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

cmake/operators.cmake

@@ -166,6 +166,8 @@ function(op_library TARGET)
       # Append first implemented MKLDNN activation operator
       if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
+      elseif(${MKLDNN_FILE} STREQUAL "conv_mkldnn_op")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, FP32);\n")
       else()
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
       endif()

paddle/fluid/API.spec

@@ -194,6 +194,8 @@ paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=Non
 paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
 paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None))
+paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))

paddle/fluid/framework/CMakeLists.txt

@@ -118,8 +118,9 @@ cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 
 cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto device_context)
+cc_library(op_kernel_type SRCS op_kernel_type.cc DEPS device_context place)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
-    shape_inference data_transform lod_tensor profiler transfer_scope_cache)
+    shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type)
 
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
 
@@ -191,7 +192,7 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
-cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
+cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto op_kernel_type)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
 
 cc_test(tuple_test SRCS tuple_test.cc )

paddle/fluid/framework/data_feed.cc

@@ -33,11 +33,7 @@ void DataFeed::AddFeedVar(Variable* var, const std::string& name) {
   CheckInit();
   for (size_t i = 0; i < use_slots_.size(); ++i) {
     if (name == use_slots_[i]) {
-      if (use_slots_is_dense_[i]) {
-        feed_vec_[i] = MixTensor(var->GetMutable<Tensor>());
-      } else {
-        feed_vec_[i] = MixTensor(var->GetMutable<LoDTensor>());
-      }
+      feed_vec_[i] = var->GetMutable<LoDTensor>();
     }
   }
 }
@@ -301,6 +297,7 @@ bool MultiSlotDataFeed::ParseOneInstance(std::vector<MultiSlotType>* instance) {
           "the data, please check if the data contains unresolvable "
           "characters.\nplease check this error line: %s",
           str);
+
       if (idx != -1) {
         (*instance)[idx].Init(all_slots_type_[i]);
         if ((*instance)[idx].GetType()[0] == 'f') {  // float
@@ -337,6 +334,7 @@ void MultiSlotDataFeed::AddInstanceToInsVec(
       (*ins_vec)[i].InitOffset();
     }
   }
+
   for (size_t i = 0; i < instance.size(); ++i) {
     (*ins_vec)[i].AddIns(instance[i]);
   }
@@ -348,36 +346,25 @@ void MultiSlotDataFeed::PutToFeedVec(
     const auto& type = ins_vec[i].GetType();
     const auto& offset = ins_vec[i].GetOffset();
     int total_instance = static_cast<int>(offset.back());
+
     if (type[0] == 'f') {  // float
       const auto& feasign = ins_vec[i].GetFloatData();
-      if (feed_vec_[i].IsDense()) {
-        int size_in_each_batch = total_instance / batch_size_;
-        float* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data<float>(
-            {batch_size_, size_in_each_batch}, platform::CPUPlace());
-        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
-      } else {
-        float* tensor_ptr = feed_vec_[i].GetLoDTensor()->mutable_data<float>(
+      float* tensor_ptr = feed_vec_[i]->mutable_data<float>(
           {total_instance, 1}, platform::CPUPlace());
       memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
-        LoD data_lod{offset};
-        feed_vec_[i].GetLoDTensor()->set_lod(data_lod);
-      }
     } else if (type[0] == 'u') {  // uint64
       // no uint64_t type in paddlepaddle
       const auto& feasign = ins_vec[i].GetUint64Data();
-      if (feed_vec_[i].IsDense()) {
-        int size_in_each_batch = total_instance / batch_size_;
-        int64_t* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data<int64_t>(
-            {batch_size_, size_in_each_batch}, platform::CPUPlace());
-        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
-      } else {
-        int64_t* tensor_ptr =
-            feed_vec_[i].GetLoDTensor()->mutable_data<int64_t>(
+      int64_t* tensor_ptr = feed_vec_[i]->mutable_data<int64_t>(
           {total_instance, 1}, platform::CPUPlace());
       memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
-        LoD data_lod{offset};
-        feed_vec_[i].GetLoDTensor()->set_lod(data_lod);
     }
+
+    LoD data_lod{offset};
+    feed_vec_[i]->set_lod(data_lod);
+    if (use_slots_is_dense_[i]) {
+      int dim = total_instance / batch_size_;
+      feed_vec_[i]->Resize({batch_size_, dim});
     }
   }
 }

paddle/fluid/framework/data_feed.h

@@ -30,35 +30,6 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-// Pack Tensor type and LoDTensor type into MixTensor type, in order
-// to record either Tensor or LoDTensor information at the same time.
-class MixTensor {
- public:
-  MixTensor() {}
-  explicit MixTensor(LoDTensor* lodtensor) {
-    is_dense_ = false;
-    lodtensor_ = lodtensor;
-  }
-  explicit MixTensor(Tensor* tensor) {
-    is_dense_ = true;
-    tensor_ = tensor;
-  }
-  bool IsDense() { return is_dense_; }
-  LoDTensor* GetLoDTensor() {
-    PADDLE_ENFORCE(!is_dense_, "Let a dense var return a LoDTensor ptr.");
-    return lodtensor_;
-  }
-  Tensor* GetTensor() {
-    PADDLE_ENFORCE(is_dense_, "Let a sparse var return a Tensor ptr.");
-    return tensor_;
-  }
-
- private:
-  bool is_dense_;
-  LoDTensor* lodtensor_;
-  Tensor* tensor_;
-};
-
 // DataFeed is the base virtual class for all ohther DataFeeds.
 // It is used to read files and parse the data for subsequent trainer.
 // Example:
@@ -133,7 +104,7 @@ class DataFeed {
       use_slots_index_;  // -1: not used; >=0: the index of use_slots_
 
   // The data read by DataFeed will be stored here
-  std::vector<MixTensor> feed_vec_;
+  std::vector<LoDTensor*> feed_vec_;
 
   // the batch size defined by user
   int default_batch_size_;

paddle/fluid/framework/data_feed_test.cc

@@ -152,21 +152,15 @@ void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
       const auto& multi_slot_desc = data_feed_desc.multi_slot_desc();
       std::map<std::string, const paddle::framework::LoDTensor*>
           lodtensor_targets;
-      std::map<std::string, const paddle::framework::Tensor*> tensor_targets;
       for (int i = 0; i < multi_slot_desc.slots_size(); ++i) {
         const auto& slot = multi_slot_desc.slots(i);
         if (slot.is_used()) {
           const auto& name = slot.name();
           readers[idx]->AddFeedVar(scope->Var(name), name);
-          if (slot.is_dense()) {
-            tensor_targets[name] =
-                &scope->FindVar(name)->Get<paddle::framework::Tensor>();
-          } else {
           lodtensor_targets[name] =
               &scope->FindVar(name)->Get<paddle::framework::LoDTensor>();
         }
       }
-      }
       readers[idx]->Start();
       while (readers[idx]->Next()) {
         int index = 0;
@@ -175,8 +169,9 @@ void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
           if (!slot.is_used()) {
             continue;
           }
+          const paddle::framework::LoDTensor* tens =
+              lodtensor_targets[slot.name()];
           if (slot.is_dense()) {  // dense branch
-            const paddle::framework::Tensor* tens = tensor_targets[slot.name()];
             if (slot.type() == "uint64") {
               const int64_t* data = tens->data<int64_t>();
               int batch_size = tens->dims()[0];
@@ -202,8 +197,6 @@ void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
               PADDLE_THROW("Error type in proto file.");
             }
           } else {  // sparse branch
-            const paddle::framework::LoDTensor* tens =
-                lodtensor_targets[slot.name()];
             if (slot.type() == "uint64") {
               const int64_t* data = tens->data<int64_t>();
               for (size_t i = 0; i < tens->NumElements(); ++i) {

paddle/fluid/framework/details/CMakeLists.txt

@@ -15,14 +15,26 @@ cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_ro
 if(WITH_GPU)
     nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
             dynload_cuda variable_visitor)
-    nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda)
+    if(WITH_DISTRIBUTE)
+        nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
+            ddim dynload_cuda selected_rows_functor sendrecvop_grpc)
+    else()
+        nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
+            ddim dynload_cuda selected_rows_functor)
+    endif()
     nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda)
     nv_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
 
 else()
     cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
              variable_visitor)
-    cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim)
+    if(WITH_DISTRIBUTE)
+        cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
+            ddim selected_rows_functor sendrecvop_grpc)
+    else()
+        cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
+            ddim selected_rows_functor)
+    endif()
     cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
     cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
 endif()

paddle/fluid/framework/details/build_strategy.cc

@@ -58,6 +58,17 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
       }
     }
 
+    CollectiveContext *context = CollectiveContext::GetInstance();
+    context->endpoints_ = strategy_.trainers_endpoints_;
+    context->trainer_id_ = strategy_.trainer_id_;
+    PADDLE_ENFORCE(strategy_.trainer_id_ >= 0, "trainer_id_ >= 0");
+    if (strategy_.trainer_id_ > 0) {
+      PADDLE_ENFORCE((unsigned)(strategy_.trainer_id_) <
+                         strategy_.trainers_endpoints_.size(),
+                     "trainer_id_ < endpoints_ size");
+    }
+    VLOG(1) << "CollectiveContext:" << context->String();
+
     // Convert graph to run on multi-devices.
     auto multi_devices_pass = AppendPass("multi_devices_pass");
     multi_devices_pass->SetNotOwned<const BuildStrategy>("strategy",
@@ -135,7 +146,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
       pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
 #endif
     } else if (pass->Type() == "sequential_execution_pass") {
-      VLOG(1) << "set enable_sequential_execution:"
+      LOG(INFO) << "set enable_sequential_execution:"
                 << enable_sequential_execution_;
 
       pass->Erase(kAllOpDescs);
@@ -143,7 +154,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
           kAllOpDescs,
           new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
     } else if (pass->Type() == "all_reduce_deps_pass") {
-      VLOG(1) << "SeqOnlyAllReduceOps:" << SeqOnlyAllReduceOps(*this)
+      LOG(INFO) << "SeqOnlyAllReduceOps:" << SeqOnlyAllReduceOps(*this)
                 << ", num_trainers:" << num_trainers_;
 
       pass->Erase(kAllOpDescs);

paddle/fluid/framework/details/build_strategy.h

@@ -74,6 +74,8 @@ struct BuildStrategy {
   bool fuse_broadcast_op_{false};
 
   int num_trainers_{1};
+  int trainer_id_{0};
+  std::vector<std::string> trainers_endpoints_;
   bool remove_unnecessary_lock_{false};
 
   // NOTE:

paddle/fluid/framework/details/reduce_and_gather.h

@@ -53,7 +53,7 @@ struct ReduceLoDTensor {
   }
 };
 
-inline void GatherSelectedRows(
+inline void GatherLocalSelectedRows(
     const std::vector<const SelectedRows *> &src_selecte_rows_,
     const std::vector<platform::Place> &in_places,
     const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes,

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -16,6 +16,12 @@
 #include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/reduce_and_gather.h"
 #include "paddle/fluid/framework/details/variable_visitor.h"
+#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
+#include "paddle/fluid/operators/distributed/collective_client.h"
+#include "paddle/fluid/operators/distributed/collective_server.h"
+#include "paddle/fluid/operators/distributed/request_handler.h"
+#endif
+#include "paddle/fluid/operators/math/selected_rows_functor.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DEFINE_bool(
@@ -26,6 +32,112 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+std::once_flag CollectiveContext::init_flag_;
+std::unique_ptr<CollectiveContext> CollectiveContext::context_;
+
+static inline std::string GetRemoteVarName(const std::string &var_name,
+                                           int trainer_id) {
+  return string::Sprintf("%s_merged_tmp@trainer_%d", var_name, trainer_id);
+}
+
+void ReduceOpHandle::Wait(
+    const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes) {
+  // TODO(gongwb): use event wait?
+  for (auto &dev_ctx : dev_ctxes) {
+    dev_ctx.second->Wait();
+  }
+}
+
+#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
+template <typename DevCtx, typename DataType>
+void ReduceOpHandle::GatherSelectedRows(
+    const std::vector<const SelectedRows *> &src_selected_rows,
+    const std::vector<platform::Place> &in_places,
+    const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes,
+    VarHandle *out_var_handle, const platform::Place &out_place,
+    SelectedRows *dst_selected_rows) {
+  const CollectiveContext &collective_context =
+      *CollectiveContext::GetInstance();
+
+  // 1. gather local selected rows, merge them
+  std::string gathered_var_name = out_var_handle->name_ + "_gathered_tmp";
+  auto scope = local_scopes_.at(out_var_handle->scope_idx_);
+  auto gathered_var_mid = scope->Var(gathered_var_name);
+  auto gathered_select_rows =
+      gathered_var_mid->GetMutable<framework::SelectedRows>();
+  GatherLocalSelectedRows(src_selected_rows, in_places, dev_ctxes, out_place,
+                          gathered_select_rows);
+  // FIXME(gongwb): remove this Wait.
+  Wait(dev_ctxes);
+
+  // merge them
+  auto merged_dev_ctx = dynamic_cast<DevCtx *>(dev_ctxes.at(out_place));
+  std::string merged_var_name =
+      GetRemoteVarName(out_var_handle->name_, collective_context.trainer_id_);
+  auto merged_select_rows =
+      scope->Var(merged_var_name)->GetMutable<SelectedRows>();
+  operators::math::scatter::MergeAdd<DevCtx, DataType> merge_func;
+  merge_func(*merged_dev_ctx, *gathered_select_rows, merged_select_rows);
+
+  // 2. start collective server if it doesn't exist
+  operators::distributed::CollectiveServer *server =
+      operators::distributed::CollectiveServer::GetInstance(
+          collective_context.endpoints_[collective_context.trainer_id_],
+          collective_context.endpoints_.size() - 1);
+
+  auto rpc_server = server->GetRPCServer();
+  rpc_server->RegisterVar(merged_var_name,
+                          operators::distributed::kRequestGetMonomerVariable,
+                          scope, merged_dev_ctx);
+
+  // 3. gather them from all remote nodes.
+  std::vector<const SelectedRows *> remote;
+  operators::distributed::CollectiveClient *client =
+      operators::distributed::CollectiveClient::GetInstance();
+
+  std::vector<operators::distributed::RemoteVar> vars;
+  for (unsigned int i = 0; i < collective_context.endpoints_.size(); i++) {
+    if (i == (unsigned)collective_context.trainer_id_) continue;
+
+    operators::distributed::RemoteVar var;
+    var.trainer_id_ = i;
+    var.var_name_ = GetRemoteVarName(out_var_handle->name_, i);
+    var.ep_ = collective_context.endpoints_[i];
+
+    vars.push_back(var);
+    VLOG(4) << "gather from:" << var.String();
+  }
+
+  // erase gathered vars
+  merged_dev_ctx->Wait();
+  scope->EraseVars(std::vector<std::string>{gathered_var_name});
+
+  PADDLE_ENFORCE(client->Gather(vars, &remote, *merged_dev_ctx, scope));
+  PADDLE_ENFORCE(remote.size() == vars.size());
+
+  // 4. merged local selected rows.
+  std::vector<const SelectedRows *> all;
+  all.resize(collective_context.endpoints_.size());
+  for (auto v : vars) {
+    all[v.trainer_id_] =
+        scope->FindVar(v.var_name_)->GetMutable<SelectedRows>();
+  }
+  all[collective_context.trainer_id_] = merged_select_rows;
+
+  merge_func(*merged_dev_ctx, all, dst_selected_rows);
+
+  rpc_server->WaitVarBarrier(merged_var_name);
+  rpc_server->ClearVar(merged_var_name);
+
+  // 5. clear mid vars
+  std::vector<std::string> tmp_vars{merged_var_name};
+  for (auto r : vars) {
+    tmp_vars.push_back(r.var_name_);
+  }
+  scope->EraseVars(tmp_vars);
+}
+#endif
+
 void ReduceOpHandle::RunImpl() {
   platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
 
@@ -90,8 +202,36 @@ void ReduceOpHandle::RunImpl() {
     this->RunAndRecordEvent([&] {
       std::vector<const SelectedRows *> in_selected_rows =
           GetInputValues<SelectedRows>(in_var_handles, var_scopes);
-      GatherSelectedRows(in_selected_rows, in_places, dev_ctxes_, t_out_p,
+
+      const CollectiveContext &collective_context =
+          *CollectiveContext::GetInstance();
+      VLOG(10) << "GatherSelectedRows CollectiveContext:"
+               << collective_context.String();
+
+      // TODO(gongwb): add cpu support
+      if (collective_context.endpoints_.size() <= 1 ||
+          is_cpu_place(in_places[0]) || is_cpu_place(t_out_p)) {
+        GatherLocalSelectedRows(in_selected_rows, in_places, dev_ctxes_,
+                                t_out_p,
+                                out_var->GetMutable<framework::SelectedRows>());
+        return;
+      }
+
+#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
+      if (framework::IsType<const float>(in_selected_rows[0]->value().type())) {
+        GatherSelectedRows<platform::CUDADeviceContext, float>(
+            in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p,
             out_var->GetMutable<framework::SelectedRows>());
+      } else if (framework::IsType<const double>(
+                     in_selected_rows[0]->value().type())) {
+        GatherSelectedRows<platform::CUDADeviceContext, double>(
+            in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p,
+            out_var->GetMutable<framework::SelectedRows>());
+      } else {
+        PADDLE_ENFORCE(false,
+                       "only support double or float when gahter SelectedRows");
+      }
+#endif
     });
   } else {
     std::vector<const LoDTensor *> lod_tensors =

paddle/fluid/framework/details/reduce_op_handle.h

@@ -30,6 +30,32 @@
 namespace paddle {
 namespace framework {
 namespace details {
+struct CollectiveContext {
+  std::vector<std::string> endpoints_;
+  int trainer_id_{0};
+
+  std::string String() const {
+    std::stringstream ss;
+    ss << "endpoints_:";
+    for (auto e : endpoints_) {
+      ss << e << ",";
+    }
+
+    ss << "trainer_id_:" << trainer_id_;
+
+    return ss.str();
+  }
+
+  static CollectiveContext *GetInstance() {
+    std::call_once(init_flag_,
+                   [&]() { context_.reset(new CollectiveContext()); });
+    return context_.get();
+  }
+
+ private:
+  static std::once_flag init_flag_;
+  static std::unique_ptr<CollectiveContext> context_;
+};
 
 struct ReduceOpHandle : public OpHandleBase {
   std::vector<Scope *> local_scopes_;
@@ -64,6 +90,19 @@ struct ReduceOpHandle : public OpHandleBase {
  protected:
   void RunImpl() override;
 
+#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
+  template <typename DevCtx, typename DataType>
+  void GatherSelectedRows(
+      const std::vector<const SelectedRows *> &src_selecte_rows_,
+      const std::vector<platform::Place> &in_places,
+      const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes,
+      VarHandle *out_var_handle, const platform::Place &out_place,
+      SelectedRows *dst_selecte_rows);
+#endif
+
+  void Wait(
+      const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes);
+
   template <typename T>
   std::vector<const T *> GetInputValues(
       const std::vector<VarHandle *> &in_var_handles,

paddle/fluid/framework/executor_thread_worker.cc

@@ -97,7 +97,7 @@ void ExecutorThreadWorker::SetDevice() {
   static unsigned concurrency_cap = std::thread::hardware_concurrency();
   int thread_id = this->thread_id_;
 
-  if (thread_id < concurrency_cap) {
+  if (static_cast<unsigned>(thread_id) < concurrency_cap) {
     unsigned proc = thread_id;
 
     cpu_set_t mask;

paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc

@@ -46,14 +46,16 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
   auto* scope = param_scope();
   PADDLE_ENFORCE(scope);
 
+  std::string type = is_conv3d() ? "conv3d" : "conv2d";
+
   GraphPatternDetector gpd;
   auto* conv_input =
       gpd.mutable_pattern()
           ->NewNode(patterns::PDNodeName(name_scope_, "conv_input"))
           ->AsInput()
-          ->assert_is_op_input("conv2d", "Input");
+          ->assert_is_op_input(type, "Input");
   patterns::ConvBias conv_bias_pattern(gpd.mutable_pattern(), name_scope_);
-  conv_bias_pattern(conv_input);
+  conv_bias_pattern(conv_input, is_conv3d());
   int found_conv_bias_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
@@ -109,7 +111,7 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
       desc.SetInput("Filter", std::vector<std::string>({conv_weight->Name()}));
       desc.SetInput("Bias", std::vector<std::string>({eltwise_bias->Name()}));
       desc.SetOutput("Output", std::vector<std::string>({eltwise_out->Name()}));
-      desc.SetType("conv2d");
+      desc.SetType(type);
 
       for (auto& attr : conv->Op()->GetAttrMap()) {
         desc.SetAttr(attr.first, attr.second);
@@ -135,3 +137,5 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
 }  // namespace paddle
 REGISTER_PASS(conv_bias_mkldnn_fuse_pass,
               paddle::framework::ir::ConvBiasFusePass);
+REGISTER_PASS(conv3d_bias_mkldnn_fuse_pass,
+              paddle::framework::ir::Conv3DBiasFusePass);

paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h

@@ -26,11 +26,19 @@ namespace ir {
 class ConvBiasFusePass : public FusePassBase {
  public:
   virtual ~ConvBiasFusePass() {}
+  virtual bool is_conv3d() const { return false; }
 
  protected:
   std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
   const std::string name_scope_{"conv_bias_mkldnn_fuse"};
 };
+/*
+* Fuse the Conv3D and Elementwise_add to a Conv3DBiasOp.
+*/
+class Conv3DBiasFusePass : public ConvBiasFusePass {
+ public:
+  bool is_conv3d() const override { return true; }
+};
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/graph.h

@@ -177,14 +177,13 @@ class Graph {
     return nullptr;
   }
 
-  const ProgramDesc &program() const { return program_; }
-  std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
-      const ProgramDesc &program);
-
   void ResolveHazard(
       const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
 
  private:
+  std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
+      const ProgramDesc &program);
+
   // This method takes ownership of `node`.
   ir::Node *AddNode(ir::Node *node) {
     PADDLE_ENFORCE(node_set_.find(node) == node_set_.end());

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -1030,10 +1030,11 @@ PDNode *patterns::ElewiseAddActInplaceGrad::operator()(
 }
 
 PDNode *patterns::ConvBias::operator()(
-    paddle::framework::ir::PDNode *conv_input) {
+    paddle::framework::ir::PDNode *conv_input, bool is_conv3d) {
+  std::string type = is_conv3d ? "conv3d" : "conv2d";
   // Create Operators
-  conv_input->assert_is_op_input("conv2d", "Input");
-  auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");
+  conv_input->assert_is_op_input(type, "Input");
+  auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op(type);
   auto *eltiwse_op =
       pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add");
   // Create variables
@@ -1041,11 +1042,11 @@ PDNode *patterns::ConvBias::operator()(
   auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
                               ->AsInput()
                               ->assert_is_persistable_var()
-                              ->assert_is_op_input("conv2d", "Filter");
+                              ->assert_is_op_input(type, "Filter");
   // intermediate variable, will be removed in the IR after fuse.
   auto *conv_out_var = pattern->NewNode(conv_out_repr())
                            ->AsIntermediate()
-                           ->assert_is_only_output_of_op("conv2d")
+                           ->assert_is_only_output_of_op(type)
                            ->assert_is_op_input("elementwise_add");
   // Bias stored in elementwise_add
   auto *eltwise_bias_var = pattern->NewNode(eltwise_bias_repr())

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -623,7 +623,7 @@ struct ElewiseAddActInplaceGrad : public PatternBase {
 struct ConvBias : public PatternBase {
   ConvBias(PDPattern* pattern, const std::string& name_scope)
       : PatternBase(pattern, name_scope, "conv_bias") {}
-  PDNode* operator()(PDNode* conv_input);
+  PDNode* operator()(PDNode* conv_input, bool is_conv3d = false);
   // declare operator node's name
   PATTERN_DECL_NODE(conv);
   PATTERN_DECL_NODE(eltwise);

paddle/fluid/framework/ir/mkldnn_placement_pass.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/ir/mkldnn_placement_pass.h"
+#include <string>
 
 namespace paddle {
 namespace framework {
@@ -21,9 +22,16 @@ namespace ir {
 std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
   VLOG(3) << "Aplies MKL-DNN placement strategy.";
+  const auto& op_types_list =
+      Get<std::unordered_set<std::string>>("mkldnn_enabled_op_types");
   for (const Node* n : graph->Nodes()) {
     if (n->IsOp() && n->RuntimeHasAttr("use_mkldnn")) {
+      if (op_types_list.empty()) {
         n->Op()->SetAttr("use_mkldnn", true);
+      } else if (std::find(op_types_list.begin(), op_types_list.end(),
+                           n->Name()) != op_types_list.end()) {
+        n->Op()->SetAttr("use_mkldnn", true);
+      }
     }
   }
   return graph;
@@ -33,5 +41,5 @@ std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
 }  // namespace framework
 }  // namespace paddle
 
