Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into linspace

CMakeLists.txt

@@ -64,6 +64,7 @@ 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)
@@ -190,6 +191,7 @@ include(configure)          # add paddle env configuration
 if(WITH_GPU)
     include(cuda)
     include(tensorrt)
+    include(anakin_subgraph)
 endif()
 if(WITH_MKL OR WITH_MKLML)
     include(external/anakin)

CONTRIBUTING.md

@@ -156,7 +156,7 @@ python \
 
 This will enable VLOG messages generated by `buddy_allocator.{h,cc}` and in the verbose range of 0 to 3, so you will see above example VLOG message, which is in level 3.  This suggests that we output overall messages in lower verbose levels, so they display with higher probability.  When coding C++, please follow the verbose level convention as follows:
 
-- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework)
-- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)
-- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform)
-- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/math)
+- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/framework)
+- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators)
+- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/platform)
+- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators/math/)

cmake/anakin_subgraph.cmake

@@ -0,0 +1,32 @@
+if(NOT WITH_GPU)
+    return()
+endif()
+
+set(ANAKIN_ROOT "/usr" CACHE PATH "ANAKIN ROOT")
+find_path(ANAKIN_INCLUDE_DIR anakin_config.h
+    PATHS ${ANAKIN_ROOT} ${ANAKIN_ROOT}/include
+    $ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/include
+    NO_DEFAULT_PATH
+)
+
+find_library(ANAKIN_LIBRARY NAMES libanakin_saber_common.so libanakin.so
+    PATHS ${ANAKIN_ROOT}
+    $ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/lib
+    NO_DEFAULT_PATH
+    DOC "Path to ANAKIN library.")
+
+if(ANAKIN_INCLUDE_DIR AND ANAKIN_LIBRARY)
+  if(WITH_DSO)
+    set(ANAKIN_FOUND ON)
+  endif(WITH_DSO)
+else()
+    set(ANAKIN_FOUND OFF)
+endif()
+
+if(ANAKIN_FOUND)
+    message(STATUS "Current ANAKIN header is ${ANAKIN_INCLUDE_DIR}/anakin_config.h. ")
+    include_directories(${ANAKIN_ROOT}/include)
+    include_directories(${ANAKIN_ROOT}/include/saber)
+    link_directories(${ANAKIN_ROOT})
+    add_definitions(-DPADDLE_WITH_ANAKIN)
+endif()

cmake/external/boost.cmake

@@ -24,7 +24,7 @@ set(BOOST_PROJECT       "extern_boost")
 # So we use 1.41.0 here.
 set(BOOST_VER           "1.41.0")
 set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
-set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
+set(BOOST_URL "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
 
 MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
 

cmake/external/grpc.cmake

@@ -44,7 +44,7 @@ ExternalProject_Add(
     # 3. keep only zlib, cares, protobuf, boringssl under "third_party",
     #    checkout and clean other dirs under third_party
     # 4. remove .git, and package the directory.
-    URL "http://paddlepaddledeps.cdn.bcebos.com/grpc-v1.10.x.tar.gz"
+    URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.10.x.tar.gz"
     URL_MD5  "1f268a2aff6759839dccd256adcc91cf"
     PREFIX          ${GRPC_SOURCES_DIR}
     UPDATE_COMMAND  ""

cmake/external/mklml.cmake

@@ -34,7 +34,7 @@ SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
 SET(TIME_VERSION "2019.0.1.20181227")
 IF(WIN32)
     SET(MKLML_VER "mklml_win_${TIME_VERSION}" CACHE STRING "" FORCE)
-    SET(MKLML_URL "https://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
+    SET(MKLML_URL "https://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
     SET(MKLML_LIB                 ${MKLML_LIB_DIR}/mklml.lib)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5md.lib)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/mklml.dll)
@@ -43,7 +43,7 @@ ELSE()
     #TODO(intel-huying):
     #  Now enable Erf function in mklml library temporarily, it will be updated as offical version later.
     SET(MKLML_VER "Glibc225_vsErf_mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
-    SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
+    SET(MKLML_URL "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
     SET(MKLML_LIB                 ${MKLML_LIB_DIR}/libmklml_intel.so)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5.so)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/libmklml_intel.so)

cmake/tensorrt.cmake

@@ -33,5 +33,6 @@ if(TENSORRT_FOUND)
     message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. "
         "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
     include_directories(${TENSORRT_INCLUDE_DIR})
+    link_directories(${TENSORRT_LIBRARY})
     add_definitions(-DPADDLE_WITH_TENSORRT)
 endif()

