Merge branch 'prv-calibration'

CMakeLists.txt

@@ -214,6 +214,7 @@ if (NOT WIN32)
 # there is no official support of warpctc, nccl, cupti in windows
 include(external/warpctc)   # download, build, install warpctc
 include(cupti)
+include(external/gzstream)
 endif (NOT WIN32)
 
 if(WITH_DISTRIBUTE)

Dockerfile

@@ -43,6 +43,8 @@ RUN wget -q https://www.python.org/ftp/python/3.7.0/Python-3.7.0.tgz && \
     CFLAGS="-Wformat" ./configure --prefix=/usr/local/ --enable-shared > /dev/null && \
     make -j8 > /dev/null && make altinstall > /dev/null
 
+RUN rm -r /root/python_build
+
 RUN apt-get update && \
     apt-get install -y --allow-downgrades patchelf \
     python3 python3-dev python3-pip \

build.sh

+if [[ $1 == "pull"  ]]; then
+	git pull
+fi
+cd build && rm -rf *
+cmake  -DCMAKE_INSTALL_PREFIX=./tmp -DWITH_GPU=OFF -DWITH_MKLDNN=ON -DWITH_TESTING=ON -DWITH_PROFILER=ON -DWITH_MKL=ON -DWITH_INFERENCE_API_TEST=ON -DCMAKE_BUILD_TYPE=Release -DON_INFER=ON ..
+make -j
+make install -j
+make -j fluid_lib_dist
+pip install --force-reinstall --user /home/chuanqiw/paddle/intelpaddle-shanghai/build/tmp/opt/paddle/share/wheels/paddlepaddle-0.0.0-cp27-cp27mu-linux_x86_64.whl 

cmake/cuda.cmake

@@ -199,8 +199,11 @@ elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
     list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
 endif()
 else(NOT WIN32)
+list(APPEND CUDA_NVCC_FLAGS  "--compiler-options;/bigobj")
 if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
-    list(APPEND CUDA_NVCC_FLAGS  "-g -G")
+  list(APPEND CUDA_NVCC_FLAGS  "-g -G")
+  # match the cl's _ITERATOR_DEBUG_LEVEL
+  list(APPEND CUDA_NVCC_FLAGS  "-D_DEBUG")
 elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
   list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
 else()

cmake/external/gzstream.cmake

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+IF(MOBILE_INFERENCE)
+    return()
+ENDIF()
+
+include (ExternalProject)
+
+# NOTE: gzstream is needed when linking with ctr reader.
+
+SET(GZSTREAM_SOURCES_DIR ${THIRD_PARTY_PATH}/gzstream)
+SET(GZSTREAM_INSTALL_DIR ${THIRD_PARTY_PATH}/install/gzstream)
+SET(GZSTREAM_INCLUDE_DIR "${GZSTREAM_INSTALL_DIR}/include/" CACHE PATH "gzstream include directory." FORCE)
+
+ExternalProject_Add(
+        extern_gzstream
+        DEPENDS zlib
+        GIT_REPOSITORY "https://github.com/jacquesqiao/gzstream.git"
+        GIT_TAG ""
+        PREFIX          ${GZSTREAM_SOURCES_DIR}
+        UPDATE_COMMAND  ""
+        CONFIGURE_COMMAND ""
+        BUILD_IN_SOURCE 1
+        BUILD_COMMAND   make EXTERN_CPPFLAGS="-I${THIRD_PARTY_PATH}/install/zlib/include" EXTERM_LDFLAGS="-L${THIRD_PARTY_PATH}/install/zlib/lib" -j8
+        INSTALL_COMMAND mkdir -p ${GZSTREAM_INSTALL_DIR}/lib/ && mkdir -p ${GZSTREAM_INSTALL_DIR}/include/
+        && cp ${GZSTREAM_SOURCES_DIR}/src/extern_gzstream/libgzstream.a ${GZSTREAM_INSTALL_DIR}/lib
+        && cp -r ${GZSTREAM_SOURCES_DIR}/src/extern_gzstream/gzstream.h ${GZSTREAM_INSTALL_DIR}/include
+)
+
+ADD_LIBRARY(gzstream STATIC IMPORTED GLOBAL)
+SET_PROPERTY(TARGET gzstream PROPERTY IMPORTED_LOCATION
+        "${GZSTREAM_INSTALL_DIR}/lib/libgzstream.a")
+
+include_directories(${GZSTREAM_INCLUDE_DIR})
+ADD_DEPENDENCIES(gzstream extern_gzstream zlib)

cmake/external/pybind11.cmake

@@ -26,7 +26,7 @@ ExternalProject_Add(
         extern_pybind
         ${EXTERNAL_PROJECT_LOG_ARGS}
         GIT_REPOSITORY  "https://github.com/pybind/pybind11.git"
-        GIT_TAG         "v2.1.1"
+        GIT_TAG         "v2.2.4"
         PREFIX          ${PYBIND_SOURCE_DIR}
         UPDATE_COMMAND  ""
         CONFIGURE_COMMAND ""

cmake/generic.cmake

@@ -349,10 +349,17 @@ function(cc_test TARGET_NAME)
     set(oneValueArgs "")
     set(multiValueArgs SRCS DEPS ARGS)
     cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    if(WIN32)
+      list(APPEND win32_deps shlwapi)
+      if("${cc_test_DEPS};" MATCHES "python;")
+        list(REMOVE_ITEM cc_test_DEPS python)
+        list(APPEND win32_deps ${PYTHON_LIBRARIES})
+      endif()
+    endif(WIN32)
     add_executable(${TARGET_NAME} ${cc_test_SRCS})
     target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     if(WIN32)
-      target_link_libraries(${TARGET_NAME} shlwapi)
+      target_link_libraries(${TARGET_NAME} ${win32_deps})
     endif(WIN32)
     add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     add_test(NAME ${TARGET_NAME}
@@ -683,7 +690,7 @@ function(py_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS ARGS ENVS)
     cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_test(NAME ${TARGET_NAME}
-             COMMAND env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
+             COMMAND ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
              FLAGS_cpu_deterministic=true
              PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
              ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}

cmake/inference_lib.cmake

@@ -186,8 +186,7 @@ set(module "inference")
 copy(inference_lib DEPS ${inference_deps}
   SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
        ${src_dir}/${module}/api/paddle_*.h
-       ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
         )
 
 set(module "platform")

cmake/operators.cmake

@@ -109,7 +109,8 @@ function(op_library TARGET)
 
     # Define operators that don't need pybind here.
     foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
-"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op")
+"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op"
+"fusion_transpose_flatten_concat_op")
         if ("${TARGET}" STREQUAL "${manual_pybind_op}")
             set(pybind_flag 1)
         endif()

paddle/fluid/API.spec

@@ -26,12 +26,19 @@ paddle.fluid.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], vara
 paddle.fluid.DistributeTranspilerConfig.__init__ 
 paddle.fluid.ParallelExecutor.__init__ ArgSpec(args=['self', 'use_cuda', 'loss_name', 'main_program', 'share_vars_from', 'exec_strategy', 'build_strategy', 'num_trainers', 'trainer_id', 'scope'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 1, 0, None))
 paddle.fluid.ParallelExecutor.run ArgSpec(args=['self', 'fetch_list', 'feed', 'feed_dict', 'return_numpy'], varargs=None, keywords=None, defaults=(None, None, True))
-paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ExecutionStrategy) -> None
-paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.GradientScaleStrategy, arg0: int) -> None
-paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ReduceStrategy, arg0: int) -> None
-paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.BuildStrategy) -> None
+paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.ExecutionStrategy) -> None
+paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.GradientScaleStrategy, arg0: int) -> None
+paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.ReduceStrategy, arg0: int) -> None
+paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy) -> None
 paddle.fluid.create_lod_tensor ArgSpec(args=['data', 'recursive_seq_lens', 'place'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.create_random_int_lodtensor ArgSpec(args=['recursive_seq_lens', 'base_shape', 'place', 'low', 'high'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.DataFeedDesc.__init__ ArgSpec(args=['self', 'proto_file'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.DataFeedDesc.desc ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.DataFeedDesc.set_batch_size ArgSpec(args=['self', 'batch_size'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.DataFeedDesc.set_dense_slots ArgSpec(args=['self', 'dense_slots_name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.DataFeedDesc.set_use_slots ArgSpec(args=['self', 'use_slots_name'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.AsyncExecutor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.AsyncExecutor.run ArgSpec(args=['self', 'program', 'data_feed', 'filelist', 'thread_num', 'fetch', 'debug'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.io.save_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.io.save_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.io.save_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
@@ -69,7 +76,7 @@ paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name']
 paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
 paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
 paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
-paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False))
+paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False))
 paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
@@ -97,8 +104,8 @@ paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_ti
 paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
-paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0))
-paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
+paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0, False))
+paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name', 'path_table', 'path_code', 'is_custom', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, False, False))
 paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None))
 paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
@@ -175,7 +182,7 @@ paddle.fluid.layers.clip ArgSpec(args=['x', 'min', 'max', 'name'], varargs=None,
 paddle.fluid.layers.clip_by_norm ArgSpec(args=['x', 'max_norm', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None))
-paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'ignore_index', 'name'], varargs=None, keywords=None, defaults=(-100, None))
 paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
@@ -187,6 +194,7 @@ paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=Non
 paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
 paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None))
+paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)

paddle/fluid/framework/CMakeLists.txt

@@ -34,6 +34,7 @@ add_subdirectory(ir)
 add_subdirectory(details)
 # ddim lib
 proto_library(framework_proto SRCS framework.proto)
+proto_library(async_executor_proto SRCS data_feed.proto)
 
 cc_library(ddim SRCS ddim.cc DEPS eigen3 boost)
 cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
@@ -116,14 +117,9 @@ cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 
-if (NOT WIN32)
-cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto)
+cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto device_context)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
     shape_inference data_transform lod_tensor profiler transfer_scope_cache)
-else()
-cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
-    shape_inference data_transform lod_tensor)
-endif(NOT WIN32)
 
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
 
@@ -140,7 +136,7 @@ endif(NOT WIN32)
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
 nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 
-py_proto_compile(framework_py_proto SRCS framework.proto)
+py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
@@ -162,18 +158,19 @@ endif(NOT WIN32)
 cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
 
 cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog)
+cc_library(variable_helper SRCS variable_helper.cc DEPS lod_tensor)
 
-cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
+cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
 
 if(WITH_DISTRIBUTE)
-  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass)
+  cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass variable_helper)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 else()
   if(NOT WIN32)
-    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator)
+    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator variable_helper)
   else(NOT WIN32)
-    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
+    cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
   endif(NOT WIN32)
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
@@ -181,8 +178,11 @@ endif()
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
         threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
         graph build_strategy
-        fast_threaded_ssa_graph_executor)
+        fast_threaded_ssa_graph_executor variable_helper)
+
+cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper)
 
+cc_test(data_feed_test SRCS data_feed_test.cc DEPS async_executor)
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)
 cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry

paddle/fluid/framework/async_executor.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/async_executor.h"
+#include "google/protobuf/io/zero_copy_stream_impl.h"
+#include "google/protobuf/message.h"
+#include "google/protobuf/text_format.h"
+
+#include "gflags/gflags.h"
+#include "paddle/fluid/framework/data_feed_factory.h"
+#include "paddle/fluid/framework/executor_thread_worker.h"
+#include "paddle/fluid/framework/feed_fetch_method.h"
+#include "paddle/fluid/framework/feed_fetch_type.h"
+#include "paddle/fluid/framework/lod_rank_table.h"
+#include "paddle/fluid/framework/lod_tensor_array.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/reader.h"
+#include "paddle/fluid/inference/io.h"
+#include "paddle/fluid/platform/place.h"
+#include "paddle/fluid/pybind/pybind.h"
+
+namespace paddle {
+namespace framework {
+AsyncExecutor::AsyncExecutor(Scope* scope, const platform::Place& place)
+    : root_scope_(scope), place_(place) {}
+
+void AsyncExecutor::CreateThreads(
+    ExecutorThreadWorker* worker, const ProgramDesc& main_program,
+    const std::shared_ptr<DataFeed>& reader,
+    const std::vector<std::string>& fetch_var_names, Scope* root_scope,
+    const int thread_index, const bool debug) {
+  worker->SetThreadId(thread_index);
+  worker->SetDebug(debug);
+  worker->SetRootScope(root_scope);
+  worker->CreateThreadResource(main_program, place_);
+  worker->SetDataFeed(reader);
+  worker->SetFetchVarNames(fetch_var_names);
+  worker->BindingDataFeedMemory();
+}
+
+void PrepareReaders(std::vector<std::shared_ptr<DataFeed>>& readers,  // NOLINT
+                    const int thread_num, const DataFeedDesc& data_feed_desc,
+                    const std::vector<std::string>& filelist) {
+  readers.resize(thread_num);
+  for (size_t i = 0; i < readers.size(); ++i) {
+    readers[i] = DataFeedFactory::CreateDataFeed(data_feed_desc.name());
+    readers[i]->Init(data_feed_desc);  // set batch_size and queue_size here
+  }
+  readers[0]->SetFileList(filelist);
+}
+
+void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
+                                const std::string& data_feed_desc_str,
+                                const std::vector<std::string>& filelist,
+                                const int thread_num,
+                                const std::vector<std::string>& fetch_var_names,
+                                const bool debug) {
+  std::vector<std::thread> threads;
+
+  auto& block = main_program.Block(0);
+  for (auto var_name : fetch_var_names) {
+    auto var_desc = block.FindVar(var_name);
+    auto shapes = var_desc->GetShape();
+    PADDLE_ENFORCE(shapes[shapes.size() - 1] == 1,
+                   "var %s: Fetched var has wrong shape, "
+                   "only variables with the last dimension size 1 supported",
+                   var_name);
+  }
+
+  DataFeedDesc data_feed_desc;
+  google::protobuf::TextFormat::ParseFromString(data_feed_desc_str,
+                                                &data_feed_desc);
+
+  int actual_thread_num = thread_num;
+  int file_cnt = filelist.size();
+  PADDLE_ENFORCE(file_cnt > 0, "File list cannot be empty");
+
+  if (actual_thread_num > file_cnt) {
+    VLOG(1) << "Thread num = " << thread_num << ", file num = " << file_cnt
+            << ". Changing thread_num = " << file_cnt;
+    actual_thread_num = file_cnt;
+  }
+
+  /*
+    readerDesc: protobuf description for reader initlization
+    argument: class_name, batch_size, use_slot, queue_size, buffer_size,
+    padding_index
+
+    reader:
+    1) each thread has a reader, reader will read input data and
+    put it into input queue
+    2) each reader has a Next() iterface, that can fetch an instance
+    from the input queue
+   */
+  // todo: should be factory method for creating datafeed
+  std::vector<std::shared_ptr<DataFeed>> readers;
+  PrepareReaders(readers, actual_thread_num, data_feed_desc, filelist);
+
+  std::vector<std::shared_ptr<ExecutorThreadWorker>> workers;
+  workers.resize(actual_thread_num);
+  for (auto& worker : workers) {
+    worker.reset(new ExecutorThreadWorker);
+  }
+
+  // prepare thread resource here
+  for (int thidx = 0; thidx < actual_thread_num; ++thidx) {
+    CreateThreads(workers[thidx].get(), main_program, readers[thidx],
+                  fetch_var_names, root_scope_, thidx, debug);
+  }
+
+  // start executing ops in multiple threads
+  for (int thidx = 0; thidx < actual_thread_num; ++thidx) {
+    threads.push_back(
+        std::thread(&ExecutorThreadWorker::TrainFiles, workers[thidx].get()));
+  }
+
+  for (auto& th : threads) {
+    th.join();
+  }
+
+  root_scope_->DropKids();
+
+  return;
+}
+
+}  // einit_modelnd namespace framework
+}  // end namespace paddle

paddle/fluid/framework/async_executor.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+  http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <map>
+#include <memory>
+#include <mutex>  // NOLINT
+#include <set>
+#include <string>
+#include <thread>  // NOLINT
+#include <typeinfo>
+#include <vector>
+#include "paddle/fluid/framework/data_feed.pb.h"
+#include "paddle/fluid/framework/executor.h"
+#include "paddle/fluid/framework/executor_thread_worker.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/scope.h"
+
+namespace paddle {
+namespace framework {
+class AsyncExecutor {
+ public:
+  AsyncExecutor(Scope* scope, const platform::Place& place);
+  virtual ~AsyncExecutor() {}
+  void RunFromFile(const ProgramDesc& main_program,
+                   const std::string& data_feed_desc_str,
+                   const std::vector<std::string>& filelist,
+                   const int thread_num,
+                   const std::vector<std::string>& fetch_names,
+                   const bool debug = false);
+
+ private:
+  void CreateThreads(ExecutorThreadWorker* worker,
+                     const ProgramDesc& main_program,
+                     const std::shared_ptr<DataFeed>& reader,
+                     const std::vector<std::string>& fetch_var_names,
+                     Scope* root_scope, const int thread_index,
+                     const bool debug);
+
+ public:
+  Scope* root_scope_;
+  platform::Place place_;
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/data_device_transform.cc

@@ -18,8 +18,8 @@ namespace framework {
 
 void TransDataDevice(const Tensor &in, const platform::Place &dst_place,
                      Tensor *out) {
-  VLOG(30) << "DeviceTransform in, src_place " << in.place()
-           << " dst_place: " << dst_place;
+  VLOG(3) << "DeviceTransform in, src_place " << in.place()
+          << " dst_place: " << dst_place;
 
   PADDLE_ENFORCE_NE(
       in.place().which(), dst_place.which(),

paddle/fluid/framework/data_device_transform_test.cu

@@ -49,10 +49,10 @@ class TestOpWithKernel : public OperatorWithKernel {
   OpKernelType GetExpectedKernelType(
       const ExecutionContext& ctx) const override {
     if (Attr<bool>("use_gpu")) {
-      VLOG(30) << "force use gpu kernel";
+      VLOG(3) << "force use gpu kernel";
       return OpKernelType(proto::VarType::FP32, platform::CUDAPlace(0));
     } else {
-      VLOG(30) << "use default kernel";
+      VLOG(3) << "use default kernel";
       return OpKernelType(proto::VarType::FP32,
                           ctx.Input<Tensor>("input")->place());
     }
@@ -148,7 +148,7 @@ TEST(Operator, CPUtoGPU) {
   // get output
   auto* output2 = scope.Var("OUT2");
   gpu_op->Run(scope, cuda_place);
-  VLOG(30) << "after gpu_op run";
+  VLOG(3) << "after gpu_op run";
 
