Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into some_small_fixes

Dockerfile

@@ -76,7 +76,8 @@ RUN easy_install -U pip && \
     pip install sphinx-rtd-theme==0.1.9 recommonmark
 
 RUN pip install pre-commit 'ipython==5.3.0' && \
-    pip install 'ipykernel==4.6.0' 'jupyter==1.0.0'
+    pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
+    pip install opencv-python
 
 #For docstring checker
 RUN pip install pylint pytest astroid isort

cmake/external/mkldnn.cmake

@@ -54,7 +54,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "db3424ad44901513c03a1ea31ccaacdf633fbe9f"
+    GIT_TAG             "a29d8487a63afca3d5b8c5bbdbb473cf8ccc6e51"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}

doc/v2/faq/build_and_install/index_cn.rst

@@ -213,3 +213,12 @@ virtualenv本身也是Python的一个包，可以用pip进行安装：
 保存并关闭文件。
 
 这样，每次打开终端时就会自动启动名为‘paddle’的Python环境了。
+
+10. 通过pip安装的PaddlePaddle在  :code:`import paddle.fluid` 报找不到 :code:`libmkldnn.so` 或 :code:`libmklml_intel.so`
+------------------------------------------------------------------------------------------
+出现这种问题的原因是在导入 :code:`paddle.fluid` 时需要加载 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so`，
+但是系统没有找到该文件。一般通过pip安装PaddlePaddle时会将 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so`
+拷贝到 :code:`/usr/local/lib` 路径下，所以解决办法是将该路径加到 :code:`LD_LIBRARY_PATH` 环境变量下，
+即： :code:`export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH` 。
+
+**注意**：如果是在虚拟环境中安装PaddlePaddle， :code:`libmkldnn.so` 和 :code:`libmklml_intel.so` 可能不在 :code:`/usr/local/lib` 路径下。
\ No newline at end of file

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -73,6 +73,9 @@ void BroadcastOpHandle::RunImpl() {
     int root_id = boost::get<platform::CUDAPlace>(in_tensor.place()).device;
     std::vector<std::function<void()>> broadcast_calls;
 
+    int type = platform::ToNCCLDataType(in_tensor.type());
+    size_t numel = static_cast<size_t>(in_tensor.numel());
+
     for (auto out_var_handle : out_var_handles) {
       Variable *out_var = var_scopes.at(out_var_handle->scope_idx_)
                               ->FindVar(out_var_handle->name_);
@@ -87,13 +90,11 @@ void BroadcastOpHandle::RunImpl() {
         send_recv_buffer = const_cast<void *>(in_tensor.data<void>());
         out_handle = out_var_handle;
       } else {
-        send_recv_buffer =
-            VariableVisitor::GetMutableTensor(out_var).mutable_data(
-                out_var_handle->place_);
+        send_recv_buffer = VariableVisitor::GetMutableTensor(out_var)
+                               .Resize(in_tensor.dims())
+                               .mutable_data(out_var_handle->place_);
       }
 
-      int type = platform::ToNCCLDataType(in_tensor.type());
-      size_t numel = static_cast<size_t>(in_tensor.numel());
       broadcast_calls.emplace_back(
           [send_recv_buffer, numel, type, root_id, &nccl_ctx] {
             PADDLE_ENFORCE(platform::dynload::ncclBcast(

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -351,7 +351,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(SSAGraph *result,
     auto &prev_grad = vars.back();
     op_handle->AddInput(prev_grad.get());
 
-    auto var = new VarHandle(vars.size() - 1, i, og, p);
+    auto var = new VarHandle(vars.size(), i, og, p);
     vars.emplace_back(var);
     op_handle->AddOutput(var);
   }
@@ -447,8 +447,7 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result,
     op_handle->AddInput(prev_grad.get());
   }
   auto &vars = result->vars_[dst_dev_id][og];
-  auto var =
-      new VarHandle(vars.size() - 1, dst_dev_id, og, places_[dst_dev_id]);
+  auto var = new VarHandle(vars.size(), dst_dev_id, og, places_[dst_dev_id]);
   vars.emplace_back(var);
   op_handle->AddOutput(var);
   return var;

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -47,7 +47,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
-  int GetVarDeviceID(const std::string &varname) const;
+  int GetVarDeviceID(const std::string &varname) const override;
 
  private:
   void CreateOpHandleIOs(SSAGraph *result, const OpDesc &op,

paddle/fluid/framework/details/op_handle_base.cc

@@ -11,8 +11,8 @@
 // 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 "paddle/fluid/framework/details/op_handle_base.h"
+#include <map>
 
 namespace paddle {
 namespace framework {
@@ -122,11 +122,16 @@ void OpHandleBase::RunAndRecordEvent(const std::function<void()> &callback) {
 #ifdef PADDLE_WITH_CUDA
   if (!events_.empty()) {  // Use event
     std::function<void()> method = callback;
-
+    // NOTE(zcd): device context must be ordered here because RecordEvent
+    // will use a mutex to ensure the safe of multi-threads.
+    std::map<platform::DeviceContext *, platform::Place> ordered_ctxes;
     for (auto &p : dev_ctxes_) {
+      ordered_ctxes.emplace(p.second, p.first);
+    }
+    for (auto &p : ordered_ctxes) {
       method = [method, p, this]() {
-        static_cast<platform::CUDADeviceContext *>(p.second)->RecordEvent(
-            events_.at(boost::get<platform::CUDAPlace>(p.first).device),
+        static_cast<platform::CUDADeviceContext *>(p.first)->RecordEvent(
+            events_.at(boost::get<platform::CUDAPlace>(p.second).device),
             method);
       };
     }

paddle/fluid/framework/framework.proto

@@ -27,6 +27,7 @@ enum AttrType {
   BOOLEANS = 7;
   BLOCK = 8;
   LONG = 9;
+  BLOCKS = 10;
 }
 
 // OpDesc describes an instance of a C++ framework::OperatorBase
@@ -46,6 +47,7 @@ message OpDesc {
     repeated bool bools = 11;
     optional int32 block_idx = 12;
     optional int64 l = 13;
+    repeated int32 blocks_idx = 14;
   };
 
   message Var {

paddle/fluid/framework/lod_tensor.cc

@@ -51,8 +51,6 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) {
 }
 
 std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
-  PADDLE_ENFORCE(t.type().hash_code() == typeid(float).hash_code());
-
   if (!platform::is_cpu_place(t.place())) {
     LoDTensor tt;
     framework::TensorCopy(t, platform::CPUPlace(), &tt);
@@ -70,7 +68,13 @@ std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
   // only print first ten elements
   int64_t size = t.numel() < 10 ? t.numel() : 10;
   for (int64_t i = 0; i < size; ++i) {
-    os << t.data<float>()[i] << " ";
+    if (t.type().hash_code() == typeid(float).hash_code()) {
+      os << t.data<float>()[i] << " ";
+    } else if (t.type().hash_code() == typeid(int64_t).hash_code()) {
+      os << t.data<int64_t>()[i] << " ";
+    } else {
+      PADDLE_THROW("LoDTensor data type not in [float, int64_t]");
+    }
   }
 
   return os;

paddle/fluid/framework/lod_tensor_test.cc

@@ -26,6 +26,20 @@
 namespace paddle {
 namespace framework {
 
+TEST(LoD, PrintLoDTensor) {
+  LoDTensor tensor1;
+  tensor1.mutable_data<float>(platform::CPUPlace());
+  tensor1.data<float>()[0] = 0.2;
+  tensor1.data<float>()[1] = 0.5;
+  LOG(INFO) << tensor1;
+
+  LoDTensor tensor2;
+  tensor2.mutable_data<int64_t>(platform::CPUPlace());
+  tensor2.data<int64_t>()[0] = 1;
+  tensor2.data<int64_t>()[1] = 2;
+  LOG(INFO) << tensor2;
+}
+
 TEST(LoD, data) {
   LoD lod{{0, 1, 2}};
   lod.push_back({0, 2, 4, 5});
@@ -37,7 +51,7 @@ TEST(LoD, data) {
   }
 }
 
-TEST(LodExpand, test) {
+TEST(LoD, ExpandLoD) {
   LoD lod{{0, 2}};
   LoDTensor tensor;
   tensor.set_lod(lod);

paddle/fluid/framework/op_desc.cc

@@ -211,6 +211,12 @@ void OpDesc::SetBlockAttr(const std::string &name, BlockDesc *block) {
   need_update_ = true;
 }
 
+void OpDesc::SetBlocksAttr(const std::string &name,
+                           std::vector<BlockDesc *> blocks) {
+  this->attrs_[name] = blocks;
+  need_update_ = true;
+}
+
 void OpDesc::SetAttrMap(
     const std::unordered_map<std::string, Attribute> &attr_map) {
   attrs_ = attr_map;
@@ -305,6 +311,13 @@ struct SetAttrDescVisitor : public boost::static_visitor<void> {
   void operator()(const std::vector<bool> &v) const {
     VectorToRepeated(v, attr_->mutable_bools());
   }
+  void operator()(const std::vector<BlockDesc *> &v) const {
+    std::vector<int> blocks_idx;
+    for (auto blk : v) {
+      blocks_idx.push_back(blk->ID());
+    }
+    VectorToRepeated(blocks_idx, attr_->mutable_blocks_idx());
+  }
   void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); }
   void operator()(int64_t v) const { attr_->set_l(v); }
   void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }

paddle/fluid/framework/op_desc.h

@@ -77,6 +77,8 @@ class OpDesc {
 
   void SetBlockAttr(const std::string &name, BlockDesc *block);
 
+  void SetBlocksAttr(const std::string &name, std::vector<BlockDesc *> blocks);
+
   Attribute GetAttr(const std::string &name) const;
 
   Attribute GetNullableAttr(const std::string &name) const;

paddle/fluid/framework/parallel_executor.cc

@@ -121,7 +121,7 @@ ParallelExecutor::ParallelExecutor(
 #endif
   }
 
-  builder_ = std::move(builder_factory.Create());
+  builder_ = builder_factory.Create();
   member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
       exec_strategy, member_->local_scopes_, places,
       builder_->Build(main_program)));

paddle/fluid/framework/type_defs.h

@@ -35,7 +35,8 @@ using VariableNameMap = std::map<std::string, std::vector<std::string>>;
 using Attribute =
     boost::variant<boost::blank, int, float, std::string, std::vector<int>,
                    std::vector<float>, std::vector<std::string>, bool,
-                   std::vector<bool>, BlockDesc*, int64_t>;
+                   std::vector<bool>, BlockDesc*, int64_t,
+                   std::vector<BlockDesc*>>;
 
 using AttributeMap = std::unordered_map<std::string, Attribute>;
 

paddle/fluid/operators/distributed/grpc_client.cc

@@ -18,6 +18,7 @@ limitations under the License. */
 
 #include <limits>
 
+#include "glog/logging.h"  // For VLOG
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/operators/distributed/request_handler.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -75,6 +76,9 @@ bool GRPCClient::AsyncSendVar(const std::string& ep,
     var_h.scope = p_scope;
     var_h.name = var_name_val;
     var_h.ctx = p_ctx;
+    var_h.method = "Send";
+
+    VLOG(3) << var_h.String() << " begin";
 
     // stub context
     SendProcessor* s = new SendProcessor(ch);
@@ -129,6 +133,9 @@ bool GRPCClient::AsyncGetVar(const std::string& ep,
     var_h.scope = p_scope;
     var_h.name = var_name_val;
     var_h.ctx = p_ctx;
+    var_h.method = "Get";
+
+    VLOG(3) << var_h.String() << " begin";
 
     // stub context
     GetProcessor* s = new GetProcessor(ch);
@@ -172,6 +179,9 @@ bool GRPCClient::AsyncPrefetchVar(const std::string& ep,
     var_h.scope = p_scope;
     var_h.name = out_var_name_val;
     var_h.ctx = p_ctx;
+    var_h.method = "Prefetch";
+
+    VLOG(3) << var_h.String() << " begin";
 
     // stub context
     GetProcessor* s = new GetProcessor(ch);
@@ -243,10 +253,11 @@ void GRPCClient::Proceed() {
     GPR_ASSERT(ok);
     PADDLE_ENFORCE(c);
     if (c->status_.ok()) {
+      VLOG(3) << c->var_h_.String() << " process";
       c->Process();
     } else {
-      LOG(FATAL) << "var: " << c->var_h_.String()
-                 << " grpc error:" << c->status_.error_message();
+      LOG(FATAL) << c->var_h_.String()
+                 << " meets grpc error:" << c->status_.error_message();
     }
     delete c;
     {

paddle/fluid/operators/distributed/grpc_client.h

@@ -47,14 +47,18 @@ namespace operators {
 namespace distributed {
 
 struct VarHandle {
+  // RPC endpoint.
   std::string ep;
   const platform::DeviceContext* ctx;
   const framework::Scope* scope;
+  // Variable name.
   std::string name;
+  // RPC method name.
+  std::string method;
 
   std::string String() const {
     std::ostringstream s;
-    s << "name:[" << name << "] ep:[" << ep << "]";
+    s << method << " name:[" << name << "], ep:[" << ep << "]";
     return s.str();
   }
 };

paddle/fluid/operators/distributed/grpc_server.cc

@@ -41,6 +41,19 @@ class RequestBase {
   virtual ~RequestBase() {}
   virtual void Process() = 0;
 
+  std::string Status2String(const std::string& method) {
+    std::string status = "Process";
+    if (status_ == FINISH) {
+      status = "Finish";
+    }
+
+    std::ostringstream s;
+    s << method << " name:[" << GetReqName() << "]"
+      << ", ep:[" << ctx_.peer() << "]"
+      << " " << status << " using req_id:" << req_id_;
+    return s.str();
+  }
+
   CallStatus Status() const {
     std::lock_guard<std::mutex> l(status_mu_);
     return status_;
@@ -272,7 +285,7 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name,
                                           int req_id) {
   std::unique_lock<std::mutex> lock(cq_mutex_);
   if (is_shut_down_) {
-    VLOG(3) << "shutdown, do not TryToRegisterNewSendOne";
+    LOG(WARNING) << "shutdown, do not TryToRegisterNewSendOne";
     return;
   }
 
@@ -306,14 +319,14 @@ void AsyncGRPCServer::HandleRequest(
   bool ok = false;
 
   while (true) {
-    VLOG(3) << "HandleRequest " << rpc_name << " wait next";
+    VLOG(4) << "HandleRequest " << rpc_name << " wait next";
     if (!cq->Next(&tag, &ok)) {
       LOG(INFO) << "CompletionQueue " << rpc_name << " shutdown!";
       break;
     }
 
     int req_id = static_cast<int>(reinterpret_cast<intptr_t>(tag));
-    VLOG(3) << "HandleRequest " << rpc_name << ", req_id:" << req_id
+    VLOG(4) << "HandleRequest " << rpc_name << ", req_id:" << req_id
             << " get next";
 
     auto& reqs = rpc_reqs_[rpc_name];
@@ -324,22 +337,21 @@ void AsyncGRPCServer::HandleRequest(
       base = reqs[req_id];
     }
 
+    VLOG(3) << base->Status2String(rpc_name);
+
     // reference:
     // https://github.com/tensorflow/tensorflow/issues/5596
     // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM
     // https://groups.google.com/forum/#!topic/grpc-io/ywATt88Ef_I
     if (!ok) {
       LOG(WARNING) << "completion queue:" << rpc_name
-                   << " recv no regular event:argument name["
-                   << base->GetReqName() << "]";
+                   << " recv no regular event"
+                   << " context:" << base->Status2String(rpc_name);
       TryToRegisterNewOne(rpc_name, req_id);
       delete base;
       continue;
     }
 
-    VLOG(3) << "queue id:" << rpc_name << ", req_id:" << req_id
-            << ", status:" << base->Status();
-
     switch (base->Status()) {
       case PROCESS: {
         base->Process();

paddle/fluid/operators/distributed/variable_response.cc

@@ -76,6 +76,8 @@ bool ReadRaw(::google::protobuf::io::CodedInputStream* input,
       if (total_written + size_to_write > length) {
         size_to_write = length - total_written;
       }
+      // This log is useful to see how long a internal block size is of rpc.
+      VLOG(7) << "copy " << size_to_write << " data to CUDAPlace";
       memory::Copy(boost::get<platform::CUDAPlace>(place),
                    reinterpret_cast<void*>(p), cpu, data, size_to_write,
                    gpu_dev_ctx.stream());
@@ -103,6 +105,8 @@ bool ReadRaw(::google::protobuf::io::CodedInputStream* input,
     }
     // TODO(gongwb): can we avoid copy?
     platform::CPUPlace cpu;
+    // This log is useful to see how long a internal block size is of rpc.
+    VLOG(7) << "copy " << size_to_write << " data to CPUPlace";
     memory::Copy(cpu, reinterpret_cast<void*>(p), cpu, data, size_to_write);
 
     p += size_to_write;

paddle/fluid/operators/listen_and_serv_op.cc

@@ -101,17 +101,16 @@ void ListenAndServOp::RunSyncLoop(
     framework::Scope *recv_scope,
     const std::vector<int> &prefetch_block_id_list) const {
   size_t num_blocks = program->Size();
+  auto optimize_blocks =
+      Attr<std::vector<framework::BlockDesc *>>(kOptimizeBlocks);
   PADDLE_ENFORCE_GE(num_blocks, 2,
                     "server program should have at least 2 blocks");
 
-  std::vector<int> optimize_block_id_list;
-  for (int blkid = 1; blkid < num_blocks; ++blkid) {
-    if (std::find(prefetch_block_id_list.begin(), prefetch_block_id_list.end(),
-                  blkid) == prefetch_block_id_list.end()) {
-      optimize_block_id_list.push_back(blkid);
-    }
+  std::vector<int> optimize_blocks_idx;
+  for (auto blk : optimize_blocks) {
+    optimize_blocks_idx.push_back(blk->ID());
   }
-  auto optimize_prepared = executor->Prepare(*program, optimize_block_id_list);
+  auto optimize_prepared = executor->Prepare(*program, optimize_blocks_idx);
   // Insert placeholder for block0 which holds current op itself.
   optimize_prepared.insert(
       optimize_prepared.begin(),
@@ -134,14 +133,14 @@ void ListenAndServOp::RunSyncLoop(
     // and this will still work.
     // The optimize blocks which have the same parent ID would run parallel
     // TODO(Yancey1989): need to use ParallelExecutor for future
-    int32_t last_parent_blkid = program->Block(1).Parent();
+    int32_t last_parent_blkid = optimize_blocks[0]->Parent();
     std::vector<size_t> parallel_blkids;
-    parallel_blkids.push_back(1);
+    parallel_blkids.push_back(optimize_blocks[0]->ID());
     double ts = GetTimestamp();
-    for (size_t i = 1; i < optimize_block_id_list.size(); ++i) {
+    for (size_t i = 1; i < optimize_blocks.size(); ++i) {
       // skip the first optimize block because it is already in the
       // parallel_blkids.
-      int blkid = optimize_block_id_list[i];
+      int blkid = optimize_blocks[i]->ID();
       if (program->Block(blkid).Parent() != last_parent_blkid) {
         ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared,
                               program, recv_scope);
@@ -261,8 +260,11 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
   rpc_service_->RegisterRPC(distributed::kRequestPrefetch,
                             request_prefetch_handler_.get());
 
-  auto *optimize_block = Attr<framework::BlockDesc *>(kOptimizeBlock);
-  auto *program = optimize_block->Program();
+  auto optimize_blocks =
+      Attr<std::vector<framework::BlockDesc *>>(kOptimizeBlocks);
+  PADDLE_ENFORCE(optimize_blocks.size() >= 1,
+                 "optimize blocks should be 1 at least on the pserver side.");
+  auto *program = optimize_blocks[0]->Program();
   framework::Executor executor(dev_place);
 
   // prepare for prefetch
@@ -339,8 +341,9 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker {
         "a map from grad name to it's optimize block id")
         .SetDefault({});
     AddAttr<bool>("sync_mode", "if works at sync_mode or not").SetDefault(true);
-    AddAttr<framework::BlockDesc *>(kOptimizeBlock,
-                                    "BlockID to run on server side.");
+    AddAttr<std::vector<framework::BlockDesc *>>(
+        kOptimizeBlocks, "Optimize blocks to run on server side.")
+        .SetDefault({});
     AddAttr<std::vector<std::string>>(kPrefetchVarNameToBlockId,
                                       "prefetch blocks to run on server side.")
         .SetDefault({});

paddle/fluid/operators/listen_and_serv_op.h

@@ -30,7 +30,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-constexpr char kOptimizeBlock[] = "OptimizeBlock";
+constexpr char kOptimizeBlocks[] = "optimize_blocks";
 constexpr char kPrefetchVarNameToBlockId[] = "prefetch_var_name_to_block_id";
 
 void RunServer(std::shared_ptr<distributed::RPCServer> service);

paddle/fluid/operators/parallel_do_op.cc

@@ -295,7 +295,7 @@ class ParallelDoGradOp : public framework::OperatorBase {
 
         auto sum_op = framework::OpRegistry::CreateOp(
             "sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}},
-            framework::AttributeMap{});
+            framework::AttributeMap{{"use_mkldnn", {false}}});
         VLOG(10) << sum_op->DebugStringEx(sub_scopes[0]);
         sum_op->Run(*sub_scopes[0], places[0]);
         WaitOnPlace(places[0]);

paddle/fluid/operators/recurrent_op.cc

@@ -429,7 +429,8 @@ class RecurrentGradOp : public RecurrentBase {
 
           auto sum_op = framework::OpRegistry::CreateOp(
               "sum", {{"X", {pg_names[param_id], new_inside_name}}},
-              {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
+              {{"Out", {pg_names[param_id]}}},
+              framework::AttributeMap{{"use_mkldnn", {false}}});
           sum_op->Run(cur_scope, place);
 
           cur_scope.Rename(new_inside_name, inside_grad_name);

paddle/fluid/operators/send_recv_op_test.cc

@@ -129,7 +129,10 @@ void StartServerNet(bool is_sparse, std::atomic<bool> *initialized) {
   // sub program run in listen_and_serv_op, for simple test we use sum
   f::ProgramDesc program;
   const auto &root_block = program.Block(0);
+  std::vector<framework::BlockDesc *> optimize_blocks;
   auto *optimize_block = program.AppendBlock(root_block);
+  optimize_blocks.push_back(optimize_block);
+
   auto *prefetch_block = program.AppendBlock(root_block);
   // X for server side tensors, RX for received tensors, must be of same shape.
   AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block,
@@ -139,7 +142,7 @@ void StartServerNet(bool is_sparse, std::atomic<bool> *initialized) {
   attrs.insert({"Fanin", 1});
   attrs.insert({"ParamList", std::vector<std::string>({"Out"})});
   attrs.insert({"GradList", std::vector<std::string>({"x1"})});
-  attrs.insert({"OptimizeBlock", optimize_block});
+  attrs.insert({"optimize_blocks", optimize_blocks});
   attrs.insert({"PrefetchBlock", prefetch_block});
   attrs.insert({"grad_to_block_id", std::vector<std::string>({""})});
   attrs.insert({"sync_mode", true});

paddle/fluid/operators/softmax_mkldnn_op.cc

@@ -27,8 +27,81 @@ using paddle::platform::MKLDNNMemDesc;
 using mkldnn::memory;  // Note: paddle has also "memory" namespace
 using mkldnn::primitive;
 using mkldnn::softmax_forward;
+using mkldnn::softmax_backward;
 using mkldnn::prop_kind;
 using mkldnn::stream;
+using platform::to_void_cast;
+
+class SoftmaxMKLDNNHandler : public platform::MKLDNNHandler {
+ public:
+  SoftmaxMKLDNNHandler(
+      std::shared_ptr<mkldnn::softmax_forward::primitive_desc> softmax_pd,
+      const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
+      const std::string& base_key)
+      : platform::MKLDNNHandler(dev_ctx, engine, base_key),
+        softmax_pd_(softmax_pd) {}
+
+  SoftmaxMKLDNNHandler(
+      std::shared_ptr<mkldnn::softmax_forward::primitive_desc> softmax_pd,
+      std::shared_ptr<mkldnn::softmax_backward::primitive_desc> softmax_bwd_pd,
+      const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
+      const std::string& base_key)
+      : platform::MKLDNNHandler(dev_ctx, engine, base_key),
+        softmax_pd_(softmax_pd),
+        softmax_bwd_pd_(softmax_bwd_pd) {
+    // If we are in Grad operatgor then update a key with BWD suffix to
+    // distinguish from FWD memory primitives
+    key_ += "-BWD";
+  }
+
+  std::shared_ptr<mkldnn::softmax_forward> AcquireSoftmax(
+      std::shared_ptr<mkldnn::memory> dst_memory_p,
+      std::shared_ptr<mkldnn::memory> src_memory_p) {
+    /*Generate key*/
+    auto prim_key = key_ + "@softmax_p";
+
+    auto softmax_p = std::static_pointer_cast<mkldnn::softmax_forward>(
+        dev_ctx_.GetBlob(prim_key));
+    PADDLE_ENFORCE((softmax_p != nullptr) || (is_reusing_ == false),
+                   "Fail to find softmax primitive in device context");
+    if (softmax_p == nullptr) {
+      softmax_p = std::make_shared<mkldnn::softmax_forward>(
+          *(softmax_pd_.get()),
+          *(static_cast<mkldnn::memory*>(src_memory_p.get())),
+          *(static_cast<mkldnn::memory*>(dst_memory_p.get())));
+      dev_ctx_.SetBlob(prim_key, softmax_p);
+    } else {
+      is_reusing_ = true;
+    }
+
+    return softmax_p;
+  }
+
+  std::shared_ptr<mkldnn::softmax_backward> AcquireSoftmaxBackward(
+      std::shared_ptr<mkldnn::memory> dst_memory_p,
+      std::shared_ptr<mkldnn::memory> diff_dst_memory_p,
+      std::shared_ptr<mkldnn::memory> diff_src_memory_p) {
+    auto prim_key = key_ + "@softmax_bwd_p";
+    auto softmax_bwd_p = std::static_pointer_cast<mkldnn::softmax_backward>(
+        dev_ctx_.GetBlob(prim_key));
+    PADDLE_ENFORCE((softmax_bwd_p != nullptr) || (is_reusing_ == false),
+                   "Fail to find softmax backward primitive in device context");
+    if (softmax_bwd_p == nullptr) {
+      softmax_bwd_p = std::make_shared<mkldnn::softmax_backward>(
+          *softmax_bwd_pd_, *(dst_memory_p.get()), *(diff_dst_memory_p.get()),
+          *(diff_src_memory_p.get()));
+      dev_ctx_.SetBlob(prim_key, softmax_bwd_p);
+    } else {
+      is_reusing_ = true;
+    }
+
+    return softmax_bwd_p;
+  }
+
+ private:
+  std::shared_ptr<mkldnn::softmax_forward::primitive_desc> softmax_pd_;
+  std::shared_ptr<mkldnn::softmax_backward::primitive_desc> softmax_bwd_pd_;
+};
 
 template <typename T>
 class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
@@ -54,56 +127,27 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
     // Same memory descriptor to be used for input and output
     memory::dims softmax_tz = {src_tz[0], src_tz[1]};
     // Generate keys for storing/retriving primitives for this operator
-    // TODO(jczaja): Each MKLDNN operator may have diffrent hashing function
-    auto gethash = [](memory::dims& operand_dims) {
-      return std::string(std::to_string(operand_dims[0]) + "-" +
-                         std::to_string(operand_dims[1]));
-    };
-    const std::string key = gethash(softmax_tz);
-    const std::string key_softmax_p = key + "@softmax_p";
-    const std::string key_softmax_src_mem_p = key + "@softmax_src_mem_p";
-    const std::string key_softmax_dst_mem_p = key + "@softmax_dst_mem_p";
-
-    std::shared_ptr<void> softmax_p = dev_ctx.GetBlob(key_softmax_p);
-    if (softmax_p == nullptr) {
-      // Currently only NC data format is supported
-      auto softmax_md =
-          MKLDNNMemDesc({softmax_tz}, memory::f32, memory::format::nc);
-      // Normalization is made after innermost dimension eg. C out of NC
-      auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring,
-                                                softmax_md, 1 /*dim: C*/);
-      // create memory primitives
-      auto softmax_src_memory_p = std::make_shared<memory>(
-          memory::primitive_desc{softmax_md, mkldnn_engine},
-          static_cast<void*>(const_cast<T*>(input_data)));
-      dev_ctx.SetBlob(key_softmax_src_mem_p, softmax_src_memory_p);
-      auto softmax_dst_memory_p = std::make_shared<memory>(
-          memory::primitive_desc{softmax_md, mkldnn_engine},
-          static_cast<void*>(output_data));
-      dev_ctx.SetBlob(key_softmax_dst_mem_p, softmax_dst_memory_p);
-
-      auto softmax_forward_pd =
-          std::make_shared<softmax_forward::primitive_desc>(softmax_desc,
-                                                            mkldnn_engine);
-      softmax_p = std::make_shared<softmax_forward>(
-          *(softmax_forward_pd.get()),
-          *(static_cast<memory*>(softmax_src_memory_p.get())),
-          *(static_cast<memory*>(softmax_dst_memory_p.get())));
-      dev_ctx.SetBlob(key_softmax_p, softmax_p);
-    } else {
-      // Primitives already exist
-      auto src_memory_p = std::static_pointer_cast<memory>(
-          dev_ctx.GetBlob(key_softmax_src_mem_p));
-      PADDLE_ENFORCE(src_memory_p != nullptr,
-                     "Fail to find softmax src mem_p in device context");
-      auto dst_memory_p = std::static_pointer_cast<memory>(
-          dev_ctx.GetBlob(key_softmax_dst_mem_p));
-      PADDLE_ENFORCE(dst_memory_p != nullptr,
-                     "Fail to find softmax dst mem_p in device context");
-      src_memory_p->set_data_handle(
-          reinterpret_cast<void*>(const_cast<T*>(input_data)));
-      dst_memory_p->set_data_handle(output_data);
-    }
+    const std::string key =
+        platform::MKLDNNHandler::GetHash(softmax_tz, ctx.op().Output("Out"));
+    const std::string key_softmax_pd = key + "@softmax_pd";
+
+    // Currently only NC data format is supported
+    auto softmax_md = MKLDNNMemDesc(
+        {softmax_tz}, platform::MKLDNNGetDataType<T>(), memory::format::nc);
+    // Normalization is made after innermost dimension eg. C out of NC
+    auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring,
+                                              softmax_md, 1 /*dim: C*/);
+    auto softmax_pd = std::make_shared<mkldnn::softmax_forward::primitive_desc>(
+        softmax_desc, mkldnn_engine);
+    dev_ctx.SetBlob(key_softmax_pd, softmax_pd);
+
+    SoftmaxMKLDNNHandler handler(softmax_pd, dev_ctx, mkldnn_engine, key);
+    auto softmax_src_memory_p =
+        handler.AcquireSrcMemory(softmax_md, to_void_cast<T>(input_data));
+    auto softmax_dst_memory_p =
+        handler.AcquireDstMemory(softmax_md, to_void_cast<T>(output_data));
+    auto softmax_p =
+        handler.AcquireSoftmax(softmax_dst_memory_p, softmax_src_memory_p);
 
     std::vector<primitive> pipeline{
         *(static_cast<softmax_forward::primitive*>(softmax_p.get()))};
@@ -120,6 +164,77 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
   }
 };
 
+template <typename T>
+class SoftmaxMKLDNNGradKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
+                   "It must use CPUPlace.");
+
+    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+    auto mkldnn_engine = dev_ctx.GetEngine();
+    const Tensor* output = ctx.Input<Tensor>("Out");
+    const T* dst_data = output->data<T>();
+
+    auto* dout = ctx.template Input<Tensor>(framework::GradVarName("Out"));
+    const auto* diff_dst_ptr = dout->template data<T>();
+
+    auto* dx =
+        ctx.template Output<framework::Tensor>(framework::GradVarName("X"));
+    T* diff_src_ptr = dx->template mutable_data<T>(ctx.GetPlace());
+
+    std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
+    std::vector<int> src_tz(dst_tz);
+    PADDLE_ENFORCE(output->dims().size() == 2UL,
+                   "The input of softmax op must be a 2D matrix.");
+    // MKL-DNN does support softmax over selected axis. Having 2D Tensor,
+    // we will make normalization after final eg. axis: 1
+    PADDLE_ENFORCE(((src_tz[0] == dst_tz[0]) && (src_tz[1] == dst_tz[1])),
+                   "Softmax input and output dimensions should match");
+    // Same memory descriptor to be used for input and output
+    memory::dims softmax_tz = {src_tz[0], src_tz[1]};
+    // Currently only supports NC data format
+    // retrieve eltwise primitive desc from device context
+    const std::string key =
+        platform::MKLDNNHandler::GetHash(softmax_tz, ctx.op().Input("Out"));
+    const std::string key_softmax_pd = key + "@softmax_pd";
+
+    auto softmax_pd =
+        std::static_pointer_cast<mkldnn::softmax_forward::primitive_desc>(
+            dev_ctx.GetBlob(key_softmax_pd));
+    PADDLE_ENFORCE(softmax_pd != nullptr,
+                   "Fail to find softmax_pd in device context");
+
+    // TODO(jczaja): Add layouts support when there is a need to do so
+    // Two dimensional softmax does support NC format
+    auto data_softmax_md = MKLDNNMemDesc(
+        {softmax_tz}, platform::MKLDNNGetDataType<T>(), memory::format::nc);
+    auto diff_softmax_md = MKLDNNMemDesc(
+        {softmax_tz}, platform::MKLDNNGetDataType<T>(), memory::format::nc);
+    // Normalization is made after innermost dimension eg. C out of NC
+    auto softmax_bwd_desc =
+        softmax_backward::desc(diff_softmax_md, data_softmax_md, 1 /* dim: C*/);
+    auto softmax_bwd_pd =
+        std::make_shared<mkldnn::softmax_backward::primitive_desc>(
+            softmax_bwd_desc, mkldnn_engine, *softmax_pd);
+
+    SoftmaxMKLDNNHandler handler(softmax_pd, softmax_bwd_pd, dev_ctx,
+                                 mkldnn_engine, key);
+    auto dst_memory_p =
+        handler.AcquireDstMemory(data_softmax_md, to_void_cast<T>(dst_data));
+    auto diff_dst_memory_p = handler.AcquireDiffDstMemory(
+        diff_softmax_md, to_void_cast<T>(diff_dst_ptr));
+    auto diff_src_memory_p = handler.AcquireDiffSrcMemory(
+        diff_softmax_md, to_void_cast<T>(diff_src_ptr));
+
+    // Get primitve from device context
+    auto softmax_bwd_p = handler.AcquireSoftmaxBackward(
+        dst_memory_p, diff_dst_memory_p, diff_src_memory_p);
+
+    std::vector<primitive> pipeline{*softmax_bwd_p};
+    stream(stream::kind::eager).submit(pipeline).wait();
+  }
+};
 }  // namespace operators
 }  // namespace paddle
 
@@ -127,3 +242,5 @@ namespace ops = paddle::operators;
 
 REGISTER_OP_KERNEL(softmax, MKLDNN, ::paddle::platform::CPUPlace,
                    ops::SoftmaxMKLDNNKernel<float>);
+REGISTER_OP_KERNEL(softmax_grad, MKLDNN, ::paddle::platform::CPUPlace,
+                   ops::SoftmaxMKLDNNGradKernel<float>);

paddle/fluid/operators/softmax_op.cc

@@ -145,16 +145,30 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
       const framework::ExecutionContext& ctx) const override {
     // choose cudnn kernel if the runtime supported.
     framework::LibraryType library_{framework::LibraryType::kPlain};
+    std::string data_format = ctx.Attr<std::string>("data_format");
+    framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
 