paddle/fluid/API.spec

@@ -10,6 +10,9 @@ paddle.fluid.default_startup_program (ArgSpec(args=[], varargs=None, keywords=No
 paddle.fluid.default_main_program (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', '5430f54ab4895f9f47db6bebbaf71659'))
 paddle.fluid.program_guard (ArgSpec(args=['main_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,)), ('document', 'b54f403e57825a1592aece03afe3afb6'))
 paddle.fluid.name_scope (ArgSpec(args=['prefix'], varargs=None, keywords=None, defaults=(None,)), ('document', '0ef753f5cec69fef9ae6ad8b867b33a2'))
+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.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.run (ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False)), ('document', 'f482e93b38b4018796969a2e1dde479d'))
@@ -44,7 +47,7 @@ paddle.fluid.AsyncExecutor.run (ArgSpec(args=['self', 'program', 'data_feed', 'f
 paddle.fluid.AsyncExecutor.save_model (ArgSpec(args=['self', 'save_path'], varargs=None, keywords=None, defaults=None), ('document', 'c8ac0dfcb3b187aba25d03af7fea56b2'))
 paddle.fluid.AsyncExecutor.stop (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '5f23d043607bb5d55e466ec3f578e093'))
 paddle.fluid.CompiledProgram.__init__ (ArgSpec(args=['self', 'program_or_graph'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.CompiledProgram.with_data_parallel (ArgSpec(args=['self', 'loss_name', 'build_strategy', 'exec_strategy', 'share_vars_from'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'e1af7fd53cf868554f312779fc803864'))
+paddle.fluid.CompiledProgram.with_data_parallel (ArgSpec(args=['self', 'loss_name', 'build_strategy', 'exec_strategy', 'share_vars_from', 'places'], varargs=None, keywords=None, defaults=(None, None, None, None, None)), ('document', 'a8c7793803cf976680d9478e378fa356'))
 paddle.fluid.CompiledProgram.with_inference_optimize (ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=None), ('document', '9e5b009d850191a010e859189c127fd8'))
 paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.ExecutionStrategy) -> None
 paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.GradientScaleStrategy, arg0: int) -> None
@@ -58,6 +61,12 @@ paddle.fluid.io.load_params (ArgSpec(args=['executor', 'dirname', 'main_program'
 paddle.fluid.io.load_persistables (ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)), ('document', '28df5bfe26ca7a077f91156abb0fe6d2'))
 paddle.fluid.io.save_inference_model (ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename', 'export_for_deployment'], varargs=None, keywords=None, defaults=(None, None, None, True)), ('document', '70f4f53f13572436ac72d1c8b5efeb9d'))
 paddle.fluid.io.load_inference_model (ArgSpec(args=['dirname', 'executor', 'model_filename', 'params_filename', 'pserver_endpoints'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '7a5255386075dac3c75b7058254fcdcb'))
+paddle.fluid.io.PyReader.__init__ (ArgSpec(args=['self', 'feed_list', 'capacity', 'use_double_buffer', 'iterable'], varargs=None, keywords=None, defaults=(True, False)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.io.PyReader.decorate_batch_generator (ArgSpec(args=['self', 'reader', 'places'], varargs=None, keywords=None, defaults=(None,)), ('document', 'a3fefec8bacd6ce83f49906a9d05e779'))
+paddle.fluid.io.PyReader.decorate_sample_generator (ArgSpec(args=['self', 'sample_generator', 'batch_size', 'drop_last', 'places'], varargs=None, keywords=None, defaults=(True, None)), ('document', '7abd9cf7d695bab5bb6cf7ded5903cb2'))
+paddle.fluid.io.PyReader.decorate_sample_list_generator (ArgSpec(args=['self', 'reader', 'places'], varargs=None, keywords=None, defaults=(None,)), ('document', 'faef298f73e91aedcfaf5d184f3109b7'))
+paddle.fluid.io.PyReader.reset (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'ff1cc1e2beb8824d453656c72c28ddfb'))
+paddle.fluid.io.PyReader.start (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'b7ea0a548991924e4cfe61a577b8e56d'))
 paddle.fluid.initializer.ConstantInitializer.__init__ (ArgSpec(args=['self', 'value', 'force_cpu'], varargs=None, keywords=None, defaults=(0.0, False)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.initializer.UniformInitializer.__init__ (ArgSpec(args=['self', 'low', 'high', 'seed'], varargs=None, keywords=None, defaults=(-1.0, 1.0, 0)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.initializer.NormalInitializer.__init__ (ArgSpec(args=['self', 'loc', 'scale', 'seed'], varargs=None, keywords=None, defaults=(0.0, 1.0, 0)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -222,6 +231,7 @@ paddle.fluid.layers.teacher_student_sigmoid_loss (ArgSpec(args=['input', 'label'
 paddle.fluid.layers.huber_loss (ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=None), ('document', '431a4301c35032166ec029f7432c80a7'))
 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.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'))
 paddle.fluid.layers.read_file (ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None), ('document', 'b0a1c2fc51c27a106da28f3308c41f5e'))
@@ -229,7 +239,7 @@ paddle.fluid.layers.shuffle (ArgSpec(args=['reader', 'buffer_size'], varargs=Non
 paddle.fluid.layers.batch (ArgSpec(args=['reader', 'batch_size'], varargs=None, keywords=None, defaults=None), ('document', 'f563d376d35e1a4c4db100fd11b381a0'))
 paddle.fluid.layers.double_buffer (ArgSpec(args=['reader', 'place', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', '07e5b796674796eb1ef3fee9c10d24e3'))
 paddle.fluid.layers.random_data_generator (ArgSpec(args=['low', 'high', 'shapes', 'lod_levels', 'for_parallel'], varargs=None, keywords=None, defaults=(True,)), ('document', '9b7f0f86ec24bbc97643cadcb6499cff'))
-paddle.fluid.layers.py_reader (ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True)), ('document', '13dabc57863f62ab3141586784ee356b'))
+paddle.fluid.layers.py_reader (ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True)), ('document', '4357643685cfd65454ba5a15f0151709'))
 paddle.fluid.layers.create_py_reader_by_data (ArgSpec(args=['capacity', 'feed_list', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, True)), ('document', '350f74d93fab9adb2ac4950f1c26416b'))
 paddle.fluid.layers.Preprocessor.__init__ (ArgSpec(args=['self', 'reader', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.Preprocessor.block (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -511,6 +521,8 @@ paddle.fluid.unique_name.guard (ArgSpec(args=['new_generator'], varargs=None, ke
 paddle.fluid.recordio_writer.convert_reader_to_recordio_file (ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', '65c7523e86f0c50bb729b01667f36310'))
 paddle.fluid.recordio_writer.convert_reader_to_recordio_files (ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', 'bc643f0f5f1b9db57ff0d8a57d379bd7'))
 paddle.fluid.Scope Scope() -> paddle.fluid.core._Scope
+paddle.fluid.install_check.run_check (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', '66b7c84a17ed32fec2df9628367be2b9'))
+paddle.reader.cache (ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None), ('document', '1676886070eb607cb608f7ba47be0d3c'))
 paddle.reader.map_readers (ArgSpec(args=['func'], varargs='readers', keywords=None, defaults=None), ('document', '77cbadb09df588e21e5cc0819b69c87d'))
 paddle.reader.buffered (ArgSpec(args=['reader', 'size'], varargs=None, keywords=None, defaults=None), ('document', '0d6186f109feceb99f60ec50a0a624cb'))
 paddle.reader.compose (ArgSpec(args=[], varargs='readers', keywords='kwargs', defaults=None), ('document', '884291104e1c3f37f33aae44b7deeb0d'))

paddle/fluid/framework/CMakeLists.txt

@@ -63,7 +63,7 @@ cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto
 cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
 nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
 
-cc_library(garbage_collector SRCS garbage_collector.cc DEPS device_context memory)
+cc_library(garbage_collector SRCS garbage_collector.cc DEPS device_context memory gflags glog)
 
 cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
 cc_test(reader_test SRCS reader_test.cc DEPS reader)
@@ -164,6 +164,8 @@ else()
   set(NGRAPH_EXE_DEPS)
 endif()
 
+cc_library(executor_gc_helper SRCS executor_gc_helper.cc DEPS scope proto_desc operator garbage_collector)
+
 if(WITH_DISTRIBUTE)
   cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog
     lod_rank_table feed_fetch_method sendrecvop_rpc  ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS})
@@ -174,7 +176,7 @@ else()
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
 
-target_link_libraries(executor garbage_collector while_op_helper)
+target_link_libraries(executor while_op_helper executor_gc_helper)
 
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
         threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor
@@ -194,6 +196,7 @@ cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_con
 cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
         proto_desc)
 cc_test(inplace_op_inference_test SRCS inplace_op_inference_test.cc DEPS op_registry proto_desc op_info memory_optimize_helper)
+
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 

paddle/fluid/framework/details/CMakeLists.txt

@@ -5,6 +5,7 @@ cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_h
 cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
 cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
 cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry)
+cc_library(fetch_barrier_op_handle SRCS fetch_barrier_op_handle.cc DEPS framework_proto scope place operator op_registry)
 
 cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper)
 cc_library(multi_devices_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper)
@@ -72,7 +73,7 @@ cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS grap
 cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
 
 cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
-        scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
+        scale_loss_grad_op_handle rpc_op_handle fetch_barrier_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
 
 cc_library(fuse_all_reduce_op_pass SRCS fuse_all_reduce_op_pass.cc DEPS graph graph_helper fused_all_reduce_op_handle)
 

