Merge remote-tracking branch 'ups/develop' into refine/op/gru

.travis.yml

@@ -27,15 +27,6 @@ script:
     # 43min timeout
     paddle/scripts/paddle_docker_build.sh ${JOB}
     if [ $? -eq 0 ] || [ $? -eq 142 ]; then true; else exit 1; fi;
-  - |
-    if [[ "$JOB" != "doc" ]]; then exit 0; fi;
-    # For document only
-    if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
-    if [[ "$TRAVIS_BRANCH" != "develop"  && ! "$TRAVIS_BRANCH" =~ ^v|release/[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi;
-    export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh
-    export DOCS_DIR=`pwd`
-    cd ..
-    curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc/
 notifications:
   email:
     on_success: change

CMakeLists.txt

@@ -65,6 +65,7 @@ option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better d
 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_SYSTEM_BLAS   "Use system blas library"           OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
 
@@ -72,6 +73,7 @@ option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VER
 if(NOT PY_VERSION)
   set(PY_VERSION 2.7)
 endif()
+set(PYBIND11_PYTHON_VERSION ${PY_VERSION})
 
 # CMAKE_BUILD_TYPE
 if(NOT CMAKE_BUILD_TYPE)
@@ -158,6 +160,7 @@ endif()
 ########################################################################################
 
 include(external/mklml)     # download mklml package
+include(external/xbyak)     # download xbyak package
 include(external/libxsmm)   # download, build, install libxsmm
 include(external/zlib)      # download, build, install zlib
 include(external/gflags)    # download, build, install gflags
@@ -174,6 +177,7 @@ include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
 include(external/cares)
+include(external/cub)
 
 if(WITH_DISTRIBUTE)
     if(WITH_GRPC)

cmake/external/cub.cmake

+if(NOT WITH_GPU)
+  return()
+endif()
+
+include(ExternalProject)
+
+set(CUB_SOURCE_DIR ${THIRD_PARTY_PATH}/cub)
+set(CUB_INCLUDE_DIR ${CUB_SOURCE_DIR}/src/extern_cub)
+
+include_directories(${CUB_INCLUDE_DIR})
+
+ExternalProject_Add(
+  extern_cub
+  ${EXTERNAL_PROJECT_LOG_ARGS}
+  GIT_REPOSITORY "https://github.com/NVlabs/cub.git"
+  GIT_TAG        "v1.8.0"
+  PREFIX         ${CUB_SOURCE_DIR}
+  UPDATE_COMMAND ""
+  CONFIGURE_COMMAND ""
+  BUILD_COMMAND     ""
+  INSTALL_COMMAND   ""
+  TEST_COMMAND      ""
+)
+
+if(${CMAKE_VERSION} VERSION_LESS "3.3.0")
+  set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cub_dummy.c)
+  file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";")
+  add_library(cub STATIC ${dummyfile})
+else()
+  add_library(cub INTERFACE)
+endif()
+
+add_dependencies(cub extern_cub)
+
+LIST(APPEND externl_project_dependencies cub)

cmake/external/xbyak.cmake

+# Copyright (c) 2017 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.
+
+set(WITH_XBYAK ON)
+if(WIN32 OR APPLE)
+    SET(WITH_XBYAK OFF CACHE STRING "Disable XBYAK in Windows and MacOS" FORCE)
+    return()
+endif()
+
+include(ExternalProject)
+
+set(XBYAK_PROJECT       extern_xbyak)
+set(XBYAK_PREFIX_DIR    ${THIRD_PARTY_PATH}/xbyak)
+set(XBYAK_INSTALL_ROOT  ${THIRD_PARTY_PATH}/install/xbyak)
+set(XBYAK_INC_DIR       ${XBYAK_INSTALL_ROOT}/include)
+
+include_directories(${XBYAK_INC_DIR})
+include_directories(${XBYAK_INC_DIR}/xbyak)
+
+add_definitions(-DPADDLE_WITH_XBYAK)
+
+# xbyak options
+add_definitions(-DXBYAK64)
+add_definitions(-DXBYAK_NO_OP_NAMES)
+
+ExternalProject_Add(
+    ${XBYAK_PROJECT}
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    DEPENDS             ""
+    GIT_REPOSITORY      "https://github.com/herumi/xbyak.git"
+    GIT_TAG             "v5.661"  # Jul 26th
+    PREFIX              ${XBYAK_PREFIX_DIR}
+    UPDATE_COMMAND      ""
+    CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${XBYAK_INSTALL_ROOT}
+    CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${XBYAK_INSTALL_ROOT}
+)
+
+if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
+    set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/xbyak_dummy.c)
+    file(WRITE ${dummyfile} "const char *dummy_xbyak = \"${dummyfile}\";")
+    add_library(xbyak STATIC ${dummyfile})
+else()
+    add_library(xbyak INTERFACE)
+endif()
+
+add_dependencies(xbyak ${XBYAK_PROJECT})
+list(APPEND external_project_dependencies xbyak)

cmake/generic.cmake

@@ -264,7 +264,10 @@ function(cc_test TARGET_NAME)
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
     if (${cc_test_SERIAL})
         set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
+
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
     endif()
   endif()
 endfunction(cc_test)
@@ -329,7 +332,10 @@ function(nv_test TARGET_NAME)
     add_test(${TARGET_NAME} ${TARGET_NAME})
     if (nv_test_SERIAL)
         set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
+
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
     endif()
   endif()
 endfunction(nv_test)
@@ -577,7 +583,9 @@ function(py_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS ARGS ENVS)
     cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_test(NAME ${TARGET_NAME}
-             COMMAND env FLAGS_init_allocated_mem=true PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
+             COMMAND env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
+             FLAGS_cpu_deterministic=true
+             PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
              ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
   endif()

doc/survey/op_fusion_design.md

+# Operator fusion  
+Fusing multiple operators together is an important method to optimize the program execution, particularly for GPU or other specialized accelerators. An obvious benefit is to avoid the overhead of saving the intermediate result back into global memory.   
+
+There are generally two ways to fuse operators, fusing directly connected operators and fusing non directly connected operators. The first method is mainly used by [NNVM Compiler](https://github.com/dmlc/tvm/) and [XLA](https://www.tensorflow.org/performance/xla/). The second method is mainly used by Dynet and TensorFlow Fold to do auto-batching. The principle of fusing operator is according to some rules to combine multiple operations into one, for example, `Y = X * W` and `Z = Y + B` can be fused to `Z = X * W + B`, and `Y1 = X1 * W` and `Y2 = X2 * W` can be fused to `[Y1;Y2] = [X1;X2] * W`. In order to get a short-term profit, we decided to try to manually specify these rules.   
+
+## Challenge
+The challenge of fusing operators is:
+  - how to make the rules.
+  - how to implement these rules efficiently.
+
+### How to make the rules?
+
+The problem of determining the best single location for a fusion operator is an NP-hard combinatorial problem. After analysis the operators of the DL model, we found there are two group of operators can be fused explicitly, one is the simple and adjacent operations, for example, `tmp = x + y` and `z = Relu(tmp)`, and the other is the operators that have the same function, for example, a serials of `SGD` or `Momentum`. They usually appear in the model in a large number. So we should think about how to fuse them separately first.
+
+### How to implement these rules efficiently?
+#### How to fuse the adjacent operations efficiently?
+Here we use a template function to represent the fused operations. The pros of using a template function are that it is simple and efficient, and the cons are that it is not easy to expand, and it can only be used to express some simple operations. So taking into account our current needs, the template function is more appropriate.
+
+#### How to fuse the operators that have the same function efficiently?
+We take SGD operator as an example, the training model may have hundreds of parameters and correspondingly have the same number of SGD operators. The expression(`w = w - lr*w_g`) of those operators is the same, so during of training, the executor will execute this expression hundreds time in CPU or other specialized accelerators. If we can fuse them and make the address of all `w` and all `w_g` continuous respectively, we only need execute one time. For some accelerators, the time of launching kernel is not neglected, so the time of hundreds of times of launching and executing kernel may be larger than launching and executing only once. There usually are many operators that similar to `SGD` in the DL model, such as `AllReduce` and `FC`.

paddle/fluid/API.spec

@@ -336,6 +336,7 @@ paddle.fluid.contrib.BeamSearchDecoder.decode ArgSpec(args=['self'], varargs=Non
 paddle.fluid.contrib.BeamSearchDecoder.early_stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.contrib.BeamSearchDecoder.read_array ArgSpec(args=['self', 'init', 'is_ids', 'is_scores'], varargs=None, keywords=None, defaults=(False, False))
 paddle.fluid.contrib.BeamSearchDecoder.update_array ArgSpec(args=['self', 'array', 'value'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.contrib.memory_usage ArgSpec(args=['program', 'batch_size'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.transpiler.DistributeTranspiler.create_splited_vars ArgSpec(args=['self', 'source_var', 'block', 'tag'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)

paddle/fluid/CMakeLists.txt

@@ -5,5 +5,7 @@ add_subdirectory(operators)
 add_subdirectory(pybind)
 add_subdirectory(string)
 add_subdirectory(recordio)
-# NOTE: please add subdirectory inference at last.
-add_subdirectory(inference)
+if(WITH_INFERENCE)
+  # NOTE: please add subdirectory inference at last.
+  add_subdirectory(inference)
+endif()

paddle/fluid/framework/details/build_strategy.h

@@ -21,6 +21,26 @@ namespace framework {
 namespace details {
 
 struct BuildStrategy {
+  // ParallelExecutor supports two modes of ReduceStrategy, kAllReduce and
+  // kReduce, for CPU and GPU. If you use kAllReduce, different threads
+  // optimize their parameters separately. If you use kReduce, the optimizations
+  // of parameters are distributed to different threads.
+  // For example, a model has 100 parameters and is running with four threads,
+  // if you choose kAllReduce, every thread is to optimize 100 parameters
+  // separately, if you choose kReduce, every thread is to optimize 25
+  // parameters.
+  // Of particular note is, if you use kReduce when using CPU training,
+  // all the parameters are shared between different threads. This feature will
+  // save memory.
+  // FIXME(zcd): The result of the two modes(kAllReduce and kReduce) maybe not
+  // equal for GPU. Because, the result of the different order of summing maybe
+  // different, for example, the result of `a+b+c+d` may be different with the
+  // result of `c+a+b+d`.
+  // For GPU, the implementation of kAllReduce and kReduce is adopted NCCL,
+  // so the result of kAllReduce and kReduce maybe not equal.
+  // For CPU, if you want to fix the order of summing to make the result
+  // of kAllReduce and kReduce no diff, you can add
+  // `FLAGS_cpu_deterministic=true` to env.
   enum class ReduceStrategy { kAllReduce = 0, kReduce = 1 };
 
   enum class GradientScaleStrategy {

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -275,7 +275,8 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
       if (strategy_.gradient_scale_ !=
           BuildStrategy::GradientScaleStrategy::kCustomized) {
         // TODO(paddle-dev): Why is there no input for this op_handle?
-        CreateScaleLossGradOp(&result);
+        auto loss_grad_name = node->Op()->OutputArgumentNames()[0];
+        CreateScaleLossGradOp(&result, loss_grad_name);
       }
       // This assumes the backward generating code will ensure IsScaleLossOp
       // is true only for the op that scale the final scalar loss.
@@ -535,7 +536,8 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
   return got == sharded_var_device.end() ? -1 : got->second;
 }
 
-void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(ir::Graph *result) const {
+void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
+    ir::Graph *result, const std::string &loss_grad_name) const {
   for (size_t i = 0; i < places_.size(); ++i) {
 // Insert ScaleCost OpHandle
 #ifdef PADDLE_WITH_CUDA
@@ -558,9 +560,9 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(ir::Graph *result) const {
     // loss->pending_ops_.emplace_back(op_handle);
     // op_handle->inputs_.emplace_back(loss);
 
-    CreateOpOutput(result, op_handle,
-                   result->CreateEmptyNode(GradVarName(loss_var_name_),
-                                           ir::Node::Type::kVariable),
+    CreateOpOutput(
+        result, op_handle,
+        result->CreateEmptyNode(loss_grad_name, ir::Node::Type::kVariable),
         places_[i], i);
   }
 }

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -75,7 +75,9 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
   void CreateComputationalOps(ir::Graph *result, ir::Node *node,
                               size_t num_places) const;
 
-  void CreateScaleLossGradOp(ir::Graph *result) const;
+  void CreateScaleLossGradOp(ir::Graph *result,
+                             const std::string &loss_grad_name) const;
+
   VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og,
                             int dst_dev_id) const;
   void CreateComputationalOp(ir::Graph *result, ir::Node *node,

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -18,6 +18,10 @@
 #include "paddle/fluid/framework/details/variable_visitor.h"
 #include "paddle/fluid/platform/profiler.h"
 
+DEFINE_bool(
+    cpu_deterministic, false,
+    "Whether to make the result of computation deterministic in CPU side.");
+
 namespace paddle {
 namespace framework {
 namespace details {
@@ -91,11 +95,33 @@ void ReduceOpHandle::RunImpl() {
   } else {
     std::vector<const LoDTensor *> lod_tensors =
         GetInputValues<LoDTensor>(in_var_handles, var_scopes);
+
     if (paddle::platform::is_cpu_place(lod_tensors[0]->place())) {
       this->RunAndRecordEvent([&] {
+        // FIXME(zcd): The order of summing is important,
+        // especially when the type of data is float or double.
+        // For example, the result of `a+b+c+d` may be different
+        // with the result of `c+a+b+d`, so the summing order should be fixed.
+        if (!FLAGS_cpu_deterministic) {
           ReduceLoDTensor func(lod_tensors,
                                out_var->GetMutable<framework::LoDTensor>());
           VisitDataType(ToDataType(lod_tensors[0]->type()), func);
+        } else {
+          // We sum lod_tensors to reduce_sum_trg which is in local_scopes_0
+          // here, but it doesn't mean reduce_sum_trg must be in local_scopes_0.
+          auto &reduce_sum_trg = *this->local_scopes_[0]
+                                      ->FindVar(kLocalExecScopeName)
+                                      ->Get<Scope *>()
+                                      ->FindVar(out_var_handle->name_)
+                                      ->GetMutable<framework::LoDTensor>();
+          ReduceLoDTensor func(lod_tensors, &reduce_sum_trg);
+          VisitDataType(ToDataType(lod_tensors[0]->type()), func);
+
+          auto trg = out_var->GetMutable<framework::LoDTensor>();
+          if (reduce_sum_trg.data<void>() != trg->data<void>()) {
+            TensorCopy(reduce_sum_trg, platform::CPUPlace(), trg);
+          }
+        }
       });
     } else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) {
 #ifdef PADDLE_WITH_CUDA

paddle/fluid/framework/executor.cc

@@ -330,12 +330,7 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
   }
 
   for (auto& op : ctx->ops_) {
-    VLOG(4) << place_ << " " << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);
-    // NOTE! Please do not delete this line, it's usefull because the debug
-    // string before and after op.run are different, after run the output
-    // will have right shape which is usefull for debug.
-    VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
 
     if (FLAGS_benchmark) {
       VLOG(2) << "Memory used after operator " + op->Type() + " running: "

paddle/fluid/framework/ir/graph.cc

@@ -182,9 +182,11 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
   }
 
   /**
-   * We only handle write after read(WAR), since it should not have a write
-   * after write in program. If there are write after write operators, we need
-   * prune them.
+   * We should handle write after read(WAR) and write after write(WAW) here.
+   * Because some of the operators of the program can be executed parallelly.
+   * So, to make the program running in the right order, we should add the
+   * dependence of WAR and WAW.
+   *
    *
    * https://en.wikipedia.org/wiki/Hazard_(computer_architecture)#Write_after_read_(WAR)
    */
@@ -201,6 +203,19 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
           (*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0];
       const auto &read_ops = (*it_old)->outputs;
 
+      PADDLE_ENFORCE(write_op, "The write_op should not be empty.");
+
+      // Add write after write dependence
+      ir::Node *upstream_op =
+          (*it_old)->inputs.empty() ? nullptr : (*it_old)->inputs[0];
+      if (upstream_op) {
+        ir::Node *dep_var = CreateControlDepVar();
+        write_op->inputs.push_back(dep_var);
+        upstream_op->outputs.push_back(dep_var);
+        dep_var->outputs.push_back(write_op);
+        dep_var->inputs.push_back(upstream_op);
+      }
+
       for (auto *read_op : read_ops) {
         // Manually add a dependency var from read_op to write_op;
         if (read_op == write_op) {

paddle/fluid/framework/operator.cc

@@ -127,7 +127,7 @@ static LoD GetLoD(const Scope& scope, const std::string& name) {
 }
 
 void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
-  VLOG(10) << "- " << DebugStringEx(&scope);
+  VLOG(4) << place << " " << DebugStringEx(&scope);
   if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
     PADDLE_THROW("Cannot run operator on place %s", place);
@@ -136,8 +136,10 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
     platform::SetDeviceId(dev_id);
 #endif
   }
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  platform::RecordEvent record_event(Type(), pool.Get(place));
   RunImpl(scope, place);
-  VLOG(10) << "+ " << DebugStringEx(&scope);
+  VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 
 bool OperatorBase::HasInputs(const std::string& name) const {
@@ -639,9 +641,6 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto* dev_ctx = pool.Get(place);
 
-  // For profiling, don't move out of this function because that will result
-  // in the failure of multi-GPU profiling.
-  platform::RecordEvent record_event(Type(), dev_ctx);
   // check if op[type] has kernel registered.
   auto& all_op_kernels = AllOpKernels();
   auto kernels_iter = all_op_kernels.find(type_);
@@ -779,6 +778,7 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
     const ExecutionContext& ctx) const {
   auto& scope = ctx.scope();
   int data_type = -1;
+  std::string last_input_name;
   for (auto& input : this->inputs_) {
     for (auto& ipt_name : input.second) {
       auto* var = scope.FindVar(ipt_name);
@@ -795,9 +795,10 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
           int tmp = static_cast<int>(ToDataType(t->type()));
           PADDLE_ENFORCE(
               tmp == data_type || data_type == -1,
-              "DataType of Paddle Op %s must be the same. Get %d != %d", Type(),
-              data_type, tmp);
+              "DataType of Paddle Op %s must be the same. Get %s(%d) != %s(%d)",
+              Type(), last_input_name, data_type, ipt_name, tmp);
           data_type = tmp;
+          last_input_name = ipt_name;
         }
       }
     }

paddle/fluid/inference/analysis/analyzer.cc

@@ -24,7 +24,7 @@
 
 namespace paddle {
 
-DEFINE_bool(inference_analysis_enable_tensorrt_subgraph_engine, false,
+DEFINE_bool(inference_analysis_enable_tensorrt_subgraph_engine, true,
             "Enable subgraph to TensorRT engine for acceleration");
 
 DEFINE_string(inference_analysis_graphviz_log_root, "./",
@@ -42,10 +42,19 @@ class DfgPassManagerImpl final : public DfgPassManager {
     // TODO(Superjomn) set the key with pass reprs.
     AddPass("fluid-to-data-flow-graph", new FluidToDataFlowGraphPass);
     if (FLAGS_inference_analysis_enable_tensorrt_subgraph_engine) {
-      auto trt_teller = [](const Node* node) {
+      auto trt_teller = [&](const Node* node) {
+        std::unordered_set<std::string> teller_set(
+            {"elementwise_add", "mul", "conv2d", "pool2d", "relu"});
         if (!node->IsFunction()) return false;
-        return static_cast<const Function*>(node)->func_type() == "mul";
+
+        const auto* func = static_cast<const Function*>(node);
+        if (teller_set.count(func->func_type()))
+          return true;
+        else {
+          return false;
+        }
       };
+
       AddPass("tensorrt-subgraph-marker",
               new TensorRTSubgraphNodeMarkPass(trt_teller));
       AddPass("tensorrt-subgraph", new TensorRTSubGraphPass(trt_teller));

paddle/fluid/inference/analysis/data_flow_graph.cc

@@ -337,6 +337,34 @@ ExtractInputAndOutputOfSubGraph(std::vector<Node *> &graph) {  // NOLINT
                         std::vector<Node *>(outputs.begin(), outputs.end()));
 }
 