 #ifdef PADDLE_WITH_CUDA
     if (platform::CanCUDNNBeUsed(ctx)) {
       library_ = framework::LibraryType::kCUDNN;
     }
 #endif
-    std::string data_format = ctx.Attr<std::string>("data_format");
-    return framework::OpKernelType(
-        framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
-        framework::StringToDataLayout(data_format), library_);
+#ifdef PADDLE_WITH_MKLDNN
+    if (library_ == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library_ = framework::LibraryType::kMKLDNN;
+      layout_ = framework::DataLayout::kMKLDNN;
+    }
+#endif
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("X")->type());
+    if (input_data_type == framework::proto::VarType::FP16) {
+      PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                     "float16 can only be used on GPU place");
+    }
+
+    return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_,
+                                   library_);
   }
 };
 

paddle/fluid/operators/sum_mkldnn_op.cc

+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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.
+
+/*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 "mkldnn.hpp"
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/operators/math/selected_rows_functor.h"
+#include "paddle/fluid/operators/sum_op.h"
+#include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/platform/mkldnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using paddle::framework::Tensor;
+using paddle::platform::MKLDNNDeviceContext;
+using paddle::platform::CPUDeviceContext;
+using framework::DataLayout;
+using mkldnn::memory;
+using mkldnn::primitive;
+using mkldnn::stream;
+using mkldnn::sum;
+using mkldnn::reorder;
+using platform::to_void_cast;
+
+template <typename T>
+class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
+                   "It must use CPUPlace.");
+    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+    const auto& mkldnn_engine = dev_ctx.GetEngine();
+    auto in_vars = ctx.MultiInputVar("X");
+
+    const int N = in_vars.size();
+    auto out_var = ctx.OutputVar("Out");
+    bool in_place = out_var == in_vars[0];
+
+    if (out_var->IsType<framework::LoDTensor>()) {
+      LoDTensor* output = ctx.Output<LoDTensor>("Out");
+      T* output_data = output->mutable_data<T>(ctx.GetPlace());
+
+      std::vector<int> dst_tz = framework::vectorize2int(output->dims());
+      auto src_tz = dst_tz;
+      memory::format output_format{memory::format::format_undef};
+      std::vector<float> scales;
+      std::vector<memory::primitive_desc> srcs_mpd;
+      std::vector<mkldnn::memory> srcs_mem;
+
+      PADDLE_ENFORCE(in_vars[0]->IsType<LoDTensor>(),
+                     "Input[0] must be LoDTensors");
+      auto& input0 = in_vars[0]->Get<LoDTensor>();
+      PADDLE_ENFORCE(input0.layout() == DataLayout::kMKLDNN &&
+                         input0.format() != memory::format::format_undef,
+                     "Wrong layout/format for inputs[0]");
+
+      memory::format input_format = input0.format();
+
+      if (src_tz.size() == 1 && (input_format == memory::format::nchw ||
+                                 input_format == memory::format::nhwc)) {
+        input_format = memory::format::x;
+      }
+      if (src_tz.size() == 2 && (input_format == memory::format::nchw ||
+                                 input_format == memory::format::nhwc)) {
+        input_format = memory::format::nc;
+      }
+
+      for (int i = in_place ? 1 : 0; i < N; i++) {
+        PADDLE_ENFORCE(in_vars[i]->IsType<LoDTensor>(),
+                       "all inputs must be all LoDTensors");
+        auto& input = in_vars[i]->Get<LoDTensor>();
+        PADDLE_ENFORCE(input.layout() == DataLayout::kMKLDNN &&
+                           input.format() != memory::format::format_undef,
+                       "Wrong layout/format for inputs");
+
+        if (input.numel() == 0) {
+          continue;
+        }
+
+        const T* input_data = input.data<T>();
+
+        auto src_md =
+            memory::desc(src_tz, memory::data_type::f32, input_format);
+        auto src_mpd = memory::primitive_desc(src_md, mkldnn_engine);
+        auto src_mem = memory(src_mpd, to_void_cast(input_data));
+        srcs_mpd.push_back(src_mpd);
+        srcs_mem.push_back(src_mem);
+        scales.push_back(1.0);
+      }
+
+      auto dst_md =
+          memory::desc(dst_tz, memory::data_type::f32, memory::format::any);
+
+      auto sum_pd = sum::primitive_desc(dst_md, scales, srcs_mpd);
+
+      std::shared_ptr<memory> dst_mem;
+      if (in_place) {
+        dst_mem.reset(new memory(sum_pd.dst_primitive_desc()));
+      } else {
+        dst_mem.reset(new memory(sum_pd.dst_primitive_desc(), output_data));
+      }
+      std::vector<mkldnn::primitive::at> inputs;
+      for (size_t i = 0; i < srcs_mem.size(); ++i) {
+        inputs.push_back(srcs_mem[i]);
+      }
+
+      auto sum_prim = mkldnn::sum(sum_pd, inputs, *dst_mem);
+      output_format = (memory::format)platform::GetMKLDNNFormat(sum_pd);
+
+      primitive reorder_prim;
+      std::shared_ptr<memory> target_mem;
+      if (in_place) {
+        output_format = input_format;
+        target_mem.reset(new memory(
+            {{{src_tz}, memory::data_type::f32, output_format}, mkldnn_engine},
+            output_data));
+        reorder_prim = reorder(*dst_mem, *target_mem);
+      }
+
+      std::vector<primitive> pipeline;
+      pipeline.push_back(sum_prim);
+      if (in_place) pipeline.push_back(reorder_prim);
+      stream(stream::kind::eager).submit(pipeline).wait();
+
+      output->set_layout(DataLayout::kMKLDNN);
+      output->set_format(output_format);
+    } else if (out_var->IsType<framework::SelectedRows>()) {
+      // TODO(@mozga-intel) Add MKLDNN SelectedRows support
+      std::unique_ptr<framework::SelectedRows> in0;
+      if (in_place) {
+        // If is in_place, we store the input[0] to in0
+        auto& in_sel0 = in_vars[0]->Get<SelectedRows>();
+        auto& rows = in_sel0.rows();
+        in0.reset(new framework::SelectedRows(rows, in_sel0.height()));
+        in0->mutable_value()->ShareDataWith(in_sel0.value());
+      }
+
+      auto get_selected_row = [&](size_t i) -> const SelectedRows& {
+        if (i == 0 && in0) {
+          return *in0.get();
+        } else {
+          return in_vars[i]->Get<SelectedRows>();
+        }
+      };
+      auto* out = ctx.Output<SelectedRows>("Out");
+      out->mutable_rows()->clear();
+      auto* out_value = out->mutable_value();
+
+      // Runtime InferShape
+      size_t first_dim = 0;
+      for (int i = 0; i < N; i++) {
+        auto& sel_row = get_selected_row(i);
+        first_dim += sel_row.rows().size();
+      }
+      auto in_dim =
+          framework::vectorize(get_selected_row(N - 1).value().dims());
+      in_dim[0] = static_cast<int64_t>(first_dim);
+
+      out_value->Resize(framework::make_ddim(in_dim));
+
+      // if all the input sparse vars are empty, no need to
+      // merge these vars.
+      if (first_dim == 0UL) {
+        return;
+      }
+      out_value->mutable_data<T>(ctx.GetPlace());
+      math::SelectedRowsAddTo<CPUDeviceContext, T> functor;
+      int64_t offset = 0;
+      for (int i = 0; i < N; i++) {
+        auto& sel_row = get_selected_row(i);
+        if (sel_row.rows().size() == 0) {
+          continue;
+        }
+        PADDLE_ENFORCE_EQ(out->height(), sel_row.height());
+        functor(ctx.template device_context<CPUDeviceContext>(), sel_row,
+                offset, out);
+        offset += sel_row.value().numel();
+      }
+    } else if (out_var->IsType<framework::LoDTensorArray>()) {
+      // TODO(@mozga-intel) Add MKLDNN LoDTensorArray support
+      auto& out_array = *out_var->GetMutable<framework::LoDTensorArray>();
+      for (size_t i = in_place ? 1 : 0; i < in_vars.size(); ++i) {
+        PADDLE_ENFORCE(in_vars[i]->IsType<framework::LoDTensorArray>(),
+                       "Only support all inputs are TensorArray");
+        auto& in_array = in_vars[i]->Get<framework::LoDTensorArray>();
+
+        for (size_t i = 0; i < in_array.size(); ++i) {
+          if (in_array[i].numel() != 0) {
+            if (i >= out_array.size()) {
+              out_array.resize(i + 1);
+            }
+            if (out_array[i].numel() == 0) {
+              framework::TensorCopy(in_array[i], in_array[i].place(),
+                                    ctx.device_context(), &out_array[i]);
+              out_array[i].set_lod(in_array[i].lod());
+            } else {
+              PADDLE_ENFORCE(out_array[i].lod() == in_array[i].lod());
+              auto in = EigenVector<T>::Flatten(in_array[i]);
+              auto result = EigenVector<T>::Flatten(out_array[i]);
+              result.device(*ctx.template device_context<MKLDNNDeviceContext>()
+                                 .eigen_device()) = result + in;
+            }
+          }
+        }
+      }
+    } else {
+      PADDLE_THROW("Unexpected branch, output variable type is %s",
+                   out_var->Type().name());
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_KERNEL(sum, MKLDNN, ::paddle::platform::CPUPlace,
+                   paddle::operators::SumMKLDNNOpKernel<float>);

paddle/fluid/operators/sum_op.cc

@@ -18,6 +18,10 @@ limitations under the License. */
 #include "paddle/fluid/framework/var_type_inference.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
 
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
 namespace paddle {
 namespace operators {
 using framework::Tensor;
@@ -63,6 +67,18 @@ class SumOp : public framework::OperatorWithKernel {
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     auto x_vars = ctx.MultiInputVar("X");
+
+    framework::LibraryType library{framework::LibraryType::kPlain};
+    framework::DataLayout layout{framework::DataLayout::kAnyLayout};
+
+#ifdef PADDLE_WITH_MKLDNN
+    if (library == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library = framework::LibraryType::kMKLDNN;
+      layout = framework::DataLayout::kMKLDNN;
+    }
+#endif
+
     if (x_vars[0]->IsType<framework::LoDTensor>()) {
       int dtype = -1;
       for (auto& x_var : x_vars) {
@@ -80,26 +96,27 @@ class SumOp : public framework::OperatorWithKernel {
                         "Sum operator should have at least one tensor");
 
       return framework::OpKernelType(
-          static_cast<framework::proto::VarType::Type>(dtype),
-          ctx.device_context());
+          static_cast<framework::proto::VarType::Type>(dtype), ctx.GetPlace(),
+          layout, library);
     } else if (x_vars[0]->IsType<framework::SelectedRows>()) {
       for (auto& var : x_vars) {
         auto& value = var->Get<framework::SelectedRows>().value();
         if (value.IsInitialized()) {
           return framework::OpKernelType(framework::ToDataType(value.type()),
-                                         ctx.device_context());
+                                         ctx.device_context(), layout, library);
         }
       }
       // if input sparse vars are not initialized, use an default kernel type.
       return framework::OpKernelType(framework::proto::VarType::FP32,
-                                     ctx.device_context());
+                                     ctx.device_context(), layout, library);
     } else if (x_vars[0]->IsType<framework::LoDTensorArray>()) {
       for (auto& x_var : x_vars) {
         auto& array = x_var->Get<framework::LoDTensorArray>();
         for (auto& each : array) {
           if (each.numel() != 0) {
             return framework::OpKernelType(framework::ToDataType(each.type()),
-                                           ctx.device_context());
+                                           ctx.device_context(), layout,
+                                           library);
           }
         }
       }
@@ -116,6 +133,9 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "(vector<Tensor>) The input tensors of sum operator.")
         .AsDuplicable();
     AddOutput("Out", "(Tensor) The output tensor of sum operator.").Reuse("X");
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Sum operator.
 