   // auto* output2_ptr = output2->Get<LoDTensor>().data<float>();
   paddle::platform::DeviceContextPool& pool =

paddle/fluid/framework/data_feed.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "google/protobuf/io/zero_copy_stream_impl.h"
+#include "google/protobuf/message.h"
+#include "google/protobuf/text_format.h"
+
+#include "gflags/gflags.h"
+#include "paddle/fluid/framework/data_feed.h"
+#include "paddle/fluid/framework/feed_fetch_method.h"
+#include "paddle/fluid/framework/feed_fetch_type.h"
+
+namespace paddle {
+namespace framework {
+
+std::vector<std::string> DataFeed::filelist_;
+size_t DataFeed::file_idx_;
+std::mutex DataFeed::mutex_for_pick_file_;
+bool DataFeed::finish_set_filelist_;
+
+void DataFeed::AddFeedVar(Variable* var, const std::string& name) {
+  CheckInit();
+  for (size_t i = 0; i < use_slots_.size(); ++i) {
+    if (name == use_slots_[i]) {
+      if (use_slots_is_dense_[i]) {
+        feed_vec_[i] = MixTensor(var->GetMutable<Tensor>());
+      } else {
+        feed_vec_[i] = MixTensor(var->GetMutable<LoDTensor>());
+      }
+    }
+  }
+}
+
+bool DataFeed::SetFileList(const std::vector<std::string>& files) {
+  std::unique_lock<std::mutex> lock(mutex_for_pick_file_);
+  CheckInit();
+  if (finish_set_filelist_) {
+    VLOG(3) << "info: you have set the filelist.";
+    return false;
+  }
+  PADDLE_ENFORCE(files.size(), "You have set an empty filelist.");
+  filelist_.assign(files.begin(), files.end());
+  file_idx_ = 0;
+
+  finish_set_filelist_ = true;
+  return true;
+}
+
+void DataFeed::SetBatchSize(int batch_size) {
+  PADDLE_ENFORCE(batch_size > 0, "Illegal batch size: %d.", batch_size);
+  default_batch_size_ = batch_size;
+}
+
+bool DataFeed::PickOneFile(std::string* filename) {
+  std::unique_lock<std::mutex> lock(mutex_for_pick_file_);
+  if (file_idx_ == filelist_.size()) {
+    return false;
+  }
+  *filename = filelist_[file_idx_++];
+  return true;
+}
+
+void DataFeed::CheckInit() {
+  PADDLE_ENFORCE(finish_init_, "Initialization did not succeed.");
+}
+
+void DataFeed::CheckSetFileList() {
+  PADDLE_ENFORCE(finish_set_filelist_, "Set filelist did not succeed.");
+}
+
+void DataFeed::CheckStart() {
+  PADDLE_ENFORCE(finish_start_, "Datafeed has not started running yet.");
+}
+
+template <typename T>
+void PrivateQueueDataFeed<T>::SetQueueSize(int queue_size) {
+  PADDLE_ENFORCE(queue_size > 0, "Illegal queue size: %d.", queue_size);
+  queue_size_ = queue_size;
+  queue_ = std::unique_ptr<paddle::operators::reader::BlockingQueue<T>>(
+      new paddle::operators::reader::BlockingQueue<T>(queue_size_));
+}
+
+template <typename T>
+bool PrivateQueueDataFeed<T>::Start() {
+  CheckSetFileList();
+  read_thread_ = std::thread(&PrivateQueueDataFeed::ReadThread, this);
+  read_thread_.detach();
+
+  finish_start_ = true;
+  return true;
+}
+
+template <typename T>
+void PrivateQueueDataFeed<T>::ReadThread() {
+  std::string filename;
+  while (PickOneFile(&filename)) {
+    file_.open(filename.c_str());  // is_text_feed
+    PADDLE_ENFORCE(file_.good(), "Open file<%s> fail.", filename.c_str());
+    T instance;
+    while (ParseOneInstance(&instance)) {
+      queue_->Send(instance);
+    }
+    file_.close();
+  }
+  queue_->Close();
+}
+
+template <typename T>
+int PrivateQueueDataFeed<T>::Next() {
+  CheckStart();
+  int index = 0;
+  T instance;
+  T ins_vec;
+  while (index < default_batch_size_) {
+    if (!queue_->Receive(&instance)) {
+      break;
+    }
+    AddInstanceToInsVec(&ins_vec, instance, index++);
+  }
+  batch_size_ = index;
+  if (batch_size_ != 0) {
+    PutToFeedVec(ins_vec);
+  }
+  return batch_size_;
+}
+
+#ifdef _WIN32
+template class PrivateQueueDataFeed<std::vector<MultiSlotType>>;
+#endif
+
+void MultiSlotDataFeed::Init(
+    const paddle::framework::DataFeedDesc& data_feed_desc) {
+  finish_init_ = false;
+  finish_set_filelist_ = false;
+  finish_start_ = false;
+
+  PADDLE_ENFORCE(data_feed_desc.has_multi_slot_desc(),
+                 "Multi_slot_desc has not been set.");
+  paddle::framework::MultiSlotDesc multi_slot_desc =
+      data_feed_desc.multi_slot_desc();
+  SetBatchSize(data_feed_desc.batch_size());
+  SetQueueSize(data_feed_desc.batch_size());
+  size_t all_slot_num = multi_slot_desc.slots_size();
+  all_slots_.resize(all_slot_num);
+  all_slots_type_.resize(all_slot_num);
+  use_slots_index_.resize(all_slot_num);
+  use_slots_.clear();
+  use_slots_is_dense_.clear();
+  for (size_t i = 0; i < all_slot_num; ++i) {
+    const auto& slot = multi_slot_desc.slots(i);
+    all_slots_[i] = slot.name();
+    all_slots_type_[i] = slot.type();
+    use_slots_index_[i] = slot.is_used() ? use_slots_.size() : -1;
+    if (slot.is_used()) {
+      use_slots_.push_back(all_slots_[i]);
+      use_slots_is_dense_.push_back(slot.is_dense());
+    }
+  }
+  feed_vec_.resize(use_slots_.size());
+  finish_init_ = true;
+}
+
+bool MultiSlotDataFeed::CheckFile(const char* filename) {
+  CheckInit();  // get info of slots
+  std::ifstream fin(filename);
+  if (!fin.good()) {
+    VLOG(1) << "error: open file<" << filename << "> fail";
+    return false;
+  }
+  std::string line;
+  int instance_cout = 0;
+  std::string all_slots_alias = "";
+  for (const auto& alias : all_slots_) {
+    all_slots_alias += alias + " ";
+  }
+  std::string use_slots_alias = "";
+  for (const auto& alias : use_slots_) {
+    use_slots_alias += alias + " ";
+  }
+  VLOG(3) << "total slots num: " << all_slots_.size();
+  VLOG(3) << "total slots alias: " << all_slots_alias;
+  VLOG(3) << "used slots num: " << use_slots_.size();
+  VLOG(3) << "used slots alias: " << use_slots_alias;
+  while (getline(fin, line)) {
+    ++instance_cout;
+    const char* str = line.c_str();
+    char* endptr = const_cast<char*>(str);
+    int len = line.length();
+    for (size_t i = 0; i < all_slots_.size(); ++i) {
+      int num = strtol(endptr, &endptr, 10);
+      if (num < 0) {
+        VLOG(0) << "error: the number of ids is a negative number: " << num;
+        VLOG(0) << "please check line<" << instance_cout << "> in file<"
+                << filename << ">";
+        return false;
+      } else if (num == 0) {
+        VLOG(0)
+            << "error: the number of ids can not be zero, you need "
+               "padding it in data generator; or if there is something wrong"
+               " with the data, please check if the data contains unresolvable "
+               "characters.";
+        VLOG(0) << "please check line<" << instance_cout << "> in file<"
+                << filename << ">";
+        return false;
+      } else if (errno == ERANGE || num > INT_MAX) {
+        VLOG(0) << "error: the number of ids greater than INT_MAX";
+        VLOG(0) << "please check line<" << instance_cout << "> in file<"
+                << filename << ">";
+        return false;
+      }
+      if (all_slots_type_[i] == "float") {
+        for (int i = 0; i < num; ++i) {
+          strtof(endptr, &endptr);
+          if (errno == ERANGE) {
+            VLOG(0) << "error: the value is out of the range of "
+                       "representable values for float";
+            VLOG(0) << "please check line<" << instance_cout << "> in file<"
+                    << filename << ">";
+            return false;
+          }
+          if (i + 1 != num && endptr - str == len) {
+            VLOG(0) << "error: there is a wrong with the number of ids.";
+            VLOG(0) << "please check line<" << instance_cout << "> in file<"
+                    << filename << ">";
+            return false;
+          }
+        }
+      } else if (all_slots_type_[i] == "uint64") {
+        for (int i = 0; i < num; ++i) {
+          strtoull(endptr, &endptr, 10);
+          if (errno == ERANGE) {
+            VLOG(0) << "error: the value is out of the range of "
+                       "representable values for uint64_t";
+            VLOG(0) << "please check line<" << instance_cout << "> in file<"
+                    << filename << ">";
+            return false;
+          }
+          if (i + 1 != num && endptr - str == len) {
+            VLOG(0) << "error: there is a wrong with the number of ids.";
+            VLOG(0) << "please check line<" << instance_cout << "> in file<"
+                    << filename << ">";
+            return false;
+          }
+        }
+      } else {
+        VLOG(0) << "error: this type<" << all_slots_type_[i]
+                << "> is not supported";
+        return false;
+      }
+    }
+    // It may be added '\t' character to the end of the output of reduce
+    // task when processes data by Hadoop(when the output of the reduce
+    // task of Hadoop has only one field, it will add a '\t' at the end
+    // of the line by default, and you can use this option to avoid it:
+    // `-D mapred.textoutputformat.ignoreseparator=true`), which does
+    // not affect the correctness of the data. Therefore, it should be
+    // judged that the data is not normal when the end of each line of
+    // data contains characters which are not spaces.
+    while (endptr - str != len) {
+      if (!isspace(*(endptr++))) {
+        VLOG(0)
+            << "error: there is some extra characters at the end of the line.";
+        VLOG(0) << "please check line<" << instance_cout << "> in file<"
+                << filename << ">";
+        return false;
+      }
+    }
+  }
+  VLOG(3) << "instances cout: " << instance_cout;
+  VLOG(3) << "The file format is correct";
+  return true;
+}
+
+bool MultiSlotDataFeed::ParseOneInstance(std::vector<MultiSlotType>* instance) {
+  std::string line;
+  if (getline(file_, line)) {
+    int use_slots_num = use_slots_.size();
+    instance->resize(use_slots_num);
+    // parse line
+    const char* str = line.c_str();
+    char* endptr = const_cast<char*>(str);
+    int pos = 0;
+    for (size_t i = 0; i < use_slots_index_.size(); ++i) {
+      int idx = use_slots_index_[i];
+      int num = strtol(&str[pos], &endptr, 10);
+      PADDLE_ENFORCE(
+          num,
+          "The number of ids can not be zero, you need padding "
+          "it in data generator; or if there is something wrong with "
+          "the data, please check if the data contains unresolvable "
+          "characters.\nplease check this error line: %s",
+          str);
+      if (idx != -1) {
+        (*instance)[idx].Init(all_slots_type_[i]);
+        if ((*instance)[idx].GetType()[0] == 'f') {  // float
+          for (int j = 0; j < num; ++j) {
+            float feasign = strtof(endptr, &endptr);
+            (*instance)[idx].AddValue(feasign);
+          }
+        } else if ((*instance)[idx].GetType()[0] == 'u') {  // uint64
+          for (int j = 0; j < num; ++j) {
+            uint64_t feasign = (uint64_t)strtoull(endptr, &endptr, 10);
+            (*instance)[idx].AddValue(feasign);
+          }
+        }
+        pos = endptr - str;
+      } else {
+        for (int j = 0; j <= num; ++j) {
+          pos = line.find_first_of(' ', pos + 1);
+        }
+      }
+    }
+  } else {
+    return false;
+  }
+  return true;
+}
+
+void MultiSlotDataFeed::AddInstanceToInsVec(
+    std::vector<MultiSlotType>* ins_vec,
+    const std::vector<MultiSlotType>& instance, int index) {
+  if (index == 0) {
+    ins_vec->resize(instance.size());
+    for (size_t i = 0; i < instance.size(); ++i) {
+      (*ins_vec)[i].Init(instance[i].GetType());
+      (*ins_vec)[i].InitOffset();
+    }
+  }
+  for (size_t i = 0; i < instance.size(); ++i) {
+    (*ins_vec)[i].AddIns(instance[i]);
+  }
+}
+
+void MultiSlotDataFeed::PutToFeedVec(
+    const std::vector<MultiSlotType>& ins_vec) {
+  for (size_t i = 0; i < use_slots_.size(); ++i) {
+    const auto& type = ins_vec[i].GetType();
+    const auto& offset = ins_vec[i].GetOffset();
+    int total_instance = static_cast<int>(offset.back());
+    if (type[0] == 'f') {  // float
+      const auto& feasign = ins_vec[i].GetFloatData();
+      if (feed_vec_[i].IsDense()) {
+        int size_in_each_batch = total_instance / batch_size_;
+        float* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data<float>(
+            {batch_size_, size_in_each_batch}, platform::CPUPlace());
+        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
+      } else {
+        float* tensor_ptr = feed_vec_[i].GetLoDTensor()->mutable_data<float>(
+            {total_instance, 1}, platform::CPUPlace());
+        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
+        LoD data_lod{offset};
+        feed_vec_[i].GetLoDTensor()->set_lod(data_lod);
+      }
+    } else if (type[0] == 'u') {  // uint64
+      // no uint64_t type in paddlepaddle
+      const auto& feasign = ins_vec[i].GetUint64Data();
+      if (feed_vec_[i].IsDense()) {
+        int size_in_each_batch = total_instance / batch_size_;
+        int64_t* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data<int64_t>(
+            {batch_size_, size_in_each_batch}, platform::CPUPlace());
+        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
+      } else {
+        int64_t* tensor_ptr =
+            feed_vec_[i].GetLoDTensor()->mutable_data<int64_t>(
+                {total_instance, 1}, platform::CPUPlace());
+        memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
+        LoD data_lod{offset};
+        feed_vec_[i].GetLoDTensor()->set_lod(data_lod);
+      }
+    }
+  }
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/data_feed.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <fstream>
+#include <memory>
+#include <mutex>  // NOLINT
+#include <string>
+#include <thread>  // NOLINT
+#include <vector>
+
+#include "paddle/fluid/framework/data_feed.pb.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/reader.h"
+#include "paddle/fluid/framework/variable.h"
+#include "paddle/fluid/operators/reader/blocking_queue.h"
+
+namespace paddle {
+namespace framework {
+
+// Pack Tensor type and LoDTensor type into MixTensor type, in order
+// to record either Tensor or LoDTensor information at the same time.
+class MixTensor {
+ public:
+  MixTensor() {}
+  explicit MixTensor(LoDTensor* lodtensor) {
+    is_dense_ = false;
+    lodtensor_ = lodtensor;
+  }
+  explicit MixTensor(Tensor* tensor) {
+    is_dense_ = true;
+    tensor_ = tensor;
+  }
+  bool IsDense() { return is_dense_; }
+  LoDTensor* GetLoDTensor() {
+    PADDLE_ENFORCE(!is_dense_, "Let a dense var return a LoDTensor ptr.");
+    return lodtensor_;
+  }
+  Tensor* GetTensor() {
+    PADDLE_ENFORCE(is_dense_, "Let a sparse var return a Tensor ptr.");
+    return tensor_;
+  }
+
+ private:
+  bool is_dense_;
+  LoDTensor* lodtensor_;
+  Tensor* tensor_;
+};
+
+// DataFeed is the base virtual class for all ohther DataFeeds.
+// It is used to read files and parse the data for subsequent trainer.
+// Example:
+//   DataFeed* reader =
+//   paddle::framework::DataFeedFactory::CreateDataFeed(data_feed_name);
+//   reader->Init(data_feed_desc); // data_feed_desc is a protobuf object
+//   reader->SetFileList(filelist);
+//   const std::vector<std::string> & use_slot_alias =
+//   reader->GetUseSlotAlias();
+//   for (auto name: use_slot_alias){ // for binding memory
+//     reader->AddFeedVar(scope->Var(name), name);
+//   }
+//   reader->Start();
+//   while (reader->Next()) {
+//      // trainer do something
+//   }
+class DataFeed {
+ public:
+  DataFeed() {}
+  virtual ~DataFeed() {}
+  virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc) = 0;
+  virtual bool CheckFile(const char* filename) {
+    PADDLE_THROW("This function(CheckFile) is not implemented.");
+  }
+  // Set filelist for DataFeed.
+  // Pay attention that it must init all readers before call this function.
+  // Otherwise, Init() function will init finish_set_filelist_ flag.
+  virtual bool SetFileList(const std::vector<std::string>& files);
+  virtual bool Start() = 0;
+  // The trainer calls the Next() function, and the DataFeed will load a new
+  // batch to the feed_vec. The return value of this function is the batch
+  // size of the current batch.
+  virtual int Next() = 0;
+  // Get all slots' alias which defined in protofile
+  virtual const std::vector<std::string>& GetAllSlotAlias() {
+    return all_slots_;
+  }
+  // Get used slots' alias which defined in protofile
+  virtual const std::vector<std::string>& GetUseSlotAlias() {
+    return use_slots_;
+  }
+  // This function is used for binding feed_vec memory
+  virtual void AddFeedVar(Variable* var, const std::string& name);
+
+ protected:
+  // The following three functions are used to check if it is executed in this
+  // order:
+  //   Init() -> SetFileList() -> Start() -> Next()
+  virtual void CheckInit();
+  virtual void CheckSetFileList();
+  virtual void CheckStart();
+  virtual void SetBatchSize(
+      int batch);  // batch size will be set in Init() function
+  // This function is used to pick one file from the global filelist(thread
+  // safe).
+  virtual bool PickOneFile(std::string* filename);
+
+  static std::vector<std::string> filelist_;
+  static size_t file_idx_;
+  static std::mutex mutex_for_pick_file_;
+
+  // the alias of used slots, and its order is determined by
+  // data_feed_desc(proto object)
+  std::vector<std::string> use_slots_;
+  std::vector<bool> use_slots_is_dense_;
+
+  // the alias of all slots, and its order is determined by data_feed_desc(proto
+  // object)
+  std::vector<std::string> all_slots_;
+  std::vector<std::string> all_slots_type_;
+  std::vector<int>
+      use_slots_index_;  // -1: not used; >=0: the index of use_slots_
+
+  // The data read by DataFeed will be stored here
+  std::vector<MixTensor> feed_vec_;
+
+  // the batch size defined by user
+  int default_batch_size_;
+  // current batch size
+  int batch_size_;
+
+  bool finish_init_;
+  static bool finish_set_filelist_;
+  bool finish_start_;
+};
+
+// PrivateQueueDataFeed is the base virtual class for ohther DataFeeds.
+// It use a read-thread to read file and parse data to a private-queue
+// (thread level), and get data from this queue when trainer call Next().
+template <typename T>
+class PrivateQueueDataFeed : public DataFeed {
+ public:
+  PrivateQueueDataFeed() {}
+  virtual ~PrivateQueueDataFeed() {}
+  virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc) = 0;
+  virtual bool Start();
+  virtual int Next();
+
+ protected:
+  // The thread implementation function for reading file and parse.
+  virtual void ReadThread();
+  // This function is used to set private-queue size, and the most
+  // efficient when the queue size is close to the batch size.
+  virtual void SetQueueSize(int queue_size);
+  // The reading and parsing method called in the ReadThread.
+  virtual bool ParseOneInstance(T* instance) = 0;
+  // This function is used to put instance to vec_ins
+  virtual void AddInstanceToInsVec(T* vec_ins, const T& instance,
+                                   int index) = 0;
+  // This function is used to put ins_vec to feed_vec
+  virtual void PutToFeedVec(const T& ins_vec) = 0;
+
+  // The thread for read files
+  std::thread read_thread_;
+  // using ifstream one line and one line parse is faster
+  // than using fread one buffer and one buffer parse.
+  //   for a 601M real data:
+  //     ifstream one line and one line parse: 6034 ms
+  //     fread one buffer and one buffer parse: 7097 ms
+  std::ifstream file_;
+  size_t queue_size_;
+  // The queue for store parsed data
+  std::unique_ptr<paddle::operators::reader::BlockingQueue<T>> queue_;
+};
+
+// This class define the data type of instance(ins_vec) in MultiSlotDataFeed
+class MultiSlotType {
+ public:
+  MultiSlotType() {}
+  ~MultiSlotType() {}
+  void Init(const std::string& type) {
+    CheckType(type);
+    if (type_[0] == 'f') {
+      float_feasign_.clear();
+    } else if (type_[0] == 'u') {
+      uint64_feasign_.clear();
+    }
+    type_ = type;
+  }
+  void InitOffset() {
+    offset_.resize(1);
+    // LoDTensor' lod is counted from 0, the size of lod
+    // is one size larger than the size of data.
+    offset_[0] = 0;
+  }
+  const std::vector<size_t>& GetOffset() const { return offset_; }
+  void AddValue(const float v) {
+    CheckFloat();
+    float_feasign_.push_back(v);
+  }
+  void AddValue(const uint64_t v) {
+    CheckUint64();
+    uint64_feasign_.push_back(v);
+  }
+  void AddIns(const MultiSlotType& ins) {
+    if (ins.GetType()[0] == 'f') {  // float
+      CheckFloat();
+      auto& vec = ins.GetFloatData();
+      offset_.push_back(offset_.back() + vec.size());
+      float_feasign_.insert(float_feasign_.end(), vec.begin(), vec.end());
+    } else if (ins.GetType()[0] == 'u') {  // uint64
+      CheckUint64();
+      auto& vec = ins.GetUint64Data();
+      offset_.push_back(offset_.back() + vec.size());
+      uint64_feasign_.insert(uint64_feasign_.end(), vec.begin(), vec.end());
+    }
+  }
+  const std::vector<float>& GetFloatData() const { return float_feasign_; }
+  const std::vector<uint64_t>& GetUint64Data() const { return uint64_feasign_; }
+  const std::string& GetType() const { return type_; }
+
+ private:
+  void CheckType(const std::string& type) const {
+    PADDLE_ENFORCE((type == "uint64") || (type == "float"),
+                   "There is no this type<%s>.", type);
+  }
+  void CheckFloat() const {
+    PADDLE_ENFORCE(type_[0] == 'f', "Add %s value to float slot.", type_);
+  }
+  void CheckUint64() const {
+    PADDLE_ENFORCE(type_[0] == 'u', "Add %s value to uint64 slot.", type_);
+  }
+  std::vector<float> float_feasign_;
+  std::vector<uint64_t> uint64_feasign_;
+  std::string type_;
+  std::vector<size_t> offset_;
+};
+
+// This DataFeed is used to feed multi-slot type data.
+// The format of multi-slot type data:
+//   [n feasign_0 feasign_1 ... feasign_n]*
+class MultiSlotDataFeed
+    : public PrivateQueueDataFeed<std::vector<MultiSlotType>> {
+ public:
+  MultiSlotDataFeed() {}
+  virtual ~MultiSlotDataFeed() {}
+  virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc);
+  virtual bool CheckFile(const char* filename);
+
+ protected:
+  virtual void AddInstanceToInsVec(std::vector<MultiSlotType>* vec_ins,
+                                   const std::vector<MultiSlotType>& instance,
+                                   int index);
+  virtual bool ParseOneInstance(std::vector<MultiSlotType>* instance);
+  virtual void PutToFeedVec(const std::vector<MultiSlotType>& ins_vec);
+};
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/data_feed.proto

