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

Author: reyoung

CMakeLists.txt

@@ -69,6 +69,7 @@ option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)
 option(WITH_GRPC     "Use grpc as the default rpc framework"            ${WITH_DISTRIBUTE})
 option(WITH_BRPC_RDMA     "Use brpc rdma as the rpc protocal"           OFF)
 option(WITH_INFERENCE    "Compile fluid inference library"              ON)
+option(WITH_INFERENCE_API_TEST   "Test fluid inference high-level api interface"  OFF)
 option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
 

doc/fluid/dev/releasing_process_en.md

@@ -1,6 +1,6 @@
 # PaddlePaddle Releasing Process
 
-PaddlePaddle manages its branches using "git-flow branching model", and [Semantic Versioning](http://semver.org/) as it's version number semantics.
+PaddlePaddle manages its branches using Trunk Based Development, and [Semantic Versioning](http://semver.org/) as it's version number semantics.
 
 Each time we release a new PaddlePaddle version, we should follow the below steps:
 

paddle/fluid/framework/details/CMakeLists.txt

@@ -28,10 +28,20 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_
 cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
 cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
 
+if(WITH_GPU)
+  cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle scale_loss_grad_op_handle rpc_op_handle
+          all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
+endif()
+
 cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
         scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle)
 
-cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto)
+if(WITH_GPU)
+  cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto reference_count_pass)
+else()
+  cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto)
+endif()
+
 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/computation_op_handle.h

@@ -32,6 +32,10 @@ struct ComputationOpHandle : public OpHandleBase {
 
   std::string Name() const override;
 
+  const Scope *GetScope() const { return scope_; }
+
+  const platform::Place &GetPlace() const { return place_; }
+
  protected:
   void RunImpl() override;
 

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -127,6 +127,9 @@ static const char kLocalScopes[] = "local_scopes";
 static const char kStrategy[] = "strategy";
 
 void MultiDevSSAGraphBuilder::Init() const {
+  all_vars_.clear();
+  balance_vars_.clear();
+
   loss_var_name_ = Get<const std::string>(kLossVarName);
   places_ = Get<const std::vector<platform::Place>>(kPlaces);
   local_scopes_ = Get<const std::vector<Scope *>>(kLocalScopes);

paddle/fluid/framework/details/multi_devices_graph_pass.h

@@ -40,12 +40,6 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
                          size_t device_id) const;
   void Init() const;
 
- private:
-  mutable std::string loss_var_name_;
-  mutable std::vector<platform::Place> places_;
-  mutable std::vector<Scope *> local_scopes_;
-  mutable std::unordered_set<std::string> grad_names_;
-
 #ifdef PADDLE_WITH_CUDA
   mutable platform::NCCLContextMap *nccl_ctxs_;
 #endif
@@ -95,13 +89,17 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
   size_t GetAppropriateDeviceID(
       const std::vector<std::string> &var_names) const;
 
- private:
+  void SetCommunicationContext(OpHandleBase *op_handle,
+                               const platform::Place &p) const;
+
+  mutable std::string loss_var_name_;
+  mutable std::vector<platform::Place> places_;
+  mutable std::vector<Scope *> local_scopes_;
+  mutable std::unordered_set<std::string> grad_names_;
+
   mutable BuildStrategy strategy_;
   mutable std::unordered_map<std::string, VarDesc *> all_vars_;
   mutable std::vector<int64_t> balance_vars_;
-
-  void SetCommunicationContext(OpHandleBase *op_handle,
-                               const platform::Place &p) const;
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/reference_count_op_handle.h

