Merge remote-tracking branch 'origin/develop' into multigpu

doc/design/block.md

@@ -243,7 +243,7 @@ class SymbolTable {
   // TODO determine whether name is generated by python or C++.
   // Currently assume that a unique name will be generated by C++ if the
   // argument name is left default.
-  VarDesc* NewVar(const string& name="");
+  VarDesc* Var(const string& name="");
 
   // find a VarDesc by name, if recursive is true, find parent's SymbolTable
   // recursively.

doc/design/scope.md

@@ -37,7 +37,7 @@ Scope is an association of a name to variable. All variables belong to `Scope`.
 ```cpp
 class Scope {
  public:
-  Variable* NewVar(const std::string& name);
+  Variable* Var(const std::string& name);
   const Variable* FindVar(const std::string& name) const;
 
  private:
@@ -98,7 +98,7 @@ class Scope {
   Variable* FindVar(const std::string& name) const;
 
   // return if already contains same name variable.
-  Variable* NewVar(const std::string& name);
+  Variable* Var(const std::string& name);
 
  private:
   std::shared_ptr<Scope> parent_;
@@ -107,7 +107,7 @@ class Scope {
 ```
 ## Only scope can create a variable
 
-To ensure `only scope can create a variable`, we should mark `Variable`'s constructor as a private member function, and Scope is a friend class of Variable. And then only `NewVar` can construct `Variable`.
+To ensure `only scope can create a variable`, we should mark `Variable`'s constructor as a private member function, and Scope is a friend class of Variable. And then only `Var` can construct `Variable`.
 
 ## When scope destroyed, all variables inside this scope should be destroyed together
 
@@ -121,4 +121,4 @@ Also, as the parent scope is a `shared_ptr`, we can only `Create()` a scope shar
 
 ## Orthogonal interface
 
-`FindVar` will return `nullptr` when `name` is not found. It can be used as `Contains` method. `NewVar` will return an `Error` when there is a name conflict locally. Combine `FindVar` and `NewVar`, we can implement `NewVar` easily.
+`FindVar` will return `nullptr` when `name` is not found. It can be used as `Contains` method. `Var` will return an `Error` when there is a name conflict locally. Combine `FindVar` and `Var`, we can implement `Var` easily.

doc/design/tensor_array.md

@@ -161,7 +161,7 @@ class TensorArray:
         @name: str
             the name of the variable to output.
         '''
-        tensor = NewVar(name)
+        tensor = Var(name)
         tensor_array_stack(self.name, tensor)
         return tensor
 

paddle/framework/backward.cc

@@ -273,18 +273,41 @@ static bool AllGradInSet(const std::vector<std::string>& names,
   return true;
 }
 
-static void CreateGradVarInBlock(BlockDescBind* block_desc,
+static void CreateGradVarInBlock(
-                                 size_t grad_op_start_index) {
+    size_t grad_op_start_index,
+    const std::unordered_map<std::string, std::string>& param_name_map,
+    BlockDescBind* block_desc,
+    std::unordered_map<std::string, GradVarInfo>* grad_var_record) {
   auto ops = block_desc->AllOps();
   for (size_t op_index = grad_op_start_index; op_index < ops.size();
        ++op_index) {
-    for (const auto& output : ops[op_index]->Outputs()) {
+    // <<<<<<< HEAD
-      for (const auto& real_output : output.second) {
+    //     for (const auto& output : ops[op_index]->Outputs()) {
-        if (!block_desc->HasVar(real_output)) {
+    //       for (const auto& real_output : output.second) {
-          block_desc->NewVar(real_output);
+    //         if (!block_desc->HasVar(real_output)) {
-        }
+    //           block_desc->Var(real_output);
+    //         }
+    //       }
+    //     }
+    // =======
+    ForEachVarName(ops[op_index]->Outputs(),
+                   [&](const std::string& grad_var_name) {
+                     if (block_desc->HasVar(grad_var_name)) {
+                       return false;
                      }
+                     block_desc->Var(grad_var_name);
+                     auto it = param_name_map.find(grad_var_name);
+                     if (it == param_name_map.end()) {
+                       return false;
                      }
+                     auto param_var_name = it->second;
+                     auto& grad_record = (*grad_var_record)[param_var_name];
+                     grad_record.name_ = grad_var_name;
+                     grad_record.block_idx_ = block_desc->ID();
+                     grad_record.op_idx_ = static_cast<int>(op_index);
+                     return false; /* not break */
+                   });
+    // >>>>>>> origin/develop
   }
 }
 
@@ -400,7 +423,8 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
   return backward_descs;
 }
 
