test=develop

CMakeLists.txt

@@ -127,6 +127,9 @@ set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
 set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING
   "A path setting fluid shared and static libraries")
 
+set(FLUID_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_inference_install_dir" CACHE STRING
+  "A path setting fluid inference shared and static libraries")
+
 if (WITH_C_API AND WITH_PYTHON)
   message(WARNING "It is suggest not embedded a python interpreter in Paddle "
     "when using C-API. It will give an unpredictable behavior when using a "

README.md

@@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
 
 
-### Latest PaddlePaddle Release: [Fluid 0.15.0](https://github.com/PaddlePaddle/Paddle/tree/v0.15.0)
+### Latest PaddlePaddle Release: [Fluid 1.0.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0)
 ### Install Latest Stable Release:
 ```
 # Linux CPU
@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==0.15.0.post85
 
 ## Installation
 
-It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/install/install_doc.html) on our website.
+It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) on our website.
 
 ## Documentation
 
-We provide [English](http://paddlepaddle.org/documentation/docs/en/0.15.0/getstarted/index_en.html) and
-[Chinese](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/index.html) documentation.
+We provide [English](http://paddlepaddle.org/documentation/docs/en/1.0.0/getstarted/index_en.html) and
+[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) documentation.
 
 - [Deep Learning 101](https://github.com/PaddlePaddle/book)
 
   You might want to start from this online interactive book that can run in a Jupyter Notebook.
 
-- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/user_guides/howto/training/cluster_howto.html)
+- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.0/user_guides/howto/training/cluster_howto.html)
 
   You can run distributed training jobs on MPI clusters.
 
-- [Python API](http://paddlepaddle.org/documentation/api/zh/0.15.0/fluid.html)
+- [Python API](http://paddlepaddle.org/documentation/api/zh/1.0/fluid.html)
 
    Our new API enables much shorter programs.
 
-- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/advanced_usage/development/contribute_to_paddle.html)
+- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.0/advanced_usage/development/contribute_to_paddle.html)
 
    We appreciate your contributions!
 

cmake/inference_lib.cmake

@@ -150,16 +150,16 @@ if (WITH_ANAKIN AND WITH_MKL)
         SRCS
         ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api
         ${ANAKIN_INSTALL_DIR} # anakin release
-        DSTS ${dst_dir}/inference/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
+        DSTS ${FLUID_INSTALL_DIR}/third_party/install/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin)
      list(APPEND inference_deps anakin_inference_lib)
 endif()
 
 set(module "inference")
 copy(inference_lib DEPS ${inference_deps}
   SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
-       ${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci
+       ${src_dir}/${module}/api/paddle_inference_api.h
        ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
 )
 
 set(module "platform")
@@ -188,18 +188,38 @@ copy(cmake_cache
 # This command generates a complete fluid library for both train and inference
 add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep}) 
 
+# Following commands generate a inference-only fluid library
+# third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR}
+copy(third_party DEPS fluid_lib_dist
+  SRCS ${FLUID_INSTALL_DIR}/third_party ${FLUID_INSTALL_DIR}/CMakeCache.txt
+  DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR}
+)
+
+# only need libpaddle_fluid.so/a and paddle_inference_api.h for inference-only library
+copy(inference_api_lib DEPS fluid_lib_dist
+  SRCS ${FLUID_INSTALL_DIR}/paddle/fluid/inference/libpaddle_fluid.*
+       ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_inference_api.h
+  DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include
+)
+
+add_custom_target(inference_lib_dist DEPENDS third_party inference_api_lib)