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+syntax = "proto2";
+package paddle.framework;
+
+message Slot {
+  required string name = 1;
+  required string type = 2;
+  optional bool is_dense = 3 [ default = false ];
+  optional bool is_used = 4 [ default = false ];
+}
+
+message MultiSlotDesc { repeated Slot slots = 1; }
+
+message DataFeedDesc {
+  optional string name = 1;
+  optional int32 batch_size = 2 [ default = 32 ];
+  optional MultiSlotDesc multi_slot_desc = 3;
+}

paddle/fluid/framework/data_feed_factory.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+  http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/data_feed_factory.h"
+#include <memory>
+#include <string>
+#include <unordered_map>
+
+#include "paddle/fluid/framework/data_feed.h"
+
+namespace paddle {
+namespace framework {
+typedef std::shared_ptr<DataFeed> (*Createdata_feedFunction)();
+typedef std::unordered_map<std::string, Createdata_feedFunction> data_feedMap;
+data_feedMap g_data_feed_map;
+
+#define REGISTER_DATAFEED_CLASS(data_feed_class)                      \
+  namespace {                                                         \
+  std::shared_ptr<DataFeed> Creator_##data_feed_class() {             \
+    return std::shared_ptr<DataFeed>(new data_feed_class);            \
+  }                                                                   \
+  class __Registerer_##data_feed_class {                              \
+   public:                                                            \
+    __Registerer_##data_feed_class() {                                \
+      g_data_feed_map[#data_feed_class] = &Creator_##data_feed_class; \
+    }                                                                 \
+  };                                                                  \
+  __Registerer_##data_feed_class g_registerer_##data_feed_class;      \
+  }  // namespace
+
+std::string DataFeedFactory::DataFeedTypeList() {
+  std::string data_feed_types;
+  for (auto iter = g_data_feed_map.begin(); iter != g_data_feed_map.end();
+       ++iter) {
+    if (iter != g_data_feed_map.begin()) {
+      data_feed_types += ", ";
+    }
+    data_feed_types += iter->first;
+  }
+  return data_feed_types;
+}
+
+std::shared_ptr<DataFeed> DataFeedFactory::CreateDataFeed(
+    std::string data_feed_class) {
+  if (g_data_feed_map.count(data_feed_class) < 1) {
+    exit(-1);
+  }
+  return g_data_feed_map[data_feed_class]();
+}
+
+REGISTER_DATAFEED_CLASS(MultiSlotDataFeed);
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/data_feed_factory.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+  http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <memory>
+#include <string>
+#include "paddle/fluid/framework/data_feed.h"
+
+namespace paddle {
+namespace framework {
+class DataFeedFactory {
+ public:
+  static std::string DataFeedTypeList();
+  static std::shared_ptr<DataFeed> CreateDataFeed(std::string data_feed_class);
+};
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/data_feed_test.cc

+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/data_feed.h"
+#include <fcntl.h>
+#include <chrono>  // NOLINT
+#include <fstream>
+#include <iostream>
+#include <map>
+#include <mutex>  // NOLINT
+#include <set>
+#include <thread>  // NOLINT
+#include <utility>
+#include <vector>
+#include "google/protobuf/io/zero_copy_stream_impl.h"
+#include "google/protobuf/text_format.h"
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/data_feed_factory.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/scope.h"
+
+paddle::framework::DataFeedDesc load_datafeed_param_from_file(
+    const char* filename) {
+  paddle::framework::DataFeedDesc data_feed_desc;
+  int file_descriptor = open(filename, O_RDONLY);
+  PADDLE_ENFORCE(file_descriptor != -1, "Can not open %s.", filename);
+  google::protobuf::io::FileInputStream fileInput(file_descriptor);
+  google::protobuf::TextFormat::Parse(&fileInput, &data_feed_desc);
+  close(file_descriptor);
+  return data_feed_desc;
+}
+
+const std::vector<std::string> load_filelist_from_file(const char* filename) {
+  std::vector<std::string> filelist;
+  std::ifstream fin(filename);
+  PADDLE_ENFORCE(fin.good(), "Can not open %s.", filename);
+  std::string line;
+  while (getline(fin, line)) {
+    filelist.push_back(line);
+  }
+  fin.close();
+  return filelist;
+}
+
+void GenerateFileForTest(const char* protofile, const char* filelist) {
+  std::ofstream w_protofile(protofile);
+  w_protofile << "name: \"MultiSlotDataFeed\"\n"
+                 "batch_size: 2\n"
+                 "multi_slot_desc {\n"
+                 "    slots {\n"
+                 "        name: \"uint64_sparse_slot\"\n"
+                 "        type: \"uint64\"\n"
+                 "        is_dense: false\n"
+                 "        is_used: true\n"
+                 "    }\n"
+                 "    slots {\n"
+                 "        name: \"float_sparse_slot\"\n"
+                 "        type: \"float\"\n"
+                 "        is_dense: false\n"
+                 "        is_used: true\n"
+                 "    }\n"
+                 "    slots {\n"
+                 "        name: \"uint64_dense_slot\"\n"
+                 "        type: \"uint64\"\n"
+                 "        is_dense: true\n"
+                 "        is_used: true\n"
+                 "    }\n"
+                 "    slots {\n"
+                 "        name: \"float_dense_slot\"\n"
+                 "        type: \"float\"\n"
+                 "        is_dense: true\n"
+                 "        is_used: true\n"
+                 "    }\n"
+                 "    slots {\n"
+                 "        name: \"not_used_slot\"\n"
+                 "        type: \"uint64\"\n"
+                 "        is_dense: false\n"
+                 "        is_used: false\n"
+                 "    }\n"
+                 "}";
+  w_protofile.close();
+  std::ofstream w_filelist(filelist);
+  int total_file = 4;
+  for (int i = 0; i < total_file; ++i) {
+    std::string filename = "TestMultiSlotDataFeed.data." + std::to_string(i);
+    w_filelist << filename;
+    if (i + 1 != total_file) {
+      w_filelist << std::endl;
+    }
+    std::ofstream w_datafile(filename.c_str());
+    w_datafile << "3 3978 620 82 1 1926.08 1 1926 1 6.02 1 1996\n"
+                  "2 1300 2983353 1 985.211 1 8 1 0.618 1 12\n"
+                  "1 19260827 2 3.14 2.718 1 27 1 2.236 1 28\n";
+    w_datafile.close();
+  }
+  w_filelist.close();
+}
+
+class MultiTypeSet {
+ public:
+  MultiTypeSet() {
+    uint64_set_.clear();
+    float_set_.clear();
+  }
+  ~MultiTypeSet() {}
+  void AddValue(uint64_t v) { uint64_set_.insert(v); }
+  void AddValue(float v) { float_set_.insert(v); }
+  const std::set<uint64_t>& GetUint64Set() const { return uint64_set_; }
+  const std::set<float>& GetFloatSet() const { return float_set_; }
+
+ private:
+  std::set<uint64_t> uint64_set_;
+  std::set<float> float_set_;
+};
+
+void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
+                          const paddle::framework::DataFeedDesc& data_feed_desc,
+                          const std::vector<std::string>& filelist,
+                          const int thread_num) {
+  int used_slot_num = 0;
+  for (auto i = 0; i < data_feed_desc.multi_slot_desc().slots_size(); ++i) {
+    if (data_feed_desc.multi_slot_desc().slots(i).is_used()) {
+      ++used_slot_num;
+    }
+  }
+  reader_elem_set->resize(used_slot_num);
+  std::vector<std::thread> threads;
+  std::vector<std::shared_ptr<paddle::framework::DataFeed>> readers;
+  readers.resize(thread_num);
+  for (int i = 0; i < thread_num; ++i) {
+    readers[i] = paddle::framework::DataFeedFactory::CreateDataFeed(
+        data_feed_desc.name());
+    readers[i]->Init(data_feed_desc);
+  }
+  readers[0]->SetFileList(filelist);
+  std::mutex mu;
+  for (int idx = 0; idx < thread_num; ++idx) {
+    threads.emplace_back(std::thread([&, idx] {
+      std::unique_ptr<paddle::framework::Scope> scope(
+          new paddle::framework::Scope());
+      const auto& multi_slot_desc = data_feed_desc.multi_slot_desc();
+      std::map<std::string, const paddle::framework::LoDTensor*>
+          lodtensor_targets;
+      std::map<std::string, const paddle::framework::Tensor*> tensor_targets;
+      for (int i = 0; i < multi_slot_desc.slots_size(); ++i) {
+        const auto& slot = multi_slot_desc.slots(i);
+        if (slot.is_used()) {
+          const auto& name = slot.name();
+          readers[idx]->AddFeedVar(scope->Var(name), name);
+          if (slot.is_dense()) {
+            tensor_targets[name] =
+                &scope->FindVar(name)->Get<paddle::framework::Tensor>();
+          } else {
+            lodtensor_targets[name] =
+                &scope->FindVar(name)->Get<paddle::framework::LoDTensor>();
+          }
+        }
+      }
+      readers[idx]->Start();
+      while (readers[idx]->Next()) {
+        int index = 0;
+        for (int k = 0; k < multi_slot_desc.slots_size(); ++k) {
+          const auto& slot = multi_slot_desc.slots(k);
+          if (!slot.is_used()) {
+            continue;
+          }
+          if (slot.is_dense()) {  // dense branch
+            const paddle::framework::Tensor* tens = tensor_targets[slot.name()];
+            if (slot.type() == "uint64") {
+              const int64_t* data = tens->data<int64_t>();
+              int batch_size = tens->dims()[0];
+              int dim = tens->dims()[1];
+              for (int i = 0; i < batch_size; ++i) {
+                for (int j = 0; j < dim; ++j) {
+                  std::lock_guard<std::mutex> lock(mu);
+                  (*reader_elem_set)[index].AddValue(
+                      (uint64_t)data[i * dim + j]);
+                }
+              }
+            } else if (slot.type() == "float") {
+              const float* data = tens->data<float>();
+              int batch_size = tens->dims()[0];
+              int dim = tens->dims()[1];
+              for (int i = 0; i < batch_size; ++i) {
+                for (int j = 0; j < dim; ++j) {
+                  std::lock_guard<std::mutex> lock(mu);
+                  (*reader_elem_set)[index].AddValue(data[i * dim + j]);
+                }
+              }
+            } else {
+              PADDLE_THROW("Error type in proto file.");
+            }
+          } else {  // sparse branch
+            const paddle::framework::LoDTensor* tens =
+                lodtensor_targets[slot.name()];
+            if (slot.type() == "uint64") {
+              const int64_t* data = tens->data<int64_t>();
+              for (size_t i = 0; i < tens->NumElements(); ++i) {
+                std::pair<size_t, size_t> element = tens->lod_element(0, i);
+                for (size_t j = element.first; j < element.second; ++j) {
+                  std::lock_guard<std::mutex> lock(mu);
+                  (*reader_elem_set)[index].AddValue((uint64_t)data[j]);
+                }
+              }
+            } else if (slot.type() == "float") {
+              const float* data = tens->data<float>();
+              for (size_t i = 0; i < tens->NumElements(); ++i) {
+                std::pair<size_t, size_t> element = tens->lod_element(0, i);
+                for (size_t j = element.first; j < element.second; ++j) {
+                  std::lock_guard<std::mutex> lock(mu);
+                  (*reader_elem_set)[index].AddValue(data[j]);
+                }
+              }
+            } else {
+              PADDLE_THROW("Error type in proto file.");
+            }
+          }  // end sparse branch
+          ++index;
+        }  // end slots loop
+      }    // end while Next()
+    }));   // end anonymous function
+  }
+  for (auto& th : threads) {
+    th.join();
+  }
+}
+
+void CheckIsUnorderedSame(const std::vector<MultiTypeSet>& s1,
+                          const std::vector<MultiTypeSet>& s2) {
+  EXPECT_EQ(s1.size(), s2.size());
+  for (size_t i = 0; i < s1.size(); ++i) {
+    // check for uint64
+    const std::set<uint64_t>& uint64_s1 = s1[i].GetUint64Set();
+    const std::set<uint64_t>& uint64_s2 = s2[i].GetUint64Set();
+    EXPECT_EQ(uint64_s1.size(), uint64_s2.size());
+    auto uint64_it1 = uint64_s1.begin();
+    auto uint64_it2 = uint64_s2.begin();
+    while (uint64_it1 != uint64_s1.end()) {
+      EXPECT_EQ(*uint64_it1, *uint64_it2);
+      ++uint64_it1;
+      ++uint64_it2;
+    }
+    // check for float
+    const std::set<float>& float_s1 = s1[i].GetFloatSet();
+    const std::set<float>& float_s2 = s2[i].GetFloatSet();
+    EXPECT_EQ(float_s1.size(), float_s2.size());
+    auto float_it1 = float_s1.begin();
+    auto float_it2 = float_s2.begin();
+    while (float_it1 != float_s1.end()) {
+      EXPECT_EQ(*float_it1, *float_it2);
+      ++float_it1;
+      ++float_it2;
+    }
+  }
+}
+
+void GetElemSetFromFile(std::vector<MultiTypeSet>* file_elem_set,
+                        const paddle::framework::DataFeedDesc& data_feed_desc,
+                        const std::vector<std::string>& filelist) {
+  int used_slot_num = 0;
+  for (auto i = 0; i < data_feed_desc.multi_slot_desc().slots_size(); ++i) {
+    if (data_feed_desc.multi_slot_desc().slots(i).is_used()) {
+      ++used_slot_num;
+    }
+  }
+  file_elem_set->resize(used_slot_num);
+  for (const auto& file : filelist) {
+    std::ifstream fin(file.c_str());
+    PADDLE_ENFORCE(fin.good(), "Can not open %s.", file.c_str());
+    while (1) {
+      bool end_flag = false;
+      int index = 0;
+      for (auto i = 0; i < data_feed_desc.multi_slot_desc().slots_size(); ++i) {
+        int num;
+        if (fin >> num) {
+          auto slot = data_feed_desc.multi_slot_desc().slots(i);
+          auto type = slot.type();
+          if (type == "uint64") {
+            while (num--) {
+              uint64_t feasign;
+              fin >> feasign;
+              if (slot.is_used()) {
+                (*file_elem_set)[index].AddValue(feasign);
+              }
+            }
+          } else if (type == "float") {
+            while (num--) {
+              float feasign;
+              fin >> feasign;
+              if (slot.is_used()) {
+                (*file_elem_set)[index].AddValue(feasign);
+              }
+            }
+          } else {
+            PADDLE_THROW("Error type in proto file.");
+          }
+          if (slot.is_used()) {
+            ++index;
+          }
+        } else {
+          end_flag = true;
+          break;
+        }
+      }
+      if (end_flag) {
+        break;
+      }
+    }
+    fin.close();
+  }
+}
+
+TEST(DataFeed, MultiSlotUnitTest) {
+  const char* protofile = "data_feed_desc.prototxt";
+  const char* filelist_name = "filelist.txt";
+  GenerateFileForTest(protofile, filelist_name);
+  const std::vector<std::string> filelist =
+      load_filelist_from_file(filelist_name);
+  paddle::framework::DataFeedDesc data_feed_desc =
+      load_datafeed_param_from_file(protofile);
+  std::vector<MultiTypeSet> reader_elem_set;
+  std::vector<MultiTypeSet> file_elem_set;
+  GetElemSetFromReader(&reader_elem_set, data_feed_desc, filelist, 4);
+  GetElemSetFromFile(&file_elem_set, data_feed_desc, filelist);
+  CheckIsUnorderedSame(reader_elem_set, file_elem_set);
+}

paddle/fluid/framework/details/CMakeLists.txt

@@ -39,11 +39,12 @@ if (WITH_GPU)
 endif()
 
 cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass)
+cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
 
 cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
         scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
 
