Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix-utest-vgg

CMakeLists.txt

@@ -47,35 +47,36 @@ find_package(Threads REQUIRED)
 
 include(simd)
 
-################################ Configurations #######################################
+################################ Exposed Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
-option(WITH_AMD_GPU     "Compile PaddlePaddle with AMD GPU"             OFF)
+option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
+option(WITH_PYTHON      "Compile PaddlePaddle with python interpreter"  ON)
+option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        OFF)
 option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FOUND})
+option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
+option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
+option(WITH_BRPC_RDMA     "Use brpc rdma as the rpc protocal"           OFF)
+option(ON_INFER         "Turn on inference optimization."               OFF)
+option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)
+################################ Internal Configurations #######################################
+option(WITH_AMD_GPU     "Compile PaddlePaddle with AMD GPU"             OFF)
 option(WITH_NGRAPH      "Compile PaddlePaddle with nGraph support."     OFF)
-option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
-option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        OFF)
-option(WITH_PYTHON      "Compile PaddlePaddle with python interpreter"  ON)
 option(WITH_PROFILER    "Compile PaddlePaddle with GPU profiler and gperftools"        OFF)
 option(WITH_JEMALLOC    "Compile PaddlePaddle with jemalloc"            OFF)
 option(WITH_COVERAGE    "Compile PaddlePaddle with code coverage"       OFF)
 option(COVERALLS_UPLOAD "Package code coverage data to coveralls"       OFF)
-option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
 option(WITH_PSLIB       "Compile with pslib support"                    OFF)
 option(WITH_CONTRIB     "Compile the third-party contributation"        OFF)
 option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
 # TODO(Superjomn) Remove WITH_ANAKIN option if not needed latter.
-option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)
 option(ANAKIN_BUILD_FAT_BIN "Build anakin cuda fat-bin lib for all device plantform, ignored when WITH_ANAKIN=OFF" OFF)
 option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plantform. ignored when WITH_ANAKIN=OFF" ON)
 option(WITH_GRPC     "Use grpc as the default rpc framework"            ${WITH_DISTRIBUTE})
-option(WITH_BRPC_RDMA     "Use brpc rdma as the rpc protocal"           OFF)
-option(ON_INFER         "Turn on inference optimization."               OFF)
-option(WITH_INFERENCE_API_TEST   "Test fluid inference high-level api interface"  OFF)
-option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
+option(WITH_INFERENCE_API_TEST   "Test fluid inference C++ high-level api interface"  OFF)
+option(WITH_HIGH_LEVEL_API_TEST   "Test fluid python high-level api interface"  OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
 option(WITH_FAST_MATH   "Make use of fast math library, might affect the precision to some extent" ON)
-option(WITH_WBAES       "Compile PaddlePaddle with WBAES support"       ON)
 
 # PY_VERSION
 if(NOT PY_VERSION)
@@ -149,7 +150,6 @@ include(external/dlpack)
 include(external/snappy)    # download snappy
 include(external/snappystream) # download snappystream
 include(external/warpctc)   # download, build, install warpctc
-include(external/wbaes)     # download wbaes
 
 if (NOT WIN32)
 # there is no official support of nccl, cupti in windows

cmake/configure.cmake

@@ -157,7 +157,3 @@ endif(WITH_BRPC_RDMA)
 if(ON_INFER)
     add_definitions(-DPADDLE_ON_INFERENCE)
 endif(ON_INFER)
-
-if(WITH_WBAES)
-    add_definitions(-DPADDLE_WITH_WBAES)
-endif(WITH_WBAES)

cmake/external/protobuf.cmake

@@ -221,6 +221,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
             -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
             -DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}
             -DCMAKE_INSTALL_LIBDIR=lib
+            -DBUILD_SHARED_LIBS=OFF
         CMAKE_CACHE_ARGS
             -DCMAKE_INSTALL_PREFIX:PATH=${PROTOBUF_INSTALL_DIR}
             -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}

cmake/external/wbaes.cmake

-# Copyright (c) 2019 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(NOT ${WITH_WBAES})
-    return()
-ENDIF(NOT ${WITH_WBAES})
-
-INCLUDE(ExternalProject)
-SET(WBAES_DST_DIR       "wbaes")
-SET(WBAES_INSTALL_ROOT  "${THIRD_PARTY_PATH}/install")
-SET(WBAES_INSTALL_DIR   ${WBAES_INSTALL_ROOT}/${WBAES_DST_DIR})
-SET(WBAES_ROOT          ${WBAES_INSTALL_DIR})
-SET(WBAES_INC_DIR       ${WBAES_ROOT}/include)
-SET(WBAES_LIB_DIR       ${WBAES_ROOT}/lib)
-
-SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${WBAES_ROOT}/lib")
-SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
-
-IF(APPLE)
-    SET(WBAES_TAG   "v1.0.0" CACHE STRING "" FORCE)
-    SET(WBAES_URL   "http://paddlepaddledeps.bj.bcebos.com/wbaes-sdk.mac.${WBAES_TAG}.tgz" CACHE STRING "" FORCE)
-    SET(WBAES_LIB   ${WBAES_LIB_DIR}/libwbaes.dylib)
-    SET(WBAES_SHARED_LIB   ${WBAES_LIB_DIR}/libwbaes.dylib)
-ELSEIF(WIN32)
-    SET(WBAES_TAG   "v1.0.0" CACHE STRING "" FORCE)
-    SET(WBAES_URL   "http://paddlepaddledeps.bj.bcebos.com/wbaes-sdk.windows-x64.${WBAES_TAG}.tgz" CACHE STRING "" FORCE)
-    SET(WBAES_LIB   ${WBAES_LIB_DIR}/libwbaes.lib)
-    SET(WBAES_SHARED_LIB   ${WBAES_LIB_DIR}/libwbaes.dll)
-ELSE()
-    SET(WBAES_TAG   "v1.0.2" CACHE STRING "" FORCE)
-    SET(WBAES_URL   "http://paddlepaddledeps.bj.bcebos.com/wbaes-sdk.linux-x86_64.${WBAES_TAG}.tgz" CACHE STRING "" FORCE)
-    SET(WBAES_LIB   ${WBAES_LIB_DIR}/libwbaes.so)
-    SET(WBAES_SHARED_LIB   ${WBAES_LIB_DIR}/libwbaes.so)
-ENDIF()
-
-SET(WBAES_PROJECT       "extern_wbaes")
-MESSAGE(STATUS "WBAES_URL: ${WBAES_URL}, WBAES_LIB: ${WBAES_LIB}")
-SET(WBAES_SOURCE_DIR    "${THIRD_PARTY_PATH}/wbaes") 
-SET(WBAES_DOWNLOAD_DIR  "${WBAES_SOURCE_DIR}/src/${WBAES_PROJECT}")
-
-ExternalProject_Add(
-    ${WBAES_PROJECT}
-    ${EXTERNAL_PROJECT_LOG_ARGS}
-    PREFIX                  ${WBAES_SOURCE_DIR}
-    URL                     ${WBAES_URL}
-    DOWNLOAD_DIR            ${WBAES_DOWNLOAD_DIR}
-    DOWNLOAD_NO_PROGRESS    1
-    CONFIGURE_COMMAND       ""
-    BUILD_COMMAND           ""
-    INSTALL_COMMAND         ""
-        ${CMAKE_COMMAND} -E copy_directory ${WBAES_DOWNLOAD_DIR}/include ${WBAES_INC_DIR} &&
-        ${CMAKE_COMMAND} -E copy_directory ${WBAES_DOWNLOAD_DIR}/lib ${WBAES_LIB_DIR}
-)
-
-INCLUDE_DIRECTORIES(${WBAES_INC_DIR})
-
-ADD_LIBRARY(wbaes SHARED IMPORTED GLOBAL)
-SET_PROPERTY(TARGET wbaes PROPERTY IMPORTED_LOCATION ${WBAES_LIB})
-SET_PROPERTY(TARGET wbaes PROPERTY IMPORTED_NO_SONAME 1)
-ADD_DEPENDENCIES(wbaes ${WBAES_PROJECT})

cmake/generic.cmake

@@ -264,14 +264,6 @@ function(cc_library TARGET_NAME)
         list(REMOVE_ITEM cc_library_DEPS warpctc)
         add_dependencies(${TARGET_NAME} warpctc)
       endif()
-      # Only deps libwbaes.so, not link
-      if("${cc_library_DEPS};" MATCHES "wbaes;")
-        list(REMOVE_ITEM cc_library_DEPS wbaes)
-        if(NOT "${TARGET_NAME}" MATCHES "dynload_wbaes")
-          list(APPEND cc_library_DEPS dynload_wbaes)
-        endif()
-        add_dependencies(${TARGET_NAME} wbaes)
-      endif()
       # Only deps libmklml.so, not link
       if("${cc_library_DEPS};" MATCHES "mklml;")
         list(REMOVE_ITEM cc_library_DEPS mklml)

cmake/inference_lib.cmake

@@ -170,14 +170,6 @@ copy(snappystream_lib
         DSTS ${dst_dir} ${dst_dir}/lib
         DEPS snappystream)
 
-if (WITH_WBAES)
-    set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/wbaes")
-    copy(wbaes_lib
-            SRCS ${WBAES_INC_DIR} ${WBAES_LIB}
-            DSTS ${dst_dir} ${dst_dir}/lib
-            DEPS wbaes)
-endif ()
-
 set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
 copy(zlib_lib
         SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}

paddle/fluid/API.spec

@@ -13,6 +13,7 @@ paddle.fluid.name_scope (ArgSpec(args=['prefix'], varargs=None, keywords=None, d
 paddle.fluid.cuda_places (ArgSpec(args=['device_ids'], varargs=None, keywords=None, defaults=(None,)), ('document', '7d9a51fc9cf3c5245b5227080a8064c3'))
 paddle.fluid.cpu_places (ArgSpec(args=['device_count'], varargs=None, keywords=None, defaults=(None,)), ('document', '4c0cd83f0b401fc2ff84c70974e5d210'))
 paddle.fluid.cuda_pinned_places (ArgSpec(args=['device_count'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd0c3ebd813c39958c92b78e3eef7e912'))
+paddle.fluid.in_dygraph_mode (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', 'f06314a1cb30c96b5808dde2219c2dae'))
 paddle.fluid.Executor.__init__ (ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.Executor.close (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'f5369953dd0c443961cf79f7a00e1a03'))
 paddle.fluid.Executor.infer_from_dataset (ArgSpec(args=['self', 'program', 'dataset', 'scope', 'thread', 'debug', 'fetch_list', 'fetch_info', 'print_period'], varargs=None, keywords=None, defaults=(None, None, None, 0, False, None, None, 100)), ('document', '9c7decb955b9c4f718114179c8985581'))
@@ -117,6 +118,8 @@ paddle.fluid.layers.reduce_mean (ArgSpec(args=['input', 'dim', 'keep_dim', 'name
 paddle.fluid.layers.reduce_max (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', '66a622db727551761ce4eb73eaa7f6a4'))
 paddle.fluid.layers.reduce_min (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', 'd50ac552b5d131468ed466d08bb2d38c'))
 paddle.fluid.layers.reduce_prod (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', 'fcd8301a0ce15f219c7a4bcd0c1e8eca'))
+paddle.fluid.layers.reduce_all (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', '646ca4d4a2cc16084f59de44b6927eca'))
+paddle.fluid.layers.reduce_any (ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)), ('document', 'f36661060aeeaf6c6b1331e41b3726fa'))
 paddle.fluid.layers.sequence_first_step (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', '2b290d3d77882bfe9bb8d331cac8cdd3'))
 paddle.fluid.layers.sequence_last_step (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', 'c16a892f44f7fe71bfa5afc32d3f34ce'))
 paddle.fluid.layers.sequence_slice (ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'fdcea0e8b5bc7d8d4b1b072c521014e6'))
@@ -124,7 +127,7 @@ paddle.fluid.layers.dropout (ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed
 paddle.fluid.layers.split (ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None)), ('document', '652625345c2acb900029c78cc75f8aa6'))
 paddle.fluid.layers.ctc_greedy_decoder (ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'ebbf2adbd79683dc93db03454dfa18c2'))
 paddle.fluid.layers.edit_distance (ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None)), ('document', '97f0262f97602644c83142789d784571'))
-paddle.fluid.layers.l2_normalize (ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None)), ('document', '6e428384ce6a77207fa2c70d9f011990'))
+paddle.fluid.layers.l2_normalize (ArgSpec(args=['x', 'axis', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(1e-12, None)), ('document', '35c6a241bcc1a1fc89508860d82ad62b'))
 paddle.fluid.layers.matmul (ArgSpec(args=['x', 'y', 'transpose_x', 'transpose_y', 'alpha', 'name'], varargs=None, keywords=None, defaults=(False, False, 1.0, None)), ('document', 'b4cbe1ac451005df6dad12e9ffdccca9'))
 paddle.fluid.layers.topk (ArgSpec(args=['input', 'k', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd3570c02f71bcd78e60b3f31dc8f5b32'))
 paddle.fluid.layers.warpctc (ArgSpec(args=['input', 'label', 'blank', 'norm_by_times', 'use_cudnn'], varargs=None, keywords=None, defaults=(0, False, False)), ('document', 'aaba49c038ba927f0a8e45c0c9a686ab'))
@@ -155,10 +158,10 @@ paddle.fluid.layers.label_smooth (ArgSpec(args=['label', 'prior_dist', 'epsilon'
 paddle.fluid.layers.roi_pool (ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0)), ('document', 'c317aa595deb31649083c8faa91cdb97'))
 paddle.fluid.layers.roi_align (ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None)), ('document', '12c5bbb8b38c42e623fbc47611d766e1'))
 paddle.fluid.layers.dice_loss (ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,)), ('document', '1ba0508d573f65feecf3564dce22aa1d'))
-paddle.fluid.layers.image_resize (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1)), ('document', '7a1966d7c3a48f1fc0881cdaf5d83b0b'))
+paddle.fluid.layers.image_resize (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1)), ('document', 'd1b08c11bb9277386fcf6ae70b6622d1'))
 paddle.fluid.layers.image_resize_short (ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)), ('document', '06211aefc50c5a3e940d7204d859cdf7'))
-paddle.fluid.layers.resize_bilinear (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1)), ('document', 'e4fb4ed511b2293b8f04f7e872afbfd7'))
-paddle.fluid.layers.resize_nearest (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True)), ('document', '735fa9758a6d7ff3b47d7b827f961c1d'))
+paddle.fluid.layers.resize_bilinear (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1)), ('document', 'c45591fbc4f64a178fbca219e1546a58'))
+paddle.fluid.layers.resize_nearest (ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True)), ('document', 'ae6d73cdc7f3a138d8a338ecdb33c1ae'))
 paddle.fluid.layers.gather (ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None), ('document', '98f1c86716b9b7f4dda83f20e2adeee2'))
 paddle.fluid.layers.scatter (ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '65f8e9d8ddfd0b412f940579c4faa342'))
 paddle.fluid.layers.sequence_scatter (ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '15b522457dfef103f0c20ca9d397678b'))
@@ -203,6 +206,7 @@ paddle.fluid.layers.gaussian_random_batch_size_like (ArgSpec(args=['input', 'sha
 paddle.fluid.layers.sum (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', 'a418e3ccb5e2ac21bd60f5cc221d5860'))
 paddle.fluid.layers.slice (ArgSpec(args=['input', 'axes', 'starts', 'ends'], varargs=None, keywords=None, defaults=None), ('document', '01dbb91e7c74cb11336cd531013de51a'))
 paddle.fluid.layers.shape (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', '17db0f814eb7bb5a3fac1ca6e60e16d8'))
+paddle.fluid.layers.rank (ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None), ('document', 'ee1386c42ecc8f424fe3fb21862fefc2'))
 paddle.fluid.layers.logical_and (ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', 'cdcf20c494c92060d10feb9374532f42'))
 paddle.fluid.layers.logical_or (ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', '0eae3f726a4afe590757552fa3ced012'))
 paddle.fluid.layers.logical_xor (ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', 'b0daaa3fa4a0aa62f9b58c43d959eb25'))
@@ -235,6 +239,7 @@ paddle.fluid.layers.huber_loss (ArgSpec(args=['input', 'label', 'delta'], vararg
 paddle.fluid.layers.kldiv_loss (ArgSpec(args=['x', 'target', 'reduction', 'name'], varargs=None, keywords=None, defaults=('mean', None)), ('document', '776d536cac47c89073abc7ee524d5aec'))
 paddle.fluid.layers.tree_conv (ArgSpec(args=['nodes_vector', 'edge_set', 'output_size', 'num_filters', 'max_depth', 'act', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(1, 2, 'tanh', None, None, None)), ('document', '34ea12ac9f10a65dccbc50100d12e607'))
 paddle.fluid.layers.npair_loss (ArgSpec(args=['anchor', 'positive', 'labels', 'l2_reg'], varargs=None, keywords=None, defaults=(0.002,)), ('document', '46994d10276dd4cb803b4062b5d14329'))
+paddle.fluid.layers.pixel_shuffle (ArgSpec(args=['x', 'upscale_factor'], varargs=None, keywords=None, defaults=None), ('document', '731b21c62a4add60a33bd76d802ffc5c'))
 paddle.fluid.layers.fsp_matrix (ArgSpec(args=['x', 'y'], varargs=None, keywords=None, defaults=None), ('document', 'b76ccca3735bea4a58a0dbf0d77c5393'))
 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)), ('document', '33bbd42027d872b3818b3d64ec52e139'))
 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)), ('document', 'b1ae2e1cc0750e58726374061ea90ecc'))
@@ -270,6 +275,7 @@ paddle.fluid.layers.has_inf (ArgSpec(args=['x'], varargs=None, keywords=None, de
 paddle.fluid.layers.has_nan (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '2e53e83127dbfd86e7098bdfe9a549e8'))
 paddle.fluid.layers.isfinite (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '0a437011c3906079fd8947ed3e52d292'))
 paddle.fluid.layers.range (ArgSpec(args=['start', 'end', 'step', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '2ec937ede953ded2fdff2675883900bb'))
+paddle.fluid.layers.linspace (ArgSpec(args=['start', 'stop', 'num', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '495e21e9a848c2d075a102802fc67756'))
 paddle.fluid.layers.While.__init__ (ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.While.block (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.Switch.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -295,12 +301,12 @@ paddle.fluid.layers.DynamicRNN.static_input (ArgSpec(args=['self', 'x'], varargs
 paddle.fluid.layers.DynamicRNN.step_input (ArgSpec(args=['self', 'x', 'level'], varargs=None, keywords=None, defaults=(0,)), ('document', '7568c5ac7622a10288d3307a94134655'))
 paddle.fluid.layers.DynamicRNN.update_memory (ArgSpec(args=['self', 'ex_mem', 'new_mem'], varargs=None, keywords=None, defaults=None), ('document', '5d83987da13b98363d6a807a52d8024f'))
 paddle.fluid.layers.StaticRNN.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.memory (ArgSpec(args=['self', 'init', 'shape', 'batch_ref', 'init_value', 'init_batch_dim_idx', 'ref_batch_dim_idx'], varargs=None, keywords=None, defaults=(None, None, None, 0.0, 0, 1)), ('document', 'c24e368e23afac1ed91a78a639d7a9c7'))
-paddle.fluid.layers.StaticRNN.output (ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.step (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.step_input (ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.step_output (ArgSpec(args=['self', 'o'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.StaticRNN.update_memory (ArgSpec(args=['self', 'mem', 'var'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.layers.StaticRNN.memory (ArgSpec(args=['self', 'init', 'shape', 'batch_ref', 'init_value', 'init_batch_dim_idx', 'ref_batch_dim_idx'], varargs=None, keywords=None, defaults=(None, None, None, 0.0, 0, 1)), ('document', '72530f299d6451a567cf4a12dc3fb1ff'))
+paddle.fluid.layers.StaticRNN.output (ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None), ('document', 'df6ceab6e6c9bd31e97914d7e7538137'))
+paddle.fluid.layers.StaticRNN.step (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6d3e0a5d9aa519a9773a36e1620ea9b7'))
+paddle.fluid.layers.StaticRNN.step_input (ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None), ('document', '903387ec11f3d0bf46821d31a68cffa5'))
+paddle.fluid.layers.StaticRNN.step_output (ArgSpec(args=['self', 'o'], varargs=None, keywords=None, defaults=None), ('document', '252890d4c3199a7623ab8667e13fd837'))
+paddle.fluid.layers.StaticRNN.update_memory (ArgSpec(args=['self', 'mem', 'var'], varargs=None, keywords=None, defaults=None), ('document', '7a0000520f179f35239956a5ba55119f'))
 paddle.fluid.layers.reorder_lod_tensor_by_rank (ArgSpec(args=['x', 'rank_table'], varargs=None, keywords=None, defaults=None), ('document', '3545f529ef04e8f6ecb76b47fa3df01a'))
 paddle.fluid.layers.Print (ArgSpec(args=['input', 'first_n', 'message', 'summarize', 'print_tensor_name', 'print_tensor_type', 'print_tensor_shape', 'print_tensor_lod', 'print_phase'], varargs=None, keywords=None, defaults=(-1, None, -1, True, True, True, True, 'both')), ('document', '5fef91b0e21c93610785f2b1f7161732'))
 paddle.fluid.layers.is_empty (ArgSpec(args=['x', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', 'bbe578dbb49ad13e15b014e98c22b519'))
@@ -359,8 +365,7 @@ paddle.fluid.layers.inverse_time_decay (ArgSpec(args=['learning_rate', 'decay_st
 paddle.fluid.layers.polynomial_decay (ArgSpec(args=['learning_rate', 'decay_steps', 'end_learning_rate', 'power', 'cycle'], varargs=None, keywords=None, defaults=(0.0001, 1.0, False)), ('document', '882634f420f626642f0874481263da40'))
 paddle.fluid.layers.piecewise_decay (ArgSpec(args=['boundaries', 'values'], varargs=None, keywords=None, defaults=None), ('document', 'c717d9d1d78a53c809d01b8bc56f3cae'))
 paddle.fluid.layers.noam_decay (ArgSpec(args=['d_model', 'warmup_steps'], varargs=None, keywords=None, defaults=None), ('document', 'd9a95746353fd574be36dc28d8726c28'))
-paddle.fluid.layers.append_LARS (ArgSpec(args=['params_grads', 'learning_rate', 'weight_decay'], varargs=None, keywords=None, defaults=None), ('document', 'd24fa1e7d62ac8a534fc6a86002f84f8'))
-paddle.fluid.layers.cosine_decay (ArgSpec(args=['learning_rate', 'step_each_epoch', 'epochs'], varargs=None, keywords=None, defaults=None), ('document', '9588c64c26ffaef3c466e404a6af9d9b'))
+paddle.fluid.layers.cosine_decay (ArgSpec(args=['learning_rate', 'step_each_epoch', 'epochs'], varargs=None, keywords=None, defaults=None), ('document', 'f8b2727bccf0f368c997d7cf05847e49'))
 paddle.fluid.layers.linear_lr_warmup (ArgSpec(args=['learning_rate', 'warmup_steps', 'start_lr', 'end_lr'], varargs=None, keywords=None, defaults=None), ('document', '2ef3f5ca5cd71ea4217c418e5a7a0565'))
 paddle.fluid.contrib.InitState.__init__ (ArgSpec(args=['self', 'init', 'shape', 'value', 'init_boot', 'need_reorder', 'dtype'], varargs=None, keywords=None, defaults=(None, None, 0.0, None, False, 'float32')), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.contrib.StateCell.__init__ (ArgSpec(args=['self', 'inputs', 'states', 'out_state', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))

paddle/fluid/framework/data_feed.cc

@@ -72,7 +72,6 @@ bool DataFeed::PickOneFile(std::string* filename) {
   }
   VLOG(3) << "file_idx_=" << *file_idx_;
   *filename = filelist_[(*file_idx_)++];
-  // LOG(ERROR) << "pick file:" << *filename;
   return true;
 }
 