+
 # paddle fluid version
-execute_process(
-  COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
-  WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
-  OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
-set(version_file ${FLUID_INSTALL_DIR}/version.txt)
-file(WRITE ${version_file}
-  "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
-  "WITH_MKL: ${WITH_MKL}\n"
-  "WITH_GPU: ${WITH_GPU}\n")
-if(WITH_GPU)
-  file(APPEND ${version_file}
-    "CUDA version: ${CUDA_VERSION}\n"
-    "CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
-endif()
+function(version version_file)
+  execute_process(
+    COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
+    OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
+  file(WRITE ${version_file}
+    "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
+    "WITH_MKL: ${WITH_MKL}\n"
+    "WITH_MKLDNN: ${WITH_MKLDNN}\n"
+    "WITH_GPU: ${WITH_GPU}\n")
+  if(WITH_GPU)
+    file(APPEND ${version_file}
+      "CUDA version: ${CUDA_VERSION}\n"
+      "CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
+  endif()
+endfunction()
+version(${FLUID_INSTALL_DIR}/version.txt)
+version(${FLUID_INFERENCE_INSTALL_DIR}/version.txt)

paddle/fluid/API.spec

@@ -75,7 +75,8 @@ paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'outp
 paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.sequence_expand_as ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.sequence_unpad ArgSpec(args=['x', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None))
 paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
@@ -84,6 +85,7 @@ paddle.fluid.layers.reduce_min ArgSpec(args=['input', 'dim', 'keep_dim', 'name']
 paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.sequence_slice ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None))
 paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,))
@@ -172,6 +174,7 @@ paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dim
 paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)

paddle/fluid/framework/executor.cc

@@ -101,7 +101,7 @@ void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope>>();
+    var->GetMutable<std::vector<framework::Scope*>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {

paddle/fluid/framework/feed_fetch_method.cc

@@ -27,8 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
   // be created.
   VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
   Variable* g_feed_value = scope->Var(var_name);
-  auto& feed_inputs =
-      *(g_feed_value->GetMutable<std::vector<paddle::framework::LoDTensor>>());
+  auto& feed_inputs = *(g_feed_value->GetMutable<FeedFetchList>());
   if (index >= feed_inputs.size()) {
     feed_inputs.resize(index + 1);
   }

paddle/fluid/framework/naive_executor.cc

@@ -37,7 +37,7 @@ static void InitializeVariable(Variable *var, proto::VarType::Type var_type) {
   } else if (var_type == proto::VarType::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
   } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope>>();
+    var->GetMutable<std::vector<framework::Scope *>>();
   } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
   } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {

paddle/fluid/framework/op_desc.h

@@ -100,16 +100,6 @@ class OpDesc {
   std::vector<std::string> InputNames() const { return MapKeys(inputs_); }
   std::vector<std::string> OutputNames() const { return MapKeys(outputs_); }
 
-  void SetInputMap(const VariableNameMap &input) {
-    this->inputs_ = input;
-    this->need_update_ = true;
-  }
-
-  void SetOutputMap(const VariableNameMap &output) {
-    this->outputs_ = output;
-    this->need_update_ = true;
-  }
-
   const VariableNameMap &Inputs() const { return inputs_; }
 
   const VariableNameMap &Outputs() const { return outputs_; }

paddle/fluid/framework/operator.cc

@@ -149,9 +149,17 @@ 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);
+
+  // The profile has a process-wide mutex, results in serious performance issue
+  // in concurrency scenerio. Here use an `if` to fix this issue.
+  // Please not remove the `if`, ask @Superjomn if there are any concern.