-set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass) 
+set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass) 
 if (WITH_GPU)
   list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
 endif()

paddle/fluid/framework/details/all_reduce_deps_pass.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <algorithm>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "paddle/fluid/framework/details/all_reduce_deps_pass.h"
+#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/details/op_graph_view.h"
+#include "paddle/fluid/framework/details/var_handle.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+static constexpr char kAllOpDescs[] = "all_op_descs";
+
+VarHandle* GetValidInput(const OpHandleBase* a) {
+  for (auto p : a->Inputs()) {
+    VarHandle* b = dynamic_cast<VarHandle*>(p);
+    if (b) {
+      return b;
+    }
+  }
+
+  return nullptr;
+}
+
+std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  auto graph_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
+
+  // get vars order
+  int order = 0;
+  std::unordered_map<std::string, int> vars;
+  // TODO(gongwb): use graph topology sort to find the order of operators.
+  //               Note that must assert topology sort is stable
+  auto& ops = Get<const std::vector<OpDesc*>>(kAllOpDescs);
+  for (auto* op_desc : ops) {
+    auto outputs = op_desc->Outputs();
+    for (auto& o_it : outputs) {
+      for (auto& v : o_it.second) {  // values
+        vars[v] = order;
+      }
+    }
+    order++;
+  }
+
+  std::vector<OpHandleBase*> dist_ops;
+  // get allreduce ops.
+  for (auto& op : graph_ops) {
+    // FIXME(gongwb):add broad cast.
+    if (op->Name() == "all_reduce" || op->Name() == "reduce") {
+      dist_ops.push_back(op);
+    }
+  }
+
+  VLOG(10) << "dist_ops size:" << dist_ops.size() << std::endl;
+
+  std::sort(dist_ops.begin(), dist_ops.end(), [&](OpHandleBase* op1,
+                                                  OpHandleBase* op2) {
+    VarHandle* i0 = dynamic_cast<VarHandle*>(GetValidInput(op1));
+    VarHandle* i1 = dynamic_cast<VarHandle*>(GetValidInput(op2));
+
+    PADDLE_ENFORCE(i0 != nullptr && i1 != nullptr, "%s convert to %s error",
+                   op1->DebugString(), op2->DebugString());
+
+    auto l_it = vars.find(i0->name_);
+    auto r_it = vars.find(i1->name_);
+
+    if (l_it->second < r_it->second) return true;
+
+    if (l_it->second == r_it->second) {
+      return i0->name_ < i1->name_;
+    }
+
+    return false;
+  });
+
+  // add dependency.
+  auto& sorted_ops = dist_ops;
+  for (size_t i = 1; i < sorted_ops.size(); ++i) {
+    auto* dep_var = new DummyVarHandle(graph->CreateControlDepVar());
+
+    auto* pre_op = sorted_ops[i - 1];
+    auto* op = sorted_ops[i];
+
+    pre_op->AddOutput(dep_var);
+    op->AddInput(dep_var);
+    graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
+
+    VLOG(10) << "add all_reduce sequential dependencies between " << pre_op
+             << " and " << op;
+
+    VLOG(10) << "pre_op:" << pre_op->DebugString()
+             << ", op:" << op->DebugString();
+  }
+
+  return graph;
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(all_reduce_deps_pass,
+              paddle::framework::details::AllReduceDepsPass)
+    .RequirePassAttr(paddle::framework::details::kAllOpDescs);

paddle/fluid/framework/details/all_reduce_deps_pass.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/pass.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+// TODO(gongwb): overlap allreduce with backward computation.
+class AllReduceDepsPass : public ir::Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -23,7 +23,7 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
                                      const std::vector<Scope *> &local_scopes,
                                      const std::vector<platform::Place> &places,
@@ -74,7 +74,7 @@ void AllReduceOpHandle::RunImpl() {
     }
 
     if (platform::is_gpu_place(lod_tensors[0]->place())) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
       int dtype = -1;
       size_t numel = 0;

paddle/fluid/framework/details/all_reduce_op_handle.h

@@ -20,7 +20,7 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -29,7 +29,7 @@ namespace framework {
 namespace details {
 
 struct AllReduceOpHandle : public OpHandleBase {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places,
                     const platform::NCCLContextMap *ctxs);
@@ -49,7 +49,7 @@ struct AllReduceOpHandle : public OpHandleBase {
  private:
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   const platform::NCCLContextMap *nccl_ctxs_;
 #endif
 };

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -60,7 +60,7 @@ void BroadcastOpHandle::BroadcastOneVar(
   PADDLE_ENFORCE_NOT_NULL(in_var);
   Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var);
   if (UNLIKELY(!in_tensor.IsInitialized())) {
-    VLOG(30) << "in var " << in_var_handle.name_ << "not inited, return!";
+    VLOG(3) << "in var " << in_var_handle.name_ << "not inited, return!";
     return;
   }
 
@@ -82,7 +82,7 @@ void BroadcastOpHandle::BroadcastOneVar(
       });
     }
   } else {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     VarHandle *out_handle = nullptr;
     int root_id = boost::get<platform::CUDAPlace>(in_tensor.place()).device;
     std::vector<std::function<void()>> broadcast_calls;

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -24,7 +24,7 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -34,7 +34,7 @@ namespace details {
 
 struct BroadcastOpHandle : public OpHandleBase {
  public:
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   BroadcastOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places,
                     const platform::NCCLContextMap *nccl_ctxs)
@@ -68,7 +68,7 @@ struct BroadcastOpHandle : public OpHandleBase {
 
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   const platform::NCCLContextMap *nccl_ctxs_;
 #endif
 

paddle/fluid/framework/details/broadcast_op_handle_test.h

@@ -42,7 +42,7 @@ struct TestBroadcastOpHandle {
   std::vector<std::unique_ptr<ir::Node>> nodes_;
   std::vector<p::Place> place_list_;
   bool use_gpu_;
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
 
@@ -50,7 +50,7 @@ struct TestBroadcastOpHandle {
     for (size_t j = 0; j < ctxs_.size(); ++j) {
       ctxs_[j]->Wait();
     }
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     if (nccl_ctxs_) {
       nccl_ctxs_->WaitAll();
     }
@@ -60,7 +60,7 @@ struct TestBroadcastOpHandle {
   void InitCtxOnGpu(bool use_gpu) {
     use_gpu_ = use_gpu;
     if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       int count = p::GetCUDADeviceCount();
       if (count <= 1) {
         LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
@@ -84,7 +84,7 @@ struct TestBroadcastOpHandle {
         place_list_.push_back(p);
         ctxs_.emplace_back(new p::CPUDeviceContext(p));
       }
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       nccl_ctxs_.reset(nullptr);
 #endif
     }
@@ -106,14 +106,14 @@ struct TestBroadcastOpHandle {
     nodes_.emplace_back(
         ir::CreateNodeForTest("node0", ir::Node::Type::kOperation));
     if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
                                          place_list_, nccl_ctxs_.get());
 #else
       PADDLE_THROW("CUDA is not support.");
 #endif
     } else {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
                                          place_list_, nccl_ctxs_.get());
 #else

paddle/fluid/framework/details/build_strategy.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
+#include "paddle/fluid/framework/details/reduce_op_handle.h"
 #include "paddle/fluid/framework/details/sequential_execution_pass.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_viz_pass.h"
@@ -24,6 +25,10 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+static inline bool SeqOnlyAllReduceOps(const BuildStrategy &strategy) {
+  return (!strategy.enable_sequential_execution_ && strategy.num_trainers_ > 1);
+}
+
 class ParallelExecutorPassBuilder : public ir::PassBuilder {
  public:
   explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy)
@@ -70,6 +75,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
     // Verify that the graph is correct for multi-device executor.
     AppendPass("multi_devices_check_pass");
 
+    if (SeqOnlyAllReduceOps(strategy)) {
+      AppendPass("all_reduce_deps_pass");
+    }
+
     if (strategy_.remove_unnecessary_lock_) {
       AppendPass("modify_op_lock_and_record_event_pass");
     }
@@ -96,7 +105,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &param_names,
     const std::vector<Scope *> &local_scopes,
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const {
 #else
     const bool use_cuda) const {
@@ -118,12 +127,23 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
       pass->Erase("local_scopes");
       pass->SetNotOwned<const std::vector<Scope *>>("local_scopes",
                                                     &local_scopes);
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
       pass->Erase("nccl_ctxs");
       pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
 #endif
     } else if (pass->Type() == "sequential_execution_pass") {
+      VLOG(1) << "set enable_sequential_execution:"
+              << enable_sequential_execution_;
+
+      pass->Erase(kAllOpDescs);
+      pass->Set<const std::vector<OpDesc *>>(
+          kAllOpDescs,
+          new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
+    } else if (pass->Type() == "all_reduce_deps_pass") {
+      VLOG(1) << "SeqOnlyAllReduceOps:" << SeqOnlyAllReduceOps(*this)
+              << ", num_trainers:" << num_trainers_;
+
       pass->Erase(kAllOpDescs);
       pass->Set<const std::vector<OpDesc *>>(
           kAllOpDescs,
@@ -144,4 +164,5 @@ USE_PASS(multi_devices_pass);
 USE_PASS(multi_devices_check_pass);
 USE_PASS(multi_devices_print_pass);
 USE_PASS(sequential_execution_pass);
+USE_PASS(all_reduce_deps_pass);
 USE_PASS(modify_op_lock_and_record_event_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -23,7 +23,7 @@
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -73,6 +73,7 @@ struct BuildStrategy {
 
   bool fuse_broadcast_op_{false};
 
+  int num_trainers_{1};
   bool remove_unnecessary_lock_{false};
 
   // NOTE:
@@ -98,7 +99,7 @@ struct BuildStrategy {
       const std::string &loss_var_name,
       const std::unordered_set<std::string> &param_names,
       const std::vector<Scope *> &local_scopes,
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const;
 #else
       const bool use_cuda) const;

paddle/fluid/framework/details/data_balance_op_handle.cc

@@ -20,7 +20,7 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 DataBalanceOpHandle::DataBalanceOpHandle(
     ir::Node *node, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places,

paddle/fluid/framework/details/data_balance_op_handle.h

@@ -19,7 +19,7 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -29,7 +29,7 @@ namespace details {
 
 struct DataBalanceOpHandle : public OpHandleBase {
  public:
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   DataBalanceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                       const std::vector<platform::Place> &places,
                       const platform::NCCLContextMap *ctxs);

paddle/fluid/framework/details/fused_broadcast_op_handle.h

@@ -25,7 +25,7 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -35,7 +35,7 @@ namespace details {
 
 struct FusedBroadcastOpHandle : public BroadcastOpHandle {
  public:
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   FusedBroadcastOpHandle(ir::Node *node,
                          const std::vector<Scope *> local_scopes,
                          const std::vector<platform::Place> &places,

paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc

@@ -44,14 +44,14 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
     nodes_.emplace_back(
         ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation));
     if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_ = new FusedBroadcastOpHandle(
           nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get());
 #else
       PADDLE_THROW("CUDA is not supported.");
 #endif
     } else {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_ = new FusedBroadcastOpHandle(
           nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get());
 #else

paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc

@@ -45,8 +45,8 @@ std::unique_ptr<ir::Graph> ModifyOpLockAndRecordEventPass::ApplyImpl(
         IsLockAndRecordEventFreeComputationOpHandle(compute_op, graph_view);
     compute_op->SetLockAndRecordEventFree(is_lock_and_record_event_free);
     if (is_lock_and_record_event_free) {
-      VLOG(100) << "Set is_lock_and_record_event_free be true in op "
-                << compute_op->DebugString();
+      VLOG(10) << "Set is_lock_and_record_event_free be true in op "
+               << compute_op->DebugString();
     }
   }
   return ir_graph;

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -142,7 +142,7 @@ void MultiDevSSAGraphBuilder::Init() const {
   places_ = Get<const std::vector<platform::Place>>(kPlaces);
   local_scopes_ = Get<const std::vector<Scope *>>(kLocalScopes);
   strategy_ = Get<const BuildStrategy>(kStrategy);
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   nccl_ctxs_ = &Get<platform::NCCLContextMap>("nccl_ctxs");
 #endif
 
@@ -399,7 +399,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
               for (size_t i = 0; i < backward_vars.size(); i += 2) {
                 auto &p_name = backward_vars[i];
                 auto &g_name = backward_vars[i + 1];
-                VLOG(100) << "Bcast " << g_name << " for parameter " << p_name;
+                VLOG(10) << "Bcast " << g_name << " for parameter " << p_name;
 
                 switch (strategy_.reduce_) {
                   case BuildStrategy::ReduceStrategy::kReduce:
@@ -431,7 +431,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
     }
   }
   bool use_gpu = false;
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   use_gpu = nccl_ctxs_ != nullptr;
 #endif
 
@@ -478,7 +478,7 @@ bool MultiDevSSAGraphBuilder::IsSparseGradient(const std::string &og) const {
 
 void MultiDevSSAGraphBuilder::SetCommunicationContext(
     OpHandleBase *op_handle, const platform::Place &p) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   if (nccl_ctxs_ == nullptr) {
     op_handle->SetDeviceContext(p,
                                 platform::DeviceContextPool::Instance().Get(p));
@@ -492,7 +492,7 @@ void MultiDevSSAGraphBuilder::SetCommunicationContext(
 void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
                                                 const std::string &p_name,
                                                 size_t src_dev_id) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   auto *op_handle = new BroadcastOpHandle(
       result->CreateEmptyNode("broadcast", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_);
@@ -522,7 +522,7 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
 void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp(
     ir::Graph *result,
     const std::vector<std::unordered_set<std::string>> &bcast_varnames) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   auto *op_handle = new FusedBroadcastOpHandle(
       result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_);
@@ -568,7 +568,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
 
 void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
                                                 const std::string &og) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
       result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_));
@@ -597,7 +597,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
 
 void MultiDevSSAGraphBuilder::InsertDataBalanceOp(
     ir::Graph *result, const std::vector<std::string> &datas) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   result->Get<GraphOps>(kGraphOps).emplace_back(new DataBalanceOpHandle(
       result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_));
@@ -694,7 +694,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result,
 VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
                                                    const std::string &og,
                                                    int dst_dev_id) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   result->Get<GraphOps>(kGraphOps).emplace_back(new ReduceOpHandle(
       result->CreateEmptyNode("reduce", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_));
@@ -809,8 +809,8 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(
           node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
       PADDLE_ENFORCE_EQ(send_param_grad.size(), 2U);
       op_dev_id = GetAppropriateDeviceID({send_param_grad[1]});
-      VLOG(100) << "send grad " << input_var_names[0] << " origin "
-                << send_param_grad[1] << " place: " << op_dev_id;
+      VLOG(10) << "send grad " << input_var_names[0] << " origin "
+               << send_param_grad[1] << " place: " << op_dev_id;
       for (auto &varname : input_var_names) {
         sharded_var_device->emplace(varname, op_dev_id);
       }
@@ -826,9 +826,9 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(
     if (recv_param_grad.size() == 2U) {
       op_dev_id =
           GetVarDeviceID(*result, recv_param_grad[1], *sharded_var_device);
-      VLOG(100) << "recv param " << recv_param_grad[0]
-                << " get grad place: " << recv_param_grad[1]
-                << " place: " << op_dev_id;
+      VLOG(10) << "recv param " << recv_param_grad[0]
+               << " get grad place: " << recv_param_grad[1]
+               << " place: " << op_dev_id;
     } else {
       op_dev_id = GetAppropriateDeviceID(output_var_names);
     }
@@ -862,7 +862,7 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(
       if (node->Op()->Type() == "fetch_barrier") {
         outvar_dev_id =
             GetVarDeviceID(*result, output->Name(), *sharded_var_device);
-        PADDLE_ENFORCE_NE(outvar_dev_id, -1);
+        PADDLE_ENFORCE_NE(outvar_dev_id, -1, "output name %s", output->Name());
       }
       p = places_[outvar_dev_id];
       ir::Node *new_node = nullptr;

paddle/fluid/framework/details/multi_devices_graph_pass.h

@@ -40,7 +40,7 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
                          size_t device_id) const;
   void Init() const;
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   mutable platform::NCCLContextMap *nccl_ctxs_;
 #endif
 

paddle/fluid/framework/details/op_registry.h

@@ -32,7 +32,9 @@ enum OpInfoFillType {
   kOpProtoAndCheckerMaker = 1,
   kGradOpDescMaker = 2,
   kVarTypeInference = 3,
-  kShapeInference = 4
+  kShapeInference = 4,
+  kEstimateFlops = 5,
+  kUnknown = -1
 };
 
 template <typename T>
@@ -48,8 +50,10 @@ struct OpInfoFillTypeID {
                                     ? kVarTypeInference
                                     : (std::is_base_of<InferShapeBase, T>::value
                                            ? kShapeInference
-                                           : static_cast<OpInfoFillType>(
-                                                 -1)))));
+                                           : (std::is_base_of<EstimateFlopsBase,
+                                                              T>::value
+                                                  ? kEstimateFlops
+                                                  : kUnknown)))));
   }
 };
 
@@ -139,6 +143,16 @@ struct OpInfoFiller<T, kShapeInference> {
   }
 };
 
+template <typename T>
+struct OpInfoFiller<T, kEstimateFlops> {
+  void operator()(const char* op_tpe, OpInfo* info) const {
+    info->estimate_flops_ = [](InferShapeContext* ctx) {
+      T estimate_flops;
+      return estimate_flops(ctx);
+    };
+  }
+};
+
 }  // namespace details
 
 }  // namespace framework

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -125,7 +125,7 @@ void ReduceOpHandle::RunImpl() {
         }
       });
     } else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       auto pre_in = pre_in_var->Get<framework::LoDTensor>();
       VariableVisitor::ShareDimsAndLoD(*pre_in_var, out_var);
       VariableVisitor::GetMutableTensor(out_var).mutable_data(

paddle/fluid/framework/details/reduce_op_handle.h

@@ -23,7 +23,7 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -35,7 +35,7 @@ struct ReduceOpHandle : public OpHandleBase {
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   const platform::NCCLContextMap *nccl_ctxs_;
   ReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                  const std::vector<platform::Place> &places,

paddle/fluid/framework/details/reduce_op_handle_test.cc

@@ -35,7 +35,7 @@ struct TestReduceOpHandle {
   std::vector<p::Place> gpu_list_;
   std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
 
