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

Author: reyoung

README.md

@@ -62,9 +62,9 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl
 ## Installation
 
 It is recommended to check out the
-[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html)
+[Docker installation guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/docker_install_en.html)
 before looking into the
-[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/build_from_source_en.html).
+[build from source guide](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html).
 
 ## Documentation
 

cmake/external/mkldnn.cmake

@@ -45,9 +45,9 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
 ELSE()
     MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
 ENDIF()
-
-SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
-SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} -Wno-error=strict-overflow")
+SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-unused-result")
+SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
+SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
 ExternalProject_Add(
     ${MKLDNN_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
@@ -61,6 +61,7 @@ ExternalProject_Add(
     CMAKE_ARGS          -DMKLROOT=${MKLML_ROOT}
     CMAKE_ARGS          -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
     CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
+    CMAKE_ARGS          -DWITH_TEST=OFF -DWITH_EXAMPLE=OFF
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
                         -DMKLROOT:PATH=${MKLML_ROOT}
 )

cmake/external/mklml.cmake

@@ -27,7 +27,7 @@ ENDIF()
 INCLUDE(ExternalProject)
 
 SET(MKLML_PROJECT       "extern_mklml")
-SET(MKLML_VER           "mklml_lnx_2018.0.1.20171007")
+SET(MKLML_VER           "mklml_lnx_2018.0.3.20180406")
 SET(MKLML_URL           "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")