paddle/fluid/framework/details/all_reduce_deps_pass.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include <algorithm>
+#include <memory>
 #include <string>
 #include <unordered_map>
 #include <unordered_set>
@@ -52,13 +53,28 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
   //               Note that must assert topology sort is stable
   auto& ops = graph->Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs);
   for (auto* op_desc : ops) {
-    auto outputs = op_desc->Outputs();
-    for (auto& o_it : outputs) {
-      for (auto& v : o_it.second) {  // values
-        vars[v] = order;
+    try {
+      bool is_bk_op =
+          static_cast<bool>(boost::get<int>(op_desc->GetAttr(
+                                OpProtoAndCheckerMaker::OpRoleAttrName())) &
+                            static_cast<int>(OpRole::kBackward));
+      if (!is_bk_op) continue;
+
+      auto backward_vars =
+          boost::get<std::vector<std::string>>(op_desc->GetNullableAttr(
+              OpProtoAndCheckerMaker::OpRoleVarAttrName()));
+      PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
+
+      auto outputs = op_desc->Outputs();
+      for (auto& o_it : outputs) {
+        for (auto& v : o_it.second) {  // values
+          vars[v] = order;
+          VLOG(1) << "in all_reduce_deps_pass:" << v;
+        }
       }
+      order++;
+    } catch (boost::bad_get e) {
     }
-    order++;
   }
 
   std::vector<OpHandleBase*> dist_ops;

paddle/fluid/framework/details/eager_deletion_pass.cc

@@ -22,14 +22,9 @@
 #include "paddle/fluid/framework/details/computation_op_handle.h"
 #include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/garbage_collector.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 
-DEFINE_double(memory_fraction_of_eager_deletion, 1.0,
-              "Fraction of eager deletion. If less than 1.0, all variables in "
-              "the program would be sorted according to its memory size, and "
-              "only the FLAGS_memory_fraction_of_eager_deletion of the largest "
-              "variables would be deleted.");
-
 namespace paddle {
 namespace framework {
 namespace details {
@@ -206,8 +201,9 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     }
   }
 
-  op_vars_map = ShrinkGCVars(op_vars_map, vars, places,
-                             FLAGS_memory_fraction_of_eager_deletion);
+  double memory_fraction = framework::GetEagerDeletionMemoryFraction();
+
+  op_vars_map = ShrinkGCVars(op_vars_map, vars, places, memory_fraction);
 
   for (auto &pair : op_vars_map) {
     auto *op = pair.first;
@@ -239,8 +235,7 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     eager_deletion_op->AddOutput(dummy_leaf);
   }
 
-  VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = "
-           << FLAGS_memory_fraction_of_eager_deletion;
+  VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = " << memory_fraction;
   VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
 
   auto while_op_eager_deletion_pass =

paddle/fluid/framework/details/early_delete_op_handle.h

@@ -1,140 +0,0 @@
-// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#pragma once
-#include <string>
-#include <vector>
-#include "paddle/fluid/framework/details/computation_op_handle.h"
-#include "paddle/fluid/framework/details/op_handle_base.h"
-#include "paddle/fluid/framework/details/var_handle.h"
-#include "paddle/fluid/framework/garbage_collector.h"
-#include "paddle/fluid/framework/lod_tensor_array.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/framework/selected_rows.h"
-#include "paddle/fluid/framework/tensor.h"
-
-namespace paddle {
-namespace framework {
-namespace details {
-
-class EarlyDeleteOpHandle : public OpHandleBase {
- public:
-  EarlyDeleteOpHandle(ir::Node* node, const Scope* scope,
-                      const platform::Place& place,
-                      const std::vector<std::string>& names,
-                      GarbageCollector* gc)
-      : OpHandleBase(node),
-        scope_(scope),
-        place_(place),
-        names_(names),
-        gc_(gc) {
-#ifdef PADDLE_WITH_CUDA
-    if (IsStreamGarabageCollector()) {
-      auto gpu_place = boost::get<platform::CUDAPlace>(place);
-      PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
-      PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
-    }
-#endif
-  }
-  ~EarlyDeleteOpHandle() {
-#ifdef PADDLE_WITH_CUDA
-    if (IsStreamGarabageCollector()) {
-      auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
-      PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
-      PADDLE_ENFORCE(cudaEventDestroy(event_));
-    }
-#endif
-  }
-
-  std::string Name() const override { return "early_delete"; }
-
- protected:
-  void RunImpl() override {
-    std::vector<std::shared_ptr<memory::Allocation>> tensors;
-    auto* local_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope*>();
-    for (auto& var_name : names_) {
-      auto* var = local_scope->FindVar(var_name);
-      PADDLE_ENFORCE(var != nullptr,
-                     string::Sprintf("Local Scope not has var %s", var_name));
-      if (var->IsType<LoDTensor>()) {
-        tensors.emplace_back(var->GetMutable<LoDTensor>()->MoveMemoryHolder());
-      } else if (var->IsType<SelectedRows>()) {
-        tensors.emplace_back(var->GetMutable<SelectedRows>()
-                                 ->mutable_value()
-                                 ->MoveMemoryHolder());
-      } else if (var->IsType<LoDTensorArray>()) {
-        LoDTensorArray* tensor_array = var->GetMutable<LoDTensorArray>();
-        for (auto& tensor : *tensor_array) {
-          tensors.emplace_back(tensor.MoveMemoryHolder());
-        }
-      }
-    }
-    if (!tensors.empty()) {
-      ClearTensors(tensors);
-    }
-  }
-
- private:
-  void ClearTensors(
-      const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
-    if (platform::is_cpu_place(place_)) {
-      ClearCPUTensors(tensors);
-    } else {
-      ClearGPUTensors(tensors);
-    }
-  }
-
-  void ClearCPUTensors(
-      const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
-    auto* gc = dynamic_cast<CPUGarbageCollector*>(gc_);
-    if (gc != nullptr) {
-      gc->Add(tensors);
-    }
-  }
-
-  void ClearGPUTensors(
-      const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
-#ifdef PADDLE_WITH_CUDA
-    auto* gc = dynamic_cast<StreamGarbageCollector*>(gc_);
-    if (gc != nullptr) {
-      auto compute_stream = dev_ctx_->stream();
-      auto callback_stream = gc->stream();
-      auto callback_func = [=]() {
-        PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream));
-        PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0));
-      };
-      gc_->Add(tensors, callback_func);
-    } else {
-      gc_->Add(tensors);
-    }
-  }
-
-  bool IsStreamGarabageCollector() const {
-    return dynamic_cast<const StreamGarbageCollector*>(gc_) != nullptr;
-#endif
-  }
-
-  const Scope* scope_;
-  const platform::Place place_;
-  std::vector<std::string> names_;
-  GarbageCollector* gc_;
-#ifdef PADDLE_WITH_CUDA
-  platform::CUDADeviceContext* dev_ctx_;
-  cudaEvent_t event_;
-#endif
-};
-
-}  // namespace details
-}  // namespace framework
-}  // namespace paddle