-REGISTER_PASS(mkldnn_placement_pass,
-              paddle::framework::ir::MKLDNNPlacementPass);
+REGISTER_PASS(mkldnn_placement_pass, paddle::framework::ir::MKLDNNPlacementPass)
+    .RequirePassAttr("mkldnn_enabled_op_types");

paddle/fluid/framework/op_kernel_type.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. */
+
+#include "paddle/fluid/framework/op_kernel_type.h"
+
+namespace paddle {
+namespace framework {
+
+size_t OpKernelType::Hash::operator()(const OpKernelType& key) const {
+  int cur_loc = 0;
+
+  int place = key.place_.which();
+  cur_loc += OpKernelType::kPlaceBits;
+
+  int data_type = static_cast<int>(key.data_type_) << cur_loc;
+  cur_loc += OpKernelType::kPrimaryDTypeBits;
+
+  int data_layout = static_cast<int>(key.data_layout_) << cur_loc;
+  cur_loc += OpKernelType::kLayoutBits;
+
+  int library_type = static_cast<int>(key.library_type_) << cur_loc;
+  cur_loc += OpKernelType::kLibBits;
+
+  int customized_value = key.customized_type_value_;
+  PADDLE_ENFORCE(customized_value < (1 << OpKernelType::kCustomizeBits));
+  customized_value = customized_value << cur_loc;
+  cur_loc += OpKernelType::kCustomizeBits;
+  PADDLE_ENFORCE(cur_loc < 64);
+
+  std::hash<int> hasher;
+  return hasher(place + data_type + data_layout + library_type +
+                customized_value);
+}
+
+bool OpKernelType::operator==(const OpKernelType& o) const {
+  return platform::places_are_same_class(place_, o.place_) &&
+         data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
+         library_type_ == o.library_type_ &&
+         customized_type_value_ == o.customized_type_value_;
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/op_kernel_type.h

@@ -24,54 +24,55 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-struct OpKernelType {
-  struct Hash {
-    size_t operator()(const OpKernelType& key) const {
-      int place = key.place_.which();
-      int data_type = static_cast<int>(key.data_type_) << LEFT_SHIFT;
-      int data_layout = static_cast<int>(key.data_layout_) << (LEFT_SHIFT * 2);
-      int library_type = static_cast<int>(key.library_type_)
-                         << (LEFT_SHIFT * 3);
-
-      std::hash<int> hasher;
-      return hasher(place + data_type + data_layout + library_type);
-    }
-  };
+class OpKernelType {
+ public:
+  constexpr static int kDefaultCustomizedTypeValue = 0;
 
-  // place, data_type, library_type kinds less than 2^8
-  constexpr static int LEFT_SHIFT = 8;
-
-  proto::VarType::Type data_type_;
-  DataLayout data_layout_;
-  platform::Place place_;
-  LibraryType library_type_;
+  // In total should be smaller than 64.
+  constexpr static int kPlaceBits = 4;
+  constexpr static int kPrimaryDTypeBits = 8;
+  constexpr static int kLayoutBits = 4;
+  constexpr static int kLibBits = 4;
+  constexpr static int kCustomizeBits = 4;
 
   OpKernelType(proto::VarType::Type data_type, platform::Place place,
                DataLayout data_layout = DataLayout::kAnyLayout,
-               LibraryType library_type = LibraryType::kPlain)
+               LibraryType library_type = LibraryType::kPlain,
+               int customized_type_value = kDefaultCustomizedTypeValue)
       : data_type_(data_type),
         data_layout_(data_layout),
         place_(place),
-        library_type_(library_type) {}
+        library_type_(library_type),
+        customized_type_value_(customized_type_value) {}
 
   OpKernelType(proto::VarType::Type data_type,
                const platform::DeviceContext& dev_ctx,
                DataLayout data_layout = DataLayout::kAnyLayout,
-               LibraryType library_type = LibraryType::kPlain)
+               LibraryType library_type = LibraryType::kPlain,
+               int customized_type_value = kDefaultCustomizedTypeValue)
       : data_type_(data_type),
         data_layout_(data_layout),
         place_(dev_ctx.GetPlace()),
-        library_type_(library_type) {}
+        library_type_(library_type),
+        customized_type_value_(customized_type_value) {}
+
+  virtual ~OpKernelType() {}
+
+  struct Hash {
+    size_t operator()(const OpKernelType& key) const;
+  };
 
   size_t hash_key() const { return Hash()(*this); }
 
-  bool operator==(const OpKernelType& o) const {
-    return platform::places_are_same_class(place_, o.place_) &&
-           data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
-           library_type_ == o.library_type_;
-  }
+  bool operator==(const OpKernelType& o) const;
 
   bool operator!=(const OpKernelType& o) const { return !(*this == o); }
+
+  proto::VarType::Type data_type_;
+  DataLayout data_layout_;
+  platform::Place place_;
+  LibraryType library_type_;
+  int customized_type_value_;
 };
 
 inline std::ostream& operator<<(std::ostream& os,

paddle/fluid/framework/op_registry.h

@@ -35,6 +35,7 @@ limitations under the License. */
 
 namespace paddle {
 namespace framework {
+
 class Registrar {
  public:
   // In our design, various kinds of classes, e.g., operators and kernels,
@@ -78,7 +79,7 @@ struct OpKernelRegistrarFunctor;
 
 template <typename PlaceType, typename T, typename Func>
 inline void RegisterKernelClass(const char* op_type, const char* library_type,
-                                Func func) {
+                                int customized_type_value, Func func) {
   std::string library(library_type);
   std::string data_layout = "ANYLAYOUT";
   if (library == "MKLDNN") {
@@ -86,7 +87,7 @@ inline void RegisterKernelClass(const char* op_type, const char* library_type,
   }
   OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(),
                    StringToDataLayout(data_layout),
-                   StringToLibraryType(library_type));
+                   StringToLibraryType(library_type), customized_type_value);
   OperatorWithKernel::AllOpKernels()[op_type][key] = func;
 }
 
@@ -95,22 +96,26 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
   using KERNEL_TYPE =
       typename std::tuple_element<I, std::tuple<KernelTypes...>>::type;
 
-  void operator()(const char* op_type, const char* library_type) const {
+  void operator()(const char* op_type, const char* library_type,
+                  int customized_type_value) const {
     using T = typename KERNEL_TYPE::ELEMENT_TYPE;
     RegisterKernelClass<PlaceType, T>(
-        op_type, library_type, [](const framework::ExecutionContext& ctx) {
+        op_type, library_type, customized_type_value,
+
+        [](const framework::ExecutionContext& ctx) {
           KERNEL_TYPE().Compute(ctx);
         });
     constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;
     OpKernelRegistrarFunctor<PlaceType, I + 1 == size, I + 1, KernelTypes...>
         func;
-    func(op_type, library_type);
+    func(op_type, library_type, customized_type_value);
   }
 };
 
 template <typename PlaceType, size_t I, typename... KernelType>
 struct OpKernelRegistrarFunctor<PlaceType, true, I, KernelType...> {
-  void operator()(const char* op_type, const char* library_type) const {}
+  void operator()(const char* op_type, const char* library_type,
+                  int customized_type_value) const {}
 };
 
 // User can register many kernel in one place. The data type could be
@@ -118,9 +123,10 @@ struct OpKernelRegistrarFunctor<PlaceType, true, I, KernelType...> {
 template <typename PlaceType, typename... KernelType>
 class OpKernelRegistrar : public Registrar {
  public:
-  explicit OpKernelRegistrar(const char* op_type, const char* library_type) {
+  explicit OpKernelRegistrar(const char* op_type, const char* library_type,
+                             int customized_type_value) {
     OpKernelRegistrarFunctor<PlaceType, false, 0, KernelType...> func;
-    func(op_type, library_type);
+    func(op_type, library_type, customized_type_value);
   }
 };
 
@@ -130,17 +136,19 @@ struct OpKernelRegistrarFunctorEx;
 template <typename PlaceType, typename... DataTypeAndKernelType>
 class OpKernelRegistrarEx : public Registrar {
  public:
-  explicit OpKernelRegistrarEx(const char* op_type, const char* library_type) {
+  explicit OpKernelRegistrarEx(const char* op_type, const char* library_type,
+                               int customized_type_value) {
     OpKernelRegistrarFunctorEx<PlaceType, false, 0, DataTypeAndKernelType...>
         func;
-    func(op_type, library_type);
+    func(op_type, library_type, customized_type_value);
   }
 };
 
 template <typename PlaceType, size_t I, typename... DataTypeAndKernelType>
 struct OpKernelRegistrarFunctorEx<PlaceType, true, I,
                                   DataTypeAndKernelType...> {
-  void operator()(const char* op_type, const char* library_type) const {}
+  void operator()(const char* op_type, const char* library_type,
+                  int customized_type_value) const {}
 };
 
 template <typename PlaceType, size_t I, typename... DataTypeAndKernelType>
@@ -153,18 +161,21 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
       typename std::tuple_element<I,
                                   std::tuple<DataTypeAndKernelType...>>::type;
 
-  void operator()(const char* op_type, const char* library_type) const {
-    RegisterKernelClass<PlaceType, T>(op_type, library_type, Functor());
+  void operator()(const char* op_type, const char* library_type,
+                  int customized_type_value) const {
+    RegisterKernelClass<PlaceType, T>(op_type, library_type,
+                                      customized_type_value, Functor());
 
     constexpr auto size =
         std::tuple_size<std::tuple<DataTypeAndKernelType...>>::value;
     OpKernelRegistrarFunctorEx<PlaceType, I + 2 >= size, I + 2,
                                DataTypeAndKernelType...>
         func;
-    func(op_type, library_type);
+    func(op_type, library_type, customized_type_value);
   }
 };
 
+// clang-format off
 /**
  * check if MACRO is used in GLOBAL NAMESPACE.
  */
@@ -199,42 +210,64 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
 /**
  * Macro to register OperatorKernel.
  */
-#define REGISTER_OP_KERNEL(op_type, library_type, place_class, ...)        \
+#define REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(op_type, library_type,             \
+                                            place_class, customized_name,      \
+                                            customized_type_value, ...)        \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                              \
-      __reg_op_kernel_##op_type##_##library_type##__,                      \
-      "REGISTER_OP_KERNEL must be called in global namespace");            \
-  static ::paddle::framework::OpKernelRegistrar<place_class, __VA_ARGS__>  \
-      __op_kernel_registrar_##op_type##_##library_type##__(#op_type,       \
-                                                           #library_type); \
-  int TouchOpKernelRegistrar_##op_type##_##library_type() {                \
-    __op_kernel_registrar_##op_type##_##library_type##__.Touch();          \
+      __reg_op_kernel_##op_type##_##library_type##_##customized_name##__,      \
+                                 "REGISTER_OP_KERNEL must be called in "       \
+                                 "global namespace");                          \
+  static ::paddle::framework::OpKernelRegistrar<place_class,                   \
+                                                __VA_ARGS__>                   \
+      __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__(\
+          #op_type, #library_type, customized_type_value);                     \
+  int TouchOpKernelRegistrar_##op_type##_##library_type##_##customized_name() {\
+    __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__   \
+        .Touch();                                                              \
     return 0;                                                                  \
   }
 
+#define REGISTER_OP_KERNEL(op_type, library_type, place_class, ...)   \
+  REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(                                \
+      op_type, library_type, place_class, DEFAULT_TYPE,               \
+      ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \
+      __VA_ARGS__)
+
 #define REGISTER_OP_CUDA_KERNEL(op_type, ...) \
   REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::CUDAPlace, __VA_ARGS__)
 
 #define REGISTER_OP_CPU_KERNEL(op_type, ...) \
   REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
 
-#define REGISTER_OP_KERNEL_EX(op_type, library_type, place_class, ...)      \
+#define REGISTER_OP_KERNEL_EX(op_type, library_type, place_class,  \
+                              customized_name,                     \
+                              customized_type_value,               \
+                              ...)                                 \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                  \
-      __reg_op_kernel_##op_type##_##library_type##__,                       \
-      "REGISTER_OP_KERNEL_EX must be called in global namespace");          \
-  static ::paddle::framework::OpKernelRegistrarEx<place_class, __VA_ARGS__> \
-      __op_kernel_registrar_##op_type##_##library_type##__(#op_type,        \
-                                                           #library_type);  \
-  int TouchOpKernelRegistrar_##op_type##_##library_type() {                 \
-    __op_kernel_registrar_##op_type##_##library_type##__.Touch();           \
+      __reg_op_kernel_##op_type##_##library_type##_##customized_name##__, \
+                                 "REGISTER_OP_KERNEL_EX must be called in "  \
+                                 "global namespace");  \
+  static ::paddle::framework::OpKernelRegistrarEx<place_class,  \
+                                                  __VA_ARGS__>  \
+      __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__(\
+          #op_type, #library_type, customized_type_value);  \
+  int TouchOpKernelRegistrar_##op_type##_##library_type##_##customized_name() {\
+    __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__   \
+        .Touch();                                                              \
     return 0;                                                                  \
   }
 
 #define REGISTER_OP_CUDA_KERNEL_FUNCTOR(op_type, ...)                 \
-  REGISTER_OP_KERNEL_EX(op_type, CUDA, ::paddle::platform::CUDAPlace, \
+  REGISTER_OP_KERNEL_EX(                                              \
+      op_type, CUDA, ::paddle::platform::CUDAPlace, DEFAULT_TYPE,     \
+      ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \
       __VA_ARGS__)
 
 #define REGISTER_OP_CPU_KERNEL_FUNCTOR(op_type, ...)                  \
-  REGISTER_OP_KERNEL_EX(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
+  REGISTER_OP_KERNEL_EX(                                              \
+      op_type, CPU, ::paddle::platform::CPUPlace, DEFAULT_TYPE,       \
+      ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \
+      __VA_ARGS__)
 
 /**
  * Macro to mark what Operator and Kernel
@@ -248,13 +281,19 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
   extern int TouchOpRegistrar_##op_type();                 \
   UNUSED static int use_op_itself_##op_type##_ = TouchOpRegistrar_##op_type()
 
-#define USE_OP_DEVICE_KERNEL(op_type, LIBRARY_TYPE)               \
+#define USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(op_type,                     \
+                                              LIBRARY_TYPE,                \
+                                              customized_name)             \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                          \
-      __use_op_kernel_##op_type##_##LIBRARY_TYPE##__,             \
+      __use_op_kernel_##op_type##_##LIBRARY_TYPE##_##customized_name##__,  \
       "USE_OP_DEVICE_KERNEL must be in global namespace");                 \
-  extern int TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE(); \
-  UNUSED static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_ = \
-      TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE()
+  extern int                                                               \
+      TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE##_##customized_name(); \
+  UNUSED static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_##DEFAULT_TYPE##_ = /* NOLINT */ \
+      TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE##_##customized_name()
+
+#define USE_OP_DEVICE_KERNEL(op_type, LIBRARY_TYPE) \
+  USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(op_type, LIBRARY_TYPE, DEFAULT_TYPE)
 
 // TODO(fengjiayi): The following macros
 // seems ugly, do we have better method?
@@ -280,6 +319,7 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
 #define USE_OP(op_type)   \
   USE_OP_ITSELF(op_type); \
   USE_OP_KERNEL(op_type)
+// clang-format off
 
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/operator_test.cc

@@ -50,6 +50,8 @@ class OpWithoutKernelCheckerMaker : public OpProtoAndCheckerMaker {
     AddInput("input", "input of test op");
     AddOutput("output", "output of test op");
     AddAttr<float>("scale", "scale of cosine op");
+    AddAttr<int>("kernel_sub_type", "kernels with different implementations.")
+        .SetDefault(0);
     AddComment("This is test op");
   }
 };
@@ -95,6 +97,8 @@ TEST(OperatorBase, all) {
 namespace paddle {
 namespace framework {
 
+static int special_type_value = 1;
+
 class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
   void Make() {
@@ -103,11 +107,14 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .GreaterThan(0.0);
+    AddAttr<int>("kernel_sub_type", "kernels with different implementations.")
+        .SetDefault(0);
     AddComment("This is test op");
   }
 };
 
 static int cpu_kernel_run_num = 0;
+static int cpu_kernel2_run_num = 0;
 
 class OpWithKernelTest : public OperatorWithKernel {
  public:
@@ -117,7 +124,10 @@ class OpWithKernelTest : public OperatorWithKernel {
   void InferShape(framework::InferShapeContext* ctx) const override {}
   OpKernelType GetExpectedKernelType(
       const ExecutionContext& ctx) const override {
-    return OpKernelType(proto::VarType::FP32, ctx.GetPlace());
+    int sub_type = ctx.Attr<int>("kernel_sub_type");
+    return OpKernelType(proto::VarType::FP32, ctx.GetPlace(),
+                        framework::DataLayout::kAnyLayout,
+                        framework::LibraryType::kPlain, sub_type);
   }
 };
 
@@ -132,6 +142,17 @@ class CPUKernelTest : public OpKernel<float> {
   }
 };
 
+template <typename T1, typename T2>
+class CPUKernel2Test : public OpKernel<float> {
+ public:
+  void Compute(const ExecutionContext& ctx) const {
+    std::cout << ctx.op().DebugString() << std::endl;
+    cpu_kernel2_run_num++;
+    ASSERT_EQ(ctx.op().Input("x"), "IN1");
+    ASSERT_EQ(ctx.op().Output("y"), "OUT1");
+  }
+};
+
 class OpKernelTestMultiInputsProtoAndCheckerMaker
     : public OpProtoAndCheckerMaker {
  public:
@@ -142,6 +163,8 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .GreaterThan(0.0);
+    AddAttr<int>("kernel_sub_type", "kernels with different implementations.")
+        .SetDefault(0);
     AddComment("This is test op");
   }
 };
@@ -189,9 +212,15 @@ class CPUKernalMultiInputsTest : public OpKernel<float> {
 REGISTER_OP_WITHOUT_GRADIENT(
     op_with_kernel, paddle::framework::OpWithKernelTest,
     paddle::framework::OpKernelTestProtoAndCheckerMaker);
+
 REGISTER_OP_CPU_KERNEL(op_with_kernel,
                        paddle::framework::CPUKernelTest<float, float>);
 
+REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(
+    op_with_kernel, CPU, paddle::platform::CPUPlace, MY_SPECIAL_NAME,
+    paddle::framework::special_type_value,
+    paddle::framework::CPUKernel2Test<float, float>);
+
 // test with single input
 TEST(OpKernel, all) {
   paddle::framework::InitDevices(true);
@@ -211,7 +240,19 @@ TEST(OpKernel, all) {
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 0);
   op->Run(scope, cpu_place);
+  // kerne_sub_type = 0, hence cpu_kernel is called, cpu_kernel2 is not called.
+  ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
+  ASSERT_EQ(paddle::framework::cpu_kernel2_run_num, 0);
+
+  attr = op_desc.mutable_attrs()->Add();
+  attr->set_name("kernel_sub_type");
+  attr->set_type(paddle::framework::proto::AttrType::INT);
+  attr->set_i(1);
+  auto op2 = paddle::framework::OpRegistry::CreateOp(op_desc);
+  op2->Run(scope, cpu_place);
+  // kerne_sub_type = 1, hence cpu_kernel2 is called, cpu_kernel is not called.
   ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
+  ASSERT_EQ(paddle::framework::cpu_kernel2_run_num, 1);
 }
 
 REGISTER_OP_WITHOUT_GRADIENT(

paddle/fluid/inference/analysis/argument.h

@@ -103,6 +103,7 @@ struct Argument {
   // Model specified with program and parameters files.
   DECL_ARGUMENT_FIELD(model_program_path, ModelProgramPath, std::string);
   DECL_ARGUMENT_FIELD(model_params_path, ModelParamsPath, std::string);
+  DECL_ARGUMENT_FIELD(model_from_memory, ModelFromMemory, bool);
 
   // The overall graph to work on.
   DECL_ARGUMENT_UNIQUE_FIELD(main_graph, MainGraph, framework::ir::Graph);
@@ -115,6 +116,10 @@ struct Argument {
   DECL_ARGUMENT_FIELD(ir_analysis_passes, IrAnalysisPasses,
                       std::vector<std::string>);
 
+  // Pass a set of op types to enable its mkldnn kernel
+  DECL_ARGUMENT_FIELD(mkldnn_enabled_op_types, MKLDNNEnabledOpTypes,
+                      std::unordered_set<std::string>);
+
   DECL_ARGUMENT_FIELD(use_gpu, UseGPU, bool);
   DECL_ARGUMENT_FIELD(gpu_device_id, GPUDeviceId, int);
   DECL_ARGUMENT_FIELD(use_tensorrt, UseTensorRT, bool);

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -63,6 +63,11 @@ void IRPassManager::CreatePasses(Argument *argument,
       pass->Set("graph_viz_path", new std::string(std::move(dot_file_path)));
       pass_num++;
     }
+    if (pass_name == "mkldnn_placement_pass") {
+      pass->Set("mkldnn_enabled_op_types",
+                new std::unordered_set<std::string>(
+                    argument->mkldnn_enabled_op_types()));
+    }
 
     if (pass_name == "tensorrt_subgraph_pass") {
       PADDLE_ENFORCE(argument->tensorrt_node_teller_valid());

paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc

@@ -178,11 +178,12 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
     output_mapping.push_back(output_name_map[name]);
   }
 
-  *block_desc.Proto()->mutable_vars() =
-      const_cast<framework::ProgramDesc *>(&graph->program())
-          ->Proto()
-          ->blocks(0)
-          .vars();
+  auto *vars = block_desc.Proto()->mutable_vars();
+  for (framework::ir::Node *node : graph->Nodes()) {
+    if (node->IsVar() && node->Var()) {
+      *vars->Add() = *node->Var()->Proto();
+    }
+  }
   PADDLE_ENFORCE(!block_desc.Proto()->vars().empty(),
                  "the block has no var-desc");
   PADDLE_ENFORCE(!output_mapping.empty());

paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc

@@ -46,7 +46,7 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
              argument->model_params_path_valid()) {
     auto program =
         LoadModel(argument->model_program_path(), argument->model_params_path(),
-                  argument->scope_ptr(), place);
+                  argument->scope_ptr(), place, argument->model_from_memory());
     argument->SetMainProgram(program.release());
   } else {
     PADDLE_THROW(
@@ -68,9 +68,14 @@ std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
 
 std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
     const std::string &program_path, const std::string &params_path,
-    framework::Scope *scope, const platform::Place &place) {
+    framework::Scope *scope, const platform::Place &place,
+    bool model_from_memory) {
   framework::Executor exe(place);
+  if (!model_from_memory) {
     return Load(&exe, scope, program_path, params_path);
+  } else {
+    return LoadFromMemory(&exe, scope, program_path, params_path);
+  }
 }
 
 std::string IrGraphBuildPass::repr() const { return "ir-graph-build-pass"; }

paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h

@@ -24,7 +24,7 @@ namespace inference {
 namespace analysis {
 
 /*
- * Load program and parameter to memory from the disk.
+ * Load program and parameter to memory from the disk or directly from memory.
  */
 class IrGraphBuildPass : public AnalysisPass {
  public:
@@ -38,7 +38,8 @@ class IrGraphBuildPass : public AnalysisPass {
       const platform::Place &place);
   std::unique_ptr<framework::ProgramDesc> LoadModel(
       const std::string &program_path, const std::string &params_path,
-      framework::Scope *scope, const platform::Place &place);
+      framework::Scope *scope, const platform::Place &place,
+      bool model_from_memory);
 
   std::string model_binary_str_;
 };

paddle/fluid/inference/api/analysis_config.cc

@@ -49,10 +49,15 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) {
   cpu_math_library_num_threads_ = other.cpu_math_library_num_threads_;
   // fields from this.
   enable_ir_optim = other.enable_ir_optim;
+  // For mkldnn
+  use_mkldnn_ = other.use_mkldnn_;
+  mkldnn_enabled_op_types_ = other.mkldnn_enabled_op_types_;
+
   use_feed_fetch_ops = other.use_feed_fetch_ops;
   use_tensorrt_ = other.use_tensorrt_;
   tensorrt_max_batchsize_ = other.tensorrt_max_batchsize_;
   tensorrt_workspace_size_ = other.tensorrt_workspace_size_;
+  model_from_memory_ = other.model_from_memory_;
 
   if (use_gpu) {
     pass_builder_.reset(new GpuPassStrategy(
@@ -76,10 +81,16 @@ contrib::AnalysisConfig::AnalysisConfig(contrib::AnalysisConfig &&other) {
   cpu_math_library_num_threads_ = other.cpu_math_library_num_threads_;
   // fields from this.
   enable_ir_optim = other.enable_ir_optim;
+  // For mkldnn
+  use_mkldnn_ = other.use_mkldnn_;
+  mkldnn_enabled_op_types_ = other.mkldnn_enabled_op_types_;
+
   use_feed_fetch_ops = other.use_feed_fetch_ops;
   use_tensorrt_ = other.use_tensorrt_;
   tensorrt_max_batchsize_ = other.tensorrt_max_batchsize_;
   tensorrt_workspace_size_ = other.tensorrt_workspace_size_;
+  model_from_memory_ = other.model_from_memory_;
+
   pass_builder_ = std::move(other.pass_builder_);
 }
 