+  if (platform::IsProfileEnabled()) {
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    platform::RecordEvent record_event(Type(), pool.Get(place));
+    RunImpl(scope, place);
+  } else {
+    RunImpl(scope, place);
+  }
   VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 

paddle/fluid/framework/tensor_util.cc

@@ -36,6 +36,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
   auto size = src.numel() * SizeOfType(src.type());
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    if (src_ptr == dst_ptr) {
+      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::CPUPlace>(src_place), src_ptr, size);
   }
@@ -71,6 +76,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
     auto stream =
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
     if (platform::is_same_place(src_place, dst_place)) {
+      if (src_ptr == dst_ptr) {
+        VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+                << dst_place;
+        return;
+      }
       memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                    stream);
     } else {
@@ -114,6 +124,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
   auto dst_ptr = dst->mutable_data(dst_place, src.type());
   auto size = src.numel() * SizeOfType(src.type());
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    if (src_ptr == dst_ptr) {
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::CPUPlace>(src_place), src_ptr, size);
   }
@@ -130,6 +145,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
     memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr);
   } else if (platform::is_gpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
+    if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) {
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
+      return;
+    }
     auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
     auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);

paddle/fluid/framework/tensor_util_test.cc

@@ -41,6 +41,11 @@ TEST(TensorCopy, Tensor) {
     EXPECT_EQ(src_ptr[i], dst_ptr[i]);
   }
 
+  TensorCopy(dst_tensor, *cpu_place, &dst_tensor);
+  for (size_t i = 0; i < 9; ++i) {
+    EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+  }
+
   EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout());
 
   Tensor slice_tensor = src_tensor.Slice(1, 2);
@@ -82,6 +87,15 @@ TEST(TensorCopy, Tensor) {
       EXPECT_EQ(src_ptr[i], dst_ptr[i]);
     }
 
+    // Copy the same tensor
+    TensorCopy(gpu_tensor, *gpu_place, gpu_ctx, &gpu_tensor);
+    gpu_ctx.Wait();
+    const int* dst_ptr_tmp = dst_tensor.data<int>();
+    EXPECT_NE(src_ptr, dst_ptr_tmp);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], dst_ptr_tmp[i]);
+    }
+
     Tensor slice_tensor = src_tensor.Slice(1, 2);
 
     // CPU Slice Tensor to GPU Tensor

paddle/fluid/framework/var_desc.h

@@ -59,6 +59,7 @@ class VarDesc {
  public:
   explicit VarDesc(const std::string &name) {
     desc_.set_name(name);
+    // TODO(paddle-dev): Why default to lodtensor.
     desc_.mutable_type()->set_type(proto::VarType::LOD_TENSOR);
   }
 

paddle/fluid/framework/variable.h

@@ -38,8 +38,12 @@ class Variable {
 
   template <typename T>
   T* GetMutable() {
-    if (!IsType<T>()) {
+    if (!holder_) {
       holder_.reset(new PlaceholderImpl<T>(new T()));
+    } else {
+      PADDLE_ENFORCE(IsType<T>(),
+                     "Variable must be type %s, the holding type is %s",
+                     typeid(T).name(), holder_->Type().name());
     }
     return static_cast<T*>(holder_->Ptr());
   }

paddle/fluid/framework/variable_test.cc

@@ -33,9 +33,10 @@ TEST(Variable, GetMutable) {
   const Tensor& tt = v->Get<Tensor>();
   EXPECT_EQ(1234, tt.content_);
 
-  std::string* s = v->GetMutable<std::string>();
-  *s = "hello";
-
-  const std::string& ss = v->Get<std::string>();
-  EXPECT_EQ("hello", ss);
+  try {
+    v->GetMutable<std::string>();
+  } catch (std::exception& e) {
+    return;
+  }
+  EXPECT_TRUE(false);
 }

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -51,9 +51,7 @@ void TestWord2vecPrediction(const std::string& model_path) {