paddle/fluid/framework/details/fetch_barrier_op_handle.cc

@@ -0,0 +1,66 @@
+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/details/fetch_barrier_op_handle.h"
+
+#include <string>
+
+namespace paddle {
+namespace framework {
+namespace details {
+FetchBarrierOpHandle::FetchBarrierOpHandle(
+    ir::Node *node, const std::vector<Scope *> &local_scopes,
+    const std::vector<platform::Place> &places)
+    // fetch_barrier op always run on place0, but output on all places.
+    : OpHandleBase(node),
+      op_(framework::OpRegistry::CreateOp(*node->Op())),
+      local_scopes_(local_scopes),
+      places_(places),
+      run_scope_(local_scopes[0]),
+      place_(places[0]) {
+  for (auto &p : places) {
+    this->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p));
+  }
+}
+
+bool FetchBarrierOpHandle::IsMultiDeviceTransfer() {
+  // override IsMultiDeviceTransfer to return true
+  return true;
+}
+
+void FetchBarrierOpHandle::RunImpl() {
+  WaitInputVarGenerated(place_);
+
+  auto run_func = [this]() {
+    op_->Run(*run_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(), place_);
+  };
+
+  if (is_lock_and_record_event_free_) {
+    run_func();
+  } else {
+    this->RunAndRecordEvent(run_func);
+  }
+}
+
+bool FetchBarrierOpHandle::NeedWait(VarHandleBase *in_var) {
+  bool need_wait =
+      in_var && in_var->GeneratedOp() &&
+      in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_.at(place_);
+  return need_wait;
+}
+
+std::string FetchBarrierOpHandle::Name() const { return op_->Type(); }
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/fetch_barrier_op_handle.h

@@ -0,0 +1,61 @@
+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/details/op_handle_base.h"
+#include "paddle/fluid/framework/feed_fetch_type.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+// **NOTE**: fetch_barrier op is special it outputs all recved variables on
+// all places if there are multiple places, must init with
+// multiple dev_ctxes_ !!!!
+
+struct FetchBarrierOpHandle : public OpHandleBase {
+ public:
+  FetchBarrierOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
+                       const std::vector<platform::Place> &places);
+
+  bool IsMultiDeviceTransfer() override;
+
+  std::string Name() const override;
+
+ protected:
+  void RunImpl() override;
+
+  bool NeedWait(VarHandleBase *in_var) override;
+
+ private:
+  std::unique_ptr<OperatorBase> op_;
+  std::vector<Scope *> local_scopes_;
+  std::vector<platform::Place> places_;
+  Scope *run_scope_;
+  platform::Place place_;
+
+  bool is_lock_and_record_event_free_{false};
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/inplace_op_pass.cc

@@ -17,6 +17,8 @@
 #include <deque>
 #include <iterator>
 #include <memory>
+#include <queue>
+#include <sstream>
 #include <stack>
 #include <string>
 #include <unordered_map>
@@ -148,12 +150,14 @@ std::unique_ptr<ir::Graph> InplacePass::ApplyImpl(
   view_.Build(graph.get());
   InitSSAGraphNodes();
 
+  auto cnt = 0;
   for (auto* op : view_.AllOps()) {
+    VLOG(4) << "Handle op " << cnt++ << ": " << op->Name();
     if (FLAGS_enable_inplace_whitelist && !whitelist_.count(op->Name()))
       continue;
     TryInplaceOpInputOutput(op, graph.get());
   }
-  graph->ResolveHazard(var_nodes_);
+  // graph->ResolveHazard(var_nodes_);
 
   return graph;
 }
@@ -264,13 +268,10 @@ void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
 void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
                                           ir::Graph* graph) const {
   VLOG(4) << "Try to inplace op " << op->Name();
-  // FIXME(liuwei1031): Graph is not aware of the existence of BlockDescs and
-  // ProgramDescs.
-  // The operations related to BlockDesc or ProgramDesc should perform on Graph
-  // or Node directly!
-  PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
-                 "op_desc is nullptr");
+  // PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
+  //               "op_desc is nullptr");
   // some pre-requirments need to meet if the op want to inplaced.
+  PADDLE_ENFORCE(op->Op() != nullptr, "op_desc is nullptr");
 
   auto* op_desc = op->Op();
   auto& infer_inplace =
@@ -281,21 +282,58 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
   PADDLE_ENFORCE(static_cast<bool>(infer_inplace),
                  "%s's infer_inplace has not been registered", op_desc->Type());
 
-  auto* block = op_desc->Block();
-  auto in_to_outs = infer_inplace(*op_desc, block);
+  auto in_to_outs = infer_inplace(*op_desc);
 
   auto& all_ops = view_.AllOps();
   auto cursor = std::find(all_ops.begin(), all_ops.end(), op);
   size_t idx = std::distance(all_ops.begin(), cursor);
 
   for (auto& pair : in_to_outs) {
-    auto& in_var_name = pair.first;
-    auto& out_var_name = pair.second;
+    auto& in_para_name = pair.first;
+    auto& out_para_name = pair.second;
+
+    auto input_vars = op->Op()->Input(in_para_name);
+    if (!input_vars.size()) {
+      VLOG(4) << "Parameter " << in_para_name << " is empty skip "
+              << in_para_name << " => " << out_para_name << " pair";
+      continue;
+    }
+    auto output_vars = op->Op()->Output(out_para_name);
+    if (!output_vars.size()) {
+      VLOG(4) << "Parameter " << out_para_name << " is empty skip "
+              << in_para_name << " => " << out_para_name << " pair";
+      continue;
+    }
+    auto in_var_name = input_vars.at(0);
+    auto out_var_name = output_vars.at(0);
     auto* in_node = view_.GetNodeByName(in_var_name, op->inputs);
     auto* out_node = view_.GetNodeByName(out_var_name, op->outputs);
 
+    VLOG(4) << "Try to inplace " << in_var_name << " with " << out_var_name;
+
+    bool can_replace = true;
+    if (in_var_name == out_var_name) {
+      can_replace = false;
+      VLOG(4) << "SKIP: Input variable " << in_var_name << " & Output variable "
+              << out_var_name << " are the same";
+    } else if (!NodeCanReused(in_node)) {
+      can_replace = false;
+      VLOG(4) << "SKIP: Input varialbe " << in_var_name << "cannot be reused";
+    } else if (!NodeCanReused(out_node)) {
+      can_replace = false;
+      VLOG(4) << "SKIP: Output variable " << out_var_name
+              << " cannot be reused";
+    } else if (details::NodeSize(*in_node->Var()) !=
+               details::NodeSize(*out_node->Var())) {
+      can_replace = false;
+      VLOG(4) << "SKIP: Input and Output varialbe size not match";
+    }
+
+    if (!can_replace) continue;
+
     // 2. there is no external pending op on the input node
-    if (view_.PendingOpsOnVar(in_node).size() > 1) {
+    // if (view_.PendingOpsOnVar(in_node).size() > 1) {
+    if (in_node->outputs.size() > 1 && !view_.CheckDeps(in_node, op)) {
       VLOG(4) << string::Sprintf(
           "Skiped pair %s => %s. %s input has external dependency."
           "inplace such pair will overwrite the memory.",
@@ -342,6 +380,97 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
   }
 }
 