@@ -102,4 +113,13 @@ void contrib::AnalysisConfig::EnableTensorRtEngine(int workspace_size,
   pass_builder()->InsertPass(1, "tensorrt_subgraph_pass");
 }
 
+void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer,
+                                             size_t prog_buffer_size,
+                                             const char *param_buffer,
+                                             size_t param_buffer_size) {
+  prog_file = std::string(prog_buffer, prog_buffer + prog_buffer_size);
+  param_file = std::string(param_buffer, param_buffer + param_buffer_size);
+  model_from_memory_ = true;
+}
+
 }  // namespace paddle

paddle/fluid/inference/api/analysis_predictor.cc

@@ -308,6 +308,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
 
   argument_.SetUseGPU(config_.use_gpu);
   argument_.SetGPUDeviceId(config_.device);
+  argument_.SetModelFromMemory(config_.model_from_memory_);
   // Analyze inference_program
   if (!config_.model_dir.empty()) {
     argument_.SetModelDir(config_.model_dir);
@@ -326,6 +327,10 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
     argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_);
   }
 
+  if (config_.use_mkldnn_) {
+    argument_.SetMKLDNNEnabledOpTypes(config_.mkldnn_enabled_op_types_);
+  }
+
   auto passes = config_.pass_builder()->AllPasses();
   if (!config_.enable_ir_optim) passes.clear();
   argument_.SetIrAnalysisPasses(passes);
@@ -448,20 +453,24 @@ bool AnalysisPredictor::LoadProgramDesc() {
     return false;
   }
 
+  // Create ProgramDesc
+  framework::proto::ProgramDesc proto;
+  if (!config_.model_from_memory()) {
     std::string pb_content;
     // Read binary
     std::ifstream fin(filename, std::ios::in | std::ios::binary);
-  PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s", filename);
+    PADDLE_ENFORCE(static_cast<bool>(fin.is_open()), "Cannot open file %s",
+                   filename);
     fin.seekg(0, std::ios::end);
-
     pb_content.resize(fin.tellg());
     fin.seekg(0, std::ios::beg);
     fin.read(&(pb_content.at(0)), pb_content.size());
     fin.close();
 
-  // Create ProgramDesc
-  framework::proto::ProgramDesc proto;
     proto.ParseFromString(pb_content);
+  } else {
+    proto.ParseFromString(config_.prog_file);
+  }
   inference_program_.reset(new framework::ProgramDesc(proto));
   return true;
 }
@@ -469,6 +478,7 @@ bool AnalysisPredictor::LoadProgramDesc() {
 bool AnalysisPredictor::LoadParameters() {
   PADDLE_ENFORCE_NOT_NULL(inference_program_.get(),
                           "The inference program should be loaded first.");
+
   const auto &global_block = inference_program_->MutableBlock(0);
 
   // create a temporary program to load parameters.

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -16,6 +16,7 @@
 #include <cassert>
 #include <memory>
 #include <string>
+#include <unordered_set>
 #include <vector>
 
 // Here we include some header files with relative paths, for that in deploy,
@@ -52,18 +53,26 @@ struct AnalysisConfig : public NativeConfig {
   bool use_tensorrt() const { return use_tensorrt_; }
 
   void EnableMKLDNN();
-  // NOTE this is just for internal development, please not use it.
-  // NOT stable yet.
   bool use_mkldnn() const { return use_mkldnn_; }
+  void SetMKLDNNOp(std::unordered_set<std::string> op_list) {
+    mkldnn_enabled_op_types_ = op_list;
+  }
+
+  // Specify the memory buffer of program and parameter
+  void SetModelBuffer(const char* prog_buffer, size_t prog_buffer_size,
+                      const char* program_buffer, size_t program_buffer_size);
+  bool model_from_memory() const { return model_from_memory_; }
 
   friend class ::paddle::AnalysisPredictor;
 
  protected:
   bool use_tensorrt_{false};
   bool use_mkldnn_{false};
+  std::unordered_set<std::string> mkldnn_enabled_op_types_;
   int tensorrt_workspace_size_;
   int tensorrt_max_batchsize_;
   std::unique_ptr<PassStrategy> pass_builder_;
+  bool model_from_memory_{false};
 };
 
 // Configurations for Anakin engine.

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -100,6 +100,7 @@ class CpuPassStrategy : public PassStrategy {
     for (auto &pass :
          std::vector<std::string>({"depthwise_conv_mkldnn_pass",    //
                                    "conv_bias_mkldnn_fuse_pass",    //
+                                   "conv3d_bias_mkldnn_fuse_pass",  //
                                    "conv_relu_mkldnn_fuse_pass",    //
                                    "conv_elementwise_add_mkldnn_fuse_pass"})) {
       passes_.push_back(pass);

paddle/fluid/inference/io.cc

@@ -69,7 +69,8 @@ bool IsPersistable(const framework::VarDesc* var) {
 void LoadPersistables(framework::Executor* executor, framework::Scope* scope,
                       const framework::ProgramDesc& main_program,
                       const std::string& dirname,
-                      const std::string& param_filename) {
+                      const std::string& param_filename,
+                      bool model_from_memory = false) {
   const framework::BlockDesc& global_block = main_program.Block(0);
 
   framework::ProgramDesc* load_program = new framework::ProgramDesc();
@@ -108,6 +109,7 @@ void LoadPersistables(framework::Executor* executor, framework::Scope* scope,
     op->SetType("load_combine");
     op->SetOutput("Out", paramlist);
     op->SetAttr("file_path", {param_filename});
+    op->SetAttr("model_from_memory", {model_from_memory});
     op->CheckAttrs();
   }
 
@@ -130,16 +132,17 @@ std::unique_ptr<framework::ProgramDesc> Load(framework::Executor* executor,
                  "model version %ld is not supported.",
                  main_program->Version());
 
-  LoadPersistables(executor, scope, *main_program, dirname, "");
+  // model_from_memory is false in seperate parameters.
+  LoadPersistables(executor, scope, *main_program, dirname, "",
+                   false /* model_from_memory */);
   return main_program;
 }
 
 std::unique_ptr<framework::ProgramDesc> Load(
     framework::Executor* executor, framework::Scope* scope,
     const std::string& prog_filename, const std::string& param_filename) {
-  std::string model_filename = prog_filename;
   std::string program_desc_str;
-  ReadBinaryFile(model_filename, &program_desc_str);
+  ReadBinaryFile(prog_filename, &program_desc_str);
 
   std::unique_ptr<framework::ProgramDesc> main_program(
       new framework::ProgramDesc(program_desc_str));
@@ -147,7 +150,22 @@ std::unique_ptr<framework::ProgramDesc> Load(
                  "model version %ld is not supported.",
                  main_program->Version());
 
-  LoadPersistables(executor, scope, *main_program, "", param_filename);
+  LoadPersistables(executor, scope, *main_program, "", param_filename,
+                   false /* model_from_memory */);
+  return main_program;
+}
+
+std::unique_ptr<framework::ProgramDesc> LoadFromMemory(
+    framework::Executor* executor, framework::Scope* scope,
+    const std::string& prog_buffer, const std::string& param_buffer) {
+  std::unique_ptr<framework::ProgramDesc> main_program(
+      new framework::ProgramDesc(prog_buffer));
+  PADDLE_ENFORCE(framework::IsProgramVersionSupported(main_program->Version()),
+                 "model version %ld is not supported.",
+                 main_program->Version());
+
+  LoadPersistables(executor, scope, *main_program, "", param_buffer,
+                   true /* model_filename */);
   return main_program;
 }
 

paddle/fluid/inference/io.h

@@ -30,7 +30,8 @@ void Init(const std::vector<std::string> argv);
 void LoadPersistables(framework::Executor* executor, framework::Scope* scope,
                       const framework::ProgramDesc& main_program,
                       const std::string& dirname,
-                      const std::string& param_filename);
+                      const std::string& param_filename,
+                      bool model_from_memory);
 
 std::unique_ptr<framework::ProgramDesc> Load(framework::Executor* executor,
                                              framework::Scope* scope,
@@ -41,6 +42,10 @@ std::unique_ptr<framework::ProgramDesc> Load(framework::Executor* executor,
                                              const std::string& prog_filename,
                                              const std::string& param_filename);
 
+std::unique_ptr<framework::ProgramDesc> LoadFromMemory(
+    framework::Executor* executor, framework::Scope* scope,
+    const std::string& prog_buffer, const std::string& param_buffer);
+
 // Save the variables from a scope to disk.
 void SaveVars(const framework::Scope& scope,
               const std::vector<std::string>& vars, const std::string& dirname,

paddle/fluid/inference/tensorrt/convert/test_prelu_op.cc

@@ -90,5 +90,4 @@ TEST(prelu_op, test_scalar) {
 }  // namespace inference
 }  // namespace paddle
 
-// USE_OP(prelu);
-USE_CPU_ONLY_OP(prelu);
+USE_OP(prelu);

paddle/fluid/inference/tensorrt/plugin/CMakeLists.txt

 nv_library(tensorrt_plugin
            SRCS trt_plugin.cc split_op_plugin.cu elementwise_op_plugin.cu prelu_op_plugin.cu
            avg_pool_op_plugin.cu
-           DEPS enforce tensorrt_engine)
+           DEPS enforce tensorrt_engine prelu)

paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.cu

@@ -14,92 +14,16 @@
 
 #include <stdio.h>
 #include <cassert>
+#include <vector>
 #include "glog/logging.h"
 #include "paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.h"
+#include "paddle/fluid/operators/math/prelu.h"
 
 namespace paddle {
 namespace inference {
 namespace tensorrt {
 namespace plugin {
 
-static const int CUDA_NUM_THREADS = 1024;
-static const int CUDA_MAX_NUM_BLOCKS = 65535;
-inline static int GET_NUM_BLOCKS(const int N) {
-  return (N + CUDA_NUM_THREADS - 1) / CUDA_NUM_THREADS;
-}
-
-__global__ void PReluChannelWiseKernel(const float *input, const float *alpha,
-                                       float *output, int channel,
-                                       size_t spatial_size) {
-  size_t offset = blockIdx.x * spatial_size;
-  const float *in = input + offset;
-  float *out = output + offset;
-  float scale = alpha[blockIdx.x % channel];
-
-  for (size_t i = threadIdx.x; i < spatial_size; i += blockDim.x) {
-    float x = in[i];
-    out[i] = (x > 0) ? x : scale * x;
-  }
-}
-
-__global__ void PReluElementWiseKernel(const float *input, const float *alpha,
-                                       float *output, size_t spatial_size) {
-  size_t offset = blockIdx.x * spatial_size;
-  const float *in = input + offset;
-  const float *scale = alpha + offset;
-  float *out = output + offset;
-
-  for (size_t i = threadIdx.x; i < spatial_size; i += blockDim.x) {
-    float x = in[i];
-    out[i] = (x > 0) ? x : scale[i] * x;
-  }
-}
-
-__global__ void PReluScalarKernel(const float *input, const float *alpha,
-                                  float *output, size_t spatial_size) {
-  size_t offset = blockIdx.x * spatial_size;
-  const float *in = input + offset;
-  float scale = *alpha;
-  float *out = output + offset;
-
-  for (size_t i = threadIdx.x; i < spatial_size; i += blockDim.x) {
-    float x = in[i];
-    out[i] = (x > 0) ? x : scale * x;
-  }
-}
-
-static inline void PReluChannelWise(cudaStream_t stream, const float *input,
-                                    const float *alpha, float *output,
-                                    int batch_size,
-                                    const nvinfer1::Dims &dims) {
-  size_t unroll = batch_size * dims.d[0];
-  size_t spatial_size = dims.d[1] * dims.d[2];
-  CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS);
-  PReluChannelWiseKernel<<<unroll, CUDA_NUM_THREADS, 0, stream>>>(
-      input, alpha, output, dims.d[0], spatial_size);
-}
-
-static inline void PReluElementWise(cudaStream_t stream, const float *input,
-                                    const float *alpha, float *output,
-                                    int batch_size,
-                                    const nvinfer1::Dims &dims) {
-  size_t unroll = batch_size * dims.d[0];
-  size_t spatial_size = dims.d[1] * dims.d[2];
-  CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS);
-  PReluElementWiseKernel<<<unroll, CUDA_NUM_THREADS, 0, stream>>>(
-      input, alpha, output, spatial_size);
-}
-
-static inline void PReluScalar(cudaStream_t stream, const float *input,
-                               const float *alpha, float *output,
-                               int batch_size, const nvinfer1::Dims &dims) {
-  size_t unroll = batch_size * dims.d[0];
-  size_t spatial_size = dims.d[1] * dims.d[2];
-  CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS);
-  PReluScalarKernel<<<unroll, CUDA_NUM_THREADS, 0, stream>>>(
-      input, alpha, output, spatial_size);
-}
-
 nvinfer1::Dims PReluPlugin::getOutputDimensions(int index,
                                                 const nvinfer1::Dims *inputDims,
                                                 int nbInputs) {
@@ -110,19 +34,31 @@ nvinfer1::Dims PReluPlugin::getOutputDimensions(int index,
   return output_dims;
 }
 
-int PReluPlugin::enqueue(int batchSize, const void *const *inputs,
+int PReluPlugin::enqueue(int batch_size, const void *const *inputs,
                          void **outputs, void *workspace, cudaStream_t stream) {
   // input dims is CHW.
   const auto &input_dims = this->getInputDims(0);
   const float *input = reinterpret_cast<const float *>(inputs[0]);
   const float *alpha = reinterpret_cast<const float *>(alpha_.get().values);
   float *output = reinterpret_cast<float **>(outputs)[0];
+
+  std::vector<int> input_shape;
+  input_shape.push_back(batch_size);
+  for (int i = 0; i < input_dims.nbDims; i++) {
+    input_shape.push_back(input_dims.d[i]);
+  }
+
   if (mode_ == "channel") {
-    PReluChannelWise(stream, input, alpha, output, batchSize, input_dims);
+    operators::math::PreluChannelWiseDirectCUDAFunctor<float>
+        prelu_channel_wise;
+    prelu_channel_wise(stream, input, alpha, output, input_shape);
   } else if (mode_ == "element") {
-    PReluElementWise(stream, input, alpha, output, batchSize, input_dims);
+    operators::math::PreluElementWiseDirectCUDAFunctor<float>
+        prelu_element_wise;
+    prelu_element_wise(stream, input, alpha, output, input_shape);
   } else {
-    PReluScalar(stream, input, alpha, output, batchSize, input_dims);
+    operators::math::PreluScalarDirectCUDAFunctor<float> prelu_scalar;
+    prelu_scalar(stream, input, alpha, output, input_shape);
   }
   return cudaGetLastError() != cudaSuccess;
 }

paddle/fluid/inference/tests/api/analyzer_dam_tester.cc

@@ -188,10 +188,16 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 }
 
 // Easy for profiling independently.
-TEST(Analyzer_dam, profile) {
+void profile(bool use_mkldnn = false) {
   contrib::AnalysisConfig cfg;
   SetConfig(&cfg);
 
+  if (use_mkldnn) {
+    cfg.EnableMKLDNN();
+    std::unordered_set<std::string> op_list = {"conv3d"};
+    cfg.SetMKLDNNOp(op_list);
+  }
+
   std::vector<PaddleTensor> outputs;
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
@@ -209,6 +215,11 @@ TEST(Analyzer_dam, profile) {
   }
 }
 
+TEST(Analyzer_dam, profile) { profile(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_dam, profile_mkldnn) { profile(true /* use_mkldnn */); }
+#endif
+
 // Check the fuse status
 TEST(Analyzer_dam, fuse_statis) {
   contrib::AnalysisConfig cfg;
@@ -222,9 +233,14 @@ TEST(Analyzer_dam, fuse_statis) {
 }
 
 // Compare result of NativeConfig and AnalysisConfig
-TEST(Analyzer_dam, compare) {
-  contrib::AnalysisConfig cfg;
+void compare(bool use_mkldnn = false) {
+  AnalysisConfig cfg;
   SetConfig(&cfg);
+  if (use_mkldnn) {
+    cfg.EnableMKLDNN();
+    std::unordered_set<std::string> op_list = {"conv3d"};
+    cfg.SetMKLDNNOp(op_list);
+  }
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
@@ -233,5 +249,10 @@ TEST(Analyzer_dam, compare) {
       reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
 }
 
+TEST(Analyzer_dam, compare) { compare(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_dam, compare_mkldnn) { compare(true /* use_mkldnn */); }
+#endif
+
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tests/api/analyzer_ner_tester.cc

@@ -93,9 +93,17 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
   }
 }
 
-void SetConfig(contrib::AnalysisConfig *cfg) {
+void SetConfig(contrib::AnalysisConfig *cfg, bool memory_load = false) {
+  if (memory_load) {
+    std::string buffer_prog, buffer_param;
+    ReadBinaryFile(FLAGS_infer_model + "/__model__", &buffer_prog);
+    ReadBinaryFile(FLAGS_infer_model + "/param", &buffer_param);
+    cfg->SetModelBuffer(&buffer_prog[0], buffer_prog.size(), &buffer_param[0],
+                        buffer_param.size());
+  } else {
     cfg->prog_file = FLAGS_infer_model + "/__model__";
     cfg->param_file = FLAGS_infer_model + "/param";
+  }
   cfg->use_gpu = false;
   cfg->device = 0;
   cfg->specify_input_name = true;
@@ -114,9 +122,9 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 }
 
 // Easy for profiling independently.
-TEST(Analyzer_Chinese_ner, profile) {
+void profile(bool memory_load = false) {
   contrib::AnalysisConfig cfg;
-  SetConfig(&cfg);
+  SetConfig(&cfg, memory_load);
   std::vector<PaddleTensor> outputs;
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
@@ -138,6 +146,12 @@ TEST(Analyzer_Chinese_ner, profile) {
   }
 }
 
+TEST(Analyzer_Chinese_ner, profile) { profile(); }
+
+TEST(Analyzer_Chinese_ner, profile_memory_load) {
+  profile(true /* memory_load */);
+}
+
 // Check the fuse status
 TEST(Analyzer_Chinese_ner, fuse_statis) {
   contrib::AnalysisConfig cfg;

paddle/fluid/inference/tests/api/config_printer.h

@@ -49,8 +49,6 @@ std::ostream &operator<<(std::ostream &os, const NativeConfig &config) {
   os << GenSpaces(num_spaces) << "device: " << config.device << "\n";
   os << GenSpaces(num_spaces)
      << "fraction_of_gpu_memory: " << config.fraction_of_gpu_memory << "\n";
-  os << GenSpaces(num_spaces) << "prog_file: " << config.prog_file << "\n";
-  os << GenSpaces(num_spaces) << "param_file: " << config.param_file << "\n";
   os << GenSpaces(num_spaces)
      << "specify_input_name: " << config.specify_input_name << "\n";
   os << GenSpaces(num_spaces)
@@ -65,6 +63,13 @@ std::ostream &operator<<(std::ostream &os,
   os << GenSpaces(num_spaces) << "contrib::AnalysisConfig {\n";
   num_spaces++;
   os << *reinterpret_cast<const NativeConfig *>(&config);
+  if (!config.model_from_memory()) {
+    os << GenSpaces(num_spaces) << "prog_file: " << config.prog_file << "\n";
+    os << GenSpaces(num_spaces) << "param_file: " << config.param_file << "\n";
+  } else {
+    os << GenSpaces(num_spaces)
+       << "prog_file and param_file: load from memory \n";
+  }
   os << GenSpaces(num_spaces) << "enable_ir_optim: " << config.enable_ir_optim
      << "\n";
   os << GenSpaces(num_spaces)

paddle/fluid/inference/utils/CMakeLists.txt

 cc_library(benchmark SRCS benchmark.cc DEPS enforce)
 cc_test(test_benchmark SRCS benchmark_tester.cc DEPS benchmark)
+cc_binary(visualizer SRCS visualizer.cc DEPS analysis
+    paddle_pass_builder ir_pass_manager pass graph_viz_pass analysis_passes)
+if(WIN32)
+  target_link_libraries(visualizer shlwapi)
+endif(WIN32)

paddle/fluid/inference/utils/visualizer.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.
+
+#include "paddle/fluid/inference/utils/visualizer.h"
+#include <gflags/gflags.h>
+#include <glog/logging.h>
+#include <fstream>
+#include <memory>
+#include "paddle/fluid/framework/ir/graph_viz_pass.h"
+#include "paddle/fluid/inference/analysis/analyzer.h"
+#include "paddle/fluid/inference/analysis/passes/ir_analysis_pass.h"
+#include "paddle/fluid/platform/init.h"
+
+DEFINE_string(model_dir, "", "model directory");
+DEFINE_string(model_program_path, "", "model program path");
+DEFINE_string(model_params_path, "", "model params path");
+
+USE_PASS(graph_viz_pass);
+USE_PASS(graph_to_program_pass);
+
+using paddle::inference::analysis::Argument;
+
+namespace paddle {
+namespace inference {
+namespace utils {
+
+void Visualizer::SetArgument(Argument *argument) { argument_ = argument; }
+
+bool Visualizer::Run() {
+  paddle::framework::InitDevices(false);
+  paddle::inference::analysis::Analyzer().Run(argument_);
+
+  return true;
+}
+
+}  // namespace utils
+}  // namespace inference
+}  // namespace paddle
+
+// Generate a dot file describing the structure of graph.
+// To use this tool, run command: ./visualizer [options...]
+// Options:
+//     --model_dir: the directory of model
+//     --model_program_path: the path of program
+//     --model_params_path: the path of params
+int main(int argc, char *argv[]) {
+  gflags::ParseCommandLineFlags(&argc, &argv, true);
+  google::InitGoogleLogging(argv[0]);
+
+  paddle::inference::analysis::Argument argument;
+  argument.SetUseGPU(false);
+  argument.SetUseTensorRT(false);
+
+  if (FLAGS_model_dir.empty()) {
+    if (FLAGS_model_program_path.empty() || FLAGS_model_params_path.empty()) {
+      LOG(ERROR) << "Please set model_dir"
+                    " or model_program_path and model_params_path";
+      return -1;
+    } else {
+      argument.SetModelProgramPath(FLAGS_model_program_path);
+      argument.SetModelParamsPath(FLAGS_model_params_path);
+    }
+  } else {
+    argument.SetModelDir(FLAGS_model_dir);
+  }
+
+  // Only 1 pass, default filename is 0_ir_origin.dot
+  // For more details, looking for paddle::inference::analysis::IRPassManager
+  argument.SetIrAnalysisPasses({"graph_viz_pass"});
+
+  std::unique_ptr<paddle::framework::Scope> scope{
+      new paddle::framework::Scope()};
+  argument.SetScopeNotOwned(
+      const_cast<paddle::framework::Scope *>(scope.get()));
+
+  paddle::inference::utils::Visualizer visualizer;
+  visualizer.SetArgument(&argument);
+  visualizer.Run();
+
+  return 0;
+}

paddle/fluid/inference/utils/visualizer.h

+// 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.
+
+#pragma once
+
+#include <string>
+#include "paddle/fluid/inference/analysis/argument.h"
+
+namespace paddle {
+namespace inference {
+namespace utils {
+
+using paddle::inference::analysis::Argument;
+
+class Visualizer final {
+ public:
+  Visualizer() = default;
+  ~Visualizer() = default;
+  Visualizer(const Visualizer &) = delete;
+  Visualizer &operator=(const Visualizer &) = delete;
+
+  void SetArgument(Argument *);
+  bool Run();
+
+ private:
+  Argument *argument_;
+};
+
+}  // namespace utils
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/operators/CMakeLists.txt