   config.model_dir = model_path;
   config.use_gpu = false;
   config.device = 0;
-  auto predictor =
-      ::paddle::CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(
-          config);
+  auto predictor = ::paddle::CreatePaddlePredictor<NativeConfig>(config);
 
   // One single batch
 

paddle/fluid/inference/api/analysis_predictor.cc

@@ -25,9 +25,11 @@
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/api/paddle_inference_pass.h"
 #include "paddle/fluid/inference/utils/singleton.h"
+#include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DECLARE_bool(profile);
+DECLARE_int32(paddle_num_threads);
 
 namespace paddle {
 
@@ -47,6 +49,9 @@ bool AnalysisPredictor::Init(
   }
 #endif
 
+  // no matter with or without MKLDNN
+  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);
+
   if (config_.use_gpu) {
     place_ = paddle::platform::CUDAPlace(config_.device);
     LOG(WARNING) << "ir optimize only supports CPU currently, enable_ir_optim "
@@ -335,6 +340,19 @@ bool AnalysisPredictor::LoadProgramDesc() {
   }
   return true;
 }
+
+AnalysisPredictor::~AnalysisPredictor() {
+#if !defined(_WIN32)
+  if (FLAGS_profile) {
+    platform::DisableProfiler(platform::EventSortingKey::kTotal,
+                              "./profile.log");
+  }
+#endif
+  if (sub_scope_) {
+    scope_->DeleteScope(sub_scope_);
+  }
+}
+
 std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone() {
   auto *x = new AnalysisPredictor(config_);
   x->Init(scope_, inference_program_);

paddle/fluid/inference/api/analysis_predictor.h

@@ -72,6 +72,7 @@ class AnalysisPredictor : public PaddlePredictor {
   template <typename T>
   void GetFetchOne(const framework::LoDTensor &fetchs,
                    PaddleTensor *output_data);
+  ~AnalysisPredictor();
 
  private:
   contrib::AnalysisConfig config_;

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -27,9 +27,7 @@ TEST(AnalysisPredictor, ZeroCopy) {
   config.model_dir = FLAGS_dirname + "/word2vec.inference.model";
   config.use_feed_fetch_ops = false;
 
-  auto predictor =
-      CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
-          config);
+  auto predictor = CreatePaddlePredictor<AnalysisConfig>(config);
 
   auto w0 = predictor->GetInputTensor("firstw");
   auto w1 = predictor->GetInputTensor("secondw");

paddle/fluid/inference/api/api_impl.cc

@@ -23,9 +23,11 @@ limitations under the License. */
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/inference/api/api_impl.h"
 #include "paddle/fluid/inference/api/helper.h"
+#include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DEFINE_bool(profile, false, "Turn on profiler for fluid");
+DECLARE_int32(paddle_num_threads);
 
 namespace paddle {
 namespace {
@@ -72,6 +74,9 @@ bool NativePaddlePredictor::Init(
   }
 #endif
 
+  // no matter with or without MKLDNN
+  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);
+
   if (config_.use_gpu) {
     place_ = paddle::platform::CUDAPlace(config_.device);
   } else {

paddle/fluid/inference/api/api_impl_tester.cc

@@ -205,7 +205,7 @@ void MainThreadsWord2Vec(bool use_gpu) {
       float* ref_data = refs[tid].data<float>();
       EXPECT_EQ(refs[tid].numel(), static_cast<int64_t>(len / sizeof(float)));
       for (int i = 0; i < refs[tid].numel(); ++i) {
-        EXPECT_NEAR(ref_data[i], data[i], ACC_DIFF);
+        EXPECT_NEAR(ref_data[i], data[i], 2e-3);
       }
     });
   }

paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc

@@ -41,11 +41,8 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
   config1.device = 0;
   config1.max_batch_size = 10;
 
-  auto predictor0 =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config0);
-  auto predictor1 =
-      CreatePaddlePredictor<MixedRTConfig,
-                            PaddleEngineKind::kAutoMixedTensorRT>(config1);
+  auto predictor0 = CreatePaddlePredictor<NativeConfig>(config0);
+  auto predictor1 = CreatePaddlePredictor<MixedRTConfig>(config1);
 
   for (int batch_id = 0; batch_id < 1; batch_id++) {
     //# 2. Prepare input.