+void GraphView::TopoSort(ir::Graph* graph) {
+  //
+  ops_.clear();
+  auto deps_num = [](ir::Node* op) {
+    auto cnt = 0;
+    for (auto& var : op->inputs)
+      if (var->inputs.size() > 0) ++cnt;
+    return cnt;
+  };
+
+  std::queue<std::pair<ir::Node*, uint32_t>> ready_ops;
+
+  int level = 0;
+  auto nodes = graph->Nodes();
+  std::unordered_map<ir::Node*, uint32_t> deps_map;
+  for (auto& node : nodes) {
+    if (node->IsOp() && node->Op() != nullptr) {
+      deps_map[node] = deps_num(node);
+      if (0 == deps_map[node]) {
+        ready_ops.push({node, level});
+      }
+    }
+  }
+
+  while (!ready_ops.empty()) {
+    auto item = ready_ops.front();
+    ready_ops.pop();
+
+    ops_.emplace_back(item.first);
+    // record level when pop from queue
+    op_level_[item.first] = item.second;
+
+    for (auto node : item.first->outputs) {
+      for (auto op : node->outputs) {
+        --deps_map[op];
+        if (deps_map[op] == 0) ready_ops.push({op, item.second + 1});
+      }
+    }
+  }
+
+  bool all_ops_checked = true;
+  for (auto& node : nodes) {
+    if (node->IsOp() && node->Op() != nullptr && deps_map[node] > 0) {
+      all_ops_checked = false;
+      break;
+    }
+  }
+
+  PADDLE_ENFORCE(all_ops_checked, "All ops deps should be 0 after analysis");
+}
+
+// return true if current op node depeneds on all other op that use the same
+// variable node
+bool GraphView::CheckDeps(ir::Node* var, ir::Node* current_op) const {
+  // get op list that rely on the same variable
+  auto op_list = var->outputs;
+  for (auto& op : op_list) {
+    if (op == current_op) continue;
+
+    VLOG(4) << "    GraphView::CheckDeps : " << op->Name() << "  & "
+            << current_op->Name();
+    if (!CheckOpDeps(op, current_op)) return false;
+    VLOG(4) << "";
+  }
+  return true;
+}
+
+// check if op2 depends on op1's output
+bool GraphView::CheckOpDeps(ir::Node* op1, ir::Node* op2) const {
+  auto print_op = [&](ir::Node* op, const char* name) {
+    std::ostringstream os;
+    os << "        " << name << " : " << op->Name() << " ";
+    os << "Input args : ";
+    for (auto& arg : op->inputs) os << arg->Name() << " ";
+    os << "Output args : ";
+    for (auto& arg : op->outputs) os << arg->Name() << " ";
+    os << "Level : " << op_level_.at(op);
+    VLOG(4) << os.str();
+  };
+  print_op(op1, "OP1");
+  print_op(op2, "OP2");
+
+  if (op1 == op2) return true;
+  if (op_level_.at(op1) >= op_level_.at(op2)) return false;
+
+  for (auto& var : op2->inputs)
+    if (var->inputs.size() > 0 && CheckOpDeps(op1, var->inputs[0])) return true;
+
+  return false;
+}
+
 ir::Node* GraphView::GetNodeByName(const std::string& name,
                                    const std::vector<ir::Node*>& nodes) const {
   // nodes should be op->inputs/outputs
@@ -387,22 +516,7 @@ void GraphView::Build(ir::Graph* g) {
   // Because we insert some new created node. Which may have data race between
   // nodes.
   // resolve data harzards depends on the var nodes in right order.
-  ops_ = SortOpLikeDescOrder(*g);
-
-  // 1. track the nodes which reused previous node in Python memory optimize.
-  // these node can not be inplaced, otherwise may generate a circle in graph.
-  std::unordered_set<std::string> all_vars;
-  for (auto& node : g->Nodes()) {
-    if (node->IsVar()) continue;
-    for (auto& out : node->outputs) {
-      if (out->IsCtrlVar() || out->Var() == nullptr) continue;
-      if (all_vars.count(out->Name())) {
-        dup_nodes_.emplace(out->Name());
-      } else {
-        all_vars.emplace(out->Name());
-      }
-    }
-  }
+  TopoSort(g);
 
   // 2. track the nodes which used by parameter server.
   // these node can not be inplaced, otherwise trainer

paddle/fluid/framework/details/inplace_op_pass.h

@@ -14,6 +14,7 @@
 
 #pragma once
 #include <map>
+#include <memory>
 #include <string>
 #include <unordered_map>
 #include <unordered_set>
@@ -50,10 +51,15 @@ class GraphView {
   // map the parameter and gradient, must be skipped.
   bool InSkipSet(const std::string& var) const;
 
+  bool CheckDeps(ir::Node* var, ir::Node* current_op) const;
+  bool CheckOpDeps(ir::Node* op1, ir::Node* op2) const;
+  void TopoSort(ir::Graph* g);
+
  private:
   std::vector<ir::Node*> ops_;
   std::unordered_set<std::string> dup_nodes_;  // mem opt affect nodes
   std::map<ir::Node*, std::unordered_set<ir::Node*>> adj_list_;
+  std::unordered_map<ir::Node*, uint32_t> op_level_;
 };
 
 // swap pairs in sequence

paddle/fluid/framework/details/memory_optimize_helper.cc

@@ -190,7 +190,7 @@ struct NodeComparator {
     auto rhs_shape = rhs_desc->GetShape();
     if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) ||
         (lhs_shape[0] != -1 && rhs_shape[0] != -1)) {
-      return NodeSize(lhs) <= NodeSize(rhs);
+      return NodeSize(lhs) == NodeSize(rhs);
     } else {
       return false;
     }
@@ -449,6 +449,7 @@ void ControlFlowGraph::LiveVariableAnalysis() {
       live_in_[op].insert(var);
     }
     for (auto& var : defs_[op]) {
+      if (uses_[op].count(var)) continue;
       live_in_[op].erase(var);
     }
 

paddle/fluid/framework/details/memory_optimize_helper_test.cc

@@ -142,15 +142,16 @@ TEST(OrderedSet, FindBestFitNode) {
   for (auto& node : nodes) {
     pool.Insert(node.get());
   }
-
+  // FIXME(liuwei1031) this API has changed,
+  // disable these tests temporarily
   // FindNextBestFitNode
-  auto* n = nodes[0].get();
-  auto* cache = pool.FindBestFitNode(n);
-  PADDLE_ENFORCE(cache->Name() == "a");
-  cache = pool.FindNextBestFitNode(n, cache);
-  PADDLE_ENFORCE(cache->Name() == "c");
-  cache = pool.FindNextBestFitNode(n, cache);
-  PADDLE_ENFORCE(cache->Name() == "b");
+  // auto* n = nodes[0].get();
+  // auto* cache = pool.FindBestFitNode(n);
+  // PADDLE_ENFORCE(cache->Name() == "a");
+  // cache = pool.FindNextBestFitNode(n, cache);
+  // PADDLE_ENFORCE(cache->Name() == "c");
+  // cache = pool.FindNextBestFitNode(n, cache);
+  // PADDLE_ENFORCE(cache->Name() == "b");
 }
 