@@ -70,7 +70,7 @@ endif()
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel concat_and_split cross_entropy softmax vol2col im2col sampler)
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions)
 if (WITH_GPU)
-  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv)
+  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv prelu)
 endif()
 
 # FIXME(typhoonzero): operator deps may not needed.

paddle/fluid/operators/activation_mkldnn_op.cc

@@ -100,8 +100,9 @@ void eltwise_forward(const framework::ExecutionContext &ctx,
   const T *x_data = x->data<T>();
   T *y_data = y->mutable_data<T>(ctx.GetPlace());
 
-  PADDLE_ENFORCE(x->dims().size() == 2 || x->dims().size() == 4,
-                 "Input dim must be with 2 or 4");
+  PADDLE_ENFORCE(
+      x->dims().size() == 2 || x->dims().size() == 3 || x->dims().size() == 4,
+      "Input dim must be with 2, 3 or 4");
 
   std::vector<int> src_tz = framework::vectorize2int(x->dims());
 

paddle/fluid/operators/activation_op.cc

@@ -76,8 +76,8 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx,
   }
 #endif
   return framework::OpKernelType(
-      framework::ToDataType(ctx.Input<framework::Tensor>(name)->type()),
-      ctx.GetPlace(), layout, library);
+      framework::GetDataTypeOfVar(ctx.InputVar(name)), ctx.GetPlace(), layout,
+      library);
 }
 
 class ActivationOp : public framework::OperatorWithKernel {

paddle/fluid/operators/activation_op.h

@@ -41,6 +41,12 @@ static std::unordered_set<std::string> InplaceOpSet = {
     "floor",   "reciprocal", "relu6", "soft_relu", "hard_sigmoid",
 };
 
+/* The following operator can be used to process SelectedRows, because the
+ * output of those operator for zero is zero too.
+ */
+static std::unordered_set<std::string> CanBeUsedBySelectedRows = {
+    "abs", "abs_grad", "square", "square_grad", "sqrt", "sqrt_grad"};
+
 static bool IsInplace(std::string op) { return InplaceOpSet.count(op); }
 
 template <typename DeviceContext, typename Functor>
@@ -50,16 +56,38 @@ class ActivationKernel
   using T = typename Functor::ELEMENT_TYPE;
 
   void Compute(const framework::ExecutionContext& context) const override {
-    auto& X = detail::Ref(context.Input<framework::Tensor>("X"),
-                          "Cannot get input tensor X, variable name = %s",
+    auto x_var = context.InputVar("X");
+    auto out_var = context.OutputVar("Out");
+    PADDLE_ENFORCE(x_var != nullptr,
+                   "Cannot get input Variable X, variable name = %s",
                    context.op().Input("X"));
+    PADDLE_ENFORCE(out_var != nullptr,
+                   "Cannot get output Variable Out, variable name = %s",
+                   context.op().Output("Out"));
+
+    framework::Tensor X, *Out;
 
-    auto& Out = detail::Ref(context.Output<framework::Tensor>("Out"),
+    if (CanBeUsedBySelectedRows.count(context.op().Type())) {
+      X = detail::Ref(
+          paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_var),
+          "Cannot get input Tensor X, variable name = %s",
+          context.op().Input("X"));
+      Out = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
+          out_var);
+    } else {
+      X = detail::Ref(context.Input<framework::Tensor>("X"),
+                      "Cannot get input Tensor X, variable name = %s",
+                      context.op().Input("X"));
+      Out = context.Output<framework::Tensor>("Out");
+    }
+
+    PADDLE_ENFORCE(Out != nullptr,
                    "Cannot get output tensor Out, variable name = %s",
                    context.op().Output("Out"));
-    Out.mutable_data<T>(context.GetPlace());
+
+    Out->mutable_data<T>(context.GetPlace());
     auto x = framework::EigenVector<T>::Flatten(X);
-    auto out = framework::EigenVector<T>::Flatten(Out);
+    auto out = framework::EigenVector<T>::Flatten(*Out);
     auto* place =
         context.template device_context<DeviceContext>().eigen_device();
     Functor functor;
@@ -78,14 +106,54 @@ class ActivationGradKernel
  public:
   using T = typename Functor::ELEMENT_TYPE;
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* Out = context.Input<framework::Tensor>("Out");
-    auto* dOut =
-        context.Input<framework::Tensor>(framework::GradVarName("Out"));
-    auto* dX = context.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto out_var = context.InputVar("Out");
+    auto out_grad_var = context.InputVar(framework::GradVarName("Out"));
+    auto x_grad_var = context.OutputVar(framework::GradVarName("X"));
+    PADDLE_ENFORCE(out_var != nullptr,
+                   "Cannot get input Variable Out, variable name = %s",
+                   context.op().Input("Out"));
+    PADDLE_ENFORCE(out_grad_var != nullptr,
+                   "Cannot get input Variable %s, variable name = %s",
+                   framework::GradVarName("Out"),
+                   context.op().Input(framework::GradVarName("Out")));
+    PADDLE_ENFORCE(x_grad_var != nullptr,
+                   "Cannot get output Variable %s, variable name = %s",
+                   framework::GradVarName("X"),
+                   context.op().Output(framework::GradVarName("X")));
+
+    framework::Tensor Out, dOut, *dX;
+    if (CanBeUsedBySelectedRows.count(context.op().Type())) {
+      Out = detail::Ref(
+          paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*out_var),
+          "Cannot get input Tensor Out, variable name = %s",
+          context.op().Input("Out"));
+      dOut =
+          detail::Ref(paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(
+                          *out_grad_var),
+                      "Cannot get input Tensor %s, variable name = %s",
+                      framework::GradVarName("Out"),
+                      context.op().Input(framework::GradVarName("Out")));
+      dX = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
+          x_grad_var);
+    } else {
+      Out = detail::Ref(context.Input<framework::Tensor>("Out"),
+                        "Cannot get input Tensor Out, variable name = %s",
+                        context.op().Input("Out"));
+      dOut = detail::Ref(
+          context.Input<framework::Tensor>(framework::GradVarName("Out")),
+          "Cannot get input Tensor %s, variable name = %s",
+          framework::GradVarName("Out"),
+          context.op().Input(framework::GradVarName("Out")));
+      dX = context.Output<framework::Tensor>(framework::GradVarName("X"));
+    }
+    PADDLE_ENFORCE(dX != nullptr,
+                   "Cannot get output tensor %s, variable name = %s",
+                   framework::GradVarName("X"),
+                   context.op().Output(framework::GradVarName("X")));
     dX->mutable_data<T>(context.GetPlace());
 
-    auto dout = framework::EigenVector<T>::Flatten(*dOut);
-    auto out = framework::EigenVector<T>::Flatten(*Out);
+    auto dout = framework::EigenVector<T>::Flatten(dOut);
+    auto out = framework::EigenVector<T>::Flatten(Out);
     auto dx = framework::EigenVector<T>::Flatten(*dX);
     auto* place =
         context.template device_context<DeviceContext>().eigen_device();
@@ -96,8 +164,19 @@ class ActivationGradKernel
     }
     bool inplace = functor.Inplace();
     if (!inplace) {
-      auto* X = context.Input<framework::Tensor>("X");
-      auto x = framework::EigenVector<T>::Flatten(*X);
+      auto x_var = context.InputVar("X");
+      PADDLE_ENFORCE(x_var != nullptr,
+                     "Cannot get input tensor X, variable name = %s",
+                     context.op().Input("X"));
+      framework::Tensor X;
+      if (CanBeUsedBySelectedRows.count(context.op().Type())) {
+        X = detail::Ref(
+            paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_var));
+      } else {
+        X = detail::Ref(context.Input<framework::Tensor>("X"));
+      }
+
+      auto x = framework::EigenVector<T>::Flatten(X);
       functor(*place, x, out, dout, dx);
     } else {
       VLOG(10) << " Inplace activation ";

paddle/fluid/operators/attention_lstm_op.cc

@@ -231,10 +231,10 @@ use lstm_x_t as input and compute as standard LSTM.
 template <typename T>
 inline void bias_relu(const int n, const T* x, const T* bias, T* y) {
   if (bias) {
-    math::vec_add_bias<T, platform::jit::avx>(n, *bias, x, y);
-    math::vec_relu<T, platform::jit::avx>(n, y, y);
+    math::vec_add_bias<T, platform::avx>(n, *bias, x, y);
+    math::vec_relu<T, platform::avx>(n, y, y);
   } else {
-    math::vec_relu<T, platform::jit::avx>(n, x, y);
+    math::vec_relu<T, platform::avx>(n, x, y);
   }
 }
 
@@ -245,7 +245,7 @@ inline void vec_softmax(const int n, const T* x, T* y) {
   for (int i = 1; i < n; ++i) {
     scalar = scalar < x[i] ? x[i] : scalar;
   }
-  math::vec_add_bias<T, platform::jit::avx>(n, -scalar, x, y);  // sub
+  math::vec_add_bias<T, platform::avx>(n, -scalar, x, y);  // sub
   math::vec_exp<T>(n, y, y);                               // exp
   // sum
   scalar = T(0);
@@ -302,13 +302,13 @@ class AttentionLSTMKernel : public framework::OpKernel<T> {
     auto& act_gate_str = ctx.Attr<std::string>("gate_activation");
     auto& act_cell_str = ctx.Attr<std::string>("cell_activation");
     auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");
-    if (platform::jit::MayIUse(platform::jit::avx)) {
-      math::VecActivations<T, platform::jit::avx> act_functor;
+    if (platform::MayIUse(platform::avx)) {
+      math::VecActivations<T, platform::avx> act_functor;
       act_gate = act_functor(act_gate_str);
       act_cell = act_functor(act_cell_str);
       act_cand = act_functor(act_cand_str);
     } else {
-      math::VecActivations<T, platform::jit::isa_any> act_functor;
+      math::VecActivations<T, platform::isa_any> act_functor;
       act_gate = act_functor(act_gate_str);
       act_cell = act_functor(act_cell_str);
       act_cand = act_functor(act_cand_str);

paddle/fluid/operators/conv_fusion_op.cu.cc

@@ -110,11 +110,7 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
 
     auto x_dims = framework::vectorize(input->dims());
     auto f_dims = framework::vectorize(filter->dims());
-    if (activation == "identity") {
-      // Only the CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM algo is
-      // enabled with CUDNN_ACTIVATION_IDENTITY in cuDNN lib.
-      algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
-    } else if (!exhaustive_search) {
+    if (!exhaustive_search) {
       CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm(
           handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
           cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
@@ -165,6 +161,29 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_LE(workspace_size_in_bytes, workspace_size_limit,
                       "workspace_size to be allocated exceeds the limit");
 
+    if ((activation == "identity") &&
+        (algo != CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM) &&
+        (!residual)) {
+      // Only the CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM algo is
+      // enabled with CUDNN_ACTIVATION_IDENTITY in cuDNN lib.
+      // But test in some case, the speed is slower, change to use
+      // cudnnConvolutionForward and cudnnAddTensor
+      // ------------- cudnn conv forward and bias add ---------------------
+      ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
+      auto cudnn_func = [&](void* cudnn_workspace) {
+        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
+            handle, &alpha, cudnn_input_desc, input_data, cudnn_filter_desc,
+            filter_data, cudnn_conv_desc, algo, cudnn_workspace,
+            workspace_size_in_bytes, &beta, cudnn_output_desc, output_data));
+      };
+      workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
+      CUDNN_ENFORCE(platform::dynload::cudnnAddTensor(
+          handle, &alpha, cudnn_bias_desc, bias_data, &alpha, cudnn_output_desc,
+          output_data));
+    } else {
+      if (activation == "identity") {
+        algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
+      }
       // ------------------- cudnn conv+bias+act forward --------------------
       ScalingParamType<T> alpha1 = 1.0f;
       ScalingParamType<T> alpha2 = residual ? 1.0f : 0.0f;
@@ -178,6 +197,7 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
       };
       workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
     }
+  }
 };
 #endif
 

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -28,6 +28,46 @@ using mkldnn::stream;
 using platform::to_void_cast;
 using platform::GetMKLDNNFormat;
 
+inline void GetWeightsTz(std::vector<int>& weights_tz, int groups,  // NOLINT
+                         bool is_conv3d) {
+  if (groups > 1) {
+    if (is_conv3d) {
+      int output = weights_tz[0];
+      int input = weights_tz[1];
+      int dimension = weights_tz[2];
+      int height = weights_tz[3];
+      int width = weights_tz[4];
+      weights_tz.resize(6);
+      weights_tz[0] = groups;
+      weights_tz[1] = output / groups;
+      weights_tz[2] = input;
+      weights_tz[3] = dimension;
+      weights_tz[4] = height;
+      weights_tz[5] = width;
+    } else {
+      int output = weights_tz[0];
+      int input = weights_tz[1];
+      int height = weights_tz[2];
+      int width = weights_tz[3];
+      weights_tz.resize(5);
+      weights_tz[0] = groups;
+      weights_tz[1] = output / groups;
+      weights_tz[2] = input;
+      weights_tz[3] = height;
+      weights_tz[4] = width;
+    }
+  }
+}
+
+inline mkldnn::memory::format GetWeightsFormat(mkldnn::memory::format format,
+                                               int groups, bool is_conv3d) {
+  if (is_conv3d) {
+    return (groups == 1) ? format : mkldnn::memory::format::goidhw;
+  } else {
+    return (groups == 1) ? format : mkldnn::memory::format::goihw;
+  }
+}
+
 template <typename T>
 class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
  public:
@@ -52,10 +92,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN &&
                        filter->format() != memory::format::format_undef,
                    "Wrong layout/format set for Filter tensor");
-    PADDLE_ENFORCE(input->dims().size() == 4,
-                   "Input must be with 4 dimensions, i.e. NCHW");
-    PADDLE_ENFORCE(filter->dims().size() == 4,
-                   "Filter must be with 4 dimensions, i.e. OIHW");
+    PADDLE_ENFORCE(input->dims().size() == 4 || input->dims().size() == 5,
+                   "Input must be with 4 or 5 dimensions, i.e. NCHW or NCDHW");
+    PADDLE_ENFORCE(filter->dims().size() == 4 || filter->dims().size() == 5,
+                   "Filter must be with 4 or 5 dimensions, i.e. OIHW or OIDHW");
     if (bias) {
       PADDLE_ENFORCE(bias->layout() == DataLayout::kMKLDNN &&
                          bias->format() != memory::format::format_undef,
@@ -71,9 +111,13 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     bool fuse_residual_conn = ctx.Attr<bool>("fuse_residual_connection");
     int groups = ctx.Attr<int>("groups");
 
+    bool is_conv3d = strides.size() == 3U;
     // TODO(tpatejko): add support for dilation
     PADDLE_ENFORCE(
-        dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1,
+        is_conv3d
+            ? dilations.size() == 3 && dilations[0] == 1 && dilations[1] == 1 &&
+                  dilations[2] == 1
+            : dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1,
         "dilation in convolution is not implemented yet");
 
     const T* input_data = input->data<T>();
@@ -83,18 +127,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     std::vector<int> weights_tz =
         paddle::framework::vectorize2int(filter->dims());
     int g = std::max(groups, 1);
-    if (g > 1) {
-      int o = weights_tz[0];
-      int i = weights_tz[1];
-      int h = weights_tz[2];
-      int w = weights_tz[3];
-      weights_tz.resize(5);
-      weights_tz[0] = g;
-      weights_tz[1] = o / g;
-      weights_tz[2] = i;
-      weights_tz[3] = h;
-      weights_tz[4] = w;
-    }
+    GetWeightsTz(weights_tz, g, is_conv3d);
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
 
     // Get unique name for storing MKLDNN primitives
@@ -105,11 +138,14 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 
     std::vector<primitive> pipeline;
 
+    auto src_format = input->format();
+    mkldnn::memory::format weights_format =
+        GetWeightsFormat(filter->format(), g, is_conv3d);
+
     auto user_src_md = platform::MKLDNNMemDesc(
-        {src_tz}, platform::MKLDNNGetDataType<T>(), input->format());
+        {src_tz}, platform::MKLDNNGetDataType<T>(), src_format);
     auto user_weights_md = platform::MKLDNNMemDesc(
-        {weights_tz}, platform::MKLDNNGetDataType<T>(),
-        (g == 1) ? filter->format() : mkldnn::memory::format::goihw);
+        {weights_tz}, platform::MKLDNNGetDataType<T>(), weights_format);
 
     /* create memory descriptor for convolution without specified format
      * ('any') which lets a primitive (convolution in this case) choose
@@ -119,10 +155,16 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto chosen_memory_format =
         platform::data_format_to_memory_format(data_format);
 
+    if (is_conv3d) {
+      chosen_memory_format =
+          platform::MKLDNNFormatForSize(src_tz.size(), chosen_memory_format);
+    }
+    weights_format = GetWeightsFormat(chosen_memory_format, g, is_conv3d);
+
     auto src_md = platform::MKLDNNMemDesc(
         src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto weights_md = platform::MKLDNNMemDesc(
-        weights_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
+        weights_tz, platform::MKLDNNGetDataType<T>(), weights_format);
     std::vector<int> bias_tz;  // TODO(mgallus): avoid empty vector creation.
                                // Currently used whenever bias is != nullptr.
     auto dst_md = platform::MKLDNNMemDesc(
@@ -263,8 +305,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                        const mkldnn::engine& engine, const bool fuse_relu,
                        const bool fuse_residual_conn,
                        mkldnn::prop_kind fwd_prop_kind) const {
-    memory::dims stride_dims = {strides[0], strides[1]};
-    memory::dims padding_dims = {paddings[0], paddings[1]};
+    memory::dims stride_dims = strides;
+    memory::dims padding_dims = paddings;
 
     auto conv_desc = mkldnn::convolution_forward::desc(
         fwd_prop_kind, mkldnn::convolution_direct, src, weights, dst,
@@ -288,8 +330,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                        const mkldnn::engine& engine, const bool fuse_relu,
                        const bool fuse_residual_conn,
                        mkldnn::prop_kind fwd_prop_kind) const {
-    memory::dims stride_dims = {strides[0], strides[1]};
-    memory::dims padding_dims = {paddings[0], paddings[1]};
+    memory::dims stride_dims = strides;
+    memory::dims padding_dims = paddings;
 
     auto conv_desc = mkldnn::convolution_forward::desc(
         fwd_prop_kind, mkldnn::convolution_direct, src, weights, bias, dst,
@@ -349,6 +391,7 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
     int groups = ctx.Attr<int>("groups");
 
+    bool is_conv3d = strides.size() == 3U;
     const T* input_data = input->data<T>();
     const T* filter_data = filter->data<T>();
     const T* output_grad_data = output_grad->data<T>();
@@ -358,8 +401,14 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
     std::vector<int> weights_tz =
         paddle::framework::vectorize2int(filter->dims());
+    int g = std::max(groups, 1);
+    GetWeightsTz(weights_tz, g, is_conv3d);
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
 
+    auto src_format = input->format();
+    mkldnn::memory::format weights_format =
+        GetWeightsFormat(filter->format(), g, is_conv3d);
+
     // Get an unique name from "argument" name of "Output" variable
     // as well as attributes of primitive to be created
     // This name will be used as key when saving info into device context
@@ -372,9 +421,9 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
 
     // Create user memory descriptors
     auto user_src_md = platform::MKLDNNMemDesc(
-        {src_tz}, platform::MKLDNNGetDataType<T>(), input->format());
+        {src_tz}, platform::MKLDNNGetDataType<T>(), src_format);
     auto user_weights_md = platform::MKLDNNMemDesc(
-        {weights_tz}, platform::MKLDNNGetDataType<T>(), filter->format());
+        {weights_tz}, platform::MKLDNNGetDataType<T>(), weights_format);
     auto user_diff_dst_md = platform::MKLDNNMemDesc(
         {dst_tz}, platform::MKLDNNGetDataType<T>(), output_grad->format());
 
@@ -386,14 +435,20 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     auto chosen_memory_format =
         platform::data_format_to_memory_format(data_format);
 
+    if (is_conv3d) {
+      chosen_memory_format =
+          platform::MKLDNNFormatForSize(src_tz.size(), chosen_memory_format);
+    }
+    weights_format = GetWeightsFormat(chosen_memory_format, g, is_conv3d);
+
     auto src_md = platform::MKLDNNMemDesc(
         src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto diff_src_md = platform::MKLDNNMemDesc(
         src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto weights_md = platform::MKLDNNMemDesc(
-        weights_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
+        weights_tz, platform::MKLDNNGetDataType<T>(), weights_format);
     auto diff_weights_md = platform::MKLDNNMemDesc(
-        weights_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
+        weights_tz, platform::MKLDNNGetDataType<T>(), weights_format);
     auto diff_dst_md = platform::MKLDNNMemDesc(
         dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
 
@@ -491,8 +546,22 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_KERNEL(conv2d, MKLDNN, ::paddle::platform::CPUPlace,
+REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN,
+                                    ::paddle::platform::CPUPlace, FP32,
+                                    ops::kConvMKLDNNFP32,
+                                    ops::ConvMKLDNNOpKernel<float>);
+
+REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv2d_grad, MKLDNN,
+                                    ::paddle::platform::CPUPlace, FP32,
+                                    ops::kConvMKLDNNFP32,
+                                    ops::ConvMKLDNNGradOpKernel<float>);
+
+REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv3d, MKLDNN,
+                                    ::paddle::platform::CPUPlace, FP32,
+                                    ops::kConvMKLDNNFP32,
                                     ops::ConvMKLDNNOpKernel<float>);
 
-REGISTER_OP_KERNEL(conv2d_grad, MKLDNN, ::paddle::platform::CPUPlace,
+REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv3d_grad, MKLDNN,
+                                    ::paddle::platform::CPUPlace, FP32,
+                                    ops::kConvMKLDNNFP32,
                                     ops::ConvMKLDNNGradOpKernel<float>);

paddle/fluid/operators/conv_op.cc

@@ -74,6 +74,8 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
 
 framework::OpKernelType ConvOp::GetExpectedKernelType(
     const framework::ExecutionContext& ctx) const {
+  int customized_type_value =
+      framework::OpKernelType::kDefaultCustomizedTypeValue;
   framework::LibraryType library{framework::LibraryType::kPlain};
   // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
   std::string data_format = ctx.Attr<std::string>("data_format");
@@ -89,6 +91,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
       platform::CanMKLDNNBeUsed(ctx)) {
     library = framework::LibraryType::kMKLDNN;
     layout = framework::DataLayout::kMKLDNN;
+    customized_type_value = kConvMKLDNNFP32;
   }
 #endif
 
@@ -105,7 +108,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
   }
 
   return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
-                                 library);
+                                 library, customized_type_value);
 }
 
 void Conv2DOpMaker::Make() {
@@ -131,14 +134,14 @@ void Conv2DOpMaker::Make() {
            "The format of output tensor is X (one-dimensional) of size equal"
            "to the number of output channels. Only used with MKL-DNN.")
       .AsDispensable();
-  AddOutput("Output",
-            "(Tensor) The output tensor of convolution operator. "
-            "The format of output tensor is also NCHW.");
   AddInput("ResidualData",
            "(Tensor) Tensor with residual data "
            "to which convolution output will be added."
            "Used with fuse_residual_connection fusion.")
       .AsDispensable();
+  AddOutput("Output",
+            "(Tensor) The output tensor of convolution operator. "
+            "The format of output tensor is also NCHW.");
   AddAttr<std::vector<int>>("strides",
                             "(vector<int> default:{1, 1}), the "
                             "strides(h_stride, w_stride) of "
@@ -229,6 +232,10 @@ $$
 }
 
 void Conv3DOpMaker::Make() {
+  AddAttr<bool>("is_test",
+                "(bool, default false) Set to true for inference only, false "
+                "for training. Some layers may run faster when this is true.")
+      .SetDefault(false);
   AddInput(
       "Input",
       "(Tensor) The input tensor of convolution operator. "
@@ -244,6 +251,11 @@ void Conv3DOpMaker::Make() {
            "is the width of the filter."
            "If the groups attribute is greater than 1, C equals the number of "
            "input image channels divided by the groups.");
+  AddInput("ResidualData",
+           "(Tensor) Tensor with residual data "
+           "to which convolution output will be added."
+           "Used with fuse_residual_connection fusion.")
+      .AsDispensable();
   AddOutput("Output",
             "(Tensor) The output tensor of convolution operator."
             "The format of output tensor is also NCDHW.");
@@ -277,6 +289,13 @@ void Conv3DOpMaker::Make() {
   AddAttr<bool>("use_mkldnn",
                 "(bool, default false) Only used in mkldnn kernel")
       .SetDefault(false);
+  AddAttr<bool>("fuse_relu", "(bool, default false) Only used in mkldnn kernel")
+      .SetDefault(false);
+  AddAttr<bool>("fuse_residual_connection",
+                "(bool, default false) Only used in mkldnn kernel. Used "
+                "whenever convolution output is as an input to residual "
+                "connection.")
+      .SetDefault(false);
   AddAttr<std::string>(
       "data_format",
       "(string, default NCHW) Only used in "
@@ -342,6 +361,8 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const {
 
 framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
     const framework::ExecutionContext& ctx) const {
+  int customized_type_value =
+      framework::OpKernelType::kDefaultCustomizedTypeValue;
   framework::LibraryType library_{framework::LibraryType::kPlain};
   // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
   std::string data_format = ctx.Attr<std::string>("data_format");
@@ -357,12 +378,13 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
       platform::CanMKLDNNBeUsed(ctx)) {
     library_ = framework::LibraryType::kMKLDNN;
     layout_ = framework::DataLayout::kMKLDNN;
+    customized_type_value = kConvMKLDNNFP32;
   }
 #endif
 
   return framework::OpKernelType(
       framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
-      layout_, library_);
+      layout_, library_, customized_type_value);
 }
 