-void AppendBackward(ProgramDescBind& program_desc, const VarDescBind& target,
+ParamGradInfoMap AppendBackward(
+    ProgramDescBind& program_desc, const VarDescBind& target,
     const std::unordered_set<std::string>& no_grad_vars) {
   std::unordered_set<std::string> no_grad_var_names;
   no_grad_var_names.reserve(no_grad_vars.size() + 1);
@@ -423,20 +447,28 @@ void AppendBackward(ProgramDescBind& program_desc, const VarDescBind& target,
   all_ops.push_back(std::move(fill_one_op));
   size_t forward_op_num = all_ops.size();
   size_t forward_block_num = program_desc.Size();
+
+  // Insert backward operators
   std::unordered_map<std::string, std::string> grad_to_var;
   auto backward_op_descs = MakeBlockBackward(program_desc, root_block_idx,
                                              &no_grad_var_names, &grad_to_var);
+
+  std::unordered_map<std::string, GradVarInfo> retv;
+
+  // Create Variable
   for (auto& ptr : backward_op_descs) {
     all_ops.push_back(std::move(ptr));
   }
-  root_block->NewVar(fill_one_op_out);
+  root_block->Var(fill_one_op_out);
 
   // create grad_var for all blocks in this program
-  CreateGradVarInBlock(root_block, forward_op_num);
+  CreateGradVarInBlock(forward_op_num, grad_to_var, root_block, &retv);
   for (size_t block_index = forward_block_num;
        block_index < program_desc.Size(); ++block_index) {
-    CreateGradVarInBlock(program_desc.Block(block_index), 0);
+    CreateGradVarInBlock(0, grad_to_var, program_desc.Block(block_index),
+                         &retv);
   }
+  return retv;
 }
 
 }  // namespace framework

paddle/framework/backward.h

@@ -14,7 +14,10 @@
 
 #pragma once
 
+#include <string>
+#include <unordered_map>
 #include <unordered_set>
+
 #include "paddle/framework/operator.h"
 #include "paddle/framework/program_desc.h"
 
@@ -27,9 +30,17 @@ extern std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars);
 
-// TODO(jiayi): Add target as parameter and generate backward op
+struct GradVarInfo {
-// according to target.
+  std::string name_;
-void AppendBackward(ProgramDescBind& program_desc, const VarDescBind& target,
+  int block_idx_;
+  int op_idx_;
+};
+
+using ParamGradInfoMap = std::unordered_map<std::string /*fwd_var_name*/,
+                                            GradVarInfo /*grad_var_info*/>;
+
+ParamGradInfoMap AppendBackward(
+    ProgramDescBind& program_desc, const VarDescBind& target,
     const std::unordered_set<std::string>& no_grad_vars);
 
 }  // namespace framework

paddle/framework/block_desc.cc

@@ -18,19 +18,22 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-VarDescBind *BlockDescBind::NewVar(const std::string &name) {
+VarDescBind *BlockDescBind::Var(const std::string &name) {
   need_update_ = true;
   auto it = vars_.find(name);
-  PADDLE_ENFORCE(it == vars_.end(), "Duplicated variable %s", name);
+  if (it != vars_.end()) {
-  auto var = new VarDescBind(name);
+    return it->second.get();
+  }
+  auto *var = new VarDescBind(name);
   vars_[name].reset(var);
   return var;
 }
 
-VarDescBind *BlockDescBind::Var(const std::string &name) const {
+VarDescBind *BlockDescBind::FindVar(const std::string &name) const {
   auto it = vars_.find(name);
-  PADDLE_ENFORCE(it != vars_.end(),
+  if (it == vars_.end()) {
-                 "Can not find variable %s in current block.", name);
+    return nullptr;
+  }
   return it->second.get();
 }
 

paddle/framework/block_desc.h

@@ -33,15 +33,6 @@ class ProgramDescBind;
 
 class BlockDescBind {
  public:
-  friend std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
-      ProgramDescBind &program_desc, int block_idx,
-      std::unordered_set<std::string> *no_grad_vars,
-      std::unordered_map<std::string, std::string> *grad_to_var);
-
-  friend void AppendBackward(
-      ProgramDescBind &program_desc, const VarDescBind &target,
-      const std::unordered_set<std::string> &no_grad_vars);
-
   BlockDescBind(ProgramDescBind *prog, BlockDesc *desc)
       : prog_(prog), desc_(desc), need_update_(false) {}
 
@@ -49,9 +40,9 @@ class BlockDescBind {
 
   int32_t Parent() const { return desc_->parent_idx(); }
 
-  VarDescBind *NewVar(const std::string &name_bytes);
+  VarDescBind *Var(const std::string &name_bytes);
 
-  VarDescBind *Var(const std::string &name_bytes) const;
+  VarDescBind *FindVar(const std::string &name_bytes) const;
 
   bool HasVar(const std::string &var_name) const;
 
@@ -69,7 +60,9 @@ class BlockDescBind {
 
   BlockDesc *Proto();
 