 }  // namespace details

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -23,6 +23,7 @@
 #include "paddle/fluid/framework/details/all_reduce_op_handle.h"
 #include "paddle/fluid/framework/details/broadcast_op_handle.h"
 #include "paddle/fluid/framework/details/computation_op_handle.h"
+#include "paddle/fluid/framework/details/fetch_barrier_op_handle.h"
 #include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
 #include "paddle/fluid/framework/details/reduce_op_handle.h"
 #include "paddle/fluid/framework/details/rpc_op_handle.h"
@@ -851,9 +852,17 @@ int DistSSAGraphBuilder::CreateRPCOp(ir::Graph *result, ir::Node *node) const {
 
   PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s",
                  node->Op()->Type());
-  result->Get<GraphOps>(kGraphOps).emplace_back(new RPCOpHandle(
-      result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
-      node->Op()->Type(), places_[op_dev_id]));
+
+  // Create fetch_barrier op handle to enable output on all devices.
+  // **NOTE** fetch_barrier should output variables list same as recv op does.
+  if (node->Op()->Type() == "fetch_barrier") {
+    result->Get<GraphOps>(kGraphOps).emplace_back(new FetchBarrierOpHandle(
+        result->CreateOpNode(node->Op()), local_scopes_, places_));
+  } else {
+    result->Get<GraphOps>(kGraphOps).emplace_back(new RPCOpHandle(
+        result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
+        node->Op()->Type(), places_[op_dev_id]));
+  }
 
   if (node->Op()->Type() == "send") {
     CreateOpHandleIOs(result, node, op_dev_id);

paddle/fluid/framework/details/op_handle_base.cc

@@ -55,7 +55,7 @@ void OpHandleBase::Run(bool use_cuda) {
         if (out_var_handle) {
           int dev_id =
               boost::get<platform::CUDAPlace>(out_var_handle->place()).device;
-          out_var_handle->SetGenerateEvent(events_[dev_id]);
+          out_var_handle->SetGenerateEvent(events_.at(dev_id));
         }
       }
     } else {
@@ -71,7 +71,7 @@ void OpHandleBase::Run(bool use_cuda) {
               "The place of input(%s) is not consistent with the "
               "place of current op(%s).",
               out_var_handle->Name(), Name());
-          out_var_handle->SetGenerateEvent(events_[dev_id]);
+          out_var_handle->SetGenerateEvent(events_.at(dev_id));
         }
       }
     }

paddle/fluid/framework/details/op_registry.h

@@ -21,6 +21,7 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/framework/grad_op_desc_maker.h"
 #include "paddle/fluid/framework/inplace_op_inference.h"
+#include "paddle/fluid/framework/no_need_buffer_vars_inference.h"
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/operator.h"
@@ -36,27 +37,86 @@ enum OpInfoFillType {
   kGradOpDescMaker = 2,
   kVarTypeInference = 3,
   kShapeInference = 4,
-  kInplaceOpInference = 5
+  kInplaceOpInference = 5,
+  kNoNeedBufferVarsInference = 6,
+  kUnknown = -1
 };
 
+namespace internal {
+template <typename T, OpInfoFillType kType>
+struct TypePair {
+  using Type = T;
+  static constexpr OpInfoFillType kFillType = kType;
+};
+
+using OpRegistryClasses = std::tuple<                                // NOLINT
+    TypePair<OperatorBase, kOperator>,                               // NOLINT
+    TypePair<OpProtoAndCheckerMaker, kOpProtoAndCheckerMaker>,       // NOLINT
+    TypePair<GradOpDescMakerBase, kGradOpDescMaker>,                 // NOLINT
+    TypePair<VarTypeInference, kVarTypeInference>,                   // NOLINT
+    TypePair<InferShapeBase, kShapeInference>,                       // NOLINT
+    TypePair<InplaceOpInference, kInplaceOpInference>,               // NOLINT
+    TypePair<NoNeedBufferVarsInference, kNoNeedBufferVarsInference>  // NOLINT
+    >;
+
+static constexpr int kOpRegistryClassNumber =
+    std::tuple_size<OpRegistryClasses>::value;
+
+template <typename T, int kPos, bool kIsBounded /* = true*/>
+struct IsMatchedBaseTypeImpl {
+  using PairType = typename std::tuple_element<kPos, OpRegistryClasses>::type;
+  static constexpr bool kValue =
+      std::is_base_of<typename PairType::Type, T>::value;
+};
+
+template <typename T, int kPos>
+struct IsMatchedBaseTypeImpl<T, kPos, false> {
+  static constexpr bool kValue = false;
+};
+
+template <typename T, int kPos>
+static inline constexpr bool IsMatchedBaseType() {
+  return IsMatchedBaseTypeImpl<
+      T, kPos, (kPos >= 0 && kPos < kOpRegistryClassNumber)>::kValue;
+}
+
+template <typename T, int kStart, int kEnd, bool kIsEnd, bool kIsMatched>
+struct OpInfoFillTypeGetterImpl {};
+
+// This case should not happen
+template <typename T, int kStart, int kEnd>
+struct OpInfoFillTypeGetterImpl<T, kStart, kEnd, true, true> {};
+
+template <typename T, int kStart, int kEnd>
+struct OpInfoFillTypeGetterImpl<T, kStart, kEnd, true, false> {
+  static constexpr OpInfoFillType kType = kUnknown;
+};
+
+template <typename T, int kStart, int kEnd>
+struct OpInfoFillTypeGetterImpl<T, kStart, kEnd, false, false> {
+  static constexpr OpInfoFillType kType =
+      OpInfoFillTypeGetterImpl<T, kStart + 1, kEnd, kStart + 1 == kEnd,
+                               IsMatchedBaseType<T, kStart + 1>()>::kType;
+};
+
+template <typename T, int kStart, int kEnd>
+struct OpInfoFillTypeGetterImpl<T, kStart, kEnd, false, true> {
+  using PairType = typename std::tuple_element<kStart, OpRegistryClasses>::type;
+  static constexpr OpInfoFillType kType = PairType::kFillType;
+};
+
+template <typename T>
+using OpInfoFillTypeGetter =
+    OpInfoFillTypeGetterImpl<T, 0, kOpRegistryClassNumber,
+                             kOpRegistryClassNumber == 0,
+                             IsMatchedBaseType<T, 0>()>;
+
+}  // namespace internal
+
 template <typename T>
 struct OpInfoFillTypeID {
   static constexpr OpInfoFillType ID() {
-    return std::is_base_of<OperatorBase, T>::value
-               ? kOperator
-               : (std::is_base_of<OpProtoAndCheckerMaker, T>::value
-                      ? kOpProtoAndCheckerMaker
-                      : (std::is_base_of<GradOpDescMakerBase, T>::value
-                             ? kGradOpDescMaker
-                             : (std::is_base_of<VarTypeInference, T>::value
-                                    ? kVarTypeInference
-                                    : (std::is_base_of<InferShapeBase, T>::value
-                                           ? kShapeInference
-                                           : (std::is_base_of<
-                                                  InplaceOpInference, T>::value
-                                                  ? kInplaceOpInference
-                                                  : static_cast<OpInfoFillType>(
-                                                        -1))))));
+    return internal::OpInfoFillTypeGetter<T>::kType;
   }
 };
 