 }  // namespace operators

paddle/fluid/operators/conv_op.h

@@ -27,6 +27,8 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+constexpr int kConvMKLDNNFP32 = 1;
+constexpr int kConvMKLDNNINT8 = 2;
 
 // Base convolution operator definations for other conv
 // like operators to reuse the implementation.

paddle/fluid/operators/cudnn_lstm_op.cu.cc

@@ -177,11 +177,19 @@ struct CudnnRNNCache {
         seed_));
 
     CUDNN_ENFORCE(platform::dynload::cudnnCreateRNNDescriptor(&rnn_desc_));
+
+#if CUDNN_VERSION >= 6000
     CUDNN_ENFORCE(platform::dynload::cudnnSetRNNDescriptor_v6(
         handle, rnn_desc_, hidden_size_, num_layers_, dropout_desc_,
         CUDNN_LINEAR_INPUT,
         is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, CUDNN_LSTM,
         CUDNN_RNN_ALGO_STANDARD, CUDNN_DATA_FLOAT));
+#else
+    CUDNN_ENFORCE(platform::dynload::cudnnSetRNNDescriptor(
+        rnn_desc_, hidden_size_, num_layers_, dropout_desc_, CUDNN_LINEAR_INPUT,
+        is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, CUDNN_LSTM,
+        CUDNN_DATA_FLOAT));
+#endif
 
     CUDNN_ENFORCE(platform::dynload::cudnnCreateFilterDescriptor(&w_desc_));
     CUDNN_ENFORCE(platform::dynload::cudnnCreateFilterDescriptor(&dw_desc_));

paddle/fluid/operators/distributed/CMakeLists.txt

@@ -13,16 +13,26 @@ set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor
 
 if(WITH_GRPC)
   grpc_library(sendrecvop_grpc SRCS grpc_bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc
-        request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc grpc_variable_response.cc grpc_serde.cc
+        request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc grpc_variable_response.cc grpc_serde.cc collective_client.cc collective_server.cc
       PROTO send_recv.proto 
-      DEPS lod_tensor selected_rows memory)
+      DEPS lod_tensor selected_rows_functor memory)
 
   set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+
   cc_test(grpc_serde_test SRCS grpc_serde_test.cc 
     DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL)
+
   cc_test(rpc_server_test SRCS rpc_server_test.cc
     DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor  proto_desc lookup_sparse_table_op SERIAL)
+
   cc_test(varhandle_test SRCS varhandle_test.cc DEPS profiler)
+
+  if(WITH_GPU)
+  cc_test(collective_server_test SRCS collective_server_test.cc 
+      DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor
+      selected_rows_functor  scope math_function SERIAL)
+  endif()
+
   cc_library(parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_grpc memory)
 else()
   set_source_files_properties(brpc_server.cc brpc_client.cc rpc_server_test.cc brpc_serde_test.cc

paddle/fluid/operators/distributed/collective_client.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.
+
+#include <condition_variable>  // NOLINT
+#include <string>
+#include "gflags/gflags.h"
+
+#include "paddle/fluid/operators/distributed/collective_client.h"
+
+DECLARE_int32(rpc_deadline);
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+std::once_flag CollectiveClient::init_flag_;
+std::unique_ptr<CollectiveClient> CollectiveClient::client_(nullptr);
+
+bool CollectiveClient::Gather(const std::vector<RemoteVar>& remote_vars,
+                              std::vector<const framework::SelectedRows*>* dst,
+                              const platform::DeviceContext& ctx,
+                              framework::Scope* scope, int64_t time_out) {
+  for (auto r : remote_vars) {
+    VLOG(50) << "begin gather from ep:" << r.String();
+    scope->Var(r.var_name_)->GetMutable<framework::SelectedRows>();
+    VarHandlePtr ptr = rpc_client_->AsyncGetMonomerVariable(
+        r.ep_, ctx, *scope, r.var_name_, time_out);
+  }
+
+  rpc_client_->Wait();
+
+  for (auto r : remote_vars) {
+    auto select_rows =
+        scope->FindVar(r.var_name_)->GetMutable<framework::SelectedRows>();
+    dst->push_back(select_rows);
+
+    VLOG(4) << "gather from ep:" << r.String()
+            << ", select_rows:" << GetSelectedRowsInfo(*select_rows);
+
+    rpc_client_->AsyncGetMonomerBarrier(r.ep_, r.var_name_);
+  }
+
+  rpc_client_->Wait();
+  return true;
+}
+
+}  // namespace distributed
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/distributed/collective_client.h

+// 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.
+
+#pragma once
+
+#include <condition_variable>  // NOLINT
+#include <string>
+#include <vector>
+#include "gflags/gflags.h"
+
+#include "paddle/fluid/framework/data_type.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/operators/detail/macros.h"
+#include "paddle/fluid/operators/distributed/request_handler.h"
+
+DECLARE_int32(rpc_deadline);
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+
+inline std::string GetSelectedRowsInfo(const framework::SelectedRows& slr) {
+  std::stringstream ss;
+  ss << ", height:" << slr.height() << ", rows:[";
+  for (unsigned int i = 0; i < slr.rows().size(); i++) {
+    if (i != slr.rows().size() - 1) {
+      ss << slr.rows()[i] << ",";
+    } else {
+      ss << slr.rows()[i];
+    }
+  }
+  ss << "], dims:" << slr.value().dims();
+  return ss.str();
+}
+
+struct RemoteVar {
+  std::string ep_;
+  std::string var_name_;
+  int trainer_id_{0};
+
+  std::string String() {
+    std::stringstream ss;
+    ss << "ep:" << ep_ << ", var_name:" << var_name_
+       << ", trainer_id:" << trainer_id_;
+
+    return ss.str();
+  }
+};
+
+class CollectiveClient {
+ public:
+  CollectiveClient() {
+    rpc_client_.reset(new RPCCLIENT_T());
+    rpc_client_->InitImpl();
+  }
+  virtual ~CollectiveClient() {}
+
+  // note this function will retain the rank order.
+  bool Gather(const std::vector<RemoteVar>& remote_vars,
+              std::vector<const framework::SelectedRows*>* dst,
+              const platform::DeviceContext& ctx, framework::Scope* scope,
+              int64_t time_out = FLAGS_rpc_deadline);
+
+  static CollectiveClient* GetInstance() {
+    std::call_once(init_flag_, [&]() {
+      if (client_.get() == nullptr) {
+        client_.reset(new CollectiveClient());
+      }
+    });
+    return client_.get();
+  }
+
+ private:
+  std::unique_ptr<RPCClient> rpc_client_;
+
+  static std::once_flag init_flag_;
+  static std::unique_ptr<CollectiveClient> client_;
+};
+}  // namespace distributed
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/distributed/collective_server.cc

+/* Copyright (c) 2016 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. */
+
+#include <stdio.h>  // for removing the port file
+#include <csignal>
+#include <cstdlib>
+#include <fstream>
+#include <thread>  // NOLINT
+#include <vector>
+
+#include "paddle/fluid/operators/distributed/collective_server.h"
+
+DEFINE_int32(collective_get_thread_num, 5, "number of threads for rpc get");
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+
+std::once_flag CollectiveServer::init_flag_;
+std::shared_ptr<CollectiveServer> CollectiveServer::collective_server_(nullptr);
+
+CollectiveServer::CollectiveServer(const std::string& end_point, int fan_in) {
+  VLOG(1) << "Create colllective server:" << end_point << ", fan_in:" << fan_in;
+  rpc_server_.reset(new RPCSERVER_T(end_point, fan_in));
+}
+
+void CollectiveServer::Stop() {
+  rpc_server_->ShutDown();
+  server_thread_->join();
+  loop_thread_->join();
+}
+
+void CollectiveServer::StartServer() {
+  get_monomer_handler_.reset(new GetMonomerHandler());
+  get_monomer_handler_->SetRPCServer(rpc_server_.get());
+
+  get_barrier_handler_.reset(new GetMonomerBarrierHandler());
+  get_barrier_handler_->SetRPCServer(rpc_server_.get());
+
+  rpc_server_->RegisterRPC(distributed::kRequestGetMonomerVariable,
+                           get_monomer_handler_.get(),
+                           FLAGS_collective_get_thread_num);
+  rpc_server_->RegisterRPC(distributed::kRequestGetMonomerBarrier,
+                           get_barrier_handler_.get(), 1);
+
+  server_thread_.reset(new std::thread([&]() { rpc_server_->StartServer(); }));
+  rpc_server_->WaitServerReady();
+
+  loop_thread_.reset(new std::thread([&]() {
+    while (true) {
+      if (rpc_server_->IsExit()) {
+        LOG(WARNING) << "get exit!rpc_processor break!";
+        break;
+      }
+      sleep(1);
+    }
+    VLOG(1) << "CollectiveServer loop_thread end";
+  }));
+}
+
+};  // namespace distributed
+};  // namespace operators
+};  // namespace paddle

paddle/fluid/operators/distributed/collective_server.h

+/* Copyright (c) 2016 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. */
+
+#pragma once
+
+#include <map>
+#include <set>
+#include <string>
+#include <thread>  // NOLINT
+#include <utility>
+#include <vector>
+
+#include "gflags/gflags.h"
+
+#include "paddle/fluid/operators/detail/macros.h"
+#include "paddle/fluid/operators/distributed/request_handler.h"
+#include "paddle/fluid/operators/distributed/request_handler_impl.h"
+#include "paddle/fluid/operators/distributed/rpc_server.h"
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+
+class CollectiveServer;
+
+class GetMonomerHandler final : public RequestHandler {
+ public:
+  GetMonomerHandler() : RequestHandler(true) {}
+  virtual ~GetMonomerHandler() {}
+  bool Handle(const std::string& var_name, framework::Scope* scope,
+              framework::Variable* var, framework::Variable** outvar,
+              const int trainer_id, const std::string& out_var_name = "",
+              const std::string& table_name = "") override {
+    VLOG(50) << "GetMonomerHandler recv " << var_name;
+
+    *outvar = scope->FindVar(var_name);
+    PADDLE_ENFORCE(outvar != nullptr, "%s not found", var_name);
+
+    return true;
+  }
+};
+
+class GetMonomerBarrierHandler final : public RequestHandler {
+ public:
+  GetMonomerBarrierHandler() : RequestHandler(true) {}
+  virtual ~GetMonomerBarrierHandler() {}
+  bool Handle(const std::string& var_name, framework::Scope* scope,
+              framework::Variable* var, framework::Variable** outvar,
+              const int trainer_id, const std::string& out_var_name = "",
+              const std::string& table_name = "") override {
+    VLOG(50) << "GetMonomerHandler recv " << var_name;
+
+    rpc_server_->IncreaseVarBarrier(var_name);
+
+    return true;
+  }
+};
+
+class CollectiveServer final {
+ public:
+  explicit CollectiveServer(const std::string& end_point, int fan_in);
+
+  virtual ~CollectiveServer() {}
+
+  void StartServer();
+
+  static CollectiveServer* GetInstance(const std::string& end_point,
+                                       int fan_in) {
+    std::call_once(init_flag_, [&]() {
+      if (collective_server_.get() == nullptr) {
+        collective_server_.reset(new CollectiveServer(end_point, fan_in));
+        collective_server_->StartServer();
+      }
+    });
+
+    return collective_server_.get();
+  }
+
+  std::shared_ptr<RPCServer> GetRPCServer() { return rpc_server_; }
+
+  void Stop();
+
+ private:
+  std::unique_ptr<GetMonomerHandler> get_monomer_handler_;
+  std::unique_ptr<GetMonomerBarrierHandler> get_barrier_handler_;
+
+  std::shared_ptr<distributed::RPCServer> rpc_server_;
+  std::shared_ptr<std::thread> server_thread_;
+  std::shared_ptr<std::thread> loop_thread_;
+
+  bool ready_{false};
+
+  static std::once_flag init_flag_;
+  static std::shared_ptr<CollectiveServer> collective_server_;
+};
+
+};  // namespace distributed
+};  // namespace operators
+};  // namespace paddle

paddle/fluid/operators/distributed/collective_server_test.cc

+/* Copyright (c) 2016 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. */
+
+#include <unistd.h>
+#include <string>
+#include <thread>  // NOLINT
+
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+
+#include "paddle/fluid/operators/detail/macros.h"
+#include "paddle/fluid/operators/distributed/collective_client.h"
+#include "paddle/fluid/operators/distributed/collective_server.h"
+#include "paddle/fluid/operators/distributed/request_handler_impl.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace framework = paddle::framework;
+namespace platform = paddle::platform;
+namespace distributed = paddle::operators::distributed;
+
+std::unique_ptr<distributed::CollectiveServer> StartServer(
+    const std::string& ep, int fan_in, framework::Scope* scope,
+    platform::DeviceContext* dev_ctx) {
+  distributed::CollectiveServer* server =
+      distributed::CollectiveServer::GetInstance(ep, fan_in);
+
+  auto rpc_server = server->GetRPCServer();
+  rpc_server->RegisterVar("var1", distributed::kRequestGetMonomerVariable,
+                          scope, dev_ctx);
+
+  std::cout << "StartServer return" << std::endl;
+  return std::unique_ptr<distributed::CollectiveServer>(server);
+}
+
+std::unique_ptr<framework::Scope> GenerateVars(platform::Place place) {
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  auto& ctx = *pool.Get(place);
+
+  framework::Scope* scope = new framework::Scope();
+  framework::Variable* var = scope->Var("var1");
+  auto* slr = var->GetMutable<framework::SelectedRows>();
+  slr->set_height(1000);
+
+  auto* tensor = slr->mutable_value();
+  auto* rows = slr->mutable_rows();
+
+  tensor->Resize(framework::make_ddim({3, 5}));
+  tensor->mutable_data<float>(place);
+
+  paddle::operators::math::set_constant(ctx, tensor, 32.7);
+  for (int i = 0; i < 3; ++i) rows->push_back(i);
+
+  std::cout << "src:" << distributed::GetSelectedRowsInfo(*slr);
+
+  return std::unique_ptr<framework::Scope>(scope);
+}
+
+void Gather(const std::vector<distributed::RemoteVar>& vars,
+            platform::DeviceContext* dev_ctx) {
+  distributed::CollectiveClient* client =
+      distributed::CollectiveClient::GetInstance();
+
+  framework::Scope* scope = new framework::Scope();
+  framework::Variable* var = scope->Var("var1");
+  var->GetMutable<framework::SelectedRows>();
+
+  std::vector<const framework::SelectedRows*> dst;
+  client->Gather(vars, &dst, *dev_ctx, scope);
+  std::cout << "dst:" << distributed::GetSelectedRowsInfo(*dst[0]);
+}
+
+TEST(PREFETCH, GPU) {
+  platform::CUDAPlace place;
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  auto& ctx = *pool.Get(place);
+
+  std::string ep = "127.0.0.1:7164";
+  auto scope = GenerateVars(place);
+
+  auto* v1 = scope->FindVar("var1");
+  std::cout << "var1:" << v1 << std::endl;
+
+  auto server = StartServer(ep, 2, scope.get(), &ctx);
+  auto rpc_server = server->GetRPCServer();
+
+  distributed::RemoteVar var;
+  var.ep_ = ep;
+  var.var_name_ = "var1";
+  var.trainer_id_ = 0;
+
+  std::vector<distributed::RemoteVar> vars{var};
+  Gather(vars, &ctx);
+  Gather(vars, &ctx);
+
+  std::cout << "begin WaitVarBarrier" << std::endl;
+  rpc_server->WaitVarBarrier("var1");
+  rpc_server->ClearRegisteredVars();
+  server->Stop();
+
+  scope.release();
+  server.release();
+}

paddle/fluid/operators/distributed/grpc_client.cc

@@ -28,11 +28,11 @@ namespace paddle {
 namespace operators {
 namespace distributed {
 
-void GRPCClient::InitImpl() { InitEventLoop(); }
-
-void GRPCClient::InitEventLoop() {
+void GRPCClient::InitImpl() {
   // start the client process thread
   // TODO(wuyi): can make this in a threadpool
+  PADDLE_ENFORCE(client_thread_ == nullptr,
+                 "please not re init proceed thread");
   client_thread_.reset(new std::thread(std::bind(&GRPCClient::Proceed, this)));
 }
 
@@ -106,6 +106,7 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep,
 
 void ProcGetResponse(const VarHandle& var_h,
                      const ::grpc::ByteBuffer& ret_msg) {
+  VLOG(100) << "ProcGetResponse";
   framework::Variable* outvar = nullptr;
   // get response's trainer_id is not used
   int trainer_id;
@@ -126,6 +127,24 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
                                      const framework::Scope& scope,
                                      const std::string& var_name,
                                      int64_t time_out) {
+  return _AsyncGetVar(ep, ctx, scope, var_name,
+                      "/sendrecv.SendRecvService/GetVariable", time_out);
+}
+
+VarHandlePtr GRPCClient::AsyncGetMonomerVariable(
+    const std::string& ep, const platform::DeviceContext& ctx,
+    const framework::Scope& scope, const std::string& var_name,
+    int64_t time_out) {
+  return _AsyncGetVar(ep, ctx, scope, var_name,
+                      "/sendrecv.SendRecvService/GetMonomerVariable", time_out);
+}
+
+VarHandlePtr GRPCClient::_AsyncGetVar(const std::string& ep,
+                                      const platform::DeviceContext& ctx,
+                                      const framework::Scope& scope,
+                                      const std::string& var_name,
+                                      const std::string& rpc_path,
+                                      int64_t time_out) {
   const platform::DeviceContext* p_ctx = &ctx;
   const std::string ep_val = ep;
   const std::string var_name_val = var_name;
@@ -136,7 +155,7 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
   VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope));
   s->Prepare(h, time_out);
 
-  framework::AsyncIO([var_name_val, s, method, p_ctx, h, this] {
+  framework::AsyncIO([var_name_val, s, method, p_ctx, h, rpc_path, this] {
     // prepare input
     sendrecv::VariableMessage req;
     req.set_varname(var_name_val);
@@ -151,8 +170,8 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
 
     platform::RecordRPCEvent record_event(method, p_ctx);
 
-    auto call = s->stub_g_.PrepareUnaryCall(
-        s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_);
+    auto call =
+        s->stub_g_.PrepareUnaryCall(s->context_.get(), rpc_path, buf, &cq_);
     call->StartCall();
     call->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
 
@@ -268,6 +287,34 @@ VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep,
   return h;
 }
 
+VarHandlePtr GRPCClient::AsyncGetMonomerBarrier(const std::string& ep,
+                                                const std::string& var_name,
+                                                int64_t time_out) {
+  const auto ch = GetChannel(ep);
+  BatchBarrierProcessor* s = new BatchBarrierProcessor(ch);
+  const std::string method = "SendMonomerFetchBarrierRPC";
+  VarHandlePtr h(
+      new VarHandle(ep, method, FETCH_BARRIER_MESSAGE, nullptr, nullptr));
+  s->Prepare(h, time_out);
+
+  VLOG(30) << s->GetVarHandlePtr()->String() << " begin";
+
+  sendrecv::VariableMessage req;
+  req.set_varname(var_name);
+
+  platform::RecordRPCEvent record_event(method, nullptr);
+
+  auto rpc = s->stub_->AsyncGetMonomerBarrier(s->context_.get(), req, &cq_);
+  rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
+  req_count_++;
+
+  if (UNLIKELY(platform::IsProfileEnabled())) {
+    h->Wait();
+  }
+
+  return h;
+}
+
 VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep,
                                            int64_t time_out) {
   const auto ch = GetChannel(ep);

paddle/fluid/operators/distributed/grpc_client.h

@@ -189,6 +189,11 @@ class GRPCClient : public RPCClient {
                            const std::string& var_name,
                            int64_t time_out = FLAGS_rpc_deadline) override;
 
+  VarHandlePtr AsyncGetMonomerVariable(
+      const std::string& ep, const platform::DeviceContext& ctx,
+      const framework::Scope& scope, const std::string& var_name,
+      int64_t time_out = FLAGS_rpc_deadline) override;
+
   VarHandlePtr AsyncPrefetchVar(const std::string& ep,
                                 const platform::DeviceContext& ctx,
                                 const framework::Scope& scope,
@@ -200,8 +205,12 @@ class GRPCClient : public RPCClient {
   VarHandlePtr AsyncSendBatchBarrier(
       const std::string& ep, int64_t time_out = FLAGS_rpc_deadline) override;
 
-  VarHandlePtr AsyncSendFetchBarrier(
-      const std::string& ep, int64_t time_out = FLAGS_rpc_deadline) override;
+  VarHandlePtr AsyncSendFetchBarrier(const std::string& ep,
+                                     int64_t time_out) override;
+
+  VarHandlePtr AsyncGetMonomerBarrier(
+      const std::string& ep, const std::string& var_name,
+      int64_t time_out = FLAGS_rpc_deadline) override;
 
   VarHandlePtr AsyncCheckpointNotify(
       const std::string& ep, const std::string& dir,
@@ -214,21 +223,22 @@ class GRPCClient : public RPCClient {
 
   void SendComplete() override;
 
- protected:
   void InitImpl() override;
 
  private:
-  // InitEventLoop should only be called by Init()
-  void InitEventLoop();
-
   void Proceed();
 
   std::shared_ptr<grpc::Channel> GetChannel(const std::string& ep);
+  VarHandlePtr _AsyncGetVar(const std::string& ep,
+                            const platform::DeviceContext& ctx,
+                            const framework::Scope& scope,
+                            const std::string& var_name, const std::string& rpc,
+                            int64_t time_out);
 
  private:
   grpc::CompletionQueue cq_;
   std::unordered_map<std::string, std::shared_ptr<grpc::Channel>> channels_;
-  std::unique_ptr<std::thread> client_thread_;
+  std::unique_ptr<std::thread> client_thread_{nullptr};
 
   // mutex for Wait client sync
   std::mutex sync_mutex_;

paddle/fluid/operators/distributed/grpc_server.cc

@@ -158,6 +158,98 @@ class RequestGet final : public RequestBase {
   ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_;
 };
 