@@ -242,6 +241,11 @@ void InMemoryDataFeed<T>::SetTrainerNum(int trainer_num) {
   trainer_num_ = trainer_num;
 }
 
+template <typename T>
+void InMemoryDataFeed<T>::SetFleetSendBatchSize(int64_t size) {
+  fleet_send_batch_size_ = size;
+}
+
 template <typename T>
 void InMemoryDataFeed<T>::PutInsToChannel(const std::string& ins_str) {
 #ifdef _LINUX
@@ -361,8 +365,13 @@ void InMemoryDataFeed<T>::GlobalShuffle() {
   VLOG(3) << "GlobalShuffle() begin, thread_id=" << thread_id_;
   auto fleet_ptr = FleetWrapper::GetInstance();
   std::vector<std::vector<T*>> send_vec(trainer_num_);
+  std::vector<int> send_index(trainer_num_);
+  uint64_t reserve_len = fleet_send_batch_size_ / trainer_num_;
   for (auto& vec : send_vec) {
-    vec.reserve(fleet_send_batch_size_);
+    vec.reserve(reserve_len);
+  }
+  for (int i = 0; i < trainer_num_; ++i) {
+    send_index[i] = i;
   }
   std::vector<std::future<int32_t>> total_status;
   auto interval = GetMemoryDataInterval();
@@ -375,7 +384,10 @@ void InMemoryDataFeed<T>::GlobalShuffle() {
     int64_t node_id = random_num % trainer_num_;
     send_vec[node_id].push_back(&((*memory_data_)[i]));
     if (i % fleet_send_batch_size_ == 0 && i != 0) {
-      for (int j = 0; j < send_vec.size(); ++j) {
+      // shuffle the sequence of sending to avoid network timeout error
+      std::random_shuffle(send_index.begin(), send_index.end());
+      for (int index = 0; index < send_index.size(); ++index) {
+        int j = send_index[index];
         std::string send_str;
         SerializeIns(send_vec[j], &send_str);
         VLOG(3) << "send str_length=" << send_str.length()
@@ -388,7 +400,10 @@ void InMemoryDataFeed<T>::GlobalShuffle() {
       }
     }
   }
-  for (int j = 0; j < send_vec.size(); ++j) {
+  // shuffle the sequence of sending to avoid network timeout error
+  std::random_shuffle(send_index.begin(), send_index.end());
+  for (int index = 0; index < send_index.size(); ++index) {
+    int j = send_index[index];
     if (send_vec[j].size() != 0) {
       std::string send_str;
       SerializeIns(send_vec[j], &send_str);
@@ -450,6 +465,17 @@ void MultiSlotDataFeed::Init(
     if (slot.is_used()) {
       use_slots_.push_back(all_slots_[i]);
       use_slots_is_dense_.push_back(slot.is_dense());
+      std::vector<int> local_shape;
+      if (slot.is_dense()) {
+        // for batch size holder if is_dense
+        if (slot.shape(0) > 0) {
+          local_shape.push_back(0);
+        }
+      }
+      for (size_t i = 0; i < slot.shape_size(); ++i) {
+        local_shape.push_back(slot.shape(i));
+      }
+      use_slots_shape_.push_back(local_shape);
     }
   }
   feed_vec_.resize(use_slots_.size());
@@ -736,8 +762,8 @@ void MultiSlotDataFeed::PutToFeedVec(
     LoD data_lod{offset};
     feed_vec_[i]->set_lod(data_lod);
     if (use_slots_is_dense_[i]) {
-      int dim = total_instance / batch_size_;
-      feed_vec_[i]->Resize({batch_size_, dim});
+      use_slots_shape_[i][0] = batch_size_;
+      feed_vec_[i]->Resize(framework::make_ddim(use_slots_shape_[i]));
     }
   }
 #endif
@@ -769,6 +795,16 @@ void MultiSlotInMemoryDataFeed::Init(
     if (slot.is_used()) {
       use_slots_.push_back(all_slots_[i]);
       use_slots_is_dense_.push_back(slot.is_dense());
+      std::vector<int> local_shape;
+      if (slot.is_dense()) {
+        if (slot.shape(0) > 0) {
+          local_shape.push_back(0);
+        }
+      }
+      for (size_t i = 0; i < slot.shape_size(); ++i) {
+        local_shape.push_back(slot.shape(i));
+      }
+      use_slots_shape_.push_back(local_shape);
     }
   }
   feed_vec_.resize(use_slots_.size());
@@ -924,8 +960,8 @@ void MultiSlotInMemoryDataFeed::PutToFeedVec(
     LoD data_lod{offset};
     feed_vec_[i]->set_lod(data_lod);
     if (use_slots_is_dense_[i]) {
-      int dim = total_instance / batch_size_;
-      feed_vec_[i]->Resize({batch_size_, dim});
+      use_slots_shape_[i][0] = batch_size_;
+      feed_vec_[i]->Resize(framework::make_ddim(use_slots_shape_[i]));
     }
   }
 #endif

paddle/fluid/framework/data_feed.h

@@ -94,6 +94,8 @@ class DataFeed {
   virtual void SetThreadNum(int thread_num) {}
   // This function will do nothing at default
   virtual void SetTrainerNum(int trainer_num) {}
+  // This function will do nothing at default
+  virtual void SetFleetSendBatchSize(int64_t size) {}
   virtual void SetFileListMutex(std::mutex* mutex) {
     mutex_for_pick_file_ = mutex;
   }
@@ -140,6 +142,7 @@ class DataFeed {
   // object)
   std::vector<std::string> all_slots_;
   std::vector<std::string> all_slots_type_;
+  std::vector<std::vector<int>> use_slots_shape_;
   std::vector<int>
       use_slots_index_;  // -1: not used; >=0: the index of use_slots_
 
@@ -212,6 +215,7 @@ class InMemoryDataFeed : public PrivateQueueDataFeed<T> {
   virtual void SetThreadId(int thread_id);
   virtual void SetThreadNum(int thread_num);
   virtual void SetTrainerNum(int trainer_num);
+  virtual void SetFleetSendBatchSize(int64_t size);
   virtual void PutInsToChannel(const std::string& ins_str);
   virtual void FillMemoryDataToChannel();
   virtual void FillChannelToMemoryData();

paddle/fluid/framework/data_feed.proto

@@ -19,6 +19,7 @@ message Slot {
   required string type = 2;
   optional bool is_dense = 3 [ default = false ];
   optional bool is_used = 4 [ default = false ];
+  repeated int32 shape = 5; // we can define N-D Tensor
 }
 
 message MultiSlotDesc { repeated Slot slots = 1; }

paddle/fluid/framework/data_set.cc

@@ -64,6 +64,17 @@ void DatasetImpl<T>::SetTrainerNum(int trainer_num) {
   }
 }
 
+// if you run distributed, and want to do global shuffle,
+// set this before global shuffle.
+// be sure you call CreateReaders before SetFleetSendBatchSize
+template <typename T>
+void DatasetImpl<T>::SetFleetSendBatchSize(int64_t size) {
+  fleet_send_batch_size_ = size;
+  for (auto reader : readers_) {
+    reader->SetFleetSendBatchSize(size);
+  }
+}
+
 template <typename T>
 void DatasetImpl<T>::SetHdfsConfig(const std::string& fs_name,
                                    const std::string& fs_ugi) {

paddle/fluid/framework/data_set.h

@@ -47,6 +47,8 @@ class Dataset {
   virtual void SetThreadNum(int thread_num) = 0;
   // set workers' num
   virtual void SetTrainerNum(int trainer_num) = 0;
+  // set fleet send batch size
+  virtual void SetFleetSendBatchSize(int64_t size) = 0;
   // set fs name and ugi
   virtual void SetHdfsConfig(const std::string& fs_name,
                              const std::string& fs_ugi) = 0;
@@ -59,6 +61,8 @@ class Dataset {
   virtual int GetThreadNum() = 0;
   // get worker num
   virtual int GetTrainerNum() = 0;
+  // get fleet send batch size
+  virtual int64_t GetFleetSendBatchSize() = 0;
   // get hdfs config
   virtual std::pair<std::string, std::string> GetHdfsConfig() = 0;
   // get data fedd desc
@@ -98,6 +102,7 @@ class DatasetImpl : public Dataset {
   virtual void SetFileList(const std::vector<std::string>& filelist);
   virtual void SetThreadNum(int thread_num);
   virtual void SetTrainerNum(int trainer_num);
+  virtual void SetFleetSendBatchSize(int64_t size);
   virtual void SetHdfsConfig(const std::string& fs_name,
                              const std::string& fs_ugi);
   virtual void SetDataFeedDesc(const std::string& data_feed_desc_str);
@@ -105,6 +110,7 @@ class DatasetImpl : public Dataset {
   virtual const std::vector<std::string>& GetFileList() { return filelist_; }
   virtual int GetThreadNum() { return thread_num_; }
   virtual int GetTrainerNum() { return trainer_num_; }
+  virtual int64_t GetFleetSendBatchSize() { return fleet_send_batch_size_; }
   virtual std::pair<std::string, std::string> GetHdfsConfig() {
     return std::make_pair(fs_name_, fs_ugi_);
   }
@@ -137,6 +143,7 @@ class DatasetImpl : public Dataset {
   std::string fs_name_;
   std::string fs_ugi_;
   unsigned int rand_seed;
+  int64_t fleet_send_batch_size_;
 };
 
 // use std::vector<MultiSlotType> as data type

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -53,6 +53,10 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
       this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
     }
   }
+  // TODO(gongwb) :polish them!
+  if (is_encoded) {
+    VLOG(1) << "Use dgc allreduce mode";
+  }
 }
 #else
 AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
@@ -86,7 +90,7 @@ void AllReduceOpHandle::RunImplEncoded() {
         paddle::framework::GradOriginalVarName(in_var_handles[i]->name());
     auto encode_var_name = original_name + g_dgc_encoded;
     auto *in_var = local_scope->FindVar(encode_var_name);
-    PADDLE_ENFORCE_NOT_NULL(in_var);
+    PADDLE_ENFORCE_NOT_NULL(in_var, "%s should not be null", encode_var_name);
     auto &in = in_var->Get<LoDTensor>();
     ins.emplace_back(&in);
 

paddle/fluid/framework/details/alloc_continuous_space_for_grad_pass.h

+//   Copyright (c) 2019 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 <algorithm>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include "paddle/fluid/framework/details/build_strategy.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+void SetFuseParameterGroupsSize(int group_size);
+int GetFuseParameterGroupsSize();
+
+void SetFuseParameterMemorySize(uint64_t memory_size);
+uint64_t GetFuseParameterMemorySize();
+
+class AllocContinuousSpaceForGradPass : public ir::Pass {
+ protected:
+  void ApplyImpl(ir::Graph *graph) const override;
+
+  template <typename AttrType>
+  void ResetAttribute(const std::string &attr_name, ir::Graph *graph) const;
+
+  void SetGroupGradsAndParams(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      const ParamsAndGrads &params_grads,
+      GroupGradsAndParams *group_grads_params) const;
+
+  void SetGroupAccordingToLayers(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      const ParamsAndGrads &params_grads,
+      GroupGradsAndParams *group_grads_params) const;
+
+  void SetGroupAccordingToMemorySize(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      GroupGradsAndParams *group_grads_params) const;
+
+  void SetGroupAccordingToGroupSize(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      GroupGradsAndParams *group_grads_params) const;
+
+ private:
+  bool IsSupportedVarType(const proto::VarType::Type &type) const;
+
+  void RecordParamsAndGrads(ir::Node *node, ParamsAndGrads *params_grads) const;
+
+  void InitFusedVarsAndAllocSpaceForVars(
+      const std::vector<platform::Place> &places,
+      const std::vector<Scope *> &local_scopes,
+      const std::unordered_map<std::string, ir::Node *> &vars,
+      const std::string &fused_var_name,
+      const ParamsAndGrads &params_grads) const;
+
+  void AppendAllocSpaceForVarsOp(const std::vector<std::string> &params_name,
+                                 const std::vector<std::string> &grads_name,
+                                 const std::string &fused_var_name,
+                                 BlockDesc *global_block) const;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/async_ssa_graph_executor.cc

@@ -64,9 +64,12 @@ void ProcessGraph(std::vector<ir::Graph *> graphs, Scope *scope) {
               node->Op()->GetNullableAttr("epmap"));
           auto height_section = boost::get<std::vector<int64_t>>(
               node->Op()->GetNullableAttr("sections"));
+          auto trainer_id =
+              boost::get<int>(node->Op()->GetNullableAttr("trainer_id"));
           send_varname_to_ctx[send_var_name] =
               operators::distributed::RpcContext(send_var_name, send_varnames,
-                                                 epmap, height_section);
+                                                 epmap, height_section,
+                                                 trainer_id);
           VLOG(3) << "find and init an send op: "
                   << send_varname_to_ctx[send_var_name];
         } else if (node->Name() == "recv") {
@@ -75,9 +78,11 @@ void ProcessGraph(std::vector<ir::Graph *> graphs, Scope *scope) {
               node->Op()->GetNullableAttr("recv_varnames"));
           auto epmap = boost::get<std::vector<std::string>>(
               node->Op()->GetNullableAttr("epmap"));
+          auto trainer_id =
+              boost::get<int>(node->Op()->GetNullableAttr("trainer_id"));
           recv_varname_to_ctx[recv_var_name] =
               operators::distributed::RpcContext(recv_var_name, recv_varnames,
-                                                 epmap, {});
+                                                 epmap, {}, trainer_id);
           nodes_to_delete.push_back(node);
           VLOG(3) << "find and remove an recv op: "
                   << recv_varname_to_ctx[recv_var_name];

paddle/fluid/framework/details/build_strategy.cc

@@ -101,8 +101,6 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
                "mode.";
         strategy_.fuse_all_optimizer_ops_ = false;
       } else {
-        VLOG(10) << "Add alloc_continuous_space_for_grad_pass";
-        AppendPass("alloc_continuous_space_for_grad_pass");
         // NOTE: fuse_all_xx_ops will count the number of xx operator first,
         // if the number is zero, fuse_all_reduce_ops will do nothing.
         // Currently, only one type of optimization algorithm can be fused.
@@ -142,6 +140,19 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
       AppendPass("memory_optimize_pass");
     }
 
+    // runtime_context_cache pass should be the last pass to enable the attr of
+    // all original and fused operators. But no operators can be enabled this
+    // attr if putting it after MultiDevPass.
+    if (strategy_.cache_runtime_context_) {
+      VLOG(10) << "Add runtime_context_cache_pass";
+      AppendPass("runtime_context_cache_pass");
+    }
+
+    if (strategy_.cache_expected_kernel_) {
+      VLOG(10) << "Add expected_kernel_cache_pass";
+      AppendPass("expected_kernel_cache_pass");
+    }
+
     AppendMultiDevPass(strategy_);
 
     if (strategy_.fuse_all_reduce_ops_) {
@@ -243,7 +254,7 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
   CreatePassesFromStrategy(false);
 
   for (std::shared_ptr<ir::Pass> &pass : pass_builder_->AllPasses()) {
-    VLOG(3) << "apply " << pass->Type();
+    VLOG(3) << "BuildStrategy::Apply pass:" << pass->Type();
     if (IsMultiDevPass(pass->Type())) {
       pass->Erase(kPlaces);
       pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
@@ -328,3 +339,5 @@ USE_PASS(graph_to_program_pass);
 USE_PASS(fuse_adam_op_pass);
 USE_PASS(fuse_sgd_op_pass);
 USE_PASS(fuse_all_reduce_op_pass);
+USE_PASS(runtime_context_cache_pass);
+USE_PASS(expected_kernel_cache_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -83,11 +83,11 @@ struct BuildStrategy {
 
   bool sync_batch_norm_{false};
 
-  bool memory_optimize_{true};
-  // TODO(dzhwinter):
-  // make enable_inplace, memory_optimize_
-  // memory_early_delete_ true by default
-  bool enable_inplace_{true};
+  // FIXME(liuwei1031) disable memory_optimzie and enable_inplace in 1.4
+  // to open them by default, we need to solve the fetch variable issue
+  bool memory_optimize_{false};
+
+  bool enable_inplace_{false};
 
   bool enable_sequential_execution_{false};
 
@@ -107,6 +107,9 @@ struct BuildStrategy {
   std::vector<std::string> trainers_endpoints_;
   bool remove_unnecessary_lock_{true};
 
+  bool cache_runtime_context_{false};
+  bool cache_expected_kernel_{true};
+
   // NOTE:
   // Before you add new options, think if it's a general strategy that works
   // with other strategy. If not, the strategy should be created through

paddle/fluid/framework/details/fuse_adam_op_pass.cc

@@ -24,7 +24,7 @@ namespace details {
 const std::string FuseAdamOpPass::GetOpType() const { return "adam"; }
 
 const std::vector<std::string> FuseAdamOpPass::GetAuxiliaryVarNames() const {
-  return {"Param", "Moment1", "Moment2", "Beta1Pow", "Beta2Pow"};
+  return {"Moment1", "Moment2", "Beta1Pow", "Beta2Pow"};
 }
 
 void FuseAdamOpPass::FuseOptimizerOps(
@@ -77,16 +77,16 @@ void FuseAdamOpPass::FuseAdamOps(
   VLOG(10) << "Insert adam to graph ";
   OpDesc adam_desc(adam_ops[0]->Op()->Block());
   adam_desc.SetType("adam");
-  adam_desc.SetInput("Param", {fused_vars_name.at("Param")});
-  adam_desc.SetInput("Grad", {fused_vars_name.at("Grad")});
+  adam_desc.SetInput(kParam, {fused_vars_name.at(kParam)});
+  adam_desc.SetInput(kGrad, {fused_vars_name.at(kGrad)});
   adam_desc.SetInput("Moment1", {fused_vars_name.at("Moment1")});
   adam_desc.SetInput("Moment2", {fused_vars_name.at("Moment2")});
   // TODO(zcd): The LearningRate, Beta1Pow, Beta2Pow should be equal.
-  adam_desc.SetInput("LearningRate", adam_ops[0]->Op()->Input("LearningRate"));
+  adam_desc.SetInput(kLearningRate, adam_ops[0]->Op()->Input(kLearningRate));
   adam_desc.SetInput("Beta1Pow", adam_ops[0]->Op()->Input("Beta1Pow"));
   adam_desc.SetInput("Beta2Pow", adam_ops[0]->Op()->Input("Beta2Pow"));
 
-  adam_desc.SetOutput("ParamOut", {fused_vars_name.at("Param")});
+  adam_desc.SetOutput("ParamOut", {fused_vars_name.at(kParam)});
   adam_desc.SetOutput("Moment1Out", {fused_vars_name.at("Moment1")});
   adam_desc.SetOutput("Moment2Out", {fused_vars_name.at("Moment2")});
   adam_desc.SetAttr("beta1", beta1);

paddle/fluid/framework/details/fuse_optimizer_op_pass.cc

@@ -29,7 +29,9 @@ void FuseOptimizerOpPass::ApplyImpl(ir::Graph *graph) const {
   auto &local_scopes = Get<const std::vector<Scope *>>(kLocalScopes);
 
   const std::string fuse_op_type = GetOpType();
-  const std::vector<std::string> aux_var_names = GetAuxiliaryVarNames();
+  std::vector<std::string> aux_var_names = GetAuxiliaryVarNames();
+  aux_var_names.emplace_back(kParam);
+  aux_var_names.emplace_back(kGrad);
 
   // Step 1: Get the specified op and auxiliary variables.
   std::vector<ir::Node *> topo_nodes = ir::TopologySortOperations(result);
@@ -61,7 +63,7 @@ void FuseOptimizerOpPass::ApplyImpl(ir::Graph *graph) const {
     result.Set(kFusedVars, new FusedVars);
   }
   std::unordered_map<std::string, std::string> fused_vars_name;
-  fused_vars_name.reserve(aux_var_names.size() + 1);
+  fused_vars_name.reserve(aux_var_names.size());
   auto &fused_var_set = result.Get<FusedVars>(kFusedVars);
   const std::string prefix(kFusedVarNamePrefix);
   // NOTE: the fused_var_name should be unique.
@@ -75,39 +77,103 @@ void FuseOptimizerOpPass::ApplyImpl(ir::Graph *graph) const {
   }
 