@@ -132,7 +152,6 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
                   framework::BlockDesc* block) const override {
     auto& inputs = op_desc.Input("X");
     auto var_type = framework::proto::VarType::SELECTED_ROWS;
-
     for (auto& name : op_desc.Input("X")) {
       VLOG(10) << name << " "
                << block->FindRecursiveOrCreateVar(name).GetType();
@@ -206,6 +225,7 @@ namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker,
                   ops::SumOpVarTypeInference);
+
 REGISTER_OP_CPU_KERNEL(
     sum, ops::SumKernel<paddle::platform::CPUDeviceContext, float>,
     ops::SumKernel<paddle::platform::CPUDeviceContext, double>,

paddle/fluid/operators/while_op.cc

@@ -203,11 +203,11 @@ class WhileGradOp : public framework::OperatorBase {
                 ->set_lod(inside_tensor.lod());
           }
         }
-
         auto new_inside_name = cur_scope.Rename(inside_grad_name);
         auto sum_op = framework::OpRegistry::CreateOp(
             "sum", {{"X", {pg_names[param_id], new_inside_name}}},
-            {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
+            {{"Out", {pg_names[param_id]}}},
+            framework::AttributeMap{{"use_mkldnn", {false}}});
         sum_op->Run(cur_scope, dev_place);
         cur_scope.Rename(new_inside_name, inside_grad_name);
       }

paddle/fluid/platform/dynload/CMakeLists.txt

 cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags enforce)
 
-list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc nccl.cc)
+list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc)
+
+# There is no macOS version of NCCL.
+if (NOT APPLE)
+  list(APPEND CUDA_SRCS nccl.cc)
+endif()
+
 if (TENSORRT_FOUND)
   list(APPEND CUDA_SRCS tensorrt.cc)
 endif()
 
-
 configure_file(cupti_lib_path.h.in ${CMAKE_CURRENT_BINARY_DIR}/cupti_lib_path.h)
 if (CUPTI_FOUND)
     list(APPEND CUDA_SRCS cupti.cc)

paddle/fluid/platform/enforce.h

@@ -44,8 +44,10 @@ limitations under the License. */
 #include "paddle/fluid/platform/dynload/cublas.h"
 #include "paddle/fluid/platform/dynload/cudnn.h"
 #include "paddle/fluid/platform/dynload/curand.h"
+#ifndef __APPLE__
 #include "paddle/fluid/platform/dynload/nccl.h"
-#endif
+#endif  // __APPLE__
+#endif  // PADDLE_WITH_CUDA
 
 namespace paddle {
 namespace platform {
@@ -174,6 +176,7 @@ inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
   throw std::runtime_error(err + string::Sprintf(args...));
 }
 
+#ifndef __APPLE__
 template <typename... Args>
 inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
     ncclResult_t stat, const Args&... args) {
@@ -184,7 +187,7 @@ inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
                              string::Sprintf(args...));
   }
 }
-
+#endif  // __APPLE__
 #endif  // PADDLE_WITH_CUDA
 
 template <typename T>

paddle/fluid/platform/mkldnn_helper.h

@@ -99,5 +99,143 @@ inline mkldnn::memory::format GetMKLDNNFormat(const mkldnn::memory memory) {
       memory.get_primitive_desc().desc().data.format);
 }
 
+inline mkldnn::memory::format GetMKLDNNFormat(
+    const mkldnn::sum::primitive_desc& memory) {
+  return static_cast<mkldnn::memory::format>(
+      memory.dst_primitive_desc().desc().data.format);
+}
+
+class MKLDNNHandler {
+ public:
+  MKLDNNHandler(const MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
+                const std::string& base_key)
+      : dev_ctx_(dev_ctx),
+        engine_(engine),
+        key_(base_key),
+        is_reusing_(false) {}
+
+  std::shared_ptr<mkldnn::memory> AcquireSrcMemory(
+      const mkldnn::memory::desc& md, void* ptr) {
+    return this->AcquireMemory(md, ptr, "@user_src_mem_p");
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireWeightsMemory(
+      const mkldnn::memory::desc& md, void* ptr) {
+    return this->AcquireMemory(md, ptr, "@user_weights_mem_p");
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireDstMemory(
+      const mkldnn::memory::desc& md, void* ptr) {
+    return this->AcquireMemory(md, ptr, "@user_dst_mem_p");
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireDiffDstMemory(
+      const mkldnn::memory::desc& md, void* ptr) {
+    return this->AcquireMemory(md, ptr, "@user_diff_dst_mem_p");
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireDiffSrcMemory(
+      const mkldnn::memory::desc& md, void* ptr) {
+    return this->AcquireMemory(md, ptr, "@user_diff_src_mem_p");
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireMemoryFromPrimitive(
+      mkldnn::memory::primitive_desc mdp, void* ptr,
+      const std::string& suffix) {
+    auto local_key = key_ + suffix;
+    auto mem_p =
+        std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
+    PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false),
+                   "Fail to find mem primitive in device context");
+    if (mem_p == nullptr) {
+      mem_p = std::make_shared<mkldnn::memory>(mdp, ptr);
+      dev_ctx_.SetBlob(local_key, mem_p);
+    } else {
+      mem_p->set_data_handle(ptr);
+      // Mark that reusing happenned. All primitives from operator instance
+      // should be reused or none of them. So we check consistency
+      is_reusing_ = true;
+    }
+    return mem_p;
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireMemory(const mkldnn::memory::desc& md,
+                                                void* ptr,
+                                                const std::string& suffix) {
+    /*Generate key*/
+    auto local_key = key_ + suffix;
+    auto mem_p =
+        std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
+    PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false),
+                   "Fail to find mem primitive in device context");
+    if (mem_p == nullptr) {
+      mem_p = std::make_shared<mkldnn::memory>(
+          mkldnn::memory::primitive_desc{md, engine_}, ptr);
+      dev_ctx_.SetBlob(local_key, mem_p);
+    } else {
+      mem_p->set_data_handle(ptr);
+      // Mark that reusing happenned. All primitives from operator instance
+      // should be reused or none of them. So we check consistency
+      is_reusing_ = true;
+    }
+    return mem_p;
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireMemory(
+      mkldnn::memory::primitive_desc& mpd,
+      mkldnn::memory::primitive_desc& user_mpd,
+      const std::shared_ptr<mkldnn::memory> user_memory_p,
+      const std::string& suffix, std::vector<mkldnn::primitive>& pipeline) {
+    // create reorder primitive if the input format is not the preferred one
+    auto local_key = key_ + suffix;
+    auto key_reorder_p = key_ + suffix + "reorder_p";
+
+    auto target_memory_p =
+        std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
+    PADDLE_ENFORCE((target_memory_p != nullptr) || (is_reusing_ == false),
+                   "Fail to find mem primitive in device context");
+    if (target_memory_p == nullptr) {
+      target_memory_p = user_memory_p;
+      std::shared_ptr<mkldnn::primitive> reorder_p;
+      if (mpd != user_mpd) {
+        target_memory_p = std::make_shared<mkldnn::memory>(mpd);
+
+        auto reorder_p =
+            std::make_shared<mkldnn::reorder>(*user_memory_p, *target_memory_p);
+        dev_ctx_.SetBlob(key_reorder_p, reorder_p);
+        pipeline.push_back(*reorder_p);
+      }
+      dev_ctx_.SetBlob(local_key, target_memory_p);
+    } else {
+      // Make reorder if needed
+      auto reorder_p = std::static_pointer_cast<mkldnn::reorder>(
+          dev_ctx_.GetBlob(key_reorder_p));
+      if (reorder_p != nullptr) {
+        pipeline.push_back(*reorder_p);
+      }
+      is_reusing_ = true;
+    }
+    return target_memory_p;
+  }
+
+  static std::string GetHash(mkldnn::memory::dims& operand_dims,
+                             const std::string& suffix) {
+    auto dims2str = [](const mkldnn::memory::dims& operand_dims) {
+      std::string dstr = "";
+      for (size_t i = 0; i < operand_dims.size(); ++i) {
+        dstr += std::to_string(operand_dims[i]) + "-";
+      }
+      return dstr;
+    };
+    return dims2str(operand_dims) + suffix;
+  };
+
+ protected:
+  const MKLDNNDeviceContext& dev_ctx_;
+  mkldnn::engine engine_;
+  std::string key_;
+  bool is_reusing_;
+};
+
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/pybind/protobuf.cc

@@ -268,7 +268,8 @@ void BindOpDesc(pybind11::module *m) {
       .value("STRINGS", pd::proto::AttrType::STRINGS)
       .value("BOOL", pd::proto::AttrType::BOOLEAN)
       .value("BOOLS", pd::proto::AttrType::BOOLEANS)
-      .value("BLOCK", pd::proto::AttrType::BLOCK);
+      .value("BLOCK", pd::proto::AttrType::BLOCK)
+      .value("BLOCKS", pd::proto::AttrType::BLOCKS);
 
   pybind11::class_<pd::OpDesc> op_desc(*m, "OpDesc", "");
   op_desc
@@ -293,6 +294,7 @@ void BindOpDesc(pybind11::module *m) {
       .def("set_attr", &pd::OpDesc::SetAttr)
       .def("attr", &pd::OpDesc::GetAttr)
       .def("set_block_attr", &pd::OpDesc::SetBlockAttr)
+      .def("set_blocks_attr", &pd::OpDesc::SetBlocksAttr)
       .def("set_serialized_attr",
            [](pd::OpDesc &self, const std::string &name,
               const pybind11::bytes &seriralized) {

paddle/fluid/pybind/pybind.cc

@@ -167,9 +167,6 @@ PYBIND11_PLUGIN(core) {
       .def("set_lod",
            [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
              // the input lod is offset-based level-of-detail info
-             LOG(WARNING)
-                 << "set_lod is deprecated and will be removed by 9.2018, "
-                    "please switch to set_recursive_sequence_lengths.";
              LoD new_lod;
              new_lod.reserve(lod.size());
              std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
@@ -196,8 +193,6 @@ PYBIND11_PLUGIN(core) {
       .def("lod",
            [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
              // output the offset-based lod info
-             LOG(WARNING) << "lod is deprecated and will be removed by 9.2018, "
-                             "please switch to recursive_sequence_lengths.";
              LoD lod = self.lod();
              std::vector<std::vector<size_t>> new_lod;
              new_lod.reserve(lod.size());

paddle/scripts/paddle_build.sh

@@ -22,7 +22,7 @@
 function print_usage() {
     echo -e "\n${RED}Usage${NONE}:
     ${BOLD}${SCRIPT_NAME}${NONE} [OPTION]"
-    
+
     echo -e "\n${RED}Options${NONE}:
     ${BLUE}build${NONE}: run build for x86 platform
     ${BLUE}build_android${NONE}: run build for android platform
@@ -133,7 +133,7 @@ EOF
         -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
         -DWITH_CONTRIB=${WITH_CONTRIB:-ON} \
-        -DWITH_ANAKIN=ON
+        -DWITH_ANAKIN=${WITH_ANAKIN:-ON}
 }
 
 function abort(){
@@ -198,7 +198,7 @@ function build_android() {
     fi
 
     ANDROID_STANDALONE_TOOLCHAIN=$ANDROID_TOOLCHAINS_DIR/$ANDROID_ARCH-android-$ANDROID_API
-    
+
     cat <<EOF
     ============================================
     Generating the standalone toolchain ...
@@ -212,13 +212,13 @@ EOF
           --arch=$ANDROID_ARCH \
           --platform=android-$ANDROID_API \
           --install-dir=$ANDROID_STANDALONE_TOOLCHAIN
-    
+
     BUILD_ROOT=${PADDLE_ROOT}/build_android
     DEST_ROOT=${PADDLE_ROOT}/install_android
-    
+
     mkdir -p $BUILD_ROOT
     cd $BUILD_ROOT
-    
+
     if [ $ANDROID_ABI == "armeabi-v7a" ]; then
       cmake -DCMAKE_SYSTEM_NAME=Android \
             -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \
@@ -286,7 +286,7 @@ function build_ios() {
           -DWITH_TESTING=OFF \
           -DWITH_SWIG_PY=OFF \
           -DCMAKE_BUILD_TYPE=Release
-    
+
     make -j 2
 }
 