paddle/fluid/inference/api/demo_ci/CMakeLists.txt

@@ -77,7 +77,7 @@ endif(NOT WIN32)
 link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib")
 link_directories("${PADDLE_LIB}/third_party/install/glog/lib")
 link_directories("${PADDLE_LIB}/third_party/install/gflags/lib")
-link_directories("${PADDLE_LIB}/paddle/fluid/inference")
+link_directories("${PADDLE_LIB}/paddle/lib")
 
 add_executable(${DEMO_NAME} ${DEMO_NAME}.cc)
 
@@ -97,10 +97,10 @@ endif()
 # Note: libpaddle_inference_api.so/a must put before libpaddle_fluid.so/a
 if(WITH_STATIC_LIB)
   set(DEPS
-      ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
+      ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
 else()
   set(DEPS
-      ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX})
+      ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX})
 endif()
 
 if (NOT WIN32)

paddle/fluid/inference/api/demo_ci/run.sh

@@ -5,12 +5,13 @@ TEST_GPU_CPU=$3 # test both GPU/CPU mode or only CPU mode
 DATA_DIR=$4 # dataset
 TENSORRT_INCLUDE_DIR=$5 # TensorRT header file dir, defalut to /usr/local/TensorRT/include
 TENSORRT_LIB_DIR=$6 # TensorRT lib file dir, default to /usr/local/TensorRT/lib
+inference_install_dir=${PADDLE_ROOT}/build/fluid_inference_install_dir
 
 cd `dirname $0`
 current_dir=`pwd`
 if [ $2 == ON ]; then
   # You can export yourself if move the install path
-  MKL_LIB=${PADDLE_ROOT}/build/fluid_install_dir/third_party/install/mklml/lib
+  MKL_LIB=${inference_install_dir}/third_party/install/mklml/lib
   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${MKL_LIB}
 fi
 if [ $3 == ON ]; then
@@ -55,7 +56,7 @@ cd build
 for WITH_STATIC_LIB in ON OFF; do
   # -----simple_on_word2vec-----
   rm -rf *
-  cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
+  cmake .. -DPADDLE_LIB=${inference_install_dir} \
     -DWITH_MKL=$TURN_ON_MKL \
     -DDEMO_NAME=simple_on_word2vec \
     -DWITH_GPU=$TEST_GPU_CPU \
@@ -75,7 +76,7 @@ for WITH_STATIC_LIB in ON OFF; do
   fi
   # ---------vis_demo---------
   rm -rf *
-  cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
+  cmake .. -DPADDLE_LIB=${inference_install_dir} \
     -DWITH_MKL=$TURN_ON_MKL \
     -DDEMO_NAME=vis_demo \
     -DWITH_GPU=$TEST_GPU_CPU \
@@ -98,7 +99,7 @@ for WITH_STATIC_LIB in ON OFF; do
   # --------tensorrt mobilenet------
   if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then
     rm -rf *
-    cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
+    cmake .. -DPADDLE_LIB=${inference_install_dir} \
       -DWITH_MKL=$TURN_ON_MKL \
       -DDEMO_NAME=trt_mobilenet_demo \
       -DWITH_GPU=$TEST_GPU_CPU \

paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc

@@ -23,7 +23,7 @@ limitations under the License. */
 #include <memory>
 #include <thread>  //NOLINT
 
-#include "paddle/fluid/inference/paddle_inference_api.h"
+#include "paddle/include/paddle_inference_api.h"
 
 DEFINE_string(dirname, "", "Directory of the inference model.");
 DEFINE_bool(use_gpu, false, "Whether use gpu.");
@@ -42,8 +42,7 @@ void Main(bool use_gpu) {
   config.use_gpu = use_gpu;
   config.fraction_of_gpu_memory = 0.15;
   config.device = 0;
-  auto predictor =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+  auto predictor = CreatePaddlePredictor<NativeConfig>(config);
 
   for (int batch_id = 0; batch_id < 3; batch_id++) {
     //# 2. Prepare input.
@@ -85,8 +84,7 @@ void MainThreads(int num_threads, bool use_gpu) {
   config.use_gpu = use_gpu;
   config.fraction_of_gpu_memory = 0.15;
   config.device = 0;
-  auto main_predictor =
-      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+  auto main_predictor = CreatePaddlePredictor<NativeConfig>(config);
 
   std::vector<std::thread> threads;
   for (int tid = 0; tid < num_threads; ++tid) {