   // Step 3: Get the fused Gradient's name
-  auto &params_grads = result.Get<ParamsAndGrads>(kParamsAndGrads);
-  if (!result.Has(kFusedGrads)) {
-    PADDLE_THROW(
-        "The alloc_continuous_space_for_grad_pass should be called before this "
-        "pass.");
-  }
-  auto &fused_grad = result.Get<FusedGrads>(kFusedGrads);
-  auto &fused_vars = result.Get<FusedVars>(kFusedVars);
-  auto iter = std::find(fused_vars.begin(), fused_vars.end(), fused_grad);
-  PADDLE_ENFORCE(iter != fused_vars.end(), "Not find the fused_grad.");
-  fused_vars_name.emplace("Grad", fused_grad);
-
-  // Step 4: Sort the parameters and auxiliary variables according
-  // to parameters' name to make variables' name correspond correctly.
-  PADDLE_ENFORCE(result.Has(kParamsAndGrads), "Does't find kParamsAndGrads.");
-  PADDLE_ENFORCE_EQ(params_grads.size(), aux_var_set.begin()->second.size(),
-                    "The size of params_grads and aux_var_set are not equal.");
-  SortParametersAndAuxVars(params_grads, &aux_var_set, &opt_ops);
-
-  // Step 5: Alloc continuous space for Parameters and AuxiliaryVar(e.g.
+  bool grad_fused = false;
+  if (result.Has(kParamsAndGrads)) {
+    auto &params_grads = result.Get<ParamsAndGrads>(kParamsAndGrads);
+    PADDLE_ENFORCE_EQ(
+        params_grads.size(), aux_var_set.at(kGrad).size(),
+        "The number of gradients and optimizer ops is not equal.");
+    std::unordered_set<std::string> opt_grad_set(aux_var_set.at(kGrad).begin(),
+                                                 aux_var_set.at(kGrad).end());
+    size_t same_grad_num = 0;
+    for (auto &p_g : params_grads) {
+      if (opt_grad_set.count(p_g.second)) {
+        ++same_grad_num;
+      }
+    }
+
+    // NOTE(zcd): the gradient of kParamsAndGrads may be different with the
+    // kGrad.
+    if (same_grad_num == aux_var_set.at(kGrad).size()) {
+      if (!result.Has(kFusedGrads)) {
+        PADDLE_THROW(
+            "The alloc_continuous_space_for_grad_pass should be called before "
+            "this pass.");
+      }
+      auto &fused_grad = result.Get<FusedGrads>(kFusedGrads);
+      auto &fused_vars = result.Get<FusedVars>(kFusedVars);
+      auto iter = std::find(fused_vars.begin(), fused_vars.end(), fused_grad);
+      PADDLE_ENFORCE(iter != fused_vars.end(), "Not find the fused_grad.");
+      fused_vars_name[kGrad] = fused_grad;
+
+      // Sort the parameters and auxiliary variables according
+      // to parameters' name to make variables' name correspond correctly.
+      SortParametersAndAuxVars(params_grads, &aux_var_set, &opt_ops);
+      grad_fused = true;
+    }
+  }
+
+  // Step 4: Alloc continuous space for Parameters and AuxiliaryVar(e.g.
   // Moment1, Moment2, Beta1Pow, Beta2Pow) of all the optimizer ops separately.
+  aux_var_names.pop_back();
+  if (!grad_fused) {
+    InitFusedGradsAndAllocSpaceForGrads(
+        places, local_scopes, aux_var_set.at(kParam), aux_var_set.at(kGrad),
+        fused_vars_name.at(kGrad), &result);
+  }
   InitFusedVarsAndAllocSpaceForVars(places, local_scopes, aux_var_names,
                                     aux_var_set, fused_vars_name);
 
-  // Step 6: Fuse optimizer Ops and Scale Ops
+  // Step 5: Fuse optimizer Ops and Scale Ops
   FuseOptimizerOps(aux_var_set, fused_vars_name, opt_ops, &result);
 
-  // Step 7: Remove optimizer Ops
+  // Step 6: Remove optimizer Ops
   for (auto &opt_op : opt_ops) {
     graph->RemoveNode(opt_op);
   }
 }
 
+void FuseOptimizerOpPass::InitFusedGradsAndAllocSpaceForGrads(
+    const std::vector<platform::Place> &places,
+    const std::vector<Scope *> &local_scopes,
+    const std::vector<std::string> &params,
+    const std::vector<std::string> &grads, const std::string &fused_grad_name,
+    ir::Graph *result) const {
+  // Get Var Nodes
+  std::unordered_map<std::string, ir::Node *> vars;
+  for (ir::Node *node : result->Nodes()) {
+    if (node->IsVar() && node->Var()) {
+      // Note: The graph may have the same name node. For example, parameter
+      // is the input of operator and it also is the output of optimizer;
+      vars.emplace(node->Var()->Name(), node);
+    }
+  }
+  // Init Grads
+  for (auto it = local_scopes.rbegin(); it != local_scopes.rend(); ++it) {
+    auto &scope = *it;
+    VLOG(10) << "Init " << fused_grad_name;
+    PADDLE_ENFORCE(scope->FindVar(fused_grad_name) == nullptr,
+                   "%s has existed in scope.", fused_grad_name);
+    scope->Var(fused_grad_name)->GetMutable<LoDTensor>();
+
+    for (auto &grad_var_name : grads) {
+      auto iter = vars.find(grad_var_name);
+      PADDLE_ENFORCE(iter != vars.end());
+      PADDLE_ENFORCE_NOT_NULL(iter->second->Var());
+      PADDLE_ENFORCE_EQ(iter->second->Var()->GetType(),
+                        proto::VarType::LOD_TENSOR);
+      scope->Var(grad_var_name)->GetMutable<LoDTensor>();
+    }
+  }
+  // Define Ops
+  ProgramDesc program_desc;
+  auto *global_block = program_desc.MutableBlock(0);
+  AppendAllocContinuousSpace(params, grads, fused_grad_name, global_block,
+                             false, false);
+  // Run Ops
+  RunInitOps(places, local_scopes, *global_block);
+}
+
 void FuseOptimizerOpPass::InitFusedVarsAndAllocSpaceForVars(
     const std::vector<platform::Place> &places,
     const std::vector<Scope *> &local_scopes,
@@ -115,37 +181,49 @@ void FuseOptimizerOpPass::InitFusedVarsAndAllocSpaceForVars(
     const std::unordered_map<std::string, std::vector<std::string>>
         &aux_var_set,
     const std::unordered_map<std::string, std::string> &fused_vars_name) const {
-  VLOG(10) << "Init FusedVars.";
-  // Alloc parameters and auxiliary vars in the respective scope.
-  size_t idx = local_scopes.size();
-  for (auto iter = local_scopes.rbegin(); iter != local_scopes.rend();
-       ++iter, --idx) {
-    auto &scope = *iter;
-    for (auto &var_name : aux_var_names) {
-      auto fused_var_name = fused_vars_name.at(var_name);
-      VLOG(10) << "Init " << fused_var_name;
-      PADDLE_ENFORCE(scope->FindVar(fused_var_name) == nullptr,
-                     "%s has exist in scope[%d]", fused_var_name, idx);
-      scope->Var(fused_var_name)->GetMutable<LoDTensor>();
-    }
+  // Init Vars
+  for (auto &var_name : aux_var_names) {
+    auto &fused_var_name = fused_vars_name.at(var_name);
+    InitVars(local_scopes, fused_var_name);
   }
-
+  // Define Ops
   ProgramDesc program_desc;
   auto *global_block = program_desc.MutableBlock(0);
   for (auto &var_name : aux_var_names) {
-    AppendAllocContinuousSpace(aux_var_set.at(var_name),
-                               fused_vars_name.at(var_name), true,
-                               global_block);
+    AppendAllocContinuousSpace(
+        aux_var_set.at(var_name), aux_var_set.at(var_name),
+        fused_vars_name.at(var_name), global_block, true);
   }
+  // Run Ops
+  RunInitOps(places, local_scopes, *global_block);
+}
 
+void FuseOptimizerOpPass::RunInitOps(const std::vector<platform::Place> &places,
+                                     const std::vector<Scope *> &local_scopes,
+                                     const BlockDesc &global_block) const {
   for (size_t i = 0; i < local_scopes.size(); ++i) {
-    for (auto &op_desc : global_block->AllOps()) {
+    for (auto &op_desc : global_block.AllOps()) {
       auto op = OpRegistry::CreateOp(*op_desc);
       op->Run(*local_scopes[i], places[i]);
     }
   }
 }
 
+void FuseOptimizerOpPass::InitVars(const std::vector<Scope *> &local_scopes,
+                                   const std::string &fused_var_name) const {
+  VLOG(10) << "Init FusedVars.";
+  // Alloc parameters and auxiliary vars in the respective scope.
+  size_t idx = local_scopes.size();
+  for (auto iter = local_scopes.rbegin(); iter != local_scopes.rend();
+       ++iter, --idx) {
+    auto &scope = *iter;
+    VLOG(10) << "Init " << fused_var_name;
+    PADDLE_ENFORCE(scope->FindVar(fused_var_name) == nullptr,
+                   "%s has exist in scope[%d]", fused_var_name, idx);
+    scope->Var(fused_var_name)->GetMutable<LoDTensor>();
+  }
+}
+
 void FuseOptimizerOpPass::SortParametersAndAuxVars(
     const std::vector<std::pair<std::string, std::string>> &params_grads,
     std::unordered_map<std::string, std::vector<std::string>> *aux_vars_set,
@@ -203,15 +281,16 @@ void FuseOptimizerOpPass::GetSpecifiedOpsAndVars(
 }
 
 void FuseOptimizerOpPass::AppendAllocContinuousSpace(
-    const std::vector<std::string> &args, const std::string &out_arg,
-    bool copy_data, BlockDesc *global_block) const {
+    const std::vector<std::string> &in_args,
+    const std::vector<std::string> &out_args, const std::string &fused_out_arg,
+    BlockDesc *global_block, bool copy_data, bool check_name) const {
   auto op_desc = global_block->AppendOp();
   op_desc->SetType("alloc_continuous_space");
-  op_desc->SetInput("Input", args);
-  op_desc->SetOutput("Output", args);
-  op_desc->SetOutput("FusedOutput", {out_arg});
+  op_desc->SetInput("Input", in_args);
+  op_desc->SetOutput("Output", out_args);
+  op_desc->SetOutput("FusedOutput", {fused_out_arg});
   op_desc->SetAttr("copy_data", copy_data);
-  op_desc->SetAttr("check_name", true);
+  op_desc->SetAttr("check_name", check_name);
 }
 
 void FuseOptimizerOpPass::InserInputAndOutputForOptOps(

paddle/fluid/framework/details/fuse_optimizer_op_pass.h

@@ -27,6 +27,10 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+constexpr char kGrad[] = "Grad";
+constexpr char kParam[] = "Param";
+constexpr char kLearningRate[] = "LearningRate";
+
 class FuseOptimizerOpPass : public ir::Pass {
  protected:
   void ApplyImpl(ir::Graph *graph) const override;
@@ -56,9 +60,18 @@ class FuseOptimizerOpPass : public ir::Pass {
       std::unordered_map<std::string, std::vector<std::string>> *aux_args_name)
       const;
 
-  void AppendAllocContinuousSpace(const std::vector<std::string> &args,
-                                  const std::string &out_arg, bool copy_data,
-                                  BlockDesc *global_block) const;
+  void AppendAllocContinuousSpace(const std::vector<std::string> &in_args,
+                                  const std::vector<std::string> &out_args,
+                                  const std::string &fused_out_arg,
+                                  BlockDesc *global_block, bool copy_data,
+                                  bool check_name = true) const;
+
+  void InitFusedGradsAndAllocSpaceForGrads(
+      const std::vector<platform::Place> &places,
+      const std::vector<Scope *> &local_scopes,
+      const std::vector<std::string> &params,
+      const std::vector<std::string> &grads, const std::string &fused_grad_name,
+      ir::Graph *result) const;
 
   void InitFusedVarsAndAllocSpaceForVars(
       const std::vector<platform::Place> &places,
@@ -68,6 +81,13 @@ class FuseOptimizerOpPass : public ir::Pass {
           &aux_var_set,
       const std::unordered_map<std::string, std::string> &fused_vars_name)
       const;
+
+  void RunInitOps(const std::vector<platform::Place> &places,
+                  const std::vector<Scope *> &local_scopes,
+                  const BlockDesc &global_block) const;
+
+  void InitVars(const std::vector<Scope *> &local_scopes,
+                const std::string &fused_var_name) const;
 };
 
 }  // namespace details

paddle/fluid/framework/details/fuse_sgd_op_pass.cc

@@ -24,7 +24,7 @@ namespace details {
 const std::string FuseSgdOpPass::GetOpType() const { return "sgd"; }
 
 const std::vector<std::string> FuseSgdOpPass::GetAuxiliaryVarNames() const {
-  return {"Param"};
+  return {};
 }
 
 void FuseSgdOpPass::FuseOptimizerOps(
@@ -50,12 +50,12 @@ void FuseSgdOpPass::FuseSgdOps(
   // Add fused scale
   OpDesc Sgd_desc(sgd_ops[0]->Op()->Block());
   Sgd_desc.SetType("sgd");
-  Sgd_desc.SetInput("Param", {fused_vars_name.at("Param")});
-  Sgd_desc.SetInput("Grad", {fused_vars_name.at("Grad")});
-  Sgd_desc.SetOutput("ParamOut", {fused_vars_name.at("Param")});
+  Sgd_desc.SetInput(kParam, {fused_vars_name.at(kParam)});
+  Sgd_desc.SetInput(kGrad, {fused_vars_name.at(kGrad)});
+  Sgd_desc.SetOutput("ParamOut", {fused_vars_name.at(kParam)});
 
   // TODO(zcd): The LearningRate, Beta1Pow, Beta2Pow should be equal.
-  Sgd_desc.SetInput("LearningRate", sgd_ops[0]->Op()->Input("LearningRate"));
+  Sgd_desc.SetInput(kLearningRate, sgd_ops[0]->Op()->Input(kLearningRate));
 
   // NOTE: multi_devices_pass requires that every op should have a role.
   Sgd_desc.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(), op_role);

paddle/fluid/framework/details/inplace_op_pass.cc

@@ -305,6 +305,12 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
 
     VLOG(4) << "Try to inplace " << in_var_name << " with " << out_var_name;
 
+    if (var_nodes_[in_var_name].back() != in_node) {
+      VLOG(4) << "SKIP since " << in_var_name
+              << " is also used as output by other ops";
+      continue;
+    }
+
     bool can_replace = true;
     if (in_var_name == out_var_name) {
       can_replace = false;
@@ -527,6 +533,9 @@ void GraphView::Build(ir::Graph* g) {
   };
   for (auto& node : g->Nodes()) {
     if (!node->IsOp()) continue;
+    // avoid optimize the variable used in sub-blocks
+    if (OpHasSubBlock(node->Op())) update_skip_set(node);
+
     if (node->Name() == "send") update_skip_set(node);
     if (node->Name() == "recv") update_skip_set(node);
     if (node->Name() == "prefetch") update_skip_set(node);

paddle/fluid/framework/details/op_registry.h

@@ -233,6 +233,12 @@ struct OpInfoFiller<T, kNoNeedBufferVarsInference> {
   }
 };
 
+// A fake OpInfoFiller of void
+template <>
+struct OpInfoFiller<void, kUnknown> {
+  void operator()(const char* op_type, OpInfo* info) const {}
+};
+
 }  // namespace details
 
 }  // namespace framework

paddle/fluid/framework/details/parallel_ssa_graph_executor.cc

@@ -106,7 +106,7 @@ ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
   VLOG(1) << "set num_threads: " << strategy_.num_threads_
           << " to run the operators of the graph on each device.";
   for (size_t i = 0; i < places.size(); ++i) {
-    executors_.emplace_back(new details::ThreadedSSAGraphExecutor(
+    executors_.emplace_back(new details::FastThreadedSSAGraphExecutor(
         strategy_, local_scopes_, {places_[i]}, graphs_.at(i).get()));
   }
 }

paddle/fluid/framework/details/parallel_ssa_graph_executor.h

@@ -14,12 +14,12 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <vector>
-
 #include "ThreadPool.h"
+#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
-#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
 #include "paddle/fluid/framework/ir/graph.h"
 
 namespace paddle {
@@ -48,7 +48,8 @@ class ParallelSSAGraphExecutor : public SSAGraphExecutor {
   std::vector<platform::Place> places_;
   std::vector<std::unique_ptr<ir::Graph>> graphs_;
 
-  std::vector<std::unique_ptr<details::ThreadedSSAGraphExecutor>> executors_;
+  std::vector<std::unique_ptr<details::FastThreadedSSAGraphExecutor>>
+      executors_;
   ExceptionHolder exception_holder_;
 };
 

paddle/fluid/framework/downpour_worker.cc

@@ -21,40 +21,40 @@ namespace framework {
 
 void DownpourWorker::Initialize(const TrainerDesc& desc) {
   param_ = desc.downpour_param();
-  for (size_t i = 0; i < param_.sparse_table_size(); ++i) {
+  for (int i = 0; i < param_.sparse_table_size(); ++i) {
     uint64_t table_id =
         static_cast<uint64_t>(param_.sparse_table(i).table_id());
     TableParameter table = param_.sparse_table(i);
     sparse_key_names_[table_id].resize(table.sparse_key_name_size());
-    for (size_t j = 0; j < table.sparse_key_name_size(); ++j) {
+    for (int j = 0; j < table.sparse_key_name_size(); ++j) {
       sparse_key_names_[table_id][j] = table.sparse_key_name(j);
     }
     sparse_value_names_[table_id].resize(table.sparse_value_name_size());
-    for (size_t j = 0; j < table.sparse_value_name_size(); ++j) {
+    for (int j = 0; j < table.sparse_value_name_size(); ++j) {
       sparse_value_names_[table_id][j] = table.sparse_value_name(j);
     }
     sparse_grad_names_[table_id].resize(table.sparse_grad_name_size());
-    for (size_t j = 0; j < table.sparse_grad_name_size(); ++j) {
+    for (int j = 0; j < table.sparse_grad_name_size(); ++j) {
       sparse_grad_names_[table_id][j] = table.sparse_grad_name(j);
     }
     label_var_name_[table_id] = table.label_var_name();
   }
 
-  for (size_t i = 0; i < param_.dense_table_size(); ++i) {
+  for (int i = 0; i < param_.dense_table_size(); ++i) {
     uint64_t table_id = static_cast<uint64_t>(param_.dense_table(i).table_id());
     auto table = param_.dense_table(i);
     dense_value_names_[table_id].resize(table.dense_value_name_size());
-    for (size_t j = 0; j < table.dense_value_name_size(); ++j) {
+    for (int j = 0; j < table.dense_value_name_size(); ++j) {
       dense_value_names_[table_id][j] = table.dense_value_name(j);
     }
     dense_grad_names_[table_id].resize(table.dense_grad_name_size());
-    for (size_t j = 0; j < table.dense_grad_name_size(); ++j) {
+    for (int j = 0; j < table.dense_grad_name_size(); ++j) {
       dense_grad_names_[table_id][j] = table.dense_grad_name(j);
     }
   }
 
   skip_ops_.resize(param_.skip_ops_size());
-  for (size_t i = 0; i < param_.skip_ops_size(); ++i) {
+  for (int i = 0; i < param_.skip_ops_size(); ++i) {
     skip_ops_[i] = param_.skip_ops(i);
   }
 
@@ -83,14 +83,14 @@ void DownpourWorker::CollectLabelInfo(size_t table_idx) {
   LoDTensor* tensor = var->GetMutable<LoDTensor>();
   int64_t* label_ptr = tensor->data<int64_t>();
 
-  int global_index = 0;
+  size_t global_index = 0;
   for (size_t i = 0; i < sparse_key_names_[table_id].size(); ++i) {
     VLOG(3) << "sparse_key_names_[" << i
             << "]: " << sparse_key_names_[table_id][i];
     Variable* fea_var = thread_scope_->FindVar(sparse_key_names_[table_id][i]);
     LoDTensor* tensor = fea_var->GetMutable<LoDTensor>();
     int64_t* ids = tensor->data<int64_t>();
-    int fea_idx = 0;
+    size_t fea_idx = 0;
     // tensor->lod()[0].size() == batch_size + 1
     for (auto lod_idx = 1u; lod_idx < tensor->lod()[0].size(); ++lod_idx) {
       for (; fea_idx < tensor->lod()[0][lod_idx]; ++fea_idx) {
@@ -138,7 +138,7 @@ void DownpourWorker::FillSparseValue(size_t table_idx) {
     auto& tensor_lod = tensor->lod()[0];
     LoD data_lod{tensor_lod};
     tensor_emb->set_lod(data_lod);
-    for (auto index = 0u; index < len; ++index) {
+    for (int index = 0; index < len; ++index) {
       if (ids[index] == 0u) {
         memcpy(ptr + table.emb_dim() * index, init_value.data() + 2,
                sizeof(float) * table.emb_dim());
@@ -192,7 +192,7 @@ void DownpourWorker::TrainFilesWithProfiler() {
     read_time += timeline.ElapsedSec();
     total_time += timeline.ElapsedSec();
     VLOG(3) << "program config size: " << param_.program_config_size();
-    for (size_t i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
+    for (int i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
          ++i) {
       uint64_t tid = static_cast<uint64_t>(
           param_.program_config(0).pull_sparse_table_id(i));
@@ -244,8 +244,8 @@ void DownpourWorker::TrainFilesWithProfiler() {
     }
 
     if (need_to_push_sparse_) {
-      for (size_t i = 0;
-           i < param_.program_config(0).push_sparse_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_sparse_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_sparse_table_id(i));
         TableParameter table;
@@ -268,8 +268,8 @@ void DownpourWorker::TrainFilesWithProfiler() {
 
     if (need_to_push_dense_) {
       timeline.Start();
-      for (size_t i = 0;
-           i < param_.program_config(0).push_dense_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_dense_table_id(i));
         fleet_ptr_->PushDenseVarsAsync(
@@ -315,8 +315,8 @@ void DownpourWorker::TrainFilesWithProfiler() {
     }
 
     if (need_to_push_dense_) {
-      for (size_t i = 0;
-           i < param_.program_config(0).push_dense_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_dense_table_id(i));
         pull_dense_worker_->IncreaseThreadVersion(thread_id_, tid);
@@ -362,7 +362,7 @@ void DownpourWorker::TrainFiles() {
   int cur_batch;
   while ((cur_batch = device_reader_->Next()) > 0) {
     // pull sparse here
-    for (size_t i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
+    for (int i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
          ++i) {
       uint64_t tid = static_cast<uint64_t>(
           param_.program_config(0).pull_sparse_table_id(i));
@@ -397,8 +397,8 @@ void DownpourWorker::TrainFiles() {
 
     if (need_to_push_sparse_) {
       // push gradients here
-      for (size_t i = 0;
-           i < param_.program_config(0).push_sparse_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_sparse_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_sparse_table_id(i));
         TableParameter table;
@@ -416,8 +416,8 @@ void DownpourWorker::TrainFiles() {
     }
 
     if (need_to_push_dense_) {
-      for (size_t i = 0;
-           i < param_.program_config(0).push_dense_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_dense_table_id(i));
         fleet_ptr_->PushDenseVarsAsync(
@@ -461,8 +461,8 @@ void DownpourWorker::TrainFiles() {
     }
 
     if (need_to_push_dense_) {
-      for (size_t i = 0;
-           i < param_.program_config(0).push_dense_table_id_size(); ++i) {
+      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
+           ++i) {
         uint64_t tid = static_cast<uint64_t>(
             param_.program_config(0).push_dense_table_id(i));
         pull_dense_worker_->IncreaseThreadVersion(thread_id_, tid);

paddle/fluid/framework/fleet/fleet_wrapper.cc

@@ -237,6 +237,7 @@ void FleetWrapper::PushDenseParamSync(
   std::vector<paddle::ps::Region> regions;
   for (auto& t : var_names) {
     Variable* var = scope.FindVar(t);
+    CHECK(var != nullptr) << "var[" << t << "] not found";
     LoDTensor* tensor = var->GetMutable<LoDTensor>();
     float* g = tensor->mutable_data<float>(place);
     paddle::ps::Region reg(g, tensor->numel());

paddle/fluid/framework/io/shell.cc

@@ -126,7 +126,7 @@ static int shell_popen_fork_internal(const char* real_cmd, bool do_read,
   }
 
   close_open_fds_internal();
-  if (execl("/bin/sh", "sh", "-c", real_cmd, NULL) < 0) {
+  if (execl("/bin/bash", "bash", "-c", real_cmd, NULL) < 0) {
     return -1;
   }
   exit(127);

paddle/fluid/framework/ir/CMakeLists.txt

@@ -68,6 +68,7 @@ pass_library(transpose_flatten_concat_fuse_pass inference)
 pass_library(identity_scale_op_clean_pass base)
 pass_library(sync_batch_norm_pass base)
 pass_library(runtime_context_cache_pass base)
+pass_library(expected_kernel_cache_pass base)
 pass_library(quant_conv2d_dequant_fuse_pass inference)
 pass_library(fillconstant_elementwisemul_fuse inference)
 

paddle/fluid/framework/ir/expected_kernel_cache_pass.cc

+/* Copyright (c) 2019 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/ir/expected_kernel_cache_pass.h"
+#include <memory>
+#include "paddle/fluid/framework/operator.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+void ExpectedKernelCachePass::ApplyImpl(ir::Graph* graph) const {
+  VLOG(3) << "Applies Expected Kernel Cache strategy.";
+  for (const Node* n : graph->Nodes()) {
+    if (n->IsOp() && n->Op()) {
+      n->Op()->SetAttr(kEnableCacheExpectedKernel, true);
+    }
+  }
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(expected_kernel_cache_pass,
+              paddle::framework::ir::ExpectedKernelCachePass);

paddle/fluid/platform/dynload/wbaes.cc → paddle/fluid/framework/ir/expected_kernel_cache_pass.h

