update to latest

CMakeLists.txt

@@ -86,6 +86,14 @@ if(ANDROID OR IOS)
         "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
     set(WITH_MKLML OFF CACHE STRING
         "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
+
+    # Compile PaddlePaddle mobile inference library
+    if (NOT WITH_C_API)
+        set(WITH_C_API ON CACHE STRING
+            "Always compile the C_API when cross-compiling for Android and iOS" FORCE)
+    endif()
+    set(MOBILE_INFERENCE ON)
+    add_definitions(-DPADDLE_MOBILE_INFERENCE)
 endif()
 
 set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
@@ -160,9 +168,11 @@ endif(USE_NNPACK)
 
 add_subdirectory(proto)
 
-# "add_subdirectory(go)" should be placed after the following loine,
-# because it depends on paddle/optimizer.
-add_subdirectory(paddle/optimizer)
+if(NOT MOBILE_INFERENCE)
+    # "add_subdirectory(go)" should be placed after the following loine,
+    # because it depends on paddle/optimizer.
+    add_subdirectory(paddle/optimizer)
+endif()
 
 # "add_subdirectory(paddle)" and "add_subdirectory(python)" should be
 # placed after this block, because they depends on it.

cmake/util.cmake

@@ -73,25 +73,43 @@ function(link_paddle_exe TARGET_NAME)
         generate_rdma_links()
     endif()
 
-    target_circle_link_libraries(${TARGET_NAME}
-        ARCHIVE_START
-        paddle_gserver
-        paddle_function
-        ARCHIVE_END
-        paddle_pserver
-        paddle_trainer_lib
-        paddle_network
-        paddle_math
-        paddle_utils
-        paddle_parameter
-        paddle_proto
-        paddle_cuda
-        paddle_optimizer
-        ${EXTERNAL_LIBS}
-        ${CMAKE_THREAD_LIBS_INIT}
-        ${CMAKE_DL_LIBS}
-        ${RDMA_LD_FLAGS}
-        ${RDMA_LIBS})
+    if(MOBILE_INFERENCE)
+        target_circle_link_libraries(${TARGET_NAME}
+            ARCHIVE_START
+            paddle_gserver
+            paddle_function
+            ARCHIVE_END
+            paddle_math
+            paddle_utils
+            paddle_parameter
+            paddle_proto
+            paddle_cuda
+            ${EXTERNAL_LIBS}
+            ${CMAKE_THREAD_LIBS_INIT}
+            ${CMAKE_DL_LIBS}
+            ${RDMA_LD_FLAGS}
+            ${RDMA_LIBS})
+    else()
+        target_circle_link_libraries(${TARGET_NAME}
+            ARCHIVE_START
+            paddle_gserver
+            paddle_function
+            ARCHIVE_END
+            paddle_pserver
+            paddle_trainer_lib
+            paddle_network
+            paddle_math
+            paddle_utils
+            paddle_parameter
+            paddle_proto
+            paddle_cuda
+            paddle_optimizer
+            ${EXTERNAL_LIBS}
+            ${CMAKE_THREAD_LIBS_INIT}
+            ${CMAKE_DL_LIBS}
+            ${RDMA_LD_FLAGS}
+            ${RDMA_LIBS})
+    endif()
 
     if(ANDROID)
         target_link_libraries(${TARGET_NAME} log)

paddle/CMakeLists.txt

@@ -1,27 +1,32 @@
 add_subdirectory(cuda)
 add_subdirectory(function)
 add_subdirectory(utils)
-add_subdirectory(testing)
 add_subdirectory(math)
-add_subdirectory(parameter)
 add_subdirectory(gserver)
-add_subdirectory(pserver)
-add_subdirectory(trainer)
-add_subdirectory(scripts)
-add_subdirectory(string)
-
-if(Boost_FOUND)
-  add_subdirectory(memory)
-  add_subdirectory(platform)
-  add_subdirectory(framework)
-  add_subdirectory(operators)
-  add_subdirectory(pybind)
-endif()
+add_subdirectory(parameter)
+add_subdirectory(testing)
 
-if(WITH_C_API)
+if(MOBILE_INFERENCE)
   add_subdirectory(capi)
-endif()
+else()
+  add_subdirectory(pserver)
+  add_subdirectory(trainer)
+  add_subdirectory(string)
+  add_subdirectory(scripts)
+
+  if(WITH_C_API)
+    add_subdirectory(capi)
+  endif()
+
+  if(Boost_FOUND)
+    add_subdirectory(memory)
+    add_subdirectory(platform)
+    add_subdirectory(framework)
+    add_subdirectory(operators)
+    add_subdirectory(pybind)
+  endif()
 