+void FilterRedundantOutputOfSubGraph(DataFlowGraph *graph) {
+  std::vector<Node *> op_nodes;
+  for (auto &node : GraphTraits<DataFlowGraph>(graph).nodes_in_TS()) {
+    if (node.type() == Node::Type::kValue || node.deleted()) {
+      continue;
+    }
+    op_nodes.push_back(&node);
+  }
+  size_t op_num = op_nodes.size();
+  for (size_t i = 0; i < op_num; i++) {
+    if (op_nodes[i]->type() == Node::Type::kFunction) continue;
+    std::unordered_set<std::string> follow_up_input_names;
+    for (size_t j = i + 1; j < op_num; j++) {
+      for (auto *in : op_nodes[j]->inlinks) {
+        follow_up_input_names.insert(in->name());
+      }
+    }
+    std::vector<Node *> filtered_subgraph_outlinks;
+    for (auto *out : op_nodes[i]->outlinks) {
+      if (follow_up_input_names.count(out->name())) {
+        filtered_subgraph_outlinks.push_back(out);
+      }
+    }
+    PADDLE_ENFORCE_GE(filtered_subgraph_outlinks.size(), 1UL);
+    op_nodes[i]->outlinks = filtered_subgraph_outlinks;
+  }
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/data_flow_graph.h

@@ -178,6 +178,7 @@ struct GraphTraits<DataFlowGraph> {
 std::pair<std::vector<Node *>, std::vector<Node *>>
 ExtractInputAndOutputOfSubGraph(std::vector<Node *> &graph);  // NOLINT
 
+void FilterRedundantOutputOfSubGraph(DataFlowGraph *graph);
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc

@@ -23,7 +23,7 @@
 namespace paddle {
 namespace inference {
 
-DEFINE_int32(tensorrt_max_batchsize, 300, "TensorRT maximum batch size");
+DEFINE_int32(tensorrt_max_batchsize, 3, "TensorRT maximum batch size");
 DEFINE_int32(tensorrt_workspace_size, 2048, "TensorRT workspace size");
 
 namespace analysis {
@@ -52,6 +52,7 @@ bool DataFlowGraphToFluidPass::Initialize(Argument *argument) {
 bool DataFlowGraphToFluidPass::Finalize() { return true; }
 
 void DataFlowGraphToFluidPass::Run(DataFlowGraph *graph) {
+  FilterRedundantOutputOfSubGraph(graph);
   LOG(INFO) << "graph.inputs " << graph->inputs.size();
   for (auto &node : GraphTraits<DataFlowGraph>(graph).nodes_in_TS()) {
     if (node.deleted()) continue;
@@ -87,34 +88,113 @@ void DataFlowGraphToFluidPass::AddFluidOp(Node *node) {
 }
 
 void CreateTrtEngineOp(Node *node, const DataFlowGraph &graph,
-                       const framework::proto::BlockDesc &block) {
+                       framework::proto::BlockDesc *block) {
   static int counter{0};
   PADDLE_ENFORCE(node->IsFunctionBlock());
   framework::OpDesc desc;
   auto *func = static_cast<FunctionBlock *>(node);
 
   // collect inputs
-  std::vector<std::string> io;
+  std::unordered_set<std::string> input_names;
   for (auto *x : func->inlinks) {
-    io.push_back(x->name());
+    input_names.insert(x->name());
   }
-  desc.SetInput("Xs", io);
+  desc.SetInput(
+      "Xs", std::vector<std::string>(input_names.begin(), input_names.end()));
 
-  // collect outputs
-  io.clear();
+  std::unordered_set<std::string> output_names;
   for (auto *x : func->outlinks) {
-    io.push_back(x->name());
+    output_names.insert(x->name());
   }
-  desc.SetOutput("Ys", io);
+
+  std::vector<std::string> output_temp(output_names.begin(),
+                                       output_names.end());
+  desc.SetOutput("Ys", output_temp);
   desc.SetType("tensorrt_engine");
 
-  PADDLE_ENFORCE(!block.vars().empty(), "the block has no var-desc");
+  std::unordered_map<std::string, std::string> output_name_map;
+
+  // The following procedure is used to rename all the intermediate
+  // variables and the output variables of the subgraph.
+  // Why we do this?
+  // During the transition from fluid OP to tensorrt OP, we map
+  // the input and output Tensor(fluid data structure) of fluid OP
+  // to the correspondin ITensor (trt data structure) through the
+  // Tensor name. When we set up ITensor for an variable, we must
+  // ensure that it has not been set before.
+  // If there is variable in the fluid graph, which is not only the
+  // input of a OP, but also the output of a Op, there will be problems.
+  // So we have to rename the variable in the subgraph to make sure
+  // it is either an OP's input or an OP's output.
+
+  auto subgraph_nodes = func->subgraph;
+  for (int index = 0; index < block->ops_size(); index++) {
+    framework::proto::OpDesc *op = block->mutable_ops(index);
+    auto correspond_node = subgraph_nodes[index];
+    PADDLE_ENFORCE_EQ(correspond_node->name(), op->type());
+
+    std::unordered_map<std::string, size_t> var2id;
+    for (auto *in_var : correspond_node->inlinks) {
+      var2id[in_var->name()] = in_var->id();
+    }
+    // rename for the input variables of op inside subgraph
+    for (int i = 0; i < op->inputs_size(); i++) {
+      framework::proto::OpDesc_Var *in_var = op->mutable_inputs(i);
+      std::vector<std::string> replaced_names;
+      for (int k = 0; k < in_var->arguments_size(); k++) {
+        std::string arg_value = in_var->arguments(k);
+        if (input_names.count(arg_value)) {
+          replaced_names.push_back(arg_value);
+        } else {
+          replaced_names.push_back(arg_value +
+                                   std::to_string(var2id[arg_value]));
+        }
+      }
+      in_var->clear_arguments();
+      for (size_t k = 0; k < replaced_names.size(); k++) {
+        in_var->add_arguments(replaced_names[k]);
+      }
+    }
+    var2id.clear();
+    for (auto out_var : correspond_node->outlinks) {
+      var2id[out_var->name()] = out_var->id();
+    }
+
+    // rename for the output variables of op inside subgraph
+    for (int i = 0; i < op->outputs_size(); i++) {
+      framework::proto::OpDesc_Var *out_var = op->mutable_outputs(i);
+      std::vector<std::string> replaced_names;
+      for (int k = 0; k < out_var->arguments_size(); k++) {
+        std::string arg_value = out_var->arguments(k);
+        if (output_names.count(arg_value)) {
+          output_name_map[arg_value] =
+              arg_value + std::to_string(var2id[arg_value]);
+        }
+        replaced_names.push_back(arg_value + std::to_string(var2id[arg_value]));
+      }
+      out_var->clear_arguments();
+      for (size_t k = 0; k < replaced_names.size(); k++) {
+        out_var->add_arguments(replaced_names[k]);
+      }
+    }
+  }
+  // When tensorrt engine runs at the end of the operation,
+  // output_mapping help us copy the data from the renamed ITensor
+  // to Tensor.
+  std::vector<std::string> output_mapping;
+  for (auto name : output_names) {
+    PADDLE_ENFORCE(output_name_map.count(name) != 0);
+    output_mapping.push_back(output_name_map[name]);
+  }
+
+  PADDLE_ENFORCE(!block->vars().empty(), "the block has no var-desc");
   // Set attrs
-  SetAttr(desc.Proto(), "subgraph", block.SerializeAsString());
+  SetAttr(desc.Proto(), "subgraph", block->SerializeAsString());
   SetAttr(desc.Proto(), "engine_uniq_key", "trt-" + std::to_string(counter++));
   SetAttr(desc.Proto(), "max_batch", FLAGS_tensorrt_max_batchsize);
   SetAttr(desc.Proto(), "max_workspace", FLAGS_tensorrt_workspace_size);
   SetAttr(desc.Proto(), "parameters", ExtractParameters(graph.nodes.nodes()));
+  SetAttr(desc.Proto(), "output_name_mapping", output_mapping);
   node->SetPbMsg(desc.Proto()->SerializeAsString());
 }
 
@@ -146,15 +226,17 @@ void DataFlowGraphToFluidPass::AddEngineOp(Node *node) {
   LOG(INFO) << "transformed variable size: "
             << block_desc.Proto()->vars().size();
   // copy ops.
+
   for (auto *node : block_node->subgraph) {
     auto *op = block_desc.AppendOp();
     PADDLE_ENFORCE(!node->pb_msg().empty());
     op->Proto()->ParseFromString(node->pb_msg());
   }
+
   *block_desc.Proto()->mutable_vars() =
       argument_->origin_program_desc->blocks(0).vars();
   PADDLE_ENFORCE(!block_desc.Proto()->vars().empty());
-  CreateTrtEngineOp(node, *argument_->main_dfg, *block_desc.Proto());
+  CreateTrtEngineOp(node, *argument_->main_dfg, block_desc.Proto());
   auto *main_block = desc_->mutable_blocks(framework::kRootBlockIndex);
   auto *op = main_block->add_ops();
   PADDLE_ENFORCE(!node->pb_msg().empty(), "failed to set desc for block");

paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc

@@ -46,9 +46,9 @@ std::string DFG_GraphvizDrawPass::Draw(DataFlowGraph *graph) {
   for (size_t i = 0; i < graph->nodes.size(); i++) {
     const Node &node = graph->nodes.Get(i);
     if (!config_.display_deleted_node && node.deleted()) continue;
-    for (auto &in : node.inlinks) {
-      if (!config_.display_deleted_node && in->deleted()) continue;
-      dot.AddEdge(in->repr(), node.repr(), {});
+    for (auto &out : node.outlinks) {
+      if (!config_.display_deleted_node && out->deleted()) continue;
+      dot.AddEdge(node.repr(), out->repr(), {});
     }
   }
   return dot.Build();

paddle/fluid/inference/analysis/subgraph_splitter.cc

@@ -76,7 +76,7 @@ void UnionFindCombine(const node_map_t &node_map, size_t a, size_t b) {
 
 std::vector<std::vector<Node *>> SubGraphSplitter::ExtractSubGraphs() {
   std::vector<Node *> marked_nodes;
-  for (auto &node : GraphTraits<DataFlowGraph>(graph_).nodes()) {
+  for (auto &node : GraphTraits<DataFlowGraph>(graph_).nodes_in_TS()) {
     if (node.attr(kMarkerAttrName).Bool()) {
       marked_nodes.push_back(&node);
     }

paddle/fluid/inference/api/CMakeLists.txt

@@ -74,9 +74,10 @@ if (WITH_ANAKIN) # only needed in CI
     target_link_libraries(inference_anakin_api anakin anakin_saber_common)
     target_link_libraries(inference_anakin_api_shared anakin anakin_saber_common)
     if (WITH_TESTING)
-        cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc
-                                  ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin
-                                  DEPS inference_anakin_api_shared)
-        target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
+        # this test is unstable, disable it first.
+        #cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc
+                                  #ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin
+                                  #DEPS inference_anakin_api_shared)
+        #target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
      endif(WITH_TESTING)
 endif()

paddle/fluid/inference/api/api.cc

@@ -12,6 +12,7 @@ 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 <glog/logging.h>
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 
 namespace paddle {
@@ -40,19 +41,36 @@ PaddleBuf::PaddleBuf(PaddleBuf&& other)
 PaddleBuf::PaddleBuf(const PaddleBuf& other) { *this = other; }
 
 PaddleBuf& PaddleBuf::operator=(const PaddleBuf& other) {
+  if (!other.memory_owned_) {
+    data_ = other.data_;
+    length_ = other.length_;
+    memory_owned_ = other.memory_owned_;
+  } else {
+    Resize(other.length());
+    memcpy(data_, other.data(), other.length());
+    length_ = other.length();
+    memory_owned_ = true;
+  }
+  return *this;
+}
+
+PaddleBuf& PaddleBuf::operator=(PaddleBuf&& other) {
   // only the buffer with external memory can be copied
-  assert(!other.memory_owned_);
   data_ = other.data_;
   length_ = other.length_;
   memory_owned_ = other.memory_owned_;
+  other.data_ = nullptr;
+  other.length_ = 0;
+  other.memory_owned_ = false;
   return *this;
 }
 
 void PaddleBuf::Resize(size_t length) {
   // Only the owned memory can be reset, the external memory can't be changed.
   if (length_ == length) return;
-  assert(memory_owned_);
+  if (memory_owned_) {
     Free();
+  }
   data_ = new char[length];
   length_ = length;
   memory_owned_ = true;
@@ -68,7 +86,7 @@ void PaddleBuf::Reset(void* data, size_t length) {
 void PaddleBuf::Free() {
   if (memory_owned_ && data_) {
     assert(length_ > 0);
-    delete static_cast<char*>(data_);
+    delete[] static_cast<char*>(data_);
     data_ = nullptr;
     length_ = 0;
   }

paddle/fluid/inference/api/paddle_inference_api.h

@@ -40,11 +40,12 @@ class PaddleBuf {
   // Copy only available when memory is managed externally.
   explicit PaddleBuf(const PaddleBuf&);
   PaddleBuf& operator=(const PaddleBuf&);
+  PaddleBuf& operator=(PaddleBuf&&);
   // Do not own the memory.
   PaddleBuf(void* data, size_t length)
       : data_(data), length_(length), memory_owned_{false} {}
   // Own memory.
-  explicit PaddleBuf(size_t length)
+  PaddleBuf(size_t length)
       : data_(new char[length]), length_(length), memory_owned_(true) {}
   // Resize to `length` bytes.
   void Resize(size_t length);

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

 # Add TRT tests
 nv_library(tensorrt_converter
   SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc
+activation_op.cc
   DEPS tensorrt_engine operator scope framework_proto op_registry)
 
 nv_test(test_op_converter SRCS test_op_converter.cc DEPS

paddle/fluid/inference/tensorrt/convert/op_converter.h

@@ -55,7 +55,6 @@ class OpConverter {
         it = Registry<OpConverter>::Lookup("fc");
       }
     }
-
     if (op_desc.Type().find("elementwise") != std::string::npos) {
       static std::unordered_set<std::string> add_tensor_op_set{
           "add", "mul", "sub", "div", "max", "min", "pow"};
@@ -72,6 +71,8 @@ class OpConverter {
                        "Unsupported elementwise type" + op_type);
         it =
             Registry<OpConverter>::Lookup("elementwise_" + op_type + "_weight");
+        PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]",
+                                op_desc.Type());
       } else {
         PADDLE_ENFORCE(add_tensor_op_set.count(op_type) > 0,
                        "Unsupported elementwise type" + op_type);

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -280,12 +280,16 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
      * ('any') which lets a primitive (convolution in this case) choose
      * the memory format preferred for best performance
      */
+    std::string data_format = ctx.Attr<std::string>("data_format");
+    auto chosen_memory_format =
+        platform::data_format_to_memory_format(data_format);
+
     auto src_md = platform::MKLDNNMemDesc(
-        src_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
+        src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto weights_md = platform::MKLDNNMemDesc(
-        weights_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
+        weights_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto dst_md = platform::MKLDNNMemDesc(
-        dst_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
+        dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
 
     // create a conv primitive descriptor and save it for usage in backward
     std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd =
@@ -423,16 +427,20 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
      * ('any') which lets a primitive (conv backward in this case) choose
      * the memory format preferred for best performance
      */
+    std::string data_format = ctx.Attr<std::string>("data_format");
+    auto chosen_memory_format =
+        platform::data_format_to_memory_format(data_format);
+
     auto src_md = platform::MKLDNNMemDesc(
-        src_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
+        src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto diff_src_md = platform::MKLDNNMemDesc(
-        src_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
+        src_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto weights_md = platform::MKLDNNMemDesc(
-        weights_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
+        weights_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto diff_weights_md = platform::MKLDNNMemDesc(
-        weights_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
+        weights_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
     auto diff_dst_md = platform::MKLDNNMemDesc(
-        dst_tz, platform::MKLDNNGetDataType<T>(), memory::format::any);
+        dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
 
     // Retrieve conv_pd from device context
     auto conv_pd =

paddle/fluid/operators/elementwise_op_function.h

@@ -534,8 +534,8 @@ void ElemwiseGradCompute(const framework::ExecutionContext& ctx,
                          const framework::Tensor& dout, int axis,
                          framework::Tensor* dx, framework::Tensor* dy,
                          DX_OP dx_op, DY_OP dy_op) {
-  const framework::DDim x_dim = x.dims();
-  const framework::DDim y_dim = y.dims();
+  const framework::DDim& x_dim = x.dims();
+  const framework::DDim& y_dim = y.dims();
   if (x.dims() == y.dims()) {
     ElemwiseGradComputeNoBroadcast<DeviceContext, T, DX_OP, DY_OP>(
         ctx, x_dim, y_dim, x, y, out, dout, axis, dx, dy, dx_op, dy_op);
@@ -558,19 +558,19 @@ void ElemwiseExplicitGradCompute(const framework::ExecutionContext& ctx,
                                  framework::Tensor* dx, framework::Tensor* dy,
                                  DX_OP dx_op, DY_OP dy_op) {
   if (dy == nullptr) {
-    const framework::DDim dx_dims = dout.dims();
+    const framework::DDim& dx_dims = dout.dims();
     auto dy_dims = dx_dims;
     ElemwiseGradComputeNoBroadcast<DeviceContext, T, DX_OP, DY_OP>(
         ctx, dx_dims, dy_dims, x, y, out, dout, axis, dx, dy, dx_op, dy_op);
   } else {
     if (dout.dims() == dy->dims()) {
-      const framework::DDim dx_dims = dout.dims();
-      const framework::DDim dy_dims = dy->dims();
+      const framework::DDim& dx_dims = dout.dims();
+      const framework::DDim& dy_dims = dy->dims();
       ElemwiseGradComputeNoBroadcast<DeviceContext, T, DX_OP, DY_OP>(
           ctx, dx_dims, dy_dims, x, y, out, dout, axis, dx, dy, dx_op, dy_op);
     } else {  // Y is a scalar
       auto dx_dims = dout.dims();
-      const framework::DDim dy_dims = dy->dims();
+      const framework::DDim& dy_dims = dy->dims();
       ElemwiseGradComputeWithBroadcast<DeviceContext, T, DX_OP, DY_OP>(
           ctx, dx_dims, dy_dims, x, y, out, dout, axis, dx, dy, dx_op, dy_op);
     }

paddle/fluid/operators/feed_op.cc

@@ -31,7 +31,6 @@ class FeedOp : public framework::OperatorBase {
                const platform::Place &place) const override {
     // get device context from pool
     auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place);
-    platform::RecordEvent record_event(Type(), dev_ctx);
 
     auto feed_var_name = Input("X");
     auto *feed_var = scope.FindVar(feed_var_name);

paddle/fluid/operators/fetch_barrier_op.cc

@@ -36,12 +36,6 @@ class FetchBarrierOp : public framework::OperatorBase {
   void RunImpl(const framework::Scope& scope,
                const platform::Place& place) const override {
     std::vector<std::string> eps = Attr<std::vector<std::string>>("endpoints");
-
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    auto& ctx = *pool.Get(place);
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 

paddle/fluid/operators/fetch_op.cc

@@ -30,9 +30,6 @@ class FetchOp : public framework::OperatorBase {
  private:
   void RunImpl(const framework::Scope &scope,
                const platform::Place &place) const override {
-    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
-    platform::RecordEvent record_event(Type(), pool.Get(place));
-
     auto fetch_var_name = Input("X");
     auto *fetch_var = scope.FindVar(fetch_var_name);
     PADDLE_ENFORCE(fetch_var != nullptr,