@@ -12,23 +12,20 @@ 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. */
 
-#ifdef PADDLE_WITH_WBAES
+#pragma once
 
-#include "paddle/fluid/platform/dynload/wbaes.h"
+#include <memory>
+#include "paddle/fluid/framework/ir/pass.h"
 
 namespace paddle {
-namespace platform {
-namespace dynload {
+namespace framework {
+namespace ir {
 
-std::once_flag wbaes_dso_flag;
-void *wbaes_dso_handle = nullptr;
+class ExpectedKernelCachePass : public Pass {
+ protected:
+  void ApplyImpl(ir::Graph* graph) const override;
+};
 
-#define DEFINE_WRAP(__name) DynLoad__##__name __name
-
-WBAES_ROUTINE_EACH(DEFINE_WRAP);
-
-}  // namespace dynload
-}  // namespace platform
+}  // namespace ir
+}  // namespace framework
 }  // namespace paddle
-
-#endif

paddle/fluid/framework/ir/graph_helper.cc

@@ -31,10 +31,10 @@ namespace paddle {
 namespace framework {
 namespace ir {
 namespace {
-void SortHelper(
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
-    ir::Node *node, std::unordered_set<ir::Node *> *visited,
-    std::vector<ir::Node *> *ret) {
+void SortHelper(const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>,
+                               ir::NodeComp> &adj_list,
+                ir::Node *node, std::unordered_set<ir::Node *> *visited,
+                std::vector<ir::Node *> *ret) {
   visited->insert(node);
 
   for (auto adj : adj_list.at(node)) {
@@ -50,7 +50,8 @@ void SortHelper(
 
 bool HasCircleHelper(
     ir::Node *node,
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
+    const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+        &adj_list,
     std::unordered_set<ir::Node *> *visited,
     std::unordered_set<ir::Node *> *in_trace,
     std::vector<std::vector<ir::Node *>> *circles) {
@@ -84,7 +85,8 @@ bool HasCircleHelper(
 }
 
 bool HasCircleInternal(
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
+    const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+        &adj_list,
     std::vector<std::vector<ir::Node *>> *circles) {
   std::unordered_set<ir::Node *> visited;
   std::unordered_set<ir::Node *> in_trace;
@@ -107,8 +109,8 @@ bool FindCircleSubGraph(const Graph &graph,
 }
 
 std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
-  std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list =
-      BuildOperationAdjList(graph);
+  std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+      adj_list = BuildOperationAdjList(graph);
   PADDLE_ENFORCE(!HasCircleInternal(adj_list, nullptr));
   std::unordered_set<ir::Node *> visited;
   std::vector<ir::Node *> ret;
@@ -117,34 +119,30 @@ std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
       SortHelper(adj_list, adj.first, &visited, &ret);
     }
   }
+
   return ret;
 }
 
 // Build operator inlink edge table.
-std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
-    const Graph &graph) {
-  std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list;
+std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+BuildOperationAdjList(const Graph &graph) {
+  std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+      adj_list;
 
   for (auto &n : graph.Nodes()) {
     if (!n->IsOp()) continue;
     if (adj_list.find(n) == adj_list.end()) {
-      adj_list[n] = std::unordered_set<ir::Node *>();
+      adj_list[n] = std::set<ir::Node *, ir::NodeComp>();
     }
-    std::vector<ir::Node *> nodes;
     for (auto &var : n->inputs) {
       for (auto &adj_n : var->inputs) {
         PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
         VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
                 << " -> " << n->Name() << reinterpret_cast<void *>(n)
                 << "  via " << var->Name() << reinterpret_cast<void *>(var);
-        nodes.push_back(adj_n);
+        adj_list[n].insert(adj_n);
       }
     }
-    std::sort(nodes.begin(), nodes.end(), [](ir::Node *node1, ir::Node *node2) {
-      return node1->id() > node2->id();
-    });
-    adj_list[n].insert(std::make_move_iterator(nodes.begin()),
-                       std::make_move_iterator(nodes.end()));
   }
   return adj_list;
 }

paddle/fluid/framework/ir/graph_helper.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <map>
 #include <memory>
+#include <set>
 #include <vector>
 
 #include "paddle/fluid/framework/ir/graph.h"
@@ -25,6 +26,13 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
+// Compare nodes via node id.
+struct NodeComp {
+  bool operator()(ir::Node *const &node1, ir::Node *const &node2) const {
+    return node1->id() < node2->id();
+  }
+};
+
 // Test if the graph contains circle.
 bool HasCircle(const Graph &graph);
 
@@ -57,8 +65,8 @@ std::vector<Node *> TopologyVarientSort(const Graph &graph, SortKind sort_kind);
 void CleanIndividualNodes(Graph *graph);
 
 // Build an adjacency list of operations for the `graph`.
-std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
-    const Graph &graph);
+std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+BuildOperationAdjList(const Graph &graph);
 
 template <typename T>
 std::vector<T *> FilterByNodeWrapper(const Graph &graph) {

paddle/fluid/framework/ir/runtime_context_cache_pass.cc

@@ -23,7 +23,7 @@ namespace ir {
 void RuntimeContextCachePass::ApplyImpl(ir::Graph* graph) const {
   VLOG(3) << "Applies Runtime Context Cache strategy.";
   for (const Node* n : graph->Nodes()) {
-    if (n->IsOp()) {
+    if (n->IsOp() && n->Op()) {
       n->Op()->SetAttr(kEnableCacheRuntimeContext, true);
     }
   }

paddle/fluid/framework/op_desc.cc

@@ -241,6 +241,7 @@ OpDesc::OpDesc(const std::string &type, const VariableNameMap &inputs,
   outputs_ = outputs;
   attrs_ = attrs;
   need_update_ = true;
+  block_ = nullptr;
 }
 
 OpDesc::OpDesc(const OpDesc &other, BlockDesc *block) {

paddle/fluid/framework/operator.cc

@@ -880,7 +880,16 @@ std::vector<KernelConfig>* OperatorWithKernel::GetKernelConfig(
 
 void OperatorWithKernel::RunImpl(const Scope& scope,
                                  const platform::Place& place) const {
-  if (!HasAttr(kEnableCacheRuntimeContext)) {
+  // To reduce the elapsed time of HasAttr, we use bool variable to record the
+  // result of HasAttr.
+  if (!enable_cache_runtime_context && HasAttr(kEnableCacheRuntimeContext))
+    enable_cache_runtime_context = true;
+  if (!enable_cache_expected_kernel && HasAttr(kEnableCacheExpectedKernel))
+    enable_cache_expected_kernel = true;
+  if (!all_kernels_must_compute_runtime_shape &&
+      HasAttr(kAllKernelsMustComputeRuntimeShape))
+    all_kernels_must_compute_runtime_shape = true;
+  if (!enable_cache_runtime_context) {
     RuntimeContext ctx(Inputs(), Outputs(), scope);
     RunImpl(scope, place, &ctx);
   } else {
@@ -899,60 +908,33 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto* dev_ctx = pool.Get(place);
 
-  // check if op[type] has kernel registered.
-  auto& all_op_kernels = AllOpKernels();
-  auto kernels_iter = all_op_kernels.find(type_);
-  if (kernels_iter == all_op_kernels.end()) {
-    PADDLE_THROW(
-        "There are no kernels which are registered in the %s operator.", type_);
+  if (!enable_cache_expected_kernel || !kernel_type_) {
+    ChooseKernel(*runtime_ctx, scope, place);
   }
 
-  OpKernelMap& kernels = kernels_iter->second;
-
-  auto expected_kernel_key = this->GetExpectedKernelType(
-      ExecutionContext(*this, scope, *dev_ctx, *runtime_ctx, nullptr));
-  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(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);
-  }
-#endif
-  if (kernel_iter == kernels.end()) {
-    PADDLE_THROW("op %s does not have kernel for %s", type_,
-                 KernelTypeToString(expected_kernel_key));
-  }
-
-  std::vector<KernelConfig>* kernel_configs =
-      GetKernelConfig(expected_kernel_key);
+  std::vector<KernelConfig>* kernel_configs = GetKernelConfig(*kernel_type_);
 
   // do data transformScope &transfer_scope;
   std::vector<std::string> transfered_inplace_vars;
-  auto* transfer_scope = PrepareData(scope, expected_kernel_key,
-                                     &transfered_inplace_vars, runtime_ctx);
+  auto* transfer_scope =
+      PrepareData(scope, *kernel_type_, &transfered_inplace_vars, runtime_ctx);
 
   // exec scope is the scope that kernel actually executed on.
   const Scope& exec_scope =
       (transfer_scope == nullptr ? scope : *transfer_scope);
 
-  if (!(expected_kernel_key.place_ == dev_ctx->GetPlace())) {
-    dev_ctx = pool.Get(expected_kernel_key.place_);
+  if (!(kernel_type_->place_ == dev_ctx->GetPlace())) {
+    dev_ctx = pool.Get(kernel_type_->place_);
   }
 
-  if (!HasAttr(kAllKernelsMustComputeRuntimeShape)) {
+  if (!all_kernels_must_compute_runtime_shape) {
     RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, *runtime_ctx);
     this->InferShape(&infer_shape_ctx);
   }
   // TODO(panyx0718): ExecutionContext should only depend on RuntimeContext
   // not Scope. Imperative mode only pass inputs and get outputs.
-  kernel_iter->second(ExecutionContext(*this, exec_scope, *dev_ctx,
-                                       *runtime_ctx, kernel_configs));
+  (*kernel_func_)(ExecutionContext(*this, exec_scope, *dev_ctx, *runtime_ctx,
+                                   kernel_configs));
 
   if (!transfered_inplace_vars.empty()) {
     // there is inplace variable has been transfered.
@@ -978,6 +960,46 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   }
 }
 
+void OperatorWithKernel::ChooseKernel(const RuntimeContext& ctx,
+                                      const Scope& scope,
+                                      const platform::Place& place) const {
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  auto* dev_ctx = pool.Get(place);
+
+  // check if op[type] has kernel registered.
+  auto& all_op_kernels = AllOpKernels();
+  auto kernels_iter = all_op_kernels.find(type_);
+  if (kernels_iter == all_op_kernels.end()) {
+    PADDLE_THROW(
+        "There are no kernels which are registered in the %s operator.", type_);
+  }
+
+  OpKernelMap& kernels = kernels_iter->second;
+
+  auto expected_kernel_key = this->GetExpectedKernelType(
+      ExecutionContext(*this, scope, *dev_ctx, ctx, nullptr));
+  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(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);
+  }
+#endif
+  if (kernel_iter == kernels.end()) {
+    PADDLE_THROW("op %s does not have kernel for %s", type_,
+                 KernelTypeToString(expected_kernel_key));
+  }
+
+  kernel_type_.reset(new OpKernelType(expected_kernel_key));
+  kernel_func_.reset(new OpKernelFunc(kernel_iter->second));
+}
+
 void OperatorWithKernel::TransferInplaceVarsBack(
     const Scope& scope, const std::vector<std::string>& inplace_vars,
     const Scope& transfer_scope) const {

paddle/fluid/framework/operator.h

@@ -70,6 +70,12 @@ constexpr char kNewGradSuffix[] = "@NEWGRAD@";
 /// this Op's execution to save the elapsed time.
 constexpr char kEnableCacheRuntimeContext[] = "@ENABLE_CACHE_RUNTIME_CONTEXT@";
 
+/// If an Op has attribtue kEnableCacheExpectedKernel, it means that in a same
+/// name scope and same place, since the expected kerenl of this Op does not
+/// change in the execution, it could be recorded only at the first iteration of
+/// this Op's execution to save the elapsed time.
+constexpr char kEnableCacheExpectedKernel[] = "@ENABLE_CACHE_EXPECTED_KERNEL@";
+
 /// If an Op has this attribute, all its kernels should calculate output
 /// variable's shape in the corresponding Compute() function. And
 /// OperatorWithKernel::RunImpl() would skip call this Op's InferShape()
@@ -491,10 +497,18 @@ class OperatorWithKernel : public OperatorBase {
                                const std::vector<std::string>& inplace_vars,
                                const Scope& exec_scope) const;
 
+  void ChooseKernel(const RuntimeContext& ctx, const Scope& scope,
+                    const platform::Place& place) const;
+
  protected:
   mutable OpKernelConfigsMap kernel_configs_map_;
+  mutable std::unique_ptr<OpKernelType> kernel_type_;
+  mutable std::unique_ptr<OpKernelFunc> kernel_func_;
   mutable std::unique_ptr<RuntimeContext> runtime_ctx_;
   mutable const Scope* pre_scope_ = nullptr;
+  mutable bool enable_cache_runtime_context = false;
+  mutable bool enable_cache_expected_kernel = false;
+  mutable bool all_kernels_must_compute_runtime_shape = false;
 };
 
 extern bool OpSupportGPU(const std::string& op_type);

paddle/fluid/framework/parallel_executor.cc

@@ -221,7 +221,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
     PADDLE_ENFORCE(!member_->use_cuda_,
                    "gpu mode does not support async_mode_ now!");
     graphs.push_back(graph);
-    for (int i = 1; i < places.size(); ++i) {
+    for (size_t i = 1; i < places.size(); ++i) {
       auto *tmp_graph = new ir::Graph(graph->OriginProgram());
       async_graphs_.emplace_back(tmp_graph);
       graphs.push_back(tmp_graph);
@@ -315,7 +315,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
     graph = build_strategy.Apply(graph, {member_->places_[0]}, loss_var_name,
                                  {member_->local_scopes_[0]}, 1,
                                  member_->use_cuda_, member_->nccl_ctxs_.get());
-    for (int i = 1; i < member_->places_.size(); ++i) {
+    for (size_t i = 1; i < member_->places_.size(); ++i) {
       graphs[i] =
           build_strategy.Apply(graphs[i], {member_->places_[i]}, loss_var_name,
                                {member_->local_scopes_[i]}, 1,

paddle/fluid/framework/trainer_desc.proto

@@ -76,7 +76,7 @@ message PullDenseWorkerParameter {
 
 message TableParameter {
   // dense table only
-  optional int64 table_id = 1;
+  optional uint64 table_id = 1;
   repeated string dense_value_name = 2;
   repeated string dense_grad_name = 3;
   repeated int32 push_dense_wait_times = 5;

paddle/fluid/framework/var_type_inference.h

@@ -45,12 +45,16 @@ class InferVarTypeContext {
 
   virtual bool HasInput(const std::string& name) const {
     PADDLE_ENFORCE_NOT_NULL(op_);
-    return op_->Inputs().count(name) > 0;
+    auto& inputs = op_->Inputs();
+    auto input = inputs.find(name);
+    return input != inputs.end() && !input->second.empty();
   }
 
   virtual bool HasOutput(const std::string& name) const {
     PADDLE_ENFORCE_NOT_NULL(op_);
-    return op_->Outputs().count(name) > 0;
+    auto& outputs = op_->Outputs();
+    auto output = outputs.find(name);
+    return output != outputs.end() && !output->second.empty();
   }
 
   virtual const std::vector<std::string>& Input(const std::string& name) const {

paddle/fluid/imperative/CMakeLists.txt

@@ -3,4 +3,7 @@ cc_library(layer SRCS layer.cc DEPS proto_desc operator device_context blas pybi
 cc_library(tracer SRCS tracer.cc DEPS proto_desc device_context pybind)
 cc_library(engine SRCS engine.cc)
 cc_library(imperative_profiler SRCS profiler.cc)
+cc_library(nccl_context SRCS nccl_context.cc DEPS device_context)
+
+cc_test(nccl_context_test SRCS nccl_context_test.cc  DEPS nccl_context)
 endif()

paddle/fluid/imperative/nccl_context.cc

+//   Copyright (c) 2019 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/imperative/nccl_context.h"
+
+namespace paddle {
+namespace imperative {
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+void NCCLParallelContext::RecvNCCLID(const std::string &ep,
+                                     ncclUniqueId *nccl_id) {
+  auto addr = paddle::string::Split(ep, ':');
+  PADDLE_ENFORCE_EQ(addr.size(), 2UL,
+                    "The endpoint should contain host and port: %s", ep);
+  std::string host = addr[0];
+  int port = std::stoi(addr[1]);
+
+  int server_fd, new_socket;
+  struct sockaddr_in address;
+  int addrlen = sizeof(address);
+  char buffer[1024] = {0};
+  int opt = 0;
+  // creating socket fd
+  if ((server_fd = socket(AF_INET, SOCK_STREAM, 0)) == 0)
+    PADDLE_THROW("create server fd failed");
+  if (setsockopt(server_fd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)))
+    PADDLE_THROW("set socket opt failed");
+
+  address.sin_family = AF_INET;
+  address.sin_addr.s_addr = INADDR_ANY;
+  address.sin_port = htons(port);
+
+  if (bind(server_fd, (struct sockaddr *)&address, sizeof(address)) < 0)
+    PADDLE_THROW("binding failed on ep: %s", ep);
+  VLOG(3) << "listening on: " << ep;
+  if (listen(server_fd, 3) < 0) PADDLE_THROW("listen on server fd failed");
+
+  if ((new_socket =
+           accept(server_fd, reinterpret_cast<struct sockaddr *>(&address),
+                  reinterpret_cast<socklen_t *>(&addrlen))) < 0)
+    PADDLE_THROW("accept the new socket fd failed");
+
+  if (read(new_socket, buffer, 1024) < 0)
+    PADDLE_THROW("reading the ncclUniqueId from socket failed");
+  VLOG(3) << "recevived the ncclUniqueId";
+  memcpy(nccl_id, buffer, NCCL_UNIQUE_ID_BYTES);
+
+  VLOG(3) << "closing the socket server: " << ep;
+  close(server_fd);
+}
+
+void NCCLParallelContext::SendNCCLID(const std::string &ep,
+                                     ncclUniqueId *nccl_id) {
+  auto addr = paddle::string::Split(ep, ':');
+  PADDLE_ENFORCE_EQ(addr.size(), 2UL,
+                    "The endpoint should contain host and port: %s", ep);
+  std::string host = addr[0];
+  int port = std::stoi(addr[1]);
+  // struct sockaddr_in address;
+  int sock = 0;
+  struct sockaddr_in serv_addr;
+  char buffer[1024] = {0};
+
+  memcpy(buffer, nccl_id, NCCL_UNIQUE_ID_BYTES);
+  if ((sock = socket(AF_INET, SOCK_STREAM, 0)) < 0)
+    PADDLE_THROW("create socket failed");
+
+  memset(&serv_addr, '0', sizeof(serv_addr));
+  serv_addr.sin_family = AF_INET;
+  serv_addr.sin_port = htons(port);
+
+  if (inet_pton(AF_INET, host.c_str(), &serv_addr.sin_addr) <= 0)
+    PADDLE_THROW("invalied address: %s", ep);
+
+  while (true) {
+    if (connect(sock, (struct sockaddr *)&serv_addr, sizeof(serv_addr)) < 0) {
+      VLOG(0) << "worker: " << ep
+              << " is not ready, will retry after 3 seconds...";
+      std::this_thread::sleep_for(std::chrono::seconds(3));
+      continue;
+    }
+    VLOG(3) << "sending the ncclUniqueId to " << ep;
+    send(sock, buffer, NCCL_UNIQUE_ID_BYTES, 0);
+    break;
+  }
+}
+
+void NCCLParallelContext::BcastNCCLId(ncclUniqueId *nccl_id, int root) {
+  if (strategy_.local_rank_ == root) {
+    for (auto ep : strategy_.trainer_endpoints_) {
+      if (ep != strategy_.current_endpoint_) SendNCCLID(ep, nccl_id);
+    }
+  } else {
+    RecvNCCLID(strategy_.current_endpoint_, nccl_id);
+  }
+}
+
+void NCCLParallelContext::Init() {
+  ncclUniqueId nccl_id;
+  ncclComm_t comm;
+  if (strategy_.local_rank_ == 0) {
+    // generate the unique ncclid on the root worker
+    platform::dynload::ncclGetUniqueId(&nccl_id);
+    BcastNCCLId(&nccl_id, 0);
+  } else {
+    BcastNCCLId(&nccl_id, 0);
+  }
+  int gpu_id = boost::get<platform::CUDAPlace>(place_).device;
+  VLOG(0) << "init nccl context nranks: " << strategy_.nranks_
+          << " local rank: " << strategy_.local_rank_ << " gpu id: " << gpu_id;
+
+  PADDLE_ENFORCE(cudaSetDevice(gpu_id));
+  PADDLE_ENFORCE(platform::dynload::ncclCommInitRank(
+      &comm, strategy_.nranks_, nccl_id, strategy_.local_rank_));
+
+  platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+  auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(pool.Get(place_));
+  dev_ctx->set_nccl_comm(comm);
+}
+#endif
+
+}  //  namespace imperative
+}  //  namespace paddle

paddle/fluid/imperative/nccl_context.h

+//   Copyright (c) 2019 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
+
+// network header files
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#include <arpa/inet.h>
+#include <netinet/in.h>
+#include <stdlib.h>
+#include <sys/socket.h>
+#endif
+
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/variable.h"
+#include "paddle/fluid/platform/device_context.h"
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+#include "paddle/fluid/platform/dynload/nccl.h"
+#endif
+#include "paddle/fluid/platform/place.h"
+#include "paddle/fluid/string/split.h"
+
+namespace paddle {
+namespace imperative {
+
+struct ParallelStrategy {
+  int nranks_{1};
+  int local_rank_{0};
+  std::vector<std::string> trainer_endpoints_{};
+  std::string current_endpoint_{""};
+};
+
+class ParallelContext {
+ public:
+  explicit ParallelContext(const ParallelStrategy& strategy,
+                           const platform::Place& place)
+      : strategy_(strategy), place_(place) {}
+
+  virtual ~ParallelContext() {}
+
+  virtual void Init() = 0;
+
+ protected:
+  ParallelStrategy strategy_;
+  platform::Place place_;
+};
+
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+class NCCLParallelContext : ParallelContext {
+ public:
+  explicit NCCLParallelContext(const ParallelStrategy& strategy,
+                               const platform::Place& place)
+      : ParallelContext(strategy, place) {}
+
+  ~NCCLParallelContext() {}
+
+  void BcastNCCLId(ncclUniqueId* nccl_id, int root);
+
+  void Init() override;
+
+ protected:
+  void RecvNCCLID(const std::string& endpoint, ncclUniqueId* nccl_id);
+
+  void SendNCCLID(const std::string& endpoint, ncclUniqueId* nccl_id);
+};
+#endif
+
+}  //  namespace imperative
+}  //  namespace paddle

paddle/fluid/imperative/nccl_context_test.cc

+//   Copyright (c) 2019 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/imperative/nccl_context.h"
+#include "gtest/gtest.h"
+#include "paddle/fluid/platform/device_context.h"
+
+namespace imperative = paddle::imperative;
+namespace platform = paddle::platform;
+
+imperative::ParallelStrategy GetStrategy(int local_rank) {
+  std::vector<std::string> eps = {"127.0.0.1:9866", "127.0.0.1:9867"};
+  imperative::ParallelStrategy strategy;
+  strategy.trainer_endpoints_ = eps;
+  strategy.current_endpoint_ = eps[local_rank];
+  strategy.nranks_ = 2;
+  strategy.local_rank_ = local_rank;
+  return strategy;
+}
+
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+void BcastNCCLId(int local_rank, ncclUniqueId *nccl_id) {
+  auto strategy = GetStrategy(local_rank);
+  platform::CUDAPlace gpu(local_rank);
+  imperative::NCCLParallelContext ctx(strategy, gpu);
+  ctx.BcastNCCLId(nccl_id, 0);
+}
+
+TEST(BcastNCCLId, Run) {
+  ncclUniqueId nccl_id;
+  platform::dynload::ncclGetUniqueId(&nccl_id);
+  std::thread t(BcastNCCLId, 0, &nccl_id);
+
+  ncclUniqueId recv_nccl_id;
+  BcastNCCLId(1, &recv_nccl_id);
+
+  t.join();
+  EXPECT_EQ(0, std::memcmp(nccl_id.internal, recv_nccl_id.internal,
+                           NCCL_UNIQUE_ID_BYTES));
+}
+#endif

paddle/fluid/imperative/tracer.cc

@@ -177,7 +177,7 @@ std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
         current_vars_map[out->Name()] = out;
       }
 