- private:
+  // FIXME(yuyang18): backward will access private data of BlockDesc.
+  // Mark it public temporary. We can fix it later.
+ public:
   ProgramDescBind *prog_;  // not_own
   BlockDesc *desc_;        // not_own
   bool need_update_;

paddle/framework/executor.cc

@@ -66,7 +66,7 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id) {
 
   // Instantiate all the vars in the global scope
   for (auto& var : block.vars()) {
-    scope->NewVar(var.name());
+    scope->Var(var.name());
   }
 
   Scope& local_scope = scope->NewScope();
@@ -78,7 +78,7 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id) {
       for (auto& var : block.ops(i).outputs()) {
         for (auto& argu : var.arguments()) {
           if (local_scope.FindVar(argu) == nullptr) {
-            local_scope.NewVar(argu);
+            local_scope.Var(argu);
           }
         }
       }

paddle/framework/executor_test.cc

@@ -46,10 +46,16 @@ void AddOp(const std::string& type, const VariableNameMap& inputs,
   // insert output
   for (auto kv : outputs) {
     for (auto v : kv.second) {
+      // <<<<<<< HEAD
+      //       auto var = block->Var(v);
+      //       var->SetType(VarDesc::LOD_TENSOR);
+      //       var->SetDataType(paddle::framework::DataType::FP32);
+      // =======
       if (!block->HasVar(v)) {
-        auto var = block->NewVar(v);
+        auto var = block->Var(v);
         var->SetDataType(paddle::framework::DataType::FP32);
       }
+      // >>>>>>> origin/develop
     }
   }
 

paddle/framework/operator.h

@@ -403,11 +403,11 @@ class CompileTimeInferShapeContext : public InferShapeContext {
 
  private:
   DDim GetDim(const std::string& name) const override {
-    return framework::make_ddim(block_.Var(name)->Shape());
+    return framework::make_ddim(block_.FindVar(name)->Shape());
   }
 
   void SetDim(const std::string& name, const DDim& dim) override {
-    block_.Var(name)->SetShape(framework::vectorize(dim));
+    block_.FindVar(name)->SetShape(framework::vectorize(dim));
   }
 
   const OpDescBind& op_;

paddle/framework/operator_test.cc

@@ -84,7 +84,7 @@ TEST(OperatorBase, all) {
   paddle::framework::Scope scope;
 
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
-  scope.NewVar("OUT1");
+  scope.Var("OUT1");
   ASSERT_EQ(paddle::framework::op_run_num, 0);
   op->Run(scope, device_context);
   ASSERT_EQ(paddle::framework::op_run_num, 1);
@@ -237,12 +237,12 @@ TEST(OpKernel, multi_inputs) {
 
   paddle::platform::CPUDeviceContext cpu_device_context;
   paddle::framework::Scope scope;
-  scope.NewVar("x0")->GetMutable<Tensor>();
+  scope.Var("x0")->GetMutable<Tensor>();
-  scope.NewVar("x1")->GetMutable<Tensor>();
+  scope.Var("x1")->GetMutable<Tensor>();
-  scope.NewVar("x2")->GetMutable<Tensor>();
+  scope.Var("x2")->GetMutable<Tensor>();
-  scope.NewVar("k0")->GetMutable<Tensor>();
+  scope.Var("k0")->GetMutable<Tensor>();
-  scope.NewVar("y0")->GetMutable<Tensor>();
+  scope.Var("y0")->GetMutable<Tensor>();
-  scope.NewVar("y1")->GetMutable<Tensor>();
+  scope.Var("y1")->GetMutable<Tensor>();
 
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   op->Run(scope, cpu_device_context);

paddle/framework/scope.cc

@@ -31,7 +31,7 @@ Scope& Scope::NewScope() const {
   return *kids_.back();
 }
 
-Variable* Scope::NewVar(const std::string& name) {
+Variable* Scope::Var(const std::string& name) {
   auto iter = vars_.find(name);
   if (iter != vars_.end()) {
     return iter->second;
@@ -42,8 +42,8 @@ Variable* Scope::NewVar(const std::string& name) {
   return v;
 }
 
-Variable* Scope::NewVar() {
+Variable* Scope::Var() {
-  return NewVar(string::Sprintf("%p.%d", this, vars_.size()));
+  return Var(string::Sprintf("%p.%d", this, vars_.size()));
 }
 
 Variable* Scope::FindVar(const std::string& name) const {
@@ -71,8 +71,8 @@ framework::Scope& GetGlobalScope() {
   static std::unique_ptr<framework::Scope> g_scope{nullptr};
   std::call_once(feed_variable_flag, [&]() {
     g_scope.reset(new framework::Scope());
-    g_scope->NewVar("feed_value");
+    g_scope->Var("feed_value");
-    g_scope->NewVar("fetch_value");
+    g_scope->Var("fetch_value");
   });
   return *(g_scope.get());
 }

paddle/framework/scope.h

@@ -45,10 +45,10 @@ class Scope {
   Scope& NewScope() const;
 