paddle/fluid/operators/fused_elemwise_activation_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 <string>
+#include <vector>
+
+#include "paddle/fluid/operators/fused_elemwise_activation_op.h"
+
+namespace paddle {
+namespace operators {
+
+class FusedElemwiseActivationOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(
+        ctx->HasInput("X"),
+        "Input(X) of FusedElemwiseActivationOp op should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput("Y"),
+        "Input(Y) of FusedElemwiseActivationOp op should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("Out"),
+        "Output(Out) of FusedElemwiseActivationOp op should not be null.");
+
+    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.");
+
+    ctx->SetOutputDim("Out", x_dim);
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    PADDLE_ENFORCE_EQ(ctx.Input<framework::Tensor>("X")->type(),
+                      ctx.Input<framework::Tensor>("Y")->type(),
+                      "The element's type of input should be the same.");
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<framework::Tensor>("X")->type());
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
+};
+
+class FusedElemwiseActivationMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "(vector<Tensor>)");
+    AddInput("Y", "(vector<Tensor>)");
+    AddOutput("Out", "vector<Tensor>");
+    AddAttr<int>("axis",
+                 "axis is used by elementwise_op, the default value is -1.")
+        .SetDefault(-1);
+    AddAttr<float>("scale",
+                   "scale is used by scale_op, the default value is 0.0.")
+        .SetDefault(0.0);
+    AddAttr<bool>("recomputation",
+                  "Whether to recompute the Out."
+                  "fused_elemwise_activation_grad has two methods to get the "
+                  "dx and dy, one "
+                  "is to use the 'Out', and the other is not to use it. "
+                  "The former method will save the time of recomputing the "
+                  "'Out', but it must occupy the memory to store the 'out'. "
+                  "While, the later method can avoid occupying the memory, "
+                  "but it must recompute the 'Out'. The default value is true.")
+        .SetDefault(true);
+    AddAttr<std::vector<std::string>>("functor_list",
+                                      "The functors that should be fused.")
+        .AddCustomChecker([&](const std::vector<std::string> &functor_list) {
+          PADDLE_ENFORCE(ValidCheck(functor_list));
+        });
+
+    AddComment(R"DOC(
+FusedElemwiseActivation Operator.
+
+At present, FusedElemwiseActivation only supports Two kinds of compound
+operators (elementwise_op and activation_op):
+
+    Z = Binary(X, Unary(Y))
+    Z = Unary(Binary(X, Y))
+
+The attributions of activation_op can be get from fused_elemwise_activation_op's
+attributions. functor_list records the functors to be fused, for example
+"scale,elementwise_add".
+
+)DOC");
+  }
+
+ private:
+  bool ValidCheck(const std::vector<std::string> &functors) {
+    std::unordered_set<std::string> unary_fun = {"scale", "relu"};
+    std::unordered_set<std::string> binary_fun = {"elementwise_add"};
+
+    std::string unary_fun_str;
+    if (binary_fun.count(functors[0])) {
+      unary_fun_str = functors[1];
+    } else if (binary_fun.count(functors[1])) {
+      unary_fun_str = functors[0];
+    } else {
+      PADDLE_THROW("%s and %s are not included in fused_list.", functors[0],
+                   functors[1]);
+    }
+    PADDLE_ENFORCE_EQ(unary_fun.count(unary_fun_str), 1,
+                      "%s is not included in fused_list.", unary_fun_str);
+    return true;
+  }
+};
+
+class FusedElemwiseActivationGradMaker
+    : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto *op_desc_ptr = new framework::OpDesc();
+    op_desc_ptr->SetType(this->ForwardOpType() + "_grad");
+
+    for (auto &input_param : this->InputNames()) {
+      op_desc_ptr->SetInput(input_param, this->Input(input_param));
+      op_desc_ptr->SetOutput(framework::GradVarName(input_param),
+                             this->InputGrad(input_param, true));
+    }
+
+    for (auto &output_param : this->OutputNames()) {
+      op_desc_ptr->SetInput(output_param, this->Output(output_param));
+      op_desc_ptr->SetInput(framework::GradVarName(output_param),
+                            this->OutputGrad(output_param));
+    }
+    op_desc_ptr->SetAttrMap(this->Attrs());
+
+    std::vector<std::string> functor_names =
+        boost::get<std::vector<std::string>>(
+            op_desc_ptr->GetAttr("functor_list"));
+    functor_names[0] += "_grad";
+    functor_names[1] += "_grad";
+    op_desc_ptr->SetAttr("functor_list", functor_names);
+    return std::unique_ptr<framework::OpDesc>(op_desc_ptr);
+  }
+};
+
+class FusedElemwiseActivationOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) should not be null");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+    auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
+
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Rank of first input must >= rank of second input.");
+
+    auto x_grad_name = framework::GradVarName("X");
+    auto y_grad_name = framework::GradVarName("Y");
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+    if (ctx->HasOutput(y_grad_name)) {
+      ctx->SetOutputDim(y_grad_name, y_dims);
+    }
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input_data_type_index = ctx.Input<framework::Tensor>("X")->type();
+    PADDLE_ENFORCE_EQ(input_data_type_index,
+                      ctx.Input<framework::Tensor>("Y")->type(),
+                      "The element's type of input should be the same.");
+    PADDLE_ENFORCE_EQ(
+        input_data_type_index,
+        ctx.Input<framework::Tensor>(framework::GradVarName("Out"))->type(),
+        "The element's type of input should be the same.");
+
+    auto input_data_type = framework::ToDataType(input_data_type_index);
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(fused_elemwise_activation, ops::FusedElemwiseActivationOp,
+                  ops::FusedElemwiseActivationMaker,
+                  ops::FusedElemwiseActivationGradMaker);
+REGISTER_OPERATOR(fused_elemwise_activation_grad,
+                  ops::FusedElemwiseActivationOpGrad);
+
+REGISTER_OP_CPU_KERNEL(
+    fused_elemwise_activation,
+    ops::FusedElemwiseActivationKernel<paddle::platform::CPUDeviceContext,
+                                       float>,
+    ops::FusedElemwiseActivationKernel<paddle::platform::CPUDeviceContext,
+                                       double>);
+
+REGISTER_OP_CPU_KERNEL(
+    fused_elemwise_activation_grad,
+    ops::FusedElemwiseActivationGradKernel<paddle::platform::CPUDeviceContext,
+                                           float>,
+    ops::FusedElemwiseActivationGradKernel<paddle::platform::CPUDeviceContext,
+                                           double>);

paddle/fluid/operators/fused_elemwise_activation_op.cu

+/* 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/fused_elemwise_activation_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    fused_elemwise_activation,
+    ops::FusedElemwiseActivationKernel<paddle::platform::CUDADeviceContext,
+                                       float>,
+    ops::FusedElemwiseActivationKernel<paddle::platform::CUDADeviceContext,
+                                       double>);
+
+REGISTER_OP_CUDA_KERNEL(
+    fused_elemwise_activation_grad,
+    ops::FusedElemwiseActivationGradKernel<paddle::platform::CUDADeviceContext,
+                                           float>,
+    ops::FusedElemwiseActivationGradKernel<paddle::platform::CUDADeviceContext,
+                                           double>);

paddle/fluid/operators/fused_elemwise_activation_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 <vector>
+#include "paddle/fluid/framework/op_desc.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/detail/safe_ref.h"
+#include "paddle/fluid/operators/elementwise_op_function.h"
+#include "paddle/fluid/operators/math/functors.h"
+
+namespace math = paddle::operators::math;
+
+namespace paddle {
+namespace operators {
+
+// CompoundFunctors
+// For example: Z = Binary(X, Unary(Y))
+template <typename T, typename BinaryFun, typename UnaryFun>
+struct BinaryCompoundFunctor {
+  BinaryCompoundFunctor(const BinaryFun &binary_fun, const UnaryFun &unary_fun)
+      : binary_fun_(binary_fun), unary_fun_(unary_fun) {}
+
+  inline HOSTDEVICE T operator()(T x, T y) {
+    return binary_fun_(x, unary_fun_(y));
+  }
+
+ private:
+  BinaryFun binary_fun_;
+  UnaryFun unary_fun_;
+};
+
+// For example: Z = Unary(Binary(X, Y))
+template <typename T, typename UnaryFun, typename BinaryFun>
+struct UnaryCompoundFunctor {
+  UnaryCompoundFunctor(const UnaryFun &unary_fun, const BinaryFun &binary_fun)
+      : unary_fun_(unary_fun), binary_fun_(binary_fun) {}
+
+  inline HOSTDEVICE T operator()(T x, T y) {
+    return unary_fun_(binary_fun_(x, y));
+  }
+
+ private:
+  UnaryFun unary_fun_;
+  BinaryFun binary_fun_;
+};
+
+// FIXME(zcd): DBinaryFun and DUnaryFun have to method to get
+// the dx, one is to use the 'out', and the other is not to use it.
+// the former method will save the time of recomputing the
+// 'out', but it must occupy the memory to store the 'out'.
+// While the later method can avoid occupying this memory,
+// but it must recompute the 'out'.
+
+template <typename T, typename DBinaryFun, typename UnaryFun,
+          bool Recomputation = true>
+struct BinaryCompoundGradDxFunctor {
+  BinaryCompoundGradDxFunctor(const DBinaryFun &d_binary_fun,
+                              const UnaryFun &unary_fun)
+      : d_binary_fun_(d_binary_fun), unary_fun_(unary_fun) {}
+
+  inline HOSTDEVICE T operator()(T x, T y, T out, T dout) {
+    if (Recomputation) {
+      return dout * d_binary_fun_(x, unary_fun_(y));
+    } else {
+      return dout * d_binary_fun_(x, unary_fun_(y), out);
+    }
+  }
+
+ private:
+  DBinaryFun d_binary_fun_;
+  UnaryFun unary_fun_;
+};
+
+template <typename T, typename DBinaryFun, typename UnaryFun,
+          typename DUnaryFun, bool Recomputation = true>
+struct BinaryCompoundGradDyFunctor {
+  BinaryCompoundGradDyFunctor(const DBinaryFun &d_binary_fun,
+                              const UnaryFun &unary_fun,
+                              const DUnaryFun &d_unary_fun)
+      : d_binary_fun_(d_binary_fun),
+        unary_fun_(unary_fun),
+        d_unary_fun_(d_unary_fun) {}
+
+  inline HOSTDEVICE T operator()(T x, T y, T out, T dout) {
+    if (Recomputation) {
+      return dout * d_binary_fun_(unary_fun_(y), x) * d_unary_fun_(y);
+    } else {
+      return dout * d_binary_fun_(unary_fun_(y), x, out) * d_unary_fun_(y);
+    }
+  }
+
+ private:
+  DBinaryFun d_binary_fun_;
+  UnaryFun unary_fun_;
+  DUnaryFun d_unary_fun_;
+};
+
+template <typename T, typename DUnaryFun, typename BinaryFun,
+          typename DBinaryFun, bool Recomputation = true>
+struct UnaryCompoundGradDxFunctor {
+  UnaryCompoundGradDxFunctor(const DUnaryFun &d_unary_fun,
+                             const BinaryFun &binary_fun,
+                             const DBinaryFun &d_binary_fun)
+      : d_unary_fun_(d_unary_fun),
+        binary_fun_(binary_fun),
+        d_binary_fun_(d_binary_fun) {}
+
+  inline HOSTDEVICE T operator()(T x, T y, T out, T dout) {
+    T base;
+    if (Recomputation) {
+      base = dout * d_unary_fun_(binary_fun_(x, y));
+    } else {
+      base = dout * d_unary_fun_(binary_fun_(x, y), out);
+    }
+    return base * d_binary_fun_(x, y);
+  }
+
+ private:
+  DUnaryFun d_unary_fun_;
+  BinaryFun binary_fun_;
+  DBinaryFun d_binary_fun_;
+};
+
+template <typename T, typename DUnaryFun, typename BinaryFun,
+          typename DBinaryFun, bool Recomputation = true>
+struct UnaryCompoundGradDyFunctor {
+  UnaryCompoundGradDyFunctor(const DUnaryFun &d_unary_fun,
+                             const BinaryFun &binary_fun,
+                             const DBinaryFun &d_binary_fun)
+      : d_unary_fun_(d_unary_fun),
+        binary_fun_(binary_fun),
+        d_binary_fun_(d_binary_fun) {}
+
+  inline HOSTDEVICE T operator()(T x, T y, T out, T dout) {
+    T base;
+    if (Recomputation) {
+      base = dout * d_unary_fun_(binary_fun_(x, y));
+    } else {
+      base = dout * d_unary_fun_(binary_fun_(x, y), out);
+    }
+    return base * d_binary_fun_(y, x);
+  }
+
+ private:
+  DUnaryFun d_unary_fun_;
+  BinaryFun binary_fun_;
+  DBinaryFun d_binary_fun_;
+};
+
+template <typename DeviceContext, typename T, typename BinaryFunctor,
+          typename UnaryFunctor>
+static void RunBinaryCompoundFunctor(const framework::ExecutionContext &ctx,
+                                     const BinaryFunctor &binary_functor,
+                                     const UnaryFunctor &unary_functor,
+                                     const framework::Tensor *in_x,
+                                     const framework::Tensor *in_y,
+                                     framework::Tensor *output) {
+  int axis = ctx.Attr<int>("axis");
+  using BinaryCompoundFunctor =
+      BinaryCompoundFunctor<T, BinaryFunctor, UnaryFunctor>;
+
+  ElementwiseComputeEx<BinaryCompoundFunctor, DeviceContext, T>(
+      ctx, in_x, in_y, axis,
+      BinaryCompoundFunctor(binary_functor, unary_functor), output);
+}
+
+template <typename DeviceContext, typename T, typename UnaryFunctor,
+          typename BinaryFunctor>
+static void RunUnaryCompoundFunctors(const framework::ExecutionContext &ctx,
+                                     const UnaryFunctor &unary_functor,
+                                     const BinaryFunctor &binary_functor,
+                                     const framework::Tensor *in_x,
+                                     const framework::Tensor *in_y,
+                                     framework::Tensor *output) {
+  int axis = ctx.Attr<int>("axis");
+
+  using UnaryCompoundFunctor =
+      UnaryCompoundFunctor<T, UnaryFunctor, BinaryFunctor>;
+
+  ElementwiseComputeEx<UnaryCompoundFunctor, DeviceContext, T>(
+      ctx, in_x, in_y, axis,
+      UnaryCompoundFunctor(unary_functor, binary_functor), output);
+}
+
+template <typename DeviceContext, typename T, typename BinaryGradFunctor,
+          typename UnaryFunctor, typename UnaryGradFunctor,
+          bool Recomputation = true>
+static void RunBinaryCompoundGradFunctors(
+    const framework::ExecutionContext &ctx,
+    const BinaryGradFunctor &binary_grad_functor,
+    const UnaryFunctor &unary_functor,
+    const UnaryGradFunctor &unary_grad_functor, const framework::Tensor *in_x,
+    const framework::Tensor *in_y, const framework::Tensor *in_out,
+    const framework::Tensor *in_out_grad, framework::Tensor *x_grad,
+    framework::Tensor *y_grad) {
+  int axis = ctx.Attr<int>("axis");
+
+  using BinaryCompoundDxFunctor =
+      BinaryCompoundGradDxFunctor<T, BinaryGradFunctor, UnaryFunctor,
+                                  Recomputation>;
+  using BinaryCompoundDyFunctor =
+      BinaryCompoundGradDyFunctor<T, BinaryGradFunctor, UnaryFunctor,
+                                  UnaryGradFunctor, Recomputation>;
+
+  ElemwiseGradCompute<DeviceContext, T, BinaryCompoundDxFunctor,
+                      BinaryCompoundDyFunctor>(
+      ctx, *in_x, *in_y, *in_out, *in_out_grad, axis, x_grad, y_grad,
+      BinaryCompoundDxFunctor(binary_grad_functor, unary_functor),
+      BinaryCompoundDyFunctor(binary_grad_functor, unary_functor,
+                              unary_grad_functor));
+}
+
+template <typename DeviceContext, typename T, typename UnaryGradFunctor,
+          typename BinaryFunctor, typename BinaryGradFunctor,
+          bool Recomputation = true>
+static void RunUnaryCompoundGradFunctors(
+    const framework::ExecutionContext &ctx,
+    const UnaryGradFunctor &unary_grad_functor,
+    const BinaryFunctor &binary_functor,
+    const BinaryGradFunctor &binary_grad_functor, const framework::Tensor *in_x,
+    const framework::Tensor *in_y, const framework::Tensor *in_out,
+    const framework::Tensor *in_out_grad, framework::Tensor *x_grad,
+    framework::Tensor *y_grad) {
+  int axis = ctx.Attr<int>("axis");
+
+  using UnaryCompoundDxFunctor =
+      UnaryCompoundGradDxFunctor<T, UnaryGradFunctor, BinaryFunctor,
+                                 BinaryGradFunctor, Recomputation>;
+  using UnaryCompoundDyFunctor =
+      UnaryCompoundGradDyFunctor<T, UnaryGradFunctor, BinaryFunctor,
+                                 BinaryGradFunctor, Recomputation>;
+
+  ElemwiseGradCompute<DeviceContext, T, UnaryCompoundDxFunctor,
+                      UnaryCompoundDyFunctor>(
+      ctx, *in_x, *in_y, *in_out, *in_out_grad, axis, x_grad, y_grad,
+      UnaryCompoundDxFunctor(unary_grad_functor, binary_functor,
+                             binary_grad_functor),
+      UnaryCompoundDyFunctor(unary_grad_functor, binary_functor,
+                             binary_grad_functor));
+}
+
+template <typename DeviceContext, typename T>
+static void RunFunctors(const framework::ExecutionContext &ctx,
+                        const framework::Tensor *in_x,
+                        const framework::Tensor *in_y,
+                        framework::Tensor *output) {
+  auto &functors = ctx.Attr<std::vector<std::string>>("functor_list");
+  auto funcs_str = functors[0] + "," + functors[1];
+  // TODO(zcd): The following code can be refined.
+  if (funcs_str == "elementwise_add,scale") {
+    // Z = Binary(X, Unary(Y))
+    T scale = static_cast<T>(ctx.Attr<float>("scale"));
+    RunBinaryCompoundFunctor<DeviceContext, T, math::AddFunctor<T>,
+                             math::ScaleFunctor<T>>(
+        ctx, math::AddFunctor<T>(), math::ScaleFunctor<T>(scale), in_x, in_y,
+        output);
+  } else if (funcs_str == "scale,elementwise_add") {
+    // Z = Unary(Binary(X, Y))
+    T scale = static_cast<T>(ctx.Attr<float>("scale"));
+    RunUnaryCompoundFunctors<DeviceContext, T, math::ScaleFunctor<T>,
+                             math::AddFunctor<T>>(
+        ctx, math::ScaleFunctor<T>(scale), math::AddFunctor<T>(), in_x, in_y,
+        output);
+  } else if (funcs_str == "elementwise_add,relu") {
+    RunBinaryCompoundFunctor<DeviceContext, T, math::AddFunctor<T>,
+                             math::ReluFunctor<T>>(
+        ctx, math::AddFunctor<T>(), math::ReluFunctor<T>(), in_x, in_y, output);
+  } else if (funcs_str == "relu,elementwise_add") {
+    RunUnaryCompoundFunctors<DeviceContext, T, math::ReluFunctor<T>,
+                             math::AddFunctor<T>>(
+        ctx, math::ReluFunctor<T>(), math::AddFunctor<T>(), in_x, in_y, output);
+  } else {
+    PADDLE_THROW("%s has not been implemented.", funcs_str);
+  }
+}
+
+template <typename DeviceContext, typename T>
+static void RunGradFunctors(const framework::ExecutionContext &ctx,
+                            const framework::Tensor *in_x,
+                            const framework::Tensor *in_y,
+                            const framework::Tensor *in_out,
+                            const framework::Tensor *in_out_grad,
+                            framework::Tensor *x_grad,
+                            framework::Tensor *y_grad) {
+  auto &functors = ctx.Attr<std::vector<std::string>>("functor_list");
+  auto funcs_str = functors[0] + "," + functors[1];
+
+  bool recomputation = ctx.Attr<bool>("recomputation");
+
+  // TODO(zcd): The following code can be refined. for example, use registion
+  if (funcs_str == "elementwise_add_grad,scale_grad") {
+    // The backward of Z = Binary(X, Unary(Y))
+    T scale = static_cast<T>(ctx.Attr<float>("scale"));
+    if (recomputation) {
+      RunBinaryCompoundGradFunctors<DeviceContext, T, math::AddGradFunctor<T>,
+                                    math::ScaleFunctor<T>,
+                                    math::ScaleGradFunctor<T>, true>(
+          ctx, math::AddGradFunctor<T>(), math::ScaleFunctor<T>(scale),
+          math::ScaleGradFunctor<T>(scale), in_x, in_y, in_out, in_out_grad,
+          x_grad, y_grad);
+    } else {
+      RunBinaryCompoundGradFunctors<DeviceContext, T, math::AddGradFunctor<T>,
+                                    math::ScaleFunctor<T>,
+                                    math::ScaleGradFunctor<T>, false>(
+          ctx, math::AddGradFunctor<T>(), math::ScaleFunctor<T>(scale),
+          math::ScaleGradFunctor<T>(scale), in_x, in_y, in_out, in_out_grad,
+          x_grad, y_grad);
+    }
+  } else if (funcs_str == "scale_grad,elementwise_add_grad") {
+    // The backward of Z = Unary(Binary(X, Y))
+    T scale = static_cast<T>(ctx.Attr<float>("scale"));
+    if (recomputation) {
+      RunUnaryCompoundGradFunctors<DeviceContext, T, math::ScaleGradFunctor<T>,
+                                   math::AddFunctor<T>, math::AddGradFunctor<T>,
+                                   true>(ctx, math::ScaleGradFunctor<T>(scale),
+                                         math::AddFunctor<T>(),
+                                         math::AddGradFunctor<T>(), in_x, in_y,
+                                         in_out, in_out_grad, x_grad, y_grad);
+    } else {
+      RunUnaryCompoundGradFunctors<DeviceContext, T, math::ScaleGradFunctor<T>,
+                                   math::AddFunctor<T>, math::AddGradFunctor<T>,
+                                   false>(ctx, math::ScaleGradFunctor<T>(scale),
+                                          math::AddFunctor<T>(),
+                                          math::AddGradFunctor<T>(), in_x, in_y,
+                                          in_out, in_out_grad, x_grad, y_grad);
+    }
+  } else if (funcs_str == "elementwise_add_grad,relu_grad") {
+    if (recomputation) {
+      RunBinaryCompoundGradFunctors<DeviceContext, T, math::AddGradFunctor<T>,
+                                    math::ReluFunctor<T>,
+                                    math::ReluGradFunctor<T>, true>(
+          ctx, math::AddGradFunctor<T>(), math::ReluFunctor<T>(),
+          math::ReluGradFunctor<T>(), in_x, in_y, in_out, in_out_grad, x_grad,
+          y_grad);
+    } else {
+      RunBinaryCompoundGradFunctors<DeviceContext, T, math::AddGradFunctor<T>,
+                                    math::ReluFunctor<T>,
+                                    math::ReluGradFunctor<T>, false>(
+          ctx, math::AddGradFunctor<T>(), math::ReluFunctor<T>(),
+          math::ReluGradFunctor<T>(), in_x, in_y, in_out, in_out_grad, x_grad,
+          y_grad);
+    }
+  } else if (funcs_str == "relu_grad,elementwise_add_grad") {
+    if (recomputation) {
+      RunUnaryCompoundGradFunctors<DeviceContext, T, math::ReluGradFunctor<T>,
+                                   math::AddFunctor<T>, math::AddGradFunctor<T>,
+                                   true>(ctx, math::ReluGradFunctor<T>(),
+                                         math::AddFunctor<T>(),
+                                         math::AddGradFunctor<T>(), in_x, in_y,
+                                         in_out, in_out_grad, x_grad, y_grad);
+    } else {
+      RunUnaryCompoundGradFunctors<DeviceContext, T, math::ReluGradFunctor<T>,
+                                   math::AddFunctor<T>, math::AddGradFunctor<T>,
+                                   false>(ctx, math::ReluGradFunctor<T>(),
+                                          math::AddFunctor<T>(),
+                                          math::AddGradFunctor<T>(), in_x, in_y,
+                                          in_out, in_out_grad, x_grad, y_grad);
+    }
+  } else {
+    PADDLE_THROW("%s has not been implemented.", funcs_str);
+  }
+}
+
+template <typename DeviceContext, typename T>
+class FusedElemwiseActivationKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto &in_x = detail::Ref(ctx.Input<framework::Tensor>("X"),
+                             "Cannot get input tensor %s, variable name = %s",
+                             "X", ctx.op().Input("X"));
+    auto &in_y = detail::Ref(ctx.Input<framework::Tensor>("Y"),
+                             "Cannot get input tensor %s, variable name = %s",
+                             "Y", ctx.op().Input("Y"));
+    auto &output = detail::Ref(ctx.Output<framework::Tensor>("Out"),
+                               "Cannot get input tensor %s, variable name = %s",
+                               "Out", ctx.op().Output("Out"));
+
+    RunFunctors<DeviceContext, T>(ctx, &in_x, &in_y, &output);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class FusedElemwiseActivationGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto &in_x = detail::Ref(ctx.Input<framework::Tensor>("X"),
+                             "Cannot get input tensor %s, variable name = %s",
+                             "X", ctx.op().Input("X"));
+    auto &in_y = detail::Ref(ctx.Input<framework::Tensor>("Y"),
+                             "Cannot get input tensor %s, variable name = %s",
+                             "Y", ctx.op().Input("Y"));
+    auto &in_out = detail::Ref(ctx.Input<framework::Tensor>("Out"),
+                               "Cannot get input tensor %s, variable name = %s",
+                               "Out", ctx.op().Input("Out"));
+    auto &in_out_grad =
+        detail::Ref(ctx.Input<framework::Tensor>(framework::GradVarName("Out")),
+                    "Cannot get input tensor %s, variable name = %s",
+                    framework::GradVarName("Out"),
+                    ctx.op().Input(framework::GradVarName("Out")));
+
+    framework::Tensor *x_grad =
+        ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    framework::Tensor *y_grad =
+        ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
+
+    RunGradFunctors<DeviceContext, T>(ctx, &in_x, &in_y, &in_out, &in_out_grad,
+                                      x_grad, y_grad);
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/load_op.cc