-      VLOG(3) << "input var name: " << out->Name()
+      VLOG(3) << "output var name: " << out->Name()
               << " inited: " << out->var_->IsInitialized()
               << " stop_grad: " << out->IsStopGradient();
     }
@@ -215,6 +215,7 @@ std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
 
   framework::Scope scope;
   op->place_ = GetExpectedPlace(expected_place, inputs);
+
   PreparedOp prepared_op = PreparedOp::Prepare(ctx, *op_kernel, op->place_);
   prepared_op.op.RuntimeInferShape(scope, op->place_, ctx);
   prepared_op.func(

paddle/fluid/inference/api/analysis_config.cc

@@ -231,6 +231,7 @@ void AnalysisConfig::Update() {
       pass_builder()->InsertPass(3, "tensorrt_subgraph_pass");
     }
     pass_builder()->DeletePass("runtime_context_cache_pass");
+    pass_builder()->DeletePass("expected_kernel_cache_pass");
   }
 
   if (use_mkldnn_) {

paddle/fluid/inference/api/analysis_predictor.cc

@@ -259,6 +259,9 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
       return false;
     }
 
+    PADDLE_ENFORCE_NOT_NULL(input_ptr);
+    PADDLE_ENFORCE_NOT_NULL(inputs[i].data.data());
+
     if (platform::is_cpu_place(place_)) {
       // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
       std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
@@ -829,6 +832,45 @@ std::string AnalysisPredictor::GetSerializedProgram() const {
   return inference_program_->Proto()->SerializeAsString();
 }
 
+// Add SaveOptimModel
+void AnalysisPredictor::SaveOptimModel(const std::string &dir) {
+  // save model
+  std::string model_name = dir + "/model";
+  std::ofstream outfile;
+  outfile.open(model_name, std::ios::out | std::ios::binary);
+  std::string inference_prog_desc = GetSerializedProgram();
+  outfile << inference_prog_desc;
+  // save params
+  framework::ProgramDesc save_program;
+  auto *save_block = save_program.MutableBlock(0);
+
+  const framework::ProgramDesc &main_program = program();
+  const framework::BlockDesc &global_block = main_program.Block(0);
+  std::vector<std::string> save_var_list;
+  for (framework::VarDesc *var : global_block.AllVars()) {
+    if (IsPersistable(var)) {
+      framework::VarDesc *new_var = save_block->Var(var->Name());
+      new_var->SetShape(var->GetShape());
+      new_var->SetDataType(var->GetDataType());
+      new_var->SetType(var->GetType());
+      new_var->SetLoDLevel(var->GetLoDLevel());
+      new_var->SetPersistable(true);
+
+      save_var_list.push_back(new_var->Name());
+    }
+  }
+  std::sort(save_var_list.begin(), save_var_list.end());
+  auto *op = save_block->AppendOp();
+  op->SetType("save_combine");
+  op->SetInput("X", save_var_list);
+  op->SetAttr("file_path", dir + "/params");
+  op->CheckAttrs();
+
+  platform::CPUPlace place;
+  framework::Executor exe(place);
+  exe.Run(save_program, scope(), 0, true, true);
+}
+
 template <>
 std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<AnalysisConfig>(
     const AnalysisConfig &config) {

paddle/fluid/inference/api/analysis_predictor.h

@@ -86,6 +86,10 @@ class AnalysisPredictor : public PaddlePredictor {
 
   bool MkldnnQuantize();
 
+  // save program to  model
+  // save parameters to params
+  void SaveOptimModel(const std::string &dir);
+
  protected:
   // For memory optimization.
   bool need_collect_var_shapes_for_memory_optim();

paddle/fluid/inference/api/api.cc

@@ -54,6 +54,7 @@ PaddleBuf &PaddleBuf::operator=(const PaddleBuf &other) {
     memory_owned_ = other.memory_owned_;
   } else {
     Resize(other.length());
+    PADDLE_ENFORCE(!(other.length() > 0 && other.data() == nullptr));
     memcpy(data_, other.data(), other.length());
     length_ = other.length();
     memory_owned_ = true;

paddle/fluid/inference/api/api_impl.cc

@@ -169,6 +169,7 @@ std::unique_ptr<PaddlePredictor> NativePaddlePredictor::Clone() {
   std::unique_ptr<PaddlePredictor> cls(new NativePaddlePredictor(config_));
   // Hot fix the bug that result diff in multi-thread.
   // TODO(Superjomn) re-implement a real clone here.
+  PADDLE_ENFORCE_NOT_NULL(dynamic_cast<NativePaddlePredictor *>(cls.get()));
   if (!dynamic_cast<NativePaddlePredictor *>(cls.get())->Init(nullptr)) {
     LOG(ERROR) << "fail to call Init";
     return nullptr;
@@ -210,6 +211,8 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
       return false;
     }
 
+    PADDLE_ENFORCE_NOT_NULL(input_ptr);
+    PADDLE_ENFORCE_NOT_NULL(inputs[i].data.data());
     if (platform::is_cpu_place(place_)) {
       // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
       std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
@@ -316,6 +319,8 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
   }
 
   std::unique_ptr<PaddlePredictor> predictor(new NativePaddlePredictor(config));
+  PADDLE_ENFORCE_NOT_NULL(
+      dynamic_cast<NativePaddlePredictor *>(predictor.get()));
   if (!dynamic_cast<NativePaddlePredictor *>(predictor.get())->Init(nullptr)) {
     return nullptr;
   }

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -86,7 +86,8 @@ const std::vector<std::string> kAnakinSubgraphPasses({
 
 GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
   passes_.assign({
-    "infer_clean_graph_pass",  //
+    "infer_clean_graph_pass",          //
+        "runtime_context_cache_pass",  //
         //   "identity_scale_op_clean_pass",              //
         "conv_affine_channel_fuse_pass",             //
         "conv_eltwiseadd_affine_channel_fuse_pass",  //
@@ -96,9 +97,9 @@ GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
         "conv_elementwise_add_act_fuse_pass",   //
         "conv_elementwise_add2_act_fuse_pass",  //
         "conv_elementwise_add_fuse_pass",       //
-        "runtime_context_cache_pass",           //
 #endif                                          //
         "transpose_flatten_concat_fuse_pass",
+        "expected_kernel_cache_pass",  //
   });
 
   use_gpu_ = true;
@@ -116,10 +117,14 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
   // NOTE the large fusions should be located in the front, so that they will
   // not be damaged by smaller ones.
   passes_.assign({
-      "infer_clean_graph_pass",         //
+      "infer_clean_graph_pass",  //
+      // TODO(luotao): runtime_context_cache_pass should be located in the
+      // front, see https://github.com/PaddlePaddle/Paddle/issues/16609,
+      // will enhance this pass later.
+      "runtime_context_cache_pass",     //
       "attention_lstm_fuse_pass",       //
-      "seqpool_concat_fuse_pass",       //
       "seqconv_eltadd_relu_fuse_pass",  //
+      // "seqpool_concat_fuse_pass",    //
       // "embedding_fc_lstm_fuse_pass", //
       "fc_lstm_fuse_pass",             //
       "mul_lstm_fuse_pass",            //
@@ -132,7 +137,7 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
       "conv_bn_fuse_pass",             //
       "conv_eltwiseadd_bn_fuse_pass",  //
       "is_test_pass",                  //
-      "runtime_context_cache_pass",    //
+      "expected_kernel_cache_pass",    //
   });
 
   use_gpu_ = false;

paddle/fluid/inference/tests/api/analyzer_dam_tester.cc

@@ -170,6 +170,15 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->SwitchIrOptim(true);
 }
 
+void SetOptimConfig(AnalysisConfig *cfg) {
+  std::string optimModelPath =
+      FLAGS_infer_model.substr(0, FLAGS_infer_model.find_last_of("/")) +
+      "/saved_optim_model";
+  cfg->SetModel(optimModelPath + "/model", optimModelPath + "/params");
+  cfg->SwitchIrOptim(true);
+  cfg->SwitchSpecifyInputNames();
+}
+
 void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
   DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
   std::vector<PaddleTensor> input_slots;
@@ -315,5 +324,44 @@ TEST(Analyzer_dam, compare_determine) {
                        input_slots_all);
 }
 
+// Save optim model
+TEST(Analyzer_dam, save_optim_model) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  std::string optimModelPath =
+      FLAGS_infer_model.substr(0, FLAGS_infer_model.find_last_of("/")) +
+      "/saved_optim_model";
+  mkdir(optimModelPath.c_str(), 0777);
+  auto predictor = CreateTestPredictor(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+      FLAGS_use_analysis);
+  (static_cast<AnalysisPredictor *>(predictor.get()))
+      ->SaveOptimModel(optimModelPath);
+}
+
+void CompareOptimAndOrig(const PaddlePredictor::Config *orig_config,
+                         const PaddlePredictor::Config *optim_config,
+                         const std::vector<std::vector<PaddleTensor>> &inputs) {
+  PrintConfig(orig_config, true);
+  PrintConfig(optim_config, true);
+  std::vector<std::vector<PaddleTensor>> orig_outputs, optim_outputs;
+  TestOneThreadPrediction(orig_config, inputs, &orig_outputs, false);
+  TestOneThreadPrediction(optim_config, inputs, &optim_outputs, false);
+  CompareResult(orig_outputs.back(), optim_outputs.back());
+}
+
+TEST(Analyzer_dam, compare_optim_orig) {
+  AnalysisConfig orig_cfg;
+  AnalysisConfig optim_cfg;
+  SetConfig(&orig_cfg);
+  SetOptimConfig(&optim_cfg);
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareOptimAndOrig(
+      reinterpret_cast<const PaddlePredictor::Config *>(&orig_cfg),
+      reinterpret_cast<const PaddlePredictor::Config *>(&optim_cfg),
+      input_slots_all);
+}
+
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tests/api/analyzer_int8_image_classification_tester.cc

@@ -23,18 +23,11 @@ namespace analysis {
 
 void SetConfig(AnalysisConfig *cfg) {
   cfg->SetModel(FLAGS_infer_model);
-  cfg->SetProgFile("__model__");
   cfg->DisableGpu();
   cfg->SwitchIrOptim();
-  cfg->SwitchSpecifyInputNames(false);
+  cfg->SwitchSpecifyInputNames();
   cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
   cfg->EnableMKLDNN();
-  cfg->pass_builder()->SetPasses(
-      {"infer_clean_graph_pass", "mkldnn_placement_pass",
-       "depthwise_conv_mkldnn_pass", "conv_bn_fuse_pass",
-       "conv_eltwiseadd_bn_fuse_pass", "conv_bias_mkldnn_fuse_pass",
-       "conv_elementwise_add_mkldnn_fuse_pass", "conv_relu_mkldnn_fuse_pass",
-       "fc_fuse_pass", "is_test_pass"});
 }
 
 template <typename T>
@@ -84,13 +77,13 @@ std::shared_ptr<std::vector<PaddleTensor>> GetWarmupData(
           std::to_string(num_images) + " is bigger than all test data size.");
 
   PaddleTensor images;
-  images.name = "input";
+  images.name = "image";
   images.shape = {num_images, 3, 224, 224};
   images.dtype = PaddleDType::FLOAT32;
   images.data.Resize(sizeof(float) * num_images * 3 * 224 * 224);
 
   PaddleTensor labels;
-  labels.name = "labels";
+  labels.name = "label";
   labels.shape = {num_images, 1};
   labels.dtype = PaddleDType::INT64;
   labels.data.Resize(sizeof(int64_t) * num_images);
@@ -132,7 +125,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs,
       images_offset_in_file + sizeof(float) * total_images * 3 * 224 * 224;
 
   TensorReader<float> image_reader(file, images_offset_in_file,
-                                   image_batch_shape, "input");
+                                   image_batch_shape, "image");
   TensorReader<int64_t> label_reader(file, labels_offset_in_file,
                                      label_batch_shape, "label");
 

paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc

@@ -32,6 +32,17 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
   SetFakeImageInput(inputs, FLAGS_infer_model);
 }
 
+void SetOptimConfig(AnalysisConfig *cfg) {
+  std::string optimModelPath =
+      FLAGS_infer_model.substr(0, FLAGS_infer_model.find_last_of("/")) +
+      "/saved_optim_model";
+  cfg->SetModel(optimModelPath + "/model", optimModelPath + "/params");
+  cfg->DisableGpu();
+  cfg->SwitchIrOptim();
+  cfg->SwitchSpecifyInputNames();
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+}
+
 // Easy for profiling independently.
 void profile(bool use_mkldnn = false) {
   AnalysisConfig cfg;
@@ -87,13 +98,51 @@ TEST(Analyzer_resnet50, compare_mkldnn) { compare(true /* use_mkldnn */); }
 TEST(Analyzer_resnet50, compare_determine) {
   AnalysisConfig cfg;
   SetConfig(&cfg);
-
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
   CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
                        input_slots_all);
 }
 
+// Save optim model
+TEST(Analyzer_resnet50, save_optim_model) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  std::string optimModelPath =
+      FLAGS_infer_model.substr(0, FLAGS_infer_model.find_last_of("/")) +
+      "/saved_optim_model";
+  mkdir(optimModelPath.c_str(), 0777);
+  auto predictor = CreateTestPredictor(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+      FLAGS_use_analysis);
+  (static_cast<AnalysisPredictor *>(predictor.get()))
+      ->SaveOptimModel(optimModelPath);
+}
+
+void CompareOptimAndOrig(const PaddlePredictor::Config *orig_config,
+                         const PaddlePredictor::Config *optim_config,
+                         const std::vector<std::vector<PaddleTensor>> &inputs) {
+  PrintConfig(orig_config, true);
+  PrintConfig(optim_config, true);
+  std::vector<std::vector<PaddleTensor>> orig_outputs, optim_outputs;
+  TestOneThreadPrediction(orig_config, inputs, &orig_outputs, false);
+  TestOneThreadPrediction(optim_config, inputs, &optim_outputs, false);
+  CompareResult(orig_outputs.back(), optim_outputs.back());
+}
+
+TEST(Analyzer_resnet50, compare_optim_orig) {
+  AnalysisConfig orig_cfg;
+  AnalysisConfig optim_cfg;
+  SetConfig(&orig_cfg);
+  SetOptimConfig(&optim_cfg);
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareOptimAndOrig(
+      reinterpret_cast<const PaddlePredictor::Config *>(&orig_cfg),
+      reinterpret_cast<const PaddlePredictor::Config *>(&optim_cfg),
+      input_slots_all);
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc

@@ -47,6 +47,7 @@ struct DataRecord {
       num_lines++;
       std::vector<std::string> data;
       split(line, '\t', &data);
+      PADDLE_ENFORCE(data.size() >= 4);
       // load title1 data
       std::vector<int64_t> title1_data;
       split_to_int64(data[0], ' ', &title1_data);

paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc

@@ -150,6 +150,9 @@ void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false) {
   if (use_mkldnn) {
     cfg->EnableMKLDNN();
   }
+  // Enable seqpool_concat_fuse_pass, disabled by default since it takes much
+  // time
+  cfg->pass_builder()->InsertPass(2, "seqpool_concat_fuse_pass");
 }
 
 void profile(bool use_mkldnn = false) {

paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc

@@ -214,28 +214,23 @@ TEST(Analyzer_Transformer, fuse_statis) {
 }
 
 // Compare result of NativeConfig and AnalysisConfig
-// void compare(bool use_mkldnn = false) {
-//   AnalysisConfig cfg;
-//   SetConfig(&cfg);
-//   if (use_mkldnn) {
-//     cfg.EnableMKLDNN();
-//   }
-//
-//   std::vector<std::vector<PaddleTensor>> input_slots_all;
-//   SetInput(&input_slots_all);
-//   CompareNativeAndAnalysis(
-//       reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
-//       input_slots_all);
-// }
-
-// TODO(yihuaxu):
-//    Disable compare and compare_mkldnn temporary, see
-//    https://github.com/paddlePaddle/Paddle/issues/16316 for details.
-// TEST(Analyzer_Transformer, compare) { compare(); }
-// #ifdef PADDLE_WITH_MKLDNN
-// TEST(Analyzer_Transformer, compare_mkldnn) { compare(true /* use_mkldnn */);
-// }
-// #endif
+void compare(bool use_mkldnn = false) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  if (use_mkldnn) {
+    cfg.EnableMKLDNN();
+  }
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
+}
+
+TEST(Analyzer_Transformer, compare) { compare(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_Transformer, compare_mkldnn) { compare(true /* use_mkldnn */); }
+#endif
 
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tests/api/int8_mkldnn_quantization.md

+# INT8 MKL-DNN quantization 
+
+This document describes how to use Paddle inference Engine to convert the FP32 model to INT8 model on ResNet-50 and MobileNet-V1. We provide the instructions on enabling INT8 MKL-DNN quantization in Paddle inference and show the ResNet-50 and MobileNet-V1 results in accuracy and performance.
+
+## 0. Install PaddlePaddle 
+Follow PaddlePaddle [installation instruction](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification#installation) to install PaddlePaddle. If you build PaddlePaddle yourself, please use the following cmake arguments. 
+```
+cmake ..  -DWITH_TESTING=ON -WITH_FLUID_ONLY=ON -DWITH_GPU=OFF -DWITH_MKL=ON  -WITH_SWIG_PY=OFF -DWITH_INFERENCE_API_TEST=ON -DON_INFER=ON
+
+```  
+Note: MKL-DNN and MKL are required.
+
+## 1. Enable INT8 MKL-DNN quantization 
+For reference, please examine the code of unit test enclosed in [analyzer_int8_image_classification_tester.cc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/api/analyzer_int8_image_classification_tester.cc).
+
+* ### Create Analysis config
+INT8 quantization is one of the optimizations in analysis config. More information about analysis config can be found [here](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/advanced_usage/deploy/inference/native_infer_en.md#upgrade-performance-based-on-contribanalysisconfig-prerelease) 
+
+* ### Create quantize config by analysis config
+We enable the MKL-DNN quantization procedure by calling an appropriate method from analysis config. Afterwards, all the required quantization parameters (quantization op names, quantization strategies etc.) can be set through quantizer config which is present in the analysis config. It is also necessary to specify a pre-processed warmup dataset and desired batch size.
+
+```cpp
+//Enable MKL-DNN quantization
+cfg.EnableMkldnnQuantizer();
+
+//use analysis config to call the MKL-DNN quantization config
+cfg.mkldnn_quantizer_config()->SetWarmupData(warmup_data); 
+cfg.mkldnn_quantizer_config()->SetWarmupBatchSize(100);
+```
+
+## 2. Accuracy and Performance benchmark
+
+We provide the results of accuracy and performance measured on Intel(R) Xeon(R) Gold 6271 on single core.
+
+   >**I. Top-1 Accuracy on Intel(R) Xeon(R) Gold 6271**
+
+| Model  | Dataset  | FP32 Accuracy  | INT8 Accuracy  | Accuracy Diff  |
+| :------------: | :------------: | :------------: | :------------: | :------------: |
+| ResNet-50  | Full ImageNet Val  | 76.63%  | 76.48%  | 0.15% |
+| MobileNet-V1 | Full ImageNet Val  | 70.78%  | 70.36%  | 0.42%  |
+
+   >**II. Throughput on Intel(R) Xeon(R) Gold 6271 (batch size 1 on single core)**
+
+| Model  | Dataset  | FP32 Throughput  | INT8 Throughput  |  Ratio(INT8/FP32)  |
+| :------------: | :------------: | :------------: | :------------: | :------------: |
+| ResNet-50  | Full ImageNet Val  |  13.17 images/s | 49.84 images/s | 3.78 |
+| MobileNet-V1 | Full ImageNet Val  | 75.49 images/s | 232.38 images/s | 3.07  |
+
+Notes:
+* Measurement of accuracy requires a model which accepts two inputs: data and labels.
+* Different sampling batch size data may cause slight difference on INT8 top accuracy.
+* CAPI performance data is better than python API performance data because of the python overhead. Especially for the small computational model, python overhead will be more obvious. 
+
+
+## 3. Commands to reproduce the above accuracy and performance benchmark
+* #### Full dataset (Single core)
+   * ##### Download full ImageNet Validation Dataset
+```bash
+cd /PATH/TO/PADDLE/build
+python ../paddle/fluid/inference/tests/api/full_ILSVRC2012_val_preprocess.py
+```
+The converted data binary file is saved by default in ~/.cache/paddle/dataset/int8/download/int8_full_val.bin
+   * ##### ResNet50 Full dataset benchmark
+```bash
+./paddle/fluid/inference/tests/api/test_analyzer_int8_resnet50 --infer_model=third_party/inference_demo/int8v2/resnet50/model --infer_data=/path/to/converted/int8_full_val.bin --batch_size=1 --paddle_num_threads=1
+```
+   * ##### Mobilenet-v1 Full dataset benchmark
+```bash
+./paddle/fluid/inference/tests/api/test_analyzer_int8_mobilenet --infer_model=third_party/inference_demo/int8v2/mobilenet/model --infer_data=/path/to/converted/int8_full_val.bin --batch_size=1 --paddle_num_threads=1
+```

paddle/fluid/inference/tests/api/tester_helper.h

@@ -55,6 +55,9 @@ DEFINE_bool(record_benchmark, false,
 DEFINE_double(accuracy, 1e-3, "Result Accuracy.");
 DEFINE_double(quantized_accuracy, 1e-2, "Result Quantized Accuracy.");
 DEFINE_bool(zero_copy, false, "Use ZeroCopy to speedup Feed/Fetch.");
+DEFINE_bool(warmup, false,
+            "Use warmup to calculate elapsed_time more accurately. "
+            "To reduce CI time, it sets false in default.");
 