   /// Create a variable with given name if it doesn't exist.
-  Variable* NewVar(const std::string& name);
+  Variable* Var(const std::string& name);
 
   /// Create a variable with a scope-unique name.
-  Variable* NewVar();
+  Variable* Var();
 
   /// Find a variable in the scope or any of its ancestors.  Returns
   /// nullptr if cannot find.

paddle/framework/scope_test.cc

@@ -23,8 +23,8 @@ TEST(Scope, VarsShadowing) {
   Scope& ss1 = s.NewScope();
   Scope& ss2 = s.NewScope();
 
-  Variable* v0 = s.NewVar("a");
+  Variable* v0 = s.Var("a");
-  Variable* v1 = ss1.NewVar("a");
+  Variable* v1 = ss1.Var("a");
 
   EXPECT_NE(v0, v1);
 
@@ -40,7 +40,7 @@ TEST(Scope, FindVar) {
   EXPECT_EQ(nullptr, s.FindVar("a"));
   EXPECT_EQ(nullptr, ss.FindVar("a"));
 
-  ss.NewVar("a");
+  ss.Var("a");
 
   EXPECT_EQ(nullptr, s.FindVar("a"));
   EXPECT_NE(nullptr, ss.FindVar("a"));
@@ -49,7 +49,7 @@ TEST(Scope, FindVar) {
 TEST(Scope, FindScope) {
   Scope s;
   Scope& ss = s.NewScope();
-  Variable* v = s.NewVar("a");
+  Variable* v = s.Var("a");
 
   EXPECT_EQ(&s, s.FindScope(v));
   EXPECT_EQ(&s, ss.FindScope(v));

paddle/gserver/layers/Layer.h

@@ -86,6 +86,7 @@ protected:
   /// Also used in 'use_mkldnn' case.
   std::vector<Argument> outputOtherDevice_;
   /// If there are several outputs, map them by each name.
+  /// MKLDNNLayer use it only to merge output grad
   std::map<std::string, Argument*> outputMap_;
   /// Used to merge grad on different devices.
   MatrixPtr tmpGrad_;
@@ -325,6 +326,11 @@ public:
     outputMap_[name] = output;
   }
 
+  /**
+   * Get the output map size, if layer has multi-output.
+   */
+  size_t getOutputMapSize() { return outputMap_.size(); }
+
   /**
    * Get the output based on layer's name.
    */

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -225,8 +225,6 @@ void MKLDNNConvLayer::resetFwdPipeline(
     MKLDNNMatrixPtr& wgt,
     MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-
   if (cvtInVal_) {
     pipeline.push_back(*cvtInVal_);
   }
@@ -245,7 +243,7 @@ void MKLDNNConvLayer::resetFwdPipeline(
 
 void MKLDNNConvLayer::resetInValue(
     std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& in) {
-  const MatrixPtr& inMat = inputLayers_[0]->getOutput().value;
+  const MatrixPtr& inMat = inputLayers_[0]->getOutputValue();
   in = MKLDNNMatrix::create(inMat, pd->src_primitive_desc());
 
   // create buffer and reorder if input value do not match
@@ -310,15 +308,20 @@ void MKLDNNConvLayer::resetOutValue(
     const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
     memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
     cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
-    if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+    if (cpuOutVal_->getPrimitiveDesc() != pd->dst_primitive_desc()) {
+      out = MKLDNNMatrix::create(nullptr, pd->dst_primitive_desc());
       cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
-      CHECK(cvtOutVal_) << "should not be emptry";
+      CHECK(cvtOutVal_) << "should not be empty";
     } else {
-      // CPU output share the same data of MKLDNN output
-      cpuOut->setData(out->getData());
       cpuOutVal_ = out;
     }
+    // when output is cpu device, change the mkldnn output value and make them
+    // share the same data. Then if next layer use inputlayer->getOuputValue()
+    // to achieve the input value, it will get the right data.
+    output_.value = std::dynamic_pointer_cast<Matrix>(cpuOutVal_);
+    return;
   }
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
 }
 
 void MKLDNNConvLayer::resetBwdWgtPD(
@@ -412,8 +415,6 @@ void MKLDNNConvLayer::resetBwdPipeline(
     MKLDNNMatrixPtr& wgt,
     MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-
   if (cvtOutGrad_) {
     pipeline.push_back(*cvtOutGrad_);
   }
@@ -446,28 +447,27 @@ void MKLDNNConvLayer::resetBwdPipeline(
 