@@ -31,9 +31,6 @@ class LoadOp : public framework::OperatorBase {
  private:
   void RunImpl(const framework::Scope &scope,
                const platform::Place &place) const override {
-    auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place);
-    platform::RecordEvent record_event(Type(), dev_ctx);
-
     // FIXME(yuyang18): We save variable to local file now, but we should change
     // it to save an output stream.
     auto filename = Attr<std::string>("file_path");

paddle/fluid/operators/math/functors.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
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+// AddFunctor
+template <typename T>
+struct AddFunctor {
+  // out = x + y;
+  inline HOSTDEVICE T operator()(T x, T y) { return x + y; }
+};
+
+template <typename T>
+struct AddGradFunctor {
+  inline HOSTDEVICE T operator()(T x, T y) { return 1; }
+
+  inline HOSTDEVICE T operator()(T x, T y, T out) const { return 1; }
+};
+
+template <typename T>
+struct ScaleFunctor {
+  explicit ScaleFunctor(const T coeff) : coeff_(coeff) {}
+
+  inline HOSTDEVICE T operator()(T ele) { return ele * coeff_; }
+
+ private:
+  T coeff_;
+};
+
+template <typename T>
+struct ScaleGradFunctor {
+  explicit ScaleGradFunctor(T coeff) : coeff_(coeff) {}
+
+  inline HOSTDEVICE T operator()(T x) { return coeff_; }
+
+  inline HOSTDEVICE T operator()(T x, T out) { return coeff_; }
+
+ private:
+  T coeff_;
+};
+
+template <typename T>
+struct ReluFunctor {
+  inline HOSTDEVICE T operator()(T x) { return x * (x > 0); }
+};
+
+template <typename T>
+struct ReluGradFunctor {
+  inline HOSTDEVICE T operator()(T x) { return x > 0 ? 1 : 0; }
+
+  inline HOSTDEVICE T operator()(T x, T out) { return x > 0 ? 1 : 0; }
+};
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/parallel_do_op.cc

@@ -163,8 +163,7 @@ class ParallelDoOp : public framework::OperatorBase {
       auto &place = places[place_idx];
       auto *cur_scope = sub_scopes[place_idx];
 
-      workers.emplace_back(
-          framework::Async([program, cur_scope, place, block, place_idx] {
+      workers.emplace_back(framework::Async([program, cur_scope, place, block] {
         framework::Executor executor(place);
         executor.Run(*program, cur_scope, block->ID(),
                      false /*create_local_scope*/);
@@ -239,8 +238,7 @@ class ParallelDoGradOp : public framework::OperatorBase {
       auto *cur_scope = sub_scopes[i];
 
       // execute
-      workers.emplace_back(
-          framework::Async([program, cur_scope, place, block, i] {
+      workers.emplace_back(framework::Async([program, cur_scope, place, block] {
         framework::Executor executor(place);
         executor.Run(*program, cur_scope, block->ID(),
                      false /*create_local_scope*/);

paddle/fluid/operators/read_op.cc

@@ -15,6 +15,7 @@
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
@@ -65,6 +66,12 @@ class ReadOp : public framework::OperatorBase {
             .GetMutable<framework::ReaderHolder>();
     std::vector<std::string> out_arg_names = Outputs("Out");
     std::vector<framework::LoDTensor> ins;
+
+    // For profiling
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    auto& ctx = *pool.Get(dev_place);
+    platform::RecordEvent record_event(Type(), &ctx);
+
     reader->ReadNext(&ins);
     if (ins.empty()) {
       if (Attr<bool>("throw_eof_exp")) {

paddle/fluid/operators/recv_op.cc

@@ -40,8 +40,6 @@ class RecvOp : public framework::OperatorBase {
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
 
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();

paddle/fluid/operators/send_barrier_op.cc

@@ -39,11 +39,6 @@ class SendBarrierOp : public framework::OperatorBase {
     std::vector<std::string> eps = Attr<std::vector<std::string>>("endpoints");
     bool sync_mode = Attr<bool>("sync_mode");
 
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    auto& ctx = *pool.Get(place);
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 

paddle/fluid/operators/send_op.cc

@@ -42,9 +42,6 @@ class SendOp : public framework::OperatorBase {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
 
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 

paddle/fluid/operators/softmax_with_cross_entropy_op.cu

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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.
@@ -14,6 +14,8 @@ limitations under the License. */
 
 #define EIGEN_USE_GPU
 
+#include <cub/cub.cuh>
+#include "paddle/fluid/operators/math/cross_entropy.h"
 #include "paddle/fluid/operators/softmax_with_cross_entropy_op.h"
 
 namespace paddle {
@@ -53,8 +55,196 @@ __global__ void SoftCrossEntropyGradientKernel(T* logit_grad,
     logit_grad[ids] = loss_grad[row_ids] * (logit_grad[ids] - labels[ids]);
   }
 }
+
 }  // namespace
 
+static __device__ __forceinline__ float real_exp(float x) { return expf(x); }
+static __device__ __forceinline__ double real_exp(double x) { return exp(x); }
+static __device__ __forceinline__ float real_log(float x) {
+  return math::TolerableValue<float>()(logf(x));
+}
+static __device__ __forceinline__ double real_log(double x) {
+  return math::TolerableValue<double>()(log(x));
+}
+
+/** In the following codes, 3 CUDA kernels are implemented to calculate softmax
+ * and loss **/
+/*
+  Supposing the x is `logits` and y is `labels`, the equations are as
+followings:
+
+  cross\_entropy_i = \sum_{j}[- y_i_j * log({e^{x_i_j}/\sum_{j}e^{x_i_j}})]
+        = \sum_{j}[- y_i_j * log({e^{x_i_j - max_i}/\sum_{j}e^{x_i_j-max_i}})]
+        = \sum_{j}[-y_i_j * (x_i_j - max_i - log\sum_{j}e^{x_i_j - max_i})]
+        = \sum_{j}[-y_i_j * (x_i_j - max_i - logDiffMaxSum_i)]
+        = \sum_{j}(-y_i_j * tmp_i_j)
+
+  softmax_i_j = e^{tmp_i_j}
+
+where:
+  max_i = \max_{j}{x_i_j}
+  logDiffMaxSum_i = log\sum_{j}e^{x_i_j - max_i}
+  tmp_i_j = x_i_j - max_i - logDiffMaxSum_i
+
+Therefore, the calculation can be separated into 3 steps:
+Step 1: row-wise operation to calculate max_i
+Step 2: row-wise operation to calculate logDiffMaxSum_i
+Step 3: caculate tmp_i_j, and finally get softmax_i_j and cross\_entropy_i
+
+To save memory, we can share memory among max_i, logDiffMaxSum_i and
+cross\_entropy_i.
+In this way, the 3 steps should be changed to:
+Step 1 (RowReductionForMax): row-wise operation to calculate max_i
+Step 2 (RowReductionForDiffMaxSum): calculate immediate result of softmax'_i_j =
+x_i_j - max_i, and row-wise operation to calculate logDiffMaxSum_i
+Step 3 (RowReductionForSoftmaxAndCrossEntropy): calculate tmp_i_j = softmax'_i_j
+- logDiffMaxSum_i, and finally get softmax_i_j and cross\_entropy_i
+*/
+
+// There are 3 kinds of reduce algorithms in cub:
+// BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY
+// BLOCK_REDUCE_RAKING
+// BLOCK_REDUCE_WARP_REDUCTIONS (default)
+template <typename T, int BlockDim>
+using BlockReduce =
+    cub::BlockReduce<T, BlockDim /*, cub::BLOCK_REDUCE_WARP_REDUCTIONS*/>;
+
+template <typename T, int BlockDim>
+using BlockReduceTempStorage = typename BlockReduce<T, BlockDim>::TempStorage;
+
+// Make sure that BlockDim <= feature_size
+// This kernel is used to calculate the max element of each row
+template <typename T, int BlockDim>
+__global__ void RowReductionForMax(const T* logits_data, T* max_data,
+                                   int feature_size) {
+  __shared__ BlockReduceTempStorage<T, BlockDim> temp_storage;
+
+  auto beg_idx = feature_size * blockIdx.x + threadIdx.x;
+  auto end_idx = feature_size * (blockIdx.x + 1);
+
+  T cur_max = logits_data[beg_idx];
+  beg_idx += BlockDim;
+  while (beg_idx < end_idx) {
+    if (cur_max < logits_data[beg_idx]) {
+      cur_max = logits_data[beg_idx];
+    }
+    beg_idx += BlockDim;
+  }
+
+  cur_max = BlockReduce<T, BlockDim>(temp_storage).Reduce(cur_max, cub::Max());
+
+  if (threadIdx.x == 0) {
+    max_data[blockIdx.x] = cur_max < -64 ? -64 : cur_max;
+  }
+}
+
+// Make sure that BlockDim <= feature_size
+template <typename T, int BlockDim>
+__global__ void RowReductionForDiffMaxSum(const T* logits_data, T* max_data,
+                                          T* softmax, int feature_size) {
+  __shared__ BlockReduceTempStorage<T, BlockDim> temp_storage;
+
+  auto beg_idx = feature_size * blockIdx.x + threadIdx.x;
+  auto end_idx = feature_size * (blockIdx.x + 1);
+
+  auto block_max = max_data[blockIdx.x];
+
+  softmax[beg_idx] = logits_data[beg_idx] - block_max;
+  T diff_max_sum = real_exp(softmax[beg_idx]);
+  beg_idx += BlockDim;
+  while (beg_idx < end_idx) {
+    softmax[beg_idx] = logits_data[beg_idx] - block_max;
+    diff_max_sum += real_exp(softmax[beg_idx]);
+    beg_idx += BlockDim;
+  }
+
+  diff_max_sum =
+      BlockReduce<T, BlockDim>(temp_storage).Reduce(diff_max_sum, cub::Sum());
+  if (threadIdx.x == 0) max_data[blockIdx.x] = real_log(diff_max_sum);
+}
+
+// Make sure that BlockDim <= feature_size
+template <typename T, int BlockDim>
+__global__ void RowReductionForSoftmaxAndCrossEntropy(const T* logits_data,
+                                                      const T* labels_data,
+                                                      T* loss_data, T* softmax,
+                                                      int feature_size) {
+  __shared__ BlockReduceTempStorage<T, BlockDim> temp_storage;
+
+  auto beg_idx = feature_size * blockIdx.x + threadIdx.x;
+  auto end_idx = feature_size * (blockIdx.x + 1);
+
+  // log_diff_max_sum shares memory with loss
+  auto block_log_diff_max_sum = loss_data[blockIdx.x];
+  auto tmp = softmax[beg_idx] - block_log_diff_max_sum;
+  softmax[beg_idx] = real_exp(tmp);
+  auto loss = -labels_data[beg_idx] * tmp;
+  beg_idx += BlockDim;
+  while (beg_idx < end_idx) {
+    tmp = softmax[beg_idx] - block_log_diff_max_sum;
+    softmax[beg_idx] = real_exp(tmp);
+    loss -= (labels_data[beg_idx] * tmp);
+    beg_idx += BlockDim;
+  }
+
+  loss = BlockReduce<T, BlockDim>(temp_storage).Reduce(loss, cub::Sum());
+  if (threadIdx.x == 0) loss_data[blockIdx.x] = loss;
+}
+
+template <typename T>
+__global__ void SetSoftmaxToOneWhenFeatureSizeIsOne(T* out, int batch_size) {
+  auto idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < batch_size) out[idx] = static_cast<T>(1);
+}
+
+template <typename T>
+static void SoftmaxWithCrossEntropyFusedKernel(const T* logits_data,
+                                               const T* labels_data,
+                                               T* softmax_data, T* loss_data,
+                                               int batch_size, int feature_size,
+                                               cudaStream_t stream) {
+  constexpr int kMaxBlockDim = 512;
+  int block_dim = feature_size >= kMaxBlockDim
+                      ? kMaxBlockDim
+                      : (1 << static_cast<int>(std::log2(feature_size)));
+
+#define CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(BlockDim)                \
+  case BlockDim:                                                              \
+    RowReductionForMax<T, BlockDim><<<batch_size, BlockDim, 0, stream>>>(     \
+        logits_data, loss_data, feature_size);                                \
+    RowReductionForDiffMaxSum<T,                                              \
+                              BlockDim><<<batch_size, BlockDim, 0, stream>>>( \
+        logits_data, loss_data, softmax_data, feature_size);                  \
+    RowReductionForSoftmaxAndCrossEntropy<                                    \
+        T, BlockDim><<<batch_size, BlockDim, 0, stream>>>(                    \
+        logits_data, labels_data, loss_data, softmax_data, feature_size);     \
+    break
+
+  switch (block_dim) {
+    CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(512);
+    CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(256);
+    CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(128);
+    CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(64);
+    CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(32);
+    CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(16);
+    CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(8);
+    CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(4);
+    CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(2);
+    case 1:
+      SetSoftmaxToOneWhenFeatureSizeIsOne<<<(batch_size + kMaxBlockDim - 1) /
+                                                kMaxBlockDim,
+                                            kMaxBlockDim, 0, stream>>>(
+          softmax_data, batch_size);
+      cudaMemsetAsync(loss_data, 0, batch_size, stream);
+      break;
+    default:
+      PADDLE_THROW("BlockDim must be 2^n in softmax_with_cross_entropy_op");
+      break;
+  }
+
+#undef CALL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL
+}
+
 template <typename T>
 class SoftmaxWithCrossEntropyCUDAKernel : public framework::OpKernel<T> {
  public:
@@ -66,14 +256,24 @@ class SoftmaxWithCrossEntropyCUDAKernel : public framework::OpKernel<T> {
     Tensor* softmax = context.Output<Tensor>("Softmax");
 
     Tensor* loss = context.Output<Tensor>("Loss");
-    softmax->mutable_data<T>(context.GetPlace());
-    loss->mutable_data<T>(context.GetPlace());
+    auto* softmax_data = softmax->mutable_data<T>(context.GetPlace());
+    auto* loss_data = loss->mutable_data<T>(context.GetPlace());
 
-    math::SoftmaxFunctor<platform::CUDADeviceContext, T>()(
-        context.cuda_device_context(), logits, softmax);
+    auto soft_label = context.Attr<bool>("soft_label");
+    if (soft_label) {
+      int batch_size = logits->dims()[0];
+      int feature_size = logits->dims()[1];
+      auto* logits_data = logits->data<T>();
+      auto* labels_data = labels->data<T>();
+      SoftmaxWithCrossEntropyFusedKernel(
+          logits_data, labels_data, softmax_data, loss_data, batch_size,
+          feature_size, context.cuda_device_context().stream());
+    } else {
+      math::SoftmaxCUDNNFunctor<T>()(context.cuda_device_context(), logits,
+                                     softmax);
       math::CrossEntropyFunctor<platform::CUDADeviceContext, T>()(
-        context.cuda_device_context(), loss, softmax, labels,
-        context.Attr<bool>("soft_label"));
+          context.cuda_device_context(), loss, softmax, labels, false);
+    }
   }
 };
 

paddle/fluid/operators/tensorrt_engine_op.cc

@@ -55,18 +55,8 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t> &shape) {
                     "TensorRT' tensor input requires at least 2 dimensions");
   PADDLE_ENFORCE_LE(shape.size(), 4UL,
                     "TensorRT' tensor input requires at most 4 dimensions");
-
-  switch (shape.size()) {
-    case 2:
-      return nvinfer1::Dims2(1, shape[1]);
-    case 3:
-      return nvinfer1::Dims3(1, shape[1], shape[2]);
-    case 4:
-      return nvinfer1::Dims4(1, shape[1], shape[2], shape[3]);
-    default:
-      return nvinfer1::Dims();
-  }
-  return nvinfer1::Dims();
+  PADDLE_ENFORCE_EQ(shape.size(), 4UL);
+  return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]);
 }
 
 }  // namespace
@@ -86,6 +76,9 @@ void TensorRTEngineKernel<DeviceContext, T>::Prepare(
     parameters.insert(param);
   }
 
+  std::vector<std::string> output_maps =
+      context.Attr<std::vector<std::string>>("output_name_mapping");
+
   // TODO(Superjomn) replace this with a different stream
   auto *engine = Singleton<TRT_EngineManager>::Global().Create(
       max_batch, max_workspace, nullptr /*engine hold its own stream*/,
@@ -97,6 +90,7 @@ void TensorRTEngineKernel<DeviceContext, T>::Prepare(
   // Add inputs
   VLOG(4) << "declare inputs";
   for (auto &input : context.Inputs("Xs")) {
+    if (parameters.count(input)) continue;
     VLOG(4) << "declare input " << input;
     auto *var = block.FindVar(input);
     // TensorRT engine need to create parameters. The parameter's description
@@ -122,7 +116,7 @@ void TensorRTEngineKernel<DeviceContext, T>::Prepare(
       block_desc, parameters, context.scope(), engine);
 