+class RequestGetMonomerVariable final : public RequestBase {
+ public:
+  explicit RequestGetMonomerVariable(GrpcService::AsyncService* service,
+                                     ::grpc::ServerCompletionQueue* cq,
+                                     RequestHandler* request_handler,
+                                     int req_id, RPCServer* rpc_server)
+      : RequestBase(service, cq, request_handler, req_id),
+        responder_(&ctx_),
+        rpc_server_(rpc_server) {
+    auto method_id =
+        static_cast<int>(distributed::GrpcMethod::kGetMonomerVariable);
+    service_->RequestAsyncUnary(
+        method_id, &ctx_, &request_, &responder_, cq_, cq_,
+        reinterpret_cast<void*>(static_cast<intptr_t>(req_id)));
+  }
+
+  virtual ~RequestGetMonomerVariable() {}
+
+  std::string GetReqName() override { return request_.varname(); }
+
+  void Process() override {
+    // proc request.
+    std::string varname = request_.varname();
+
+    rpc_server_->WaitVarCond(varname);
+    MonomerHandle h = rpc_server_->GetMonomer(varname);
+
+    auto scope = h.scope_;
+    auto invar = scope->FindVar(varname);
+    framework::Variable* outvar = nullptr;
+
+    request_handler_->Handle(varname, scope, invar, &outvar,
+                             request_.trainer_id());
+
+    if (outvar) {
+      SerializeToByteBuffer(varname, outvar, *h.dev_ctx_, &reply_);
+    }
+    Finish(reply_, &responder_);
+  }
+
+ protected:
+  sendrecv::VariableMessage request_;
+  ::grpc::ByteBuffer reply_;
+  ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_;
+  RPCServer* rpc_server_{nullptr};
+};
+
+class RequestGetMonomerBarrier final : public RequestBase {
+ public:
+  explicit RequestGetMonomerBarrier(GrpcService::AsyncService* service,
+                                    ::grpc::ServerCompletionQueue* cq,
+                                    RequestHandler* request_handler, int req_id,
+                                    RPCServer* rpc_server)
+      : RequestBase(service, cq, request_handler, req_id),
+        responder_(&ctx_),
+        rpc_server_(rpc_server) {
+    auto method_id =
+        static_cast<int>(distributed::GrpcMethod::kGetMonomerBarrier);
+    service_->RequestAsyncUnary(
+        method_id, &ctx_, &request_, &responder_, cq_, cq_,
+        reinterpret_cast<void*>(static_cast<intptr_t>(req_id)));
+  }
+
+  virtual ~RequestGetMonomerBarrier() {}
+
+  std::string GetReqName() override { return request_.varname(); }
+
+  void Process() override {
+    // proc request.
+    std::string varname = request_.varname();
+    VLOG(4) << "RequestGetMonomerBarrier " << varname;
+
+    rpc_server_->WaitVarCond(varname);
+    MonomerHandle h = rpc_server_->GetMonomer(varname);
+
+    framework::Scope* scope = nullptr;
+    framework::Variable* invar = nullptr;
+    framework::Variable* outvar = nullptr;
+
+    request_handler_->Handle(varname, scope, invar, &outvar,
+                             request_.trainer_id());
+
+    Finish(reply_, &responder_);
+  }
+
+ protected:
+  sendrecv::VariableMessage request_;
+  sendrecv::VoidMessage reply_;
+  ServerAsyncResponseWriter<sendrecv::VoidMessage> responder_;
+  RPCServer* rpc_server_{nullptr};
+};
+
 class RequestPrefetch final : public RequestBase {
  public:
   explicit RequestPrefetch(GrpcService::AsyncService* service,
@@ -249,7 +341,7 @@ class RequestCheckpointNotify final : public RequestBase {
 };
 
 void AsyncGRPCServer::WaitServerReady() {
-  VLOG(4) << "AsyncGRPCServer is wait server ready";
+  VLOG(4) << "AsyncGRPCServer is waiting server ready";
   std::unique_lock<std::mutex> lock(this->mutex_ready_);
   condition_ready_.wait(lock, [=] { return this->ready_ == 1; });
   VLOG(4) << "AsyncGRPCServer WaitSeverReady";
@@ -368,6 +460,12 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name,
     b = new RequestSend(&service_, cq.get(), handler, req_id);
   } else if (rpc_name == kRequestGet) {
     b = new RequestGet(&service_, cq.get(), handler, req_id);
+  } else if (rpc_name == kRequestGetMonomerVariable) {
+    b = new RequestGetMonomerVariable(&service_, cq.get(), handler, req_id,
+                                      this);
+  } else if (rpc_name == kRequestGetMonomerBarrier) {
+    b = new RequestGetMonomerBarrier(&service_, cq.get(), handler, req_id,
+                                     this);
   } else if (rpc_name == kRequestPrefetch) {
     b = new RequestPrefetch(&service_, cq.get(), handler, req_id);
   } else if (rpc_name == kRequestCheckpoint) {
@@ -378,7 +476,7 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name,
 
   reqs[req_id] = b;
 
-  VLOG(4) << "Create RequestSend status:" << b->Status();
+  VLOG(4) << "TryToRegisterNewOne status:" << b->Status();
 }
 
 void AsyncGRPCServer::HandleRequest(

paddle/fluid/operators/distributed/grpc_service.h

@@ -81,10 +81,12 @@ enum class GrpcMethod {
   kGetVariable,
   kPrefetchVariable,
   kCheckpointNotify,
+  kGetMonomerVariable,
+  kGetMonomerBarrier,
 };
 
 static const int kGrpcNumMethods =
-    static_cast<int>(GrpcMethod::kCheckpointNotify) + 1;
+    static_cast<int>(GrpcMethod::kGetMonomerBarrier) + 1;
 
 inline const char* GrpcMethodName(GrpcMethod id) {
   switch (id) {
@@ -92,6 +94,10 @@ inline const char* GrpcMethodName(GrpcMethod id) {
       return "/sendrecv.SendRecvService/SendVariable";
     case GrpcMethod::kGetVariable:
       return "/sendrecv.SendRecvService/GetVariable";
+    case GrpcMethod::kGetMonomerVariable:
+      return "/sendrecv.SendRecvService/GetMonomerVariable";
+    case GrpcMethod::kGetMonomerBarrier:
+      return "/sendrecv.SendRecvService/GetMonomerBarrier";
     case GrpcMethod::kPrefetchVariable:
       return "/sendrecv.SendRecvService/PrefetchVariable";
     case GrpcMethod::kCheckpointNotify:

paddle/fluid/operators/distributed/request_handler.h

@@ -37,6 +37,8 @@ namespace distributed {
 
 constexpr char kRequestSend[] = "RequestSend";
 constexpr char kRequestGet[] = "RequestGet";
+constexpr char kRequestGetMonomerVariable[] = "RequestGetMonomerVariable";
+constexpr char kRequestGetMonomerBarrier[] = "RequestGetMonomerBarrier";
 constexpr char kRequestPrefetch[] = "RequestPrefetch";
 constexpr char kRequestCheckpoint[] = "RequestCheckpoint";
 constexpr char kRequestPassBarrier[] = "RequestPassBarrier";

paddle/fluid/operators/distributed/rpc_client.h

@@ -45,6 +45,11 @@ class RPCClient {
                                    const std::string& var_name,
                                    int64_t time_out = FLAGS_rpc_deadline) = 0;
 
+  virtual VarHandlePtr AsyncGetMonomerVariable(
+      const std::string& ep, const platform::DeviceContext& ctx,
+      const framework::Scope& scope, const std::string& var_name,
+      int64_t time_out = FLAGS_rpc_deadline) = 0;
+
   virtual VarHandlePtr AsyncPrefetchVar(
       const std::string& ep, const platform::DeviceContext& ctx,
       const framework::Scope& scope, const std::string& in_var_name,
@@ -57,6 +62,10 @@ class RPCClient {
   virtual VarHandlePtr AsyncSendFetchBarrier(
       const std::string& ep, int64_t time_out = FLAGS_rpc_deadline) = 0;
 
+  virtual VarHandlePtr AsyncGetMonomerBarrier(
+      const std::string& ep, const std::string& var_name,
+      int64_t time_out = FLAGS_rpc_deadline) = 0;
+
   virtual VarHandlePtr AsyncCheckpointNotify(
       const std::string& ep, const std::string& dir,
       int64_t time_out = FLAGS_rpc_deadline) = 0;
@@ -87,8 +96,9 @@ class RPCClient {
     }
   }
 
- protected:
   virtual void InitImpl() {}
+
+ protected:
   // each trainer have exact one trainer id, it should be static
   static int trainer_id_;
 

paddle/fluid/operators/distributed/rpc_server.cc

@@ -132,6 +132,96 @@ void RPCServer::WaitCond(const std::string& rpc_name) {
       lock, [=] { return (cur_cond_.load() == cond || exit_flag_.load()); });
 }
 
+void RPCServer::RegisterVar(const std::string& var_name,
+                            const std::string& rpc_name,
+                            framework::Scope* scope,
+                            platform::DeviceContext* dev_ctx) {
+  MonomerHandle h;
+  h.var_name_ = var_name;
+  h.rpc_name_ = rpc_name;
+  h.scope_ = scope;
+  h.dev_ctx_ = dev_ctx;
+
+  {
+    std::unique_lock<std::mutex> lock(mutex_);
+    if (var_map_.find(var_name) != var_map_.end()) {
+      PADDLE_ENFORCE(false, "%s alreay in var_map", var_name);
+    }
+    var_map_[var_name] = h;
+  }
+
+  rpc_cond_.notify_all();
+  VLOG(4) << "RegisterVar context:" << h.String();
+}
+
+void RPCServer::IncreaseVarBarrier(const std::string& var_name) {
+  int b = 0;
+  MonomerHandle h;
+  {
+    std::unique_lock<std::mutex> lock(mutex_);
+    b = ++var_map_[var_name].barrier_;
+    h = var_map_[var_name];
+  }
+
+  if (b >= client_num_) {
+    barrier_cond_.notify_all();
+  }
+
+  VLOG(4) << "IncreaseVarBarrier context:" << h.String();
+}
+
+void RPCServer::WaitVarBarrier(const std::string& var_name) {
+  VLOG(4) << "WaitBarrier var_name:" << var_name;
+
+  std::unique_lock<std::mutex> lock(mutex_);
+  barrier_cond_.wait(lock, [&]() {
+    return ((var_map_[var_name].barrier_ >= client_num_ && client_num_ != 0) ||
+            exit_flag_.load());
+  });
+
+  VLOG(4) << "WaitBarrier context: " << var_map_[var_name].String();
+}
+
+void RPCServer::SetVarCond(const std::string& var_name) {
+  VLOG(4) << "SetVarCond var_name:" << var_name;
+  {
+    std::unique_lock<std::mutex> lock(mutex_);
+    if (var_map_.find(var_name) != var_map_.end()) {
+      rpc_cond_.notify_all();
+    }
+  }
+}
+
+void RPCServer::WaitVarCond(const std::string& var_name) {
+  VLOG(4) << "WaitVarCond var_name:" << var_name;
+
+  std::unique_lock<std::mutex> lock(mutex_);
+  rpc_cond_.wait(lock, [=] {
+    return (var_map_.find(var_name) != var_map_.end() || exit_flag_.load());
+  });
+
+  VLOG(4) << "WaitVarCond var_name:" << var_name << " end";
+}
+
+MonomerHandle RPCServer::GetMonomer(const std::string& var_name) {
+  MonomerHandle h;
+  {
+    std::unique_lock<std::mutex> lock(mutex_);
+    h = var_map_[var_name];
+  }
+
+  return h;
+}
+
+void RPCServer::ClearRegisteredVars() {
+  std::unique_lock<std::mutex> lock(mutex_);
+  var_map_.clear();
+}
+
+void RPCServer::ClearVar(const std::string& var_name) {
+  std::unique_lock<std::mutex> lock(mutex_);
+  var_map_.erase(var_name);
+}
 }  // namespace distributed
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/distributed/rpc_server.h

@@ -21,12 +21,30 @@
 #include <utility>
 #include <vector>
 
+#include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/operators/distributed/request_handler.h"
+#include "paddle/fluid/platform/device_context.h"
 
 namespace paddle {
 namespace operators {
 namespace distributed {
 
+struct MonomerHandle {
+  std::string var_name_;
+  std::string rpc_name_;
+  framework::Scope* scope_{nullptr};
+  platform::DeviceContext* dev_ctx_{nullptr};
+  int64_t barrier_{0};
+
+  std::string String() {
+    std::stringstream ss;
+    ss << "var_name:" << var_name_ << ", rpc_name:" << rpc_name_
+       << ", scope:" << scope_ << ", dev_ctx:" << dev_ctx_
+       << ", barrier_:" << barrier_;
+    return ss.str();
+  }
+};
+
 class RPCServer {
  public:
   explicit RPCServer(const std::string& address, int client_num)
@@ -67,6 +85,16 @@ class RPCServer {
   void WaitCond(const std::string& rpc_name);
   void IncreaseBatchBarrier(const std::string rpc_name);
 
+  void RegisterVar(const std::string& var_name, const std::string& rpc_name,
+                   framework::Scope* scope, platform::DeviceContext* dev_ctx);
+  void IncreaseVarBarrier(const std::string& var_name);
+  void WaitVarBarrier(const std::string& var_name);
+  void SetVarCond(const std::string& var_name);
+  void WaitVarCond(const std::string& var_name);
+  void ClearRegisteredVars();
+  void ClearVar(const std::string& var_name);
+  MonomerHandle GetMonomer(const std::string& var_name);
+
   void Complete();
 
   void ResetBarrierCounter();
@@ -95,6 +123,9 @@ class RPCServer {
   std::unordered_map<std::string, RequestHandler*> rpc_call_map_;
   std::unordered_map<std::string, int> rpc_thread_num_;
   friend class RequestHandler;
+
+  // TODO(gongwb): use more cond to notify or wait;
+  std::unordered_map<std::string, MonomerHandle> var_map_;
 };
 
 };  // namespace distributed

paddle/fluid/operators/distributed/send_recv.proto.in

@@ -28,6 +28,9 @@ service SendRecvService {
   rpc PrefetchVariable(VariableMessage) returns (VariableMessage) {}
 
   rpc CheckpointNotify(VariableMessage) returns (VoidMessage) {}
+
+  rpc GetMonomerVariable(VariableMessage) returns (VariableMessage) {}
+  rpc GetMonomerBarrier(VariableMessage) returns (VoidMessage) {}
 }
 
 // VariableMessage is serialized paddle variable message.

paddle/fluid/operators/elementwise/elementwise_mul_op.h

@@ -60,15 +60,37 @@ template <typename DeviceContext, typename T>
 class ElementwiseMulKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* x = ctx.Input<framework::LoDTensor>("X");
+    auto x_var = ctx.InputVar("X");
+    PADDLE_ENFORCE(x_var != nullptr,
+                   "Cannot get input Variable X, variable name = %s",
+                   ctx.op().Input("X"));
     auto* y = ctx.Input<framework::LoDTensor>("Y");
-    auto* z = ctx.Output<framework::LoDTensor>("Out");
+
+    framework::Tensor x, *z;
+    if (x_var->IsType<framework::SelectedRows>()) {
+      PADDLE_ENFORCE(y->dims().size() == 1 && y->dims()[0] == 1,
+                     "For elementwise_op, if X is Sparse, Y must be scalar.");
+      auto& x_sele = x_var->Get<framework::SelectedRows>();
+      auto out_sele = ctx.Output<framework::SelectedRows>("Out");
+      x = x_sele.value();
+      out_sele->set_rows(x_sele.rows());
+      out_sele->set_height(x_sele.height());
+      out_sele->mutable_value()->Resize(x_sele.value().dims());
+      out_sele->mutable_value()->mutable_data(ctx.GetPlace(), x.type());
+      z = ctx.Output<framework::SelectedRows>("Out")->mutable_value();
+    } else if (x_var->IsType<framework::LoDTensor>()) {
+      x = x_var->Get<framework::LoDTensor>();
+      z = ctx.Output<framework::LoDTensor>("Out");
+    } else {
+      PADDLE_THROW("X's type[%s] is not supported by elementwise_op.",
+                   x_var->Type().name());
+    }
 
     z->mutable_data<T>(ctx.GetPlace());
-    if (x->numel() == y->numel()) {
-      elementwise_mul<DeviceContext, T>(ctx, x, y, z);
+    if (x.numel() == y->numel()) {
+      elementwise_mul<DeviceContext, T>(ctx, &x, y, z);
     } else {
-      default_elementwise_mul<DeviceContext, T>(ctx, x, y, z);
+      default_elementwise_mul<DeviceContext, T>(ctx, &x, y, z);
     }
   }
 };

paddle/fluid/operators/elementwise/elementwise_op.h

@@ -40,21 +40,28 @@ class ElementwiseOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of elementwise op should not be null.");
 
-    PADDLE_ENFORCE(
-        ctx->GetInputsVarType("X").front() ==
-            framework::proto::VarType::LOD_TENSOR,
-        "The input var's type should be LoDTensor, but the received is %s",
-        ctx->Inputs("X").front(), ctx->GetInputsVarType("X").front());
     PADDLE_ENFORCE(
         ctx->GetInputsVarType("Y").front() ==
             framework::proto::VarType::LOD_TENSOR,
-        "The input var's type should be LoDTensor, but the received is %s",
-        ctx->Inputs("Y").front(), ctx->GetInputsVarType("Y").front());
+        "The input var's type should be LoDTensor, but the received is %s [%s]",
+        ctx->GetInputsVarType("Y").front(), ctx->Inputs("Y").front());
 
+    if (ctx->GetInputsVarType("X").front() ==
+        framework::proto::VarType::LOD_TENSOR) {
       auto x_dim = ctx->GetInputDim("X");
       auto y_dim = ctx->GetInputDim("Y");
       PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
                         "Rank of first input must >= rank of second input.");
+    } else if (ctx->GetInputsVarType("X").front() ==
+               framework::proto::VarType::SELECTED_ROWS) {
+      PADDLE_ENFORCE((ctx->GetInputDim("Y").size() == 1u) &&
+                         (ctx->GetInputDim("Y")[0] == 1),
+                     "For elementwise_op, if X is Sparse, "
+                     "Y must be scalar.");
+    } else {
+      PADDLE_THROW("X's type[%s] is not supported by elementwise_op.",
+                   ctx->GetInputsVarType("X").front());
+    }
 
     ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");

paddle/fluid/operators/fused/fused_embedding_fc_lstm_op.cc

@@ -217,13 +217,13 @@ class FusedEmbeddingFCLSTMKernel : public framework::OpKernel<T> {
   auto& act_gate_str = ctx.Attr<std::string>("gate_activation");               \
   auto& act_cell_str = ctx.Attr<std::string>("cell_activation");               \
   auto& act_cand_str = ctx.Attr<std::string>("candidate_activation");          \
-  if (platform::jit::MayIUse(platform::jit::avx)) {                            \
-    math::VecActivations<T, platform::jit::avx> act_functor;                   \
+  if (platform::MayIUse(platform::avx)) {                                      \
+    math::VecActivations<T, platform::avx> act_functor;                        \
     act_gate = act_functor(act_gate_str);                                      \
     act_cell = act_functor(act_cell_str);                                      \
     act_cand = act_functor(act_cand_str);                                      \
   } else {                                                                     \
-    math::VecActivations<T, platform::jit::isa_any> act_functor;               \
+    math::VecActivations<T, platform::isa_any> act_functor;                    \
     act_gate = act_functor(act_gate_str);                                      \
     act_cell = act_functor(act_cell_str);                                      \
     act_cand = act_functor(act_cand_str);                                      \

paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc

@@ -151,11 +151,11 @@ class FusionSeqExpandConcatFCOpKernel : public framework::OpKernel<T> {
 
     std::function<void(const int, const T*, T*)> fc_act;
     auto& fc_act_str = ctx.Attr<std::string>("fc_activation");
-    if (platform::jit::MayIUse(platform::jit::avx)) {
-      math::VecActivations<T, platform::jit::avx> act_functor;
+    if (platform::MayIUse(platform::avx)) {
+      math::VecActivations<T, platform::avx> act_functor;
       fc_act = act_functor(fc_act_str);
     } else {
-      math::VecActivations<T, platform::jit::isa_any> act_functor;
+      math::VecActivations<T, platform::isa_any> act_functor;
       fc_act = act_functor(fc_act_str);
     }
 

paddle/fluid/operators/get_tensor_from_selected_rows_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. */
+
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/tensor_util.h"
+
+namespace paddle {
+namespace operators {
+
+class GetTensorFromSelectedRowsOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "GetTensorFromSelectedRowsOp must has input X.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "GetTensorFromSelectedRowsOp must has output Out.");
+    PADDLE_ENFORCE(
+        ctx->GetInputsVarType("X").front() ==
+            framework::proto::VarType::SELECTED_ROWS,
+        "The input X's type should be SelectedRows, but the received is %s",
+        ctx->Inputs("X").front(), ctx->GetInputsVarType("X").front());
+    PADDLE_ENFORCE(
+        ctx->GetOutputsVarType("Out").front() ==
+            framework::proto::VarType::LOD_TENSOR,
+        "The output Out's type should be LoDTensor, but the received is %s",
+        ctx->Outputs("Out").front(), ctx->GetOutputsVarType("Out").front());
+
+    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    return framework::OpKernelType(
+        framework::GetDataTypeOfVar(ctx.InputVar("X")), ctx.device_context());
+  }
+};
+
+class GetTensorFromSelectedRowsKernel {
+ public:
+  void operator()(const framework::ExecutionContext &ctx) const {
+    auto *x = ctx.Input<framework::SelectedRows>("X");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
+
+    out->Resize(x->value().dims());
+    out->mutable_data(ctx.GetPlace(), x->value().type());
+    framework::TensorCopy(x->value(), ctx.GetPlace(), ctx.device_context(),
+                          out);
+  }
+};
+
+class GetTensorFromSelectedRowsOpProtoMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "The input type is SelectedRows.");
+    AddOutput("Out", "The output type is LoDTensor.");
+    AddComment(
+        R"DOC(
+GetTensorFromSelectedRows Operator
+
+GetTensorFromSelectedRows is used to get the tensor from SelectedRows.
+
+)DOC");
+  }
+};
+
+class GetTensorFromSelectedRowsOpVarTypeInference
+    : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const final {
+    auto out_var_name = op_desc.Output("Out").front();
+    auto in_var_name = op_desc.Input("X").front();
+
+    auto out_var = block->FindRecursiveOrCreateVar(out_var_name);
+    auto in_var = block->FindRecursiveOrCreateVar(in_var_name);
+    out_var.SetType(framework::proto::VarType::LOD_TENSOR);
+    out_var.SetDataType(in_var.GetDataType());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(get_tensor_from_selected_rows,
+                  ops::GetTensorFromSelectedRowsOp,
+                  ops::GetTensorFromSelectedRowsOpProtoMaker,
+                  ops::GetTensorFromSelectedRowsOpVarTypeInference);
+
+REGISTER_OP_CPU_KERNEL_FUNCTOR(get_tensor_from_selected_rows, float,
+                               ops::GetTensorFromSelectedRowsKernel, double,
+                               ops::GetTensorFromSelectedRowsKernel, int,
+                               ops::GetTensorFromSelectedRowsKernel, int64_t,
+                               ops::GetTensorFromSelectedRowsKernel);
+
+#ifdef PADDLE_WITH_CUDA
+REGISTER_OP_CUDA_KERNEL_FUNCTOR(get_tensor_from_selected_rows, float,
+                                ops::GetTensorFromSelectedRowsKernel, double,
+                                ops::GetTensorFromSelectedRowsKernel, int,
+                                ops::GetTensorFromSelectedRowsKernel, int64_t,
+                                ops::GetTensorFromSelectedRowsKernel);
+#endif

paddle/fluid/operators/hierarchical_sigmoid_op.cc

@@ -150,14 +150,14 @@ class HierarchicalSigmoidGradOp : public framework::OperatorWithKernel {
                    "Output(W@Grad should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
                    "Output(X@Grad should not be null.");
-    if (!ctx->Attrs().Get<bool>("is_sparse")) {
+
     if (ctx->HasOutput(framework::GradVarName("Bias"))) {
       ctx->SetOutputDim(framework::GradVarName("Bias"),
                         ctx->GetInputDim("Bias"));
     }
     ctx->SetOutputDim(framework::GradVarName("W"), ctx->GetInputDim("W"));
-    }
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", /*->*/ framework::GradVarName("X"));
   }
 
  protected:

paddle/fluid/operators/hierarchical_sigmoid_op.h

@@ -185,7 +185,6 @@ class HierarchicalSigmoidGradOpKernel : public framework::OpKernel<T> {
           ctx.Output<framework::SelectedRows>(framework::GradVarName("W"));
       w_grad->set_rows(real_rows);
       // Build a map of id -> row_index to speed up finding the index of one id
-      w_grad->SyncIndex();
       w_grad->set_height(w.dims()[0]);
       auto* w_grad_value = w_grad->mutable_value();
       framework::DDim temp_dim(w.dims());

paddle/fluid/operators/load_combine_op.cc

@@ -32,16 +32,26 @@ class LoadCombineOp : public framework::OperatorBase {
                const platform::Place &place) const override {
     auto filename = Attr<std::string>("file_path");
     auto load_as_fp16 = Attr<bool>("load_as_fp16");
-
-    std::ifstream fin(filename);
-    PADDLE_ENFORCE(static_cast<bool>(fin),
-                   "Cannot open file %s for load_combine op", filename);
-
+    auto model_from_memory = Attr<bool>("model_from_memory");
     auto out_var_names = Outputs("Out");
     PADDLE_ENFORCE_GT(
         static_cast<int>(out_var_names.size()), 0,
         "The number of output variables should be greater than 0.");
-
+    if (!model_from_memory) {
+      std::ifstream fin(filename);
+      PADDLE_ENFORCE(static_cast<bool>(fin),
+                     "Cannot open file %s for load_combine op", filename);
+      LoadParamsFromBuffer(scope, place, &fin, load_as_fp16, out_var_names);
+    } else {
+      PADDLE_ENFORCE(!filename.empty(), "Cannot load file from memory");
+      std::stringstream fin(filename);
+      LoadParamsFromBuffer(scope, place, &fin, load_as_fp16, out_var_names);
+    }
+  }
+  void LoadParamsFromBuffer(
+      const framework::Scope &scope, const platform::Place &place,
+      std::istream *buffer, bool load_as_fp16,
+      const std::vector<std::string> &out_var_names) const {
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &dev_ctx = *pool.Get(place);
 
@@ -54,11 +64,10 @@ class LoadCombineOp : public framework::OperatorBase {
       auto *tensor = out_var->GetMutable<framework::LoDTensor>();
 
       // Error checking
-      PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot read more from file %s",
-                     filename);
+      PADDLE_ENFORCE(static_cast<bool>(buffer), "Cannot read more");
 
       // Get data from fin to tensor
-      DeserializeFromStream(fin, tensor, dev_ctx);
+      DeserializeFromStream(*buffer, tensor, dev_ctx);
 
       auto in_dtype = framework::ToDataType(tensor->type());
       auto out_dtype =
@@ -103,11 +112,17 @@ class LoadCombineOpProtoMaker : public framework::OpProtoAndCheckerMaker {
                          "LoDTensors will be loaded from \"file_path\".")
         .AddCustomChecker(
             [](const std::string &path) { return !path.empty(); });
+    AddAttr<bool>("model_from_memory",
+                  "(boolean, default false)"
+                  "If true, file_path is in memory, and LoDTensors will be "
+                  "loaded directly from memory")
+        .SetDefault(false);
     AddComment(R"DOC(
 LoadCombine Operator.
 
