Merge develop

test=develop

Author: sneaxiy

CMakeLists.txt

@@ -64,6 +64,7 @@ option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
 option(WITH_PSLIB       "Compile with pslib support"                    OFF)
 option(WITH_CONTRIB     "Compile the third-party contributation"        OFF)
 option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
+# TODO(Superjomn) Remove WITH_ANAKIN option if not needed latter.
 option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)
 option(ANAKIN_BUILD_FAT_BIN "Build anakin cuda fat-bin lib for all device plantform, ignored when WITH_ANAKIN=OFF" OFF)
 option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plantform. ignored when WITH_ANAKIN=OFF" ON)
@@ -190,6 +191,7 @@ include(configure)          # add paddle env configuration
 if(WITH_GPU)
     include(cuda)
     include(tensorrt)
+    include(anakin_subgraph)
 endif()
 if(WITH_MKL OR WITH_MKLML)
     include(external/anakin)

CONTRIBUTING.md

@@ -156,7 +156,7 @@ python \
 
 This will enable VLOG messages generated by `buddy_allocator.{h,cc}` and in the verbose range of 0 to 3, so you will see above example VLOG message, which is in level 3.  This suggests that we output overall messages in lower verbose levels, so they display with higher probability.  When coding C++, please follow the verbose level convention as follows:
 
-- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework)
-- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)
-- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform)
-- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/math)
+- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/framework)
+- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators)
+- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/platform)
+- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators/math/)

cmake/anakin_subgraph.cmake

@@ -0,0 +1,32 @@
+if(NOT WITH_GPU)
+    return()
+endif()
+
+set(ANAKIN_ROOT "/usr" CACHE PATH "ANAKIN ROOT")
+find_path(ANAKIN_INCLUDE_DIR anakin_config.h
+    PATHS ${ANAKIN_ROOT} ${ANAKIN_ROOT}/include
+    $ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/include
+    NO_DEFAULT_PATH
+)
+
+find_library(ANAKIN_LIBRARY NAMES libanakin_saber_common.so libanakin.so
+    PATHS ${ANAKIN_ROOT}
+    $ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/lib
+    NO_DEFAULT_PATH
+    DOC "Path to ANAKIN library.")
+
+if(ANAKIN_INCLUDE_DIR AND ANAKIN_LIBRARY)
+  if(WITH_DSO)
+    set(ANAKIN_FOUND ON)
+  endif(WITH_DSO)
+else()
+    set(ANAKIN_FOUND OFF)
+endif()
+
+if(ANAKIN_FOUND)
+    message(STATUS "Current ANAKIN header is ${ANAKIN_INCLUDE_DIR}/anakin_config.h. ")
+    include_directories(${ANAKIN_ROOT}/include)
+    include_directories(${ANAKIN_ROOT}/include/saber)
+    link_directories(${ANAKIN_ROOT})
+    add_definitions(-DPADDLE_WITH_ANAKIN)
+endif()

cmake/tensorrt.cmake

@@ -33,5 +33,6 @@ if(TENSORRT_FOUND)
     message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. "
         "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
     include_directories(${TENSORRT_INCLUDE_DIR})
+    link_directories(${TENSORRT_LIBRARY})
     add_definitions(-DPADDLE_WITH_TENSORRT)
 endif()

paddle/fluid/API.spec

@@ -520,6 +520,7 @@ paddle.fluid.unique_name.guard (ArgSpec(args=['new_generator'], varargs=None, ke
 paddle.fluid.recordio_writer.convert_reader_to_recordio_file (ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', '65c7523e86f0c50bb729b01667f36310'))
 paddle.fluid.recordio_writer.convert_reader_to_recordio_files (ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', 'bc643f0f5f1b9db57ff0d8a57d379bd7'))
 paddle.fluid.Scope Scope() -> paddle.fluid.core._Scope
+paddle.fluid.install_check.run_check (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', '66b7c84a17ed32fec2df9628367be2b9'))
 paddle.reader.cache (ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None), ('document', '1676886070eb607cb608f7ba47be0d3c'))
 paddle.reader.map_readers (ArgSpec(args=['func'], varargs='readers', keywords=None, defaults=None), ('document', '77cbadb09df588e21e5cc0819b69c87d'))
 paddle.reader.buffered (ArgSpec(args=['reader', 'size'], varargs=None, keywords=None, defaults=None), ('document', '0d6186f109feceb99f60ec50a0a624cb'))

paddle/fluid/framework/details/CMakeLists.txt

@@ -5,6 +5,7 @@ cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_h
 cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
 cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
 cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry)
+cc_library(fetch_barrier_op_handle SRCS fetch_barrier_op_handle.cc DEPS framework_proto scope place operator op_registry)
 
 cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper)
 cc_library(multi_devices_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper)
@@ -72,7 +73,7 @@ cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS grap
 cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
 