   // Add outputs
-  for (auto &output : context.Outputs("Ys")) {
+  for (auto &output : output_maps) {
     engine->DeclareOutput(output);
   }
 

paddle/fluid/operators/tensorrt_engine_op.h

@@ -66,8 +66,17 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_LE(FLAGS_tensorrt_engine_batch_size,
                       context.Attr<int>("max_batch"));
 
+    std::vector<std::string> output_maps =
+        context.Attr<std::vector<std::string>>("output_name_mapping");
+
+    auto params = context.Attr<std::vector<std::string>>("parameters");
+    std::unordered_set<std::string> parameters;
+    for (const auto& param : params) {
+      parameters.insert(param);
+    }
     // Convert input tensor from fluid to engine.
     for (const auto& x : context.Inputs("Xs")) {
+      if (parameters.count(x)) continue;
       // convert input and copy to TRT engine's buffer
       auto& t = inference::analysis::GetFromScope<framework::LoDTensor>(
           context.scope(), x);
@@ -82,10 +91,12 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
     // Execute the engine.
     PADDLE_ENFORCE_GT(FLAGS_tensorrt_engine_batch_size, 0);
     engine->Execute(FLAGS_tensorrt_engine_batch_size);
+
     // Convert output tensor from engine to fluid
+    int output_index = 0;
     for (const auto& y : context.Outputs("Ys")) {
       // convert output and copy to fluid.
-      nvinfer1::ITensor* trt_t = engine->GetITensor(y);
+      nvinfer1::ITensor* trt_t = engine->GetITensor(output_maps[output_index]);
       auto dims = trt_t->getDimensions();
       // Use the output ITensor's dims to reshape the Fluid Tensor.
       std::vector<int> ddim(dims.d, dims.d + dims.nbDims);
@@ -102,7 +113,7 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
       // TODO(Superjomn) change this float to dtype size.
       auto size = inference::analysis::AccuDims(dims.d, dims.nbDims) *
                   FLAGS_tensorrt_engine_batch_size;
-      engine->GetOutputInCPU(y,
+      engine->GetOutputInCPU(output_maps[output_index],
                              fluid_t->mutable_data<float>(platform::CPUPlace()),
                              size * sizeof(float));
       //} else {
@@ -110,6 +121,7 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
       // y, fluid_t->mutable_data<float>(platform::CUDAPlace()),
       // size * sizeof(float));
       //}
+      output_index += 1;
     }
 
     cudaStreamSynchronize(*engine->stream());

paddle/fluid/operators/tensorrt_engine_op_test.cc

@@ -103,6 +103,9 @@ TEST(TensorRTEngineOp, manual) {
   SetAttr<std::string>(engine_op_desc.Proto(), "engine_uniq_key", "a_engine");
   SetAttr<std::vector<std::string>>(engine_op_desc.Proto(), "parameters",
                                     std::vector<std::string>({}));
+  SetAttr<std::vector<std::string>>(engine_op_desc.Proto(),
+                                    "output_name_mapping",
+                                    std::vector<std::string>({"z0"}));
 
   LOG(INFO) << "create engine op";
   auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto());
@@ -196,6 +199,10 @@ void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) {
       std::vector<std::string>({"y0", "y1", "y2", "y3"}));
   SetAttr<std::string>(engine_op_desc.Proto(), "engine_uniq_key", "b_engine");
 
+  SetAttr<std::vector<std::string>>(engine_op_desc.Proto(),
+                                    "output_name_mapping",
+                                    std::vector<std::string>({"z3"}));
+
   auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto());
 
   // Execute them.

paddle/fluid/platform/CMakeLists.txt

@@ -18,7 +18,11 @@ else()
 endif()
 cc_test(enforce_test SRCS enforce_test.cc DEPS stringpiece enforce)
 
-cc_library(cpu_info SRCS cpu_info.cc DEPS gflags glog enforce)
+set(CPU_INFO_DEPS gflags glog enforce)
+IF(WITH_XBYAK)
+    list(APPEND CPU_INFO_DEPS xbyak)
+ENDIF()
+cc_library(cpu_info SRCS cpu_info.cc DEPS ${CPU_INFO_DEPS})
 cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info)
 
 nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog enforce)

paddle/fluid/platform/cpu_info.cc

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/platform/cpu_info.h"
+#include "xbyak/xbyak.h"
+#include "xbyak/xbyak_util.h"
 
 #ifdef __APPLE__
 #include <sys/sysctl.h>
@@ -98,5 +100,39 @@ size_t CUDAPinnedMaxChunkSize() {
   return CUDAPinnedMaxAllocSize() / 256;
 }
 
+#ifdef PADDLE_WITH_XBYAK
+namespace jit {
+
+static Xbyak::util::Cpu cpu;
+bool MayIUse(const cpu_isa_t cpu_isa) {
+  using namespace Xbyak::util;  // NOLINT
+  switch (cpu_isa) {
+    case sse42:
+      return cpu.has(Cpu::tSSE42);
+    case avx2:
+      return cpu.has(Cpu::tAVX2);
+    case avx512_common:
+      return cpu.has(Cpu::tAVX512F);
+    case avx512_core:
+      return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) &&
+             cpu.has(Cpu::tAVX512VL) && cpu.has(Cpu::tAVX512DQ);
+    case avx512_core_vnni:
+      return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) &&
+             cpu.has(Cpu::tAVX512VL) && cpu.has(Cpu::tAVX512DQ) &&
+             cpu.has(Cpu::tAVX512_VNNI);
+    case avx512_mic:
+      return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512CD) &&
+             cpu.has(Cpu::tAVX512ER) && cpu.has(Cpu::tAVX512PF);
+    case avx512_mic_4ops:
+      return true && MayIUse(avx512_mic) && cpu.has(Cpu::tAVX512_4FMAPS) &&
+             cpu.has(Cpu::tAVX512_4VNNIW);
+    case isa_any:
+      return true;
+  }
+  return false;
+}
+
+}  // namespace jit
+#endif
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/cpu_info.h

@@ -37,5 +37,25 @@ size_t CUDAPinnedMinChunkSize();
 //! Get the maximum chunk size for buddy allocator.
 size_t CUDAPinnedMaxChunkSize();
 
+#ifdef PADDLE_WITH_XBYAK
+namespace jit {
+
+typedef enum {
+  isa_any,
+  sse42,
+  avx2,
+  avx512_common,
+  avx512_core,
+  avx512_core_vnni,
+  avx512_mic,
+  avx512_mic_4ops,
+} cpu_isa_t;  // Instruction set architecture
+
+// May I use some instruction
+inline bool MayIUse(const cpu_isa_t cpu_isa);
+
+}  // namespace jit
+#endif
+
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/mkldnn_helper.h

@@ -223,7 +223,7 @@ class MKLDNNHandler {
   static std::string GetHash(mkldnn::memory::dims& operand_dims,  // NOLINT
                              const std::string& suffix) {
     return dims2str(operand_dims) + suffix;
-  };
+  }
 
  protected:
   static std::string dims2str(const mkldnn::memory::dims& operand_dims) {
@@ -251,5 +251,17 @@ inline mkldnn::memory::format MKLDNNFormatForSize(
   return data_format;
 }
 
+inline mkldnn::memory::format data_format_to_memory_format(
+    const std::string& data_format) {
+  switch (framework::StringToDataLayout(data_format)) {
+    case framework::DataLayout::kNHWC:
+      return mkldnn::memory::format::nhwc;
+    case framework::DataLayout::kNCHW:
+      return mkldnn::memory::format::nchw;
+    default:
+      return mkldnn::memory::format::any;
+  }
+}
+
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/profiler.cc

@@ -110,6 +110,8 @@ Event::Event(EventType type, std::string name, uint32_t thread_id,
   has_cuda_ = dev_ctx ? platform::is_gpu_place(dev_ctx->GetPlace()) : false;
   if (has_cuda_) {
     auto* cuda_dev_ctx = static_cast<const CUDADeviceContext*>(dev_ctx);
+    PADDLE_ENFORCE(cudaSetDevice(
+        boost::get<platform::CUDAPlace>(cuda_dev_ctx->GetPlace()).device));
     PADDLE_ENFORCE(cudaGetDevice(&device_));
     PADDLE_ENFORCE(cudaEventCreate(&event_));
     auto stream = cuda_dev_ctx->stream();
@@ -176,6 +178,7 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
 
 RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
     : is_enabled_(false), start_ns_(PosixInNsec()) {
+  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled) return;
   is_enabled_ = true;
   dev_ctx_ = dev_ctx;
@@ -186,6 +189,7 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
 }
 
 RecordEvent::~RecordEvent() {
+  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled || !is_enabled_) return;
   DeviceTracer* tracer = GetDeviceTracer();
   if (tracer) {
@@ -198,6 +202,7 @@ RecordEvent::~RecordEvent() {
 
 RecordBlock::RecordBlock(int block_id)
     : is_enabled_(false), start_ns_(PosixInNsec()) {
+  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled) return;
   is_enabled_ = true;
   SetCurBlock(block_id);
@@ -205,6 +210,7 @@ RecordBlock::RecordBlock(int block_id)
 }
 
 RecordBlock::~RecordBlock() {
+  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled || !is_enabled_) return;
   DeviceTracer* tracer = GetDeviceTracer();
   if (tracer) {

paddle/fluid/pybind/pybind.cc

@@ -394,8 +394,10 @@ All parameter, weight, gradient are variables in Paddle.
     InferenceOptimize(*(origin.Proto()), &pruned_desc);
     return new ProgramDesc(pruned_desc);
   });
-  m.def("empty_var_name", []() { return framework::kEmptyVarName; });
-  m.def("grad_var_suffix", []() { return framework::kGradVarSuffix; });
+  m.def("empty_var_name",
+        []() { return std::string(framework::kEmptyVarName); });
+  m.def("grad_var_suffix",
+        []() { return std::string(framework::kGradVarSuffix); });
   m.def_submodule(
        "var_names",
        "The module will return special predefined variable name in Paddle")

paddle/scripts/paddle_build.sh

@@ -419,6 +419,25 @@ EOF
     linkchecker doc/v2/en/html/index.html
     linkchecker doc/v2/cn/html/index.html
     linkchecker doc/v2/api/en/html/index.html
+
+    if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
+
+    # Deploy to the the content server if its a "develop" or "release/version" branch
+    # The "develop_doc" branch is reserved to test full deploy process without impacting the real content.
+    if [ "$TRAVIS_BRANCH" == "develop_doc" ]; then
+        PPO_SCRIPT_BRANCH=develop
+    elif [[ "$TRAVIS_BRANCH" == "develop"  ||  "$TRAVIS_BRANCH" =~ ^v|release/[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then
+        PPO_SCRIPT_BRANCH=master
+    else
+        # Early exit, this branch doesn't require documentation build
+        return 0;
+    fi
+     # Fetch the paddlepaddle.org deploy_docs.sh from the appopriate branch
+    export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/$PPO_SCRIPT_BRANCH/scripts/deploy/deploy_docs.sh
+    export PYTHONPATH=$PYTHONPATH:${PADDLE_ROOT}/build/python:/paddle/build/python
+    cd ..
+    curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH ${PADDLE_ROOT} ${PADDLE_ROOT}/build/doc/ ${PPO_SCRIPT_BRANCH}
+    cd -
 }
 
 function gen_html() {

paddle/scripts/paddle_docker_build.sh

@@ -52,6 +52,9 @@ EOL
     ${DOCKER_CMD} run -it \
         ${DOCKER_ENV} \
         -e SCRIPT_NAME=$0 \
+        -e CONTENT_DEC_PASSWD=$CONTENT_DEC_PASSWD \
+        -e TRAVIS_BRANCH=$TRAVIS_BRANCH \
+        -e TRAVIS_PULL_REQUEST=$TRAVIS_PULL_REQUEST \
         -v $PADDLE_ROOT:/paddle \
         -v ${HOME}/.ccache:/root/.ccache \
         -w /paddle \

python/paddle/dataset/cifar.py

@@ -28,11 +28,12 @@ images per class.
 
 """
 
-import cPickle
 import itertools
 import numpy
 import paddle.dataset.common
 import tarfile
+from six.moves import zip
+from six.moves import cPickle as pickle
 
 __all__ = ['train100', 'test100', 'train10', 'test10', 'convert']
 
@@ -48,7 +49,7 @@ def reader_creator(filename, sub_name, cycle=False):
         data = batch['data']
         labels = batch.get('labels', batch.get('fine_labels', None))
         assert labels is not None
-        for sample, label in itertools.izip(data, labels):
+        for sample, label in zip(data, labels):
             yield (sample / 255.0).astype(numpy.float32), int(label)
 
     def reader():
@@ -58,7 +59,7 @@ def reader_creator(filename, sub_name, cycle=False):
 
             while True:
                 for name in names:
-                    batch = cPickle.load(f.extractfile(name))
+                    batch = pickle.load(f.extractfile(name))
                     for item in read_batch(batch):
                         yield item
                 if not cycle:

python/paddle/dataset/common.py

@@ -20,9 +20,8 @@ import shutil
 import sys
 import importlib
 import paddle.dataset
-import cPickle
+import six.moves.cPickle as pickle
 import glob
-import cPickle as pickle
 
 __all__ = [
     'DATA_HOME',
@@ -75,13 +74,13 @@ def download(url, module_name, md5sum, save_name=None):
     retry_limit = 3
     while not (os.path.exists(filename) and md5file(filename) == md5sum):
         if os.path.exists(filename):
-            print "file md5", md5file(filename), md5sum
+            print("file md5", md5file(filename), md5sum)
         if retry < retry_limit:
             retry += 1
         else:
             raise RuntimeError("Cannot download {0} within retry limit {1}".
                                format(url, retry_limit))
-        print "Cache file %s not found, downloading %s" % (filename, url)
+        print("Cache file %s not found, downloading %s" % (filename, url))
         r = requests.get(url, stream=True)
         total_length = r.headers.get('content-length')
 
@@ -104,8 +103,9 @@ def download(url, module_name, md5sum, save_name=None):
 
 
 def fetch_all():
-    for module_name in filter(lambda x: not x.startswith("__"),
-                              dir(paddle.dataset)):
+    for module_name in [
+            x for x in dir(paddle.dataset) if not x.startswith("__")
+    ]:
         if "fetch" in dir(
                 importlib.import_module("paddle.dataset.%s" % module_name)):
             getattr(
@@ -114,8 +114,9 @@ def fetch_all():
 
 
 def fetch_all_recordio(path):
-    for module_name in filter(lambda x: not x.startswith("__"),
-                              dir(paddle.dataset)):
+    for module_name in [
+            x for x in dir(paddle.dataset) if not x.startswith("__")
+    ]:
         if "convert" in dir(
                 importlib.import_module("paddle.dataset.%s" % module_name)) and \
                 not module_name == "common":
@@ -126,7 +127,7 @@ def fetch_all_recordio(path):
                 "convert")(ds_path)
 
 
-def split(reader, line_count, suffix="%05d.pickle", dumper=cPickle.dump):
+def split(reader, line_count, suffix="%05d.pickle", dumper=pickle.dump):
     """
     you can call the function as:
 
@@ -167,7 +168,7 @@ def split(reader, line_count, suffix="%05d.pickle", dumper=cPickle.dump):
 def cluster_files_reader(files_pattern,
                          trainer_count,
                          trainer_id,
-                         loader=cPickle.load):
+                         loader=pickle.load):
     """
     Create a reader that yield element from the given files, select
     a file set according trainer count and trainer_id
@@ -188,7 +189,7 @@ def cluster_files_reader(files_pattern,
         my_file_list = []
         for idx, fn in enumerate(file_list):
             if idx % trainer_count == trainer_id:
-                print "append file: %s" % fn
+                print("append file: %s" % fn)
                 my_file_list.append(fn)
         for fn in my_file_list:
             with open(fn, "r") as f:
@@ -221,7 +222,7 @@ def convert(output_path, reader, line_count, name_prefix):
         for l in lines:
             # FIXME(Yancey1989):
             # dumps with protocol: pickle.HIGHEST_PROTOCOL
-            writer.write(cPickle.dumps(l))
+            writer.write(pickle.dumps(l))
         writer.close()
 
     lines = []

python/paddle/dataset/conll05.py

@@ -24,18 +24,19 @@ import tarfile
 import gzip
 import itertools
 import paddle.dataset.common
+from six.moves import zip
 
 __all__ = ['test, get_dict', 'get_embedding', 'convert']
 
 DATA_URL = 'http://www.cs.upc.edu/~srlconll/conll05st-tests.tar.gz'
 DATA_MD5 = '387719152ae52d60422c016e92a742fc'
-WORDDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/wordDict.txt'
+WORDDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FwordDict.txt'
 WORDDICT_MD5 = 'ea7fb7d4c75cc6254716f0177a506baa'
-VERBDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/verbDict.txt'
+VERBDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FverbDict.txt'
 VERBDICT_MD5 = '0d2977293bbb6cbefab5b0f97db1e77c'
-TRGDICT_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/targetDict.txt'
+TRGDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FtargetDict.txt'
 TRGDICT_MD5 = 'd8c7f03ceb5fc2e5a0fa7503a4353751'
-EMB_URL = 'http://paddlepaddle.bj.bcebos.com/demo/srl_dict_and_embedding/emb'
+EMB_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2Femb'
 EMB_MD5 = 'bf436eb0faa1f6f9103017f8be57cdb7'
 
 UNK_IDX = 0
@@ -87,12 +88,12 @@ def corpus_reader(data_path, words_name, props_name):
             sentences = []
             labels = []
             one_seg = []
-            for word, label in itertools.izip(words_file, props_file):
+            for word, label in zip(words_file, props_file):
                 word = word.strip()
                 label = label.strip().split()
 
                 if len(label) == 0:  # end of sentence
-                    for i in xrange(len(one_seg[0])):
+                    for i in range(len(one_seg[0])):
                         a_kind_lable = [x[i] for x in one_seg]
                         labels.append(a_kind_lable)
 

python/paddle/dataset/flowers.py

@@ -28,10 +28,9 @@ Graphics and Image Processing (2008)
 http://www.robots.ox.ac.uk/~vgg/publications/papers/nilsback08.{pdf,ps.gz}.
 