@@ -43,7 +43,7 @@ struct TestReduceOpHandle {
     for (size_t j = 0; j < ctxs_.size(); ++j) {
       ctxs_[j]->Wait();
     }
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     if (nccl_ctxs_) {
       nccl_ctxs_->WaitAll();
     }
@@ -53,7 +53,7 @@ struct TestReduceOpHandle {
   void InitCtxOnGpu(bool use_gpu) {
     use_gpu_ = use_gpu;
     if (use_gpu) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       int count = p::GetCUDADeviceCount();
       if (count <= 1) {
         LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
@@ -77,7 +77,7 @@ struct TestReduceOpHandle {
         gpu_list_.push_back(p);
         ctxs_.emplace_back(new p::CPUDeviceContext(p));
       }
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       nccl_ctxs_.reset(nullptr);
 #endif
     }
@@ -99,14 +99,14 @@ struct TestReduceOpHandle {
 
     nodes.emplace_back(new ir::Node("node"));
     if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_,
                                           gpu_list_, nccl_ctxs_.get()));
 #else
       PADDLE_THROW("CUDA is not support.");
 #endif
     } else {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_,
                                           gpu_list_, nccl_ctxs_.get()));
 #else

paddle/fluid/framework/details/reference_count_pass.cc

@@ -140,8 +140,8 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
         if (next_compute_op != nullptr) {
           if (compute_ref_cnt_map.count(next_compute_op)) {
             compute_ref_cnt_map[next_compute_op]->AddVar(var_name);
-            VLOG(50) << "Add reference count of " << var_name << " to Operator "
-                     << next_compute_op->Name();
+            VLOG(5) << "Add reference count of " << var_name << " to Operator "
+                    << next_compute_op->Name();
           } else {
             // Create new reference_count_op_handle
             ir::Node *ref_cnt_node = graph->CreateEmptyNode(

paddle/fluid/framework/details/scale_loss_grad_op_handle.cc

@@ -51,7 +51,7 @@ void ScaleLossGradOpHandle::RunImpl() {
                         ->stream();
       memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
                    platform::CPUPlace(), &coeff_, sizeof(float), stream);
-      VLOG(100) << place_ << "RUN Scale loss grad op";
+      VLOG(10) << place_ << "RUN Scale loss grad op";
     });
 #endif
   }

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc

@@ -16,7 +16,7 @@
 #include <stdexcept>
 #include <string>
 #include <vector>
-#include "paddle/fluid/framework/executor.h"
+#include "paddle/fluid/framework/variable_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 #ifdef PADDLE_WITH_CUDA
 #include "paddle/fluid/framework/details/reference_count_op_handle.h"

paddle/fluid/framework/details/sequential_execution_pass.cc

@@ -94,8 +94,8 @@ std::unique_ptr<ir::Graph> SequentialExecutionPass::ApplyImpl(
     op_node_list[i - 1]->outputs.push_back(dep_var);
     dep_var->outputs.push_back(op_node_list[i]);
     dep_var->inputs.push_back(op_node_list[i - 1]);
-    VLOG(100) << "Add dependencies between " << op_node_list[i - 1]->Name()
-              << " and " << op_node_list[i]->Name();
+    VLOG(10) << "Add dependencies between " << op_node_list[i - 1]->Name()
+             << " and " << op_node_list[i]->Name();
   }
   return graph;
 }

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -210,16 +210,16 @@ void ThreadedSSAGraphExecutor::RunOp(
     details::OpHandleBase *op) {
   auto op_run = [ready_var_q, op, this] {
     try {
-      if (VLOG_IS_ON(100)) {
-        VLOG(100) << op << " " << op->Name() << " : " << op->DebugString();
+      if (VLOG_IS_ON(10)) {
+        VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
       }
       if (LIKELY(!strategy_.dry_run_)) {
         op->Run(strategy_.use_cuda_);
       }
-      VLOG(100) << op << " " << op->Name() << " Done ";
+      VLOG(10) << op << " " << op->Name() << " Done ";
       running_ops_--;
       ready_var_q->Extend(op->Outputs());
-      VLOG(100) << op << " " << op->Name() << "Signal posted";
+      VLOG(10) << op << " " << op->Name() << "Signal posted";
     } catch (...) {
       exception_holder_.Catch(std::current_exception());
     }

paddle/fluid/framework/executor.cc

@@ -21,6 +21,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/transfer_scope_cache.h"
+#include "paddle/fluid/framework/variable_helper.h"
 #include "paddle/fluid/operators/detail/macros.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -46,7 +47,7 @@ ExecutorPrepareContext::ExecutorPrepareContext(
 }
 
 ExecutorPrepareContext::~ExecutorPrepareContext() {
-  VLOG(50) << "destroy ExecutorPrepareContext";
+  VLOG(5) << "destroy ExecutorPrepareContext";
 }
 
 template <typename RefCntMap>
@@ -63,7 +64,7 @@ static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
         if ((it->second)-- == 1) {
           auto* var = scope.FindVar(name);
           if (var != nullptr) {
-            VLOG(100) << "Erase tensor \'" << name << "\'";
+            VLOG(10) << "Erase tensor \'" << name << "\'";
             if (var->IsType<LoDTensor>()) {
               erase_tensors.insert(var->GetMutable<LoDTensor>());
             } else if (var->IsType<SelectedRows>()) {
@@ -114,36 +115,6 @@ void Executor::Close() {
 #endif
 }
 
-void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
-  if (var_type == proto::VarType::LOD_TENSOR) {
-    var->GetMutable<LoDTensor>();
-  } else if (var_type == proto::VarType::SELECTED_ROWS) {
-    var->GetMutable<SelectedRows>();
-  } else if (var_type == proto::VarType::FEED_MINIBATCH) {
-    var->GetMutable<FeedFetchList>();
-  } else if (var_type == proto::VarType::FETCH_LIST) {
-    var->GetMutable<FeedFetchList>();
-  } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope*>>();
-  } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
-    var->GetMutable<LoDRankTable>();
-  } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
-    var->GetMutable<LoDTensorArray>();
-  } else if (var_type == proto::VarType::PLACE_LIST) {
-    var->GetMutable<platform::PlaceList>();
-  } else if (var_type == proto::VarType::READER) {
-    var->GetMutable<ReaderHolder>();
-  } else if (var_type == proto::VarType::RAW) {
-    // GetMutable will be called in operator
-  } else {
-    PADDLE_THROW(
-        "Variable type %d is not in "
-        "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
-        "LOD_RANK_TABLE, PLACE_LIST, READER, RAW]",
-        var_type);
-  }
-}
-
 void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
                                int block_id) {
   auto& global_block = pdesc.Block(block_id);
@@ -162,21 +133,21 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
       if (var->Persistable()) {
         auto* ptr = const_cast<Scope*>(ancestor_scope)->Var(var->Name());
         InitializeVariable(ptr, var->GetType());
-        VLOG(30) << "Create Variable " << var->Name()
-                 << " global, which pointer is " << ptr;
+        VLOG(3) << "Create Variable " << var->Name()
+                << " global, which pointer is " << ptr;
       } else {
         auto* ptr = scope->Var(var->Name());
         InitializeVariable(ptr, var->GetType());
-        VLOG(30) << "Create Variable " << var->Name()
-                 << " locally, which pointer is " << ptr;
+        VLOG(3) << "Create Variable " << var->Name()
+                << " locally, which pointer is " << ptr;
       }
     }
   } else {
     for (auto& var : global_block.AllVars()) {
       auto* ptr = scope->Var(var->Name());
       InitializeVariable(ptr, var->GetType());
-      VLOG(30) << "Create variable " << var->Name() << ", which pointer is "
-               << ptr;
+      VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
+              << ptr;
     }
   }
 }
@@ -307,7 +278,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
     int i = 0;
     for (auto& feed_target : (*feed_targets)) {
       std::string var_name = feed_target.first;
-      VLOG(30) << "feed target's name: " << var_name;
+      VLOG(3) << "feed target's name: " << var_name;
 
       // prepend feed op
       auto* op = global_block->PrependOp();
@@ -330,7 +301,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
     int i = 0;
     for (auto& fetch_target : (*fetch_targets)) {
       std::string var_name = fetch_target.first;
-      VLOG(30) << "fetch target's name: " << var_name;
+      VLOG(3) << "fetch target's name: " << var_name;
 
       // append fetch op
       auto* op = global_block->AppendOp();
@@ -392,8 +363,8 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
 
   int64_t max_memory_size = GetEagerDeletionThreshold();
   std::unique_ptr<GarbageCollector<Tensor>> gc;
-  // WhileOp would set keep_kids to false
-  // WhileGradOp would need the scopes created in WhileOp
+  // WhileOp would set keep_kids to true,
+  // because WhileGradOp needs the scopes created in WhileOp.
   // Perhaps, we should not perform eager deletion in WhileOp
   // The scopes and variables created by WhileOp would be deleted
   // in WhileGradOp.
@@ -482,7 +453,7 @@ void Executor::RunPreparedContext(
 
 void Executor::EnableMKLDNN(const ProgramDesc& program) {
 #ifdef PADDLE_WITH_MKLDNN
-  VLOG(30) << "use_mkldnn=True";
+  VLOG(3) << "use_mkldnn=True";
   for (size_t bid = 0; bid < program.Size(); ++bid) {
     auto* block = const_cast<ProgramDesc&>(program).MutableBlock(bid);
     for (auto* op : block->AllOps()) {

paddle/fluid/framework/executor.h

@@ -26,7 +26,6 @@ limitations under the License. */
 
 namespace paddle {
 namespace framework {
-extern void InitializeVariable(Variable* var, proto::VarType::Type var_type);
 
 template <typename T>
 std::unordered_map<std::string, T> GetNonPersistableReferenceCount(

paddle/fluid/framework/executor_thread_worker.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/executor_thread_worker.h"
+#include "google/protobuf/io/zero_copy_stream_impl.h"
+#include "google/protobuf/message.h"
+#include "google/protobuf/text_format.h"
+
+#include "gflags/gflags.h"
+#include "paddle/fluid/framework/feed_fetch_method.h"
+#include "paddle/fluid/framework/feed_fetch_type.h"
+#include "paddle/fluid/framework/lod_rank_table.h"
+#include "paddle/fluid/framework/lod_tensor_array.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/reader.h"
+#include "paddle/fluid/framework/variable_helper.h"
+#include "paddle/fluid/inference/io.h"
+#include "paddle/fluid/platform/place.h"
+#include "paddle/fluid/pybind/pybind.h"
+namespace paddle {
+namespace framework {
+
+void ExecutorThreadWorker::CreateThreadOperators(const ProgramDesc& program) {
+  auto& block = program.Block(0);
+  op_names_.clear();
+  for (auto& op_desc : block.AllOps()) {
+    std::unique_ptr<OperatorBase> local_op = OpRegistry::CreateOp(*op_desc);
+    op_names_.push_back(op_desc->Type());
+    OperatorBase* local_op_ptr = local_op.release();
+    ops_.push_back(local_op_ptr);
+    continue;
+  }
+}
+
+void ExecutorThreadWorker::CreateThreadResource(
+    const framework::ProgramDesc& program,
+    const paddle::platform::Place& place) {
+  CreateThreadScope(program);
+  CreateThreadOperators(program);
+  SetMainProgram(program);
+  SetPlace(place);
+}
+
+void ExecutorThreadWorker::CreateThreadScope(const ProgramDesc& program) {
+  auto& block = program.Block(0);
+
+  PADDLE_ENFORCE_NOT_NULL(
+      root_scope_, "root_scope should be set before creating thread scope");
+
+  thread_scope_ = &root_scope_->NewScope();
+  for (auto& var : block.AllVars()) {
+    if (var->Persistable()) {
+      auto* ptr = root_scope_->Var(var->Name());
+      InitializeVariable(ptr, var->GetType());
+    } else {
+      auto* ptr = thread_scope_->Var(var->Name());
+      InitializeVariable(ptr, var->GetType());
+    }
+  }
+}
+
+void ExecutorThreadWorker::SetDataFeed(
+    const std::shared_ptr<DataFeed>& datafeed) {
+  thread_reader_ = datafeed;
+}
+
+void ExecutorThreadWorker::BindingDataFeedMemory() {
+  const std::vector<std::string>& input_feed =
+      thread_reader_->GetUseSlotAlias();
+  for (auto name : input_feed) {
+    thread_reader_->AddFeedVar(thread_scope_->Var(name), name);
+  }
+}
+
+void ExecutorThreadWorker::SetFetchVarNames(
+    const std::vector<std::string>& fetch_var_names) {
+  fetch_var_names_.clear();
+  fetch_var_names_.insert(fetch_var_names_.end(), fetch_var_names.begin(),
+                          fetch_var_names.end());
+}
+
+void ExecutorThreadWorker::SetDevice() {
+#if defined _WIN32 || defined __APPLE__
+  return;
+#else
+  static unsigned concurrency_cap = std::thread::hardware_concurrency();
+  int thread_id = this->thread_id_;
+
+  if (thread_id < concurrency_cap) {
+    unsigned proc = thread_id;
+
+    cpu_set_t mask;
+    CPU_ZERO(&mask);
+    CPU_SET(proc, &mask);
+
+    if (-1 == sched_setaffinity(0, sizeof(mask), &mask)) {
+      VLOG(1) << "WARNING: Failed to set thread affinity for thread "
+              << thread_id;
+    } else {
+      CPU_ZERO(&mask);
+      if ((0 != sched_getaffinity(0, sizeof(mask), &mask)) ||
+          (CPU_ISSET(proc, &mask) == 0)) {
+        VLOG(3) << "WARNING: Failed to set thread affinity for thread "
+                << thread_id;
+      }
+    }
+  } else {
+    VLOG(1) << "WARNING: Failed to set thread affinity for thread "
+            << thread_id;
+  }
+#endif
+}
+
+template <typename T>
+void print_lod_tensor(std::string var_name, const LoDTensor& lod_tensor) {
+  auto inspect = lod_tensor.data<T>();
+  auto element_num = lod_tensor.numel();
+
+  std::ostringstream sstream;
+  sstream << var_name << " (element num " << element_num << "): [";
+  sstream << inspect[0];
+  for (int j = 1; j < element_num; ++j) {
+    sstream << " " << inspect[j];
+  }
+  sstream << "]";
+
+  std::cout << sstream.str() << std::endl;
+}
+
+void print_fetch_var(Scope* scope, std::string var_name) {
+  const LoDTensor& tensor = scope->FindVar(var_name)->Get<LoDTensor>();
+
+  if (std::type_index(tensor.type()) ==
+      std::type_index(typeid(platform::float16))) {
+    print_lod_tensor<platform::float16>(var_name, tensor);
+  } else if (std::type_index(tensor.type()) == std::type_index(typeid(float))) {
+    print_lod_tensor<float>(var_name, tensor);
+  } else if (std::type_index(tensor.type()) ==
+             std::type_index(typeid(double))) {
+    print_lod_tensor<double>(var_name, tensor);
+  } else if (std::type_index(tensor.type()) == std::type_index(typeid(int))) {
+    print_lod_tensor<int>(var_name, tensor);
+  } else if (std::type_index(tensor.type()) ==
+             std::type_index(typeid(int64_t))) {
+    print_lod_tensor<int64_t>(var_name, tensor);
+  } else if (std::type_index(tensor.type()) == std::type_index(typeid(bool))) {
+    print_lod_tensor<bool>(var_name, tensor);
+  } else if (std::type_index(tensor.type()) ==
+             std::type_index(typeid(uint8_t))) {
+    print_lod_tensor<uint8_t>(var_name, tensor);
+  } else if (std::type_index(tensor.type()) ==
+             std::type_index(typeid(int16_t))) {
+    print_lod_tensor<int16_t>(var_name, tensor);
+  } else if (std::type_index(tensor.type()) ==
+             std::type_index(typeid(int8_t))) {
+    print_lod_tensor<int8_t>(var_name, tensor);
+  } else {
+    VLOG(1) << "print_fetch_var: unrecognized data type:"
+            << tensor.type().name();
+  }
+
+  return;
+}
+
+void ExecutorThreadWorker::TrainFiles() {
+  // todo: configurable
+  SetDevice();
+
+  int fetch_var_num = fetch_var_names_.size();
+  fetch_values_.clear();
+  fetch_values_.resize(fetch_var_num);
+
+  thread_reader_->Start();
+
+  int cur_batch;
+  int batch_cnt = 0;
+  while ((cur_batch = thread_reader_->Next()) > 0) {
+    // executor run here
+    for (auto& op : ops_) {
+      op->Run(*thread_scope_, place_);
+    }
+
+    ++batch_cnt;
+    thread_scope_->DropKids();
+
+    if (debug_ == false || thread_id_ != 0) {
+      continue;
+    }
+
+    for (int i = 0; i < fetch_var_num; ++i) {
+      print_fetch_var(thread_scope_, fetch_var_names_[i]);
+    }  // end for (int i = 0...)
+  }    // end while ()
+}
+
+void ExecutorThreadWorker::SetThreadId(int tid) { thread_id_ = tid; }
+
+void ExecutorThreadWorker::SetPlace(const platform::Place& place) {
+  place_ = place;
+}
+
+void ExecutorThreadWorker::SetMainProgram(
+    const ProgramDesc& main_program_desc) {
+  main_program_.reset(new ProgramDesc(main_program_desc));
+}
+
+void ExecutorThreadWorker::SetRootScope(Scope* g_scope) {
+  root_scope_ = g_scope;
+}
+
+}  // einit_modelnd namespace framework
+}  // end namespace paddle

paddle/fluid/framework/executor_thread_worker.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+  http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <map>
+#include <memory>
+#include <mutex>  // NOLINT
+#include <set>
+#include <string>
+#include <thread>  // NOLINT
+#include <vector>
+#include "paddle/fluid/framework/data_feed.h"
+#include "paddle/fluid/framework/executor.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/scope.h"
+
+namespace paddle {
+namespace framework {
+void CreateTensor(Variable* var, proto::VarType::Type var_type);
+
+class ExecutorThreadWorker {
+ public:
+  ExecutorThreadWorker()
+      : thread_id_(-1), root_scope_(NULL), thread_scope_(NULL), debug_(false) {}
+  ~ExecutorThreadWorker() {}
+
+  void CreateThreadResource(const framework::ProgramDesc& program,
+                            const paddle::platform::Place& place);
+  void SetThreadId(int tid);
+  void SetDebug(const bool debug) { debug_ = debug; }
+  void SetRootScope(Scope* g_scope);
+  // set cpu device in this function
+  // cpu binding is used by default
+  void SetDevice();
+  // since we read data into memory that can not be accessed by program
+  // we need to bind memory of data with corresponding variables in program
+  // this function should be called after data feed is set
+  void BindingDataFeedMemory();
+  // set data feed declared in executor
+  void SetDataFeed(const std::shared_ptr<DataFeed>& datafeed);
+  // A multi-thread training function
+  void TrainFiles();
+  // set fetch variable names from python interface assigned by users
+  void SetFetchVarNames(const std::vector<std::string>& fetch_var_names);
+
+ private:
+  void CreateThreadScope(const framework::ProgramDesc& program);
+  void CreateThreadOperators(const framework::ProgramDesc& program);
+  void SetMainProgram(const ProgramDesc& main_program_desc);
+  void SetPlace(const paddle::platform::Place& place);
+
+ protected:
+  // thread index
+  std::shared_ptr<DataFeed> thread_reader_;  // shared queue, thread buffer
+  int thread_id_;
+  // operator name
+  std::vector<std::string> op_names_;
+  // thread level, local operators for forward and backward
+  std::vector<OperatorBase*> ops_;
+  // main program for training
+  std::unique_ptr<framework::ProgramDesc> main_program_;
+  // execution place
+  platform::Place place_;
+  // root scope for model parameters
+  Scope* root_scope_;
+  // a thread scope, father scope is global score which is shared
+  Scope* thread_scope_;
+
+ private:
+  std::vector<std::string> fetch_var_names_;
+  std::vector<std::vector<float>> fetch_values_;
+  bool debug_;
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/feed_fetch_method.cc

@@ -25,7 +25,7 @@ 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
   // be created.
-  VLOG(30) << "SetFeedVariable name=" << var_name << " index=" << index;
+  VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
   Variable* g_feed_value = scope->Var(var_name);
   auto& feed_inputs = *(g_feed_value->GetMutable<FeedFetchList>());
   if (index >= feed_inputs.size()) {
@@ -47,8 +47,8 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
                  typeid(FeedFetchList).name());
   auto& fetch_outputs = *g_fetch_value->GetMutable<FeedFetchList>();
   auto& tensor = fetch_outputs[index];
-  VLOG(30) << "Fetch " << var_name << " with index " << index
-           << " shape= " << tensor.dims();
+  VLOG(3) << "Fetch " << var_name << " with index " << index
+          << " shape= " << tensor.dims();
   PADDLE_ENFORCE_LT(index, fetch_outputs.size());
   return tensor;
 }

paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc

@@ -147,19 +147,19 @@ void PrepareParameters(Graph* graph, const Param& param) {
   scope->Var(param.LSTMX)->GetMutable<LoDTensor>();
   scope->Var(param.LSTMOUT)->GetMutable<LoDTensor>();
 
-#define GATE_W(name__)                                                \
-  auto* W_##name__##_w0 = scope->FindVar(#name__ ".w_0");             \
-  auto* W_##name__##_w1 = scope->FindVar(#name__ ".w_1");             \
-  auto* W_##name__##_b0 = scope->FindVar(#name__ ".b_0");             \
-  CHECK_P3(W_##name__##_w0, W_##name__##_w1, W_##name__##_b0);        \
-  VLOG(40) << #name__ "_w0"                                           \
-           << " shape: " << W_##name__##_w0->Get<LoDTensor>().dims(); \
-  VLOG(40) << #name__ "_w1"                                           \
-           << " shape: " << W_##name__##_w1->Get<LoDTensor>().dims(); \
-  VLOG(40) << #name__ "_b0"                                           \
-           << " shape: " << W_##name__##_b0->Get<LoDTensor>().dims(); \
-  auto& W_##name__##_w0_t = W_##name__##_w0->Get<LoDTensor>();        \
-  auto& W_##name__##_w1_t = W_##name__##_w1->Get<LoDTensor>();        \
+#define GATE_W(name__)                                               \
+  auto* W_##name__##_w0 = scope->FindVar(#name__ ".w_0");            \
+  auto* W_##name__##_w1 = scope->FindVar(#name__ ".w_1");            \
+  auto* W_##name__##_b0 = scope->FindVar(#name__ ".b_0");            \
+  CHECK_P3(W_##name__##_w0, W_##name__##_w1, W_##name__##_b0);       \
+  VLOG(4) << #name__ "_w0"                                           \
+          << " shape: " << W_##name__##_w0->Get<LoDTensor>().dims(); \
+  VLOG(4) << #name__ "_w1"                                           \
+          << " shape: " << W_##name__##_w1->Get<LoDTensor>().dims(); \
+  VLOG(4) << #name__ "_b0"                                           \
+          << " shape: " << W_##name__##_b0->Get<LoDTensor>().dims(); \
+  auto& W_##name__##_w0_t = W_##name__##_w0->Get<LoDTensor>();       \
+  auto& W_##name__##_w1_t = W_##name__##_w1->Get<LoDTensor>();       \
   auto& W_##name__##_b0_t = W_##name__##_b0->Get<LoDTensor>();
 
   GATE_W(forget);
@@ -208,7 +208,7 @@ void PrepareLSTMWeight(const LoDTensor& W_forget_w0,
   int D = W_forget_w0.dims()[0];
   int M = W_forget_w1.dims()[0];
   out->Resize(make_ddim({D + M, 4 * D}));
-  VLOG(30) << "LSTMWeight resized to " << out->dims();
+  VLOG(3) << "LSTMWeight resized to " << out->dims();
 
   float* out_data = out->mutable_data<float>(platform::CPUPlace());
   std::array<const float*, 4> tensors{

paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc

@@ -57,7 +57,7 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
   int found_conv_bias_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
-    VLOG(40) << "handle ConvBias fuse";
+    VLOG(4) << "handle ConvBias fuse";
     GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight,
                               conv_bias_pattern);                      // Filter
     GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_bias_pattern);  // tmp
@@ -74,7 +74,7 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
     // check if fuse can be done and if MKL-DNN should be used
     FuseOptions fuse_option = FindFuseOption(*conv, *eltwise);
     if (fuse_option == DO_NOT_FUSE || fuse_option == FUSE_NATIVE) {
-      VLOG(30) << "do not perform conv+bias fuse";
+      VLOG(3) << "do not perform conv+bias fuse";
       return;
     }
 

paddle/fluid/framework/ir/conv_bn_fuse_pass.cc

@@ -121,7 +121,7 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
   int found_conv_bn_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
-    VLOG(40) << "handle ConvBN fuse";
+    VLOG(4) << "handle ConvBN fuse";
 
     // conv, batch_norm,
     // conv_weight, conv_out,
@@ -133,7 +133,7 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
     // check if fuse can be done and if MKL-DNN should be used
     FuseOptions fuse_option = FindFuseOption(*conv, *batch_norm);
     if (fuse_option == DO_NOT_FUSE) {
-      VLOG(30) << "do not perform conv+bn fuse";
+      VLOG(3) << "do not perform conv+bn fuse";
       return;
     }
 
@@ -241,7 +241,7 @@ std::unique_ptr<ir::Graph> ConvEltwiseAddBNFusePass::ApplyImpl(
   int found_conv_bn_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
-    VLOG(40) << "handle ConvBN fuse";
+    VLOG(4) << "handle ConvBN fuse";
 
     // conv, batch_norm,
     // conv_weight, conv_out,

paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc

@@ -38,7 +38,7 @@ std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
   int found_conv_relu_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
-    VLOG(40) << "handle ConvReLU fuse";
+    VLOG(4) << "handle ConvReLU fuse";
     GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight,
                               conv_relu_pattern);                      // Filter
     GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_relu_pattern);  // tmp
@@ -48,7 +48,7 @@ std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
 
     FuseOptions fuse_option = FindFuseOption(*conv, *relu);
     if (fuse_option == DO_NOT_FUSE) {
-      VLOG(30) << "do not perform conv+relu fuse";
+      VLOG(3) << "do not perform conv+relu fuse";
       return;
     }
 

paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc

@@ -39,7 +39,7 @@ std::unique_ptr<ir::Graph> DepthwiseConvMKLDNNPass::ApplyImpl(
   int found_depthwise_conv_mkldnn_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
-    VLOG(30) << "handle DepthwiseConvMKLDNN fuse";
+    VLOG(3) << "handle DepthwiseConvMKLDNN fuse";
     GET_NODE(depthwise_conv, (*pattern));
     depthwise_conv->Op()->SetType("conv2d");
     found_depthwise_conv_mkldnn_count++;

paddle/fluid/framework/ir/fc_fuse_pass.cc

@@ -39,7 +39,7 @@ std::unique_ptr<ir::Graph> FCFusePass::ApplyImpl(
   int found_fc_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
-    VLOG(40) << "handle FC fuse";
+    VLOG(4) << "handle FC fuse";
     GET_IR_NODE_FROM_SUBGRAPH(w, w, fc_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(fc_bias, bias, fc_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(fc_out, Out, fc_pattern);

paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc

@@ -61,7 +61,7 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddAct(
 
   auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph,
                      Graph *g) {
-    VLOG(40) << "handle FuseElewiseAddAct fuse";
+    VLOG(4) << "handle FuseElewiseAddAct fuse";
     GET_IR_NODE_FROM_SUBGRAPH(ele_y, ele_y, elewise_add_act_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(ele_out, elewise_add_out,
                               elewise_add_act_pattern);
@@ -77,10 +77,10 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddAct(
     Node *elewise_add_act_node = CreateFuseElewiseAddActNode(
         g, act, ele_add, ele_x_n, ele_y_n, ele_out_n, act_out_n);
 
-    VLOG(40) << "\n\t " << ele_x_n << " and " << ele_y_n << " -> "
-             << ele_add->Name() << " -> " << ele_out_n << "\n"
-             << "\t " << ele_out_n << " -> " << act->Name() << " -> "
-             << act_out_n;
+    VLOG(4) << "\n\t " << ele_x_n << " and " << ele_y_n << " -> "
+            << ele_add->Name() << " -> " << ele_out_n << "\n"
+            << "\t " << ele_out_n << " -> " << act->Name() << " -> "
+            << act_out_n;
 
     ReLinkNodes(g, ele_out, ele_add, act, elewise_add_act_node);
     found_elewise_add_act_count++;
@@ -113,7 +113,7 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseActElewiseAdd(
 
   auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph,
                      Graph *g) {
-    VLOG(40) << "handle FuseElewiseAddAct fuse";
+    VLOG(4) << "handle FuseElewiseAddAct fuse";
     GET_IR_NODE_FROM_SUBGRAPH(act_out, act_out, act_elewise_add_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(ele_x, ele_x, act_elewise_add_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(ele_out, elewise_add_out,
@@ -129,9 +129,9 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseActElewiseAdd(
     Node *elewise_add_act_node = CreateFuseElewiseAddActNode(
         g, ele_add, act, elewise_add_x_n, act_i_n, act_o_n, elewise_add_out_n);
 
-    VLOG(40) << "\n\t " << act_i_n << " -> " << act->Name() << " -> " << act_o_n
-             << "\n\t " << act_o_n << " and " << elewise_add_x_n << " -> "
-             << ele_add->Name() << " -> " << elewise_add_out_n;
+    VLOG(4) << "\n\t " << act_i_n << " -> " << act->Name() << " -> " << act_o_n
+            << "\n\t " << act_o_n << " and " << elewise_add_x_n << " -> "
+            << ele_add->Name() << " -> " << elewise_add_out_n;
 
     ReLinkNodes(g, act_out, act, ele_add, elewise_add_act_node);
     found_elewise_add_act_count++;
@@ -165,7 +165,7 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad(
 
   auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph,
                      Graph *g) {
-    VLOG(40) << "handle FuseElewiseAddActGrad1 fuse";
+    VLOG(4) << "handle FuseElewiseAddActGrad1 fuse";
     GET_IR_NODE_FROM_SUBGRAPH(act_out, act_out, elewise_add_act_grad_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(act_grad, act_grad, elewise_add_act_grad_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(d_itermediate_out, d_itermediate_out,
@@ -208,10 +208,10 @@ std::unique_ptr<ir::Graph> FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad(
 
     auto fused_node = g->CreateOpNode(&desc);
 
-    VLOG(40) << "\n\t " << d_act_out_n << " and " << act_out_n << " -> "
-             << act_grad->Name() << " -> " << d_itermediate_out_n << "\n\t "
-             << d_itermediate_out_n << " and " << act_out_n << " -> "
-             << ele_add_grad->Name() << " -> " << d_itermediate_out_n;
+    VLOG(4) << "\n\t " << d_act_out_n << " and " << act_out_n << " -> "
+            << act_grad->Name() << " -> " << d_itermediate_out_n << "\n\t "
+            << d_itermediate_out_n << " and " << act_out_n << " -> "
+            << ele_add_grad->Name() << " -> " << d_itermediate_out_n;
 
     ReLinkNodes(g, d_itermediate_out, act_grad, ele_add_grad, fused_node);
     found_elewise_add_act_count++;

paddle/fluid/framework/ir/graph.cc

@@ -90,7 +90,7 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
 
 std::map<std::string, std::vector<ir::Node *>> Graph::InitFromProgram(
     const ProgramDesc &program) {
-  VLOG(30) << "block in program:" << program_.Size();
+  VLOG(3) << "block in program:" << program_.Size();
   std::unordered_map<std::string, VarDesc *> all_vars;
   // var nodes for each var name, will have multiple versions in SSA
   std::map<std::string, std::vector<ir::Node *>> var_nodes;
@@ -158,7 +158,7 @@ void Graph::ResolveHazard(
     auto it_old = versions.rbegin();
     ++it_old;
     for (; it_old != versions.rend(); it_new = it_old, ++it_old) {
-      VLOG(30) << "deal with var: " << (*it_new)->Name();
+      VLOG(3) << "deal with var: " << (*it_new)->Name();
       ir::Node *write_op =
           (*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0];
       const auto &read_ops = (*it_old)->outputs;

paddle/fluid/framework/ir/graph.h

@@ -89,7 +89,7 @@ class Graph {
                    attr_name);
     attrs_[attr_name] = attr;
     attr_dels_[attr_name] = [attr, attr_name]() {
-      VLOG(30) << "deleting " << attr_name;
+      VLOG(3) << "deleting " << attr_name;
       delete attr;
     };
   }

paddle/fluid/framework/ir/graph_helper.cc

@@ -40,9 +40,8 @@ void SortHelper(
     }
   }
 
-  VLOG(30) << "topology sort insert: " << node->Name()
-           << reinterpret_cast<void *>(node) << " input "
-           << node->inputs.size();
+  VLOG(3) << "topology sort insert: " << node->Name()
+          << reinterpret_cast<void *>(node) << " input " << node->inputs.size();
   ret->push_back(node);
 }
 
@@ -111,9 +110,9 @@ std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
     for (auto &var : n->inputs) {
       for (auto &adj_n : var->inputs) {
         PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
-        VLOG(40) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
-                 << " -> " << n->Name() << reinterpret_cast<void *>(n)
-                 << "  via " << var->Name() << reinterpret_cast<void *>(var);
+        VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
+                << " -> " << n->Name() << reinterpret_cast<void *>(n)
+                << "  via " << var->Name() << reinterpret_cast<void *>(var);
         adj_list[n].insert(adj_n);
       }
     }

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -92,19 +92,19 @@ void GraphPatternDetector::operator()(Graph *graph,
   PrettyLogEndl(Style::detail(), "---  detect %d subgraphs", subgraphs.size());
   int id = 0;
   for (auto &g : subgraphs) {
-    VLOG(30) << "optimizing #" << id++ << " subgraph";
+    VLOG(3) << "optimizing #" << id++ << " subgraph";
     handler(g, graph);
   }
 }
 
 bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
-  VLOG(30) << "mark pdnodes in graph";
+  VLOG(3) << "mark pdnodes in graph";
   if (graph.Nodes().empty()) return false;
 
   for (auto &node : GraphTraits::DFS(graph)) {
     for (const auto &pdnode : pattern_.nodes()) {
       if (pdnode->Tell(&node)) {
-        VLOG(40) << "pdnode " << pdnode->name() << " marked";
+        VLOG(4) << "pdnode " << pdnode->name() << " marked";
         pdnodes2nodes_[pdnode.get()].insert(&node);
       }
     }
@@ -112,7 +112,7 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
   // Check to early stop if some PDNode can't find matched Node.
   for (auto &pdnode : pattern_.nodes()) {
     if (!pdnodes2nodes_.count(pdnode.get())) {
-      VLOG(40) << pdnode->name() << " can't find matched Node, early stop";
+      VLOG(4) << pdnode->name() << " can't find matched Node, early stop";
       // return false;
     }
   }
@@ -121,7 +121,7 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
       GetMarkedNodes(const_cast<Graph *>(&graph)).insert(n);
     }
   }
-  VLOG(30) << pdnodes2nodes_.size() << " nodes marked";
+  VLOG(3) << pdnodes2nodes_.size() << " nodes marked";
 
   return !pdnodes2nodes_.empty();
 }
@@ -215,7 +215,7 @@ GraphPatternDetector::DetectPatterns() {
   // Extend a PDNode to subgraphs by deducing the connection relations defined
   // in edges of PDNodes.
   for (const auto &edge : pattern_.edges()) {
-    VLOG(40) << "check " << edge.first->name() << " -> " << edge.second->name();
+    VLOG(4) << "check " << edge.first->name() << " -> " << edge.second->name();
     // TODO(Superjomn) Fix bug here, the groups might be duplicate here.
     // Each role has two PDNodes, which indicates two roles.
     // Detect two Nodes that can match these two roles and they are connected.
@@ -226,7 +226,7 @@ GraphPatternDetector::DetectPatterns() {
     // source -> target
     for (Node *source : pdnodes2nodes_[edge.first]) {
       for (Node *target : pdnodes2nodes_[edge.second]) {
-        VLOG(80) << "check " << source->id() << " -- " << target->id();
+        VLOG(8) << "check " << source->id() << " -- " << target->id();
         // TODO(Superjomn) add some prune strategies.
         for (const auto &group : pre_groups) {
           if (IsNodesLink(source, target)) {
@@ -243,13 +243,12 @@ GraphPatternDetector::DetectPatterns() {
         }
       }
     }
-    VLOG(30) << "step " << step << " get records: " << cur_groups.size();
+    VLOG(3) << "step " << step << " get records: " << cur_groups.size();
     for (auto &group : cur_groups) {
       for (auto &item : group.roles) {
-        VLOG(40) << "node " << item.second->id() << " as "
-                 << item.first->name();
+        VLOG(4) << "node " << item.second->id() << " as " << item.first->name();
       }
-      VLOG(40) << "=========================================================";
+      VLOG(4) << "=========================================================";
     }
   }
 

paddle/fluid/framework/ir/graph_viz_pass.cc

@@ -41,7 +41,7 @@ std::string FormatName(const Node* node) {
 std::unique_ptr<ir::Graph> GraphVizPass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
   const std::string graph_viz_path = Get<std::string>(kGraphVizPath);
-  VLOG(30) << "draw IR graph viz to " << graph_viz_path;
+  VLOG(3) << "draw IR graph viz to " << graph_viz_path;
   std::unique_ptr<std::ostream> fout(new std::ofstream(graph_viz_path));
   PADDLE_ENFORCE(fout->good());
   std::ostream& sout = *fout;

paddle/fluid/framework/ir/is_test_pass_tester.cc

@@ -15,7 +15,10 @@
 #include "paddle/fluid/framework/ir/is_test_pass.h"
 
 #include <gtest/gtest.h>
-
+#ifdef _WIN32
+#undef FALSE
+#undef TRUE
+#endif
 namespace paddle {
 namespace framework {
 namespace ir {

paddle/fluid/framework/ir/mkldnn_placement_pass.cc

@@ -20,7 +20,7 @@ namespace ir {
 
 std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
-  VLOG(30) << "Aplies MKL-DNN placement strategy.";
+  VLOG(3) << "Aplies MKL-DNN placement strategy.";
   for (const Node* n : graph->Nodes()) {
     if (n->IsOp() && n->Op()->HasAttr("use_mkldnn")) {
       n->Op()->SetAttr("use_mkldnn", true);

paddle/fluid/framework/ir/multi_batch_merge_pass.cc

@@ -62,7 +62,7 @@ VarDesc UpdateGradVarDesc(
         string::Sprintf("%s.repeat.%d", var_desc->Name(), repeat);
     VarDesc repeated_var = CopyVarDesc(var_desc);
     repeated_var.SetName(new_gname);
-    VLOG(30) << "update " << var_desc->Name() << " to repeat " << repeat;
+    VLOG(3) << "update " << var_desc->Name() << " to repeat " << repeat;
     return repeated_var;
   }
   return *var_desc;
@@ -78,7 +78,7 @@ std::unique_ptr<Graph> BatchMergePass::ApplyImpl(
 
   std::vector<ir::Node*> nodes = TopologySortOperations(*graph);
   auto origin_nodes = graph->ReleaseNodes();
-  VLOG(30) << "origin nodes count: " << origin_nodes.size();
+  VLOG(3) << "origin nodes count: " << origin_nodes.size();
   ir::Graph& result = *graph;
 
   // 1. record op nodes of different roles
@@ -137,8 +137,8 @@ std::unique_ptr<Graph> BatchMergePass::ApplyImpl(
             "%s.repeat.%d", repeated_op.Input("Variance")[0], i);
         bn_vars_need_rename.insert(repeated_op.Input("Mean")[0]);
         bn_vars_need_rename.insert(repeated_op.Input("Variance")[0]);
-        VLOG(30) << "renaming " << repeated_op.Input("Mean")[0] << " to "
-                 << new_mean_name;
+        VLOG(3) << "renaming " << repeated_op.Input("Mean")[0] << " to "
+                << new_mean_name;
         repeated_op.RenameInput(repeated_op.Input("Mean")[0], new_mean_name);
         repeated_op.RenameInput(repeated_op.Input("Variance")[0], new_var_name);
         repeated_op.RenameOutput(repeated_op.Output("MeanOut")[0],

paddle/fluid/framework/ir/pass.h

@@ -76,7 +76,7 @@ class Pass {
                    attr_name);
     attrs_[attr_name] = attr;
     attr_dels_[attr_name] = [attr, attr_name]() {
-      VLOG(30) << "deleting " << attr_name;
+      VLOG(3) << "deleting " << attr_name;
       delete attr;
     };
   }

paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc

@@ -196,7 +196,7 @@ std::unique_ptr<ir::Graph> SeqConcatFcFusePass::ApplyImpl(
 
   detector(graph.get(), [&](const GraphPatternDetector::subgraph_t& subgraph,
                             Graph* graph) {
-    VLOG(40) << "get one concat pattern";
+    VLOG(4) << "get one concat pattern";
     // fc
     GET_NODE(fc_w, detector.pattern());
     GET_NODE(fc_bias, detector.pattern());

paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc

@@ -60,7 +60,7 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope) {
 
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
-    VLOG(40) << "handle SeqConv EltAdd Relu fuse";
+    VLOG(4) << "handle SeqConv EltAdd Relu fuse";
     GET_IR_NODE_FROM_SUBGRAPH(seqconv, seqconv, fuse_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(seqconv_weight, seqconv_weight, fuse_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(seqconv_out, seqconv_out, fuse_pattern);

paddle/fluid/framework/lod_rank_table.cc

@@ -31,7 +31,7 @@ void LoDRankTable::Reset(const LoD& lod, size_t level) {
     TableItem item;
     item.index = i;
     item.length = vec[i + 1] - vec[i];
-    VLOG(100) << "Add item to rank table " << item.index << " " << item.length;
+    VLOG(10) << "Add item to rank table " << item.index << " " << item.length;
     items_.emplace_back(item);
   }
   // NOTE(yuyang18):

paddle/fluid/framework/lod_tensor.cc

@@ -26,10 +26,8 @@ limitations under the License. */
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/memory/memory.h"
 
-#if !defined(_WIN32)
 #include "paddle/fluid/recordio/scanner.h"
 #include "paddle/fluid/recordio/writer.h"
-#endif  // _WIN32
 
 namespace paddle {
 namespace framework {
@@ -305,7 +303,6 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
   TensorFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
 }
 
-#if !defined(_WIN32)
 void WriteToRecordIO(recordio::Writer *writer,
                      const std::vector<LoDTensor> &tensor,
                      const platform::DeviceContext &dev_ctx) {
@@ -335,19 +332,7 @@ bool ReadFromRecordIO(recordio::Scanner *scanner,
 
   return true;
 }
-#else
-class Writer {};
-class Scanner {};
-void WriteToRecordIO(recordio::Writer *writer,
-                     const std::vector<LoDTensor> &tensor,
-                     const platform::DeviceContext &dev_ctx) {}
-bool ReadFromRecordIO(recordio::Scanner *scanner,
-                      const platform::DeviceContext &dev_ctx,
-                      std::vector<LoDTensor> *result_ptr) {
-  PADDLE_ENFORCE("windows didn't supported recordio!.");
-  return true;
-}
-#endif  // _WIN32
+
 std::vector<LoDTensor> LoDTensor::SplitLoDTensor(
     const std::vector<platform::Place> places) const {
   check_memory_size();

paddle/fluid/framework/lod_tensor_test.cc

@@ -274,7 +274,6 @@ TEST(LoD, ConvertToOffsetBasedLoD) {
   EXPECT_EQ(offset_lod, expected);
 }
 
-#if !defined(_WIN32)
 template <typename T>
 static void TestRecordIO() {
   LoDTensor tensor;
@@ -321,7 +320,6 @@ TEST(LoDTensor, RecordIO) {
   TestRecordIO<float>();
   TestRecordIO<double>();
 }
-#endif  // !defined(_WIN32)
 
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/mixed_vector_test.cc

@@ -51,7 +51,7 @@ TEST(mixed_vector, InitWithCount) {
 TEST(mixed_vector, ForEach) {
   vec<int> tmp;
   for (auto& v : tmp) {
-    VLOG(30) << v;
+    VLOG(3) << v;
   }
 }
 

paddle/fluid/framework/naive_executor.cc

@@ -21,42 +21,11 @@
 #include "paddle/fluid/framework/naive_executor.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/reader.h"
+#include "paddle/fluid/framework/variable_helper.h"
 #include "paddle/fluid/string/pretty_log.h"
 
 namespace paddle {
 namespace framework {
-
-// These code can be shared with Executor.
-static void InitializeVariable(Variable *var, proto::VarType::Type var_type) {
-  if (var_type == proto::VarType::LOD_TENSOR) {
-    var->GetMutable<LoDTensor>();
-  } else if (var_type == proto::VarType::SELECTED_ROWS) {
-    var->GetMutable<SelectedRows>();
-  } else if (var_type == proto::VarType::FEED_MINIBATCH) {
-    var->GetMutable<FeedFetchList>();
-  } else if (var_type == proto::VarType::FETCH_LIST) {
-    var->GetMutable<FeedFetchList>();
-  } else if (var_type == proto::VarType::STEP_SCOPES) {
-    var->GetMutable<std::vector<framework::Scope *>>();
-  } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
-    var->GetMutable<LoDRankTable>();
-  } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
-    var->GetMutable<LoDTensorArray>();
-  } else if (var_type == proto::VarType::PLACE_LIST) {
-    var->GetMutable<platform::PlaceList>();
-  } else if (var_type == proto::VarType::READER) {
-    var->GetMutable<ReaderHolder>();
-  } else if (var_type == proto::VarType::RAW) {
-    // GetMutable will be called in operator
-  } else {
-    PADDLE_THROW(
-        "Variable type %d is not in "
-        "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
-        "LOD_RANK_TABLE, PLACE_LIST, READER, CHANNEL, RAW]",
-        var_type);
-  }
-}
-
 void NaiveExecutor::Prepare(Scope *scope, const ProgramDesc &program_desc,
                             int block_id, bool with_feed_fetch_ops) {
   if (!scope) {

paddle/fluid/framework/op_desc.cc

@@ -81,13 +81,35 @@ class CompileTimeInferShapeContext : public InferShapeContext {
                    "The %s[%d] is @EMPTY@", out, j);
     auto *in_var = block_.FindVarRecursive(Inputs(in)[i]);
     auto *out_var = block_.FindVarRecursive(Outputs(out)[j]);
-    if (in_var->GetType() != proto::VarType::LOD_TENSOR) {
-      VLOG(30) << "input " << in << " is not LodTensor";
+    if (in_var->GetType() != proto::VarType::LOD_TENSOR &&
+        in_var->GetType() != proto::VarType::LOD_TENSOR_ARRAY) {
+      VLOG(3) << "input " << in << " is not LodTensor or LodTensorArray.";
       return;
     }
     out_var->SetLoDLevel(in_var->GetLoDLevel());
   }
 
+  void DecreaseLoDLevel(const std::string &in, const std::string &out,
+                        size_t i = 0, size_t j = 0) const override {
+    PADDLE_ENFORCE_LT(i, Inputs(in).size());
+    PADDLE_ENFORCE_LT(j, Outputs(out).size());
+    PADDLE_ENFORCE(Inputs(in)[i] != framework::kEmptyVarName,
+                   "The %s[%d] is @EMPTY@", in, i);
+    PADDLE_ENFORCE(Outputs(out)[j] != framework::kEmptyVarName,
+                   "The %s[%d] is @EMPTY@", out, j);
+    auto *in_var = block_.FindVarRecursive(Inputs(in)[i]);
+    auto *out_var = block_.FindVarRecursive(Outputs(out)[j]);
+    PADDLE_ENFORCE(out_var->GetType() == proto::VarType::LOD_TENSOR_ARRAY ||
+                       out_var->GetType() == proto::VarType::LOD_TENSOR,
+                   "The input %s should be LodTensorArray or LodTensor.",
+                   out_var->Name());
+    PADDLE_ENFORCE(in_var->GetType() == proto::VarType::LOD_TENSOR,
+                   "The input %s should be LodTensor.", in_var->Name());
+    if (in_var->GetLoDLevel() > 0) {
+      out_var->SetLoDLevel(in_var->GetLoDLevel() - 1);
+    }
+  }
+
   bool IsRuntime() const override;
 
  protected:
@@ -241,38 +263,38 @@ void OpDesc::SetAttr(const std::string &name, const Attribute &v) {
     const proto::OpProto::Attr &attr = GetProtoAttr(name);
     switch (attr.type()) {
       case proto::AttrType::BOOLEANS: {
-        VLOG(110) << "SetAttr: " << Type() << ", " << name
-                  << " from INTS to BOOLEANS";
+        VLOG(11) << "SetAttr: " << Type() << ", " << name
+                 << " from INTS to BOOLEANS";
         this->attrs_[name] = std::vector<bool>();
         break;
       }
       case proto::AttrType::INTS: {
-        VLOG(110) << "SetAttr: " << Type() << ", " << name
-                  << " from INTS to INTS";
+        VLOG(11) << "SetAttr: " << Type() << ", " << name
+                 << " from INTS to INTS";
         this->attrs_[name] = std::vector<int>();
         break;
       }
       case proto::AttrType::LONGS: {
-        VLOG(110) << "SetAttr: " << Type() << ", " << name
-                  << " from LONGS to LONGS";
+        VLOG(11) << "SetAttr: " << Type() << ", " << name
+                 << " from LONGS to LONGS";
         this->attrs_[name] = std::vector<int64_t>();
         break;
       }
       case proto::AttrType::FLOATS: {
-        VLOG(110) << "SetAttr: " << Type() << ", " << name
-                  << " from INTS to FLOATS";
+        VLOG(11) << "SetAttr: " << Type() << ", " << name
+                 << " from INTS to FLOATS";
         this->attrs_[name] = std::vector<float>();
         break;
       }
       case proto::AttrType::STRINGS: {
-        VLOG(110) << "SetAttr: " << Type() << ", " << name
-                  << " from INTS to STRINGS";
+        VLOG(11) << "SetAttr: " << Type() << ", " << name
+                 << " from INTS to STRINGS";
         this->attrs_[name] = std::vector<std::string>();
         break;
       }
       case proto::AttrType::BLOCKS: {
-        VLOG(110) << "SetAttr: " << Type() << ", " << name
-                  << " from INTS to BLOCKS";
+        VLOG(11) << "SetAttr: " << Type() << ", " << name
+                 << " from INTS to BLOCKS";
         this->SetBlocksAttr(name, std::vector<BlockDesc *>());
         return;
       }
@@ -505,13 +527,13 @@ void OpDesc::CheckAttrs() {
 }
 
 void OpDesc::InferShape(const BlockDesc &block) const {
-  VLOG(30) << "CompileTime infer shape on " << Type();
+  VLOG(3) << "CompileTime infer shape on " << Type();
   InitInferShapeFuncs();
   auto &infer_shape = OpInfoMap::Instance().Get(this->Type()).infer_shape_;
   PADDLE_ENFORCE(static_cast<bool>(infer_shape),
                  "%s's infer_shape has not been registered", this->Type());
   CompileTimeInferShapeContext ctx(*this, block);
-  if (VLOG_IS_ON(100)) {
+  if (VLOG_IS_ON(10)) {
     std::ostringstream sout;
     auto inames = this->InputArgumentNames();
     sout << " From [";
@@ -522,7 +544,7 @@ void OpDesc::InferShape(const BlockDesc &block) const {
     std::copy(onames.begin(), onames.end(),
               std::ostream_iterator<std::string>(sout, ", "));
     sout << "]";
-    VLOG(100) << sout.str();
+    VLOG(10) << sout.str();
   }
   infer_shape(&ctx);
 }
@@ -613,7 +635,7 @@ DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const {
     auto shape = var->GetShape();
     res = shape.empty() ? make_ddim({0UL}) : make_ddim(shape);
   } catch (...) {
-    VLOG(50) << "GetDim of variable " << name << " error";
+    VLOG(5) << "GetDim of variable " << name << " error";
     std::rethrow_exception(std::current_exception());
   }
   return res;
@@ -630,7 +652,7 @@ std::vector<DDim> CompileTimeInferShapeContext::GetRepeatedDims(
       res.push_back(s.empty() ? make_ddim({0UL}) : make_ddim(s));
     }
   } catch (...) {
-    VLOG(50) << "GetRepeatedDim of variable " << name << " error.";
+    VLOG(5) << "GetRepeatedDim of variable " << name << " error.";
     std::rethrow_exception(std::current_exception());
   }
   return res;

paddle/fluid/framework/op_info.h

@@ -31,6 +31,12 @@ class InferShapeBase {
   virtual void operator()(InferShapeContext*) const = 0;
 };
 
+class EstimateFlopsBase {
+ public:
+  virtual ~EstimateFlopsBase() = default;
+  virtual size_t operator()(InferShapeContext*) const = 0;
+};
+
 struct OpInfo {
   OpCreator creator_;
   GradOpMakerFN grad_op_maker_;
@@ -38,6 +44,7 @@ struct OpInfo {
   OpAttrChecker* checker_{nullptr};
   InferVarTypeFN infer_var_type_;
   InferShapeFN infer_shape_;
+  EstimateFlopsFN estimate_flops_;
 
   bool HasOpProtoAndChecker() const {
     return proto_ != nullptr && checker_ != nullptr;

paddle/fluid/framework/op_registry.cc

@@ -46,9 +46,9 @@ static VariableNameMap ConvertOpDescVarsToVarNameMap(
 
 std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
     const proto::OpDesc& op_desc) {
-  VLOG(10) << "CreateOp directly from OpDesc is deprecated. It should only be"
-              "used in unit tests. Use CreateOp(const OpDesc& op_desc) "
-              "instead.";
+  VLOG(1) << "CreateOp directly from OpDesc is deprecated. It should only be"
+             "used in unit tests. Use CreateOp(const OpDesc& op_desc) "
+             "instead.";
   VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
   VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
   AttributeMap attrs;

paddle/fluid/framework/operator.cc

@@ -139,7 +139,7 @@ static LoD GetLoD(const Scope& scope, const std::string& name) {
 }
 
 void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
-  VLOG(40) << place << " " << DebugStringEx(&scope);
+  VLOG(4) << place << " " << DebugStringEx(&scope);
   if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
     PADDLE_THROW("Cannot run operator on place %s", place);
@@ -149,20 +149,17 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
 #endif
   }
 
-// The profile has a process-wide mutex, results in serious performance issue
-// in concurrency scenerio. Here use an `if` to fix this issue.
-// Please not remove the `if`, ask @Superjomn if there are any concern.
-#ifndef _WIN32
+  // The profile has a process-wide mutex, results in serious performance issue
+  // in concurrency scenerio. Here use an `if` to fix this issue.
+  // Please not remove the `if`, ask @Superjomn if there are any concern.
   if (platform::IsProfileEnabled()) {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     platform::RecordEvent record_event(Type(), pool.Get(place));
     RunImpl(scope, place);
-  } else
-#endif
-  {
+  } else {
     RunImpl(scope, place);
   }
-  VLOG(30) << place << " " << DebugStringEx(&scope);
+  VLOG(3) << place << " " << DebugStringEx(&scope);
 }
 
 bool OperatorBase::HasInputs(const std::string& name) const {
@@ -626,6 +623,11 @@ class RuntimeInferShapeContext : public InferShapeContext {
       out_tensor->set_layout(in_tensor.layout());
   }
 
+  void DecreaseLoDLevel(const std::string& in, const std::string& out,
+                        size_t i = 0, size_t j = 0) const override {
+    PADDLE_THROW("DecreaseLoDLevel is only used in compile time.");
+  }
+
   bool IsRuntime() const override { return true; }
 
  protected:
@@ -719,14 +721,14 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
 
   auto expected_kernel_key =
       this->GetExpectedKernelType(ExecutionContext(*this, scope, *dev_ctx));
-  VLOG(30) << "expected_kernel_key:" << expected_kernel_key;
+  VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
 
   auto kernel_iter = kernels.find(expected_kernel_key);
 #ifdef PADDLE_WITH_MKLDNN
   // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
   if (kernel_iter == kernels.end() &&
       expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
-    VLOG(30) << "missing MKLDNN kernel: fallbacking to PLAIN one";
+    VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
     expected_kernel_key.library_type_ = LibraryType::kPlain;
     expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
     kernel_iter = kernels.find(expected_kernel_key);
@@ -778,8 +780,7 @@ void OperatorWithKernel::TransferInplaceVarsBack(
     const Scope& scope, const std::vector<std::string>& inplace_vars,
     const Scope& transfer_scope) const {
   for (auto& var_name : inplace_vars) {
-    VLOG(30) << "share inplace var " + var_name +
-                    " back to it's original scope";
+    VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
     auto* original_tensor =
         GetMutableLoDTensorOrSelectedRowsValueFromVar(scope.FindVar(var_name));
     auto* var = transfer_scope.FindVar(var_name);
@@ -820,8 +821,8 @@ Scope* OperatorWithKernel::TryTransferData(
         transfered_inplace_vars->emplace_back(var_name);
       }
 