 DECLARE_bool(profile);
 DECLARE_int32(paddle_num_threads);
@@ -316,7 +319,8 @@ void PredictionRun(PaddlePredictor *predictor,
                    int num_threads, int tid) {
   int num_times = FLAGS_repeat;
   int iterations = inputs.size();  // process the whole dataset ...
-  if (FLAGS_iterations > 0 && FLAGS_iterations < inputs.size())
+  if (FLAGS_iterations > 0 &&
+      FLAGS_iterations < static_cast<int64_t>(inputs.size()))
     iterations =
         FLAGS_iterations;  // ... unless the number of iterations is set
   outputs->resize(iterations);
@@ -329,14 +333,14 @@ void PredictionRun(PaddlePredictor *predictor,
 #endif
   if (!FLAGS_zero_copy) {
     run_timer.tic();
-    for (size_t i = 0; i < iterations; i++) {
+    for (int i = 0; i < iterations; i++) {
       for (int j = 0; j < num_times; j++) {
         predictor->Run(inputs[i], &(*outputs)[i], FLAGS_batch_size);
       }
     }
     elapsed_time = run_timer.toc();
   } else {
-    for (size_t i = 0; i < iterations; i++) {
+    for (int i = 0; i < iterations; i++) {
       ConvertPaddleTensorToZeroCopyTensor(predictor, inputs[i]);
       run_timer.tic();
       for (int j = 0; j < num_times; j++) {
@@ -366,9 +370,10 @@ void TestOneThreadPrediction(
     const std::vector<std::vector<PaddleTensor>> &inputs,
     std::vector<std::vector<PaddleTensor>> *outputs, bool use_analysis = true) {
   auto predictor = CreateTestPredictor(config, use_analysis);
-  PredictionWarmUp(predictor.get(), inputs, outputs, FLAGS_paddle_num_threads,
-                   0);
-  PredictionRun(predictor.get(), inputs, outputs, FLAGS_paddle_num_threads, 0);
+  if (FLAGS_warmup) {
+    PredictionWarmUp(predictor.get(), inputs, outputs, 1, 0);
+  }
+  PredictionRun(predictor.get(), inputs, outputs, 1, 0);
 }
 
 void TestMultiThreadPrediction(
@@ -395,7 +400,10 @@ void TestMultiThreadPrediction(
             ->SetMkldnnThreadID(static_cast<int>(tid) + 1);
       }
 #endif
-      PredictionWarmUp(predictor.get(), inputs, &outputs_tid, num_threads, tid);
+      if (FLAGS_warmup) {
+        PredictionWarmUp(predictor.get(), inputs, &outputs_tid, num_threads,
+                         tid);
+      }
       PredictionRun(predictor.get(), inputs, &outputs_tid, num_threads, tid);
     });
   }

paddle/fluid/op_use_default_grad_op_maker.spec

-abs
-acos
-asin
-atan
 attention_lstm
-brelu
 conv_shift
-cos
 cos_sim
 dequantize
-elu
 fc
 flatten
 fsp
@@ -21,14 +14,8 @@ fusion_seqconv_eltadd_relu
 fusion_seqexpand_concat_fc
 fusion_seqpool_concat
 fusion_squared_mat_sub
-gelu
 gru
-hard_shrink
 hierarchical_sigmoid
-leaky_relu
-log
-logsigmoid
-lookup_table
 lrn
 lstm_unit
 lstmp
@@ -39,10 +26,11 @@ modified_huber_loss
 nce
 pool2d
 pool3d
-pow
 prelu
 quantize
 rank_loss
+reduce_all
+reduce_any
 reduce_max
 reduce_mean
 reduce_min
@@ -51,26 +39,10 @@ reduce_sum
 requantize
 reshape
 rnn_memory_helper
-round
 sequence_softmax
-sin
-softplus
-softshrink
-softsign
-space_to_depth
 spp
-square
-squared_l2_distance
-squared_l2_norm
 squeeze
-stanh
-swish
-tanh_shrink
-teacher_student_sigmoid_loss
 tensor_array_to_tensor
-thresholded_relu
 transpose
-tree_conv
 unpool
 unsqueeze
-warpctc

paddle/fluid/operators/activation_cudnn_op.cu.cc

@@ -12,6 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <memory>
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/activation_op.h"
 #include "paddle/fluid/platform/cudnn_desc.h"
@@ -82,6 +85,8 @@ template <typename T>
 struct CudnnReluGradFunctor : public CudnnActivationGradFunctor<T> {
   explicit CudnnReluGradFunctor(const CUDADeviceContext& ctx)
       : CudnnActivationGradFunctor<T>(ctx, 0.0, CUDNN_ACTIVATION_RELU) {}
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename T>
@@ -94,6 +99,8 @@ struct CudnnRelu6GradFunctor : public CudnnActivationGradFunctor<T> {
   explicit CudnnRelu6GradFunctor(const CUDADeviceContext& ctx)
       : CudnnActivationGradFunctor<T>(ctx, 6.0, CUDNN_ACTIVATION_CLIPPED_RELU) {
   }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename T>
@@ -105,6 +112,8 @@ template <typename T>
 struct CudnnSigmoidGradFunctor : public CudnnActivationGradFunctor<T> {
   explicit CudnnSigmoidGradFunctor(const CUDADeviceContext& ctx)
       : CudnnActivationGradFunctor<T>(ctx, 0.0, CUDNN_ACTIVATION_SIGMOID) {}
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename T>
@@ -116,6 +125,8 @@ template <typename T>
 struct CudnnTanhGradFunctor : public CudnnActivationGradFunctor<T> {
   explicit CudnnTanhGradFunctor(const CUDADeviceContext& ctx)
       : CudnnActivationGradFunctor<T>(ctx, 0.0, CUDNN_ACTIVATION_TANH) {}
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
 };
 
 template <typename Functor>
@@ -140,10 +151,13 @@ class CudnnActivationGradKernel
  public:
   using T = typename Functor::ELEMENT_TYPE;
   void Compute(const framework::ExecutionContext& context) const override {
+    static_assert(Functor::FwdDeps() == kDepOut, "Forward deps must be Out.");
+
     const framework::Tensor *X, *Out, *dOut;
     X = Out = dOut = nullptr;
     framework::Tensor* dX = nullptr;
-    ExtractActivationGradTensor(context, &X, &Out, &dOut, &dX);
+    ExtractActivationGradTensor<Functor::FwdDeps()>(context, &X, &Out, &dOut,
+                                                    &dX);
     dX->mutable_data<T>(context.GetPlace());
     auto& dev_ctx = context.template device_context<CUDADeviceContext>();
     Functor functor(dev_ctx);

paddle/fluid/operators/activation_op.cc

@@ -15,7 +15,9 @@ limitations under the License. */
 #include "paddle/fluid/operators/activation_op.h"
 #include <memory>
 #include <string>
+#include <type_traits>
 #include <unordered_map>
+#include <vector>
 #include "paddle/fluid/operators/mkldnn/mkldnn_activation_op.h"
 #include "paddle/fluid/platform/port.h"
 #ifdef PADDLE_WITH_CUDA
@@ -27,6 +29,25 @@ namespace operators {
 
 using paddle::framework::Tensor;
 
+template <typename GradFunctor>
+static constexpr bool CanInplaceAct() {
+  return GradFunctor::FwdDeps() == kDepOut || GradFunctor::FwdDeps() == kNoDeps;
+}
+
+std::unique_ptr<std::unordered_set<std::string>> GetInplaceOpSet() {
+  std::unique_ptr<std::unordered_set<std::string>> ret(
+      new std::unordered_set<std::string>());
+#define INSERT_INTO_INPLACE_OP_SET(op_type, __omitted, fwd_functor, \
+                                   bwd_functor)                     \
+  if (CanInplaceAct<bwd_functor<float>>()) {                        \
+    ret->insert(#op_type);                                          \
+  }
+
+  FOR_EACH_ACTIVATION_OP(INSERT_INTO_INPLACE_OP_SET);
+#undef INSERT_INTO_INPLACE_OP_SET
+  return ret;
+}
+
 #define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT)                    \
   class OP_NAME##OpMaker                                                     \
       : public ::paddle::framework::OpProtoAndCheckerMaker {                 \
@@ -50,26 +71,32 @@ using paddle::framework::Tensor;
     }                                                                        \
   }
 
-#define REGISTER_ACTIVATION_OP_GRAD_MAKER(OP_NAME, KERNEL_TYPE)              \
-  class OP_NAME##GradMaker                                                   \
-      : public ::paddle::framework::SingleGradOpDescMaker {                  \
-   public:                                                                   \
-    using ::paddle::framework::SingleGradOpDescMaker::SingleGradOpDescMaker; \
-                                                                             \
-   protected:                                                                \
-    std::unique_ptr<::paddle::framework::OpDesc> Apply() const override {    \
-      auto* op = new ::paddle::framework::OpDesc();                          \
-      op->SetType(#KERNEL_TYPE "_grad");                                     \
-      op->SetInput("Out", Output("Out"));                                    \
-      op->SetInput(::paddle::framework::GradVarName("Out"),                  \
-                   OutputGrad("Out"));                                       \
-                                                                             \
-      op->SetAttrMap(Attrs());                                               \
-                                                                             \
-      op->SetOutput(::paddle::framework::GradVarName("X"), InputGrad("X"));  \
-      return std::unique_ptr<::paddle::framework::OpDesc>(op);               \
-    }                                                                        \
+template <ActBwdOpFwdDeps kDepValue>
+class ActivationGradOpDescMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    std::unique_ptr<framework::OpDesc> op(new framework::OpDesc());
+    op->SetType(ForwardOpType() + "_grad");
+    op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    op->SetAttrMap(Attrs());
+
+    if (static_cast<int>(kDepValue) &
+        static_cast<int>(ActBwdOpFwdDeps::kDepX)) {
+      op->SetInput("X", Input("X"));
+    }
+
+    if (static_cast<int>(kDepValue) &
+        static_cast<int>(ActBwdOpFwdDeps::kDepOut)) {
+      op->SetInput("Out", Output("Out"));
+    }
+
+    return op;
   }
+};
 
 framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx,
                                       const framework::OperatorWithKernel& oper,
@@ -129,14 +156,15 @@ class ActivationOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
-    ctx->ShareDim("Out", framework::GradVarName("X"));
-    ctx->ShareLoD("Out", framework::GradVarName("X"));
+    auto out_grad_name = framework::GradVarName("Out");
+    ctx->ShareDim(out_grad_name, framework::GradVarName("X"));
+    ctx->ShareLoD(out_grad_name, framework::GradVarName("X"));
   }
 
  protected:
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
-    return GetKernelType(ctx, *this, "Out");
+    return GetKernelType(ctx, *this, framework::GradVarName("Out"));
   }
 };
 
@@ -558,79 +586,27 @@ REGISTER_ACTIVATION_OP_MAKER(Log, LogDoc);
 REGISTER_ACTIVATION_OP_MAKER(Square, SquareDoc);
 REGISTER_ACTIVATION_OP_MAKER(Softplus, SoftplusDoc);
 REGISTER_ACTIVATION_OP_MAKER(Softsign, SoftsignDoc);
-
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Sigmoid, sigmoid);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Relu, relu);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Gelu, gelu);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Exp, exp);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Tanh, tanh);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Ceil, ceil);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Floor, floor);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Sqrt, sqrt);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(SoftRelu, soft_relu);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Relu6, relu6);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(Reciprocal, reciprocal);
-REGISTER_ACTIVATION_OP_GRAD_MAKER(HardSigmoid, hard_sigmoid);
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 
-#define FOR_EACH_INPLACE_OP_FUNCTOR(__macro) \
-  __macro(Sigmoid, sigmoid);                 \
-  __macro(Relu, relu);                       \
-  __macro(Exp, exp);                         \
-  __macro(Tanh, tanh);                       \
-  __macro(Ceil, ceil);                       \
-  __macro(Floor, floor);                     \
-  __macro(Sqrt, sqrt);                       \
-  __macro(SoftRelu, soft_relu);              \
-  __macro(Relu6, relu6);                     \
-  __macro(Reciprocal, reciprocal);           \
-  __macro(HardSigmoid, hard_sigmoid);
-
-#define FOR_EACH_OP_FUNCTOR(__macro) \
-  __macro(LogSigmoid, logsigmoid);   \
-  __macro(SoftShrink, softshrink);   \
-  __macro(Abs, abs);                 \
-  __macro(Cos, cos);                 \
-  __macro(Acos, acos);               \
-  __macro(Sin, sin);                 \
-  __macro(Asin, asin);               \
-  __macro(Atan, atan);               \
-  __macro(Round, round);             \
-  __macro(Log, log);                 \
-  __macro(Square, square);           \
-  __macro(Gelu, gelu);               \
-  __macro(BRelu, brelu);             \
-  __macro(Pow, pow);                 \
-  __macro(STanh, stanh);             \
-  __macro(Softplus, softplus);       \
-  __macro(Softsign, softsign);       \
-  __macro(LeakyRelu, leaky_relu);    \
-  __macro(TanhShrink, tanh_shrink);  \
-  __macro(ELU, elu);                 \
-  __macro(HardShrink, hard_shrink);  \
-  __macro(Swish, swish);             \
-  __macro(ThresholdedRelu, thresholded_relu);
-
-#define REGISTER_INPLACE_ACTIVATION_OP(OP_NAME, KERNEL_TYPE)                   \
-  REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp,            \
-                    ::paddle::operators::OP_NAME##OpMaker,                     \
-                    ::paddle::operators::ActivationOpInferVarType,             \
-                    ::paddle::operators::OP_NAME##GradMaker,                   \
-                    ::paddle::framework::SingleOpInplaceInToOut);              \
-  REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad, \
-                    ::paddle::framework::SingleOpInplaceInToOut)
-
-#define REGISTER_ACTIVATION_OP(OP_NAME, KERNEL_TYPE)                    \
-  REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp,     \
-                    ::paddle::operators::OP_NAME##OpMaker,              \
-                    ::paddle::operators::ActivationOpInferVarType,      \
-                    ::paddle::framework::DefaultGradOpDescMaker<true>); \
-  REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad)
-
-#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor)   \
+#define REGISTER_ACTIVATION_OP(KERNEL_TYPE, OP_NAME, functor, grad_functor) \
+  REGISTER_OPERATOR(                                                        \
+      KERNEL_TYPE, ops::ActivationOp, ops::OP_NAME##OpMaker,                \
+      ops::ActivationOpInferVarType,                                        \
+      ops::ActivationGradOpDescMaker<ops::grad_functor<float>::FwdDeps()>,  \
+      std::conditional<ops::CanInplaceAct<ops::grad_functor<float>>(),      \
+                       ::paddle::framework::SingleOpInplaceInToOut,         \
+                       void>::type);                                        \
+  REGISTER_OPERATOR(                                                        \
+      KERNEL_TYPE##_grad, ops::ActivationOpGrad,                            \
+      std::conditional<ops::CanInplaceAct<ops::grad_functor<float>>(),      \
+                       ::paddle::framework::SingleOpInplaceInToOut,         \
+                       void>::type)
+
+#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, op_name, functor,        \
+                                       grad_functor)                      \
   REGISTER_OP_CPU_KERNEL(                                                 \
       act_type, ops::ActivationKernel<paddle::platform::CPUDeviceContext, \
                                       ops::functor<float>>,               \
@@ -643,6 +619,5 @@ namespace ops = paddle::operators;
       ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,       \
                                 ops::grad_functor<double>>);
 
-FOR_EACH_OP_FUNCTOR(REGISTER_ACTIVATION_OP);
-FOR_EACH_INPLACE_OP_FUNCTOR(REGISTER_INPLACE_ACTIVATION_OP);
-FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);
+FOR_EACH_ACTIVATION_OP(REGISTER_ACTIVATION_OP);
+FOR_EACH_ACTIVATION_OP(REGISTER_ACTIVATION_CPU_KERNEL);

paddle/fluid/operators/activation_op.cu

@@ -15,7 +15,8 @@ limitations under the License. */
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
-#define REGISTER_ACTIVATION_CUDA_KERNEL(act_type, functor, grad_functor)    \
+#define REGISTER_ACTIVATION_CUDA_KERNEL(act_type, op_name, functor,         \
+                                        grad_functor)                       \
   REGISTER_OP_CUDA_KERNEL(                                                  \
       act_type,                                                             \
       ops::ActivationKernel<plat::CUDADeviceContext, ops::functor<float>>,  \
@@ -30,4 +31,4 @@ namespace plat = paddle::platform;
       ops::ActivationGradKernel<plat::CUDADeviceContext,                    \
                                 ops::grad_functor<plat::float16>>);
 
-FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL);
+FOR_EACH_ACTIVATION_OP(REGISTER_ACTIVATION_CUDA_KERNEL);

paddle/fluid/operators/affine_grid_op.h

@@ -121,9 +121,11 @@ class AffineGridOpKernel : public framework::OpKernel<T> {
     // TODO(wanghaoshuang): Refine batched matrix multiply
     auto blas = math::GetBlas<DeviceContext, T>(ctx);
     for (int i = 0; i < n; ++i) {
-      Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3});
+      Tensor sliced_grid = grid.Slice(i, i + 1).Resize(
+          {static_cast<int64_t>(h) * static_cast<int64_t>(w), 3});
       Tensor sliced_theta = theta->Slice(i, i + 1).Resize({2, 3});
-      Tensor sliced_out = output->Slice(i, i + 1).Resize({h * w, 2});
+      Tensor sliced_out = output->Slice(i, i + 1).Resize(
+          {static_cast<int64_t>(h) * static_cast<int64_t>(w), 2});
       blas.MatMul(sliced_grid, false, sliced_theta, true, T(1), &sliced_out,
                   T(0));
     }
@@ -161,8 +163,10 @@ class AffineGridGradOpKernel : public framework::OpKernel<T> {
     // TODO(wanghaoshuang): Refine batched matrix multiply
     auto blas = math::GetBlas<DeviceContext, T>(ctx);
     for (int i = 0; i < n; ++i) {
-      Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3});
-      Tensor sliced_out_grad = output_grad->Slice(i, i + 1).Resize({h * w, 2});
+      Tensor sliced_grid = grid.Slice(i, i + 1).Resize(
+          {static_cast<int64_t>(h) * static_cast<int64_t>(w), 3});
+      Tensor sliced_out_grad = output_grad->Slice(i, i + 1).Resize(
+          {static_cast<int64_t>(h) * static_cast<int64_t>(w), 2});
       Tensor sliced_theta_grad = theta_grad->Slice(i, i + 1).Resize({2, 3});
       blas.MatMul(sliced_out_grad, true, sliced_grid, false, T(1),
                   &sliced_theta_grad, T(0));

paddle/fluid/operators/controlflow/conditional_block_op.cc

@@ -12,6 +12,9 @@ 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 <memory>
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/var_type.h"
@@ -174,24 +177,41 @@ class ConditionalBlockGradOp : public ConditionalOp {
 
       framework::Executor exec(dev_place);
       auto *block = Attr<framework::BlockDesc *>("sub_block");
-      exec.Run(*block->Program(), &cur_scope, block->ID(), false);
 
-      AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("Input"),
-                                  Outputs(framework::GradVarName("Input")));
+      const auto &ins = Inputs("Input");
+      const auto &d_ins = Outputs(framework::GradVarName("Input"));
+      const auto &conds = Inputs("Cond");
+      const auto &d_conds = Outputs(framework::GradVarName("Cond"));
+
+      std::vector<std::string> ins_conds_grads;
+      ins_conds_grads.reserve(ins.size() + conds.size());
+      for (auto &in : ins) {
+        ins_conds_grads.emplace_back(framework::GradVarName(in));
+      }
+      for (auto &cond : conds) {
+        ins_conds_grads.emplace_back(framework::GradVarName(cond));
+      }
+
+      exec.Run(*block->Program(), &cur_scope, block->ID(), false, true,
+               ins_conds_grads);
+
+      AssignLocalGradientToGlobal(dev_place, cur_scope, ins_conds_grads.data(),
+                                  ins.size(), d_ins);
 