 void MKLDNNConvLayer::resetOutGrad(
     std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD, MKLDNNMatrixPtr& out) {
-  const MatrixPtr& outMat = output_.grad;
-  out = MKLDNNMatrix::create(outMat, wgtPD->diff_dst_primitive_desc());
-  CHECK(outVal_ != nullptr &&
-        out->getPrimitiveDesc() == outVal_->getPrimitiveDesc())
-      << "primitive desc of out grad and value should be equal";
-
-  // TODO(TJ): merge outgrad
-  // create reorder if has output grad does not match
   cpuOutGrad_ = nullptr;
   cvtOutGrad_ = nullptr;
-  if (!outputIsOnlyMKLDNN()) {
+  CHECK(outVal_ != nullptr &&
+        outVal_->getPrimitiveDesc() == wgtPD->diff_dst_primitive_desc())
+      << "primitive desc of out grad and value should be equal";
+  if (outputIsOnlyMKLDNN()) {
+    MKLDNNLayer::resetOutGrad(out, outVal_->getPrimitiveDesc());
+  } else {
     const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
-    outMat->setData(cpuOut->getData());
     // same PrimitiveDesc with cpuInVal_
     CHECK(cpuOutVal_);
     cpuOutGrad_ = MKLDNNMatrix::create(cpuOut, cpuOutVal_->getPrimitiveDesc());
-    if (cpuOutGrad_->getPrimitiveDesc() == out->getPrimitiveDesc()) {
+    // create reorder if primitive desc does not match
-      out = cpuOutGrad_;
+    if (cpuOutGrad_->getPrimitiveDesc() != outVal_->getPrimitiveDesc()) {
-    } else {
+      out = MKLDNNMatrix::create(output_.grad, outVal_->getPrimitiveDesc());
-      out = MKLDNNMatrix::create(nullptr, wgtPD->diff_dst_primitive_desc());
       cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
       CHECK(cvtOutGrad_);
+    } else {
+      // share the same data of CPU output
+      output_.grad->setData(cpuOut->getData());
+      out = cpuOutGrad_;
     }
   }
 }
@@ -496,32 +496,30 @@ void MKLDNNConvLayer::resetWgtBiasGrad(
 void MKLDNNConvLayer::resetInGrad(
     std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
     MKLDNNMatrixPtr& in) {
+  in = nullptr;
+  cpuInGrad_ = nullptr;
+  cvtInGrad_ = nullptr;
   if (dataPD == nullptr) {
     return;
   }
 
-  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
+  if (inputIsOnlyMKLDNN()) {
-  in = MKLDNNMatrix::create(inputLayers_[0]->getOutput().grad,
+    MKLDNNLayer::resetInGrad(in, dataPD->diff_src_primitive_desc());
-                            dataPD->diff_src_primitive_desc());
     CHECK(nullptr != inVal_ &&
           in->getPrimitiveDesc() == inVal_->getPrimitiveDesc())
         << "primitive desc of input grad and value should be equal";
-
+  } else {
-  // create reorder if has output grad does not match
-  cpuInGrad_ = nullptr;
-  cvtInGrad_ = nullptr;
-  if (!inputIsOnlyMKLDNN()) {
     const MatrixPtr& cpuIn = getInputGrad(0, CPU_DEVICE);
     // same PrimitiveDesc with cpuInVal_
     CHECK(cpuInVal_);
     cpuInGrad_ = MKLDNNMatrix::create(cpuIn, cpuInVal_->getPrimitiveDesc());
-    if (cpuInGrad_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+    in = cpuInGrad_;
-      const MatrixPtr& dnnIn = getInputGrad(0, MKLDNN_DEVICE);
+    // create reorder if PrimitiveDesc does not match
-      in = MKLDNNMatrix::create(dnnIn, in->getPrimitiveDesc());
+    if (cpuInGrad_->getPrimitiveDesc() != dataPD->diff_src_primitive_desc()) {
+      in = MKLDNNMatrix::create(getInputGrad(0, MKLDNN_DEVICE),
+                                dataPD->diff_src_primitive_desc());
       cvtInGrad_ = MKLDNNMatrix::createReorder(in, cpuInGrad_);
       CHECK(cvtInGrad_);
-    } else {
-      in = cpuInGrad_;
     }
   }
 }

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -180,10 +180,10 @@ void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
 void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
   out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
   if (!outputIsOnlyMKLDNN()) {
-    // fc cpu output value do not need create convert
+    // fc cpu output value do not need create convert, just share data
-    // just share point
     getOutput(CPU_DEVICE).value->setData(out->getData());
   }
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
 }
 
 void MKLDNNFcLayer::resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
@@ -214,8 +214,6 @@ void MKLDNNFcLayer::resetFwdPipeline(
     MKLDNNMatrixPtr& wgt,
     MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-
   if (bias) {
     fwd_.reset(new fc_fwd(*pd, *in, *wgt, *bias, *out));
   } else {
@@ -237,19 +235,14 @@ void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
 }
 