 """
-import cPickle
 import itertools
 import functools
-from common import download
+from .common import download
 import tarfile
 import scipy.io as scio
 from paddle.dataset.image import *
@@ -39,6 +38,8 @@ from paddle.reader import *
 import os
 import numpy as np
 from multiprocessing import cpu_count
+from six.moves import cPickle as pickle
+from six.moves import zip
 __all__ = ['train', 'test', 'valid']
 
 DATA_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/102flowers.tgz'
@@ -116,10 +117,10 @@ def reader_creator(data_file,
                 file = file.strip()
                 batch = None
                 with open(file, 'r') as f:
-                    batch = cPickle.load(f)
+                    batch = pickle.load(f)
                 data = batch['data']
                 labels = batch['label']
-                for sample, label in itertools.izip(data, batch['label']):
+                for sample, label in zip(data, batch['label']):
                     yield sample, int(label) - 1
             if not cycle:
                 break

python/paddle/dataset/image.py

@@ -36,7 +36,7 @@ except ImportError:
     cv2 = None
 import os
 import tarfile
-import cPickle
+import six.moves.cPickle as pickle
 
 __all__ = [
     "load_image_bytes", "load_image", "resize_short", "to_chw", "center_crop",
@@ -86,10 +86,10 @@ def batch_images_from_tar(data_file,
                 output = {}
                 output['label'] = labels
                 output['data'] = data
-                cPickle.dump(
+                pickle.dump(
                     output,
                     open('%s/batch_%d' % (out_path, file_id), 'w'),
-                    protocol=cPickle.HIGHEST_PROTOCOL)
+                    protocol=pickle.HIGHEST_PROTOCOL)
                 file_id += 1
                 data = []
                 labels = []
@@ -97,10 +97,10 @@ def batch_images_from_tar(data_file,
         output = {}
         output['label'] = labels
         output['data'] = data
-        cPickle.dump(
+        pickle.dump(
             output,
             open('%s/batch_%d' % (out_path, file_id), 'w'),
-            protocol=cPickle.HIGHEST_PROTOCOL)
+            protocol=pickle.HIGHEST_PROTOCOL)
 
     with open(meta_file, 'a') as meta:
         for file in os.listdir(out_path):

python/paddle/dataset/imdb.py

@@ -42,13 +42,13 @@ def tokenize(pattern):
         # sequential access of member files, other than
         # tarfile.extractfile, which does random access and might
         # destroy hard disks.
-        tf = tarf.next()
+        tf = next(tarf)
         while tf != None:
             if bool(pattern.match(tf.name)):
                 # newline and punctuations removal and ad-hoc tokenization.
                 yield tarf.extractfile(tf).read().rstrip("\n\r").translate(
                     None, string.punctuation).lower().split()
-            tf = tarf.next()
+            tf = next(tarf)
 
 
 def build_dict(pattern, cutoff):
@@ -62,11 +62,11 @@ def build_dict(pattern, cutoff):
             word_freq[word] += 1
 
     # Not sure if we should prune less-frequent words here.
-    word_freq = filter(lambda x: x[1] > cutoff, word_freq.items())
+    word_freq = [x for x in list(word_freq.items()) if x[1] > cutoff]
 
     dictionary = sorted(word_freq, key=lambda x: (-x[1], x[0]))
     words, _ = list(zip(*dictionary))
-    word_idx = dict(zip(words, xrange(len(words))))
+    word_idx = dict(list(zip(words, list(range(len(words))))))
     word_idx['<unk>'] = len(words)
     return word_idx
 

python/paddle/dataset/imikolov.py

@@ -64,11 +64,11 @@ def build_dict(min_word_freq=50):
             # remove <unk> for now, since we will set it as last index
             del word_freq['<unk>']
 
-        word_freq = filter(lambda x: x[1] > min_word_freq, word_freq.items())
+        word_freq = [x for x in list(word_freq.items()) if x[1] > min_word_freq]
 
         word_freq_sorted = sorted(word_freq, key=lambda x: (-x[1], x[0]))
         words, _ = list(zip(*word_freq_sorted))
-        word_idx = dict(zip(words, xrange(len(words))))
+        word_idx = dict(list(zip(words, list(range(len(words))))))
         word_idx['<unk>'] = len(words)
 
     return word_idx

python/paddle/dataset/mnist.py

@@ -65,7 +65,7 @@ def reader_creator(image_filename, label_filename, buffer_size):
 
                 images = images / 255.0 * 2.0 - 1.0
 
-                for i in xrange(buffer_size):
+                for i in range(buffer_size):
                     yield images[i, :], int(labels[i])
         finally:
             try:

python/paddle/dataset/movielens.py

@@ -187,7 +187,7 @@ def max_movie_id():
     Get the maximum value of movie id.
     """
     __initialize_meta_info__()
-    return reduce(__max_index_info__, MOVIE_INFO.viewvalues()).index
+    return reduce(__max_index_info__, list(MOVIE_INFO.values())).index
 
 
 def max_user_id():
@@ -195,7 +195,7 @@ def max_user_id():
     Get the maximum value of user id.
     """
     __initialize_meta_info__()
-    return reduce(__max_index_info__, USER_INFO.viewvalues()).index
+    return reduce(__max_index_info__, list(USER_INFO.values())).index
 
 
 def __max_job_id_impl__(a, b):
@@ -210,7 +210,7 @@ def max_job_id():
     Get the maximum value of job id.
     """
     __initialize_meta_info__()
-    return reduce(__max_job_id_impl__, USER_INFO.viewvalues()).job_id
+    return reduce(__max_job_id_impl__, list(USER_INFO.values())).job_id
 
 
 def movie_categories():
@@ -243,7 +243,7 @@ def unittest():
     for test_count, _ in enumerate(test()()):
         pass
 
-    print train_count, test_count
+    print(train_count, test_count)
 
 
 def fetch():

python/paddle/dataset/mq2007.py

@@ -26,7 +26,7 @@ http://research.microsoft.com/en-us/um/beijing/projects/letor/LETOR4.0/Data/MQ20
 import os
 import functools
 import rarfile
-from common import download
+from .common import download
 import numpy as np
 
 # URL = "http://research.microsoft.com/en-us/um/beijing/projects/letor/LETOR4.0/Data/MQ2007.rar"
@@ -330,4 +330,4 @@ if __name__ == "__main__":
     mytest = functools.partial(
         __reader__, filepath="MQ2007/MQ2007/Fold1/sample", format="listwise")
     for label, query in mytest():
-        print label, query
+        print(label, query)

python/paddle/dataset/sentiment.py

@@ -43,11 +43,11 @@ def download_data_if_not_yet():
             nltk.data.path.append(paddle.dataset.common.DATA_HOME)
         movie_reviews.categories()
     except LookupError:
-        print "Downloading movie_reviews data set, please wait....."
+        print("Downloading movie_reviews data set, please wait.....")
         nltk.download(
             'movie_reviews', download_dir=paddle.dataset.common.DATA_HOME)
-        print "Download data set success....."
-        print "Path is " + nltk.data.find('corpora/movie_reviews').path
+        print("Download data set success.....")
+        print("Path is " + nltk.data.find('corpora/movie_reviews').path)
 
 
 def get_word_dict():
@@ -64,7 +64,7 @@ def get_word_dict():
         for field in movie_reviews.fileids(category):
             for words in movie_reviews.words(field):
                 word_freq_dict[words] += 1
-    words_sort_list = word_freq_dict.items()
+    words_sort_list = list(word_freq_dict.items())
     words_sort_list.sort(cmp=lambda a, b: b[1] - a[1])
     for index, word in enumerate(words_sort_list):
         words_freq_sorted.append((word[0], index))
@@ -80,7 +80,8 @@ def sort_files():
     files_list = list()
     neg_file_list = movie_reviews.fileids('neg')
     pos_file_list = movie_reviews.fileids('pos')
-    files_list = list(chain.from_iterable(zip(neg_file_list, pos_file_list)))
+    files_list = list(
+        chain.from_iterable(list(zip(neg_file_list, pos_file_list))))
     return files_list
 
 

python/paddle/dataset/tests/common_test.py

@@ -36,7 +36,7 @@ class TestCommon(unittest.TestCase):
     def test_split(self):
         def test_reader():
             def reader():
-                for x in xrange(10):
+                for x in range(10):
                     yield x
 
             return reader
@@ -49,7 +49,7 @@ class TestCommon(unittest.TestCase):
 
     def test_cluster_file_reader(self):
         _, temp_path = tempfile.mkstemp()
-        for x in xrange(5):
+        for x in range(5):
             with open(temp_path + '/%05d.test' % x) as f:
                 f.write('%d\n' % x)
         reader = paddle.dataset.common.cluster_files_reader(
@@ -63,7 +63,7 @@ class TestCommon(unittest.TestCase):
 
         def test_reader():
             def reader():
-                for x in xrange(record_num):
+                for x in range(record_num):
                     yield x
 
             return reader

python/paddle/dataset/tests/imikolov_test.py

@@ -59,7 +59,7 @@ class TestMikolov(unittest.TestCase):
         self.assertEqual(first_line, read_line)
 
     def test_total(self):
-        _, idx = zip(*WORD_DICT.items())
+        _, idx = list(zip(*list(WORD_DICT.items())))
         self.assertEqual(sorted(idx)[-1], len(WORD_DICT) - 1)
 
 

python/paddle/dataset/tests/test_sentiment.py

@@ -24,9 +24,8 @@ from nltk.corpus import movie_reviews
 class TestSentimentMethods(unittest.TestCase):
     def test_get_word_dict(self):
         word_dict = st.get_word_dict()[0:10]
-        test_word_list = [(u',', 0), (u'the', 1), (u'.', 2), (u'a', 3),
-                          (u'and', 4), (u'of', 5), (u'to', 6), (u"'", 7),
-                          (u'is', 8), (u'in', 9)]
+        test_word_list = [(',', 0), ('the', 1), ('.', 2), ('a', 3), ('and', 4),
+                          ('of', 5), ('to', 6), ("'", 7), ('is', 8), ('in', 9)]
         for idx, each in enumerate(word_dict):
             self.assertEqual(each, test_word_list[idx])
         self.assertTrue("/root/.cache/paddle/dataset" in nltk.data.path)

python/paddle/dataset/uci_housing.py

@@ -49,9 +49,12 @@ def feature_range(maximums, minimums):
     import matplotlib.pyplot as plt
     fig, ax = plt.subplots()
     feature_num = len(maximums)
-    ax.bar(range(feature_num), maximums - minimums, color='r', align='center')
+    ax.bar(list(range(feature_num)),
+           maximums - minimums,
+           color='r',
+           align='center')
     ax.set_title('feature scale')
-    plt.xticks(range(feature_num), feature_names)
+    plt.xticks(list(range(feature_num)), feature_names)
     plt.xlim([-1, feature_num])
     fig.set_figheight(6)
     fig.set_figwidth(10)
@@ -71,7 +74,7 @@ def load_data(filename, feature_num=14, ratio=0.8):
     maximums, minimums, avgs = data.max(axis=0), data.min(axis=0), data.sum(
         axis=0) / data.shape[0]
     feature_range(maximums[:-1], minimums[:-1])
-    for i in xrange(feature_num - 1):
+    for i in range(feature_num - 1):
         data[:, i] = (data[:, i] - avgs[i]) / (maximums[i] - minimums[i])
     offset = int(data.shape[0] * ratio)
     UCI_TRAIN_DATA = data[:offset]

python/paddle/dataset/wmt14.py

@@ -40,7 +40,7 @@ URL_TRAIN = ('http://paddlepaddle.cdn.bcebos.com/demo/'
              'wmt_shrinked_data/wmt14.tgz')
 MD5_TRAIN = '0791583d57d5beb693b9414c5b36798c'
 # BLEU of this trained model is 26.92
-URL_MODEL = 'http://paddlepaddle.bj.bcebos.com/demo/wmt_14/wmt14_model.tar.gz'
+URL_MODEL = 'http://paddlemodels.bj.bcebos.com/wmt%2Fwmt14.tgz'
 MD5_MODEL = '0cb4a5366189b6acba876491c8724fa3'
 
 START = "<s>"
@@ -154,8 +154,8 @@ def get_dict(dict_size, reverse=True):
     tar_file = paddle.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN)
     src_dict, trg_dict = __read_to_dict(tar_file, dict_size)
     if reverse:
-        src_dict = {v: k for k, v in src_dict.items()}
-        trg_dict = {v: k for k, v in trg_dict.items()}
+        src_dict = {v: k for k, v in list(src_dict.items())}
+        trg_dict = {v: k for k, v in list(trg_dict.items())}
     return src_dict, trg_dict
 
 

python/paddle/dataset/wmt16.py

@@ -70,7 +70,9 @@ def __build_dict(tar_file, dict_size, save_path, lang):
         fout.write("%s\n%s\n%s\n" % (START_MARK, END_MARK, UNK_MARK))
         for idx, word in enumerate(
                 sorted(
-                    word_dict.iteritems(), key=lambda x: x[1], reverse=True)):
+                    iter(list(word_dict.items())),
+                    key=lambda x: x[1],
+                    reverse=True)):
             if idx + 3 == dict_size: break
             fout.write("%s\n" % (word[0]))
 

python/paddle/fluid/__init__.py

@@ -14,49 +14,49 @@
 
 from __future__ import print_function
 # import all class inside framework into fluid module
-import framework
-from framework import *
+from . import framework
+from .framework import *
 # import all class inside executor into fluid module
-import executor
-from executor import *
-
-import trainer
-from trainer import Trainer
-from trainer import BeginEpochEvent
-from trainer import EndEpochEvent
-from trainer import BeginStepEvent
-from trainer import EndStepEvent
-from trainer import CheckpointConfig
-
-import inferencer
-from inferencer import Inferencer
-
-import io
-import evaluator
-import initializer
-import layers
-import contrib
-import nets
-import optimizer
-import backward
-import regularizer
-import average
-import metrics
-import transpiler
-from param_attr import ParamAttr, WeightNormParamAttr
-from data_feeder import DataFeeder
-from core import LoDTensor, LoDTensorArray, CPUPlace, CUDAPlace, CUDAPinnedPlace, Scope
-from transpiler import DistributeTranspiler, InferenceTranspiler, \
+from . import executor
+from .executor import *
+
+from . import trainer
+from .trainer import Trainer
+from .trainer import BeginEpochEvent
+from .trainer import EndEpochEvent
+from .trainer import BeginStepEvent
+from .trainer import EndStepEvent
+from .trainer import CheckpointConfig
+
+from . import inferencer
+from .inferencer import Inferencer
+
+from . import io
+from . import evaluator
+from . import initializer
+from . import layers
+from . import contrib
+from . import nets
+from . import optimizer
+from . import backward
+from . import regularizer
+from . import average
+from . import metrics
+from . import transpiler
+from .param_attr import ParamAttr, WeightNormParamAttr
+from .data_feeder import DataFeeder
+from .core import LoDTensor, LoDTensorArray, CPUPlace, CUDAPlace, CUDAPinnedPlace, Scope
+from .transpiler import DistributeTranspiler, InferenceTranspiler, \
     memory_optimize, release_memory, DistributeTranspilerConfig
-from concurrency import (Go, make_channel, channel_send, channel_recv,
+from .concurrency import (Go, make_channel, channel_send, channel_recv,
                           channel_close, Select)
-from lod_tensor import create_lod_tensor, create_random_int_lodtensor
-import clip
-import profiler
-import unique_name
-import recordio_writer
-import parallel_executor
-from parallel_executor import *
+from .lod_tensor import create_lod_tensor, create_random_int_lodtensor
+from . import clip
+from . import profiler
+from . import unique_name
+from . import recordio_writer
+from . import parallel_executor
+from .parallel_executor import *
 from paddle.fluid.layers.math_op_patch import monkey_patch_variable
 
 Tensor = LoDTensor
@@ -99,8 +99,8 @@ def __bootstrap__():
         None
     """
     import sys
-    import core
     import os
+    from . import core
 
     in_test = 'unittest' in sys.modules
 
@@ -123,7 +123,8 @@ def __bootstrap__():
     read_env_flags = [
         'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir',
         'eager_delete_scope', 'use_mkldnn', 'initial_cpu_memory_in_mb',
-        'init_allocated_mem', 'free_idle_memory', 'paddle_num_threads'
+        'init_allocated_mem', 'free_idle_memory', 'paddle_num_threads',
+        'cpu_deterministic'
     ]
     if core.is_compiled_with_dist():
         read_env_flags.append('rpc_deadline')

python/paddle/fluid/annotations.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from __future__ import print_function
 import functools
 import sys
 
@@ -28,7 +29,7 @@ def deprecated(since, instead, extra_message=""):
 
         @functools.wraps(func)
         def wrapper(*args, **kwargs):
-            print >> sys.stderr, err_msg
+            print(err_msg, file=sys.stderr)
             return func(*args, **kwargs)
 
         wrapper.__doc__ += "\n    "

python/paddle/fluid/backward.py

@@ -16,7 +16,8 @@ from paddle.fluid import framework as framework
 from . import core
 import collections
 import copy
-import unique_name
+import six
+from . import unique_name
 
 __all__ = ['append_backward']
 
@@ -44,17 +45,25 @@ def _create_op_desc_(op_type, inputs, outputs, attrs):
     """
     op_desc = core.OpDesc()
     op_desc.set_type(op_type)
-    for para, args in inputs.iteritems():
-        op_desc.set_input(para, args)
-    for para, args in outputs.iteritems():
-        op_desc.set_output(para, args)
+    for para, args in list(inputs.items()):
+        op_desc.set_input(
+            para,
+            list(
+                map(lambda arg: arg.decode() if isinstance(arg, six.binary_type) else arg,
+                    args)))
+    for para, args in list(outputs.items()):
+        op_desc.set_output(
+            para,
+            list(
+                map(lambda arg: arg.decode() if isinstance(arg, six.binary_type) else arg,
+                    args)))
 
     op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
 
     if op_role_attr_name not in attrs:
         attrs[
             op_role_attr_name] = core.op_proto_and_checker_maker.OpRole.Backward
-    for name, val in attrs.iteritems():
+    for name, val in list(attrs.items()):
         if isinstance(val, framework.Block):
             op_desc.set_block_attr(name, val.desc)
         else:
@@ -105,7 +114,9 @@ def _strip_grad_suffix_(name):
     e.g. x@GRAD ==> x
          y@GRAD@RENAME@1 ==> y
     """
-    pos = name.find(core.grad_var_suffix())
+    if isinstance(name, six.text_type):
+        name = name.encode()
+    pos = name.find(six.b(core.grad_var_suffix()))
     return name[:pos] if pos != -1 else name
 
 
@@ -114,7 +125,9 @@ def _append_grad_suffix_(name):
     Append grad suffix to the given variable name
     e.g. x ==> x@GRAD
     """
-    return name + core.grad_var_suffix()
+    if isinstance(name, six.text_type):
+        name = name.encode()
+    return name + six.b(core.grad_var_suffix())
 
 
 def _addup_repetitive_outputs_(op_descs):
@@ -174,7 +187,7 @@ def _addup_repetitive_outputs_(op_descs):
                     op_desc.set_output(param_name, arg_names)
                     renamed_vars[var_name].append(new_name)
 
-    for var_name, inputs in renamed_vars.iteritems():
+    for var_name, inputs in list(renamed_vars.items()):
         if len(inputs) > 1:
             pending_sum_ops.append(
                 (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]},
@@ -198,16 +211,19 @@ def _remove_no_grad_branch_(op_descs, no_grad_set):
         out_arg_names = op_desc.output_arg_names()
         if len(out_arg_names) == 0 or _all_in_set_(out_arg_names, no_grad_set):
             return True
-        if _all_in_set_(
-                filter(lambda name: name.find(core.grad_var_suffix()) != -1,
-                       op_desc.input_arg_names()), no_grad_set):
+        if _all_in_set_([
+                name for name in op_desc.input_arg_names()
+                if name.find(core.grad_var_suffix()) != -1
+        ], no_grad_set):
             no_grad_set.update(out_arg_names)
             return True
         return False
 
     # Remove ops whose outputs are all in no_grad_dict
-    op_descs = filter(
-        lambda op_desc: not _op_can_be_removed_(op_desc, no_grad_set), op_descs)
+    op_descs = [
+        op_desc for op_desc in op_descs
+        if not _op_can_be_removed_(op_desc, no_grad_set)
+    ]
     # Insert fill_zeros_like_op
     to_insert = []
     for idx, op_desc in enumerate(op_descs):
@@ -217,12 +233,12 @@ def _remove_no_grad_branch_(op_descs, no_grad_set):
                     "X": [_strip_grad_suffix_(arg)]
                 }, {"Out": [arg]}, {}), idx))
 
-    map(lambda p: op_descs.insert(p[1], p[0]), reversed(to_insert))
+    list([op_descs.insert(p[1], p[0]) for p in reversed(to_insert)])
 
     return op_descs
 
 
-import proto.framework_pb2 as framework_pb2
+from .proto import framework_pb2
 
 
 def serialize_op_decs(op_desc):
@@ -244,8 +260,10 @@ def _callback_lookup_(op):
     if op.type == 'parallel_do' and op.attr('use_nccl'):
         all_vars = op.block.vars
         param_names = set(op.input('parameters'))
-        param_names = filter(lambda name: all_vars[name].stop_gradient is False,
-                             param_names)
+        param_names = [
+            name for name in param_names
+            if all_vars[name].stop_gradient is False
+        ]
         param_grad_names = [n + "@GRAD" for n in param_names]
 
         class ParallelDoCallBack(object):
@@ -399,7 +417,7 @@ def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
                 continue
             block.desc.var(grad_var_name)
             new_vars.add(grad_var_name)
-            if not grad_to_var.has_key(grad_var_name):
+            if grad_var_name not in grad_to_var:
                 continue
             grad_info_map[grad_to_var[grad_var_name]] = (grad_var_name, block)
         # infer_shape and infer_type
@@ -427,7 +445,7 @@ def _rename_grad_(block, start_op_idx, grad_to_var, target_grad_map):
                 op_desc.rename_output(name, new_name)
                 var_map[name] = new_name
 
-    for g, ng in var_map.iteritems():
+    for g, ng in list(var_map.items()):
         if g in grad_to_var:
             grad_to_var[ng] = grad_to_var[g]
             grad_to_var.pop(g)
@@ -439,7 +457,7 @@ def _get_stop_gradients_(program):
     for block in program.blocks:
         assert isinstance(block, framework.Block)
         block_no_grad_set = set()
-        for var in block.vars.itervalues():
+        for var in list(block.vars.values()):
             assert isinstance(var, framework.Variable)
             if var.stop_gradient:
                 block_no_grad_set.add(_append_grad_suffix_(var.name))
@@ -535,7 +553,7 @@ def append_backward(loss, parameter_list=None, no_grad_set=None,
         no_grad_set = set()
     no_grad_set = copy.copy(no_grad_set)
     no_grad_dict = _get_stop_gradients_(program)
-    no_grad_dict[0].update(map(_append_grad_suffix_, no_grad_set))
+    no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set)))
 
     grad_info_map = dict()
     root_block = program.block(0)
@@ -558,7 +576,7 @@ def append_backward(loss, parameter_list=None, no_grad_set=None,
 
     block_no_grad_set = set(map(_strip_grad_suffix_, no_grad_dict[0]))
     op_path = _find_op_path_(root_block, [loss], [], block_no_grad_set)
-    no_grad_dict[0].update(map(_append_grad_suffix_, block_no_grad_set))
+    no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set)))
 