-      AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("Cond"),
-                                  Outputs(framework::GradVarName("Cond")));
+      AssignLocalGradientToGlobal(dev_place, cur_scope,
+                                  ins_conds_grads.data() + ins.size(),
+                                  conds.size(), d_conds);
     }
   }
 
  private:
   void AssignLocalGradientToGlobal(
       const platform::Place &place, const framework::Scope &cur_scope,
-      const std::vector<std::string> &p_names,
+      const std::string *p_grad_names, size_t p_grad_names_num,
       const std::vector<std::string> &pg_names) const {
-    for (size_t i = 0; i < p_names.size(); ++i) {
+    for (size_t i = 0; i < p_grad_names_num; ++i) {
       auto out_grad_name = pg_names[i];
-      auto in_grad_name = framework::GradVarName(p_names[i]);
+      const auto &in_grad_name = p_grad_names[i];
       auto *in_var = cur_scope.FindVar(in_grad_name);
       if (in_var == nullptr) {
         continue;

paddle/fluid/operators/detection/gpc.cc

@@ -24,6 +24,7 @@
  **/
 
 #include "paddle/fluid/operators/detection/gpc.h"
+#include "paddle/fluid/platform/enforce.h"
 
 namespace gpc {
 
@@ -689,6 +690,7 @@ static bbox *create_contour_bboxes(gpc_polygon *p) {
 
   gpc_malloc<bbox>(box, p->num_contours * sizeof(bbox),
                    const_cast<char *>("Bounding box creation"));
+  PADDLE_ENFORCE_NOT_NULL(box);
 
   /* Construct contour bounding boxes */
   for (c = 0; c < p->num_contours; c++) {
@@ -852,6 +854,7 @@ void gpc_add_contour(gpc_polygon *p, gpc_vertex_list *new_contour, int hole) {
   /* Create an extended hole array */
   gpc_malloc<int>(extended_hole, (p->num_contours + 1) * sizeof(int),
                   const_cast<char *>("contour hole addition"));
+  PADDLE_ENFORCE_NOT_NULL(extended_hole);
 
   /* Create an extended contour array */
   gpc_malloc<gpc_vertex_list>(extended_contour,
@@ -969,6 +972,7 @@ void gpc_polygon_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip,
   /* Build scanbeam table from scanbeam tree */
   gpc_malloc<double>(sbt, sbt_entries * sizeof(double),
                      const_cast<char *>("sbt creation"));
+  PADDLE_ENFORCE_NOT_NULL(sbt);
   build_sbt(&scanbeam, sbt, sbtree);
   scanbeam = 0;
   free_sbtree(&sbtree);
@@ -1604,6 +1608,7 @@ void gpc_tristrip_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip,
   /* Build scanbeam table from scanbeam tree */
   gpc_malloc<double>(sbt, sbt_entries * sizeof(double),
                      const_cast<char *>("sbt creation"));
+  PADDLE_ENFORCE_NOT_NULL(sbt);
   build_sbt(&scanbeam, sbt, sbtree);
   scanbeam = 0;
   free_sbtree(&sbtree);

paddle/fluid/operators/distributed/CMakeLists.txt

@@ -9,6 +9,9 @@ else()
 endif()
 configure_file(send_recv.proto.in ${CMAKE_CURRENT_SOURCE_DIR}/send_recv.proto @ONLY)
 
+cc_library(async_sparse_param_update_recorder SRCS async_sparse_param_update_recorder.cc DEPS enforce simple_threadpool)
+cc_test(async_sparse_param_update_recorder_test SRCS async_sparse_param_update_recorder_test.cc DEPS async_sparse_param_update_recorder)
+
 # FIXME(typhoonzero): use add_subdirectory once we clean the dependency of these files
 set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
 if(WITH_GRPC)
@@ -20,7 +23,7 @@ if(WITH_GRPC)
         collective_client.cc collective_server.cc
         ${GRPC_SRCS}
       PROTO send_recv.proto 
-      DEPS lod_tensor selected_rows_functor memory scope ${GRPC_DEPS})
+      DEPS lod_tensor selected_rows_functor memory scope ${GRPC_DEPS} async_sparse_param_update_recorder)
 
   set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
   set(RPC_DEPS sendrecvop_rpc ${GRPC_DEPS})

paddle/fluid/operators/distributed/async_sparse_param_update_recorder.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/operators/distributed/async_sparse_param_update_recorder.h"
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+
+std::once_flag AsyncSparseParamUpdateRecorder::init_flag_;
+std::unique_ptr<AsyncSparseParamUpdateRecorder>
+    AsyncSparseParamUpdateRecorder::recorder_(nullptr);
+
+}  // namespace distributed
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/distributed/async_sparse_param_update_recorder.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 <functional>
+#include <future>  // NOLINT
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include <ThreadPool.h>
+
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+
+class ConcurrentSet {
+ public:
+  ConcurrentSet() : pool_(new ::ThreadPool(1)) {}
+  ~ConcurrentSet() {}
+
+  std::future<void> Update(const std::vector<int64_t>& rows) {
+    auto task = [this, rows] {
+      if (VLOG_IS_ON(3)) {
+        std::ostringstream sstream;
+        sstream << "[";
+        for (auto& id : rows) {
+          sstream << id << ", ";
+        }
+        sstream << "]";
+        VLOG(3) << "update ids -> " << sstream.str();
+      }
+      for (auto row : rows) {
+        set_.insert(row);
+      }
+    };
+    return pool_->enqueue(std::move(task));
+  }
+
+  std::future<void> GetAndClear(std::vector<int64_t>* result) {
+    auto task = [this, &result] {
+      result->clear();
+      for (auto& id : set_) {
+        result->push_back(id);
+      }
+      if (VLOG_IS_ON(3)) {
+        std::ostringstream sstream;
+        sstream << "[";
+        for (auto& id : *result) {
+          sstream << id << ", ";
+        }
+        sstream << "]";
+        VLOG(3) << "result ids size: " << result->size() << " "
+                << sstream.str();
+      }
+      set_.clear();
+    };
+    return pool_->enqueue(std::move(task));
+  }
+
+ private:
+  std::unordered_set<int64_t> set_;
+  std::unique_ptr<::ThreadPool> pool_{nullptr};
+};
+
+class AsyncSparseParamUpdateRecorder {
+  using TrainerToRows = std::vector<std::unique_ptr<ConcurrentSet>>;
+
+ public:
+  AsyncSparseParamUpdateRecorder(
+      int trainer_num,
+      const std::unordered_map<std::string, std::string>& grad_to_param)
+      : trainer_num_(trainer_num), grad_to_param_(grad_to_param) {
+    if (VLOG_IS_ON(3)) {
+      std::ostringstream sstream;
+      sstream << "[";
+      for (auto& item : grad_to_param) {
+        sstream << item.first << ":" << item.second << ", ";
+      }
+      sstream << "]";
+      VLOG(3) << "trainer_num: " << trainer_num
+              << " grad_to_param_: " << sstream.str();
+    }
+    for (auto& iter : grad_to_param) {
+      param_to_grad_[iter.second] = iter.first;
+      auto& param_name = iter.second;
+      param_to_updated_rows_[param_name] = TrainerToRows();
+      auto& trainer_to_rows = param_to_updated_rows_[param_name];
+      for (auto i = 0; i < trainer_num; ++i) {
+        trainer_to_rows.emplace_back(new ConcurrentSet());
+      }
+    }
+  }
+
+  ~AsyncSparseParamUpdateRecorder() = default;
+
+  void Update(const std::string& grad_name,
+              const std::vector<int64_t>& update_rows) {
+    VLOG(3) << "update grad: " << grad_name
+            << " row size: " << update_rows.size();
+    auto& param_name = grad_to_param_.at(grad_name);
+    auto& trainer_to_rows = param_to_updated_rows_.at(param_name);
+
+    std::vector<std::future<void>> fs;
+    for (auto& set : trainer_to_rows) {
+      fs.push_back(set->Update(update_rows));
+    }
+    for (auto& f : fs) {
+      f.wait();
+    }
+  }
+
+  void GetAndClear(const std::string& param_name, int trainer_id,
+                   std::vector<int64_t>* result) {
+    VLOG(3) << "GetAndClear param: " << param_name
+            << " for trainer: " << trainer_id;
+    PADDLE_ENFORCE_LT(trainer_id, trainer_num_);
+    param_to_updated_rows_.at(param_name)[trainer_id]
+        ->GetAndClear(result)
+        .wait();
+  }
+
+  bool HasParam(const std::string& param_name) {
+    return param_to_grad_.find(param_name) != param_to_grad_.end();
+  }
+
+  bool HasGrad(const std::string& grad_name) {
+    return grad_to_param_.find(grad_name) != grad_to_param_.end();
+  }
+
+ private:
+  const int trainer_num_;
+  std::unordered_map<std::string, std::string> grad_to_param_;
+  std::unordered_map<std::string, std::string> param_to_grad_;
+  std::unordered_map<std::string, TrainerToRows> param_to_updated_rows_;
+
+  // init recorder
+ public:
+  static void Init(
+      int trainer_num,
+      const std::unordered_map<std::string, std::string>& grad_to_param) {
+    InitImpl(trainer_num, grad_to_param);
+  }
+
+  static AsyncSparseParamUpdateRecorder* GetInstance() {
+    return recorder_.get();
+  }
+
+ private:
+  // Init is called by GetInstance.
+  static void InitImpl(
+      int trainer_num,
+      const std::unordered_map<std::string, std::string>& grad_to_param) {
+    if (recorder_ == nullptr) {
+      recorder_.reset(
+          new AsyncSparseParamUpdateRecorder(trainer_num, grad_to_param));
+    }
+  }
+
+  static std::once_flag init_flag_;
+  static std::unique_ptr<AsyncSparseParamUpdateRecorder> recorder_;
+};
+
+}  // namespace distributed
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/distributed/async_sparse_param_update_recorder_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/operators/distributed/async_sparse_param_update_recorder.h"
+
+#include <algorithm>
+
+#include "gtest/gtest.h"
+
+namespace paddle {
+namespace operators {
+namespace distributed {
+
+TEST(ConcurrentSet, All) {
+  ConcurrentSet concurrent_set;
+  std::vector<int64_t> in1 = {1, 2, 3, 4};
+  std::vector<int64_t> in2 = {2, 3, 5, 6};
+
+  std::vector<std::future<void>> futures;
+  futures.push_back(concurrent_set.Update(in1));
+  futures.push_back(concurrent_set.Update(in2));
+
+  for (auto &f : futures) {
+    f.wait();
+  }
+
+  std::unordered_set<int64_t> in;
+  std::copy(in1.begin(), in1.end(), std::inserter(in, in.begin()));
+  std::copy(in2.begin(), in2.end(), std::inserter(in, in.begin()));
+
+  std::vector<int64_t> ret;
+  concurrent_set.GetAndClear(&ret).wait();
+
+  std::unordered_set<int64_t> out;
+  std::copy(ret.begin(), ret.end(), std::inserter(out, out.begin()));
+
+  EXPECT_EQ(in, out);
+
+  concurrent_set.GetAndClear(&ret).wait();
+  EXPECT_EQ(ret.size(), 0);
+}
+
+TEST(AsyncSparseParamUpdateRecorder, All) {
+  std::unordered_map<std::string, std::string> grad_to_param;
+  grad_to_param["grad1"] = "param1";
+  grad_to_param["grad2"] = "param2";
+
+  int trainer_num = 10;
+
+  AsyncSparseParamUpdateRecorder recorder(trainer_num, grad_to_param);
+  std::vector<int64_t> in1 = {1, 2, 3, 4};
+  std::vector<int64_t> in2 = {2, 3, 5, 6};
+
+  std::unordered_set<int64_t> in;
+  std::copy(in1.begin(), in1.end(), std::inserter(in, in.begin()));
+  std::copy(in2.begin(), in2.end(), std::inserter(in, in.begin()));
+
+  recorder.Update("grad1", in1);
+  recorder.Update("grad1", in2);
+
+  EXPECT_TRUE(recorder.HasParam("param1"));
+  EXPECT_TRUE(recorder.HasParam("param2"));
+  EXPECT_FALSE(recorder.HasParam("param3"));
+
+  EXPECT_TRUE(recorder.HasGrad("grad1"));
+  EXPECT_TRUE(recorder.HasGrad("grad2"));
+  EXPECT_FALSE(recorder.HasGrad("grad3"));
+
+  std::vector<int64_t> ret;
+  EXPECT_ANY_THROW(recorder.GetAndClear("param1", trainer_num, &ret));
+
+  for (int i = 0; i < trainer_num; ++i) {
+    std::vector<int64_t> ret;
+    std::unordered_set<int64_t> out;
+
+    recorder.GetAndClear("param1", i, &ret);
+    std::copy(ret.begin(), ret.end(), std::inserter(out, out.begin()));
+
+    EXPECT_EQ(in, out);
+
+    recorder.GetAndClear("param1", i, &ret);
+    EXPECT_EQ(ret.size(), 0);
+  }
+}
+
+}  // namespace distributed
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/distributed/brpc/brpc_client.cc

@@ -234,6 +234,7 @@ VarHandlePtr BRPCClient::AsyncGetVar(const std::string& ep,
                                      const framework::Scope& scope,
                                      const std::string& var_name,
                                      const std::string& out_var_name,
+                                     const std::string& table_name,
                                      int64_t time_out) {
   return _AsyncGetVar(ep, ctx, scope, var_name, out_var_name, kGetRPC,
                       time_out);

paddle/fluid/operators/distributed/brpc/brpc_client.h

@@ -21,8 +21,10 @@ limitations under the License. */
 #include <functional>
 #include <iostream>
 #include <map>
+#include <memory>
 #include <mutex>  // NOLINT
 #include <string>
+#include <unordered_map>
 #include <vector>
 
 #include "brpc/channel.h"
@@ -66,6 +68,7 @@ class BRPCClient : public RPCClient {
                            const framework::Scope& scope,
                            const std::string& var_name,
                            const std::string& out_var_name,
+                           const std::string& table_name = "",
                            int64_t time_out = FLAGS_rpc_deadline) override;
 
   VarHandlePtr AsyncGetMonomerBarrier(
@@ -107,13 +110,11 @@ class BRPCClient : public RPCClient {
   void SendComplete() override;
 
  private:
-  VarHandlePtr _AsyncGetVar(const std::string& ep,
-                            const platform::DeviceContext& ctx,
-                            const framework::Scope& scope,
-                            const std::string& var_name,
-                            const std::string& out_var_name,
-                            const std::string& method_name,
-                            int64_t time_out = FLAGS_rpc_deadline);
+  VarHandlePtr _AsyncGetVar(
+      const std::string& ep, const platform::DeviceContext& ctx,
+      const framework::Scope& scope, const std::string& var_name,
+      const std::string& out_var_name, const std::string& method_name,
+      const std::string& table_name, int64_t time_out = FLAGS_rpc_deadline);
 
   void Proceed();
   ChannelQueuePtr GetChannel(const std::string& ep);

paddle/fluid/operators/distributed/communicator.cc

@@ -32,6 +32,9 @@ DEFINE_int32(communicator_send_queue_size, 20,
 DEFINE_int32(communicator_max_send_grad_num_before_recv, 20,
              "max grad num to send before recv parameters");
 DEFINE_int32(communicator_thread_pool_size, 5, "thread num to do send or recv");
+DEFINE_int32(communicator_send_wait_times, 5,
+             "times that send thread will wait if merge num does not reach "
+             "max_merge_var_num");
 DEFINE_int32(communicator_max_merge_var_num, 20,
              "max var num to merge and send");
 DEFINE_bool(communicator_fake_rpc, false,
@@ -65,6 +68,8 @@ Communicator::Communicator(const RpcCtxMap &send_varname_to_ctx,
           << FLAGS_communicator_max_send_grad_num_before_recv;
   VLOG(0) << "communicator_thread_pool_size: "
           << FLAGS_communicator_thread_pool_size;
+  VLOG(0) << "communicator_send_wait_times: "
+          << FLAGS_communicator_send_wait_times;
   VLOG(0) << "communicator_max_merge_var_num: "
           << FLAGS_communicator_max_merge_var_num;
   VLOG(0) << "communicator_fake_rpc: " << FLAGS_communicator_fake_rpc;
@@ -101,20 +106,32 @@ void Communicator::SendThread() {
           VLOG(3) << var_name << " merge and send";
           std::vector<std::shared_ptr<Variable>> vars;
           size_t merged_var_num = 0;
-          while (var_queue->Size() > 0 &&
-                 merged_var_num < FLAGS_communicator_max_merge_var_num) {
-            vars.push_back(var_queue->Pop());
-            // only count the send number of the first var
-            if (var_name == send_varname_to_queue_.begin()->first) {
-              grad_num_.fetch_add(1, std::memory_order_relaxed);
+          size_t wait_times = 0;
+          while (merged_var_num < FLAGS_communicator_max_merge_var_num) {
+            if (var_queue->Size() == 0) {
+              VLOG(3) << "wait_times -> " << wait_times;
+              if (wait_times >= FLAGS_communicator_send_wait_times) {
+                break;
+              }
+              std::this_thread::sleep_for(std::chrono::milliseconds(10));
+              wait_times++;
+              continue;
+            } else {
+              wait_times = 0;
+
+              vars.push_back(var_queue->Pop());
+              // only count the send number of the first var
+              if (var_name == send_varname_to_queue_.begin()->first) {
+                grad_num_.fetch_add(1, std::memory_order_relaxed);
+              }
+              merged_var_num++;
             }
-            merged_var_num++;
           }
           auto before_merge = GetCurrentUS();
           MergeVars(var_name, vars, send_scope_.get());
           auto after_merge = GetCurrentUS();
-          VLOG(3) << "merge " << var_name << " use time "
-                  << after_merge - before_merge;
+          VLOG(3) << "merge " << merged_var_num << " " << var_name
+                  << " use time " << after_merge - before_merge;
           auto send_functor = distributed::ParameterSend<float>();
           auto &ctx = send_varname_to_ctx_.at(var_name);
           if (!FLAGS_communicator_fake_rpc) {

paddle/fluid/operators/distributed/communicator.h

@@ -109,7 +109,7 @@ inline void MergeVars(const std::string& var_name,
   auto* out_var = scope->Var(var_name);
   if (var0->IsType<framework::LoDTensor>()) {
     auto dims = var0->Get<framework::LoDTensor>().dims();
-    VLOG(3) << "merge " << var_name << " LoDTensor " << dims;
+    VLOG(3) << "merge " << var_name << " LoDTensor dims " << dims;
 
     // init output tensor
     auto* out_t = out_var->GetMutable<framework::LoDTensor>();

paddle/fluid/operators/distributed/grpc/grpc_client.cc

@@ -128,9 +128,11 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep,
                                      const framework::Scope& scope,
                                      const std::string& var_name,
                                      const std::string& out_varname,
+                                     const std::string& table_name,
                                      int64_t time_out) {
   return _AsyncGetVar(ep, ctx, scope, kGetRPC, var_name, out_varname,
-                      "/sendrecv.SendRecvService/GetVariable", time_out);
+                      "/sendrecv.SendRecvService/GetVariable", table_name,
+                      time_out);
 }
 
 VarHandlePtr GRPCClient::AsyncGetVarNoBarrier(
@@ -142,7 +144,7 @@ VarHandlePtr GRPCClient::AsyncGetVarNoBarrier(
 
   return _AsyncGetVar(
       ep, ctx, scope, kGetNoBarrierRPC, var_name_no_barrier, out_varname,
-      "/sendrecv.SendRecvService/GetVariableNoBarrier", time_out);
+      "/sendrecv.SendRecvService/GetVariableNoBarrier", "", time_out);
 }
 
 VarHandlePtr GRPCClient::AsyncGetMonomerVariable(
@@ -150,18 +152,21 @@ VarHandlePtr GRPCClient::AsyncGetMonomerVariable(
     const framework::Scope& scope, const std::string& var_name,
     int64_t time_out) {
   return _AsyncGetVar(ep, ctx, scope, kGetMonomerRPC, var_name, var_name,
-                      "/sendrecv.SendRecvService/GetMonomerVariable", time_out);
+                      "/sendrecv.SendRecvService/GetMonomerVariable", "",
+                      time_out);
 }
 
 VarHandlePtr GRPCClient::_AsyncGetVar(
     const std::string& ep, const platform::DeviceContext& ctx,
     const framework::Scope& scope, const std::string& method,
     const std::string& var_name, const std::string& out_varname,
-    const std::string& rpc_path, int64_t time_out) {
+    const std::string& rpc_path, const std::string& table_name,
+    int64_t time_out) {
   const platform::DeviceContext* p_ctx = &ctx;
   const std::string ep_val = ep;
   const std::string var_name_val = var_name;
   const std::string out_varname_val = out_varname;
+  const std::string table_name_val = table_name;
   const framework::Scope* p_scope = &scope;
   const auto ch = GetChannel(ep_val);
   GetProcessor* s = new GetProcessor(ch);
@@ -169,32 +174,33 @@ VarHandlePtr GRPCClient::_AsyncGetVar(
   VarHandlePtr h(new VarHandle(ep, method, out_varname_val, p_ctx, p_scope));
   s->Prepare(h, time_out);
 
-  framework::AsyncIO(
-      [var_name_val, out_varname_val, s, method, p_ctx, h, rpc_path, this] {
-        // prepare input
-        sendrecv::VariableMessage req;
-        req.set_varname(var_name_val);
-        req.set_out_varname(out_varname_val);
-        req.set_trainer_id(trainer_id_);
-        ::grpc::ByteBuffer buf;
-        RequestToByteBuffer<sendrecv::VariableMessage>(req, &buf);
+  framework::AsyncIO([var_name_val, out_varname_val, table_name_val, s, method,
+                      p_ctx, h, rpc_path, this] {
+    // prepare input
+    sendrecv::VariableMessage req;
+    req.set_varname(var_name_val);
+    req.set_out_varname(out_varname_val);
+    req.set_trainer_id(trainer_id_);
+    req.set_table_name(table_name_val);
+    ::grpc::ByteBuffer buf;
+    RequestToByteBuffer<sendrecv::VariableMessage>(req, &buf);
 
-        VLOG(3) << s->GetVarHandlePtr()->String() << " begin";
+    VLOG(3) << s->GetVarHandlePtr()->String() << " begin";
 
-        // stub context
-        s->response_call_back_ = ProcGetResponse;
+    // stub context
+    s->response_call_back_ = ProcGetResponse;
 
-        platform::RecordRPCEvent record_event(method);
+    platform::RecordRPCEvent record_event(method);
 
-        auto call =
-            s->stub_g_.PrepareUnaryCall(s->context_.get(), rpc_path, buf, &cq_);
-        call->StartCall();
-        call->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
+    auto call =
+        s->stub_g_.PrepareUnaryCall(s->context_.get(), rpc_path, buf, &cq_);
+    call->StartCall();
+    call->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
 
-        if (UNLIKELY(platform::IsProfileEnabled())) {
-          h->Wait();
-        }
-      });
+    if (UNLIKELY(platform::IsProfileEnabled())) {
+      h->Wait();
+    }
+  });
 
   req_count_++;
 

paddle/fluid/operators/distributed/grpc/grpc_client.h

@@ -23,9 +23,11 @@ limitations under the License. */
 #include <functional>
 #include <iostream>
 #include <map>
+#include <memory>
 #include <mutex>  // NOLINT
 #include <string>
 #include <thread>  // NOLINT
+#include <unordered_map>
 #include <vector>
 