@@ -149,9 +209,21 @@ struct OpInfoFiller<T, kShapeInference> {
 template <typename T>
 struct OpInfoFiller<T, kInplaceOpInference> {
   void operator()(const char* op_type, OpInfo* info) const {
-    info->infer_inplace_ = [](const OpDesc& op_desc, BlockDesc* block) {
+    info->infer_inplace_ = [](const OpDesc& op_desc) {
       T infer;
-      return infer(op_desc, block);
+      return infer(op_desc);
+    };
+  }
+};
+
+template <typename T>
+struct OpInfoFiller<T, kNoNeedBufferVarsInference> {
+  void operator()(const char* op_type, OpInfo* info) const {
+    info->infer_no_need_buffer_vars_ = [](const VariableNameMap& inputs,
+                                          const VariableNameMap& outputs,
+                                          const AttributeMap& attrs) {
+      T infer(inputs, outputs, attrs);
+      return infer();
     };
   }
 };

paddle/fluid/framework/details/reference_count_pass.cc

@@ -193,6 +193,79 @@ ExtractComputationOpFromLastLivedVar(VarHandle *var, size_t scope_idx,
   return shrink_func(computation_op);
 }
 
+/**
+ * Shrink op dependencies according to no need buffer vars.
+ *
+ * If some ops do not need Tensor buffer of any input,
+ * just remove the dependency of this op, i.e, decrease reference count.
+ *
+ * For example, input Y of elementwise_add_grad op is only used to infer shape
+ * and lod of Y@GRAD, we do not need the buffer of input Y. Data buffer of
+ * input Y can be collected before elementwise_add_grad op runs.
+ *
+ * This method returns whether the dependency count decreases to 0, and
+ * shrinks op dependency if possible.
+ */
+static bool ShrinkNoNeedBufferVarOpDependency(
+    const std::string &var_name,
+    std::unordered_set<ComputationOpHandle *> *op_handles) {
+  std::vector<ComputationOpHandle *> skip_ops;
+  for (auto *op_handle : *op_handles) {
+    auto *op_base = op_handle->GetOp();
+    auto &inferer = op_base->Info().NoNeedBufferVarsInferer();
+    if (!inferer) {
+      continue;
+    }
+
+    std::unordered_set<std::string> no_need_buffer_vars =
+        inferer(op_base->Inputs(), op_base->Outputs(), op_base->Attrs());
+
+    // Check whether var_name occurs in other inputs or outputs of the op
+    // If it occurs, we cannot decrease the dependency number.
+    bool occurred_in_other_vars = false;
+    for (auto &in_pair : op_base->Inputs()) {
+      if (no_need_buffer_vars.count(in_pair.first) > 0) {
+        continue;
+      }
+
+      auto &args = in_pair.second;
+      auto iter = std::find(args.begin(), args.end(), var_name);
+      if (iter != args.end()) {
+        occurred_in_other_vars = true;
+        break;
+      }
+    }
+
+    if (occurred_in_other_vars) {
+      continue;
+    }
+
+    for (auto &out_pair : op_base->Outputs()) {
+      auto &args = out_pair.second;
+      auto iter = std::find(args.begin(), args.end(), var_name);
+      if (iter != args.end()) {
+        occurred_in_other_vars = true;
+        break;
+      }
+    }
+
+    if (!occurred_in_other_vars) {
+      VLOG(2) << "Shrink var " << var_name << " in op " << op_handle->Name();
+      skip_ops.emplace_back(op_handle);
+    }
+  }
+
+  if (skip_ops.size() == op_handles->size()) {
+    op_handles->clear();
+    return true;
+  } else {
+    for (auto *skip_op : skip_ops) {
+      op_handles->erase(skip_op);
+    }
+    return false;
+  }
+}
+
 std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
   auto &ref_cnts = Get<std::vector<ReferenceCountMap>>(kGlobalReferenceCount);
@@ -229,17 +302,43 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
         continue;
       }
 
-      bool ok;
-      auto result = ExtractComputationOpFromLastLivedVar(
-          name_var_pair.second.back(), i, shrink_func, &ok);
+      auto &var_name = name_var_pair.first;
+      auto &var_handles = name_var_pair.second;
+
+      for (auto iter = var_handles.rbegin(); iter != var_handles.rend();
+           ++iter) {
+        bool ok;
+        auto result =
+            ExtractComputationOpFromLastLivedVar(*iter, i, shrink_func, &ok);
+
+        // Seldomly, some vars may have no pending or preceding computation ops
+        // Just break;
+        if (!ok) break;
+        VLOG(10) << "Extract " << result.size() << " ops of var " << var_name;
+
+        size_t original_op_deps = result.size();
+        // If all ops do not need buffer of var_name, calculate reference count
+        // of the previous version of var_name.
+        if (ShrinkNoNeedBufferVarOpDependency(var_name, &result)) {
+          VLOG(10) << "Try to precede reference count computing at var "
+                   << var_name;
+          continue;
+        }
+
+        size_t final_op_deps = result.size();
+        if (final_op_deps < original_op_deps) {
+          VLOG(5) << "Shrink op deps from " << original_op_deps << " to "
+                  << final_op_deps;
+        }
 
-      if (ok) {
-        auto &var_name = name_var_pair.first;
         PADDLE_ENFORCE(!result.empty(), "Last living ops of %s cannot be empty",
                        var_name);
         ref_cnts[i].emplace(var_name, result.size());
         last_live_ops_of_vars[i].emplace(var_name, std::move(result));
       }
+
+      // Seldomly, all preceding trying failed.
+      // Just skip this corner case
     }
   }
 

paddle/fluid/framework/executor.cc

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <unordered_set>
 #include <utility>
 
+#include "paddle/fluid/framework/executor_gc_helper.h"
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/framework/lod_rank_table.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
@@ -48,97 +49,23 @@ namespace {
 int kProgramId = -1;
 }  // namespace
 