     _append_backward_ops_(root_block, op_path, root_block, no_grad_dict,
                           grad_to_var, callbacks)
@@ -572,8 +590,6 @@ def append_backward(loss, parameter_list=None, no_grad_set=None,
 
     program.current_block_idx = current_block_idx
     program._sync_with_cpp()
-    # FIXME(zcd): prevent loss.grad optimized by mem_opt.
-    loss.block.var(_append_grad_suffix_(loss.name)).persistable = True
 
     if parameter_list is not None:
         parameters = parameter_list
@@ -699,7 +715,7 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
         no_grad_set = set()
     no_grad_set = copy.copy(no_grad_set)
     no_grad_dict = _get_stop_gradients_(prog)
-    no_grad_dict[0].update(map(_append_grad_suffix_, no_grad_set))
+    no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set)))
 
     fwd_op_num = block.desc.op_size()
 
@@ -733,7 +749,7 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
 
     block_no_grad_set = set(map(_strip_grad_suffix_, no_grad_dict[0]))
     op_path = _find_op_path_(block, targets, inputs, block_no_grad_set)
-    no_grad_dict[0].update(map(_append_grad_suffix_, block_no_grad_set))
+    no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set)))
     grad_to_var = dict()
     grad_info_map = dict()
     _append_backward_ops_(block, op_path, block, no_grad_dict, grad_to_var)

python/paddle/fluid/clip.py

@@ -13,10 +13,11 @@
 # limitations under the License.
 
 import copy
+import six
 
 import functools
-import layers
-import framework
+from . import layers
+from . import framework
 from . import core
 
 __all__ = [
@@ -80,8 +81,7 @@ def error_clip_callback(block, context):
     # the context is a grad_to_var map
     grad_to_var = context
     op_desc = block.desc.op(block.desc.op_size() - 1)
-    for grad_n in filter(lambda n: grad_to_var.has_key(n),
-                         op_desc.output_arg_names()):
+    for grad_n in [n for n in op_desc.output_arg_names() if n in grad_to_var]:
         fwd_var = block._var_recursive(grad_to_var[grad_n])
         error_clip = getattr(fwd_var, "error_clip", None)
         if not (error_clip is None or isinstance(error_clip,
@@ -247,8 +247,8 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
     """
 
     def __init__(self, clip_norm, group_name="default_group"):
-        if not isinstance(group_name, basestring):
-            raise TypeError("'group_name' must be a basestring.")
+        if not isinstance(group_name, six.string_types):
+            raise TypeError("'group_name' must be a %s." % (six.string_types))
 
         self.clip_norm = clip_norm
         self.group_name = group_name
@@ -284,7 +284,7 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
                 x=clip_var,
                 y=layers.elementwise_max(
                     x=clip_var, y=group_norm_var))
-            assert group_scale_var.shape == (1L, )
+            assert group_scale_var.shape == (1, )
             self.context[group_scale_name] = group_scale_var
 
         new_grad = layers.elementwise_mul(
@@ -313,7 +313,7 @@ def set_gradient_clip(clip, param_list=None, program=None):
         program = framework.default_main_program()
     if param_list is None:
         param_list = program.block(0).all_parameters()
-    if all(isinstance(elem, basestring) for elem in param_list):
+    if all(isinstance(elem, six.string_types) for elem in param_list):
         param_list = [program.block(0).var(elem) for elem in param_list]
     if not all(isinstance(elem, framework.Parameter) for elem in param_list):
         raise TypeError(

python/paddle/fluid/concurrency.py

@@ -12,11 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from layers.control_flow import BlockGuard, equal
+from .layers.control_flow import BlockGuard, equal
 from .framework import Operator
-from layer_helper import LayerHelper, unique_name
-from layers import fill_constant
-import core
+from .layer_helper import LayerHelper, unique_name
+from .layers import fill_constant
+from . import core
 
 __all__ = [
     'Go', 'make_channel', 'channel_send', 'channel_recv', 'channel_close',

python/paddle/fluid/contrib/__init__.py

@@ -12,7 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import decoder
-from decoder import *
+from . import decoder
+from .decoder import *
+from . import memory_usage_calc
+from .memory_usage_calc import *
 
-__all__ = decoder.__all__
+__all__ = decoder.__all__ + memory_usage_calc.__all__

python/paddle/fluid/contrib/decoder/__init__.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import beam_search_decoder
-from beam_search_decoder import *
+from . import beam_search_decoder
+from .beam_search_decoder import *
 
 __all__ = beam_search_decoder.__all__

python/paddle/fluid/contrib/decoder/beam_search_decoder.py

@@ -22,6 +22,7 @@ This API is still under active development and may change drastically.
 
 import contextlib
 import numpy as np
+import six
 
 from ... import layers
 from ...framework import Variable
@@ -191,7 +192,7 @@ class StateCell(object):
         self._helper = LayerHelper('state_cell', name=name)
         self._cur_states = {}
         self._state_names = []
-        for state_name, state in states.items():
+        for state_name, state in six.iteritems(states):
             if not isinstance(state, InitState):
                 raise ValueError('state must be an InitState object.')
             self._cur_states[state_name] = state
@@ -346,7 +347,7 @@ class StateCell(object):
         if self._in_decoder and not self._switched_decoder:
             self._switch_decoder()
 
-        for input_name, input_value in inputs.items():
+        for input_name, input_value in six.iteritems(inputs):
             if input_name not in self._inputs:
                 raise ValueError('Unknown input %s. '
                                  'Please make sure %s in input '
@@ -361,7 +362,7 @@ class StateCell(object):
         if self._in_decoder and not self._switched_decoder:
             self._switched_decoder()
 
-        for state_name, decoder_state in self._states_holder.items():
+        for state_name, decoder_state in six.iteritems(self._states_holder):
             if id(self._cur_decoder_obj) not in decoder_state:
                 raise ValueError('Unknown decoder object, please make sure '
                                  'switch_decoder been invoked.')
@@ -671,7 +672,7 @@ class BeamSearchDecoder(object):
             feed_dict = {}
             update_dict = {}
 
-            for init_var_name, init_var in self._input_var_dict.items():
+            for init_var_name, init_var in six.iteritems(self._input_var_dict):
                 if init_var_name not in self.state_cell._inputs:
                     raise ValueError('Variable ' + init_var_name +
                                      ' not found in StateCell!\n')
@@ -721,7 +722,8 @@ class BeamSearchDecoder(object):
                     self.state_cell.update_states()
                     self.update_array(prev_ids, selected_ids)
                     self.update_array(prev_scores, selected_scores)
-                    for update_name, var_to_update in update_dict.items():
+                    for update_name, var_to_update in six.iteritems(
+                            update_dict):
                         self.update_array(var_to_update, feed_dict[update_name])
 
     def read_array(self, init, is_ids=False, is_scores=False):

python/paddle/fluid/contrib/memory_usage_calc.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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.
+"""
+This module privides a memory usage calculate function for user.
+The purpose of this API is to allow users to estimate memory usage of
+a program under a special batch size, then user can set appropriate 
+batch size to fully utilize a GPU. 
+
+This API is still under active development and may change drastically.
+"""
+
+from .. import core
+from ..framework import Program, Variable
+
+__all__ = ['memory_usage']
+
+dtype_to_size = {
+    core.VarDesc.VarType.FP16: 2,
+    core.VarDesc.VarType.FP32: 4,
+    core.VarDesc.VarType.FP64: 8,
+    core.VarDesc.VarType.INT16: 2,
+    core.VarDesc.VarType.INT32: 4,
+    core.VarDesc.VarType.INT64: 8,
+    core.VarDesc.VarType.BOOL: 1,
+    core.VarDesc.VarType.UINT8: 1,
+}
+
+DEBUG = False
+
+
+def memory_usage(program, batch_size):
+    """
+    Get the estimate memory usage of program with input batch size.
+
+    Args:
+        program(Program): The current Program.
+        batch_size(int): The current input data batch_size.  
+    
+    Returns:
+        min_total_memory(float): the estimate memory usage lower bound.
+        max_total_memory(float): the estimate memory usage upper bound.
+        unit_str(string): the unit of estimate usage result.
+    
+    Examples:
+        
+        >>> import paddle.fluid as fluid
+        >>> lower_usage, upper_usage, unit = fluid.contrib.memory_usage(
+                fluid.default_main_program(), batch_size=10)
+        >>> print "memory usage is about %.3f - %.3f %s" % \
+                (lower_usage, upper_usage, unit)
+
+    """
+
+    # Parameters check
+    if not isinstance(program, Program):
+        raise TypeError(
+            "Calculating Memory Usage requires Program as its Parameter."
+            "But you passed in %s" % (type(prgram)))
+    if batch_size <= 0:
+        raise ValueError("The batch size need to be positive.")
+
+    # Get the var_name list of first block and calculate
+    total_memory = 0.0
+    for var in program.global_block().vars.itervalues():
+        data_count = 1
+        for x in var.shape:
+            if x == -1:
+                data_count *= batch_size
+            else:
+                data_count *= x
+        var_memory = data_count * dtype_to_size[var.dtype]
+        if DEBUG:
+            print "%s memory usage: %d" % (var.name, var_memory)
+        total_memory += var_memory
+    if DEBUG:
+        print "total memory usage: %.2f" % (total_memory)
+
+    # Convert appropriate unit
+    unit_str = "B"
+    if total_memory > 1024:
+        total_memory /= 1024
+        unit_str = "KB"
+        if total_memory > 1024:
+            total_memory /= 1024
+            unit_str = "MB"
+
+    # Append extra memory consumption (5% - 10%)
+    min_total_memory = total_memory * 1.05
+    max_total_memory = total_memory * 1.1
+
+    return min_total_memory, max_total_memory, unit_str

python/paddle/fluid/data_feeder.py

@@ -12,14 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from __future__ import print_function
-import core
+from . import core
 import numpy
 import os
-import six.moves as six
+import six
+from six.moves import zip, range, xrange
 import multiprocessing
 
-from framework import Variable, default_main_program
+from .framework import Variable, default_main_program
 
 __all__ = ['DataFeeder']
 
@@ -53,7 +53,7 @@ class DataToLoDTensorConverter(object):
         self.data = []
         self.lod = []
 
-        for i in six.range(lod_level):
+        for i in six.moves.range(lod_level):
             self.lod.append([])
 
     def feed(self, data):
@@ -142,7 +142,7 @@ class DataFeeder(object):
         if program is None:
             program = default_main_program()
         for each_var in feed_list:
-            if isinstance(each_var, basestring):
+            if isinstance(each_var, six.string_types):
                 each_var = program.block(0).var(each_var)
             if not isinstance(each_var, Variable):
                 raise TypeError("Feed list should contain a list of variable")
@@ -174,7 +174,7 @@ class DataFeeder(object):
             dict: the result of conversion.
         """
         converter = []
-        for lod_level, shape, dtype in six.zip(
+        for lod_level, shape, dtype in six.moves.zip(
                 self.feed_lod_level, self.feed_shapes, self.feed_dtypes):
             converter.append(
                 DataToLoDTensorConverter(
@@ -187,10 +187,12 @@ class DataFeeder(object):
             assert len(each_sample) == len(converter), (
                 "The number of fields in data (%s) does not match " +
                 "len(feed_list) (%s)") % (len(each_sample), len(converter))
-            for each_converter, each_slot in six.zip(converter, each_sample):
+            for each_converter, each_slot in six.moves.zip(converter,
+                                                           each_sample):
                 each_converter.feed(each_slot)
         ret_dict = {}
-        for each_name, each_converter in six.zip(self.feed_names, converter):
+        for each_name, each_converter in six.moves.zip(self.feed_names,
+                                                       converter):
             ret_dict[each_name] = each_converter.done()
         return ret_dict
 
@@ -212,12 +214,14 @@ class DataFeeder(object):
         if isinstance(self.place, core.CUDAPlace):
             places = [
                 core.CUDAPlace(i)
-                for i in six.xrange(self._get_number_of_places_(num_places))
+                for i in six.moves.xrange(
+                    self._get_number_of_places_(num_places))
             ]
         else:
             places = [
                 core.CPUPlace()
-                for _ in six.xrange(self._get_number_of_places_(num_places))
+                for _ in six.moves.xrange(
+                    self._get_number_of_places_(num_places))
             ]
 
         if len(iterable) != len(places):
@@ -227,7 +231,7 @@ class DataFeeder(object):
                              "must be same.")
 
         place = self.place
-        for p, batch in six.zip(places, iterable):
+        for p, batch in six.moves.zip(places, iterable):
             self.place = p
             yield self.feed(batch)
         self.place = place

python/paddle/fluid/debugger.py

@@ -14,8 +14,8 @@
 
 import sys
 import re
-from graphviz import GraphPreviewGenerator
-import proto.framework_pb2 as framework_pb2
+from .graphviz import GraphPreviewGenerator
+from .proto import framework_pb2
 from google.protobuf import text_format
 
 _vartype2str_ = [

python/paddle/fluid/evaluator.py

@@ -15,11 +15,11 @@
 import warnings
 import numpy as np
 
-import layers
-from framework import Program, Variable, program_guard
-import unique_name
-from layer_helper import LayerHelper
-from initializer import Constant
+from . import layers
+from .framework import Program, Variable, program_guard
+from . import unique_name
+from .layer_helper import LayerHelper
+from .initializer import Constant
 
 __all__ = [
     'ChunkEvaluator',

python/paddle/fluid/executor.py

@@ -14,7 +14,8 @@
 
 import numpy as np
 import contextlib
-from framework import Program, default_main_program, Variable
+import six
+from .framework import Program, default_main_program, Variable
 from . import core
 
 __all__ = [
@@ -204,19 +205,19 @@ def fetch_var(name, scope=None, return_numpy=True):
 
 
 def _get_program_cache_key(feed, fetch_list):
-    feed_var_names = feed.keys()
+    feed_var_names = list(feed.keys())
 
     def to_name_str(var):
         if isinstance(var, Variable):
             return var.desc.name()
         elif isinstance(var, str):
             return var
-        elif isinstance(var, basestring):
+        elif isinstance(var, six.string_types):
             return str(var)
         else:
             raise TypeError(str(var) + " should be Variable or str")
 
-    fetch_var_names = map(to_name_str, fetch_list)
+    fetch_var_names = list(map(to_name_str, fetch_list))
 
     return str(feed_var_names + fetch_var_names)
 
@@ -345,7 +346,7 @@ class Executor(object):
     def _fetch_data(self, fetch_list, fetch_var_name, scope):
         outs = [
             core.get_fetch_variable(scope, fetch_var_name, i)
-            for i in xrange(len(fetch_list))
+            for i in range(len(fetch_list))
         ]
         return outs
 

python/paddle/fluid/framework.py

@@ -15,21 +15,22 @@
 import collections
 import contextlib
 import re
+import six
 
 import numpy as np
 
-import proto.framework_pb2 as framework_pb2
+from .proto import framework_pb2
 try:
     from . import core
-except ImportError, e:
+except ImportError as e:
     raise ImportError(
         """NOTE: You may need to run \"export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH\"
     if you encounters \"libmkldnn.so not found\" errors. If you have python
     installed in other directory, replace \"/usr/local/lib\" with your own
     directory. The original error is: \n""" + e.message)
-except Exception, e:
+except Exception as e:
     raise e
-import unique_name
+from . import unique_name
 
 __all__ = [
     'Program',
@@ -86,7 +87,7 @@ def convert_np_dtype_to_dtype_(np_dtype):
     elif dtype == np.uint8:
         return core.VarDesc.VarType.UINT8
     else:
-        raise ValueError("Not supported numpy dtype " + str(dtype))
+        raise ValueError("Not supported numpy dtype " + six.binary_type(dtype))
 
 
 def dtype_is_floating(dtype):
@@ -197,6 +198,7 @@ class Variable(object):
         if name is None:
             name = unique_name.generate('_generated_var')
         is_new_var = False
+        name = name if isinstance(name, six.binary_type) else name.encode()
         self.desc = self.block.desc.find_var(name)
 
         if self.desc is None:
@@ -290,13 +292,13 @@ class Variable(object):
         assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                                bool)
         protostr = self.desc.serialize_to_string()
-        proto = framework_pb2.VarDesc.FromString(str(protostr))
+        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
         res_str = _debug_string_(proto, throw_on_error)
         if with_details:
             additional_attr = ("error_clip", "stop_gradient")
             for attr_name in additional_attr:
-                res_str += "%s: %s\n" % (attr_name,
-                                         str(getattr(self, attr_name)))
+                res_str += "%s: %s\n" % (
+                    attr_name, six.binary_type(getattr(self, attr_name)))
         return res_str
 
     __repr__ = __str__
@@ -369,7 +371,7 @@ def get_all_op_protos():
     protostrs = core.get_all_op_protos()
     ret_values = []
     for pbstr in protostrs:
-        op_proto = framework_pb2.OpProto.FromString(str(pbstr))
+        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
         ret_values.append(op_proto)
     return ret_values
 
@@ -472,7 +474,6 @@ class Operator(object):
                  inputs=None,
                  outputs=None,
                  attrs=None):
-
         self.block = block
         self.desc = desc
         self.attrs = attrs
@@ -523,10 +524,19 @@ class Operator(object):
                             % (in_proto.name, len(in_args)))
                     in_arg_names = []
                     for arg in in_args:
-                        if isinstance(arg, basestring):
+                        if isinstance(arg, six.string_types):
                             in_arg_names.append(arg)
+                        elif isinstance(arg, six.binary_type):
+                            in_arg_names.append(arg.decode())
                         else:
+                            if isinstance(arg.name, six.string_types):
                                 in_arg_names.append(arg.name)
+                            elif isinstance(arg.name, six.binary_type):
+                                in_arg_names.append(arg.name.decode())
+                            else:
+                                raise TypeError(
+                                    "arguments require unicode, str or bytes, but get %s instead."
+                                    % (type(arg.name)))
                     self.desc.set_input(in_proto.name, in_arg_names)
                 else:
                     self.desc.set_input(in_proto.name, [])
@@ -541,8 +551,9 @@ class Operator(object):
             if not given == need:
                 raise ValueError(("Incorrect setting for output(s) of "
                                   "operator \"%s\". Need: [%s] Given: [%s]") %
-                                 (type, ", ".join(str(e) for e in need),
-                                  ", ".join(str(e) for e in given)))
+                                 (type,
+                                  ", ".join(six.binary_type(e) for e in need),
+                                  ", ".join(six.binary_type(e) for e in given)))
 
             for out_proto in proto.outputs:
                 out_args = outputs[out_proto.name]
@@ -554,7 +565,14 @@ class Operator(object):
                         (out_proto.name, len(out_args)))
                 out_arg_names = []
                 for arg in out_args:
+                    if isinstance(arg.name, six.string_types):
                         out_arg_names.append(arg.name)
+                    elif isinstance(arg.name, six.binary_type):
+                        out_arg_names.append(arg.name.decode())
+                    else:
+                        raise TypeError(
+                            "arguments require unicode, str or bytes, but get %s instead."
+                            % (type(arg.name)))
                     arg.op = self
                 self.desc.set_output(out_proto.name, out_arg_names)
 
@@ -590,7 +608,7 @@ class Operator(object):
 
         """
         protostr = self.desc.serialize_to_string()
-        proto = framework_pb2.OpDesc.FromString(str(protostr))
+        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
         return _debug_string_(proto, throw_on_error)
 
     def __str__(self):
@@ -845,7 +863,7 @@ class Block(object):
             re_add_indent = re.compile(r"\n(.)")
             res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                 self.idx, self.parent_idx)
-            for var in self.vars.itervalues():
+            for var in list(self.vars.values()):
                 res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
                     r"\n    \1", var.to_string(throw_on_error, with_details))
             for op in self.ops:
@@ -854,7 +872,8 @@ class Block(object):
             res_str += "\n}"
         else:
             protostr = self.desc.serialize_to_string()
-            proto = framework_pb2.BlockDesc.FromString(str(protostr))
+            proto = framework_pb2.BlockDesc.FromString(
+                six.binary_type(protostr))
             res_str = _debug_string_(proto, throw_on_error)
         return res_str
 