@@ -331,14 +331,14 @@ EOF
 function bind_test() {
     # the number of process to run tests
     NUM_PROC=6
-    
+
     # calculate and set the memory usage for each process
     MEM_USAGE=$(printf "%.2f" `echo "scale=5; 1.0 / $NUM_PROC" | bc`)
     export FLAGS_fraction_of_gpu_memory_to_use=$MEM_USAGE
-    
+
     # get the CUDA device count
     CUDA_DEVICE_COUNT=$(nvidia-smi -L | wc -l)
-    
+
     for (( i = 0; i < $NUM_PROC; i++ )); do
         cuda_list=()
         for (( j = 0; j < $CUDA_DEVICE_COUNT; j++ )); do

python/paddle/fluid/backward.py

@@ -132,9 +132,9 @@ def _addup_repetitive_outputs_(op_descs):
     for idx, op_desc in enumerate(op_descs):
         for var_name in op_desc.input_arg_names():
             if len(renamed_vars[var_name]) > 1:
-                pending_sum_ops.append(
-                    (_create_op_desc_("sum", {"X": renamed_vars[var_name]},
-                                      {"Out": [var_name]}, {}), idx))
+                pending_sum_ops.append((_create_op_desc_(
+                    "sum", {"X": renamed_vars[var_name]}, {"Out": [var_name]},
+                    {"use_mkldnn": False}), idx))
                 renamed_vars[var_name] = [var_name]
         for var_name in op_desc.output_arg_names():
             if var_name == core.empty_var_name(
@@ -161,8 +161,9 @@ def _addup_repetitive_outputs_(op_descs):
                 renamed_vars[var_name].append(new_name)
     for var_name, inputs in renamed_vars.iteritems():
         if len(inputs) > 1:
-            pending_sum_ops.append((_create_op_desc_(
-                "sum", {"X": inputs}, {"Out": [var_name]}, {}), len(op_descs)))
+            pending_sum_ops.append(
+                (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]},
+                                  {"use_mkldnn": False}), len(op_descs)))
     # sum_op descs are sorted according to their insert position
     for p in reversed(pending_sum_ops):
         op_descs.insert(p[1], p[0])

python/paddle/fluid/framework.py

@@ -558,15 +558,20 @@ class Operator(object):
                 if (attr_name not in self.attrs) or (
                         self.attrs[attr_name] is None):
                     continue
-                if isinstance(self.attrs[attr_name], Block):
+                attr_val = self.attrs[attr_name]
+                if isinstance(attr_val, Block):
                     self.desc.set_block_attr(attr_name,
                                              self.attrs[attr_name].desc)
-                elif isinstance(self.attrs[attr_name], core.BlockDesc) or \
-                        isinstance(self.attrs[attr_name], core.ProgramDesc):
+                elif isinstance(attr_val, list) and attr_val and \
+                      all(isinstance(v, Block) for v in attr_val):
+                    self.desc.set_blocks_attr(attr_name,
+                                              [v.desc for v in attr_val])
+                elif isinstance(attr_val, core.BlockDesc) or \
+                        isinstance(attr_val, core.ProgramDesc):
                     self.desc.set_serialized_attr(
-                        attr_name, self.attrs[attr_name].serialize_to_string())
+                        attr_name, attr_val.serialize_to_string())
                 else:
-                    self.desc.set_attr(attr_name, self.attrs[attr_name])
+                    self.desc.set_attr(attr_name, attr_val)
         self.desc.check_attrs()
         if self.has_kernel(type):
             self.desc.infer_var_type(self.block.desc)
@@ -715,6 +720,9 @@ class Operator(object):
         self.attrs[name] = val
         if isinstance(val, Block):
             self.desc.set_block_attr(name, val.desc)
+        elif isinstance(val, list) and val and all(
+                isinstance(v, Block) for v in val):
+            self.desc.set_blocks_attr(name, [v.desc for v in val])
         elif isinstance(val, core.BlockDesc) or \
                 isinstance(val, core.ProgramDesc):
             self.desc.set_serialized_attr(name, val.serialize_to_string())
@@ -1387,7 +1395,11 @@ class Program(object):
         * Set for_test to True when we want to clone the program for testing.
 
         Notes: This API DOES NOT prune any operator. Use
-        :code:`clone(for_test=True)` before backward and optimization please.
+        :code:`clone(for_test=True)` before backward and optimization please. e.g.
+
+            >>> test_program = fluid.default_main_program().clone(for_test=True)
+            >>> optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
+            >>> optimizer.minimize()
 
         Args:
             for_test(bool): True if change the :code:`is_test` attribute of

python/paddle/fluid/layers/io.py

@@ -186,7 +186,6 @@ class ListenAndServ(object):
         main_program = self.helper.main_program
         current_block = main_program.current_block()
         parent_block = self.parent_block()
-        empty_block = Program().global_block()
 
         parent_block.append_op(
             type='listen_and_serv',
@@ -195,8 +194,9 @@ class ListenAndServ(object):
             attrs={
                 'endpoint': self.endpoint,
                 'Fanin': self.fan_in,
-                'OptimizeBlock': current_block,
-                'PrefetchBlock': empty_block,
+                'optimize_blocks': [
+                    current_block
+                ],  # did not support multiple optimize blocks in layers
                 'sync_mode': True,  # did not support async now in layers
                 'grad_to_block_id': [""]
             })

python/paddle/fluid/layers/nn.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-All layers just related to the neural network. 
+All layers just related to the neural network.
 """
 
 from ..layer_helper import LayerHelper
@@ -110,14 +110,14 @@ def fc(input,
     """
     **Fully Connected Layer**
 
-    This function creates a fully connected layer in the network. It can take 
-    multiple tensors as its inputs. It creates a variable called weights for 
-    each input tensor, which represents a fully connected weight matrix from 
-    each input unit to each output unit. The fully connected layer multiplies 
-    each input tensor with its coresponding weight to produce an output Tensor. 
-    If multiple input tensors are given, the results of multiple multiplications 
-    will be sumed up. If bias_attr is not None, a bias variable will be created 
-    and added to the output. Finally, if activation is not None, it will be applied 
+    This function creates a fully connected layer in the network. It can take
+    multiple tensors as its inputs. It creates a variable called weights for
+    each input tensor, which represents a fully connected weight matrix from
+    each input unit to each output unit. The fully connected layer multiplies
+    each input tensor with its coresponding weight to produce an output Tensor.
+    If multiple input tensors are given, the results of multiple multiplications
+    will be sumed up. If bias_attr is not None, a bias variable will be created
+    and added to the output. Finally, if activation is not None, it will be applied
     to the output as well.
 
     This process can be formulated as follows:
@@ -199,7 +199,10 @@ def fc(input,
     else:
         pre_bias = helper.create_tmp_variable(dtype)
         helper.append_op(
-            type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
+            type="sum",
+            inputs={"X": mul_results},
+            outputs={"Out": pre_bias},
+            attrs={"use_mkldnn": use_mkldnn})
     # add bias
     pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
     # add activation
@@ -1248,7 +1251,7 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True):
     """
     This function computes the softmax activation among all time-steps for each
     sequence. The dimension of each time-step should be 1. Thus, the shape of
-    input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N` 
+    input Tensor can be either :math:`[N, 1]` or :math:`[N]`, where :math:`N`
     is the sum of the length of all sequences.
 
     For i-th sequence in a mini-batch:
@@ -1829,11 +1832,11 @@ def pool2d(input,
     ${comment}
 
     Args:
-        input (Variable): The input tensor of pooling operator. The format of 
-                          input tensor is NCHW, where N is batch size, C is 
-                          the number of channels, H is the height of the 
+        input (Variable): The input tensor of pooling operator. The format of
+                          input tensor is NCHW, where N is batch size, C is
+                          the number of channels, H is the height of the
                           feature, and W is the width of the feature.
-        pool_size (int): The side length of pooling windows. All pooling 
+        pool_size (int): The side length of pooling windows. All pooling
                          windows are squares with pool_size on a side.
         pool_type: ${pooling_type_comment}
         pool_stride (int): stride of the pooling layer.
@@ -1842,7 +1845,7 @@ def pool2d(input,
         use_cudnn: ${use_cudnn_comment}
         ceil_mode: ${ceil_mode_comment}
         use_mkldnn: ${use_mkldnn_comment}
-        name (str|None): A name for this layer(optional). If set None, the 
+        name (str|None): A name for this layer(optional). If set None, the
                         layer will be named automatically.
 
     Returns:
@@ -1860,10 +1863,10 @@ def pool2d(input,
           data = fluid.layers.data(
               name='data', shape=[3, 32, 32], dtype='float32')
           conv2d = fluid.layers.pool2d(
-                            input=data, 
-                            pool_size=2, 
-                            pool_type='max', 
-                            pool_stride=1, 
+                            input=data,
+                            pool_size=2,
+                            pool_type='max',
+                            pool_stride=1,
                             global_pooling=False)
     """
     if pool_type not in ["max", "avg"]:
@@ -2228,14 +2231,14 @@ def beam_search_decode(ids, scores, name=None):
     This layers is to pack the output of beam search layer into sentences and
     associated scores. It is usually called after the beam search layer.
     Typically, the output of beam search layer is a tensor of selected ids, with
-    a tensor of the score of each id. Beam search layer's output ids, however, 
-    are generated directly during the tree search, and they are stacked by each 
-    level of the search tree. Thus we need to reorganize them into sentences, 
+    a tensor of the score of each id. Beam search layer's output ids, however,
+    are generated directly during the tree search, and they are stacked by each
+    level of the search tree. Thus we need to reorganize them into sentences,
     based on the score of each id. This layer takes the output of beam search
     layer as input and repack them into sentences.
 
     Args:
-        ids (Variable): The selected ids, output of beam search layer. 
+        ids (Variable): The selected ids, output of beam search layer.
         scores (Variable): The associated scores of the ids, out put of beam
             search layer.
         name (str): The name of this layer. It is optional.
@@ -2243,7 +2246,7 @@ def beam_search_decode(ids, scores, name=None):
     Returns:
         tuple(Variable): a tuple of two output tensors: sentence_ids, sentence_scores.
         sentence_ids is a tensor with shape [size, length], where size is the
-        beam size of beam search, and length is the length of each sentence. 
+        beam size of beam search, and length is the length of each sentence.
         Note that the length of sentences may vary.
         sentence_scores is a tensor with the same shape as sentence_ids.
 
@@ -2920,7 +2923,7 @@ def reduce_mean(input, dim=None, keep_dim=False, name=None):
             `None`, compute the mean over all elements of :attr:`input`
             and return a variable with a single element, otherwise it
             must be in the range :math:`[-rank(input), rank(input))`. If
-            :math:`dim[i] < 0`, the dimension to reduce is 
+            :math:`dim[i] < 0`, the dimension to reduce is
             :math:`rank(input) + dim[i]`.
         keep_dim (bool): Whether to reserve the reduced dimension in the
             output Tensor. The result tensor will have one fewer dimension
@@ -3391,16 +3394,16 @@ def topk(input, k, name=None):
     Args:
         input(Variable): The input variable which can be a vector or Tensor with
             higher rank.
-        k(int):  The number of top elements to look for along the last dimension 
+        k(int):  The number of top elements to look for along the last dimension
                  of input.
         name(str|None): A name for this layer(optional). If set None, the layer
-                       will be named automatically. 
+                       will be named automatically.
                        Default: None
 
     Returns:
-        Tuple[Variable]: A tuple with two elements. Each element is a Variable. 
-        The first one is k largest elements along each last 
-        dimensional slice. The second one is indices of values 
+        Tuple[Variable]: A tuple with two elements. Each element is a Variable.
+        The first one is k largest elements along each last
+        dimensional slice. The second one is indices of values
         within the last dimension of input.
 
     Raises:
@@ -3595,15 +3598,15 @@ def warpctc(input, label, blank=0, norm_by_times=False):
          It's shape is [Lp, num_classes + 1], where Lp is the sum of all input
          sequences' length and num_classes is the true number of classes.
          (not including the blank label).
-       label (Variable): The ground truth of variable-length sequence, 
+       label (Variable): The ground truth of variable-length sequence,
          which is a 2-D Tensor with LoD information. It is of the shape [Lg, 1],
          where Lg is th sum of all labels' length.
        blank (int, default 0): The blank label index of Connectionist
          Temporal Classification (CTC) loss, which is in the
          half-opened interval [0, num_classes + 1).
-       norm_by_times(bool, default false): Whether to normalize the gradients 
-         by the number of time-step, which is also the sequence's length. 
-         There is no need to normalize the gradients if warpctc layer was 
+       norm_by_times(bool, default false): Whether to normalize the gradients
+         by the number of time-step, which is also the sequence's length.
+         There is no need to normalize the gradients if warpctc layer was
          follewed by a mean_op.
 
     Returns:
@@ -3709,8 +3712,8 @@ def nce(input,
         input (Variable): input variable.
         label (Variable): label.
         num_total_classes (int):${num_total_classes_comment}
-        sample_weight (Variable|None): A Variable of shape [batch_size, 1] 
-            storing a weight for each sample. The default weight for each 
+        sample_weight (Variable|None): A Variable of shape [batch_size, 1]
+            storing a weight for each sample. The default weight for each
             sample is 1.0.
         param_attr (ParamAttr|None): attributes for parameter
         bias_attr (ParamAttr|None): attributes for bias
@@ -4100,7 +4103,7 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
     This layer computes the smooth L1 loss for Variable :attr:`x` and :attr:`y`.
     It takes the first dimension of :attr:`x` and :attr:`y` as batch size.
     For each instance, it computes the smooth L1 loss element by element first
-    and then sums all the losses. So the shape of ouput Variable is 
+    and then sums all the losses. So the shape of ouput Variable is
     [batch_size, 1].
 