 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 fused_broadcast_op_handle)
+        scale_loss_grad_op_handle rpc_op_handle fetch_barrier_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
 
 cc_library(fuse_all_reduce_op_pass SRCS fuse_all_reduce_op_pass.cc DEPS graph graph_helper fused_all_reduce_op_handle)
 

paddle/fluid/framework/details/all_reduce_deps_pass.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include <algorithm>
+#include <memory>
 #include <string>
 #include <unordered_map>
 #include <unordered_set>
@@ -52,13 +53,28 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
   //               Note that must assert topology sort is stable
   auto& ops = graph->Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs);
   for (auto* op_desc : ops) {
-    auto outputs = op_desc->Outputs();
-    for (auto& o_it : outputs) {
-      for (auto& v : o_it.second) {  // values
-        vars[v] = order;
+    try {
+      bool is_bk_op =
+          static_cast<bool>(boost::get<int>(op_desc->GetAttr(
+                                OpProtoAndCheckerMaker::OpRoleAttrName())) &
+                            static_cast<int>(OpRole::kBackward));
+      if (!is_bk_op) continue;
+
+      auto backward_vars =
+          boost::get<std::vector<std::string>>(op_desc->GetNullableAttr(
+              OpProtoAndCheckerMaker::OpRoleVarAttrName()));
+      PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
+
+      auto outputs = op_desc->Outputs();
+      for (auto& o_it : outputs) {
+        for (auto& v : o_it.second) {  // values
+          vars[v] = order;
+          VLOG(1) << "in all_reduce_deps_pass:" << v;
+        }
       }
+      order++;
+    } catch (boost::bad_get e) {
     }
-    order++;
   }
 
   std::vector<OpHandleBase*> dist_ops;

paddle/fluid/framework/details/fetch_barrier_op_handle.cc

@@ -0,0 +1,66 @@
+//   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/fetch_barrier_op_handle.h"
+
+#include <string>
+
+namespace paddle {
+namespace framework {
+namespace details {
+FetchBarrierOpHandle::FetchBarrierOpHandle(
+    ir::Node *node, const std::vector<Scope *> &local_scopes,
+    const std::vector<platform::Place> &places)
+    // fetch_barrier op always run on place0, but output on all places.
+    : OpHandleBase(node),
+      op_(framework::OpRegistry::CreateOp(*node->Op())),
+      local_scopes_(local_scopes),
+      places_(places),
+      run_scope_(local_scopes[0]),
+      place_(places[0]) {
+  for (auto &p : places) {
+    this->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p));
+  }
+}
+
+bool FetchBarrierOpHandle::IsMultiDeviceTransfer() {
+  // override IsMultiDeviceTransfer to return true
+  return true;
+}
+
+void FetchBarrierOpHandle::RunImpl() {
+  WaitInputVarGenerated(place_);
+
+  auto run_func = [this]() {
+    op_->Run(*run_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(), place_);
+  };
+
+  if (is_lock_and_record_event_free_) {
+    run_func();
+  } else {
+    this->RunAndRecordEvent(run_func);
+  }
+}
+
+bool FetchBarrierOpHandle::NeedWait(VarHandleBase *in_var) {
+  bool need_wait =
+      in_var && in_var->GeneratedOp() &&
+      in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_.at(place_);
+  return need_wait;
+}
+
+std::string FetchBarrierOpHandle::Name() const { return op_->Type(); }
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/fetch_barrier_op_handle.h

@@ -0,0 +1,61 @@
+//   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/op_handle_base.h"
+#include "paddle/fluid/framework/feed_fetch_type.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+// **NOTE**: fetch_barrier op is special it outputs all recved variables on
+// all places if there are multiple places, must init with
+// multiple dev_ctxes_ !!!!
+
+struct FetchBarrierOpHandle : public OpHandleBase {
+ public:
+  FetchBarrierOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
+                       const std::vector<platform::Place> &places);
+
+  bool IsMultiDeviceTransfer() override;
+
+  std::string Name() const override;
+
+ protected:
+  void RunImpl() override;
+
+  bool NeedWait(VarHandleBase *in_var) override;
+
+ private:
+  std::unique_ptr<OperatorBase> op_;
+  std::vector<Scope *> local_scopes_;
+  std::vector<platform::Place> places_;
+  Scope *run_scope_;
+  platform::Place place_;
+
+  bool is_lock_and_record_event_free_{false};
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle