for merge

CMakeLists.txt

@@ -63,7 +63,28 @@ if(WIN32)
         set(CMAKE_C_FLAGS_RELEASE  "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT")
         set(CMAKE_CXX_FLAGS_DEBUG  "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd")
         set(CMAKE_CXX_FLAGS_RELEASE   "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
+        foreach(flag_var
+            CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+            CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO
+            CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
+            CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO)
+            if(${flag_var} MATCHES "/MD")
+                string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
             endif()
+        endforeach(flag_var)
+    endif()
+
+    # windows build turn off warnings.
+    foreach(flag_var
+        CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+        CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO
+        CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
+        CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO)
+        string(REGEX REPLACE "/W[1-4]" " /W0 " ${flag_var} "${${flag_var}}")
+    endforeach(flag_var)
+    foreach(flag_var CMAKE_CXX_FLAGS CMAKE_C_FLAGS)
+        set(${flag_var} "${${flag_var}} /w")
+    endforeach(flag_var)
 
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /wd4068 /wd4129 /wd4244 /wd4267 /wd4297 /wd4530 /wd4577 /wd4819 /wd4838 /MP")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4068 /wd4129 /wd4244 /wd4267 /wd4297 /wd4530 /wd4577 /wd4819 /wd4838 /MP")

cmake/external/cryptopp.cmake

@@ -22,23 +22,8 @@ SET(CRYPTOPP_TAG        CRYPTOPP_8_2_0)
 
 IF(WIN32)
   SET(CRYPTOPP_LIBRARIES "${CRYPTOPP_INSTALL_DIR}/lib/cryptopp-static.lib" CACHE FILEPATH "cryptopp library." FORCE)
-  SET(CRYPTOPP_CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
-  set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /MT")
-  set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /MTd")
-  set(CompilerFlags
-        CMAKE_CXX_FLAGS
-        CMAKE_CXX_FLAGS_DEBUG
-        CMAKE_CXX_FLAGS_RELEASE
-        CMAKE_C_FLAGS
-        CMAKE_C_FLAGS_DEBUG
-        CMAKE_C_FLAGS_RELEASE
-        )
-  foreach(CompilerFlag ${CompilerFlags})
-    string(REPLACE "/MD" "/MT" ${CompilerFlag} "${${CompilerFlag}}")
-  endforeach()
 ELSE(WIN32)
   SET(CRYPTOPP_LIBRARIES "${CRYPTOPP_INSTALL_DIR}/lib/libcryptopp.a" CACHE FILEPATH "cryptopp library." FORCE)
-  SET(CRYPTOPP_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS})
 ENDIF(WIN32)
 
 set(CRYPTOPP_CMAKE_ARGS ${COMMON_CMAKE_ARGS}
@@ -48,7 +33,7 @@ set(CRYPTOPP_CMAKE_ARGS ${COMMON_CMAKE_ARGS}
                         -DCMAKE_INSTALL_LIBDIR=${CRYPTOPP_INSTALL_DIR}/lib
                         -DCMAKE_INSTALL_PREFIX=${CRYPTOPP_INSTALL_DIR}
                         -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
-                        -DCMAKE_CXX_FLAGS=${CRYPTOPP_CMAKE_CXX_FLAGS}
+                        -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
                         -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
                         -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
                         -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}

cmake/flags.cmake

@@ -90,20 +90,6 @@ macro(safe_set_nvflag flag_name)
     endif()
 endmacro()
 
-macro(safe_set_static_flag) # set c_flags and cxx_flags to static or shared
-    if (BUILD_SHARED_LIBS) 
-        return() # if build shared libs, the flags keep same with '/MD'
-    endif(BUILD_SHARED_LIBS)
-    foreach(flag_var
-        CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
-        CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO
-        CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
-        CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO)
-      if(${flag_var} MATCHES "/MD")
-        string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
-      endif(${flag_var} MATCHES "/MD")
-    endforeach(flag_var)
-endmacro()
 
 CHECK_CXX_SYMBOL_EXISTS(UINT64_MAX "stdint.h" UINT64_MAX_EXISTS)
 if(NOT UINT64_MAX_EXISTS)
@@ -229,20 +215,3 @@ endforeach()
 
 set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${SAFE_GPU_COMMON_FLAGS}")
 
-
-if(WIN32)
-    # windows build turn off warnings.
-    if(MSVC_STATIC_CRT)
-        safe_set_static_flag()
-    endif()
-    foreach(flag_var
-        CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
-        CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO
-        CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
-        CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO)
-        string(REGEX REPLACE "/W[1-4]" " /W0 " ${flag_var} "${${flag_var}}")
-    endforeach(flag_var)
-    foreach(flag_var CMAKE_CXX_FLAGS CMAKE_C_FLAGS)
-        set(${flag_var} "${${flag_var}} /w")
-    endforeach(flag_var)
-endif()

cmake/generic.cmake

@@ -446,6 +446,9 @@ function(nv_library TARGET_NAME)
         message(FATAL "Please specify source file or library in nv_library.")
       endif()
     endif(nv_library_SRCS)
+    if (WIN32)
+      set_target_properties(${TARGET_NAME} PROPERTIES VS_USER_PROPS ${WIN_PROPS})
+    endif(WIN32)
   endif()
 endfunction(nv_library)
 
@@ -461,6 +464,9 @@ function(nv_binary TARGET_NAME)
       add_dependencies(${TARGET_NAME} ${nv_binary_DEPS})
       common_link(${TARGET_NAME})
     endif()
+    if (WIN32)
+      set_target_properties(${TARGET_NAME} PROPERTIES VS_USER_PROPS ${WIN_PROPS})
+    endif(WIN32)
   endif()
 endfunction(nv_binary)
 
@@ -482,6 +488,9 @@ function(nv_test TARGET_NAME)
     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 ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
+    if (WIN32)
+      set_target_properties(${TARGET_NAME} PROPERTIES VS_USER_PROPS ${WIN_PROPS})
+    endif(WIN32)
   endif()
 endfunction(nv_test)
 
@@ -712,6 +721,7 @@ function(proto_library TARGET_NAME)
   set(proto_hdrs)
   paddle_protobuf_generate_cpp(proto_srcs proto_hdrs ${proto_library_SRCS})
   cc_library(${TARGET_NAME} SRCS ${proto_srcs} DEPS ${proto_library_DEPS} protobuf)
+  add_dependencies(extern_xxhash ${TARGET_NAME})
 endfunction()
 
 function(py_proto_compile TARGET_NAME)

cmake/inference_lib.cmake

@@ -19,12 +19,11 @@ set(PADDLE_INSTALL_DIR "${CMAKE_BINARY_DIR}/paddle_install_dir" CACHE STRING
 set(PADDLE_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/paddle_inference_install_dir" CACHE STRING
   "A path setting paddle inference shared and static libraries")
   
-# TODO(zhaolong)
-# At present, the size of static lib in Windows exceeds the system limit,
-# so the generation of static lib is temporarily turned off.
+# At present, the size of static lib in Windows is very large,
+# so we need to crop the library size.
 if(WIN32)
     #todo: remove the option 
-    option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static."   OFF)
+    option(WITH_STATIC_LIB "Compile demo with static/shared library, default use dynamic."   OFF)
     if(NOT PYTHON_EXECUTABLE)
         FIND_PACKAGE(PythonInterp REQUIRED)
     endif()
@@ -165,25 +164,22 @@ copy_part_of_thrid_party(inference_lib_dist ${PADDLE_INFERENCE_INSTALL_DIR})
 set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
 if(WIN32)
     if(WITH_STATIC_LIB)
-        set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/${CMAKE_BUILD_TYPE}/libpaddle_fluid.lib)
+        set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/${CMAKE_BUILD_TYPE}/libpaddle_fluid.lib
+                             ${PADDLE_BINARY_DIR}/paddle/fluid/inference/${CMAKE_BUILD_TYPE}/paddle_fluid.*)
     else()
         set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/${CMAKE_BUILD_TYPE}/paddle_fluid.dll
                              ${PADDLE_BINARY_DIR}/paddle/fluid/inference/${CMAKE_BUILD_TYPE}/paddle_fluid.lib)
     endif()
-else(WIN32)
-    set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*)
-endif(WIN32)
-
-if(WIN32 AND NOT WITH_STATIC_LIB)
     copy(inference_lib_dist
             SRCS  ${src_dir}/inference/api/paddle_*.h ${paddle_fluid_lib}
             DSTS  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/lib
             ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/lib)
-else()
+else(WIN32)
+    set(paddle_fluid_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*)
     copy(inference_lib_dist
                 SRCS  ${src_dir}/inference/api/paddle_*.h ${paddle_fluid_lib}
                 DSTS  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/lib)
-endif()
+endif(WIN32)
 
 copy(inference_lib_dist
         SRCS  ${CMAKE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
@@ -199,7 +195,11 @@ set(PADDLE_INFERENCE_C_INSTALL_DIR "${CMAKE_BINARY_DIR}/paddle_inference_c_insta
 copy_part_of_thrid_party(inference_lib_dist ${PADDLE_INFERENCE_C_INSTALL_DIR})
 
 set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
-set(paddle_fluid_c_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/capi/libpaddle_fluid_c.*)
+if(WIN32)
+  set(paddle_fluid_c_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/capi/${CMAKE_BUILD_TYPE}/paddle_fluid_c.*)
+else(WIN32)
+  set(paddle_fluid_c_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/capi/libpaddle_fluid_c.*)
+endif(WIN32)
 
 copy(inference_lib_dist
       SRCS  ${src_dir}/inference/capi/paddle_c_api.h  ${paddle_fluid_c_lib}
@@ -211,12 +211,12 @@ add_custom_target(fluid_lib_dist ALL DEPENDS ${fluid_lib_deps})
 
 set(dst_dir "${PADDLE_INSTALL_DIR}/paddle/fluid")
 set(module "inference")
-if(WIN32 AND NOT WITH_STATIC_LIB)
+if(WIN32)
         copy(fluid_lib_dist
                 SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/api/paddle_*.h ${paddle_fluid_lib}
                 DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
                 )
-else()
+        else()
         copy(fluid_lib_dist
                 SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/api/paddle_*.h ${paddle_fluid_lib}
                 DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} 

cmake/init.cmake

@@ -26,4 +26,7 @@ if(WITH_GPU)
     set(CMAKE_CUDA_FLAGS_MINSIZEREL "-O1 -DNDEBUG")
 endif()
 
+if(WIN32)
+    set(WIN_PROPS ${CMAKE_SOURCE_DIR}/cmake/paddle_win.props)
+endif()
 

cmake/operators.cmake

@@ -127,7 +127,8 @@ function(op_library TARGET)
 "tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op"
 "fusion_transpose_flatten_concat_op" "fusion_conv_inception_op"
 "sync_batch_norm_op" "dgc_op" "fused_fc_elementwise_layernorm_op"
-"multihead_matmul_op" "fusion_group_op" "fused_bn_activation_op" "fused_embedding_eltwise_layernorm_op" "fusion_gru_op")
+"multihead_matmul_op" "fusion_group_op" "fused_bn_activation_op" "fused_embedding_eltwise_layernorm_op" "fusion_gru_op"
+"fused_bn_add_activation_op")
         if ("${TARGET}" STREQUAL "${manual_pybind_op}")
             set(pybind_flag 1)
         endif()

cmake/paddle_win.props

+<?xml version="1.0" encoding="utf-8"?>
+<Project xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+    <ItemDefinitionGroup>
+        <CudaCompile>
+            <!-- Project schema: Host properties -->
+            <UseHostDefines>true</UseHostDefines>
+            <Emulation>false</Emulation>
+            <HostDebugInfo Condition="'$(Configuration)' == 'Debug'">true</HostDebugInfo>
+            <HostDebugInfo Condition="'$(Configuration)' != 'Debug'">false</HostDebugInfo>
+            <FastMath>false</FastMath>
+            <Optimization>InheritFromHost</Optimization>
+            <Runtime>InheritFromHost</Runtime>
+            <RuntimeChecks>InheritFromHost</RuntimeChecks>
+            <TypeInfo>InheritFromHost</TypeInfo>
+            <Warning>InheritFromHost</Warning>
+
+            <BaseCommandLineTemplate>-ccbin "%(VCBinDir)" -x cu [GenerateRelocatableDeviceCode] [Include] [RequiredIncludes] [InterleaveSourceInPTX] [GPUDebugInfo] [GenerateLineInfo] [Keep] [KeepDir] [MaxRegCount] [PtxAsOptionV] [TargetMachinePlatform] [NvccCompilation] [CudaRuntime] [AdditionalOptions]</BaseCommandLineTemplate>
+            <BuildCommandLineTemplate>--use-local-env</BuildCommandLineTemplate>
+            <BuildDynamicCommandLineTemplate>[CodeGeneration]</BuildDynamicCommandLineTemplate>
+            <CleanCommandLineTemplate>-clean</CleanCommandLineTemplate>
+            <!-- <HostCommandLineTemplate>-Xcompiler &quot;/EHsc [Warning] /nologo [Optimization] [ProgramDataBaseFileName] $(CudaForceSynchronousPdbWrites) /Zi [RuntimeChecks] [Runtime] [TypeInfo]&quot;</HostCommandLineTemplate> -->
+            <HostCommandLineTemplate>-Xcompiler &quot;/EHsc [Warning] /nologo [Optimization] [ProgramDataBaseFileName] $(CudaForceSynchronousPdbWrites) [RuntimeChecks] [Runtime] [TypeInfo]&quot;</HostCommandLineTemplate>
+
+            <DriverApiCommandLineTemplate>%(BaseCommandLineTemplate) [CompileOut] "%(FullPath)"</DriverApiCommandLineTemplate>
+            <RuntimeApiCommandLineTemplate>%(BaseCommandLineTemplate) [HostDebugInfo] [Emulation] [FastMath] [Defines] %(HostCommandLineTemplate) [CompileOut] "%(FullPath)"</RuntimeApiCommandLineTemplate>
+
+            <CommandLineTemplate>
+# (Approximate command-line.  Settings inherited from host are not visible below.)
+# (Please see the output window after a build for the full command-line)
+
+# Driver API (NVCC Compilation Type is .cubin, .gpu, or .ptx)
+set CUDAFE_FLAGS=--sdk_dir "$(WindowsSdkDir)"
+"$(CudaToolkitNvccPath)" %(BuildCommandLineTemplate) %(DriverApiCommandLineTemplate)
+
+# Runtime API (NVCC Compilation Type is hybrid object or .c file)
+set CUDAFE_FLAGS=--sdk_dir "$(WindowsSdkDir)"
+"$(CudaToolkitNvccPath)" %(BuildCommandLineTemplate) %(RuntimeApiCommandLineTemplate)
+            </CommandLineTemplate>
+            <ExecutionDescription>Compiling CUDA source file %(Identity)...</ExecutionDescription>
+            <ExclusionDescription>Skipping CUDA source file %(Identity) (excluded from build).</ExclusionDescription>
+
+            <!-- Miscellaneous -->
+            <PropsCacheOutputFile>%(Filename)%(Extension).cache</PropsCacheOutputFile>
+            <PropsCacheOutputPath>$(IntDir)%(PropsCacheOutputFile)</PropsCacheOutputPath>
+
+            <CudaCompileCoreProject>$(MSBuildProjectFullPath)</CudaCompileCoreProject>
+        </CudaCompile>
+
+        <CudaLink>
+            <PerformDeviceLink>true</PerformDeviceLink>
+            <LinkOut>$(IntDir)$(TargetName).device-link.obj</LinkOut>
+
+            <AdditionalLibraryDirectories></AdditionalLibraryDirectories>
+            <UseHostLibraryDirectories>true</UseHostLibraryDirectories>
+            <AdditionalDependencies></AdditionalDependencies>
+            <UseHostLibraryDependencies>true</UseHostLibraryDependencies>
+
+            <GPUDebugInfo>InheritFromProject</GPUDebugInfo>
+            <Optimization>InheritFromProject</Optimization>
+
+            <!-- Implicitly inherited from the project via @(CudaCompile) -->
+            <CodeGeneration></CodeGeneration>
+            <RuntimeChecks></RuntimeChecks>
+            <Runtime></Runtime>
+            <TargetMachinePlatform></TargetMachinePlatform>
+            <TypeInfo></TypeInfo>
+            <Warning></Warning>
+
+            <Inputs></Inputs>
+
+            <!-- <HostCommandLineTemplate>-Xcompiler &quot;/EHsc [Warning] /nologo [Optimization] /Zi [RuntimeChecks] [Runtime] [TypeInfo]&quot;</HostCommandLineTemplate> -->
+            <HostCommandLineTemplate>-Xcompiler &quot;/EHsc [Warning] /nologo [Optimization] [RuntimeChecks] [Runtime] [TypeInfo]&quot;</HostCommandLineTemplate>
+            <LinkCommandLineTemplate>"$(CudaToolkitNvccPath)" -dlink [LinkOut] %(HostCommandLineTemplate) [AdditionalLibraryDirectories] [AdditionalDependencies] [AdditionalOptions] [CodeGeneration] [GPUDebugInfo] [TargetMachinePlatform] [Inputs]</LinkCommandLineTemplate>
+            <CommandLineTemplate>
+# (Approximate command-line.  Settings inherited from host are not visible below.)
+# (Please see the output window after a build for the full command-line)
+
+%(LinkCommandLineTemplate)
+            </CommandLineTemplate>
+        </CudaLink>
+
+        <Link>
+            <AdditionalLibraryDirectories>%(AdditionalLibraryDirectories);$(CudaToolkitLibDir)</AdditionalLibraryDirectories>
+        </Link>
+
+        <ClCompile>
+            <AdditionalIncludeDirectories>%(AdditionalIncludeDirectories);$(CudaToolkitIncludeDir)</AdditionalIncludeDirectories>
+        </ClCompile>
+    </ItemDefinitionGroup>
+</Project>
+

cmake/third_party.cmake

@@ -39,6 +39,7 @@ set(third_party_deps)
 #            REPOSITORY ${TARGET_REPOSITORY}
 #            TAG        ${TARGET_TAG}
 #            DIR        ${TARGET_SOURCE_DIR})
+
 FUNCTION(cache_third_party TARGET)
     SET(options "")
     SET(oneValueArgs URL REPOSITORY TAG DIR)
@@ -269,6 +270,10 @@ if(WITH_PSLIB)
     endif()
 endif(WITH_PSLIB)
 
+if(NOT WIN32 AND NOT APPLE)
+    include(external/gloo)
+    list(APPEND third_party_deps extern_gloo)
+endif()
 
 if(WITH_BOX_PS)
     include(external/box_ps)
@@ -276,10 +281,6 @@ if(WITH_BOX_PS)
 endif(WITH_BOX_PS)
 
 if(WITH_DISTRIBUTE)
-    if(WITH_GLOO)
-        include(external/gloo)
-        list(APPEND third_party_deps extern_gloo)
-    endif()
 
     if(WITH_GRPC)
         list(APPEND third_party_deps extern_grpc)

paddle/fluid/framework/attribute.cc

@@ -14,8 +14,6 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/attribute.h"
 
-#include <vector>
-
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/block_desc.h

@@ -30,6 +30,8 @@ namespace paddle {
 namespace framework {
 
 class ProgramDesc;
+class OpDesc;
+class VarDesc;
 
 // Each Protobuf Message, we provide a XXXBind class. In that class, we optimize
 // read/write speed. Only when we want the protobuf message, the local changes

paddle/fluid/framework/c/c_api.cc

@@ -12,17 +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 <string>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/c/c_api.h"
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/program_desc.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/platform/init.h"
 
 extern "C" {
 

paddle/fluid/framework/c/c_api.h

@@ -24,6 +24,15 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+class OpInfoMap;
+}  // namespace framework
+namespace platform {
+class DeviceContextPool;
+}  // namespace platform
+}  // namespace paddle
+
 #ifdef __cplusplus
 extern "C" {
 #endif

paddle/fluid/framework/channel.h

@@ -277,7 +277,7 @@ class ChannelObject {
     size_t finished = 0;
     while (finished < n && WaitForWrite(lock)) {
       size_t m =
-          std::min(n - finished, capacity_ + reading_count_ - data_.size());
+          (std::min)(n - finished, capacity_ + reading_count_ - data_.size());
       for (size_t i = 0; i < m; i++) {
         data_.push_back(std::move(p[finished++]));
       }

paddle/fluid/framework/copy_same_tensor_test.cc

@@ -12,8 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include <cstring>
 #include <random>
+
 #include "gflags/gflags.h"
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/tensor.h"

paddle/fluid/framework/data_device_transform.h

@@ -21,6 +21,8 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class Tensor;
+
 void TransDataDevice(const Tensor& in, const platform::Place& dst_place,
                      Tensor* out);
 

paddle/fluid/framework/data_feed.h

@@ -41,6 +41,15 @@ limitations under the License. */
 #include "paddle/fluid/framework/variable.h"
 #include "paddle/fluid/string/string_helper.h"
 
+namespace paddle {
+namespace framework {
+class DataFeedDesc;
+class LoDTensor;
+class Scope;
+class Variable;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 
@@ -418,6 +427,7 @@ class MultiSlotType {
 
   std::string DebugString() {
     std::stringstream ss;
+
     ss << "\ntype: " << type_ << "\n";
     ss << "offset: ";
     ss << "[";

paddle/fluid/framework/data_feed_factory.cc

@@ -17,10 +17,10 @@ limitations under the License. */
 #include <string>
 #include <unordered_map>
 
-#include "paddle/fluid/framework/data_feed.h"
-
 namespace paddle {
 namespace framework {
+class DataFeed;
+
 typedef std::shared_ptr<DataFeed> (*Createdata_feedFunction)();
 typedef std::unordered_map<std::string, Createdata_feedFunction> data_feedMap;
 data_feedMap g_data_feed_map;

paddle/fluid/framework/data_feed_factory.h

@@ -16,10 +16,13 @@ limitations under the License. */
 
 #include <memory>
 #include <string>
+
 #include "paddle/fluid/framework/data_feed.h"
 
 namespace paddle {
 namespace framework {
+class DataFeed;
+
 class DataFeedFactory {
  public:
   static std::string DataFeedTypeList();

paddle/fluid/framework/data_layout_transform.cc

@@ -13,8 +13,8 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/data_layout_transform.h"
+
 #include <string>
-#include <vector>
 
 #include "paddle/fluid/operators/math/math_function.h"
 #ifdef PADDLE_WITH_MKLDNN

paddle/fluid/framework/data_layout_transform.h

@@ -17,10 +17,18 @@
 #include <map>
 #include <unordered_map>
 #include <vector>
+
 #include "paddle/fluid/framework/op_kernel_type.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/variable.h"
 
+namespace paddle {
+namespace framework {
+class OpKernelType;
+class Tensor;
+}  // namespace framework
+}  // namespace paddle
+
 #ifdef PADDLE_WITH_MKLDNN
 #include "paddle/fluid/platform/mkldnn_helper.h"
 #endif

paddle/fluid/framework/data_layout_transform_test.cc

@@ -15,7 +15,6 @@
 #include "paddle/fluid/framework/data_layout_transform.h"
 
 #include "gtest/gtest.h"
-#include "paddle/fluid/platform/device_context.h"
 
 TEST(DataTransform, DataLayoutFunction) {
   auto place = paddle::platform::CPUPlace();

paddle/fluid/framework/data_transform.cc

@@ -18,8 +18,13 @@ limitations under the License. */
 #include "paddle/fluid/framework/data_layout_transform.h"
 #include "paddle/fluid/framework/data_type_transform.h"
 
+namespace paddle {
+namespace framework {
+class Variable;
+}  // namespace framework
+}  // namespace paddle
+
 #ifdef PADDLE_WITH_MKLDNN
-#include <algorithm>
 #include "paddle/fluid/platform/mkldnn_helper.h"
 #endif
 

paddle/fluid/framework/data_transform.h

@@ -30,6 +30,10 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class OpKernelType;
+class Tensor;
+class Variable;
+
 void TransformData(const OpKernelType &expected_kernel_type,
                    const OpKernelType &kernel_type_for_var,
                    const Tensor &input_tensor, Tensor *out);

paddle/fluid/framework/data_type.cc

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/data_type.h"
-#include <stdint.h>
 #include <string>
 #include <unordered_map>
 

paddle/fluid/framework/data_type.h

@@ -15,12 +15,19 @@ limitations under the License. */
 #pragma once
 #include <string>
 #include <typeindex>
-#include "paddle/fluid/framework/framework.pb.h"
-#include "paddle/fluid/platform/enforce.h"
 
+#include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/platform/bfloat16.h"
+#include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/float16.h"
 
+namespace paddle {
+namespace platform {
+struct bfloat16;
+struct float16;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/data_type_test.cc

@@ -14,9 +14,17 @@
 #include "paddle/fluid/framework/data_type.h"
 
 #include <string>
+
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/tensor.h"
 
+namespace paddle {
+namespace platform {
+struct bfloat16;
+struct float16;
+}  // namespace platform
+}  // namespace paddle
+
 TEST(DataType, float16) {
   using paddle::framework::Tensor;
   using paddle::platform::CPUPlace;

paddle/fluid/framework/data_type_transform.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <utility>
+
 #include "paddle/fluid/framework/op_kernel_type.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/variable.h"
@@ -23,6 +24,9 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class OpKernelType;
+class Tensor;
+
 using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
 
 void TransDataType(const OpKernelType& kernel_type_for_var,

paddle/fluid/framework/dataset_factory.cc

@@ -13,12 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/dataset_factory.h"
-#include <memory>
 #include <string>
 #include <unordered_map>
 
-#include "paddle/fluid/framework/data_set.h"
-
 namespace paddle {
 namespace framework {
 typedef std::unique_ptr<Dataset> (*CreateDatasetFunction)();

paddle/fluid/framework/dataset_factory.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <memory>
 #include <string>
+
 #include "paddle/fluid/framework/data_set.h"
 
 namespace paddle {

paddle/fluid/framework/ddim.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <stdexcept>
 #include <string>
 #include <vector>
+
 #include "paddle/fluid/framework/dim.h"
 
 namespace paddle {

paddle/fluid/framework/ddim_test.cc

@@ -12,7 +12,6 @@ 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 <sstream>
-#include <vector>
 
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/ddim.h"

paddle/fluid/framework/details/all_reduce_op_handle.h

@@ -20,6 +20,17 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class NCCLCommunicator;
+}  // namespace platform
+}  // namespace paddle
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/framework/details/nccl_op_handle.h"
 #include "paddle/fluid/platform/nccl_helper.h"

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -24,6 +24,20 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+namespace details {
+struct VarHandle;
+}  // namespace details
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+struct NCCLContextMap;
+}  // namespace platform
+}  // namespace paddle
+
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/framework/details/broadcast_op_handle_test.h

@@ -21,13 +21,15 @@
 
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/details/broadcast_op_handle.h"
-
 #include "paddle/fluid/platform/device_context.h"
 
 namespace paddle {
 namespace framework {
 namespace details {
 
+struct DummyVarHandle;
+struct VarHandle;
+
 namespace f = paddle::framework;
 namespace p = paddle::platform;
 

paddle/fluid/framework/details/build_strategy.h

@@ -27,6 +27,18 @@
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Graph;
+class PassBuilder;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class NCCLCommunicator;
+}  // namespace platform
+}  // namespace paddle
+
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/framework/details/computation_op_handle.cc

@@ -19,6 +19,8 @@
 namespace paddle {
 namespace framework {
 namespace details {
+struct VarHandleBase;
+
 ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope,
                                          platform::Place place,
                                          size_t scope_idx)

paddle/fluid/framework/details/computation_op_handle.h

@@ -24,9 +24,21 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+class OperatorBase;
+class Scope;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {
+struct VarHandleBase;
+
 class ComputationOpHandle : public OpHandleBase {
  public:
   ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,

paddle/fluid/framework/details/eager_deletion_op_handle.h

@@ -19,12 +19,23 @@
 #include <string>
 #include <unordered_set>
 #include <vector>
+
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/ir/memory_optimize_pass/reference_count_pass_helper.h"
 
+namespace paddle {
+namespace platform {
+class CUDADeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 class Scope;
+class GarbageCollector;
+namespace ir {
+class Node;
+}  // namespace ir
 
 namespace ir {
 class MemOptVarInfo;

paddle/fluid/framework/details/exception_holder_test.cc

@@ -13,8 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/exception_holder.h"
-#include <memory>
-#include <unordered_map>
+
 #include "gtest/gtest.h"
 #include "paddle/fluid/memory/allocation/allocator.h"
 

paddle/fluid/framework/details/fetch_async_op_handle.cc

@@ -15,9 +15,14 @@
 #include "paddle/fluid/framework/details/fetch_async_op_handle.h"
 #include <string>
 #include <utility>
-#include <vector>
 #include "paddle/fluid/platform/profiler.h"
 
+namespace paddle {
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/fetch_async_op_handle.h

@@ -22,6 +22,18 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+class LoDTensor;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/fetch_barrier_op_handle.cc

@@ -19,6 +19,8 @@
 namespace paddle {
 namespace framework {
 namespace details {
+struct VarHandleBase;
+
 FetchBarrierOpHandle::FetchBarrierOpHandle(
     ir::Node *node, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places)

paddle/fluid/framework/details/fetch_barrier_op_handle.h

@@ -24,6 +24,15 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {
@@ -32,6 +41,8 @@ namespace details {
 // all places if there are multiple places, must init with
 // multiple dev_ctxes_ !!!!
 
+struct VarHandleBase;
+
 struct FetchBarrierOpHandle : public OpHandleBase {
  public:
   FetchBarrierOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,

paddle/fluid/framework/details/fetch_op_handle.h

@@ -22,6 +22,17 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/fused_all_reduce_op_handle.h

@@ -17,10 +17,22 @@
 #include <string>
 #include <utility>
 #include <vector>
+
 #include "paddle/fluid/framework/details/all_reduce_op_handle.h"
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class NCCLCommunicator;
+}  // namespace platform
+}  // namespace paddle
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/framework/details/nccl_op_handle.h"
 #include "paddle/fluid/platform/nccl_helper.h"

paddle/fluid/framework/details/fused_broadcast_op_handle.h

@@ -25,6 +25,17 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+struct NCCLContextMap;
+}  // namespace platform
+}  // namespace paddle
+
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc

@@ -17,11 +17,20 @@
 #include <unordered_map>
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/details/broadcast_op_handle_test.h"
+#include "paddle/fluid/framework/details/op_handle_base.h"
+
+namespace paddle {
+namespace framework {
+class Scope;
+}  // namespace framework
+}  // namespace paddle
 
 namespace paddle {
 namespace framework {
 namespace details {
 
+struct VarHandle;
+
 struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
   std::vector<std::string> out_varnames_;
   std::vector<std::unique_ptr<ir::Node>> nodes_;

paddle/fluid/framework/details/gather_op_handle.h

@@ -24,6 +24,14 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/gather_op_handle_test.cc

@@ -17,11 +17,11 @@
 #include <unordered_map>
 #include "gtest/gtest.h"
 
-#include "paddle/fluid/platform/device_context.h"
-
 namespace paddle {
 namespace framework {
 namespace details {
+struct DummyVarHandle;
+
 namespace f = paddle::framework;
 namespace p = paddle::platform;
 

paddle/fluid/framework/details/multi_devices_helper.h

@@ -20,16 +20,21 @@
 #include <unordered_set>
 #include <utility>
 #include <vector>
+
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/details/var_handle.h"
-
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/pass.h"
 #include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/platform/place.h"
 
-#include "paddle/fluid/framework/ir/graph.h"
-#include "paddle/fluid/framework/ir/pass.h"
+namespace paddle {
+namespace framework {
+class OpDesc;
+}  // namespace framework
+}  // namespace paddle
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/details/nan_inf_utils_detail.h

@@ -19,6 +19,12 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/place.h"
 
+namespace paddle {
+namespace framework {
+class Tensor;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/op_handle_base.h

@@ -24,10 +24,22 @@
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/macros.h"
 
+namespace paddle {
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 
 class Scope;
+namespace details {
+struct VarHandleBase;
+}  // namespace details
+namespace ir {
+class Node;
+}  // namespace ir
 
 namespace details {
 

paddle/fluid/framework/details/reduce_op_handle.h

@@ -24,6 +24,21 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+class SelectedRows;
+namespace details {
+struct VarHandle;
+}  // namespace details
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+struct NCCLContextMap;
+}  // namespace platform
+}  // namespace paddle
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/framework/details/rpc_op_handle.h

@@ -24,6 +24,16 @@
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/scope.h"
 
+namespace paddle {
+namespace framework {
+class OpDesc;
+class Scope;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/scale_loss_grad_op_handle.cc

@@ -13,9 +13,17 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
+
 #include <string>
+
 #include "paddle/fluid/platform/profiler.h"
 
+namespace paddle {
+namespace framework {
+class Tensor;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/scale_loss_grad_op_handle.h

@@ -21,6 +21,18 @@
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/scope_buffered_monitor.h

@@ -17,7 +17,9 @@
 #include <unordered_set>
 #include <utility>
 #include <vector>
+
 #include "paddle/fluid/framework/scope.h"
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/share_tensor_buffer_functor.cc

@@ -22,6 +22,17 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/enforce.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+class Tensor;
+class Variable;
+namespace ir {
+class MemOptVarInfo;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/share_tensor_buffer_functor.h

@@ -25,6 +25,15 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/variable.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+namespace ir {
+class MemOptVarInfo;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/share_tensor_buffer_op_handle.cc

@@ -23,10 +23,20 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/enforce.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class MemOptVarInfo;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {
 
+class ComputationOpHandle;
+
 ComputationOpHandle *GetUniquePendingComputationOpHandle(
     ShareTensorBufferOpHandle *share_tensor_op) {
   ComputationOpHandle *result_op = nullptr;

paddle/fluid/framework/details/share_tensor_buffer_op_handle.h

@@ -22,10 +22,22 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/details/share_tensor_buffer_functor.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+namespace ir {
+class MemOptVarInfo;
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {
 
+class ComputationOpHandle;
+
 class ShareTensorBufferOpHandle : public OpHandleBase {
  public:
   ShareTensorBufferOpHandle(

paddle/fluid/framework/details/sparse_all_reduce_op_handle.h

@@ -23,6 +23,17 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class NCCLCommunicator;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/var_handle.h

@@ -24,6 +24,14 @@
 #include "paddle/fluid/platform/macros.h"
 #include "paddle/fluid/platform/place.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/variable_visitor.cc

@@ -15,6 +15,14 @@
 #include "paddle/fluid/framework/details/variable_visitor.h"
 
 #include "paddle/fluid/framework/selected_rows.h"
+
+namespace paddle {
+namespace framework {
+class LoDTensor;
+class Variable;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/variable_visitor.h

@@ -17,6 +17,13 @@
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/variable.h"
 
+namespace paddle {
+namespace framework {
+class Tensor;
+class Variable;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/device_worker.cc

@@ -13,11 +13,13 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/device_worker.h"
-#include "xxhash.h"  // NOLINT
 
 namespace paddle {
 namespace framework {
 
+class LoDTensor;
+class Scope;
+
 void DeviceWorker::SetRootScope(Scope* root_scope) { root_scope_ = root_scope; }
 
 void DeviceWorker::SetDataFeed(DataFeed* data_feed) {

paddle/fluid/framework/device_worker.h

@@ -39,6 +39,18 @@ limitations under the License. */
 #include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/platform/timer.h"
 
+namespace paddle {
+namespace framework {
+class LoDTensor;
+class ProgramDesc;
+class Scope;
+class Tensor;
+}  // namespace framework
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
@@ -62,7 +74,9 @@ class PullDenseWorker {
   virtual void Initialize(const TrainerDesc& param);
 #ifdef PADDLE_WITH_CUDA
   void AddStream(const cudaStream_t stream) { copy_streams_.push_back(stream); }
+#endif
 
+#if (defined PADDLE_WITH_CUDA) || (defined PADDLE_WITH_XPU)
   void AddPlace(const paddle::platform::Place place) {
     places_.push_back(place);
   }
@@ -123,9 +137,9 @@ class PullDenseWorker {
 
 #ifdef PADDLE_WITH_CUDA
   std::vector<cudaStream_t> copy_streams_;
+#endif
   std::vector<paddle::platform::Place> places_;
   std::vector<Scope*> thread_scopes_;
-#endif
 };
 
 // should incorporate different type of device
@@ -149,6 +163,7 @@ class DeviceWorker {
   virtual void SetDataFeed(DataFeed* data_feed);
   virtual void SetWorkerNum(int num) {}
   virtual void CacheProgram(const ProgramDesc& main_program) {}
+  virtual void GetXpuOpIndex() {}
   virtual void SetNeedDumpField(bool need_dump_field) {
     need_dump_field_ = need_dump_field;
   }

paddle/fluid/framework/device_worker_factory.cc

@@ -20,6 +20,8 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class DeviceWorker;
+
 typedef std::shared_ptr<DeviceWorker> (*Createdevice_workerFunction)();
 typedef std::unordered_map<std::string, Createdevice_workerFunction>
     device_workerMap;

paddle/fluid/framework/device_worker_factory.h

@@ -16,11 +16,14 @@ limitations under the License. */
 
 #include <memory>
 #include <string>
+
 #include "paddle/fluid/framework/device_worker.h"
 
 namespace paddle {
 namespace framework {
 
+class DeviceWorker;
+
 class DeviceWorkerFactory {
  public:
   static std::string DeviceWorkerTypeList();

paddle/fluid/framework/device_worker_test.cc

@@ -13,9 +13,10 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/device_worker.h"
+
 #include <gtest/gtest.h>
+
 #include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/trainer.h"
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/distributed_strategy.proto

@@ -128,6 +128,7 @@ message DistributedStrategy {
   optional int32 conv_workspace_size_limit = 22 [ default = 4000 ];
   optional bool cudnn_batchnorm_spatial_persistent = 23 [ default = true ];
   optional bool adaptive_localsgd = 24 [ default = false ];
+  optional bool fp16_allreduce = 25 [ default = false ];
 
   optional RecomputeConfig recompute_configs = 101;
   optional AMPConfig amp_configs = 102;

paddle/fluid/framework/dlpack_tensor.cc

@@ -11,10 +11,17 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
+#include "paddle/fluid/framework/dlpack_tensor.h"
 #include <unordered_map>
-
 #include "paddle/fluid/framework/data_type.h"
-#include "paddle/fluid/framework/dlpack_tensor.h"
+
+namespace paddle {
+namespace platform {
+struct bfloat16;
+struct float16;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/dlpack_tensor.h

@@ -15,11 +15,14 @@
 #pragma once
 
 #include <dlpack/dlpack.h>
+
 #include "paddle/fluid/framework/tensor.h"
 
 namespace paddle {
 namespace framework {
 
+class Tensor;
+
 class DLPackTensor {
  public:
   using LaneType = decltype(::DLTensor::dtype.lanes);  // uint16_t

paddle/fluid/framework/dlpack_tensor_test.cc

@@ -17,6 +17,12 @@
 #include <gtest/gtest.h>
 #include <vector>
 
+namespace paddle {
+namespace platform {
+struct float16;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/downpour_worker.cc

@@ -13,10 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/device_worker.h"
-#include "paddle/fluid/framework/device_worker_factory.h"
-#include "paddle/fluid/framework/fleet/fleet_wrapper.h"
 #include "paddle/fluid/platform/cpu_helper.h"
-#include "paddle/fluid/string/string_helper.h"
+
+namespace paddle {
+namespace framework {
+class LoDTensor;
+class Variable;
+}  // namespace framework
+}  // namespace paddle
 
 #if defined _WIN32 || defined __APPLE__
 #else

paddle/fluid/framework/downpour_worker_opt.cc

@@ -12,18 +12,16 @@ 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 <set>
-#include <unordered_map>
-#include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/device_worker.h"
-#include "paddle/fluid/framework/device_worker_factory.h"
-#include "paddle/fluid/framework/fleet/fleet_wrapper.h"
 #include "paddle/fluid/platform/cpu_helper.h"
-#include "paddle/fluid/platform/lodtensor_printer.h"
 
 namespace paddle {
 namespace framework {
 
+class OpDesc;
+class OperatorBase;
+class ProgramDesc;
+
 bool HasDependentOutput(const OpDesc& op_desc,
                         const std::unordered_set<std::string>& dependent_vars) {
   for (auto& var : op_desc.Outputs()) {

paddle/fluid/framework/eigen.h

@@ -14,6 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include <stdint.h>
+
 #include "paddle/fluid/framework/tensor.h"
 #include "unsupported/Eigen/CXX11/Tensor"
 

paddle/fluid/framework/executor.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <string>
 #include <unordered_map>
 #include <vector>
+
 #include "paddle/fluid/framework/data_set.h"
 #include "paddle/fluid/framework/executor_gc_helper.h"
 #include "paddle/fluid/framework/garbage_collector.h"
@@ -32,6 +33,11 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class Dataset;
+class ProgramDesc;
+class Scope;
+class TrainerBase;
+
 struct ExecutorPrepareContext {
   ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id);
 

paddle/fluid/framework/executor_gc_helper.cc

@@ -13,16 +13,19 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/executor_gc_helper.h"
+
 #include <deque>
 #include <string>
-#include <unordered_map>
 #include <unordered_set>
 #include <utility>
-#include <vector>
+
 #include "glog/logging.h"
-#include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/lod_tensor_array.h"
-#include "paddle/fluid/framework/selected_rows.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/framework/framework.pb.h"
+#include "paddle/fluid/framework/no_need_buffer_vars_inference.h"
+#include "paddle/fluid/framework/op_info.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/var_desc.h"
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {

paddle/fluid/framework/executor_gc_helper.h

@@ -18,6 +18,7 @@
 #include <string>
 #include <unordered_map>
 #include <vector>
+
 #include "paddle/fluid/framework/garbage_collector.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/scope.h"
@@ -26,6 +27,10 @@ namespace paddle {
 namespace framework {
 
 // Result map: op -> variable names that can be deleted after op runs
+class GarbageCollector;
+class OperatorBase;
+class Scope;
+
 std::unordered_map<const OperatorBase *, std::vector<std::string>>
 GetUnusedVars(const BlockDesc &block,
               const std::vector<std::unique_ptr<OperatorBase>> &ops,

paddle/fluid/framework/feed_fetch_method.cc

@@ -13,16 +13,17 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/feed_fetch_method.h"
+
 #include <string>
-#include <vector>
+
 #include "glog/logging.h"
-#include "paddle/fluid/framework/var_type.h"
-#include "paddle/fluid/framework/variable.h"
-#include "paddle/fluid/platform/place.h"
 
 namespace paddle {
 namespace framework {
 
+class LoDTensor;
+class Variable;
+
 void SetFeedVariable(Scope* scope, const LoDTensor& input,
                      const std::string& var_name, size_t index) {
   // If var_name Variable is not found in GlobalScope, a new variable will

paddle/fluid/framework/feed_fetch_method.h

@@ -15,12 +15,16 @@ limitations under the License. */
 #pragma once
 
 #include <string>
+
 #include "paddle/fluid/framework/feed_fetch_type.h"
 #include "paddle/fluid/framework/scope.h"
 
 namespace paddle {
 namespace framework {
 
+class LoDTensor;
+class Scope;
+
 void SetFeedVariable(Scope* scope, const LoDTensor& input,
                      const std::string& var_name, size_t index);
 

paddle/fluid/framework/fleet/fleet_wrapper.cc

@@ -29,12 +29,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/fleet/fleet_wrapper.h"
 #include <algorithm>
 #include <utility>
-#include "paddle/fluid/framework/channel.h"
-#include "paddle/fluid/framework/data_feed.h"
-#include "paddle/fluid/framework/io/fs.h"
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/platform/timer.h"
 
 namespace paddle {
 namespace framework {
@@ -751,7 +745,57 @@ void FleetWrapper::PushDenseVarsAsync(
     push_sparse_status->push_back(std::move(status));
   }
 }
+#endif
+
+#ifdef PADDLE_WITH_XPU
+void FleetWrapper::PushDenseVarsAsync(
+    const Scope& scope, const uint64_t table_id,
+    const std::vector<std::string>& var_names,
+    std::vector<::std::future<int32_t>>* push_sparse_status,
+    float scale_datanorm, int batch_size,
+    const paddle::platform::Place& place) {
+#ifdef PADDLE_WITH_PSLIB
+  std::vector<paddle::ps::Region> regions;
+  for (auto& t : var_names) {
+    Variable* var = scope.FindVar(t);
+    LoDTensor* tensor = var->GetMutable<LoDTensor>();
+    int count = tensor->numel();
+    float* g_data = tensor->data<float>();
+
+    Variable* pin_var = scope.FindVar(t + "pin");
+    LoDTensor* pin_tensor = pin_var->GetMutable<LoDTensor>();
+    float* pin_g =
+        pin_tensor->mutable_data<float>(tensor->dims(), platform::CPUPlace());
+    memory::Copy(platform::CPUPlace(), pin_g,
+                 BOOST_GET_CONST(platform::XPUPlace, place), g_data,
+                 sizeof(float) * count);
+
+    float* g = pin_g;
+    if (scale_datanorm >= 0) {
+      if (t.find(".batch_size@GRAD") != std::string::npos ||
+          t.find(".batch_sum@GRAD") != std::string::npos) {
+        Eigen::Map<Eigen::MatrixXf> mat(g, 1, count);
+        float scale = 1.0 / batch_size;
+        mat *= scale;
+      } else if (t.find(".batch_square_sum@GRAD") != std::string::npos) {
+        VLOG(3) << "epsilon: " << scale_datanorm;
+        for (int i = 0; i < count; ++i) {
+          g[i] = (g[i] - batch_size * scale_datanorm) / batch_size +
+                 batch_size * scale_datanorm;
+        }
+      }
+    }
+    paddle::ps::Region reg(g, count);
+    regions.emplace_back(std::move(reg));
+  }
 
+  auto status = pslib_ptr_->_worker_ptr->push_dense(regions.data(),
+                                                    regions.size(), table_id);
+  if (push_sparse_status) {
+    push_sparse_status->push_back(std::move(status));
+  }
+#endif
+}
 #endif
 void FleetWrapper::PushDenseVarsAsync(
     const Scope& scope, const uint64_t table_id,

paddle/fluid/framework/fleet/fleet_wrapper.h

@@ -35,6 +35,12 @@ limitations under the License. */
 #include "paddle/fluid/framework/variable_helper.h"
 #include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
 
+namespace paddle {
+namespace framework {
+class Scope;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 
@@ -154,6 +160,14 @@ class FleetWrapper {
       float scale_datanorm, int batch_size,
       const paddle::platform::Place& place, cudaStream_t stream,
       cudaEvent_t event);
+#endif
+#ifdef PADDLE_WITH_XPU
+  void PushDenseVarsAsync(
+      const Scope& scope, const uint64_t table_id,
+      const std::vector<std::string>& var_names,
+      std::vector<::std::future<int32_t>>* push_sparse_status,
+      float scale_datanorm, int batch_size,
+      const paddle::platform::Place& place);
 #endif
   void PushDenseVarsAsync(
       const Scope& scope, const uint64_t table_id,

paddle/fluid/framework/fleet/gloo_wrapper.cc

@@ -10,10 +10,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/fleet/gloo_wrapper.h"
-#include <thread>  // NOLINT
-#include <vector>
 #include "paddle/fluid/framework/io/fs.h"
-#include "paddle/fluid/platform/errors.h"
 #include "paddle/fluid/string/string_helper.h"
 
 namespace gloo {

paddle/fluid/framework/fleet/heter_wrapper.cc

@@ -27,15 +27,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/fleet/heter_wrapper.h"
-#include <algorithm>
-#include <utility>
-#include "paddle/fluid/framework/channel.h"
-#include "paddle/fluid/framework/data_feed.h"
-#include "paddle/fluid/framework/device_worker.h"
-#include "paddle/fluid/framework/io/fs.h"
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/platform/timer.h"
 #ifdef PADDLE_WITH_PSLIB
 
 namespace paddle {
@@ -122,29 +113,66 @@ void HeterWrapper::SerializeToReq(const std::string& varname, Scope* scope,
   if (platform::is_cpu_place(tensor->place())) {
     memcpy(data_ptr, tensor->data<void>(),
            tensor->numel() * SizeOfType(tensor->type()));
-  }
+  } else {
 #ifdef PADDLE_WITH_CUDA
-  else {
     memory::Copy(platform::CPUPlace(), data_ptr,
                  BOOST_GET_CONST(platform::CUDAPlace, tensor->place()),
                  tensor->data<void>(),
                  tensor->numel() * SizeOfType(tensor->type()), nullptr);
-  }
 #endif
+#ifdef PADDLE_WITH_XPU
+    memory::Copy(platform::CPUPlace(), data_ptr,
+                 BOOST_GET_CONST(platform::XPUPlace, tensor->place()),
+                 tensor->data<void>(),
+                 tensor->numel() * SizeOfType(tensor->type()));
+#endif
+  }
 }
 
-// void HeterWrapper::DeSerializeToTensor(Scope* scope,
-// const HeterRequest* request) {
 #ifdef PADDLE_WITH_CUDA
 void HeterWrapper::DeSerializeToTensor(Scope* scope,
                                        const VariableMessage& req_var,
                                        platform::Place place,
                                        cudaStream_t stream) {
+  // const VariableMessage& req_var = request->vars();
+  auto* var = scope->FindVar(req_var.varname());
+  auto* tensor = var->GetMutable<LoDTensor>();
+
+  std::vector<int> vec_dim;
+  for (auto& x : req_var.dims()) {
+    vec_dim.push_back(x);
+  }
+  tensor->Resize(make_ddim(vec_dim));
+
+  LoD lod;
+  for (int i = 0; i < req_var.lod_level(); ++i) {
+    framework::Vector<size_t> v;
+    for (int j = 0; j < req_var.lod(i).lod_data_size(); ++j) {
+      v.push_back(req_var.lod(i).lod_data(j));
+    }
+    lod.push_back(v);
+  }
+  tensor->set_lod(lod);
+
+  void* tensor_data =
+      tensor->mutable_data(place, ToVarType(req_var.data_type()));
+
+#ifdef PADDLE_WITH_CUDA
+  memory::Copy(BOOST_GET_CONST(platform::CUDAPlace, place), tensor_data,
+               platform::CPUPlace(), req_var.data().data(),
+               tensor->numel() * SizeOfType(tensor->type()), stream);
 #else
+  memcpy(tensor_data, req_var.data().data(),
+         tensor->numel() * SizeOfType(tensor->type()));
+#endif
+}
+#endif
+
+// void HeterWrapper::DeSerializeToTensor(Scope* scope,
+// const HeterRequest* request) {
 void HeterWrapper::DeSerializeToTensor(Scope* scope,
                                        const VariableMessage& req_var,
                                        platform::Place place) {
-#endif
   // const VariableMessage& req_var = request->vars();
   auto* var = scope->FindVar(req_var.varname());
   auto* tensor = var->GetMutable<LoDTensor>();
@@ -168,10 +196,10 @@ void HeterWrapper::DeSerializeToTensor(Scope* scope,
   void* tensor_data =
       tensor->mutable_data(place, ToVarType(req_var.data_type()));
 
-#ifdef PADDLE_WITH_CUDA
-  memory::Copy(BOOST_GET_CONST(platform::CUDAPlace, place), tensor_data,
+#ifdef PADDLE_WITH_XPU
+  memory::Copy(BOOST_GET_CONST(platform::XPUPlace, place), tensor_data,
                platform::CPUPlace(), req_var.data().data(),
-               tensor->numel() * SizeOfType(tensor->type()), stream);
+               tensor->numel() * SizeOfType(tensor->type()));
 #else
   memcpy(tensor_data, req_var.data().data(),
          tensor->numel() * SizeOfType(tensor->type()));
@@ -192,7 +220,8 @@ framework::proto::VarType::Type HeterWrapper::ToVarType(
     case VariableMessage::BOOL:
       return framework::proto::VarType::BOOL;  // NOLINT
     default:
-      VLOG(0) << "Not support type " << type;
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "ToVarType:Unsupported type %d", type));
   }
 }
 
@@ -239,7 +268,7 @@ void HeterWrapper::CallRemoteXpu(std::shared_ptr<HeterTask> task,
   request.set_cur_batch(task->cur_batch_);
 
   OnHeterRpcDone* done = new OnHeterRpcDone([this, task, worker](void* done) {
-    auto* closure = (OnHeterRpcDone*)done;
+    auto* closure = reinterpret_cast<OnHeterRpcDone*>(done);
     if (closure->cntl.Failed()) {
       VLOG(0) << "call xpu fail: " << closure->cntl.ErrorText();
     } else {

paddle/fluid/framework/fleet/nccl_wrapper.cc

@@ -13,9 +13,6 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/fleet/nccl_wrapper.h"
-#include <utility>
-#include "paddle/fluid/framework/data_feed.h"
-#include "paddle/fluid/framework/scope.h"
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/fleet/nccl_wrapper.h

@@ -21,6 +21,7 @@ limitations under the License. */
 #include <random>
 #include <string>
 #include <vector>
+
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/variable_helper.h"
@@ -29,6 +30,12 @@ limitations under the License. */
 #endif
 #include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
 
+namespace paddle {
+namespace framework {
+class Scope;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/garbage_collector.h

@@ -19,9 +19,16 @@
 #include <memory>
 #include <mutex>  // NOLINT
 #include <utility>
+
 #include "gflags/gflags.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/generator.h

@@ -16,7 +16,6 @@ limitations under the License. */
 
 #include <glog/logging.h>
 #include <stdint.h>
-
 #include <atomic>
 #include <deque>
 #include <iostream>  // temp for debug

paddle/fluid/framework/hetercpu_worker.cc

@@ -12,13 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/framework/device_worker.h"
-#include "paddle/fluid/framework/device_worker_factory.h"
-#include "paddle/fluid/framework/fleet/fleet_wrapper.h"
-#include "paddle/fluid/framework/fleet/heter_wrapper.h"
-#include "paddle/fluid/platform/cpu_helper.h"
-#include "paddle/fluid/string/string_helper.h"
-
 #ifdef PADDLE_WITH_PSLIB
 
 #if defined _WIN32 || defined __APPLE__

paddle/fluid/framework/heterxpu_trainer.cc

@@ -22,9 +22,11 @@ limitations under the License. */
 #include "paddle/fluid/framework/device_worker_factory.h"
 #include "paddle/fluid/framework/fleet/fleet_wrapper.h"
 #include "paddle/fluid/framework/trainer.h"
-#if (defined PADDLE_WITH_CUDA) && (defined PADDLE_WITH_PSLIB)
+#if (defined PADDLE_WITH_CUDA || defined PADDLE_WITH_XPU) && \
+    (defined PADDLE_WITH_PSLIB)
+#ifdef PADDLE_WITH_CUDA
 #include "paddle/fluid/platform/cuda_device_guard.h"
-
+#endif
 namespace paddle {
 namespace framework {
 
@@ -44,6 +46,7 @@ void HeterXpuTrainer::Initialize(const TrainerDesc& trainer_desc,
   int place_num = trainer_desc.worker_places_size();
   for (int i = 0; i < place_num; ++i) {
     int num = trainer_desc.worker_places(i);
+#ifdef PADDLE_WITH_CUDA
     platform::CUDAPlace place = platform::CUDAPlace(num);
     platform::CUDADeviceGuard guard(place.device);
     cudaStream_t stream;
@@ -54,6 +57,11 @@ void HeterXpuTrainer::Initialize(const TrainerDesc& trainer_desc,
     PADDLE_ENFORCE_CUDA_SUCCESS(
         cudaEventCreateWithFlags(&event, cudaEventDisableTiming));
     events_.push_back(event);
+#endif
+#ifdef PADDLE_WITH_XPU
+    platform::XPUPlace place = platform::XPUPlace(num);
+    places_.push_back(place);
+#endif
   }
   // thread_num_ = trainer_desc.thread_num();
   // SetDataset(dataset);
@@ -105,11 +113,17 @@ void HeterXpuTrainer::Initialize(const TrainerDesc& trainer_desc,
 void HeterXpuTrainer::CreateThreadParam(const ProgramDesc& program, int num) {
   auto place = places_[num];
   Scope* scope = place_scopes_[num];
+#ifdef PADDLE_WITH_CUDA
   auto stream = copy_streams_[num];
   auto event = events_[num];
-
   auto dev_id = BOOST_GET_CONST(platform::CUDAPlace, place).device;
   platform::CUDADeviceGuard guard(dev_id);
+#endif
+
+#ifdef PADDLE_WITH_XPU
+  xpu_set_device(BOOST_GET_CONST(platform::XPUPlace, place).device);
+#endif
+
   auto& block = program.Block(0);
   for (auto& var : block.AllVars()) {
     if (var->Persistable()) {
@@ -126,13 +140,28 @@ void HeterXpuTrainer::CreateThreadParam(const ProgramDesc& program, int num) {
       HeterMemCpy<cpp_type>(thread_tensor, root_tensor, place, stream); \
     }                                                                   \
   } while (0)
+
+#define HeterMemcpyXpuFunc(cpp_type, proto_type)                \
+  do {                                                          \
+    if (root_tensor->type() == proto_type) {                    \
+      HeterMemCpy<cpp_type>(thread_tensor, root_tensor, place); \
+    }                                                           \
+  } while (0)
+#ifdef PADDLE_WITH_CUDA
       _ForEachDataType_(HeterMemcpyFunc);
+#endif
+#ifdef PADDLE_WITH_XPU
+      _ForEachDataType_(HeterMemcpyXpuFunc);
+#endif
     }
   }
+#ifdef PADDLE_WITH_CUDA
   PADDLE_ENFORCE_CUDA_SUCCESS(cudaEventRecord(event, stream));
   cudaEventSynchronize(event);
+#endif
 }
 
+#ifdef PADDLE_WITH_CUDA
 template <typename T>
 void HeterXpuTrainer::HeterMemCpy(LoDTensor* thread_tensor,
                                   LoDTensor* root_tensor,
@@ -151,6 +180,27 @@ void HeterXpuTrainer::HeterMemCpy(LoDTensor* thread_tensor,
                  root_ptr, sizeof(T) * root_tensor->numel(), stream);
   }
 }
+#endif
+
+#ifdef PADDLE_WITH_XPU
+template <typename T>
+void HeterXpuTrainer::HeterMemCpy(LoDTensor* thread_tensor,
+                                  LoDTensor* root_tensor,
+                                  const paddle::platform::Place& thread_place) {
+  T* thread_ptr =
+      thread_tensor->mutable_data<T>(root_tensor->dims(), thread_place);
+  T* root_ptr = root_tensor->data<T>();
+  if (platform::is_cpu_place(root_tensor->place())) {
+    memory::Copy(BOOST_GET_CONST(platform::XPUPlace, thread_place), thread_ptr,
+                 platform::CPUPlace(), root_ptr,
+                 sizeof(T) * root_tensor->numel());
+  } else {
+    memory::Copy(BOOST_GET_CONST(platform::XPUPlace, thread_place), thread_ptr,
+                 BOOST_GET_CONST(platform::XPUPlace, root_tensor->place()),
+                 root_ptr, sizeof(T) * root_tensor->numel());
+  }
+}
+#endif
 
 void HeterXpuTrainer::DumpWork(int tid) {}
 
@@ -181,13 +231,16 @@ void HeterXpuTrainer::InitOtherEnv(const ProgramDesc& main_program) {
     CreateThreadParam(main_program, i);
     pull_dense_worker_->AddThreadScope(scope);
     pull_dense_worker_->AddPlace(places_[i]);
+#ifdef PADDLE_WITH_CUDA
     pull_dense_worker_->AddStream(copy_streams_[i]);
+#endif
   }
-
   pull_dense_worker_->Start();
+#ifdef PADDLE_WITH_CUDA
   for (auto& stream : copy_streams_) {
     cudaStreamSynchronize(stream);
   }
+#endif
   op_names_.clear();
   for (auto& op_desc : block.AllOps()) {
     std::unique_ptr<OperatorBase> local_op = OpRegistry::CreateOp(*op_desc);
@@ -230,10 +283,12 @@ void HeterXpuTrainer::InitOtherEnv(const ProgramDesc& main_program) {
         OperatorBase* local_op_ptr = local_op.release();
         (context->ops_).push_back(local_op_ptr);
       }
+#ifdef PADDLE_WITH_CUDA
       auto dev_id = BOOST_GET_CONST(platform::CUDAPlace, place).device;
       platform::CUDADeviceGuard guard(dev_id);
       PADDLE_ENFORCE_CUDA_SUCCESS(
           cudaEventCreateWithFlags(&context->event_, cudaEventDisableTiming));
+#endif
       object_pool_.Push(context);
     }
   }
@@ -277,12 +332,25 @@ int HeterXpuTrainer::EndPass(const HeterRequest* request,
     }                                                                          \
   } while (0)
       _ForEachDataType_(MergeCallback);
-      if (platform::is_gpu_place(thread_tensor->place())) {
+      if (!platform::is_cpu_place(thread_tensor->place())) {
+#ifdef PADDLE_WITH_CUDA
         auto dev_id =
             BOOST_GET_CONST(platform::CUDAPlace, thread_tensor->place()).device;
         platform::CUDADeviceGuard guard(dev_id);
         cudaMemset(thread_tensor->data<void>(), 0,
                    thread_tensor->numel() * SizeOfType(thread_tensor->type()));
+#endif
+#ifdef PADDLE_WITH_XPU
+        auto place = thread_tensor->place();
+        xpu_set_device(BOOST_GET_CONST(platform::XPUPlace, place).device);
+        platform::DeviceContextPool& pool =
+            platform::DeviceContextPool::Instance();
+        platform::DeviceContext* dev_ctx = pool.Get(place);
+        const platform::XPUDeviceContext* xpu_ctx =
+            reinterpret_cast<const platform::XPUDeviceContext*>(dev_ctx);
+        xpu::memset(xpu_ctx->x_context(), thread_tensor->data<void>(), 0,
+                    thread_tensor->numel() * SizeOfType(thread_tensor->type()));
+#endif
       } else {
         memset(thread_tensor->data<void>(), 0,
                thread_tensor->numel() * SizeOfType(thread_tensor->type()));
@@ -291,12 +359,25 @@ int HeterXpuTrainer::EndPass(const HeterRequest* request,
     auto* merge_var = response->add_vars();
     heter_ptr_->SerializeToReq(need_merge_var_names_[i], root_scope_,
                                merge_var);
-    if (platform::is_gpu_place(root_tensor->place())) {
+    if (!platform::is_cpu_place(root_tensor->place())) {
+#ifdef PADDLE_WITH_CUDA
       auto dev_id =
           BOOST_GET_CONST(platform::CUDAPlace, root_tensor->place()).device;
       platform::CUDADeviceGuard guard(dev_id);
       cudaMemset(root_tensor->data<void>(), 0,
                  root_tensor->numel() * SizeOfType(root_tensor->type()));
+#endif
+#ifdef PADDLE_WITH_XPU
+      auto place = root_tensor->place();
+      xpu_set_device(BOOST_GET_CONST(platform::XPUPlace, place).device);
+      platform::DeviceContextPool& pool =
+          platform::DeviceContextPool::Instance();
+      platform::DeviceContext* dev_ctx = pool.Get(place);
+      const platform::XPUDeviceContext* xpu_ctx =
+          reinterpret_cast<const platform::XPUDeviceContext*>(dev_ctx);
+      xpu::memset(xpu_ctx->x_context(), root_tensor->data<void>(), 0,
+                  root_tensor->numel() * SizeOfType(root_tensor->type()));
+#endif
     } else {
       memset(root_tensor->data<void>(), 0,
              root_tensor->numel() * SizeOfType(root_tensor->type()));
@@ -334,7 +415,7 @@ int HeterXpuTrainer::RunTask(const HeterRequest* request,
   std::shared_ptr<HeterServiceContext> context = object_pool_.Get();
 
   if (!context->scope_) {
-    int num = rand() % places_.size();
+    int num = rand_r() % places_.size();
     context->place_num_ = num;
     auto place = places_[num];
     context->scope_ = &(place_scopes_[num]->NewScope());
@@ -356,11 +437,12 @@ int HeterXpuTrainer::RunTask(const HeterRequest* request,
       OperatorBase* local_op_ptr = local_op.release();
       (context->ops_).push_back(local_op_ptr);
     }
-
+#ifdef PADDLE_WITH_CUDA
     auto dev_id = BOOST_GET_CONST(platform::CUDAPlace, place).device;
     platform::CUDADeviceGuard guard(dev_id);
     PADDLE_ENFORCE_CUDA_SUCCESS(
         cudaEventCreateWithFlags(&context->event_, cudaEventDisableTiming));
+#endif
   }
 
   context->Reset();
@@ -369,15 +451,22 @@ int HeterXpuTrainer::RunTask(const HeterRequest* request,
     auto deserial_timer =
         std::make_shared<paddle::ps::CostTimer>("xpu_service_deserial");
     for (int i = 0; i < request->vars_size(); ++i) {
+#ifdef PADDLE_WITH_CUDA
       heter_ptr_->DeSerializeToTensor(context->scope_, request->vars(i), place,
                                       copy_streams_[context->place_num_]);
+#endif
+#ifdef PADDLE_WITH_XPU
+      heter_ptr_->DeSerializeToTensor(context->scope_, request->vars(i), place);
+#endif
     }
+#ifdef PADDLE_WITH_CUDA
     PADDLE_ENFORCE_CUDA_SUCCESS(
         cudaEventRecord(context->event_, copy_streams_[context->place_num_]));
     while (cudaEventQuery(context->event_) != cudaSuccess) {
       VLOG(3) << "wait for kernel";
       bthread_yield();
     }
+#endif
   }
 
   {
@@ -388,6 +477,7 @@ int HeterXpuTrainer::RunTask(const HeterRequest* request,
       op->Run(*(context->scope_), place);
     }
   }
+#ifdef PADDLE_WITH_CUDA
   auto* dev_ctx = static_cast<platform::CUDADeviceContext*>(
       platform::DeviceContextPool::Instance().Get(place));
   PADDLE_ENFORCE_CUDA_SUCCESS(
@@ -401,6 +491,10 @@ int HeterXpuTrainer::RunTask(const HeterRequest* request,
       bthread_yield();
     }
   }
+#endif
+#ifdef PADDLE_WITH_XPU
+  xpu_wait();
+#endif
 
   for (int i = 0; i < trainer_desc_.xpu_send_list_size(); ++i) {
     const std::string& varname = trainer_desc_.xpu_send_list(i);
@@ -417,11 +511,19 @@ int HeterXpuTrainer::RunTask(const HeterRequest* request,
        ++i) {
     uint64_t tid =
         static_cast<uint64_t>(param_.program_config(0).push_dense_table_id(i));
+#ifdef PADDLE_WITH_CUDA
     fleet_ptr_->PushDenseVarsAsync(
         *(context->scope_), tid, dense_grad_names_[tid],
         &(context->push_dense_status_), scale_datanorm_, request->cur_batch(),
         places_[context->place_num_], copy_streams_[context->place_num_],
         context->event_);
+#endif
+#ifdef PADDLE_WITH_XPU
+    fleet_ptr_->PushDenseVarsAsync(
+        *(context->scope_), tid, dense_grad_names_[tid],
+        &(context->push_dense_status_), scale_datanorm_, request->cur_batch(),
+        places_[context->place_num_]);
+#endif
   }
   for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
        ++i) {
@@ -463,7 +565,6 @@ void HeterXpuTrainer::Finalize() {
   pull_dense_worker_->Stop();
   root_scope_->DropKids();
 }
-
 }  // namespace framework
 }  // namespace paddle
 #endif

paddle/fluid/framework/inlined_vector.h

@@ -15,6 +15,7 @@
 #pragma once
 #include <cstdint>
 #include <vector>
+
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {

paddle/fluid/framework/inlined_vector_test.cc

@@ -13,10 +13,10 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/inlined_vector.h"
+
 #include <cstdlib>
 #include <ctime>
-#include <iostream>
-#include <vector>
+
 #include "gtest/gtest.h"
 
 namespace paddle {

paddle/fluid/framework/io/crypto/cipher_utils.cc

@@ -15,8 +15,6 @@
 #include "paddle/fluid/framework/io/crypto/cipher_utils.h"
 
 #include <cryptopp/osrng.h>
-
-#include <fstream>
 #include <sstream>
 
 #include "paddle/fluid/platform/enforce.h"

paddle/fluid/framework/io/crypto/cipher_utils.h

@@ -17,6 +17,7 @@
 #include <sstream>
 #include <string>
 #include <unordered_map>
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/io/fs.h

@@ -18,6 +18,7 @@
 #include <memory>
 #include <string>
 #include <vector>
+
 #include "glog/logging.h"
 #include "paddle/fluid/framework/io/shell.h"
 #include "paddle/fluid/string/string_helper.h"

paddle/fluid/framework/io/shell.h

@@ -32,6 +32,7 @@
 #include <string>
 #include <utility>
 #include <vector>
+
 #include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/string/string_helper.h"
 

paddle/fluid/framework/ir/attention_lstm_fuse_pass.h

@@ -20,6 +20,8 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
+class Graph;
+
 class AttentionLSTMFusePass : public FusePassBase {
  protected:
   void ApplyImpl(ir::Graph* graph) const override;