doc/fluid/design/concepts/lod_tensor.md

@@ -155,7 +155,7 @@ into offsets
    3  2+3 4+5 1+9 2+10 3+12
 ```
 
-so we know that the first sentence is from word 0 to word 3, and the second sentence from work 3 to word 5.
+so we know that the first sentence is from word 0 to word 3, and the second sentence from word 3 to word 5.
 
 Similarly, the lengths in the top level LoD
 

paddle/fluid/framework/CMakeLists.txt

@@ -57,7 +57,7 @@ cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor
 cc_library(attribute SRCS attribute.cc DEPS framework_proto boost)
 cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc
 device_context)
-cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
+cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute glog)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 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)

paddle/fluid/framework/data_device_transform_test.cu

@@ -32,8 +32,7 @@ struct AddFunctor {
 
 class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
-  OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  void Make() {
     AddInput("input", "input1 of test op");
     AddOutput("output", "output of test op");
     AddAttr<bool>("use_gpu", "force to use gpu kernel").SetDefault(false);

paddle/fluid/framework/details/computation_op_handle.h

@@ -36,7 +36,7 @@ struct ComputationOpHandle : public OpHandleBase {
  protected:
   void RunImpl() override;
 
-  virtual bool NeedWait(VarHandleBase *in_var);
+  bool NeedWait(VarHandleBase *in_var) override;
 
  private:
   std::unique_ptr<OperatorBase> op_;

paddle/fluid/framework/details/fetch_op_handle.h

@@ -42,7 +42,7 @@ struct FetchOpHandle : public OpHandleBase {
  protected:
   void RunImpl() override;
 
-  virtual void WaitInputVarGenerated(const platform::Place &place);
+  void WaitInputVarGenerated(const platform::Place &place) override;
 
  private:
   FeedFetchList *data_;

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -37,20 +37,26 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
     const std::vector<Scope *> &local_scopes,
-    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale)
+    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale,
+    bool balance_parameter_opt_between_cards)
     : loss_var_name_(loss_var_name),
       places_(places),
       local_scopes_(local_scopes),
-      nccl_ctxs_(nccl_ctxs) {
+      nccl_ctxs_(nccl_ctxs),
+      balance_parameter_opt_between_cards_(
+          balance_parameter_opt_between_cards) {
 #else
 MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale)
+    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale,
+    bool balance_parameter_opt_between_cards)
     : loss_var_name_(loss_var_name),
       places_(places),
-      local_scopes_(local_scopes) {
+      local_scopes_(local_scopes),
+      balance_parameter_opt_between_cards_(
+          balance_parameter_opt_between_cards) {
 #endif
   for (auto &p : params) {
     grad_names_.insert(GradVarName(p));
@@ -124,6 +130,12 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   // Find "send" op first for split is in front of send.
   OpDesc *send_op = GetSendOpDesc(program);
 
+  size_t cur_device_id = 0;
+  std::vector<std::unordered_set<std::string>> var_name_on_devices;
+  std::vector<std::unordered_set<std::string>> bcast_var_name_set;
+  var_name_on_devices.resize(places_.size());
+  bcast_var_name_set.resize(places_.size());
+
   bool is_forwarding = true;
   for (auto *op : program.Block(0).AllOps()) {
     if (op->Type() == "send") {
@@ -139,24 +151,47 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
       }
       is_forwarding = false;
     } else {
-      CreateComputationalOps(&result, *op, places_.size());
+      int op_dev_id = GetOpDeviceID(var_name_on_devices, *op);
+      if (op_dev_id == -1) {  // var on all device
+        CreateComputationalOps(&result, *op, places_.size());
+      } else {
+        CreateComputationalOp(&result, *op, op_dev_id);
+        for (auto &var_name : op->OutputArgumentNames()) {
+          var_name_on_devices[op_dev_id].emplace(var_name);
+        }
+      }
       if (!is_forwarding && places_.size() > 1) {
         // Currently, we assume that once gradient is generated, it can be
         // broadcast, and each gradient is only broadcast once.
         for (auto &og : op->OutputArgumentNames()) {
           if (IsParameterGradientOnce(og, &og_has_been_broadcast)) {
-            if (IsSparseGradient(var_types, og)) {
-              CreateReduceOp(&result, og, 0);
-              CreateBroadcastOp(&result, og, 0);
+            if (balance_parameter_opt_between_cards_) {
+              CreateReduceOp(&result, og, cur_device_id);
+              var_name_on_devices[cur_device_id].emplace(og);
+              bcast_var_name_set[cur_device_id].emplace(
+                  og.substr(0, og.size() - strlen(kGradVarSuffix)));
+              cur_device_id = (cur_device_id + 1) % places_.size();
             } else {
-              InsertNCCLAllReduceOp(&result, og);
+              if (IsSparseGradient(var_types, og)) {
+                CreateReduceOp(&result, og, 0);
+                CreateBroadcastOp(&result, og, 0);
+              } else {
+                InsertNCCLAllReduceOp(&result, og);
+              }
             }
           }
         }
       }
     }
   }
 
+  // Insert BCast Ops
+  for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
+    auto &to_bcast_set = bcast_var_name_set[dev_id];
+    for (auto &bcast_name : to_bcast_set) {
+      CreateBroadcastOp(&result, bcast_name, dev_id);
+    }
+  }
   /*
     Dependency graph has been constructed. However, there are still data
     harzaeds need to be handled.
@@ -265,6 +300,26 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
   return is_pg_once;
 }
 
+int MultiDevSSAGraphBuilder::GetOpDeviceID(
+    const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
+    const OpDesc &op) const {
+  if (!balance_parameter_opt_between_cards_) {
+    return -1;
+  }
+
+  int var_dev_id = -1;
+  for (auto &var_name : op.InputArgumentNames()) {
+    if (var_dev_id != -1) break;
+    for (size_t i = 0; i < var_name_on_devices.size(); ++i) {
+      if (var_name_on_devices[i].count(var_name)) {
+        var_dev_id = static_cast<int>(i);
+        break;
+      }
+    }
+  }
+  return var_dev_id;
+}
+
 void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const {
   for (size_t i = 0; i < places_.size(); ++i) {
 // Insert ScaleCost OpHandle

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -36,13 +36,15 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
                           platform::NCCLContextMap *nccl_ctxs,
-                          bool use_default_grad_scale);
+                          bool use_default_grad_scale,
+                          bool balance_parameter_opt_between_cards);
 #else
   MultiDevSSAGraphBuilder(const std::vector<platform::Place> &places,
                           const std::string &loss_var_name,
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
-                          bool use_default_grad_scale);
+                          bool use_default_grad_scale,
+                          bool balance_parameter_opt_between_cards);
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
@@ -60,6 +62,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #ifdef PADDLE_WITH_CUDA
   platform::NCCLContextMap *nccl_ctxs_;
 #endif
+  bool balance_parameter_opt_between_cards_;
   bool use_default_grad_scale_;
 
   bool IsScaleLossOp(const OpDesc &op) const;
@@ -84,6 +87,10 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
       const std::string &og,
       std::unordered_set<std::string> *og_has_been_broadcast) const;
 
+  int GetOpDeviceID(
+      const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
+      const OpDesc &op) const;
+
   void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const;
 
   void CreateBroadcastOp(SSAGraph *result, const std::string &p_name,