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

Author: reyoung

benchmark/fluid/README.md

@@ -29,9 +29,11 @@ Currently supported `--model` argument include:
     You can choose to use GPU/CPU training. With GPU training, you can specify
     `--gpus <gpu_num>` to run multi GPU training.
 * Run distributed training with parameter servers:
+    * see [run_fluid_benchmark.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/fluid/run_fluid_benchmark.sh) as an example.
     * start parameter servers:
         ```bash
         PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist  --device GPU --update_method pserver
+        sleep 15
         ```
     * start trainers:
         ```bash

benchmark/fluid/run_fluid_benchmark.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+
+PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model resnet --device CPU --update_method pserver --iterations=10000 &
+
+sleep 15
+
+CUDA_VISIBLE_DEVICES=0,1 PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model resnet --device GPU --update_method pserver --iterations=10000 --gpus 2 &
+
+CUDA_VISIBLE_DEVICES=2,3 PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=1 python fluid_benchmark.py --model resnet --device GPU --update_method pserver --iterations=10000 --gpus 2 &

paddle/fluid/framework/CMakeLists.txt

@@ -87,7 +87,7 @@ cc_library(executor SRCS executor.cc DEPS op_registry device_context scope
 framework_proto glog lod_rank_table feed_fetch_method)
 
 
-cc_library(parallel_executor SRCS parallel_executor.cc DEPS multi_devices_graph_builder threaded_ssa_graph_executor scope_buffered_ssa_graph_executor)
+cc_library(parallel_executor SRCS parallel_executor.cc DEPS graph_builder_factory threaded_ssa_graph_executor scope_buffered_ssa_graph_executor)
 
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)

paddle/fluid/framework/details/CMakeLists.txt

@@ -7,6 +7,7 @@ cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place
 
 cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base)
 cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph)
+cc_library(ssa_graph_printer SRCS ssa_graph_printer.cc DEPS ssa_graph_builder)
 
 cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
 
@@ -28,6 +29,9 @@ cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope d
 cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle
         scale_loss_grad_op_handle rpc_op_handle ${multi_devices_graph_builder_deps} reduce_op_handle broadcast_op_handle)
 
+
+cc_library(graph_builder_factory SRCS graph_builder_factory.cc DEPS multi_devices_graph_builder ssa_graph_printer)
+
 cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph framework_proto)
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
         simple_threadpool device_context)

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -59,8 +59,8 @@ struct BroadcastOpHandle : public OpHandleBase {
   void RunImpl() override;
 
  private:
-  const std::vector<Scope *> &local_scopes_;
-  const std::vector<platform::Place> &places_;
+  std::vector<Scope *> local_scopes_;
+  std::vector<platform::Place> places_;
 #ifdef PADDLE_WITH_CUDA
   const platform::NCCLContextMap *nccl_ctxs_;
 #endif

paddle/fluid/framework/details/build_strategy.h

@@ -14,6 +14,8 @@
 
 #pragma once
 
+#include <string>
+
 namespace paddle {
 namespace framework {
 namespace details {
@@ -29,6 +31,8 @@ struct BuildStrategy {
 
   ReduceStrategy reduce_{ReduceStrategy::kAllReduce};
   GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice};
+
+  std::string debug_graphviz_path_{""};
 };
 
 }  // namespace details

paddle/fluid/framework/details/graph_builder_factory.cc

@@ -0,0 +1,47 @@
+// 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/details/graph_builder_factory.h"
+#include <fstream>
+#include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
+#include "paddle/fluid/framework/details/ssa_graph_printer.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+std::unique_ptr<SSAGraphBuilder> SSAGraphBuilderFactory::Create() {
+  std::unique_ptr<SSAGraphBuilder> res(
+#ifdef PADDLE_WITH_CUDA
+      new MultiDevSSAGraphBuilder(places_, loss_var_name_, param_names_,
+                                  local_scopes_, nccl_ctxs_, strategy_)
+#else
+      new MultiDevSSAGraphBuilder(places_, loss_var_name_, param_names_,
+                                  local_scopes_, strategy_)
+#endif
+          );  // NOLINT
+
+  if (!strategy_.debug_graphviz_path_.empty()) {
+    std::unique_ptr<std::ostream> fout(
+        new std::ofstream(strategy_.debug_graphviz_path_));
+    PADDLE_ENFORCE(fout->good());
+    std::unique_ptr<GraphvizSSAGraphPrinter> graphviz_printer(
+        new GraphvizSSAGraphPrinter());
+    res.reset(new SSAGraghBuilderWithPrinter(
+        std::move(fout), std::move(graphviz_printer), std::move(res)));
+  }
+  return res;
+}
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/graph_builder_factory.h

@@ -0,0 +1,67 @@
+// 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 <memory>
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/details/build_strategy.h"
+#include "paddle/fluid/framework/details/ssa_graph_builder.h"
+#include "paddle/fluid/platform/place.h"
+
+#ifdef PADDLE_WITH_CUDA
+#include "paddle/fluid/platform/nccl_helper.h"
+#endif
+
+namespace paddle {
+namespace framework {
+class Scope;
+namespace details {
+
+class SSAGraphBuilderFactory {
+ public:
+  SSAGraphBuilderFactory(const std::vector<platform::Place>& places,
+                         const std::string& loss_var_name,
+                         const std::unordered_set<std::string>& param_names,
+                         const std::vector<Scope*>& local_scopes,
+                         const BuildStrategy& strategy)
+      : places_(places),
+        loss_var_name_(loss_var_name),
+        param_names_(param_names),
+        local_scopes_(local_scopes),
+        strategy_(strategy) {}
+
+#ifdef PADDLE_WITH_CUDA
+  void SetNCCLContextMap(platform::NCCLContextMap* nccl_ctxs) {
+    nccl_ctxs_ = nccl_ctxs;
+  }
+#endif
+
+  std::unique_ptr<SSAGraphBuilder> Create();
+
+ private:
+  std::vector<platform::Place> places_;
+  std::string loss_var_name_;
+  std::unordered_set<std::string> param_names_;
+  std::vector<Scope*> local_scopes_;
+  BuildStrategy strategy_;
+
+#ifdef PADDLE_WITH_CUDA
+  platform::NCCLContextMap* nccl_ctxs_;
+#endif
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -30,10 +30,6 @@
 #include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h"
 #endif
 
-DEFINE_string(ssa_graph_path, "/tmp/ssa_graph.dot",
-              "the ssa graph path only print with GLOG_v=10,"
-              "default /tmp/graph.dot");
-
 namespace paddle {
 namespace framework {
 namespace details {
@@ -277,11 +273,6 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
    */
   AddOutputToLeafOps(&result);
 
-  if (VLOG_IS_ON(10)) {
-    std::ofstream fout(FLAGS_ssa_graph_path);
-    PrintGraphviz(*graph, fout);
-  }
-
   return std::unique_ptr<SSAGraph>(graph);
 }
 

paddle/fluid/framework/details/nccl_all_reduce_op_handle.h

@@ -41,8 +41,8 @@ struct NCCLAllReduceOpHandle : public OpHandleBase {
   void RunImpl() override;
 
  private:
-  const std::vector<Scope *> &local_scopes_;
-  const std::vector<platform::Place> &places_;
+  std::vector<Scope *> local_scopes_;
+  std::vector<platform::Place> places_;
   const platform::NCCLContextMap &nccl_ctxs_;
 };