-static std::unordered_map<std::string, size_t> GetNonPersistableReferenceCounts(
-    const BlockDesc& block, const std::vector<std::string>& skip_var_list) {
-  std::unordered_map<std::string, size_t> ref_cnts;
-  std::unordered_set<std::string> skip_vars(skip_var_list.begin(),
-                                            skip_var_list.end());
-
-  auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) {
-    for (auto& name_pair : name_map) {
-      for (auto& name : name_pair.second) {
-        if (skip_vars.count(name)) continue;
-        auto* var_desc = block.FindVar(name);
-        if (var_desc == nullptr || var_desc->Persistable()) continue;
-        auto type = var_desc->Proto()->type().type();
-        if (type != proto::VarType::LOD_TENSOR &&
-            type != proto::VarType::SELECTED_ROWS &&
-            type != proto::VarType::LOD_TENSOR_ARRAY) {
-          continue;
-        }
-        ++ref_cnts[name];
-      }
-    }
-  };
-
-  for (auto op_desc : block.AllOps()) {
-    update_ref_cnts(op_desc, op_desc->Inputs());
-    update_ref_cnts(op_desc, op_desc->Outputs());
-  }
-  return ref_cnts;
-}
-
 ExecutorPrepareContext::ExecutorPrepareContext(
-    const framework::ProgramDesc& prog, size_t block_id,
-    const std::vector<std::string>& keep_vars, bool force_disable_gc)
-    : prog_(prog), block_id_(block_id), force_disable_gc_(force_disable_gc) {
-  if (GetEagerDeletionThreshold() >= 0 && !force_disable_gc_) {
-    global_ref_cnts_ =
-        GetNonPersistableReferenceCounts(prog.Block(block_id), keep_vars);
+    const framework::ProgramDesc& prog, size_t block_id)
+    : prog_(prog), block_id_(block_id) {}
+
+void ExecutorPrepareContext::PrepareUnusedVars(
+    const std::vector<std::string>& keep_vars, bool force_disable_gc) {
+  force_disable_gc_ = force_disable_gc;
+  if (GetEagerDeletionThreshold() < 0 || force_disable_gc_) {
+    return;
   }
+  unused_vars_ = GetUnusedVars(prog_.Block(block_id_), ops_, keep_vars);
 }
 
 ExecutorPrepareContext::~ExecutorPrepareContext() {
   VLOG(5) << "destroy ExecutorPrepareContext";
 }
 
-static void DeleteUnusedTensors(
-    const Scope& scope, const OperatorBase* op, GarbageCollector* gc,
-    std::unordered_map<std::string, size_t>* ref_cnts) {
-  std::deque<std::shared_ptr<memory::Allocation>> garbages;
-
-  auto handler = [&](const VariableNameMap& name_map) {
-    for (auto& name_pair : name_map) {
-      for (auto& name : name_pair.second) {
-        auto it = ref_cnts->find(name);
-        if (it == ref_cnts->end()) continue;
-        if (--(it->second) != 0) {
-          continue;
-        }
-        auto* var = scope.FindVar(name);
-        if (var == nullptr) {
-          continue;
-        }
-
-        VLOG(2) << "Erase variable " << name;
-        if (var->IsType<LoDTensor>()) {
-          garbages.emplace_back(
-              var->GetMutable<LoDTensor>()->MoveMemoryHolder());
-        } else if (var->IsType<SelectedRows>()) {
-          garbages.emplace_back(var->GetMutable<SelectedRows>()
-                                    ->mutable_value()
-                                    ->MoveMemoryHolder());
-        } else if (var->IsType<LoDTensorArray>()) {
-          auto* lod_tensor_arr = var->GetMutable<LoDTensorArray>();
-          for (auto& t : *lod_tensor_arr) {
-            garbages.emplace_back(t.MoveMemoryHolder());
-          }
-        } else {
-          PADDLE_THROW("Type %s of %s is not supported eager deletion",
-                       framework::ToTypeName(var->Type()), name);
-        }
-      }
-    }
-  };
-
-  handler(op->Inputs());
-  handler(op->Outputs());
-
-  if (!garbages.empty()) {
-    gc->Add(std::move(garbages));
-  }
-}
-
 Executor::Executor(const platform::Place& place) : place_(place) {}
 
 void Executor::Close() {
@@ -362,8 +289,8 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
 std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
     const ProgramDesc& program, int block_id,
     const std::vector<std::string>& skip_ref_cnt_vars, bool force_disable_gc) {
-  std::unique_ptr<ExecutorPrepareContext> ctx(new ExecutorPrepareContext(
-      program, block_id, skip_ref_cnt_vars, force_disable_gc));
+  std::unique_ptr<ExecutorPrepareContext> ctx(
+      new ExecutorPrepareContext(program, block_id));
   PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
   auto& block = program.Block(block_id);
   for (auto& op_desc : block.AllOps()) {
@@ -375,6 +302,7 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
         ctx->prog_.Block(ctx->block_id_), &ctx->ops_);
   }
 #endif
+  ctx->PrepareUnusedVars(skip_ref_cnt_vars, force_disable_gc);
   return ctx;
 }
 
@@ -389,19 +317,17 @@ std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
   std::vector<std::shared_ptr<ExecutorPrepareContext>> result;
   size_t idx = 0;
   for (auto& bid : block_ids) {
-    ExecutorPrepareContext* ctx;
-    if (skip_ref_cnt_vars.empty()) {
-      ctx = new ExecutorPrepareContext(program, bid, std::vector<std::string>(),
-                                       force_disable_gc);
-    } else {
-      ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx],
-                                       force_disable_gc);
-    }
     PADDLE_ENFORCE_LT(static_cast<size_t>(bid), program.Size());
+    auto* ctx = new ExecutorPrepareContext(program, bid);
     auto& block = program.Block(bid);
     for (auto& op_desc : block.AllOps()) {
       ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
     }
+    if (skip_ref_cnt_vars.empty()) {
+      ctx->PrepareUnusedVars(std::vector<std::string>(), force_disable_gc);
+    } else {
+      ctx->PrepareUnusedVars(skip_ref_cnt_vars[idx], force_disable_gc);
+    }
     result.push_back(std::shared_ptr<ExecutorPrepareContext>(ctx));
     ++idx;
   }
@@ -425,7 +351,6 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
   // FIXME(zjl): recurrent_op is rather complex, we would
   // disable gc forcely in recurrent_op
   if (!ctx->force_disable_gc_ && max_memory_size >= 0) {
-    ctx->ResetReferenceCount();
 #ifdef PADDLE_WITH_CUDA
     if (platform::is_gpu_place(place_)) {
       if (IsFastEagerDeletionModeEnabled()) {
@@ -453,8 +378,7 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
     op->Run(*local_scope, place_);
 
     if (gc) {
-      DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
-                          &(ctx->runtime_ref_cnts_));
+      DeleteUnusedTensors(*local_scope, op.get(), ctx->unused_vars_, gc.get());
     }
   }
 

paddle/fluid/framework/executor.h

@@ -30,22 +30,20 @@ namespace paddle {
 namespace framework {
 
 struct ExecutorPrepareContext {
-  ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id,
-                         const std::vector<std::string>& skip_ref_cnt_vars =
-                             std::vector<std::string>(),
-                         bool force_disable_gc = false);
+  ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id);
 
   ~ExecutorPrepareContext();
 
-  void ResetReferenceCount() { runtime_ref_cnts_ = global_ref_cnts_; }
+  void PrepareUnusedVars(const std::vector<std::string>& keep_vars,
+                         bool force_disable_gc = false);
 
   const framework::ProgramDesc& prog_;
-  size_t block_id_;
-  bool force_disable_gc_;
+  const size_t block_id_;
+
   std::vector<std::unique_ptr<OperatorBase>> ops_;
 
-  std::unordered_map<std::string, size_t> global_ref_cnts_;
-  std::unordered_map<std::string, size_t> runtime_ref_cnts_;
+  std::unordered_map<OperatorBase*, std::vector<std::string>> unused_vars_;
+  bool force_disable_gc_{false};
 };
 
 class Executor {