnnLayer::init(const LayerMap& layerMap,
+                       const ParameterMap& parameterMap) {
+  CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
+                          << "Please set WITH_MKLDNN=ON";
+  // TODO(TJ): deivecId
+  return Layer::init(layerMap, parameterMap);
+}
+
+void MkldnnLayer::resetForwardFC(int bs,
+                                 int ic,
+                                 int ih,
+                                 int iw,
+                                 real* botData,
+                                 int oc,
+                                 real* topData,
+                                 real* wgtData,
+                                 real* biasData) {
+  bool hasSpatial = ih == 1 && iw == 1 ? false : true;
+  engine_ = CpuEngine::Instance().getEngine();
+
+  mem::desc botMD = hasSpatial ? createMD({bs, ic, ih, iw}, format::nchw)
+                               : createMD({bs, ic}, format::nc);
+  mem::desc wgtMD = hasSpatial ? createMD({oc, ic, ih, iw}, format::oihw)
+                               : createMD({oc, ic}, format::oi);
+  mem::desc biasMD = biasData != NULL ? createMD({oc}, format::x)
+                                      : createMD({}, format::format_undef);
+  mem::desc topMD = createMD({bs, oc}, format::nc);
+
+  mkldnn::prop_kind pk = mkldnn::prop_kind::forward;
+  fc_fwd::desc fwdDesc = biasData != NULL
+                             ? fc_fwd::desc(pk, botMD, wgtMD, biasMD, topMD)
+                             : fc_fwd::desc(pk, botMD, wgtMD, topMD);
+  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
+
+  mem bot = mem(mem::primitive_desc(botMD, engine_), botData);
+  mem wgt = mem(mem::primitive_desc(wgtMD, engine_), wgtData);
+  mem top = mem(mem::primitive_desc(topMD, engine_), topData);
+
+  if (biasData != NULL) {
+    mem bias = mem(mem::primitive_desc(biasMD, engine_), biasData);
+    fwd_.reset(new fc_fwd(fwdPD, bot, wgt, bias, top));
+  } else {
+    fwd_.reset(new fc_fwd(fwdPD, bot, wgt, top));
+  }
+  pipelineFwd_.clear();
+  pipelineFwd_.push_back(*fwd_);
+}
+
+void MkldnnLayer::mkldnnForwardFC(int bs,
+                                  int ic,
+                                  int ih,
+                                  int iw,
+                                  real* botData,
+                                  int oc,
+                                  real* topData,
+                                  real* wgtData,
+                                  real* biasData) {
+  // if input size changed, reset it
+  resetForwardFC(bs, ic, ih, iw, botData, oc, topData, wgtData, biasData);
+
+  // just forward
+  // update botdata
+  stream_->submit(pipelineFwd_);
+}
+
+mem::desc MkldnnLayer::createMD(mem::dims dims,
+                                mem::format fmt,
+                                mem::data_type type) {
+  // TODO(TJ): isFmtSuppoted(fmt)
+  return mem::desc(dims, type, fmt);
+}
+
+}  // namespace paddle

paddle/gserver/layers/MkldnnLayer.h

@@ -29,20 +29,65 @@ typedef std::shared_ptr<MkldnnLayer> MkldnnLayerPtr;
  *
  */
 class MkldnnLayer : public Layer {
+protected:
+  // batch size
+  int bs_;
+  // input image channel, height and width
+  int ic_, ih_, iw_;
+  // output image channel, height and width
+  int oc_, oh_, ow_;
+
+  // mkldnn engine, stream and primivtives
+  mkldnn::engine engine_;
+  std::shared_ptr<MkldnnStream> stream_;
+
+  std::shared_ptr<mkldnn::primitive> fwd_;
+  std::vector<mkldnn::primitive> pipelineFwd_;
+  std::vector<mkldnn::primitive> pipelineBwd_;
+
 public:
-  explicit MkldnnLayer(const LayerConfig& config) : Layer(config) {}
+  explicit MkldnnLayer(const LayerConfig& config)
+      : Layer(config),
+        bs_(0),
+        ic_(0),
+        ih_(0),
+        iw_(0),
+        oc_(0),
+        oh_(0),
+        ow_(0),
+        engine_(mkldnn::engine::cpu, 0),
+        stream_(nullptr) {}
 
   ~MkldnnLayer() {}
 
-  virtual bool init(const LayerMap& layerMap,
-                    const ParameterMap& parameterMap) {
-    CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
-                            << "Please set WITH_MKLDNN=ON";
-    // TODO(TJ): deivecId
-    return Layer::init(layerMap, parameterMap);
-  }
+  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void resetForwardFC(int bs,
+                      int ic,
+                      int ih,
+                      int iw,
+                      real* botData,
+                      int oc,
+                      real* topData,
+                      real* wgtData,
+                      real* biasData);
+
+  void mkldnnForwardFC(int bs,
+                       int ic,
+                       int ih,
+                       int iw,
+                       real* botData,
+                       int oc,
+                       real* topData,
+                       real* wgtData,
+                       real* biasData);
 
-  void resetOutput(size_t height, size_t width) { ; }
+  // TODO(TJ): move to MkldnnMatrix
+  // create memory desc
+  inline mkldnn::memory::desc createMD(
+      mkldnn::memory::dims dims,
+      mkldnn::memory::format fmt,
+      mkldnn::memory::data_type type = mkldnn::memory::data_type::f32);
 };
 
 }  // namespace paddle