@@ -0,0 +1,123 @@
+// 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 <atomic>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "paddle/fluid/framework/details/op_handle_base.h"
+#include "paddle/fluid/framework/garbage_collector.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/tensor.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+using ReferenceCountMap = std::unordered_map<std::string, int>;
+using AtomicReferenceCountMap =
+    std::unordered_map<std::string, std::atomic<int>>;
+using DeviceReferenceCountMap =
+    std::unordered_map<int, std::unique_ptr<ReferenceCountMap>>;
+using AtomicDeviceReferenceCountMap =
+    std::unordered_map<int, std::unique_ptr<AtomicReferenceCountMap>>;
+using DeviceGarbageCollectorMap =
+    std::unordered_map<int,
+                       std::unique_ptr<GarbageCollector<framework::Tensor>>>;
+
+class ReferenceCountOpHandle : public OpHandleBase {
+ public:
+  ReferenceCountOpHandle(ir::Node *node, const Scope *scope,
+                         const platform::CUDAPlace &place,
+                         const std::vector<std::string> &var_names,
+                         GarbageCollector<Tensor> *gc,
+                         AtomicReferenceCountMap *ref_cnts)
+      : OpHandleBase(node),
+        scope_(scope),
+        var_names_(var_names),
+        gc_(gc),
+        ref_cnts_(ref_cnts) {
+    dev_ctx_ = static_cast<platform::CUDADeviceContext *>(
+        platform::DeviceContextPool::Instance().Get(place));
+    if (IsStreamGarabageCollector()) {
+      PADDLE_ENFORCE(cudaSetDevice(place.device));
+      PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
+    }
+  }
+
+  ~ReferenceCountOpHandle() {
+    if (IsStreamGarabageCollector()) {
+      auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
+      PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
+      PADDLE_ENFORCE(cudaEventDestroy(event_));
+    }
+  }
+
+  std::string Name() const override { return "reference_count"; }
+
+ protected:
+  void RunImpl() override {
+    auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
+    std::vector<LoDTensor *> tensors;
+    for (auto &name : var_names_) {
+      auto it = ref_cnts_->find(name);
+      if (it == ref_cnts_->end()) continue;
+
+      auto *var = exec_scope->FindVar(name);
+      if (var == nullptr || !var->IsType<LoDTensor>()) continue;
+
+      if (it->second.fetch_sub(1) <= 1) {
+        tensors.emplace_back(var->GetMutable<LoDTensor>());
+      }
+    }
+
+    if (!tensors.empty()) {
+      ClearTensors(tensors);
+    }
+  }
+
+ private:
+  void ClearTensors(const std::vector<LoDTensor *> &tensors) {
+    auto *gc = dynamic_cast<StreamGarbageCollector<Tensor> *>(gc_);
+    if (gc != nullptr) {
+      auto compute_stream = dev_ctx_->stream();
+      auto callback_stream = gc->stream();
+      auto callback_func = [=]() {
+        PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream));
+        PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0));
+      };
+      gc_->Add(tensors, callback_func);
+    } else {
+      gc_->Add(tensors);
+    }
+  }
+
+  bool IsStreamGarabageCollector() const {
+    return dynamic_cast<const StreamGarbageCollector<Tensor> *>(gc_) != nullptr;
+  }
+
+  const Scope *scope_;
+  platform::CUDADeviceContext *dev_ctx_;
+  std::vector<std::string> var_names_;
+  GarbageCollector<Tensor> *gc_;       // not own
+  AtomicReferenceCountMap *ref_cnts_;  // not own
+  cudaEvent_t event_;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/reference_count_pass.cc