 #include "grpc++/channel.h"
@@ -187,6 +189,7 @@ class GRPCClient : public RPCClient {
                            const framework::Scope& scope,
                            const std::string& var_name,
                            const std::string& out_varname,
+                           const std::string& table_name = "",
                            int64_t time_out = FLAGS_rpc_deadline) override;
 
   VarHandlePtr AsyncGetVarNoBarrier(
@@ -239,7 +242,8 @@ class GRPCClient : public RPCClient {
       const std::string& ep, const platform::DeviceContext& ctx,
       const framework::Scope& scope, const std::string& method,
       const std::string& var_name, const std::string& out_varname,
-      const std::string& rpc_path, int64_t time_out = FLAGS_rpc_deadline);
+      const std::string& rpc_path, const std::string& table_name = "",
+      int64_t time_out = FLAGS_rpc_deadline);
 
  private:
   grpc::CompletionQueue cq_;

paddle/fluid/operators/distributed/grpc/grpc_serde.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 #include <nccl.h>
 #endif
 #include <limits>
+#include <memory>
 #include <thread>  // NOLINT
 
 #include "google/protobuf/io/coded_stream.h"
@@ -104,8 +105,10 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
   e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber,
                             payload->memory_size());
   if (payload->memory_size() >= std::numeric_limits<int>::max()) {
-    LOG(FATAL) << "AppendZeroCopy varname:" << name
-               << ", vlen:" << payload->memory_size();
+    LOG(FATAL) << "FATAL error: varname:" << name
+               << ", vlen:" << payload->memory_size()
+               << " >= std::numeric_limits<int>::max():"
+               << std::numeric_limits<int>::max() << ", so exit!";
   }
   // steal reference of tensor data
   ::grpc::Slice slices[4];  // metadata, tensor, rows meta, rows

paddle/fluid/operators/distributed/grpc/grpc_server.cc

@@ -137,6 +137,7 @@ class RequestGet final : public RequestBase {
     // proc request.
     std::string varname = request_.varname();
     std::string out_varname = request_.out_varname();
+    std::string table_name = request_.table_name();
     int trainer_id = request_.trainer_id();
 
     VLOG(4) << "RequestGet " << out_varname << " from " << varname;
@@ -145,19 +146,23 @@ class RequestGet final : public RequestBase {
     framework::Variable* invar = nullptr;
     framework::Variable* outvar = nullptr;
 
-    request_handler_->Handle(varname, scope, invar, &outvar, trainer_id,
-                             out_varname);
+    tmp_scope_ = std::move(scope->NewTmpScope());
+    request_handler_->Handle(varname, tmp_scope_.get(), invar, &outvar,
+                             trainer_id, out_varname, table_name);
 
+    VLOG(1) << "before SerializeToByteBuffer";
     if (outvar) {
       SerializeToByteBuffer(out_varname, outvar, *request_handler_->dev_ctx(),
                             &reply_);
     }
+    VLOG(1) << "after SerializeToByteBuffer";
     Finish(reply_, &responder_);
   }
 
  protected:
   sendrecv::VariableMessage request_;
   ::grpc::ByteBuffer reply_;
+  std::unique_ptr<framework::Scope> tmp_scope_;
   ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_;
 };
 

paddle/fluid/operators/distributed/parameter_recv.cc

@@ -42,27 +42,23 @@ using DDim = framework::DDim;
 template <typename T>
 void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
                                   const framework::Scope &scope) {
-  VLOG(3) << "ParameterRecv in";
+  VLOG(3) << "ParameterRecv in " << rpc_ctx.var_name;
   std::unique_ptr<framework::Scope> local_scope = scope.NewTmpScope();
 
   platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
   auto &cpu_ctx = *pool.Get(platform::CPUPlace());
 
   distributed::RPCClient *rpc_client =
-      distributed::RPCClient::GetInstance<RPCCLIENT_T>(0);
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>(rpc_ctx.trainer_id);
 
   auto *recv_var = scope.FindVar(rpc_ctx.var_name);
 
-  std::vector<framework::Tensor *> recved_tensors;
-
   // recv all vars to local scope
   if (recv_var->IsType<framework::LoDTensor>()) {
     std::vector<distributed::VarHandlePtr> rets;
     for (size_t i = 0; i < rpc_ctx.splited_var_names.size(); i++) {
       auto &recv_var_name = rpc_ctx.splited_var_names[i];
-      framework::Tensor *t =
-          local_scope->Var(recv_var_name)->GetMutable<framework::LoDTensor>();
-      recved_tensors.push_back(t);
+      local_scope->Var(recv_var_name);
       VLOG(3) << "recv " << recv_var_name << " from " << rpc_ctx.epmap[i];
       rets.push_back(rpc_client->AsyncGetVar(rpc_ctx.epmap[i], cpu_ctx,
                                              *local_scope.get(), recv_var_name,
@@ -78,23 +74,61 @@ void ParameterRecv<T>::operator()(const RpcContext &rpc_ctx,
   // concat recved tensor into one var
   {
     size_t output_offset = 0;
+    size_t row_offset = 0;
     framework::Tensor *recv_tensor =
         recv_var->GetMutable<framework::LoDTensor>();
     auto dev_ctx = paddle::platform::CPUDeviceContext();
     int64_t recv_numel = 0;
-    for (auto *in : recved_tensors) {
-      recv_numel += in->numel();
-      auto in_stride = framework::stride_numel(in->dims());
-      auto out_stride = framework::stride_numel(recv_tensor->dims());
-      StridedNumelCopyWithAxis<T>(
-          dev_ctx, 0, recv_tensor->data<T>() + output_offset, out_stride,
-          in->data<T>(), in_stride, in_stride[0]);
-      output_offset += in_stride[0];
+    for (auto &recv_var_name : rpc_ctx.splited_var_names) {
+      auto *recv_var = local_scope->FindVar(recv_var_name);
+      if (recv_var->IsType<framework::LoDTensor>()) {
+        auto &in = recv_var->Get<framework::LoDTensor>();
+        recv_numel += in.numel();
+        auto in_stride = framework::stride_numel(in.dims());
+        auto out_stride = framework::stride_numel(recv_tensor->dims());
+        StridedNumelCopyWithAxis<T>(
+            dev_ctx, 0, recv_tensor->data<T>() + output_offset, out_stride,
+            in.data<T>(), in_stride, in_stride[0]);
+        output_offset += in_stride[0];
+      } else if (recv_var->IsType<framework::SelectedRows>()) {
+        auto &recv_slr = recv_var->Get<framework::SelectedRows>();
+        auto &recv_dims = recv_tensor->dims();
+        int64_t width = recv_dims[1];
+        recv_numel += recv_slr.height() * width;
+        PADDLE_ENFORCE_EQ(recv_slr.value().dims()[1], width);
+        PADDLE_ENFORCE_EQ(recv_slr.value().dims()[0], recv_slr.rows().size());
+        VLOG(3) << "recv slr " << recv_var_name << " dims "
+                << recv_slr.value().dims();
+        if (VLOG_IS_ON(3)) {
+          std::ostringstream sstream;
+          sstream << "[";
+          for (auto &row_id : recv_slr.rows()) {
+            sstream << row_id << ", ";
+          }
+          sstream << "]";
+          VLOG(3) << "recv_slr size: " << recv_slr.rows().size() << " "
+                  << sstream.str();
+        }
+
+        for (auto i = 0; i < recv_slr.rows().size(); ++i) {
+          auto row_id = recv_slr.rows()[i] + row_offset;
+          PADDLE_ENFORCE_LT(row_id, recv_dims[0]);
+          memcpy(recv_tensor->data<T>() + row_id * width,
+                 recv_slr.value().data<T>() + i * width, sizeof(T) * width);
+        }
+        row_offset += recv_slr.height();
+      } else {
+        PADDLE_THROW("unsupported recieved var type");
+      }
+    }
+    auto numel = recv_tensor->numel();
+    if (recv_numel != numel) {
+      LOG(FATAL) << "recv_numel: " << recv_numel << " acture numel: " << numel;
     }
-    PADDLE_ENFORCE_EQ(recv_numel, recv_tensor->numel());
+    PADDLE_ENFORCE_EQ(recv_numel, numel);
   }
 
-  VLOG(3) << "ParameterRecv out";
+  VLOG(3) << "ParameterRecv out " << rpc_ctx.var_name;
 }
 
 template struct ParameterRecv<float>;

paddle/fluid/operators/distributed/parameter_send.cc

@@ -47,7 +47,7 @@ void ParameterSend<T>::operator()(const RpcContext &rpc_ctx,
   auto &cpu_ctx = *pool.Get(platform::CPUPlace());
 
   distributed::RPCClient *rpc_client =
-      distributed::RPCClient::GetInstance<RPCCLIENT_T>(0);
+      distributed::RPCClient::GetInstance<RPCCLIENT_T>(rpc_ctx.trainer_id);
 
   auto *send_var = scope.FindVar(rpc_ctx.var_name);
   size_t out_num = rpc_ctx.splited_var_names.size();

paddle/fluid/operators/distributed/request_handler.h

@@ -18,7 +18,9 @@
 #include <condition_variable>  // NOLINT
 
 #include <functional>
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <utility>
 #include <vector>
 
@@ -180,6 +182,10 @@ class RequestHandler {
     grad_to_prepared_ctx_ = g;
   }
 
+  void SetSparseGradToParam(std::unordered_map<std::string, std::string>* g) {
+    sparse_grad_to_param_ = g;
+  }
+
   void SetRPCServer(RPCServer* rpc_server) { rpc_server_ = rpc_server; }
 
   // Get attributes.
@@ -228,6 +234,7 @@ class RequestHandler {
   std::unordered_map<std::string,
                      std::shared_ptr<framework::ExecutorPrepareContext>>*
       grad_to_prepared_ctx_;
+  std::unordered_map<std::string, std::string>* sparse_grad_to_param_;
 
   RPCServer* rpc_server_;
 };

paddle/fluid/operators/distributed/request_handler_impl.cc

@@ -22,6 +22,7 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/variable_helper.h"
+#include "paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h"
 #include "paddle/fluid/operators/distributed/rpc_server.h"
 #include "paddle/fluid/string/piece.h"
 #include "paddle/fluid/string/printf.h"
@@ -59,6 +60,12 @@ bool RequestSendHandler::Handle(const std::string& varname,
             "async mode should not recv BATCH_BARRIER_MESSAGE or "
             "COMPLETE_MESSAGE");
       }
+      if (AsyncSparseParamUpdateRecorder::GetInstance()->HasGrad(varname)) {
+        auto& grad_slr =
+            scope->FindVar(varname)->Get<framework::SelectedRows>();
+        AsyncSparseParamUpdateRecorder::GetInstance()->Update(varname,
+                                                              grad_slr.rows());
+      }
       executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(),
                                     scope);
       return true;
@@ -82,8 +89,9 @@ bool RequestGetHandler::Handle(const std::string& varname,
                                const int trainer_id,
                                const std::string& out_var_name,
                                const std::string& table_name) {
-  VLOG(4) << "RequestGetHandler:" << varname
-          << " out_var_name: " << out_var_name;
+  VLOG(3) << "RequestGetHandler:" << varname
+          << " out_var_name: " << out_var_name << " trainer_id: " << trainer_id
+          << " table_name: " << table_name;
 
   if (sync_mode_) {
     if (varname == FETCH_BARRIER_MESSAGE) {
@@ -108,7 +116,42 @@ bool RequestGetHandler::Handle(const std::string& varname,
         VLOG(3) << "copying " << varname << " to " << param_bak_name;
         framework::TensorCopy(t_orig, dev_ctx_->GetPlace(), t);
       }
-      *outvar = scope_->FindVar(varname);
+      if (AsyncSparseParamUpdateRecorder::GetInstance()->HasParam(varname) &&
+          !table_name.empty()) {
+        std::vector<int64_t> updated_rows;
+        AsyncSparseParamUpdateRecorder::GetInstance()->GetAndClear(
+            varname, trainer_id, &updated_rows);
+        if (VLOG_IS_ON(3)) {
+          std::ostringstream sstream;
+          sstream << "[";
+          for (auto& row_id : updated_rows) {
+            sstream << row_id << ", ";
+          }
+          sstream << "]";
+          VLOG(3) << "updated_rows size: " << updated_rows.size() << " "
+                  << sstream.str();
+        }
+        auto& origin_tensor =
+            scope_->FindVar(varname)->Get<framework::LoDTensor>();
+        auto* origin_tensor_data = origin_tensor.data<float>();
+        auto& dims = origin_tensor.dims();
+        *outvar = scope->Var();
+        auto* out_slr = (*outvar)->GetMutable<framework::SelectedRows>();
+        out_slr->set_rows(updated_rows);
+        out_slr->set_height(dims[0]);
+        auto out_dims = framework::make_ddim(
+            {static_cast<int64_t>(updated_rows.size()), dims[1]});
+        auto* data = out_slr->mutable_value()->mutable_data<float>(
+            out_dims, origin_tensor.place());
+        auto width = dims[1];
+        for (auto i = 0; i < updated_rows.size(); ++i) {
+          PADDLE_ENFORCE_LT(updated_rows[i], dims[0]);
+          memcpy(data + i * width, origin_tensor_data + updated_rows[i] * width,
+                 sizeof(float) * width);
+        }
+      } else {
+        *outvar = scope_->FindVar(varname);
+      }
     }
   }
   return true;

paddle/fluid/operators/distributed/rpc_client.h

@@ -15,6 +15,7 @@
 #pragma once
 
 #include <condition_variable>  // NOLINT
+#include <memory>
 #include <string>
 #include "gflags/gflags.h"
 
@@ -44,6 +45,7 @@ class RPCClient {
                                    const framework::Scope& scope,
                                    const std::string& var_name,
                                    const std::string& out_varname,
+                                   const std::string& table_name = "",
                                    int64_t time_out = FLAGS_rpc_deadline) = 0;
 
   virtual VarHandlePtr AsyncGetVarNoBarrier(
@@ -96,6 +98,7 @@ class RPCClient {
   // Init is called by GetInstance.
   template <typename T>
   static void Init(int trainer_id) {
+    VLOG(0) << "init rpc client with trainer_id " << trainer_id;
     trainer_id_ = trainer_id;
     if (rpc_client_.get() == nullptr) {
       rpc_client_.reset(new T());

paddle/fluid/operators/distributed/rpc_common.h

@@ -27,23 +27,26 @@ struct RpcContext {
 
   RpcContext(const std::string &name, const std::vector<std::string> &names,
              const std::vector<std::string> &emap,
-             const std::vector<int64_t> &sections)
+             const std::vector<int64_t> &sections, int id)
       : var_name(name),
         splited_var_names(names),
         epmap(emap),
-        height_sections(sections) {}
+        height_sections(sections),
+        trainer_id(id) {}
 
   RpcContext(const RpcContext &ctx) {
     var_name = ctx.var_name;
     splited_var_names = ctx.splited_var_names;
     epmap = ctx.epmap;
     height_sections = ctx.height_sections;
+    trainer_id = ctx.trainer_id;
   }
 
   std::string var_name;
   std::vector<std::string> splited_var_names;
   std::vector<std::string> epmap;
   std::vector<int64_t> height_sections;
+  int trainer_id;
 };
 
 inline std::ostream &operator<<(std::ostream &os, const RpcContext &rpc_ctx) {

paddle/fluid/operators/distributed_ops/CMakeLists.txt

@@ -2,9 +2,9 @@ include(operators)
 
 set(DISTRIBUTE_DEPS "")
 if(WITH_GRPC)
-    set(DISTRIBUTE_DEPS sendrecvop_rpc parameter_send parameter_recv communicator grpc++_unsecure grpc_unsecure gpr cares zlib protobuf node)
+    set(DISTRIBUTE_DEPS sendrecvop_rpc parameter_send parameter_recv communicator async_sparse_param_update_recorder grpc++_unsecure grpc_unsecure gpr cares zlib protobuf node)
 else()
-    set(DISTRIBUTE_DEPS sendrecvop_rpc parameter_send parameter_recv communicator brpc leveldb snappystream snappy protobuf ssl crypto zlib node)
+    set(DISTRIBUTE_DEPS sendrecvop_rpc parameter_send parameter_recv communicator async_sparse_param_update_recorder brpc leveldb snappystream snappy protobuf ssl crypto zlib node)
     if(WITH_BRPC_RDMA)
         find_library(IBVERBS_LIBRARY NAMES ibverbs)
         ADD_LIBRARY(ibverbs SHARED IMPORTED GLOBAL)

paddle/fluid/operators/distributed_ops/listen_and_serv_op.cc

@@ -24,8 +24,10 @@ limitations under the License. */
 #include "paddle/fluid/operators/distributed/distributed.h"
 #include "paddle/fluid/operators/math/math_function.h"
 
+#include "paddle/fluid/operators/distributed/async_sparse_param_update_recorder.h"
 #include "paddle/fluid/operators/distributed/request_handler_impl.h"
 #include "paddle/fluid/operators/distributed_ops/listen_and_serv_op.h"
+
 #include "paddle/fluid/platform/profiler.h"
 
 DEFINE_int32(rpc_send_thread_num, 12, "number of threads for rpc send");
@@ -292,6 +294,8 @@ static void FillRequestCtx(
     std::unordered_map<std::string,
                        std::shared_ptr<framework::ExecutorPrepareContext>>
         *prefetch_ctx,
+    std::unordered_map<std::string, std::string>
+        *sparse_grad_name_to_param_name,
     std::shared_ptr<framework::ExecutorPrepareContext> checkpoint_ctx,
     distributed::RPCServer *rpc_server) {
   h->SetScope(scope);
@@ -299,6 +303,7 @@ static void FillRequestCtx(
   h->SetExecutor(executor);
   h->SetProgram(program);
   h->SetPrefetchPreparedCtx(prefetch_ctx);
+  h->SetSparseGradToParam(sparse_grad_name_to_param_name);
   h->SetRPCServer(rpc_server);
   h->SetCheckpointNotifyPreparedCtx(checkpoint_ctx);
 }
@@ -414,10 +419,24 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
     prefetch_var_name_to_prepared_ctx[prefetch_var_name] = prefetch_prepared[i];
   }
 
-  auto f =
-      std::bind(FillRequestCtx, std::placeholders::_1, &recv_scope, &dev_ctx,
-                &executor, program, &prefetch_var_name_to_prepared_ctx,
-                ckpt_pre_context, rpc_service_.get());
+  // parse attr of kSparseGradToParam  sparse_grad_name -> param_name
+  std::unordered_map<std::string, std::string> sparse_grad_name_to_param_name;
+  auto sparse_grad_name_to_param_name_str =
+      Attr<std::vector<std::string>>(kSparseGradToParam);
+  for (const auto &sparse_grad_name_and_param_name :
+       sparse_grad_name_to_param_name_str) {
+    std::vector<std::string> pieces;
+    split(sparse_grad_name_and_param_name, ':', &pieces);
+    PADDLE_ENFORCE_EQ(pieces.size(), 2);
+    VLOG(3) << "after split, sparse_grad_name = " << pieces[0]
+            << ", param_name = " << pieces[1];
+    sparse_grad_name_to_param_name[pieces[0]] = pieces[1];
+  }
+
+  auto f = std::bind(
+      FillRequestCtx, std::placeholders::_1, &recv_scope, &dev_ctx, &executor,
+      program, &prefetch_var_name_to_prepared_ctx,
+      &sparse_grad_name_to_param_name, ckpt_pre_context, rpc_service_.get());
 
   f(request_send_handler_.get());
   f(request_get_handler_.get());
@@ -445,6 +464,8 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
     RunSyncLoop(&executor, program, &recv_scope, &dev_ctx,
                 prefetch_block_id_list, checkpoint_block_id);
   } else {
+    distributed::AsyncSparseParamUpdateRecorder::Init(
+        fan_in, sparse_grad_name_to_param_name);
     RunAsyncLoop(&executor, program, &recv_scope);
   }
 }
@@ -475,6 +496,10 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<std::vector<std::string>>(kPrefetchVarNameToBlockId,
                                       "prefetch blocks to run on server side.")
         .SetDefault({});
+    AddAttr<std::vector<std::string>>(
+        kSparseGradToParam,
+        "sparse grad name to param name. like: 'emb@Grad:emb'")
+        .SetDefault({});
     AddAttr<int>("Fanin", "How many clients send to this server.")
         .SetDefault(1);
     AddAttr<int>(kCheckpointBlockId,

paddle/fluid/operators/distributed_ops/listen_and_serv_op.h

@@ -16,8 +16,10 @@ limitations under the License. */
 
 #include <stdint.h>
 #include <atomic>
+#include <memory>
 #include <set>
 #include <string>
+#include <unordered_map>
 #include <utility>
 #include <vector>
 
@@ -35,6 +37,7 @@ namespace operators {
 constexpr char kOptimizeBlocks[] = "optimize_blocks";
 constexpr char kPrefetchVarNameToBlockId[] = "prefetch_var_name_to_block_id";
 constexpr char kCheckpointBlockId[] = "checkpint_block_id";
+constexpr char kSparseGradToParam[] = "sparse_grad_to_param";
 
 void RunServer(std::shared_ptr<distributed::RPCServer> service);
 

paddle/fluid/operators/distributed_ops/recv_op.cc

@@ -50,17 +50,18 @@ class RecvOp : public framework::OperatorBase {
 
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &ctx = *pool.Get(place);
+    auto trainer_id = Attr<int>("trainer_id");
 
     distributed::RPCClient *rpc_client =
-        distributed::RPCClient::GetInstance<RPCCLIENT_T>(
-            Attr<int>("trainer_id"));
+        distributed::RPCClient::GetInstance<RPCCLIENT_T>(trainer_id);
 
     std::vector<std::string> recv_varnames =
         Attr<std::vector<std::string>>("recv_varnames");
 
     if (recv_varnames.size() > 0) {
       auto recv_functor = distributed::ParameterRecv<float>();
-      auto rpc_ctx = distributed::RpcContext(outs[0], recv_varnames, epmap, {});
+      auto rpc_ctx = distributed::RpcContext(outs[0], recv_varnames, epmap, {},
+                                             trainer_id);
       recv_functor(rpc_ctx, scope);
     } else {
       if (with_barrier) {

paddle/fluid/operators/load_op.cc

@@ -29,7 +29,7 @@ class LoadOp : public framework::OperatorWithKernel {
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
     framework::OpKernelType kt = framework::OpKernelType(
-        framework::proto::VarType::FP32, platform::CPUPlace());
+        framework::proto::VarType::FP32, ctx.GetPlace());
     return kt;
   }
 };