-LoadCombine operator loads LoDTensor variables from a file. The file should 
-contain one or more LoDTensors serialized using the SaveCombine operator. The 
+LoadCombine operator loads LoDTensor variables from a file, which could be 
+loaded in memory already. The file should contain one or more LoDTensors 
+serialized using the SaveCombine operator. The
 LoadCombine operator applies a deserialization strategy to appropriately load 
 the LodTensors, and this strategy complements the serialization strategy used 
 in the SaveCombine operator. Hence, the LoadCombine operator is tightly coupled

paddle/fluid/operators/math/CMakeLists.txt

@@ -59,6 +59,7 @@ math_library(matrix_bit_code)
 
 math_library(unpooling)
 math_library(vol2col)
+math_library(prelu)
 
 cc_test(math_function_test SRCS math_function_test.cc DEPS math_function)
 cc_test(selected_rows_functor_test SRCS selected_rows_functor_test.cc DEPS selected_rows_functor)

paddle/fluid/operators/math/cpu_vec.h

@@ -77,7 +77,7 @@ inline void vec_scal<double>(const int n, const double a, double* x) {
 #endif
 
 // MKL scal only support inplace, choose this if src and dst are not equal
-template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 inline void vec_scal(const int n, const T a, const T* x, T* y) {
   for (int i = 0; i < n; ++i) {
     y[i] = a * x[i];
@@ -85,12 +85,12 @@ inline void vec_scal(const int n, const T a, const T* x, T* y) {
 }
 
 template <>
-inline void vec_scal<float, platform::jit::avx>(const int n, const float a,
+inline void vec_scal<float, platform::avx>(const int n, const float a,
                                            const float* x, float* y) {
 #ifdef __AVX__
   constexpr int block = YMM_FLOAT_BLOCK;
   if (n < block) {
-    vec_scal<float, platform::jit::isa_any>(n, a, x, y);
+    vec_scal<float, platform::isa_any>(n, a, x, y);
     return;
   }
   const int rest = n % block;
@@ -114,24 +114,24 @@ inline void vec_scal<float, platform::jit::avx>(const int n, const float a,
     y[i] = a * x[i];
   }
 #else
-  vec_scal<float, platform::jit::isa_any>(n, a, x, y);
+  vec_scal<float, platform::isa_any>(n, a, x, y);
 #endif
 }
 
 template <>
-inline void vec_scal<float, platform::jit::avx2>(const int n, const float a,
+inline void vec_scal<float, platform::avx2>(const int n, const float a,
                                             const float* x, float* y) {
-  vec_scal<float, platform::jit::avx>(n, a, x, y);
+  vec_scal<float, platform::avx>(n, a, x, y);
 }
 
 template <>
-inline void vec_scal<float, platform::jit::avx512f>(const int n, const float a,
+inline void vec_scal<float, platform::avx512f>(const int n, const float a,
                                                const float* x, float* y) {
   // TODO(TJ): enable me
-  vec_scal<float, platform::jit::avx2>(n, a, x, y);
+  vec_scal<float, platform::avx2>(n, a, x, y);
 }
 
-template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 inline void vec_bias_sub(const int n, const T a, const T* x, T* y) {
   for (int i = 0; i < n; ++i) {
     y[i] = a - x[i];
@@ -139,12 +139,12 @@ inline void vec_bias_sub(const int n, const T a, const T* x, T* y) {
 }
 
 template <>
-inline void vec_bias_sub<float, platform::jit::avx>(const int n, const float a,
+inline void vec_bias_sub<float, platform::avx>(const int n, const float a,
                                                const float* x, float* y) {
 #ifdef __AVX__
   constexpr int block = YMM_FLOAT_BLOCK;
   if (n < block) {
-    vec_bias_sub<float, platform::jit::isa_any>(n, a, x, y);
+    vec_bias_sub<float, platform::isa_any>(n, a, x, y);
     return;
   }
   const int rest = n % block;
@@ -168,27 +168,25 @@ inline void vec_bias_sub<float, platform::jit::avx>(const int n, const float a,
     y[i] = a - x[i];
   }
 #else
-  vec_bias_sub<float, platform::jit::isa_any>(n, a, x, y);
+  vec_bias_sub<float, platform::isa_any>(n, a, x, y);
 #endif
 }
 
 template <>
-inline void vec_bias_sub<float, platform::jit::avx2>(const int n, const float a,
+inline void vec_bias_sub<float, platform::avx2>(const int n, const float a,
                                                 const float* x, float* y) {
-  vec_bias_sub<float, platform::jit::avx>(n, a, x, y);
+  vec_bias_sub<float, platform::avx>(n, a, x, y);
 }
 
 template <>
-inline void vec_bias_sub<float, platform::jit::avx512f>(const int n,
-                                                        const float a,
-                                                        const float* x,
-                                                        float* y) {
+inline void vec_bias_sub<float, platform::avx512f>(const int n, const float a,
+                                                   const float* x, float* y) {
   // TODO(TJ): enable me
-  vec_bias_sub<float, platform::jit::avx2>(n, a, x, y);
+  vec_bias_sub<float, platform::avx2>(n, a, x, y);
 }
 
 // out = x*y + (1-x)*z
-template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 inline void vec_cross(const int n, const T* x, const T* y, const T* z, T* out) {
   for (int i = 0; i < n; ++i) {
     out[i] = x[i] * y[i] + (static_cast<T>(1) - x[i]) * z[i];
@@ -196,13 +194,13 @@ inline void vec_cross(const int n, const T* x, const T* y, const T* z, T* out) {
 }
 
 template <>
-inline void vec_cross<float, platform::jit::avx>(const int n, const float* x,
+inline void vec_cross<float, platform::avx>(const int n, const float* x,
                                             const float* y, const float* z,
                                             float* out) {
 #ifdef __AVX__
   constexpr int block = YMM_FLOAT_BLOCK;
   if (n < block) {
-    vec_cross<float, platform::jit::isa_any>(n, x, y, z, out);
+    vec_cross<float, platform::isa_any>(n, x, y, z, out);
     return;
   }
   const int rest = n % block;
@@ -228,25 +226,26 @@ inline void vec_cross<float, platform::jit::avx>(const int n, const float* x,
     out[i] = x[i] * y[i] + (1.f - x[i]) * z[i];
   }
 #else
-  vec_cross<float, platform::jit::isa_any>(n, x, y, z, out);
+  vec_cross<float, platform::isa_any>(n, x, y, z, out);
 #endif
 }
 
 template <>
-inline void vec_cross<float, platform::jit::avx2>(const int n, const float* x,
-                                                  const float* y,
-                                                  const float* z, float* out) {
-  vec_cross<float, platform::jit::avx>(n, x, y, z, out);
+inline void vec_cross<float, platform::avx2>(const int n, const float* x,
+                                             const float* y, const float* z,
+                                             float* out) {
+  vec_cross<float, platform::avx>(n, x, y, z, out);
 }
 
 template <>
-inline void vec_cross<float, platform::jit::avx512f>(
-    const int n, const float* x, const float* y, const float* z, float* out) {
+inline void vec_cross<float, platform::avx512f>(const int n, const float* x,
+                                                const float* y, const float* z,
+                                                float* out) {
   // TODO(TJ): enable me
-  vec_cross<float, platform::jit::avx>(n, x, y, z, out);
+  vec_cross<float, platform::avx>(n, x, y, z, out);
 }
 
-template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 inline void vec_add_bias(const int n, const T a, const T* x, T* y) {
   for (int i = 0; i < n; ++i) {
     y[i] = x[i] + a;
@@ -254,12 +253,12 @@ inline void vec_add_bias(const int n, const T a, const T* x, T* y) {
 }
 
 template <>
-inline void vec_add_bias<float, platform::jit::avx>(const int n, const float a,
+inline void vec_add_bias<float, platform::avx>(const int n, const float a,
                                                const float* x, float* y) {
 #ifdef __AVX__
   constexpr int block = YMM_FLOAT_BLOCK;
   if (n < block) {
-    vec_add_bias<float, platform::jit::isa_any>(n, a, x, y);
+    vec_add_bias<float, platform::isa_any>(n, a, x, y);
     return;
   }
   const int rest = n % block;
@@ -283,32 +282,30 @@ inline void vec_add_bias<float, platform::jit::avx>(const int n, const float a,
     y[i] = x[i] + a;
   }
 #else
-  vec_add_bias<float, platform::jit::isa_any>(n, a, x, y);
+  vec_add_bias<float, platform::isa_any>(n, a, x, y);
 #endif
 }
 
 template <>
-inline void vec_add_bias<float, platform::jit::avx2>(const int n, const float a,
+inline void vec_add_bias<float, platform::avx2>(const int n, const float a,
                                                 const float* x, float* y) {
-  vec_add_bias<float, platform::jit::avx>(n, a, x, y);
+  vec_add_bias<float, platform::avx>(n, a, x, y);
 }
 
 template <>
-inline void vec_add_bias<float, platform::jit::avx512f>(const int n,
-                                                        const float a,
-                                                        const float* x,
-                                                        float* y) {
+inline void vec_add_bias<float, platform::avx512f>(const int n, const float a,
+                                                   const float* x, float* y) {
   // TODO(TJ): enable me
-  vec_add_bias<float, platform::jit::avx2>(n, a, x, y);
+  vec_add_bias<float, platform::avx2>(n, a, x, y);
 }
 
-template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 inline void vec_identity(const int n, const T* x, T* y) {
   // do nothing
   return;
 }
 
-template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 inline void vec_sigmoid(const int n, const T* x, T* y) {
   const T min = SIGMOID_THRESHOLD_MIN;
   const T max = SIGMOID_THRESHOLD_MAX;
@@ -323,12 +320,12 @@ inline void vec_sigmoid(const int n, const T* x, T* y) {
 }
 
 template <>
-inline void vec_sigmoid<float, platform::jit::avx>(const int n, const float* x,
+inline void vec_sigmoid<float, platform::avx>(const int n, const float* x,
                                               float* y) {
 #ifdef __AVX__
   constexpr int block = YMM_FLOAT_BLOCK;
   if (n < block) {
-    vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+    vec_sigmoid<float, platform::isa_any>(n, x, y);
     return;
   }
   const int rest = n % block;
@@ -377,25 +374,24 @@ inline void vec_sigmoid<float, platform::jit::avx>(const int n, const float* x,
     y[i] = 1.f / (1.f + y[i]);
   }
 #else
-  vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
+  vec_sigmoid<float, platform::isa_any>(n, x, y);
 #endif
 }
 
 template <>
-inline void vec_sigmoid<float, platform::jit::avx2>(const int n, const float* x,
+inline void vec_sigmoid<float, platform::avx2>(const int n, const float* x,
                                                float* y) {
-  vec_sigmoid<float, platform::jit::avx>(n, x, y);
+  vec_sigmoid<float, platform::avx>(n, x, y);
 }
 
 template <>
-inline void vec_sigmoid<float, platform::jit::avx512f>(const int n,
-                                                       const float* x,
+inline void vec_sigmoid<float, platform::avx512f>(const int n, const float* x,
                                                   float* y) {
   // TODO(TJ): enable me
-  vec_sigmoid<float, platform::jit::avx2>(n, x, y);
+  vec_sigmoid<float, platform::avx2>(n, x, y);
 }
 
-template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 inline void vec_tanh(const int n, const T* x, T* y) {
   vec_scal<T, isa>(n, static_cast<T>(2), x, y);
   vec_sigmoid<T, isa>(n, y, y);
@@ -404,7 +400,7 @@ inline void vec_tanh(const int n, const T* x, T* y) {
 }
 
 // TODO(TJ): make relu clip
-template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 inline void vec_relu(const int n, const T* x, T* y) {
   for (int i = 0; i < n; ++i) {
     y[i] = x[i] > 0 ? x[i] : 0;
@@ -412,12 +408,12 @@ inline void vec_relu(const int n, const T* x, T* y) {
 }
 
 template <>
-inline void vec_relu<float, platform::jit::avx>(const int n, const float* x,
+inline void vec_relu<float, platform::avx>(const int n, const float* x,
                                            float* y) {
 #ifdef __AVX__
   constexpr int block = YMM_FLOAT_BLOCK;
   if (n < block * 4) {
-    vec_relu<float, platform::jit::isa_any>(n, x, y);
+    vec_relu<float, platform::isa_any>(n, x, y);
     return;
   }
 
@@ -441,26 +437,26 @@ inline void vec_relu<float, platform::jit::avx>(const int n, const float* x,
 #undef MOVE_ONE_STEP
 
 #else
-  vec_relu<float, platform::jit::isa_any>(n, x, y);
+  vec_relu<float, platform::isa_any>(n, x, y);
 #endif
 }
 
 template <>
-inline void vec_relu<float, platform::jit::avx2>(const int n, const float* x,
+inline void vec_relu<float, platform::avx2>(const int n, const float* x,
                                             float* y) {
-  vec_relu<float, platform::jit::avx>(n, x, y);
+  vec_relu<float, platform::avx>(n, x, y);
 }
 
 template <>
-inline void vec_relu<float, platform::jit::avx512f>(const int n, const float* x,
+inline void vec_relu<float, platform::avx512f>(const int n, const float* x,
                                                float* y) {
   // TODO(TJ): enable me
-  vec_relu<float, platform::jit::avx2>(n, x, y);
+  vec_relu<float, platform::avx2>(n, x, y);
 }
 
 // TODO(TJ): optimize double of sigmoid, tanh and relu if necessary
 
-template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+template <typename T, platform::cpu_isa_t isa = platform::isa_any>
 class VecActivations {
  public:
   std::function<void(const int, const T*, T*)> operator()(

paddle/fluid/operators/math/cpu_vec_test.cc

@@ -104,38 +104,42 @@ void TestAndBench(const int n, std::function<void(const int, const T*, T*)> tgt,
 }
 
 TEST(CpuVecTest, sigmoid) {
-  namespace jit = paddle::platform::jit;
+  namespace platform = paddle::platform;
   using namespace paddle::operators::math;  // NOLINT
   for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
     TestAndBench<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
-    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
-    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
-    TestAndBench<float>(sz, vec_sigmoid<float, jit::avx512f>,
+    TestAndBench<float>(sz, vec_sigmoid<float, platform::avx>,
+                        ref_sigmoid<float>);
+    TestAndBench<float>(sz, vec_sigmoid<float, platform::avx2>,
+                        ref_sigmoid<float>);
+    TestAndBench<float>(sz, vec_sigmoid<float, platform::avx512f>,
                         ref_sigmoid<float>);
   }
   TestAndBench<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
 }
 
 TEST(CpuVecTest, tanh) {
-  namespace jit = paddle::platform::jit;
+  namespace platform = paddle::platform;
   using namespace paddle::operators::math;  // NOLINT
   for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
     TestAndBench<float>(sz, vec_tanh<float>, ref_tanh<float>);
-    TestAndBench<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
-    TestAndBench<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
-    TestAndBench<float>(sz, vec_tanh<float, jit::avx512f>, ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, platform::avx>, ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, platform::avx2>, ref_tanh<float>);
+    TestAndBench<float>(sz, vec_tanh<float, platform::avx512f>,
+                        ref_tanh<float>);
   }
   TestAndBench<double>(30, vec_tanh<double>, ref_tanh<double>);
 }
 
 TEST(CpuVecTest, relu) {
-  namespace jit = paddle::platform::jit;
+  namespace platform = paddle::platform;
   using namespace paddle::operators::math;  // NOLINT
   for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
     TestAndBench<float>(sz, vec_relu<float>, ref_relu<float>);
-    TestAndBench<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
-    TestAndBench<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
-    TestAndBench<float>(sz, vec_relu<float, jit::avx512f>, ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, platform::avx>, ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, platform::avx2>, ref_relu<float>);
+    TestAndBench<float>(sz, vec_relu<float, platform::avx512f>,
+                        ref_relu<float>);
   }
   TestAndBench<double>(30, vec_relu<double>, ref_relu<double>);
 }
@@ -162,38 +166,40 @@ void TestInplace(const int n, std::function<void(const int, const T*, T*)> tgt,
 }
 
 TEST(CpuVecTest, inplace_sigmoid) {
-  namespace jit = paddle::platform::jit;
+  namespace platform = paddle::platform;
   using namespace paddle::operators::math;  // NOLINT
   for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
     TestInplace<float>(sz, vec_sigmoid<float>, ref_sigmoid<float>);
-    TestInplace<float>(sz, vec_sigmoid<float, jit::avx>, ref_sigmoid<float>);
-    TestInplace<float>(sz, vec_sigmoid<float, jit::avx2>, ref_sigmoid<float>);
-    TestInplace<float>(sz, vec_sigmoid<float, jit::avx512f>,
+    TestInplace<float>(sz, vec_sigmoid<float, platform::avx>,
+                       ref_sigmoid<float>);
+    TestInplace<float>(sz, vec_sigmoid<float, platform::avx2>,
+                       ref_sigmoid<float>);
+    TestInplace<float>(sz, vec_sigmoid<float, platform::avx512f>,
                        ref_sigmoid<float>);
   }
   TestInplace<double>(30, vec_sigmoid<double>, ref_sigmoid<double>);
 }
 
 TEST(CpuVecTest, inplace_tanh) {
-  namespace jit = paddle::platform::jit;
+  namespace platform = paddle::platform;
   using namespace paddle::operators::math;  // NOLINT
   for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
     TestInplace<float>(sz, vec_tanh<float>, ref_tanh<float>);
-    TestInplace<float>(sz, vec_tanh<float, jit::avx>, ref_tanh<float>);
-    TestInplace<float>(sz, vec_tanh<float, jit::avx2>, ref_tanh<float>);
-    TestInplace<float>(sz, vec_tanh<float, jit::avx512f>, ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, platform::avx>, ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, platform::avx2>, ref_tanh<float>);
+    TestInplace<float>(sz, vec_tanh<float, platform::avx512f>, ref_tanh<float>);
   }
   TestInplace<double>(30, vec_tanh<double>, ref_tanh<double>);
 }
 
 TEST(CpuVecTest, inplace_relu) {
-  namespace jit = paddle::platform::jit;
+  namespace platform = paddle::platform;
   using namespace paddle::operators::math;  // NOLINT
   for (auto sz : {1, 2, 15, 16, 30, 32, 128, 200, 512}) {
     TestInplace<float>(sz, vec_relu<float>, ref_relu<float>);
-    TestInplace<float>(sz, vec_relu<float, jit::avx>, ref_relu<float>);
-    TestInplace<float>(sz, vec_relu<float, jit::avx2>, ref_relu<float>);
-    TestInplace<float>(sz, vec_relu<float, jit::avx512f>, ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, platform::avx>, ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, platform::avx2>, ref_relu<float>);
+    TestInplace<float>(sz, vec_relu<float, platform::avx512f>, ref_relu<float>);
   }
   TestInplace<double>(30, vec_relu<double>, ref_relu<double>);
 }

paddle/fluid/operators/math/jit_code.cc

@@ -22,7 +22,7 @@ namespace math {
 namespace jitkernel {
 namespace gen {
 
-using namespace platform::jit;  // NOLINT
+using namespace platform;  // NOLINT
 
 bool VXXJitCode::init(int d, int scalar_index) {
   // It's not necessary to use avx512 since it would slow down the frequency

paddle/fluid/operators/math/jit_code.h

@@ -179,7 +179,7 @@ class VActJitCode : public JitCode {
   template <typename JMM>
   void exp_jmm(JMM& dst, JMM& src, int src_idx = 11, int fx_idx = 12,  // NOLINT
                int fy_idx = 13, int mask_idx = 14, int tmp_idx = 15) {
-    using namespace platform::jit;  // NOLINT
+    using namespace platform;  // NOLINT
     // check all idx can not equal
     JMM jmm_src = JMM(src_idx);
     JMM jmm_fx = JMM(fx_idx);

paddle/fluid/operators/math/jit_gen.cc

@@ -36,7 +36,7 @@ void JitCode::preCode() {
   for (int i = 0; i < num_g_abi_regs; ++i) {
     push(Xbyak::Reg64(g_abi_regs[i]));
   }
-  if (platform::jit::MayIUse(platform::jit::avx512f)) {
+  if (platform::MayIUse(platform::avx512f)) {
     mov(reg_EVEX_max_8b_offt, 2 * EVEX_max_8b_offt);
   }
 }

paddle/fluid/operators/math/jit_kernel.cc

@@ -21,8 +21,6 @@ namespace operators {
 namespace math {
 namespace jitkernel {
 
-namespace jit = platform::jit;
-
 KernelPool& KernelPool::Instance() {
   static thread_local KernelPool g_jit_kernels;
   return g_jit_kernels;

paddle/fluid/operators/math/jit_kernel_blas.cc

@@ -30,7 +30,6 @@ namespace paddle {
 namespace operators {
 namespace math {
 namespace jitkernel {
-namespace jit = platform::jit;
 
 #ifdef PADDLE_WITH_MKLML
 template <typename T>
@@ -125,7 +124,7 @@ bool VMulKernelImpl<float>::useJIT(int d) {
 #ifdef PADDLE_WITH_MKLML
 template <>
 bool VMulKernelImpl<float>::useMKL(int d) {
-  return jit::MayIUse(jit::avx512f) && d > 512;
+  return platform::MayIUse(platform::avx512f) && d > 512;
 }
 
 template <>

paddle/fluid/operators/math/jit_kernel_crf_decode.cc

@@ -25,10 +25,8 @@ namespace operators {
 namespace math {
 namespace jitkernel {
 
-namespace jit = platform::jit;
-
 /* CRF Decode JitKernel */
-template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+template <typename T, platform::cpu_isa_t isa, jit_block>
 class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
  public:
   explicit CRFDecodeKernelImpl(int tag_num) : CRFDecodeKernel<T>() {
@@ -101,7 +99,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
 
 #define INTRIAVX_FLOAT(block)                                                  \
   template <>                                                                  \
-  CRFDecodeKernelImpl<float, jit::avx, block>::CRFDecodeKernelImpl(            \
+  CRFDecodeKernelImpl<float, platform::avx, block>::CRFDecodeKernelImpl(       \
       int tag_num)                                                             \
       : CRFDecodeKernel<float>() {                                             \
     this->num_ = tag_num;                                                      \
@@ -109,7 +107,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
     this->rest_ = this->num_ % YMM_FLOAT_BLOCK;                                \
   }                                                                            \
   template <>                                                                  \
-  void CRFDecodeKernelImpl<float, jit::avx, block>::Compute(                   \
+  void CRFDecodeKernelImpl<float, platform::avx, block>::Compute(              \
       const int seq_len, const float* x, const float* w, float* alpha,         \
       int* track) const {                                                      \
     INIT_ALPHA(YMM_FLOAT_BLOCK)                                                \
@@ -204,7 +202,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
 