     Args:
@@ -4109,14 +4112,14 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
         y (Variable): A tensor with rank at least 2. The target value of smooth
             L1 loss op with same shape as :attr:`x`.
         inside_weight (Variable|None):  A tensor with rank at least 2. This
-            input is optional and should have same shape with :attr:`x`. If 
-            provided, the result of (:attr:`x` - :attr:`y`) will be multiplied 
+            input is optional and should have same shape with :attr:`x`. If
+            provided, the result of (:attr:`x` - :attr:`y`) will be multiplied
             by this tensor element by element.
         outside_weight (Variable|None): A tensor with rank at least 2. This
-            input is optional and should have same shape with :attr:`x`. If 
-            provided, the out smooth L1 loss will be multiplied by this tensor 
+            input is optional and should have same shape with :attr:`x`. If
+            provided, the out smooth L1 loss will be multiplied by this tensor
             element by element.
-        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float 
+        sigma (float|None): Hyper parameter of smooth L1 loss layer. A float
            scalar with default value 1.0.
 
     Returns:
@@ -4316,10 +4319,10 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
 def lod_reset(x, y=None, target_lod=None):
     """
     Set LoD of :attr:`x` to a new one specified by :attr:`y` or
-    :attr:`target_lod`. When :attr:`y` provided, :attr:`y.lod` would be 
-    considered as target LoD first, otherwise :attr:`y.data` would be 
-    considered as target LoD. If :attr:`y` is not provided, target LoD should 
-    be specified by :attr:`target_lod`. If target LoD is specified by 
+    :attr:`target_lod`. When :attr:`y` provided, :attr:`y.lod` would be
+    considered as target LoD first, otherwise :attr:`y.data` would be
+    considered as target LoD. If :attr:`y` is not provided, target LoD should
+    be specified by :attr:`target_lod`. If target LoD is specified by
     :attr:`Y.data` or :attr:`target_lod`, only one level LoD is supported.
 
     .. code-block:: text
@@ -4373,7 +4376,7 @@ def lod_reset(x, y=None, target_lod=None):
 
     Args:
         x (Variable): Input variable which could be a Tensor or LodTensor.
-        y (Variable|None): If provided, output's LoD would be derived 
+        y (Variable|None): If provided, output's LoD would be derived
                            from :attr:`y`.
         target_lod (list|tuple|None): One level LoD which should be considered
                                       as target LoD when :attr:`y` not provided.
@@ -4689,7 +4692,7 @@ def image_resize(input,
     """
     **Resize a Batch of Images**
 
-    The input must be a tensor of the shape (num_batches, channels, in_h, in_w), 
+    The input must be a tensor of the shape (num_batches, channels, in_h, in_w),
     and the resizing only applies on the last two dimensions(hight and width).
 
     Supporting resample methods:
@@ -4785,9 +4788,9 @@ def resize_bilinear(input, out_shape=None, scale=None, name=None):
 
 def image_resize_short(input, out_short_len, resample='BILINEAR'):
     """
-    Resize a batch of images. The short edge of input images will be 
-    resized to the given 'out_short_len'. The long edge of input images 
-    will be resized proportionately to make images' length-width ratio 
+    Resize a batch of images. The short edge of input images will be
+    resized to the given 'out_short_len'. The long edge of input images
+    will be resized proportionately to make images' length-width ratio
     constant.
 
     Args:
@@ -4820,7 +4823,7 @@ def gather(input, index):
     """
     **Gather Layer**
 
-    Output is obtained by gathering entries of the outer-most dimension 
+    Output is obtained by gathering entries of the outer-most dimension
     of X indexed by `index` and concatenate them together.
 
     .. math::
@@ -4845,7 +4848,7 @@ def gather(input, index):
                        [5, 6]]
 
     Args:
-        input (Variable): The source input with rank>=1. 
+        input (Variable): The source input with rank>=1.
         index (Variable): The index input with rank=1.
 
     Returns:
@@ -4910,16 +4913,16 @@ def random_crop(x, shape, seed=None):
     return out
 
 
-def log(x):
+def log(input):
     """
     Calculates the natural log of the given input tensor, element-wise.
 
     .. math::
 
-        Out = \\ln(x)
+        Out = \\ln(input)
 
     Args:
-        x (Variable): Input tensor. 
+        input (Variable): Input tensor.
 
     Returns:
         Variable: The natural log of the input tensor computed element-wise.
@@ -4928,27 +4931,27 @@ def log(x):
 
         .. code-block:: python
 
-            output = fluid.layers.log(x)
+            output = fluid.layers.log(input)
     """
     helper = LayerHelper('log', **locals())
-    dtype = helper.input_dtype()
+    dtype = helper.input_dtype(input_param_name='x')
     out = helper.create_tmp_variable(dtype)
-    helper.append_op(type="log", inputs={"X": input}, outputs={"Out": out})
+    helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
     return out
 
 
-def relu(x):
+def relu(input):
     """
     Relu takes one input data (Tensor) and produces one output data (Tensor)
-    where the rectified linear function, y = max(0, x), is applied to
+    where the rectified linear function, y = max(0, input), is applied to
     the tensor elementwise.
 
     .. math::
 
-        Out = \\max(0, x)
+        Out = \\max(0, input)
 
     Args:
-        x (Variable): The input tensor. 
+        input (Variable): The input tensor.
 
     Returns:
         Variable: The output tensor with the same shape as input.
@@ -4957,27 +4960,27 @@ def relu(x):
 
         .. code-block:: python
 
-            output = fluid.layers.relu(x)
+            output = fluid.layers.relu(input)
     """
     helper = LayerHelper('relu', **locals())
-    dtype = helper.input_dtype()
+    dtype = helper.input_dtype(input_param_name='x')
     out = helper.create_tmp_variable(dtype)
-    helper.append_op(type="relu", inputs={"X": input}, outputs={"Out": out})
+    helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out})
     return out
 
 
 def mean_iou(input, label, num_classes):
     """
     Mean Intersection-Over-Union is a common evaluation metric for
-    semantic image segmentation, which first computes the IOU for each 
-    semantic class and then computes the average over classes. 
-    IOU is defined as follows: 
+    semantic image segmentation, which first computes the IOU for each
+    semantic class and then computes the average over classes.
+    IOU is defined as follows:
 
     .. math::
 
         IOU = \\frac{true\_positiv}{(true\_positive + false\_positive + false\_negative)}.
 
-    The predictions are accumulated in a confusion matrix and mean-IOU 
+    The predictions are accumulated in a confusion matrix and mean-IOU
     is then calculated from it.
 
 
@@ -4990,7 +4993,7 @@ def mean_iou(input, label, num_classes):
     Returns:
         mean_iou (Variable): A Tensor representing the mean intersection-over-union with shape [1].
         out_wrong(Variable): A Tensor with shape [num_classes]. The wrong numbers of each class.
-        out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class. 
+        out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class.
 
     Examples:
 

python/paddle/fluid/layers/tensor.py

@@ -230,7 +230,11 @@ def sums(input, out=None):
     helper = LayerHelper('sum', **locals())
     if out is None:
         out = helper.create_tmp_variable(dtype=helper.input_dtype())
-    helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out})
+    helper.append_op(
+        type='sum',
+        inputs={'X': input},
+        outputs={'Out': out},
+        attrs={'use_mkldnn': False})
     return out
 
 
@@ -382,7 +386,7 @@ def argmin(x, axis=0):
     """
     **argmin**
 
-    This function computes the indices of the min elements 
+    This function computes the indices of the min elements
     of the input tensor's element along the provided axis.
 
     Args:
@@ -397,7 +401,7 @@ def argmin(x, axis=0):
         .. code-block:: python
 
           out = fluid.layers.argmin(x=in, axis=0)
-          out = fluid.layers.argmin(x=in, axis=-1)  
+          out = fluid.layers.argmin(x=in, axis=-1)
     """
     helper = LayerHelper("arg_min", **locals())
     out = helper.create_tmp_variable(VarDesc.VarType.INT64)
@@ -413,7 +417,7 @@ def argmax(x, axis=0):
     """
     **argmax**
 
-    This function computes the indices of the max elements 
+    This function computes the indices of the max elements
     of the input tensor's element along the provided axis.
 
     Args:
@@ -428,7 +432,7 @@ def argmax(x, axis=0):
         .. code-block:: python
 
           out = fluid.layers.argmax(x=in, axis=0)
-          out = fluid.layers.argmax(x=in, axis=-1)  
+          out = fluid.layers.argmax(x=in, axis=-1)
     """
     helper = LayerHelper("arg_max", **locals())
     out = helper.create_tmp_variable(VarDesc.VarType.INT64)
@@ -497,9 +501,9 @@ def reverse(x, axis):
 
     Args:
         x(Vairbale): the input to be reversed.
-        axis(int|tuple|list): Axis that along which order of elements 
-                    is reversed. If it is a tuple or a list, reversing 
-                    will be apply on each axis in the tuple or list.  
+        axis(int|tuple|list): Axis that along which order of elements
+                    is reversed. If it is a tuple or a list, reversing
+                    will be apply on each axis in the tuple or list.
 
     Returns:
         Variable: The reversed tensor.
@@ -530,9 +534,9 @@ def save(x, file_path, overwrite=True):
     Args:
         x(variable): The Tensor/LoDTensor to be saved.
         file_path(str): The file path where the variable will be saved.
-        overwrite(bool): Whether or not cover the given file when it has already 
-            existed. If it's set 'False' and the file is existed, a runtime 
-            error will be thrown. 
+        overwrite(bool): Whether or not cover the given file when it has already
+            existed. If it's set 'False' and the file is existed, a runtime
+            error will be thrown.
     """
     helper = LayerHelper("save", **locals())
     helper.append_op(
@@ -552,8 +556,8 @@ def save_combine(x, file_path, overwrite=True):
                  a single file.
         file_path(str): The file path where variables will be saved.
         overwrite(bool): Whether or not cover the given file when it has already
-            existed. If it's set 'False' and the file is existed, a runtime 
-            error will be thrown. 
+            existed. If it's set 'False' and the file is existed, a runtime
+            error will be thrown.
 
     Returns:
         There is no return value.

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -15,7 +15,7 @@ if(NOT WITH_DISTRIBUTE)
 endif(NOT WITH_DISTRIBUTE)
 
 list(REMOVE_ITEM TEST_OPS test_seq_concat_op) # FIXME(helin): https://github.com/PaddlePaddle/Paddle/issues/8290
-list(REMOVE_ITEM TEST_OPS test_modified_huber_loss_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5184 
+list(REMOVE_ITEM TEST_OPS test_modified_huber_loss_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5184
 list(REMOVE_ITEM TEST_OPS test_lstm_unit_op) # # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/5185
 list(REMOVE_ITEM TEST_OPS test_nce) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/7778
 list(REMOVE_ITEM TEST_OPS test_recurrent_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/6152
@@ -43,8 +43,6 @@ list(REMOVE_ITEM TEST_OPS test_warpctc_op)
 list(REMOVE_ITEM TEST_OPS test_dist_train)
 list(REMOVE_ITEM TEST_OPS test_parallel_executor_crf)
 list(REMOVE_ITEM TEST_OPS test_parallel_executor_fetch_feed)
-# TODO(wuyi): this test hungs on CI, will add it back later
-list(REMOVE_ITEM TEST_OPS test_listen_and_serv_op)
 foreach(TEST_OP ${TEST_OPS})
     py_test_modules(${TEST_OP} MODULES ${TEST_OP})
 endforeach(TEST_OP)
@@ -52,3 +50,4 @@ py_test_modules(test_warpctc_op MODULES test_warpctc_op ENVS FLAGS_warpctc_dir=$
 py_test_modules(test_dist_train MODULES test_dist_train SERIAL)
 py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SERIAL)
 py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL)
+set_tests_properties(test_listen_and_serv_op PROPERTIES TIMEOUT 20)

python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py

@@ -94,7 +94,7 @@ class TestListenAndServOp(OpTest):
         self._wait_ps_ready(p1.pid)
 
         # raise SIGTERM to pserver
-        os.kill(p1.pid, signal.SIGKILL)
+        os.kill(p1.pid, signal.SIGINT)
         p1.join()
 
         # run pserver on CPU in async mode
@@ -102,7 +102,7 @@ class TestListenAndServOp(OpTest):
         self._wait_ps_ready(p2.pid)
 
         # raise SIGTERM to pserver
-        os.kill(p2.pid, signal.SIGKILL)
+        os.kill(p2.pid, signal.SIGTERM)
         p2.join()
 