-if(WITH_SWIG_PY)
-  add_subdirectory(api)
+  if(WITH_SWIG_PY)
+    add_subdirectory(api)
+  endif()
 endif()

paddle/capi/CMakeLists.txt

@@ -37,9 +37,7 @@ set(PADDLE_CAPI_INFER_LIBS
     paddle_cuda
     paddle_function
     paddle_gserver
-    paddle_proto
-    paddle_pserver
-    paddle_network)
+    paddle_proto)
 
 cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
 

paddle/capi/tests/CMakeLists.txt

@@ -4,11 +4,12 @@ add_unittest(capi_test_mats test_Vector.cpp
 target_include_directories(capi_test_mats PUBLIC ${PADDLE_CAPI_INC_PATH})
 target_link_libraries(capi_test_mats paddle_capi)
 
-
-add_unittest_without_exec(capi_test_gradientMachine test_GradientMachine.cpp)
-target_include_directories(capi_test_gradientMachine PUBLIC
-  ${PADDLE_CAPI_INC_PATH})
-target_link_libraries(capi_test_gradientMachine paddle_capi)
-add_test(NAME capi_test_gradientMachine
-  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine
-  WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests)
+if(NOT MOBILE_INFERENCE)
+    add_unittest_without_exec(capi_test_gradientMachine test_GradientMachine.cpp)
+    target_include_directories(capi_test_gradientMachine PUBLIC
+      ${PADDLE_CAPI_INC_PATH})
+    target_link_libraries(capi_test_gradientMachine paddle_capi)
+    add_test(NAME capi_test_gradientMachine
+      COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine
+      WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests)
+endif()

paddle/framework/backward.cc

@@ -302,7 +302,7 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
     return grad_op_descs;  // empty vector
   }
 
-  grad_op_descs = OpRegistry::CreateGradOpDescs(*op_desc);
+  grad_op_descs = OpRegistry::CreateGradOpDescs(op_desc.get());
 
   std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
   for (auto& desc : grad_op_descs) {

paddle/framework/backward_test.cc

@@ -58,6 +58,8 @@ class MulOpMaker : public OpProtoAndCheckerMaker {
     AddInput("X", "A");
     AddInput("Y", "B");
     AddOutput("Out", "Out");
+    AddAttr<int>("x_num_col_dims", "").SetDefault(1).EqualGreaterThan(1);
+    AddAttr<int>("y_num_col_dims", "").SetDefault(1).EqualGreaterThan(1);
     AddComment("Mul");
   }
 };
@@ -440,6 +442,28 @@ TEST(Backward, simple_single_op) {
             std::vector<std::string>({f::GradVarName("b")}));
 }
 
+TEST(Backward, default_attribute) {
+  f::ProgramDesc *program_desc = GetNewProgramDesc();
+  f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
+  f::BlockDescBind *block = program.Block(0);
+  f::OpDescBind *op = block->AppendOp();
+  op->SetType("mul");
+  op->SetInput("X", {"x"});
+  op->SetInput("Y", {"y"});
+  op->SetOutput("Out", {"out"});
+
+  AppendBackward(program, {});
+
+  ASSERT_EQ(block->AllOps().size(), 2UL);
+  EXPECT_EQ(boost::get<int>(op->GetAttr("x_num_col_dims")), 1);
+  EXPECT_EQ(boost::get<int>(op->GetAttr("y_num_col_dims")), 1);
+
+  f::OpDescBind *grad_op = block->AllOps()[1];
+  ASSERT_EQ(grad_op->Type(), "mul_grad");
+  EXPECT_EQ(boost::get<int>(grad_op->GetAttr("x_num_col_dims")), 1);
+  EXPECT_EQ(boost::get<int>(grad_op->GetAttr("y_num_col_dims")), 1);
+}
+
 TEST(Backward, simple_mult_op) {
   f::ProgramDesc *program_desc = GetNewProgramDesc();
   f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);