 #define INTRIAVX512_FLOAT(block)                                               \
   template <>                                                                  \
-  CRFDecodeKernelImpl<float, jit::avx512f, block>::CRFDecodeKernelImpl(        \
+  CRFDecodeKernelImpl<float, platform::avx512f, block>::CRFDecodeKernelImpl(   \
       int tag_num)                                                             \
       : CRFDecodeKernel<float>() {                                             \
     this->num_ = tag_num;                                                      \
@@ -212,7 +210,7 @@ class CRFDecodeKernelImpl : public CRFDecodeKernel<T> {
     this->rest_ = this->num_ % ZMM_FLOAT_BLOCK;                                \
   }                                                                            \
   template <>                                                                  \
-  void CRFDecodeKernelImpl<float, jit::avx512f, block>::Compute(               \
+  void CRFDecodeKernelImpl<float, platform::avx512f, block>::Compute(          \
       const int seq_len, const float* x, const float* w, float* alpha,         \
       int* track) const {                                                      \
     INIT_ALPHA(ZMM_FLOAT_BLOCK)                                                \
@@ -270,14 +268,14 @@ INTRIAVX_FLOAT(kEQ16);
 INTRIAVX_FLOAT(kGT16);
 #endif
 #ifdef __AVX2__
-INTRIAVX2_FLOAT(jit::avx2, kEQ8);
-INTRIAVX2_FLOAT(jit::avx2, kGT8LT16);
-INTRIAVX2_FLOAT(jit::avx2, kEQ16);
-INTRIAVX2_FLOAT(jit::avx2, kGT16);
+INTRIAVX2_FLOAT(platform::avx2, kEQ8);
+INTRIAVX2_FLOAT(platform::avx2, kGT8LT16);
+INTRIAVX2_FLOAT(platform::avx2, kEQ16);
+INTRIAVX2_FLOAT(platform::avx2, kGT16);
 #endif
 #ifdef __AVX512F__
-INTRIAVX2_FLOAT(jit::avx512f, kEQ8);
-INTRIAVX2_FLOAT(jit::avx512f, kGT8LT16);
+INTRIAVX2_FLOAT(platform::avx512f, kEQ8);
+INTRIAVX2_FLOAT(platform::avx512f, kGT8LT16);
 INTRIAVX512_FLOAT(kEQ16);
 INTRIAVX512_FLOAT(kGT16);
 #endif

paddle/fluid/operators/math/jit_kernel_exp.cc

@@ -29,7 +29,6 @@ namespace paddle {
 namespace operators {
 namespace math {
 namespace jitkernel {
-namespace jit = platform::jit;
 
 #ifdef PADDLE_WITH_MKLML
 // try to use MKL to speedup

paddle/fluid/operators/math/jit_kernel_layer_norm.cc

@@ -22,10 +22,8 @@ namespace operators {
 namespace math {
 namespace jitkernel {
 
-namespace jit = platform::jit;
-
 /* Layer Norm JitKernel */
-template <typename T, platform::jit::cpu_isa_t isa, jit_block>
+template <typename T, platform::cpu_isa_t isa, jit_block>
 class LayerNormKernelImpl : public LayerNormKernel<T> {
  public:
   explicit LayerNormKernelImpl(int right) : LayerNormKernel<T>() {
@@ -90,7 +88,7 @@ class LayerNormKernelImpl : public LayerNormKernel<T> {
     this->end_ = this->num_ - this->rest_;                                     \
   }                                                                            \
   template <>                                                                  \
-  void LayerNormKernelImpl<float, jit::avx, block>::Compute(                   \
+  void LayerNormKernelImpl<float, platform::avx, block>::Compute(              \
       float* x, float* out, float* mean, float* var, const float* scale,       \
       const float* bias, int height, const float epsilon) const {              \
     __m256 sum;                                                                \
@@ -219,16 +217,16 @@ class LayerNormKernelImpl : public LayerNormKernel<T> {
   }
 
 #ifdef __AVX__
-INTRIAVX_FLOAT(jit::avx, kEQ8);
-INTRIAVX_FLOAT(jit::avx, kGT8LT16);
-INTRIAVX_FLOAT(jit::avx, kEQ16);
-INTRIAVX_FLOAT(jit::avx, kGT16);
+INTRIAVX_FLOAT(platform::avx, kEQ8);
+INTRIAVX_FLOAT(platform::avx, kGT8LT16);
+INTRIAVX_FLOAT(platform::avx, kEQ16);
+INTRIAVX_FLOAT(platform::avx, kGT16);
 #endif
 #ifdef __AVX2__
-INTRIAVX_FLOAT(jit::avx2, kEQ8);
-INTRIAVX_FLOAT(jit::avx2, kGT8LT16);
-INTRIAVX_FLOAT(jit::avx2, kEQ16);
-INTRIAVX_FLOAT(jit::avx2, kGT16);
+INTRIAVX_FLOAT(platform::avx2, kEQ8);
+INTRIAVX_FLOAT(platform::avx2, kGT8LT16);
+INTRIAVX_FLOAT(platform::avx2, kEQ16);
+INTRIAVX_FLOAT(platform::avx2, kGT16);
 #endif
 
 #undef INTRIAVX_FLOAT

paddle/fluid/operators/math/jit_kernel_macro.h

@@ -92,7 +92,6 @@ namespace jitkernel {
                           JITKERNEL_DECLARE, JITKERNEL_FIND_KEY,     \
                           JITKERNEL_IMPL)
 
-namespace jit = platform::jit;
 // TODO(TJ): below defines are deprecated, would be remove recently
 #define SEARCH_BLOCK(macro_, ker, dtype, isa)              \
   if (d < YMM_FLOAT_BLOCK) {                               \
@@ -108,14 +107,14 @@ namespace jit = platform::jit;
   }
 
 #define SEARCH_ISA_BLOCK(macro_, ker, dtype)             \
-  if (jit::MayIUse(jit::avx512f)) {                 \
-    SEARCH_BLOCK(macro_, ker, dtype, jit::avx512f); \
-  } else if (jit::MayIUse(jit::avx2)) {             \
-    SEARCH_BLOCK(macro_, ker, dtype, jit::avx2);    \
-  } else if (jit::MayIUse(jit::avx)) {              \
-    SEARCH_BLOCK(macro_, ker, dtype, jit::avx);     \
+  if (platform::MayIUse(platform::avx512f)) {            \
+    SEARCH_BLOCK(macro_, ker, dtype, platform::avx512f); \
+  } else if (platform::MayIUse(platform::avx2)) {        \
+    SEARCH_BLOCK(macro_, ker, dtype, platform::avx2);    \
+  } else if (platform::MayIUse(platform::avx)) {         \
+    SEARCH_BLOCK(macro_, ker, dtype, platform::avx);     \
   } else {                                               \
-    SEARCH_BLOCK(macro_, ker, dtype, jit::isa_any); \
+    SEARCH_BLOCK(macro_, ker, dtype, platform::isa_any); \
   }
 
 #define JITKERNEL_KEY(ker_key, dtype_key) \
@@ -156,10 +155,10 @@ namespace jit = platform::jit;
                                   marco_declare, macro_key, macro_impl)
 
 #define FOR_EACH_ISA(macro_, block) \
-  macro_(jit::avx512f, block);      \
-  macro_(jit::avx2, block);         \
-  macro_(jit::avx, block);          \
-  macro_(jit::isa_any, block)
+  macro_(platform::avx512f, block); \
+  macro_(platform::avx2, block);    \
+  macro_(platform::avx, block);     \
+  macro_(platform::isa_any, block)
 
 #define FOR_EACH_BLOCK(macro_, isa) \
   macro_(isa, kLT8);                \
@@ -169,10 +168,10 @@ namespace jit = platform::jit;
   macro_(isa, kGT16)
 
 #define FOR_EACH_ISA_BLOCK(macro_)           \
-  FOR_EACH_BLOCK(macro_, jit::avx512f); \
-  FOR_EACH_BLOCK(macro_, jit::avx2);    \
-  FOR_EACH_BLOCK(macro_, jit::avx);     \
-  FOR_EACH_BLOCK(macro_, jit::isa_any)
+  FOR_EACH_BLOCK(macro_, platform::avx512f); \
+  FOR_EACH_BLOCK(macro_, platform::avx2);    \
+  FOR_EACH_BLOCK(macro_, platform::avx);     \
+  FOR_EACH_BLOCK(macro_, platform::isa_any)
 
 }  // namespace jitkernel
 }  // namespace math

paddle/fluid/operators/math/jit_kernel_test.cc

@@ -705,7 +705,7 @@ TEST(JitKernel, pool) {
   jit::lstm_attr_t attr(frame_size, act_gate, act_cand, act_cell, false);
 
   // empty call it to avoid unknown flag 'use_pinned_memory' on Mac
-  paddle::platform::jit::MayIUse(paddle::platform::jit::avx);
+  paddle::platform::MayIUse(paddle::platform::avx);
   const auto& plstm1 =
       jit::KernelPool::Instance()
           .template Get<jit::LSTMKernel<float>, const jit::lstm_attr_t&>(attr);

paddle/fluid/operators/math/matrix_bit_code.cc

@@ -89,6 +89,8 @@ template <typename T>
 void MatrixBitCodeFunctor<T>::Mul(framework::Tensor* tmat,
                                   const framework::Tensor& weight,
                                   const framework::Tensor& input) {
+  auto blas =
+      GetBlas<platform::CPUDeviceContext, T>(platform::CPUDeviceContext());
   size_t num_samples = tmat->dims()[0];
   size_t tmat_width = tmat->dims()[1];
   size_t input_width = input.dims()[1];
@@ -99,13 +101,12 @@ void MatrixBitCodeFunctor<T>::Mul(framework::Tensor* tmat,
   for (size_t i = 0; i < num_samples; ++i) {
     auto code = code_table_->get_code(i);
     int code_length = code->get_length();
+    const T* input_row = input_value + input_width * i;
     for (int j = 0; j < code_length; ++j) {
       size_t index = code->calc_index(j);
+      const T* weight_row = weight_value + weight_width * index;
       T sum = static_cast<T>(0.0);
-      for (size_t k = 0; k < input_width; ++k) {
-        sum += weight_value[weight_width * index + k] *
-               input_value[input_width * i + k];
-      }
+      sum = blas.DOT(input_width, weight_row, input_row);
       tmat_value[i * tmat_width + j] += sum;
     }
   }
@@ -115,6 +116,8 @@ template <typename T>
 void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor& tmat,
                                             framework::Tensor* weight,
                                             const framework::Tensor& input) {
+  auto blas =
+      GetBlas<platform::CPUDeviceContext, T>(platform::CPUDeviceContext());
   size_t num_samples = tmat.dims()[0];
   size_t input_width = input.dims()[1];
   size_t tmat_width = tmat.dims()[1];
@@ -122,17 +125,26 @@ void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor& tmat,
   auto tmat_value = tmat.data<T>();
   auto weight_value = weight->data<T>();
   auto input_value = input.data<T>();
+
+  std::unordered_map<int, std::vector<std::pair<T, const T*>>> ops;
+
   for (size_t i = 0; i < num_samples; ++i) {
     auto code = code_table_->get_code(i);
     int code_length = code->get_length();
+    const T* input_value_row = input_value + input_width * i;
+    const T* tmat_row = tmat_value + i * tmat_width;
     for (int j = 0; j < code_length; ++j) {
-      size_t index = code->calc_index(j);
-
-      for (size_t k = 0; k < input_width; ++k) {
-        weight_value[weight_width * index + k] +=
-            tmat_value[i * tmat_width + j] * input_value[input_width * i + k];
+      ops[code->calc_index(j)].emplace_back(tmat_row[j], input_value_row);
     }
   }
+  for (auto& op : ops) {
+    auto& op_in_row = op.second;
+    for (auto& pair : op_in_row) {
+      auto& scale = pair.first;
+      auto* input_row = pair.second;
+      T* weight_row = weight_value + op.first * weight_width;
+      blas.AXPY(input_width, scale, input_row, weight_row);
+    }
   }
 }
 
@@ -140,6 +152,8 @@ template <typename T>
 void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor& tmat,
                                             framework::SelectedRows* weight,
                                             const framework::Tensor& input) {
+  auto blas =
+      GetBlas<platform::CPUDeviceContext, T>(platform::CPUDeviceContext());
   size_t num_samples = tmat.dims()[0];
   size_t input_width = input.dims()[1];
   size_t tmat_width = tmat.dims()[1];
@@ -147,17 +161,28 @@ void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor& tmat,
   auto tmat_value = tmat.data<T>();
   auto weight_value = weight->mutable_value()->data<T>();
   auto input_value = input.data<T>();
+
+  std::unordered_map<int, std::vector<std::pair<T, const T*>>> ops;
+  ops.reserve(weight->rows().size());
+
   for (size_t i = 0; i < num_samples; ++i) {
     auto code = code_table_->get_code(i);
     int code_length = code->get_length();
+    const T* input_value_row = input_value + input_width * i;
+    const T* tmat_row = tmat_value + i * tmat_width;
     for (int j = 0; j < code_length; ++j) {
-      size_t index = code->calc_index(j);
-      for (size_t k = 0; k < input_width; ++k) {
-        int64_t row_index = weight->GetIndexFromId(static_cast<int64_t>(index));
-        weight_value[row_index * weight_width + k] +=
-            tmat_value[i * tmat_width + j] * input_value[input_width * i + k];
+      ops[code->calc_index(j)].emplace_back(tmat_row[j], input_value_row);
+    }
   }
+
+  for (auto& row : weight->rows()) {
+    auto& op_in_row = ops[row];
+    for (auto& pair : op_in_row) {
+      auto& scale = pair.first;
+      auto* input_row = pair.second;
+      blas.AXPY(input_width, scale, input_row, weight_value);
     }
+    weight_value += weight_width;
   }
 }
 

paddle/fluid/operators/math/matrix_bit_code.h

@@ -13,10 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <unordered_map>
+#include <utility>
+#include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/platform/device_context.h"
 
 #if defined(_WIN32)

paddle/fluid/operators/math/prelu.cu

+/* Copyright (c) 2016 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. */
+
+#include "paddle/fluid/operators/math/prelu.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+static const int CUDA_NUM_THREADS = 1024;
+static const int CUDA_MAX_NUM_BLOCKS = 65535;
+inline static int GET_NUM_BLOCKS(const int N) {
+  return (N + CUDA_NUM_THREADS - 1) / CUDA_NUM_THREADS;
+}
+
+template <typename T>
+__global__ void PReluChannelWiseKernel(const T *input, const T *alpha,
+                                       T *output, int channel,
+                                       size_t spatial_size) {
+  size_t offset = blockIdx.x * spatial_size;
+  const T *in = input + offset;
+  T *out = output + offset;
+  T scale = alpha[blockIdx.x % channel];
+
+  for (size_t i = threadIdx.x; i < spatial_size; i += blockDim.x) {
+    T x = in[i];
+    out[i] = (x > 0) ? x : scale * x;
+  }
+}
+
+template <typename T>
+__global__ void PReluElementWiseKernel(const T *input, const T *alpha,
+                                       T *output, size_t spatial_size) {
+  size_t offset = blockIdx.x * spatial_size;
+  const T *in = input + offset;
+  const T *scale = alpha + offset;
+  T *out = output + offset;
+
+  for (size_t i = threadIdx.x; i < spatial_size; i += blockDim.x) {
+    T x = in[i];
+    out[i] = (x > 0) ? x : scale[i] * x;
+  }
+}
+
+template <typename T>
+__global__ void PReluScalarKernel(const T *input, const T *alpha, T *output,
+                                  size_t spatial_size) {
+  size_t offset = blockIdx.x * spatial_size;
+  const T *in = input + offset;
+  T scale = *alpha;
+  T *out = output + offset;
+
+  for (size_t i = threadIdx.x; i < spatial_size; i += blockDim.x) {
+    T x = in[i];
+    out[i] = (x > 0) ? x : scale * x;
+  }
+}
+
+template <typename T>
+static inline void PReluChannelWise(cudaStream_t stream, const T *input,
+                                    const T *alpha, T *output,
+                                    std::vector<int> input_shape) {
+  size_t unroll = input_shape[0] * input_shape[1];
+  size_t spatial_size = input_shape[2] * input_shape[3];
+  CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS);
+  PReluChannelWiseKernel<<<unroll, CUDA_NUM_THREADS, 0, stream>>>(
+      input, alpha, output, input_shape[1], spatial_size);
+}
+
+template <typename T>
+static inline void PReluElementWise(cudaStream_t stream, const T *input,
+                                    const T *alpha, T *output,
+                                    std::vector<int> input_shape) {
+  size_t unroll = input_shape[0] * input_shape[1];
+  size_t spatial_size = input_shape[2] * input_shape[3];
+  CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS);
+  PReluElementWiseKernel<<<unroll, CUDA_NUM_THREADS, 0, stream>>>(
+      input, alpha, output, spatial_size);
+}
+
+template <typename T>
+static inline void PReluScalar(cudaStream_t stream, const T *input,
+                               const T *alpha, T *output,
+                               std::vector<int> input_shape) {
+  size_t unroll = input_shape[0] * input_shape[1];
+  size_t spatial_size = input_shape[2] * input_shape[3];
+  CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS);
+  PReluScalarKernel<<<unroll, CUDA_NUM_THREADS, 0, stream>>>(
+      input, alpha, output, spatial_size);
+}
+
+template <typename T>
+void PreluChannelWiseDirectCUDAFunctor<T>::operator()(
+    cudaStream_t stream, const T *input, const T *alpha, T *output,
+    std::vector<int> input_shape) {
+  size_t unroll = input_shape[0] * input_shape[1];
+  size_t spatial_size = input_shape[2] * input_shape[3];
+  CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS);
+  PReluChannelWiseKernel<<<unroll, CUDA_NUM_THREADS, 0, stream>>>(
+      input, alpha, output, input_shape[1], spatial_size);
+}
+
+template <typename T>
+void PreluElementWiseDirectCUDAFunctor<T>::operator()(
+    cudaStream_t stream, const T *input, const T *alpha, T *output,
+    std::vector<int> input_shape) {
+  size_t unroll = input_shape[0] * input_shape[1];
+  size_t spatial_size = input_shape[2] * input_shape[3];
+  CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS);
+  PReluElementWiseKernel<<<unroll, CUDA_NUM_THREADS, 0, stream>>>(
+      input, alpha, output, spatial_size);
+}
+
+template <typename T>
+void PreluScalarDirectCUDAFunctor<T>::operator()(cudaStream_t stream,
+                                                 const T *input, const T *alpha,
+                                                 T *output,
+                                                 std::vector<int> input_shape) {
+  size_t unroll = input_shape[0] * input_shape[1];
+  size_t spatial_size = input_shape[2] * input_shape[3];
+  CHECK_LT(unroll, CUDA_MAX_NUM_BLOCKS);
+  PReluScalarKernel<<<unroll, CUDA_NUM_THREADS, 0, stream>>>(
+      input, alpha, output, spatial_size);
+}
+
+template class PreluChannelWiseDirectCUDAFunctor<float>;
+template class PreluChannelWiseDirectCUDAFunctor<double>;
+
+template class PreluElementWiseDirectCUDAFunctor<float>;
+template class PreluElementWiseDirectCUDAFunctor<double>;
+
+template class PreluScalarDirectCUDAFunctor<float>;
+template class PreluScalarDirectCUDAFunctor<double>;
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/prelu.h

+/* Copyright (c) 2016 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. */
+
+#pragma once
+#include <vector>
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/platform/cudnn_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+#ifdef PADDLE_WITH_CUDA
+template <typename T>
+class PreluChannelWiseDirectCUDAFunctor {
+ public:
+  void operator()(cudaStream_t stream, const T *input, const T *alpha,
+                  T *output, std::vector<int> input_shape);
+};
+
+template <typename T>
+class PreluElementWiseDirectCUDAFunctor {
+ public:
+  void operator()(cudaStream_t stream, const T *input, const T *alpha,
+                  T *output, std::vector<int> input_shape);
+};
+
+template <typename T>
+class PreluScalarDirectCUDAFunctor {
+ public:
+  void operator()(cudaStream_t stream, const T *input, const T *alpha,
+                  T *output, std::vector<int> input_shape);
+};
+#endif
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/softmax_impl.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/tensor.h"
 

paddle/fluid/operators/merge_selected_rows_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. */
+
+#include "paddle/fluid/operators/merge_selected_rows_op.h"
+
+namespace paddle {
+namespace operators {
+
+class MergeSelectedRowsOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of MergeSelectedRowsOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of MergeSelectedRowsOp should not be null.");
+    ctx->ShareDim("X", /*->*/ "Out");
+  }
+};
+
+class MergeSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "The input type is SelectedRows, and the selected rows may be "
+             "duplicated.");
+    AddOutput("Out",
+              "The output type is SelectedRows, and the selected rows are not "
+              "duplicated.");
+    AddComment(
+        R"DOC(
+MergeSelectedRows Operator.
+
+MergeSelectedRows is used to merge the duplicated rows of the input.
+)DOC");
+  }
+};
+
+class MergeSelectedRowsOpInferVarType
+    : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return std::unordered_map<std::string, std::string>{{"X", /*->*/ "Out"}};
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+REGISTER_OPERATOR(merge_selected_rows, ops::MergeSelectedRowsOp,
+                  ops::MergeSelectedRowsOpMaker,
+                  ops::MergeSelectedRowsOpInferVarType);
+
+REGISTER_OP_CPU_KERNEL(
+    merge_selected_rows,
+    ops::MergeSelectedRowsKernel<plat::CPUDeviceContext, float>,
+    ops::MergeSelectedRowsKernel<plat::CPUDeviceContext, double>);

paddle/fluid/operators/merge_selected_rows_op.cu.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. */
+
+#include "paddle/fluid/operators/merge_selected_rows_op.h"
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
+REGISTER_OP_CUDA_KERNEL(
+    merge_selected_rows,
+    ops::MergeSelectedRowsKernel<plat::CUDADeviceContext, float>,
+    ops::MergeSelectedRowsKernel<plat::CUDADeviceContext, double>);

paddle/fluid/operators/merge_selected_rows_op.h

+/* 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. */
+
+#pragma once
+#include <string>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/selected_rows_functor.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class MergeSelectedRowsKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* x = context.Input<framework::SelectedRows>("X");
+    auto* out = context.Output<framework::SelectedRows>("Out");
+
+    math::scatter::MergeAdd<DeviceContext, T> merge_func;
+    merge_func(context.template device_context<DeviceContext>(), *x, out);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/prelu_op.cc

@@ -58,7 +58,7 @@ class PReluOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
-        platform::CPUPlace());
+        ctx.device_context());
   }
 };
 

paddle/fluid/operators/prelu_op.cu

+/* Copyright (c) 2016 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. */
+
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/prelu.h"
+#include "paddle/fluid/operators/prelu_op.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename DeviceContext, typename T>
+class CUDAPReluKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* x = context.Input<Tensor>("X");
+    auto* alpha = context.Input<Tensor>("Alpha");
+    auto* out = context.Output<Tensor>("Out");
+
+    const T* x_ptr = x->data<T>();
+    T* o_ptr = out->mutable_data<T>(context.GetPlace());
+
+    const T* alpha_ptr = alpha->data<T>();
+    auto& mode = context.Attr<std::string>("mode");
+
+    int numel = x->numel();
+    auto dim = x->dims();
+    std::vector<int> input_shape = framework::vectorize2int(dim);
+
+    if (mode == "channel") {
+      math::PreluChannelWiseDirectCUDAFunctor<T> prelu_channel_wise;
+      prelu_channel_wise(context.cuda_device_context().stream(), x_ptr,
+                         alpha_ptr, o_ptr, input_shape);
+    } else if (mode == "element") {
+      math::PreluElementWiseDirectCUDAFunctor<T> prelu_element_wise;
+      prelu_element_wise(context.cuda_device_context().stream(), x_ptr,
+                         alpha_ptr, o_ptr, input_shape);
+    } else {
+      math::PreluScalarDirectCUDAFunctor<T> prelu_scalar;
+      prelu_scalar(context.cuda_device_context().stream(), x_ptr, alpha_ptr,
+                   o_ptr, input_shape);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    prelu, ops::CUDAPReluKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::CUDAPReluKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/platform/cpu_info.cc

@@ -123,7 +123,6 @@ size_t CUDAPinnedMaxChunkSize() {
   return CUDAPinnedMaxAllocSize() / 256;
 }
 
-namespace jit {
 #ifdef PADDLE_WITH_XBYAK
 static Xbyak::util::Cpu cpu;
 bool MayIUse(const cpu_isa_t cpu_isa) {
@@ -165,6 +164,5 @@ bool MayIUse(const cpu_isa_t cpu_isa) {
 }
 #endif
 
-}  // namespace jit
 }  // namespace platform
 }  // namespace paddle