@@ -0,0 +1,150 @@
+// 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 <string>
+#include <vector>
+
+#include "paddle/fluid/framework/details/computation_op_handle.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/details/reference_count_pass.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  auto &ref_cnts = Get<DeviceReferenceCountMap>(kGlobalReferenceCount);
+  auto &cur_ref_cnts = Get<AtomicDeviceReferenceCountMap>(kCurReferenceCount);
+  auto &gcs = Get<DeviceGarbageCollectorMap>(kGarbageCollector);
+
+  // It is not easy to find the right reference counts of varaibles in graph
+  // Step 1: Find all variables in computation ops
+  // Step 2: Find all variables in non-computation ops which refers to variables
+  // in computation ops
+  std::unordered_set<std::string> names;
+  auto get_ref_cnts_from_compute_op = [&](
+      const std::unique_ptr<OpHandleBase> &op,
+      const std::vector<VarHandleBase *> &vars) {
+    std::vector<std::string> var_names_in_op;
+    auto *compute_op = dynamic_cast<ComputationOpHandle *>(op.get());
+    if (compute_op == nullptr ||
+        !platform::is_gpu_place(compute_op->GetPlace()))
+      return var_names_in_op;
+    auto place = boost::get<platform::CUDAPlace>(compute_op->GetPlace());
+    for (VarHandleBase *var_handle_base : vars) {
+      auto *var_handle = dynamic_cast<VarHandle *>(var_handle_base);
+      if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue;
+
+      if (!platform::is_gpu_place(var_handle->place_) ||
+          boost::get<platform::CUDAPlace>(var_handle->place_) != place)
+        continue;
+
+      VarDesc *var_desc = var_handle->Node()->Var();
+      auto var_name = var_handle->Node()->Name();
+
+      // This is wierd but there is really some variables without var_desc
+      // in computation_op
+      if (var_desc == nullptr) {
+        if (compute_op->Node()->Op()->Block()->FindVar(var_name) == nullptr)
+          continue;
+      } else {
+        if (var_desc->Persistable() ||
+            var_desc->Proto()->type().type() != proto::VarType::LOD_TENSOR)
+          continue;
+      }
+
+      // compute op only runs in one device
+      if (ref_cnts[place.device]->count(var_name))
+        ++(*ref_cnts[place.device])[var_name];
+      else
+        (*ref_cnts[place.device])[var_name] = 1;
+
+      names.insert(var_name);
+      var_names_in_op.push_back(var_name);
+    }
+    return var_names_in_op;
+  };
+
+  auto update_ref_cnts_from_non_compute_op = [&](
+      const std::unique_ptr<OpHandleBase> &op,
+      const std::vector<VarHandleBase *> &vars) {
+    if (dynamic_cast<ComputationOpHandle *>(op.get()) != nullptr) return;
+    for (VarHandleBase *var_handle_base : vars) {
+      auto *var_handle = dynamic_cast<VarHandle *>(var_handle_base);
+      if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue;
+
+      auto var_name = var_handle->Node()->Name();
+      auto var_place = var_handle->place_;
+      if (!platform::is_gpu_place(var_place)) continue;
+      auto place = boost::get<platform::CUDAPlace>(var_place);
+      if (names.count(var_name) == 0) continue;
+      if (ref_cnts.count(place.device) &&
+          ref_cnts[place.device]->count(var_name)) {
+        ++(*ref_cnts[place.device])[var_name];
+      }
+    }
+  };
+
+  std::unordered_map<OpHandleBase *, ReferenceCountOpHandle *>
+      compute_ref_cnt_map;
+  auto &all_ops = graph->Get<GraphOps>(kGraphOps);
+  for (auto &op : all_ops) {
+    auto in_var_names = get_ref_cnts_from_compute_op(op, op->Inputs());
+    auto out_var_names = get_ref_cnts_from_compute_op(op, op->Outputs());
+    if (in_var_names.empty() && out_var_names.empty()) continue;
+    in_var_names.insert(in_var_names.end(), out_var_names.begin(),
+                        out_var_names.end());
+    auto *compute_op = dynamic_cast<ComputationOpHandle *>(op.get());
+    auto place = boost::get<platform::CUDAPlace>(compute_op->GetPlace());
+    ir::Node *ref_cnt_node =
+        graph->CreateEmptyNode("reference_count", ir::Node::Type::kOperation);
+    auto *ref_cnt_handle = new ReferenceCountOpHandle(
+        ref_cnt_node, compute_op->GetScope(), place, in_var_names,
+        gcs[place.device].get(), cur_ref_cnts[place.device].get());
+    auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
+    compute_op->AddOutput(dep_var);
+    ref_cnt_handle->AddInput(dep_var);
+    graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
+    compute_ref_cnt_map[compute_op] = ref_cnt_handle;
+  }
+
+  for (auto &op : all_ops) {
+    update_ref_cnts_from_non_compute_op(op, op->Inputs());
+    update_ref_cnts_from_non_compute_op(op, op->Outputs());
+  }
+
+  std::vector<std::unique_ptr<OpHandleBase>> new_all_ops;
+  new_all_ops.reserve(compute_ref_cnt_map.size() + all_ops.size());
+  for (auto &op : all_ops) {
+    new_all_ops.emplace_back(std::move(op));
+    auto it = compute_ref_cnt_map.find(new_all_ops.back().get());
+    if (it != compute_ref_cnt_map.end()) {
+      new_all_ops.emplace_back(it->second);
+    }
+  }
+
+  all_ops.swap(new_all_ops);
+  return graph;
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(reference_count_pass,
+              paddle::framework::details::ReferenceCountPass)
+    .RequirePassAttr(paddle::framework::details::kGlobalReferenceCount)
+    .RequirePassAttr(paddle::framework::details::kCurReferenceCount)
+    .RequirePassAttr(paddle::framework::details::kGarbageCollector);