paddle/framework/block_desc.cc

@@ -74,6 +74,12 @@ void BlockDescBind::Sync() {
     for (auto &op_desc : ops_) {
       op_field.AddAllocated(op_desc->Proto());
     }
+    auto &var_field = *this->desc_->mutable_vars();
+    var_field.Clear();
+    var_field.Reserve(static_cast<int>(vars_.size()));
+    for (auto &var_desc : vars_) {
+      var_field.AddAllocated(var_desc.second->Proto());
+    }
     need_update_ = false;
   }
 }

paddle/framework/block_desc.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <deque>
+#include <memory>
 #include <unordered_map>
 #include <vector>
 #include "paddle/framework/op_desc.h"

paddle/framework/framework.proto

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 syntax = "proto2";
+option optimize_for = LITE_RUNTIME;
 package paddle.framework;
 
 enum AttrType {

paddle/framework/op_desc.h

@@ -52,8 +52,6 @@ class OpDescBind {
   void SetOutput(const std::string &param_name,
                  const std::vector<std::string> &args);
 
-  std::string DebugString() { return this->Proto()->DebugString(); }
-
   bool HasAttr(const std::string &name) const {
     return attrs_.find(name) != attrs_.end();
   }
@@ -97,6 +95,11 @@ class OpDescBind {
 
   const VariableNameMap &Outputs() const { return outputs_; }
 
+  AttributeMap *MutableAttrMap() {
+    this->need_update_ = true;
+    return &this->attrs_;
+  }
+
  private:
   template <typename MapType>
   static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {

paddle/framework/op_registry.cc

@@ -60,9 +60,14 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDescBind& op_desc) {
 }
 
 std::vector<std::unique_ptr<OpDescBind>> OpRegistry::CreateGradOpDescs(
-    const OpDescBind& op_desc) {
-  auto& info = OpInfoMap::Instance().Get(op_desc.Type());
-  return info.grad_op_maker_(op_desc);
+    OpDescBind* op_desc) {
+  auto& info = OpInfoMap::Instance().Get(op_desc->Type());
+
+  if (info.Checker() != nullptr) {
+    info.Checker()->Check(*op_desc->MutableAttrMap());
+  }
+
+  return info.grad_op_maker_(*op_desc);
 }
 
 }  // namespace framework

paddle/framework/op_registry.h

@@ -80,7 +80,7 @@ class OpRegistry {
   static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
   static std::vector<std::unique_ptr<OpDescBind>> CreateGradOpDescs(
-      const OpDescBind& op_desc);
+      OpDescBind* op_desc);
 
   static std::unique_ptr<OperatorBase> CreateOp(const OpDescBind& op_desc);
 };

paddle/framework/operator.cc

@@ -205,13 +205,13 @@ void OperatorBase::GenerateTemporaryNames() {
 }
 
 template <>
-const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const {
+const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
   auto* var = InputVar(name);
   return var == nullptr ? nullptr : GetTensorFromVar(var);
 }
 
 template <>
-const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
     const std::string& name) const {
   auto names = op().Inputs(name);
   std::vector<const Tensor*> res;
@@ -225,13 +225,13 @@ const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
 }
 
 template <>
-Tensor* InferShapeContext::Output<Tensor>(const std::string& name) const {
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
   auto var = OutputVar(name);
   return var == nullptr ? nullptr : var->GetMutable<LoDTensor>();
 }
 
 template <>
-std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
     const std::string& name) const {
   auto names = op().Outputs(name);
   std::vector<Tensor*> res;

paddle/framework/operator.h

@@ -57,7 +57,6 @@ inline std::string GradVarName(const std::string& var_name) {
 }
 
 class OperatorBase;
-class InferShapeContext;
 class ExecutionContext;
 
 extern const Tensor* GetTensorFromVar(const Variable* var);
@@ -169,10 +168,11 @@ class NOP : public OperatorBase {
   }
 };
 
-class InferShapeContext {
+class ExecutionContext {
  public:
-  InferShapeContext(const OperatorBase& op, const Scope& scope)
-      : op_(op), scope_(scope) {}
+  ExecutionContext(const OperatorBase& op, const Scope& scope,
+                   const platform::DeviceContext& device_context)
+      : op_(op), scope_(scope), device_context_(device_context) {}
 
   const OperatorBase& op() const { return op_; }
 
@@ -278,31 +278,6 @@ class InferShapeContext {
     out_tensor->set_lod(in_tensor.lod());
   }
 
- private:
-  const OperatorBase& op_;
-  const Scope& scope_;
-};
-
-template <>
-const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const;
-
-template <>
-const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
-    const std::string& name) const;
-
-template <>
-Tensor* InferShapeContext::Output<Tensor>(const std::string& name) const;
-
-template <>
-std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
-    const std::string& name) const;
-
-class ExecutionContext : public InferShapeContext {
- public:
-  ExecutionContext(const OperatorBase& op, const Scope& scope,
-                   const platform::DeviceContext& device_context)
-      : InferShapeContext(op, scope), device_context_(device_context) {}
-
   template <typename PlaceType,
             typename DeviceType = typename platform::EigenDeviceConverter<
                 PlaceType>::EigenDeviceType>
@@ -315,10 +290,26 @@ class ExecutionContext : public InferShapeContext {
   }
 
  private:
+  const OperatorBase& op_;
+  const Scope& scope_;
   const platform::DeviceContext& device_context_;
 };
 
-class CompileTimeInferShapeContext : public InferShapeContextBase {
+template <>
+const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const;
+
+template <>
+const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
+    const std::string& name) const;
+
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const;
+
+class CompileTimeInferShapeContext : public InferShapeContext {
  public:
   CompileTimeInferShapeContext(const OpDescBind& op, const BlockDescBind& block)
       : op_(op), block_(block) {}
@@ -414,7 +405,7 @@ class CompileTimeInferShapeContext : public InferShapeContextBase {
   const BlockDescBind& block_;
 };
 
-class RuntimeInferShapeContext : public InferShapeContextBase {
+class RuntimeInferShapeContext : public InferShapeContext {
  public:
   RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope)
       : op_(op), scope_(scope) {}
@@ -612,7 +603,7 @@ class OperatorWithKernel : public OperatorBase {
                        });
   }
 
-  virtual void InferShape(InferShapeContextBase* ctx) const = 0;
+  virtual void InferShape(InferShapeContext* ctx) const = 0;
 
  protected:
   // indicate kernel DataType by input data. Defaultly all input data must be

paddle/framework/operator_test.cc

@@ -113,7 +113,7 @@ class OpWithKernelTest : public OperatorWithKernel {
   using OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {}
+  void InferShape(framework::InferShapeContext* ctx) const override {}
   DataType IndicateDataType(const ExecutionContext& ctx) const override {
     return DataType::FP32;
   }

paddle/framework/program_desc.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <memory>
 #include <vector>
 #include "paddle/framework/framework.pb.h"
 #include "paddle/platform/macros.h"
@@ -31,8 +32,6 @@ class ProgramDescBind {
 
   BlockDescBind *Block(size_t idx) { return blocks_[idx].get(); }
 
-  std::string DebugString() { return Proto()->DebugString(); }
-
   size_t Size() const { return blocks_.size(); }
 
   ProgramDesc *Proto();

paddle/framework/shape_inference.h

@@ -20,11 +20,11 @@ namespace paddle {
 namespace framework {
 
 // TODO(longfei): Once after both CompileTimeInferShapeContext and
-// RuntimeInferShapeContext get merged, we can rename InferShapeContextBase into
+// RuntimeInferShapeContext get merged, we can rename InferShapeContext into
 // InferShapeContext so to replace the current InferShapeContext.
-class InferShapeContextBase {
+class InferShapeContext {
  public:
-  virtual ~InferShapeContextBase() {}
+  virtual ~InferShapeContext() {}
   virtual bool HasInput(const std::string &name) const = 0;
   virtual bool HasOutput(const std::string &name) const = 0;
 

paddle/framework/tensor.h

@@ -95,6 +95,19 @@ class Tensor {
   template <typename T>
   inline void CopyFrom(const Tensor& src, const platform::Place& dst_place);
 
+  /**
+   * @brief   Copy the content of an external vector to a tensor.
+   *
+   * @param[in] src   The external vector.
+   * @param[in] ctx   The device context contains place where to store.
+   *
+   * * @note    CopyFromVector assumes that the tensor has been resized
+   *            before invoking.
+   */
+  template <typename T>
+  inline void CopyFromVector(const std::vector<T>& src,
+                             const platform::Place& dst_place);
+
   /**
    * @brief   Return the slice of the tensor.
    *

paddle/framework/tensor_impl.h

@@ -123,6 +123,29 @@ inline void Tensor::CopyFrom(const Tensor& src,
 #endif
 }
 
+template <typename T>
+inline void Tensor::CopyFromVector(const std::vector<T>& src,
+                                   const platform::Place& dst_place) {
+  auto src_ptr = static_cast<const void*>(src.data());
+  platform::CPUPlace src_place;
+  auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
+  auto size = src.size() * sizeof(T);
+
+  if (platform::is_cpu_place(dst_place)) {
+    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr, src_place,
+                 src_ptr, size);
+  }
+#ifdef PADDLE_WITH_CUDA
+  else if (platform::is_gpu_place(dst_place)) {
+    memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr, src_place,
+                 src_ptr, size, 0);
+  }
+  PADDLE_ENFORCE(cudaStreamSynchronize(0),
+                 "cudaStreamSynchronize failed in Tensor CopyFromVector");
+
+#endif
+}
+
 template <typename T>
 inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   check_memory_size<T>();

paddle/framework/tensor_test.cc

@@ -263,6 +263,93 @@ TEST(Tensor, CopyFrom) {
 #endif
 }
 
+TEST(Tensor, CopyFromVector) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  {
+    std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+    Tensor cpu_tensor;
+
+    // Copy to CPU Tensor
+    cpu_tensor.Resize(make_ddim({3, 3}));
+    auto cpu_place = new paddle::platform::CPUPlace();
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+
+    // Compare Tensors
+    const int* cpu_ptr = cpu_tensor.data<int>();
+    const int* src_ptr = src_vec.data();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+    }
+
+    src_vec.erase(src_vec.begin(), src_vec.begin() + 5);
+    cpu_tensor.Resize(make_ddim({2, 2}));
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+    cpu_ptr = cpu_tensor.data<int>();
+    src_ptr = src_vec.data();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    for (size_t i = 0; i < 5; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+    }
+
+    delete cpu_place;
+  }
+
+#ifdef PADDLE_WITH_CUDA
+  {
+    std::vector<int> src_vec = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+    Tensor cpu_tensor;
+    Tensor gpu_tensor;
+    Tensor dst_tensor;
+
+    // Copy to CPU Tensor
+    cpu_tensor.Resize(make_ddim({3, 3}));
+    auto cpu_place = new paddle::platform::CPUPlace();
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+
+    // Copy to GPUTensor
+    gpu_tensor.Resize(make_ddim({3, 3}));
+    auto gpu_place = new paddle::platform::GPUPlace();
+    gpu_tensor.CopyFromVector<int>(src_vec, *gpu_place);
+    // Copy from GPU to CPU tensor for comparison
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+
+    // Compare Tensors
+    const int* src_ptr = src_vec.data();
+    const int* cpu_ptr = cpu_tensor.data<int>();
+    const int* dst_ptr = dst_tensor.data<int>();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    ASSERT_NE(src_ptr, dst_ptr);
+    for (size_t i = 0; i < 9; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+      EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+    }
+
+    src_vec.erase(src_vec.begin(), src_vec.begin() + 5);
+
+    cpu_tensor.Resize(make_ddim({2, 2}));
+    cpu_tensor.CopyFromVector<int>(src_vec, *cpu_place);
+    gpu_tensor.Resize(make_ddim({2, 2}));
+    gpu_tensor.CopyFromVector<int>(src_vec, *gpu_place);
+    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place);
+
+    src_ptr = src_vec.data();
+    cpu_ptr = cpu_tensor.data<int>();
+    dst_ptr = dst_tensor.data<int>();
+    ASSERT_NE(src_ptr, cpu_ptr);
+    ASSERT_NE(src_ptr, dst_ptr);
+    for (size_t i = 0; i < 5; ++i) {
+      EXPECT_EQ(src_ptr[i], cpu_ptr[i]);
+      EXPECT_EQ(src_ptr[i], dst_ptr[i]);
+    }
+
+    delete cpu_place;
+    delete gpu_place;
+  }
+#endif
+}
+
 TEST(Tensor, ReshapeToMatrix) {
   using namespace paddle::framework;
   using namespace paddle::platform;

paddle/framework/type_defs.h

@@ -15,6 +15,7 @@
 #pragma once
 #include <functional>
 #include <map>
+#include <memory>
 #include "paddle/platform/variant.h"
 
 namespace paddle {

paddle/gserver/CMakeLists.txt

@@ -60,6 +60,36 @@ if(NOT WITH_PYTHON)
             dataproviders/PyDataProvider.h)
 endif()
 