 

python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+import unittest
+
+from test_sum_op import TestSumOp
+
+
+class TestMKLDNN(TestSumOp):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_sum_op.py

@@ -20,12 +20,15 @@ from op_test import OpTest
 class TestSumOp(OpTest):
     def setUp(self):
         self.op_type = "sum"
+        self.use_mkldnn = False
+        self.init_kernel_type()
         x0 = np.random.random((3, 4)).astype('float32')
         x1 = np.random.random((3, 4)).astype('float32')
         x2 = np.random.random((3, 4)).astype('float32')
         self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]}
         y = x0 + x1 + x2
         self.outputs = {'Out': y}
+        self.attrs = {'use_mkldnn': self.use_mkldnn}
 
     def test_check_output(self):
         self.check_output()
@@ -33,6 +36,9 @@ class TestSumOp(OpTest):
     def test_check_grad(self):
         self.check_grad(['x0'], 'Out')
 
+    def init_kernel_type(self):
+        pass
+
 
 if __name__ == "__main__":
     unittest.main()

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -396,7 +396,7 @@ class DistributeTranspiler(object):
                     return varname
             return ""
 
-        def __clone_lr_op_sub_block__(op, program, new_block):
+        def __clone_lr_op_sub_block__(op, program, lr_block):
             if not op.has_attr('sub_block'):
                 return
 
@@ -405,36 +405,41 @@ class DistributeTranspiler(object):
             assert isinstance(origin_block, Block)
             # we put the new sub block to new block to follow the block
             # hierarchy of the original blocks
-            new_sub_block = program.create_block(new_block.idx)
+            new_sub_block = program.create_block(lr_block.idx)
 
             # clone vars
             for var in origin_block.vars:
                 new_sub_block.clone_variable(var)
 
             # clone ops
-            for op in origin_block.ops:
-                self._clone_lr_op(program, new_sub_block, op)
+            for origin_op in origin_block.ops:
+                cloned_op = self._clone_lr_op(program, new_sub_block, origin_op)
                 # clone sub_block of op
-                __clone_lr_op_sub_block__(op, program, new_sub_block)
+                __clone_lr_op_sub_block__(cloned_op, program, new_sub_block)
 
             # reset the block of op
             op.set_attr('sub_block', new_sub_block)
 
         # append lr decay ops to the child block if exists
         lr_ops = self._get_lr_ops()
+        # record optimize blocks and we can run them on pserver parallel
+        optimize_blocks = []
         if len(lr_ops) > 0:
             lr_decay_block = pserver_program.create_block(
                 pserver_program.num_blocks - 1)
+            optimize_blocks.append(lr_decay_block)
             for _, op in enumerate(lr_ops):
-                self._append_pserver_non_opt_ops(lr_decay_block, op)
+                cloned_op = self._append_pserver_non_opt_ops(lr_decay_block, op)
                 # append sub blocks to pserver_program in lr_decay_op
-                __clone_lr_op_sub_block__(op, pserver_program, lr_decay_block)
+                __clone_lr_op_sub_block__(cloned_op, pserver_program,
+                                          lr_decay_block)
 
         # append op to the current block
         grad_to_block_id = []
         pre_block_idx = pserver_program.num_blocks - 1
         for idx, opt_op in enumerate(opt_op_on_pserver):
             per_opt_block = pserver_program.create_block(pre_block_idx)
+            optimize_blocks.append(per_opt_block)
             # append grad merging ops before clip and weight decay
             for _, op in enumerate(self.optimize_ops):
                 # find the origin @GRAD var before clipping
@@ -453,6 +458,7 @@ class DistributeTranspiler(object):
         if global_ops:
             opt_state_block = pserver_program.create_block(
                 pserver_program.num_blocks - 1)
+            optimize_blocks.append(opt_state_block)
             for glb_op in global_ops:
                 __append_optimize_op__(glb_op, opt_state_block,
                                        grad_to_block_id, None)
@@ -474,11 +480,11 @@ class DistributeTranspiler(object):
             assert len(prefetch_var_name_to_block_id) == 0
 
         attrs = {
-            "OptimizeBlock": pserver_program.block(1),
+            "optimize_blocks": optimize_blocks,
             "endpoint": endpoint,
             "Fanin": self.trainer_num,
             "sync_mode": self.sync_mode,
-            "grad_to_block_id": grad_to_block_id
+            "grad_to_block_id": grad_to_block_id,
         }
         if len(prefetch_var_name_to_block_id) > 0:
             attrs['prefetch_var_name_to_block_id'] \
@@ -872,7 +878,8 @@ class DistributeTranspiler(object):
             table_opt_block.append_op(
                 type="sum",
                 inputs={"X": pserver_side_table_grad_list},
-                outputs={"Out": [grad_var]})
+                outputs={"Out": [grad_var]},
+                attrs={"use_mkldnn": False})
         else:
             # in async_mode, for table gradient, it also need to be splited to each parameter server
             origin_grad_name = grad_var.name
@@ -1104,7 +1111,8 @@ class DistributeTranspiler(object):
             optimize_block.append_op(
                 type="sum",
                 inputs={"X": vars2merge},
-                outputs={"Out": merged_var})
+                outputs={"Out": merged_var},
+                attrs={"use_mkldnn": False})
             # TODO(panyx0718): What if it's SELECTED_ROWS.
             if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS:
                 optimize_block.append_op(
@@ -1209,7 +1217,7 @@ class DistributeTranspiler(object):
                 if var not in program.global_block().vars:
                     block.clone_variable(var)
 
-        block.append_op(
+        return block.append_op(
             type=op.type, inputs=inputs, outputs=outputs, attrs=op.attrs)
 
     def _append_pserver_non_opt_ops(self, optimize_block, opt_op):
@@ -1247,7 +1255,7 @@ class DistributeTranspiler(object):
                 elif not program.global_block().vars.has_key(var.name):
                     program.global_block().clone_variable(var)
 
-        optimize_block.append_op(
+        return optimize_block.append_op(
             type=opt_op.type,
             inputs=inputs,
             outputs=outputs,

python/paddle/reader/decorator.py

@@ -336,7 +336,7 @@ def _buf2lines(buf, line_break="\n"):
 
 class PipeReader:
     """
-        PipeReader read data by stream from a command, take it's 
+        PipeReader read data by stream from a command, take it's
         stdout into a pipe buffer and redirect it to the parser to
         parse, then yield data as your desired format.
 
@@ -352,7 +352,7 @@ class PipeReader:
         An example:
 
         .. code-block:: python
-    
+
            def example_reader():
                for f in myfiles:
                    pr = PipeReader("cat %s"%f)

python/paddle/v2/dataset/cifar.py

@@ -43,7 +43,7 @@ CIFAR100_URL = URL_PREFIX + 'cifar-100-python.tar.gz'
 CIFAR100_MD5 = 'eb9058c3a382ffc7106e4002c42a8d85'
 
 
-def reader_creator(filename, sub_name):
+def reader_creator(filename, sub_name, cycle=False):
     def read_batch(batch):
         data = batch['data']
         labels = batch.get('labels', batch.get('fine_labels', None))
@@ -56,10 +56,13 @@ def reader_creator(filename, sub_name):
             names = (each_item.name for each_item in f
                      if sub_name in each_item.name)
 
-            for name in names:
-                batch = cPickle.load(f.extractfile(name))
-                for item in read_batch(batch):
-                    yield item
+            while True:
+                for name in names:
+                    batch = cPickle.load(f.extractfile(name))
+                    for item in read_batch(batch):
+                        yield item
+                if not cycle:
+                    break
 
     return reader
 
@@ -94,34 +97,40 @@ def test100():
         'test')
 
 
-def train10():
+def train10(cycle=False):
     """
     CIFAR-10 training set creator.
 
     It returns a reader creator, each sample in the reader is image pixels in
     [0, 1] and label in [0, 9].
 
+    :param cycle: whether to cycle through the dataset
+    :type cycle: bool
     :return: Training reader creator
     :rtype: callable
     """
     return reader_creator(
         paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5),
-        'data_batch')
+        'data_batch',
+        cycle=cycle)
 
 
-def test10():
+def test10(cycle=False):
     """
     CIFAR-10 test set creator.
 
     It returns a reader creator, each sample in the reader is image pixels in
     [0, 1] and label in [0, 9].
 
+    :param cycle: whether to cycle through the dataset
+    :type cycle: bool
     :return: Test reader creator.
     :rtype: callable
     """
     return reader_creator(
         paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5),
-        'test_batch')
+        'test_batch',
+        cycle=cycle)
 
 
 def fetch():

python/paddle/v2/dataset/flowers.py

@@ -76,7 +76,8 @@ def reader_creator(data_file,
                    dataset_name,
                    mapper,
                    buffered_size=1024,
-                   use_xmap=True):
+                   use_xmap=True,
+                   cycle=False):
     '''
     1. read images from tar file and
         merge images into batch files in 102flowers.tgz_batch/
@@ -96,6 +97,8 @@ def reader_creator(data_file,
     :type mapper: callable
     :param buffered_size: the size of buffer used to process images
     :type buffered_size: int
+    :param cycle: whether to cycle through the dataset
+    :type cycle: bool
     :return: data reader
     :rtype: callable
     '''
@@ -108,15 +111,18 @@ def reader_creator(data_file,
     file_list = batch_images_from_tar(data_file, dataset_name, img2label)
 
     def reader():
-        for file in open(file_list):
-            file = file.strip()
-            batch = None
-            with open(file, 'r') as f:
-                batch = cPickle.load(f)
-            data = batch['data']
-            labels = batch['label']
-            for sample, label in itertools.izip(data, batch['label']):
-                yield sample, int(label) - 1
+        while True:
+            for file in open(file_list):
+                file = file.strip()
+                batch = None
+                with open(file, 'r') as f:
+                    batch = cPickle.load(f)
+                data = batch['data']
+                labels = batch['label']
+                for sample, label in itertools.izip(data, batch['label']):
+                    yield sample, int(label) - 1
+            if not cycle:
+                break
 
     if use_xmap:
         cpu_num = int(os.environ.get('CPU_NUM', cpu_count()))
@@ -125,7 +131,7 @@ def reader_creator(data_file,
         return map_readers(mapper, reader)
 
 
-def train(mapper=train_mapper, buffered_size=1024, use_xmap=True):
+def train(mapper=train_mapper, buffered_size=1024, use_xmap=True, cycle=False):
     '''
     Create flowers training set reader.
     It returns a reader, each sample in the reader is
@@ -138,17 +144,23 @@ def train(mapper=train_mapper, buffered_size=1024, use_xmap=True):
     :type mapper: callable
     :param buffered_size: the size of buffer used to process images
     :type buffered_size: int
+    :param cycle: whether to cycle through the dataset
+    :type cycle: bool
     :return: train data reader
     :rtype: callable
     '''
     return reader_creator(
         download(DATA_URL, 'flowers', DATA_MD5),
         download(LABEL_URL, 'flowers', LABEL_MD5),
-        download(SETID_URL, 'flowers', SETID_MD5), TRAIN_FLAG, mapper,
-        buffered_size, use_xmap)
+        download(SETID_URL, 'flowers', SETID_MD5),
+        TRAIN_FLAG,
+        mapper,
+        buffered_size,
+        use_xmap,
+        cycle=cycle)
 
 
-def test(mapper=test_mapper, buffered_size=1024, use_xmap=True):
+def test(mapper=test_mapper, buffered_size=1024, use_xmap=True, cycle=False):
     '''
     Create flowers test set reader.
     It returns a reader, each sample in the reader is
@@ -161,14 +173,20 @@ def test(mapper=test_mapper, buffered_size=1024, use_xmap=True):
     :type mapper: callable
     :param buffered_size: the size of buffer used to process images
     :type buffered_size: int
+    :param cycle: whether to cycle through the dataset
+    :type cycle: bool
     :return: test data reader
     :rtype: callable
     '''
     return reader_creator(
         download(DATA_URL, 'flowers', DATA_MD5),
         download(LABEL_URL, 'flowers', LABEL_MD5),
-        download(SETID_URL, 'flowers', SETID_MD5), TEST_FLAG, mapper,
-        buffered_size, use_xmap)
+        download(SETID_URL, 'flowers', SETID_MD5),
+        TEST_FLAG,
+        mapper,
+        buffered_size,
+        use_xmap,
+        cycle=cycle)
 
 
 def valid(mapper=test_mapper, buffered_size=1024, use_xmap=True):