-      VLOG(30) << "Transform Variable " << var_name << " from "
-               << kernel_type_for_var << " to " << expected_kernel_key;
+      VLOG(3) << "Transform Variable " << var_name << " from "
+              << kernel_type_for_var << " to " << expected_kernel_key;
 
       // In the inference scenerio, the scopes will be reused across the
       // batches, so the `new_scope` here will result in GPU memroy explosion

paddle/fluid/framework/operator.h

@@ -71,7 +71,7 @@ class OperatorBase;
 class ExecutionContext;
 
 /**
- * OperatorBase has the basic element that Net will call to do computation.
+ * OperatorBase has the basic elements that Net will call to do computation.
  * Only CreateOperator from OpRegistry will new Operator directly. User
  * should always construct a proto message OpDesc and call
  * OpRegistry::CreateOp(op_desc) to get an Operator instance.

paddle/fluid/framework/parallel_executor.cc

@@ -20,7 +20,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/ir/graph.h"
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -54,7 +54,7 @@ class ParallelExecutorPrivate {
   Scope *global_scope_;  // not owned
   std::unique_ptr<details::SSAGraphExecutor> executor_;
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
   bool own_local_scope_;
@@ -104,7 +104,7 @@ ParallelExecutor::ParallelExecutor(
 
   if (member_->use_cuda_) {
 // Bcast Parameters to all GPUs
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
     ncclUniqueId *nccl_id = nullptr;
     if (nccl_id_var != nullptr) {
@@ -124,7 +124,7 @@ ParallelExecutor::ParallelExecutor(
 
 // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
 // ncclOp
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
       main_program, member_->places_, loss_var_name, params,
       member_->local_scopes_, member_->use_cuda_, member_->nccl_ctxs_.get());
@@ -208,12 +208,12 @@ void ParallelExecutor::BCastParamsToDevices(
 
     auto &main_tensor = main_var->Get<LoDTensor>();
     if (!main_tensor.IsInitialized()) {
-      VLOG(30) << "one in var not inited, return!";
+      VLOG(3) << "one in var not inited, return!";
       continue;
     }
     auto &dims = main_tensor.dims();
     if (paddle::platform::is_gpu_place(main_tensor.place())) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       std::vector<void *> buffers;
       size_t numel = main_tensor.numel();
       ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type());

paddle/fluid/framework/scope.cc

@@ -162,7 +162,7 @@ Variable* Scope::VarInternal(const std::string& name) {
 
   v = new Variable();
   vars_[name].reset(v);
-  VLOG(30) << "Create variable " << name;
+  VLOG(3) << "Create variable " << name;
   v->name_ = &(vars_.find(name)->first);
   return v;
 }

paddle/fluid/framework/selected_rows.cc

@@ -206,7 +206,7 @@ void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value,
   PADDLE_ENFORCE(value->IsInitialized(),
                  "The value tensor should be initialized.");
   if (ids.numel() == 0) {
-    VLOG(30) << "keys is empty, please check data!";
+    VLOG(3) << "keys is empty, please check data!";
   } else {
     int64_t value_width = value_->numel() / value_->dims()[0];
     PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0],

paddle/fluid/framework/selected_rows.h

@@ -120,8 +120,22 @@ class SelectedRows {
    */
   int64_t AutoGrownIndex(int64_t key, bool auto_grown, bool is_test = false);
 
-  void SyncIndex();
+  /*
+   * @brief Get the index of the key from id_to_index_ map.
+   */
+  inline int64_t GetIndexFromId(int64_t key) {
+    auto iter = id_to_index_.find(key);
+    if (iter == id_to_index_.end()) {
+      return -1;
+    } else {
+      return iter->second;
+    }
+  }
 
+  void SyncIndex();
+  /*
+   * @brief Get complete Dims before
+   */
   DDim GetCompleteDims() const {
     std::vector<int64_t> dims = vectorize(value_->dims());
     dims[0] = height_;
@@ -133,9 +147,10 @@ class SelectedRows {
   // SelectedRows are simply concated when adding together. Until a
   // SelectedRows add a Tensor, will the duplicate rows be handled.
   Vector<int64_t> rows_;
-  std::unordered_map<int64_t, int64_t> id_to_index_;
+  std::unordered_map<int64_t, int64_t>
+      id_to_index_;  // should not be used when rows_ has duplicate member
   std::unique_ptr<Tensor> value_{nullptr};
-  int64_t height_;
+  int64_t height_;  // height indicates the underline tensor's height
   std::unique_ptr<RWLock> rwlock_{nullptr};
 };
 

paddle/fluid/framework/shape_inference.h

@@ -62,6 +62,9 @@ class InferShapeContext {
   virtual void ShareLoD(const std::string &in, const std::string &out,
                         size_t i = 0, size_t j = 0) const = 0;
 
+  virtual void DecreaseLoDLevel(const std::string &in, const std::string &out,
+                                size_t i = 0, size_t j = 0) const = 0;
+
   virtual bool IsRuntime() const = 0;
 
   std::vector<InferShapeVarPtr> GetInputVarPtrs(const std::string &name);

paddle/fluid/framework/tensor_util.cc

@@ -22,8 +22,8 @@ namespace framework {
 
 void TensorCopy(const Tensor& src, const platform::Place& dst_place,
                 const platform::DeviceContext& ctx, Tensor* dst) {
-  VLOG(30) << "TensorCopy " << src.dims() << " from " << src.place() << " to "
-           << dst_place;
+  VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to "
+          << dst_place;
   src.check_memory_size();
 
   dst->Resize(src.dims());
@@ -37,8 +37,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
     if (src_ptr == dst_ptr) {
-      VLOG(30) << "Skip copy the same data async from " << src_place << " to "
-               << dst_place;
+      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+              << dst_place;
       return;
     }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
@@ -77,8 +77,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
         reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
     if (platform::is_same_place(src_place, dst_place)) {
       if (src_ptr == dst_ptr) {
-        VLOG(30) << "Skip copy the same data async from " << src_place << " to "
-                 << dst_place;
+        VLOG(3) << "Skip copy the same data async from " << src_place << " to "
+                << dst_place;
         return;
       }
       memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
@@ -114,8 +114,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
 
 void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
                     Tensor* dst) {
-  VLOG(30) << "TensorCopySync " << src.dims() << " from " << src.place()
-           << " to " << dst_place;
+  VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place()
+          << " to " << dst_place;
   src.check_memory_size();
   dst->Resize(src.dims());
   dst->set_layout(src.layout());
@@ -125,8 +125,8 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
   auto size = src.numel() * SizeOfType(src.type());
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
     if (src_ptr == dst_ptr) {
-      VLOG(30) << "Skip copy the same data from " << src_place << " to "
-               << dst_place;
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
       return;
     }
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
@@ -146,8 +146,8 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
   } else if (platform::is_gpu_place(src_place) &&
              platform::is_gpu_place(dst_place)) {
     if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) {
-      VLOG(30) << "Skip copy the same data from " << src_place << " to "
-               << dst_place;
+      VLOG(3) << "Skip copy the same data from " << src_place << " to "
+              << dst_place;
       return;
     }
     auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);

paddle/fluid/framework/threadpool.cc

@@ -39,7 +39,7 @@ void ThreadPool::Init() {
     int num_threads = std::thread::hardware_concurrency();
     if (FLAGS_dist_threadpool_size > 0) {
       num_threads = FLAGS_dist_threadpool_size;
-      VLOG(10) << "set dist_threadpool_size to " << num_threads;
+      VLOG(1) << "set dist_threadpool_size to " << num_threads;
     }
     PADDLE_ENFORCE_GT(num_threads, 0);
     threadpool_.reset(new ThreadPool(num_threads));

paddle/fluid/framework/type_defs.h

@@ -54,5 +54,7 @@ using InferVarTypeFN =
 
 using InferShapeFN = std::function<void(InferShapeContext*)>;
 
+using EstimateFlopsFN = std::function<void(InferShapeContext*)>;
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/var_desc.cc

@@ -61,10 +61,10 @@ size_t VarDesc::GetTensorDescNum() const {
 void VarDesc::SetShapes(
     const std::vector<std::vector<int64_t>> &multiple_dims) {
   if (multiple_dims.size() != GetTensorDescNum()) {
-    VLOG(30) << "WARNING: The number of given shapes(" << multiple_dims.size()
-             << ") doesn't match the existing tensor number("
-             << GetTensorDescNum()
-             << "). The Reader is going to be reinitialized.";
+    VLOG(3) << "WARNING: The number of given shapes(" << multiple_dims.size()
+            << ") doesn't match the existing tensor number("
+            << GetTensorDescNum()
+            << "). The Reader is going to be reinitialized.";
     SetTensorDescNum(multiple_dims.size());
   }
   std::vector<proto::VarType::TensorDesc *> tensors = mutable_tensor_descs();
@@ -94,11 +94,11 @@ void VarDesc::SetDataType(proto::VarType::Type data_type) {
 void VarDesc::SetDataTypes(
     const std::vector<proto::VarType::Type> &multiple_data_type) {
   if (multiple_data_type.size() != GetTensorDescNum()) {
-    VLOG(30) << "WARNING: The number of given data types("
-             << multiple_data_type.size()
-             << ") doesn't match the existing tensor number("
-             << GetTensorDescNum()
-             << "). The Reader is going to be reinitialized.";
+    VLOG(3) << "WARNING: The number of given data types("
+            << multiple_data_type.size()
+            << ") doesn't match the existing tensor number("
+            << GetTensorDescNum()
+            << "). The Reader is going to be reinitialized.";
     SetTensorDescNum(multiple_data_type.size());
   }
   std::vector<proto::VarType::TensorDesc *> tensor_descs =
@@ -139,11 +139,11 @@ void VarDesc::SetLoDLevel(int32_t lod_level) {
 
 void VarDesc::SetLoDLevels(const std::vector<int32_t> &multiple_lod_level) {
   if (multiple_lod_level.size() != GetTensorDescNum()) {
-    VLOG(30) << "WARNING: The number of given lod_levels("
-             << multiple_lod_level.size()
-             << ") doesn't match the existing tensor number("
-             << GetTensorDescNum()
-             << "). The Reader is going to be reinitialized.";
+    VLOG(3) << "WARNING: The number of given lod_levels("
+            << multiple_lod_level.size()
+            << ") doesn't match the existing tensor number("
+            << GetTensorDescNum()
+            << "). The Reader is going to be reinitialized.";
     SetTensorDescNum(multiple_lod_level.size());
   }
   switch (desc_.type().type()) {

paddle/fluid/framework/variable_helper.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/variable_helper.h"
+
+#include <vector>
+
+#include "paddle/fluid/framework/feed_fetch_type.h"
+#include "paddle/fluid/framework/lod_rank_table.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/lod_tensor_array.h"
+#include "paddle/fluid/framework/reader.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/selected_rows.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace framework {
+void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
+  if (var_type == proto::VarType::LOD_TENSOR) {
+    var->GetMutable<LoDTensor>();
+  } else if (var_type == proto::VarType::SELECTED_ROWS) {
+    var->GetMutable<SelectedRows>();
+  } else if (var_type == proto::VarType::FEED_MINIBATCH) {
+    var->GetMutable<FeedFetchList>();
+  } else if (var_type == proto::VarType::FETCH_LIST) {
+    var->GetMutable<FeedFetchList>();
+  } else if (var_type == proto::VarType::STEP_SCOPES) {
+    var->GetMutable<std::vector<framework::Scope*>>();
+  } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
+    var->GetMutable<LoDRankTable>();
+  } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
+    var->GetMutable<LoDTensorArray>();
+  } else if (var_type == proto::VarType::PLACE_LIST) {
+    var->GetMutable<platform::PlaceList>();
+  } else if (var_type == proto::VarType::READER) {
+    var->GetMutable<ReaderHolder>();
+  } else if (var_type == proto::VarType::RAW) {
+    // GetMutable will be called in operator
+  } else {
+    PADDLE_THROW(
+        "Variable type %d is not in "
+        "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
+        "LOD_RANK_TABLE, PLACE_LIST, READER, RAW]",
+        var_type);
+  }
+}
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/variable_helper.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+#pragma once
+
+#include "paddle/fluid/framework/framework.pb.h"
+#include "paddle/fluid/framework/variable.h"
+namespace paddle {
+namespace framework {
+void InitializeVariable(Variable *var, proto::VarType::Type var_type);
+}
+}

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -35,4 +35,5 @@ function(inference_analysis_test TARGET)
   endif()
 endfunction(inference_analysis_test)
 
-inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS reset_tensor_array paddle_inference_api)
+inference_analysis_test(test_analyzer SRCS analyzer_tester.cc
+    EXTRA_DEPS reset_tensor_array paddle_inference_api)

paddle/fluid/inference/analysis/analysis_pass.h

@@ -46,8 +46,6 @@ class AnalysisPass {
  protected:
   // User should implement these.
   virtual void RunImpl(Argument* argument) = 0;
-
-  Argument* argument_{nullptr};
 };
 
 }  // namespace analysis

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -19,6 +19,7 @@
 #include "paddle/fluid/inference/analysis/ut_helper.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/api/paddle_inference_pass.h"
+#include "paddle/fluid/platform/port.h"
 
 namespace paddle {
 namespace inference {
@@ -75,7 +76,8 @@ void TestWord2vecPrediction(const std::string& model_path) {
                      0.000932706};
   const size_t num_elements = outputs.front().data.length() / sizeof(float);
   // The outputs' buffers are in CPU memory.
-  for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
+  for (size_t i = 0; i < std::min(static_cast<size_t>(5UL), num_elements);
+       i++) {
     LOG(INFO) << "data: "
               << static_cast<float*>(outputs.front().data.data())[i];
     PADDLE_ENFORCE(static_cast<float*>(outputs.front().data.data())[i],

paddle/fluid/inference/analysis/passes/CMakeLists.txt

 cc_library(ir_graph_build_pass SRCS ir_graph_build_pass.cc DEPS analysis_pass argument ir_pass_manager)
 cc_library(ir_analysis_pass SRCS ir_analysis_pass.cc DEPS analysis_pass argument ir_pass_manager)
-cc_library(analysis_passes SRCS passes.cc DEPS ir_graph_build_pass ir_analysis_pass)
+cc_library(ir_params_sync_among_devices_pass SRCS ir_params_sync_among_devices_pass.cc DEPS analysis_pass argument ir_pass_manager)
+cc_library(analysis_passes SRCS passes.cc DEPS ir_graph_build_pass ir_analysis_pass ir_params_sync_among_devices_pass)
 
 set(analysis_deps ${analysis_deps}
         ir_graph_build_pass

paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc

@@ -61,6 +61,7 @@ void IrAnalysisComposePass::InitTensorRTAttrs(Argument *argument) {
 void IrAnalysisComposePass::ApplyIrPasses(Argument *argument) {
   std::vector<std::string> passes({
       "ir_graph_build_pass", "ir_analysis_pass",
+      "ir_params_sync_among_devices_pass",
   });
   for (const auto &pass : passes) {
     VLOG(2) << "Run pass " << pass;

paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc

@@ -36,12 +36,7 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
   // so that the parameters will on the same device, or they will keep copying
   // between difference devices.
   platform::Place place;
-  if (argument->use_gpu()) {
-    PADDLE_ENFORCE(argument->gpu_device_id_valid());
-    place = platform::CUDAPlace(argument->gpu_device_id());
-  } else {
-    place = platform::CPUPlace();
-  }
+  place = platform::CPUPlace();
 
   if (argument->model_dir_valid()) {
     auto program =

paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h"
+#include "paddle/fluid/framework/data_layout.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+void IrParamsSyncAmongDevicesPass::RunImpl(Argument *argument) {
+  PADDLE_ENFORCE(argument->scope_valid());
+  PADDLE_ENFORCE(argument->use_gpu_valid());
+
+  platform::Place place;
+
+  // The parameters are on the cpu, therefore, synchronization is not necessary.
+  if (!argument->use_gpu()) return;
+
+  LOG(INFO) << "Sync params from CPU to GPU";
+
+  PADDLE_ENFORCE(argument->gpu_device_id_valid());
+  place = platform::CUDAPlace(argument->gpu_device_id());
+
+  auto *scope = argument->scope_ptr();
+  std::vector<std::string> all_vars = scope->LocalVarNames();
+
+  // We get all the vars from local_scope instead of the ProgramDesc.
+  // Because there exists the case that new parameter variables are not added to
+  // the program in the analysis pass.
+  for (auto &var_name : all_vars) {
+    auto *var = scope->FindLocalVar(var_name);
+    PADDLE_ENFORCE(var != nullptr);
+    if (var->IsType<framework::LoDTensor>() ||
+        var->IsType<framework::Tensor>()) {
+      auto *t = var->GetMutable<framework::LoDTensor>();
+
+      platform::CPUPlace cpu_place;
+      framework::LoDTensor temp_tensor;
+      temp_tensor.Resize(t->dims());
+      temp_tensor.mutable_data<float>(cpu_place);
+
+      // Copy the parameter data to a tmp tensor.
+      TensorCopySync(*t, cpu_place, &temp_tensor);
+      // Reallocation the space on GPU
+      t->mutable_data<float>(place);
+
+      // Copy parameter data to newly allocated GPU space.
+      TensorCopySync(temp_tensor, place, t);
+    }
+  }
+}
+
+std::string IrParamsSyncAmongDevicesPass::repr() const {
+  return "ir-params-sync-among-devices-pass";
+}
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/inference/analysis/analysis_pass.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+/*
+ * Sync parameter from CPU to GPU.
+ */
+class IrParamsSyncAmongDevicesPass : public AnalysisPass {
+ public:
+  void RunImpl(Argument *argument) override;
+  std::string repr() const override;
+};
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/analysis/passes/passes.cc

@@ -16,6 +16,7 @@
 #include "paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc"
 #include "paddle/fluid/inference/analysis/passes/ir_analysis_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h"
+#include "paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h"
 
 namespace paddle {
 namespace inference {
@@ -27,6 +28,9 @@ PassRegistry::PassRegistry() {
                   std::unique_ptr<AnalysisPass>(new IrGraphBuildPass));
   passes_.emplace("ir_analysis_compose_pass",
                   std::unique_ptr<AnalysisPass>(new IrAnalysisComposePass));
+  passes_.emplace(
+      "ir_params_sync_among_devices_pass",
+      std::unique_ptr<AnalysisPass>(new IrParamsSyncAmongDevicesPass));
 }
 
 }  // namespace analysis