@@ -898,7 +917,8 @@ class Block(object):
         Returns:
             Variable: the Variable with the giving name.
         """
-        if not isinstance(name, basestring):
+        if not isinstance(name, six.string_types):
+            if not isinstance(name, six.binary_type):
                 raise TypeError(
                     "var require string as parameter, but get %s instead." %
                     (type(name)))
@@ -949,10 +969,10 @@ class Block(object):
         raise ValueError("Var {0} is not found recursively".format(name))
 
     def all_parameters(self):
-        return list(self._iter_parameters())
+        return list(self.iter_parameters())
 
-    def _iter_parameters(self):
-        return (item[1] for item in self.vars.iteritems()
+    def iter_parameters(self):
+        return (item[1] for item in list(self.vars.items())
                 if isinstance(item[1], Parameter))
 
     def create_var(self, *args, **kwargs):
@@ -1038,7 +1058,26 @@ class Block(object):
         global_block = self.program.global_block()
         param = Parameter(global_block, *args, **kwargs)
         if 'initializer' in kwargs:
-            kwargs['initializer'](param, self)
+
+            def _is_inited_by(block, var):
+                init_ops = []
+                for op in block.ops:
+                    if var.name in op.output_arg_names:
+                        init_ops.append(op)
+                return init_ops
+
+            initializer = kwargs['initializer']
+            init_ops = _is_inited_by(global_block, param)
+            init_ops_len = len(init_ops)
+            if init_ops_len > 1:
+                raise RuntimeError("param " + param.name +
+                                   " is inited by multiple init ops " + str(
+                                       init_ops))
+            elif init_ops_len == 1:
+                #TODO already inited, do nothing, should log a warning
+                pass
+            else:
+                initializer(param, self)
         return param
 
     def append_op(self, *args, **kwargs):
@@ -1113,7 +1152,7 @@ class Block(object):
                 self.create_var(name=var.name(), desc=var, type=var.type())
 
         # sync variables removed from c++ end
-        for var in self.vars.keys():
+        for var in list(self.vars.keys()):
             if not self.desc.find_var(var):
                 self.vars.pop(var)
 
@@ -1185,7 +1224,7 @@ class Block(object):
         if not isinstance(other, Block):
             raise TypeError(
                 "_copy_param_info_from should be invoked with Block")
-        for p in other._iter_parameters():
+        for p in other.iter_parameters():
             assert isinstance(p, Parameter)
             v = self.vars.get(p.name, None)
             if v is None:
@@ -1384,7 +1423,8 @@ class Program(object):
                 res_str += block.to_string(throw_on_error, with_details)
         else:
             protostr = self.desc.serialize_to_string()
-            proto = framework_pb2.ProgramDesc.FromString(str(protostr))
+            proto = framework_pb2.ProgramDesc.FromString(
+                six.binary_type(protostr))
             res_str = _debug_string_(proto, throw_on_error)
         return res_str
 
@@ -1482,7 +1522,7 @@ class Program(object):
         else:
             p = Program()
             p.desc = core.ProgramDesc(self.desc)
-            p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())]
+            p.blocks = [Block(p, i) for i in range(self.desc.num_blocks())]
             p._sync_with_cpp()
 
         p._copy_param_info_from(self)
@@ -1534,7 +1574,7 @@ class Program(object):
             targets_idx.append([t.block.idx, t.idx])
         res = Program()
         res.desc = core.prune(self.desc, targets_idx)
-        res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())]
+        res.blocks = [Block(res, i) for i in range(res.desc.num_blocks())]
         res._sync_with_cpp()
         return res
 
@@ -1575,13 +1615,13 @@ class Program(object):
                 root_block._remove_var(var.name())
 
         # change all `is_test` attributes to True
-        for i in xrange(res.desc.num_blocks()):
+        for i in range(res.desc.num_blocks()):
             block = res.desc.block(i)
-            for j in xrange(block.op_size()):
+            for j in range(block.op_size()):
                 op = block.op(j)
                 if op.has_attr('is_test'):
                     op.set_attr('is_test', True)
-        res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())]
+        res.blocks = [Block(res, i) for i in range(res.desc.num_blocks())]
         res._sync_with_cpp()
         return res
 
@@ -1594,14 +1634,14 @@ class Program(object):
         and deserialization.
 
         Args:
-            binary_str(str): The binary prootbuf string.
+            binary_str_type(str): The binary prootbuf string.
 
         Returns:
             Program: A deserialized program desc.
         """
         p = Program()
         p.desc = core.ProgramDesc(binary_str)
-        p.blocks = [Block(p, i) for i in xrange(p.desc.num_blocks())]
+        p.blocks = [Block(p, i) for i in range(p.desc.num_blocks())]
         p._sync_with_cpp()
         return p
 
@@ -1629,7 +1669,7 @@ class Program(object):
         self._seed = seed
 
     def __repr__(self):
-        return str(self)
+        return self.__str__()
 
     def global_block(self):
         """
@@ -1740,7 +1780,7 @@ class Program(object):
         if len(self.blocks) != len(other.blocks):
             raise ValueError("_copy_param_info_from should be invoked with two "
                              "program, with represent the same topology")
-        for var in other.global_block().vars.itervalues():
+        for var in list(other.global_block().vars.values()):
             if var.is_data:
                 self.global_block().var(var.name).is_data = True
 
@@ -1752,7 +1792,7 @@ class Program(object):
             iterable: The generator will yield every variable in this program.
         """
         for each_block in self.blocks:
-            for each_var in each_block.vars.itervalues():
+            for each_var in list(each_block.vars.values()):
                 yield each_var
 
 
@@ -1826,8 +1866,8 @@ class Parameter(Variable):
             additional_attr = ("trainable", "optimize_attr", "regularizer",
                                "gradient_clip_attr", "do_model_average")
             for attr_name in additional_attr:
-                res_str += "%s: %s\n" % (attr_name,
-                                         str(getattr(self, attr_name)))
+                res_str += "%s: %s\n" % (
+                    attr_name, six.binary_type(getattr(self, attr_name)))
         else:
             res_str = Variable.to_string(self, throw_on_error, False)
         return res_str

python/paddle/fluid/graphviz.py

@@ -14,12 +14,13 @@
 
 import os
 import random
+import six
 import subprocess
 import logging
 
 
 def crepr(v):
-    if type(v) is str or type(v) is unicode:
+    if isinstance(v, six.string_types):
         return '"%s"' % v
     return str(v)
 
@@ -104,7 +105,7 @@ class Graph(object):
 
     def _rank_repr(self):
         ranks = sorted(
-            self.rank_groups.items(),
+            list(self.rank_groups.items()),
             cmp=lambda a, b: a[1].priority > b[1].priority)
         repr = []
         for x in ranks:
@@ -148,7 +149,7 @@ class Node(object):
             name=self.name,
             label=self.label,
             extra=',' + ','.join("%s=%s" % (key, crepr(value))
-                                 for key, value in self.attrs.items())
+                                 for key, value in list(self.attrs.items()))
             if self.attrs else "")
         return reprs
 
@@ -172,7 +173,7 @@ class Edge(object):
             target=self.target.name,
             extra="" if not self.attrs else
             "[" + ','.join("{}={}".format(attr[0], crepr(attr[1]))
-                           for attr in self.attrs.items()) + "]")
+                           for attr in list(self.attrs.items())) + "]")
         return repr
 
 

python/paddle/fluid/inferencer.py

@@ -14,14 +14,14 @@
 
 import contextlib
 
-import core
-
-import executor
-import framework
-import io
-import parallel_executor
-import unique_name
-from trainer import check_and_get_place
+from . import core
+
+from . import executor
+from . import framework
+from . import io
+from . import parallel_executor
+from . import unique_name
+from .trainer import check_and_get_place
 
 __all__ = ['Inferencer', ]
 

python/paddle/fluid/initializer.py

@@ -12,11 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import framework
+from . import framework
 import numpy as np
 import contextlib
-from framework import convert_np_dtype_to_dtype_
-from core import VarDesc
+from .framework import convert_np_dtype_to_dtype_
+from .core import VarDesc
 
 __all__ = [
     'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'MSRA',

python/paddle/fluid/io.py

@@ -16,6 +16,7 @@ import os
 import errno
 import time
 import shutil
+import six
 
 from paddle.fluid.evaluator import Evaluator
 from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable
@@ -163,7 +164,7 @@ def save_vars(executor,
         save_vars(
             executor,
             dirname=dirname,
-            vars=filter(predicate, main_program.list_vars()),
+            vars=list(filter(predicate, main_program.list_vars())),
             filename=filename)
     else:
         save_program = Program()
@@ -369,7 +370,7 @@ def load_vars(executor,
         load_vars(
             executor,
             dirname=dirname,
-            vars=filter(predicate, main_program.list_vars()),
+            vars=list(filter(predicate, main_program.list_vars())),
             filename=filename)
     else:
         load_prog = Program()
@@ -599,13 +600,25 @@ def save_inference_model(dirname,
             # "./infer_model".
 
     """
-    if isinstance(feeded_var_names, basestring):
+    if isinstance(feeded_var_names, six.binary_type):
         feeded_var_names = [feeded_var_names]
+    elif isinstance(feeded_var_names, six.text_type):
+        feeded_var_names = [feeded_var_names.encode()]
     else:
         if len(feeded_var_names) > 0:
+            # TODO(paddle-dev): polish these code blocks
             if not (bool(feeded_var_names) and all(
-                    isinstance(name, basestring) for name in feeded_var_names)):
-                raise ValueError("'feed_var_names' should be a list of str.")
+                    isinstance(name, six.binary_type)
+                    for name in feeded_var_names)):
+                if not (all(
+                        isinstance(name, six.text_type)
+                        for name in feeded_var_names)):
+                    raise ValueError(
+                        "'feed_var_names' should be a list of str.")
+                else:
+                    feeded_var_names = [
+                        name.encode() for name in feeded_var_names
+                    ]
 
     if isinstance(target_vars, Variable):
         target_vars = [target_vars]

python/paddle/fluid/layer_helper.py

@@ -14,12 +14,14 @@
 
 import copy
 import itertools
+import six
 
-from framework import Variable, Parameter, default_main_program, default_startup_program, dtype_is_floating
-import unique_name
+from .framework import Variable, Parameter, default_main_program, default_startup_program, dtype_is_floating
+from . import unique_name
 from paddle.fluid.initializer import Constant, Xavier
-from param_attr import ParamAttr, WeightNormParamAttr
-import core
+from .param_attr import ParamAttr, WeightNormParamAttr
+from . import core
+from six.moves import zip
 
 
 class LayerHelper(object):
@@ -83,7 +85,7 @@ class LayerHelper(object):
             raise ValueError("parameter number mismatch")
         elif len(param_attr) == 1 and length != 1:
             tmp = [None] * length
-            for i in xrange(length):
+            for i in range(length):
                 tmp[i] = copy.deepcopy(param_attr[0])
             param_attr = tmp
         return param_attr
@@ -91,7 +93,7 @@ class LayerHelper(object):
     def iter_inputs_and_params(self, input_param_name='input'):
         inputs = self.multiple_input(input_param_name)
         param_attrs = self.multiple_param_attr(len(inputs))
-        for ipt, param_attr in itertools.izip(inputs, param_attrs):
+        for ipt, param_attr in zip(inputs, param_attrs):
             yield ipt, param_attr
 
     def input_dtype(self, input_param_name='input'):
@@ -218,7 +220,7 @@ class LayerHelper(object):
                 norm = __norm_op(reshape, dim=0, block=block)
                 __reshape_op(norm, out=out, shape=out_shape, block=block)
             else:
-                perm = range(len(x.shape))
+                perm = list(range(len(x.shape)))
                 perm[0], perm[dim] = dim, 0
                 transpose = __transpose_op(x, perm, block=block)
                 norm = __norm_op(transpose, dim=0, block=block)
@@ -397,8 +399,10 @@ class LayerHelper(object):
         act = self.kwargs.get('act', None)
         if act is None:
             return input_var
-        if isinstance(act, basestring):
+        if isinstance(act, six.string_types):
             act = {'type': act}
+        else:
+            raise TypeError(str(act) + " should be unicode or str")
 
         if 'use_cudnn' in self.kwargs and self.kwargs.get('use_cudnn'):
             act['use_cudnn'] = self.kwargs.get('use_cudnn')

python/paddle/fluid/layers/__init__.py

@@ -12,25 +12,25 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import ops
-from ops import *
-import nn
-from nn import *
-import io
-from io import *
-import tensor
-from tensor import *
-import control_flow
-from control_flow import *
-import device
-from device import *
-import math_op_patch
-from math_op_patch import *
-import detection
-from detection import *
-import metric_op
-from metric_op import *
-from learning_rate_scheduler import *
+from . import ops
+from .ops import *
+from . import nn
+from .nn import *
+from . import io
+from .io import *
+from . import tensor
+from .tensor import *
+from . import control_flow
+from .control_flow import *
+from . import device
+from .device import *
+from . import math_op_patch
+from .math_op_patch import *
+from . import detection
+from .detection import *
+from . import metric_op
+from .metric_op import *
+from .learning_rate_scheduler import *
 
 __all__ = []
 __all__ += nn.__all__

python/paddle/fluid/layers/control_flow.py

@@ -13,15 +13,16 @@
 # limitations under the License.
 import contextlib
 
-from layer_function_generator import autodoc, templatedoc
-from tensor import assign, fill_constant
+from .layer_function_generator import autodoc, templatedoc
+from .tensor import assign, fill_constant
 from .. import core
 from ..framework import Program, Variable, Operator
 from ..layer_helper import LayerHelper, unique_name
 from ..initializer import force_init_on_cpu
-from ops import logical_and, logical_not, logical_or
+from .ops import logical_and, logical_not, logical_or
 import numpy
 import warnings
+from functools import reduce
 
 __all__ = [
     'While',
@@ -601,7 +602,7 @@ class StaticRNN(object):
         boot_memories = []
         pre_memories = []
         memories = []
-        for _, mem in self.memories.iteritems():
+        for _, mem in list(self.memories.items()):
             boot_memories.append(mem.init)
             pre_memories.append(mem.pre_mem.name)
             mem_var = rnn_block.var(mem.mem.name)
@@ -1512,7 +1513,7 @@ class IfElse(object):
     def __call__(self):
         if self.status != self.OUT_IF_ELSE_BLOCKS:
             raise ValueError("IfElse::__call__ must be out of sub-block")
-        false_len, true_len = map(len, self.output_table)
+        false_len, true_len = list(map(len, self.output_table))
         if false_len == 0 and true_len == 0:
             raise ValueError("Must invoke true_block/false_block before "
                              "__call__")

python/paddle/fluid/layers/detection.py

@@ -15,12 +15,13 @@
 All layers just related to the detection neural network.
 """
 
-from layer_function_generator import generate_layer_fn
-from layer_function_generator import autodoc, templatedoc
+from .layer_function_generator import generate_layer_fn
+from .layer_function_generator import autodoc, templatedoc
 from ..layer_helper import LayerHelper
-import tensor
-import nn
+from . import tensor
+from . import nn
 import math
+from functools import reduce
 
 __all__ = [
     'prior_box',
@@ -1032,7 +1033,7 @@ def multi_box_head(inputs,
         min_sizes = []
         max_sizes = []
         step = int(math.floor(((max_ratio - min_ratio)) / (num_layer - 2)))
-        for ratio in xrange(min_ratio, max_ratio + 1, step):
+        for ratio in range(min_ratio, max_ratio + 1, step):
             min_sizes.append(base_size * ratio / 100.)
             max_sizes.append(base_size * (ratio + step) / 100.)
         min_sizes = [base_size * .10] + min_sizes

python/paddle/fluid/layers/device.py

@@ -15,7 +15,7 @@
 All util layers.
 """
 
-from layer_function_generator import autodoc
+from .layer_function_generator import autodoc
 from ..framework import unique_name
 from ..layer_helper import LayerHelper
 from ..annotations import deprecated

python/paddle/fluid/layers/io.py

@@ -16,8 +16,8 @@ import multiprocessing
 import threading
 
 from ..data_feeder import DataFeeder
-from control_flow import BlockGuard
-from layer_function_generator import templatedoc
+from .control_flow import BlockGuard
+from .layer_function_generator import templatedoc
 from .. import core
 from ..executor import global_scope
 from ..framework import convert_np_dtype_to_dtype_, default_main_program, \
@@ -69,7 +69,7 @@ def data(name,
     """
     helper = LayerHelper('data', **locals())
     shape = list(shape)
-    for i in xrange(len(shape)):
+    for i in range(len(shape)):
         if shape[i] is None:
             shape[i] = -1
             append_batch_size = False
@@ -1005,7 +1005,7 @@ class Preprocessor(object):
         source_lod_levels = self.underlying_reader.desc.lod_levels()
         self.source_var_names = [
             unique_name("preprocessor_source")
-            for _ in xrange(len(source_shapes))
+            for _ in range(len(source_shapes))
         ]
         source_vars = []
         for var_name, shape, dtype, lod_level in zip(

python/paddle/fluid/layers/layer_function_generator.py

@@ -12,11 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import re
-import cStringIO
 import functools
 import warnings
 import string
 
+from six.moves import cStringIO
 from ..proto import framework_pb2
 from ..framework import OpProtoHolder, Variable
 from ..layer_helper import LayerHelper
@@ -70,7 +70,7 @@ def _generate_doc_string_(op_proto):
     if not isinstance(op_proto, framework_pb2.OpProto):
         raise TypeError("OpProto should be `framework_pb2.OpProto`")
 
-    buf = cStringIO.StringIO()
+    buf = cStringIO()
     buf.write(escape_math(op_proto.comment))
     buf.write('\nArgs:\n')
     for each_input in op_proto.inputs:
@@ -119,9 +119,9 @@ def generate_layer_fn(op_type):
     """
     op_proto = OpProtoHolder.instance().get_op_proto(op_type)
     not_intermediate_outputs = \
-        filter(lambda output: not output.intermediate, op_proto.outputs)
+        [output for output in op_proto.outputs if not output.intermediate]
     intermediate_outputs = \
-        filter(lambda output: output.intermediate, op_proto.outputs)
+        [output for output in op_proto.outputs if output.intermediate]
 
     if len(not_intermediate_outputs) != 1:
         raise ValueError("Only one non intermediate output operator can be",

python/paddle/fluid/layers/learning_rate_scheduler.py

@@ -20,10 +20,10 @@ User can also implement their own learning_rate_decay
 strategy according to this module.
 """
 
-import control_flow
-import nn
-import ops
-import tensor
+from . import control_flow
+from . import nn
+from . import ops
+from . import tensor
 from ..initializer import init_on_cpu
 from ..framework import default_main_program, Parameter
 

python/paddle/fluid/layers/math_op_patch.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from ..framework import Variable, unique_name
-from layer_function_generator import OpProtoHolder
+from .layer_function_generator import OpProtoHolder
 from ..initializer import force_init_on_cpu
 
 

python/paddle/fluid/layers/metric_op.py

@@ -20,7 +20,7 @@ from ..layer_helper import LayerHelper
 from ..initializer import Normal, Constant
 from ..framework import Variable
 from ..param_attr import ParamAttr
-import nn
+from . import nn
 
 __all__ = ['accuracy', 'auc']
 

python/paddle/fluid/layers/nn.py

@@ -33,11 +33,12 @@ from ..layer_helper import LayerHelper
 from ..initializer import Normal, Constant
 from ..framework import Variable
 from ..param_attr import ParamAttr
-from layer_function_generator import autodoc, templatedoc
-from tensor import concat
-import utils
+from .layer_function_generator import autodoc, templatedoc
+from .tensor import concat
+from . import utils
 import random
 from .. import unique_name
+from functools import reduce
 
 __all__ = [
     'fc',
@@ -949,6 +950,10 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
     helper = LayerHelper('dropout', **locals())
     out = helper.create_tmp_variable(dtype=x.dtype)
     mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True)
+
+    if (seed is None or seed == 0) and helper.main_program.random_seed != 0:
+        seed = helper.main_program.random_seed
+
     helper.append_op(
         type='dropout',
         inputs={'X': [x]},
@@ -4845,7 +4850,7 @@ def dice_loss(input, label, epsilon=0.00001):
             loss = fluid.layers.dice_loss(input=predictions, label=label, 2)
     """
     label = one_hot(label, depth=input.shape[-1])
-    reduce_dim = range(1, len(input.shape))
+    reduce_dim = list(range(1, len(input.shape)))
     inse = reduce_sum(input * label, dim=reduce_dim)
     dice_denominator = reduce_sum(
         input, dim=reduce_dim) + reduce_sum(

python/paddle/fluid/layers/ops.py

@@ -11,7 +11,7 @@
 # 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.
-from layer_function_generator import generate_layer_fn
+from .layer_function_generator import generate_layer_fn
 
 __activations__ = [
     'sigmoid',