 void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
-  // TODO(TJ): merge outgrad
-  int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
-  output_.grad->setData(getOutput(device).grad->getData());
-  // for MKLDNN device:
-  // can not directly cast outputgrad to mkldnnmatrix,
-  // since each layer can not write the inputgrad to mkldnn inputgrad.
-  // So just create from matrix with outputvalue format.
-  // for CPU device:
-  // fc do not need to convert from cpu device since output is always nc format
-  // only need create from cpu device
   CHECK(outVal_);
-  out =
+  if (outputIsOnlyMKLDNN()) {
-      MKLDNNMatrix::create(getOutput(device).grad, outVal_->getPrimitiveDesc());
+    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,
@@ -267,13 +260,11 @@ void MKLDNNFcLayer::resetWgtBiasGrad(MKLDNNMatrixPtr& wgt,
 
 void MKLDNNFcLayer::resetInGrad(MKLDNNMatrixPtr& in) {
   in = nullptr;
-  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inputLayers_[0]->getOutput().grad == nullptr) {
-  if (inGrad == nullptr) {
     return;
   }
-  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
   CHECK(inVal_);
-  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+  MKLDNNLayer::resetInGrad(in, inVal_->getPrimitiveDesc());
 }
 
 void MKLDNNFcLayer::resetBwdWgtPD(
@@ -314,7 +305,6 @@ void MKLDNNFcLayer::resetBwdPipeline(
     MKLDNNMatrixPtr& wgt,
     MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
-  pipeline.clear();
   CHECK(inVal_);
   if (bias) {
     bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt, *bias));

paddle/gserver/layers/MKLDNNLayer.h

@@ -65,6 +65,17 @@ protected:
   MKLDNNMatrixPtr biasVal_;
   MKLDNNMatrixPtr biasGrad_;
 
+  // merge grad primitive
+  std::shared_ptr<mkldnn::primitive> mergeGrad_;
+  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)
       : Layer(config),
@@ -99,6 +110,7 @@ public:
     if (!Layer::init(layerMap, parameterMap)) {
       return false;
     }
+    setOutputMap();
     checkCPUOutputsNumber();
 
     stream_.reset(new MKLDNNStream());
@@ -118,12 +130,9 @@ public:
         VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward";
         // reset when input total sizes changed, not only the batchsize
         inputElemenCnt_ = elemenCnt;
+        pipelineFwd_.clear();
         reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
         resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
-        if (outVal_) {
-          // change original output value to mkldnn output value
-          output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
-        }
         convertWeightsFromPaddle();
         needResetBwd_ = true;
       }
@@ -144,9 +153,18 @@ public:
   void backward(const UpdateCallback& callback) override {
     if (needResetBwd_) {
       VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward";
+      pipelineBwd_.clear();
+      pipelineMergeGrad_.clear();
+      mergeGrad_ = nullptr;
       resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
       needResetBwd_ = false;
     }
+
+    // merge grad must before backward activation
+    if (mergeGrad_) {
+      REGISTER_TIMER_INFO("MergeBpGrad", getName().c_str());
+      stream_->submit(pipelineMergeGrad_);
+    }
     {
       REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
       backwardActivation();
@@ -247,6 +265,76 @@ protected:
     }
   }
 
+  /**
+   * reset the output grad matrix from primitive desc.
+   * and reset the merge grad primitive if needed.
+   * note: when this layer has serval outputs,
+   *       it could not be mixed with cpu device,
+   *       since it can not get memory desc from cpu device.
+   */
+  virtual void resetOutGrad(MKLDNNMatrixPtr& out,
+                            mkldnn::memory::primitive_desc pd) {
+    CHECK(outputIsOnlyMKLDNN()) << "do not support mixed with other device yet";
+    mergeGrad_ = nullptr;
+    pipelineMergeGrad_.clear();
+    out = MKLDNNMatrix::create(output_.grad, pd);
+    if (outputMap_.size() <= 1) {
+      return;
+    }
+    std::vector<double> scales(outputMap_.size(), 1.0);
+    std::vector<mkldnn::memory::primitive_desc> srcPDs;
+    std::vector<mkldnn::primitive::at> srcs;
+    for (auto it = outputMap_.begin(); it != outputMap_.end(); ++it) {
+      MKLDNNMatrixPtr src =
+          std::dynamic_pointer_cast<MKLDNNMatrix>(it->second->grad);
+      VLOG(MKLDNN_BASE) << getName() << " has output grad " << it->first;
+      CHECK(src) << "should be MKLDNNMatrix";
+      auto srcDims = src->getDims();
+      auto dstDims = out->getDims();
+      CHECK_EQ(srcDims.size(), dstDims.size());
+      for (size_t i = 0; i < srcDims.size(); ++i) {
+        CHECK_EQ(srcDims[i], dstDims[i]);
+      }
+      srcPDs.push_back(src->getPrimitiveDesc());
+      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_ = nullptr;
+    tmpCvt_ = nullptr;
+    if (out->getPrimitiveDesc() != srcPDs[0]) {
+      tmpOutGrad_ = MKLDNNMatrix::create(nullptr, srcPDs[0]);
+      tmpCvt_ = MKLDNNMatrix::createReorder(tmpOutGrad_, out);
+      CHECK(tmpCvt_);
+      pipelineMergeGrad_.push_back(*tmpCvt_);
+    } else {
+      tmpOutGrad_ = out;
+    }
+
+    auto sumPD = mkldnn::sum::primitive_desc(
+        tmpOutGrad_->getMemoryDesc(), scales, srcPDs);
+    mergeGrad_.reset(new mkldnn::sum(sumPD, srcs, *tmpOutGrad_));
+    pipelineMergeGrad_.insert(pipelineMergeGrad_.begin(), *mergeGrad_);
+  }
+
+  /**
+   * reset input grad from primitive desc.
+   * this function is avaiable for input is only mkldnn
+   * or input do not care cpu device
+   */
+  virtual void resetInGrad(MKLDNNMatrixPtr& in,
+                           mkldnn::memory::primitive_desc pd) {
+    LayerPtr& input = inputLayers_[0];
+    const MatrixPtr& grad =
+        input->getOutputMapSize() > 1 ? nullptr : input->getOutput().grad;
+    in = MKLDNNMatrix::create(grad, pd);
+    Argument& arg = input->getOutput(this->getName());
+    arg.grad = std::dynamic_pointer_cast<Matrix>(in);
+  }
+
   /**
    * print info about sizes
    */
@@ -334,6 +422,16 @@ private:
     }
   }
 
+  /**
+   * Set output map of prev layers.
+   */
+  void setOutputMap() {
+    outputMap_.clear();
+    for (size_t i = 0; i < inputLayers_.size(); ++i) {
+      inputLayers_[i]->setOutput(getName(), &tmpInArg_);
+    }
+  }
+
   /**
    * Check the cpu device number of outputOtherDevice_.
    * should have only one at most.

paddle/gserver/layers/MKLDNNPoolLayer.cpp

@@ -142,14 +142,16 @@ void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
     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 {
-      // CPU output share the same data of MKLDNN output
-      cpuOut->setData(out->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,
@@ -187,7 +189,6 @@ void MKLDNNPoolLayer::resetFwdPipeline(
     std::shared_ptr<pool_fwd::primitive_desc>& pd,
     MKLDNNMatrixPtr& in,
     MKLDNNMatrixPtr& out) {
-  pipeline.clear();
   fwd_ = workspace_
              ? std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out, *workspace_))
              : std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out));
@@ -205,17 +206,17 @@ void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
   resetInGrad(in);
 }
 void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
-  CHECK(outVal_) << "Should have output value";
-  out = MKLDNNMatrix::create(output_.grad, outVal_->getPrimitiveDesc());
-
-  // create reorder if output value has cpu device and pd do not match
   cpuOutGrad_ = nullptr;
   cvtOutGrad_ = nullptr;
-  if (!outputIsOnlyMKLDNN()) {
+  CHECK(outVal_);
+  if (outputIsOnlyMKLDNN()) {
+    MKLDNNLayer::resetOutGrad(out, outVal_->getPrimitiveDesc());
+  } else {
     const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
     cpuOutGrad_ = MKLDNNMatrix::create(
         cpuOut, memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_);
-    if (cpuOutGrad_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+    if (cpuOutGrad_->getPrimitiveDesc() != outVal_->getPrimitiveDesc()) {
+      out = MKLDNNMatrix::create(output_.grad, outVal_->getPrimitiveDesc());
       cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
       CHECK(cvtOutGrad_) << "should not be emptry";
     } else {
@@ -228,12 +229,11 @@ void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
 
 void MKLDNNPoolLayer::resetInGrad(MKLDNNMatrixPtr& in) {
   in = nullptr;
-  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inputLayers_[0]->getOutput().grad == nullptr) {
-  if (inGrad == nullptr) {
     return;
   }
   CHECK(inVal_);
-  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+  MKLDNNLayer::resetInGrad(in, inVal_->getPrimitiveDesc());
 }
 
 void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
@@ -261,7 +261,6 @@ void MKLDNNPoolLayer::resetBwdPipeline(
     std::shared_ptr<pool_bwd::primitive_desc>& pd,
     MKLDNNMatrixPtr& in,
     MKLDNNMatrixPtr& out) {
-  pipeline.clear();
   if (cvtOutGrad_) {
     pipeline.push_back(*cvtOutGrad_);
   }

paddle/gserver/tests/MKLDNNTester.cpp

@@ -124,8 +124,8 @@ void MKLDNNTester::randomTopDiffs() {
 void MKLDNNTester::checkForward() {
   VLOG(MKLDNN_ALL) << "Check Forward";
   printTopDatas();
-  double delta = compareMatrix(dnnLayer_->getOutput(CPU_DEVICE).value,
+  double delta =
-                               refLayer_->getOutputValue());
+      compareMatrix(dnnLayer_->getOutputValue(), refLayer_->getOutputValue());
   EXPECT_LE(fabs(delta), eps_);
 }
 