+if(MOBILE_INFERENCE)
+    # Remove evaluators
+    list(REMOVE_ITEM GSERVER_SOURCES
+         layers/ValidationLayer.cpp
+         evaluators/Evaluator.cpp
+         evaluators/DetectionMAPEvaluator.cpp
+         evaluators/CTCErrorEvaluator.cpp
+         evaluators/ChunkEvaluator.cpp)
+
+    # Remove dataproviders
+    list(REMOVE_ITEM GSERVER_SOURCES
+         dataproviders/DataProvider.cpp
+         dataproviders/MultiDataProvider.cpp
+         dataproviders/ProtoDataProvider.cpp
+         dataproviders/PyDataProvider2.cpp
+         dataproviders/PyDataProvider.cpp)
+
+    # Remove useless gradientmachines
+    list(REMOVE_ITEM GSERVER_SOURCES
+         gradientmachines/MultiNetwork.cpp
+         gradientmachines/RecurrentGradientMachine.cpp
+         gradientmachines/ParallelNeuralNetwork.cpp
+         gradientmachines/GradientMachineMode.cpp
+         gradientmachines/MultiGradientMachine.cpp)
+
+    # Remove useless layers
+    list(REMOVE_ITEM GSERVER_SOURCES
+    	 layers/RecurrentLayerGroup.cpp)
+endif()
+
 if(WITH_GPU)
     cuda_add_library(paddle_gserver ${GSERVER_SOURCES})
 else()

paddle/gserver/gradientmachines/GradientMachine.cpp

@@ -17,12 +17,15 @@ limitations under the License. */
 #include <fstream>
 #include "paddle/utils/Logging.h"
 
+#include "NeuralNetwork.h"
+#include "hl_gpu.h"
+
+#ifndef PADDLE_MOBILE_INFERENCE
 #include "GradientMachineMode.h"
 #include "MultiGradientMachine.h"
 #include "MultiNetwork.h"
-#include "NeuralNetwork.h"
 #include "ParallelNeuralNetwork.h"
-#include "hl_gpu.h"
+#endif
 
 namespace paddle {
 
@@ -30,13 +33,16 @@ GradientMachine* GradientMachine::create(
     const ModelConfig& config,
     int mode,
     const std::vector<ParameterType>& parameterTypes) {
+#ifndef PADDLE_MOBILE_INFERENCE
   if (auto gm = IGradientMachineMode::tryCreateGradientMachine(mode, config)) {
     return gm;
   }
   if (FLAGS_trainer_count > 1) {
     return new MultiGradientMachine(config, FLAGS_use_gpu);
   }
+#endif
   if (FLAGS_trainer_count == 1) {  // single
+#ifndef PADDLE_MOBILE_INFERENCE
     NeuralNetwork* nn;
     if (config.type() == "multi_nn") {
       /* multi submodel calculate, thread(s) will be initialized inside */
@@ -48,6 +54,9 @@ GradientMachine* GradientMachine::create(
       /* single thread calculate */
       nn = NeuralNetwork::create(config);
     }
+#else
+    NeuralNetwork* nn = NeuralNetwork::create(config);
+#endif
     ParamInitCallback testParamInitCb = [](int paramId, Parameter* para) {
       para->enableType(PARAMETER_VALUE);
     };

paddle/gserver/gradientmachines/GradientMachine.h

@@ -20,13 +20,16 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "TrainerConfig.pb.h"
 #include "paddle/gserver/dataproviders/DataProvider.h"
-#include "paddle/gserver/evaluators/Evaluator.h"
 #include "paddle/gserver/layers/Layer.h"
 #include "paddle/math/Matrix.h"
 #include "paddle/parameter/Parameter.h"
 #include "paddle/parameter/ParameterUpdaterBase.h"
 #include "paddle/utils/Thread.h"
 
+#ifndef PADDLE_MOBILE_INFERENCE
+#include "paddle/gserver/evaluators/Evaluator.h"
+#endif
+
 namespace paddle {
 /**
  * @brief A gradient machine is capable of calculating some outputs given
@@ -147,6 +150,7 @@ public:
 
   virtual void onPassEnd() = 0;
 
+#ifndef PADDLE_MOBILE_INFERENCE
   /**
    * Create an evaluator which can be used for eval()
    */
@@ -156,6 +160,7 @@ public:
    * evaluate using the given evaluator
    */
   virtual void eval(Evaluator* evaluator) const = 0;
+#endif
 
   std::vector<ParameterPtr>& getParameters() { return parameters_; }