paddle/operators/cond_op.cc

@@ -134,7 +134,7 @@ void CondOp::PrepareDataForSubnet(
   for (int i = 0; i < BRANCH_NUM; ++i) {
     for (auto& output : (*sub_net_op_[i]).Outputs()) {
       for (auto& var_name : output.second) {
-        sub_scopes[i]->NewVar(var_name);
+        sub_scopes[i]->Var(var_name);
       }
     }
   }

paddle/operators/dynamic_recurrent_op.cc

@@ -30,7 +30,7 @@ namespace detail {
 inline void CreateVariables(Scope& scope,
                             const std::vector<std::string>& var_names) {
   for (const auto& name : var_names) {
-    scope.NewVar(name);
+    scope.Var(name);
   }
 }
 
@@ -136,7 +136,7 @@ void DynamicRecurrentOp::WriteStepInputs() const {
       auto& step_scope = cache_.GetScope(step);
       Variable* var = step_scope.FindVar(item.first);
       if (var == nullptr) {
-        var = step_scope.NewVar(item.first);
+        var = step_scope.Var(item.first);
       }
       var->GetMutable<LoDTensor>()->ShareDataWith<value_type>(tensor);
     }

paddle/operators/dynamic_recurrent_op_test.cc

@@ -36,7 +36,7 @@ void OpDescNewVar(const std::string& param_name,
 // create a LoD tensor in scope with specific dims
 LoDTensor* CreateVar(Scope& scope, std::string name, framework::DDim dims,
                      const platform::Place& place) {
-  auto* var = scope.NewVar(name);
+  auto* var = scope.Var(name);
   auto* tensor = var->GetMutable<LoDTensor>();
   tensor->Resize(dims);
   tensor->mutable_data<float>(place);
@@ -85,7 +85,7 @@ class DynamicRecurrentOpTestHelper : public ::testing::Test {
 
   void CreateGlobalVariables() {
     platform::CPUPlace place;
-    scope.NewVar("step_scopes");
+    scope.Var("step_scopes");
     CreateVar(scope, "boot_mem", framework::make_ddim({10, 20}), place);
     CreateVar(scope, "out0", framework::make_ddim({10, 20}), place);
     auto* in0 = CreateVar(scope, "in0", framework::make_ddim({10, 8}), place);

paddle/operators/recurrent_op.cc

@@ -70,14 +70,14 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope,
         // the weight are located in parent scope
         for (auto& var_name : input.second) {
           if (!step_scope.FindVar(var_name)) {
-            step_scope.NewVar(var_name)->GetMutable<LoDTensor>();
+            step_scope.Var(var_name)->GetMutable<LoDTensor>();
           }
         }
       }
       // create stepnet's outputs
       for (const auto& output : (*stepnet_)->Outputs()) {
         for (auto& var_name : output.second) {
-          step_scope.NewVar(var_name);
+          step_scope.Var(var_name);
         }
       }
       step_scopes->emplace_back(&step_scope);
@@ -87,7 +87,7 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope,
 
 void RecurrentAlgorithm::InitMemories(Scope* step_scope) const {
   for (auto& attr : arg_->memories) {
-    auto* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable<LoDTensor>();
+    auto* pre_mem = step_scope->Var(attr.pre_var)->GetMutable<LoDTensor>();
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
                    "memory [%s]'s boot variable [%s] not exists", attr.var,
                    attr.boot_var);
@@ -167,9 +167,9 @@ void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
                    "memory variable [%s] does not exists", attr.var);
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
                    "boot variable [%s] does not exists", attr.boot_var);
-    auto* mem_grad = step_scope->NewVar(attr.var)->GetMutable<LoDTensor>();
+    auto* mem_grad = step_scope->Var(attr.var)->GetMutable<LoDTensor>();
     auto* boot_mem_grad =
-        step_scope->NewVar(attr.boot_var)->GetMutable<LoDTensor>();
+        step_scope->Var(attr.boot_var)->GetMutable<LoDTensor>();
     boot_mem_grad->Resize(mem_grad->dims());
     boot_mem_grad->ShareDataWith<float>(*mem_grad);
   }

paddle/operators/rnn/recurrent_op_utils.cc

@@ -40,7 +40,7 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
     f::DDim step_dims = slice_ddim(dims, 1, dims.size());
     for (size_t j = 0; j < seq_len; j++) {
       Tensor* step_input =
-          step_scopes[j]->NewVar(inlinks[i])->GetMutable<Tensor>();
+          step_scopes[j]->Var(inlinks[i])->GetMutable<Tensor>();
       // The input of operators of each step is Tensor here.
       // Maybe need to modify Slice function.
       *step_input = input->Slice<float>(j, j + 1);