Merge remote-tracking branch 'ups/develop' into refine/im2col

Author: tensor-tang

AUTHORS.md

@@ -46,6 +46,7 @@
 | tianbingsz | Tian-Bing Xu |
 | tpatejko | Tomasz Patejko |
 | typhoonzero | Yi Wu |
+| velconia | Qi-Yang Min |
 | wanghaoshuang | Hao-Shuang Wang |
 | wangyang59 | Yang Wang |
 | wangzhen-nlp | Zhen Wang |

benchmark/fluid/fluid_benchmark.py

@@ -85,8 +85,7 @@ def dist_transpile(trainer_id, args):
         trainer_id,
         pservers=pserver_endpoints,
         trainers=trainers,
-        sync_mode=not args.async_mode,
-        slice_var_up=not args.no_split_var)
+        sync_mode=not args.async_mode)
     if training_role == "PSERVER":
         pserver_program = t.get_pserver_program(current_endpoint)
         pserver_startup_program = t.get_startup_program(current_endpoint,

cmake/external/grpc.cmake

@@ -50,7 +50,7 @@ ExternalProject_Add(
     UPDATE_COMMAND  ""
     CONFIGURE_COMMAND ""
     BUILD_IN_SOURCE 1
-    PATCH_COMMAND git apply ${PADDLE_SOURCE_DIR}/patches/grpc/fix_too_early_destory.patch
+    PATCH_COMMAND cp ${PADDLE_SOURCE_DIR}/patches/grpc/grpc_library.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/grpc_library.h && cp ${PADDLE_SOURCE_DIR}/patches/grpc/completion_queue.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/completion_queue.h
     # NOTE(yuyang18):
     # Disable -Werror, otherwise the compile will fail in MacOS.
     # It seems that we cannot configure that by make command.

cmake/inference_lib.cmake

@@ -148,18 +148,11 @@ if (WITH_ANAKIN AND WITH_GPU)
      list(APPEND inference_deps anakin_inference_lib)
 endif()
 
-copy(inference_api_lib DEPS paddle_inference_api paddle_inference_api_shared
-  SRCS ${src_dir}/${module}/paddle_inference_api.h 
-       ${src_dir}/${module}/demo_ci
-       ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libpaddle_inference_api*
-  DSTS ${dst_dir}/inference ${dst_dir}/inference ${dst_dir}/inference
-)
-list(APPEND inference_deps inference_api_lib)
-
 set(module "inference")
 copy(inference_lib DEPS ${inference_deps}
   SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
-  DSTS ${dst_dir}/${module} ${dst_dir}/${module}
+       ${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci
+  DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
 )
 
 set(module "platform")

paddle/fluid/framework/CMakeLists.txt

@@ -8,9 +8,9 @@ cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
 cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context)
 if(WITH_GPU)
-  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type)
+  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context)
 else()
-  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type)
+  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context)
 endif()
 
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
@@ -110,7 +110,7 @@ cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
 cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
-      
+
 # cc_test(channel_test SRCS channel_test.cc)
 cc_test(tuple_test SRCS tuple_test.cc )
 

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -715,6 +715,7 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
       result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
       node->Op()->Type(), places_[op_dev_id]));
 
+  // TODO(panyx0718): This might not be needed anymore.
   if (node->Op()->Type() == "send_barrier") {
     ConnectOp(result, result->Get<GraphOps>("ops").back().get(), "send");
   } else if (node->Op()->Type() == "recv") {

paddle/fluid/framework/ir/graph.cc

@@ -24,6 +24,68 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
+std::vector<std::string> FindDistTrainSendVars(
+    const std::vector<ir::Node *> &nodes) {
+  std::vector<std::string> send_vars;
+  // since parameters are all in block 0,
+  // it's enough to only scan send ops in block 0
+  for (auto &node : nodes) {
+    auto op_vars = node->Op()->InputArgumentNames();
+    send_vars.reserve(send_vars.size() +
+                      std::distance(op_vars.begin(), op_vars.end()));
+    send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end());
+  }
+  return send_vars;
+}
+
+std::vector<std::string> FindDistTrainRecvVars(
+    const std::vector<ir::Node *> &nodes) {
+  std::vector<std::string> recv_vars;
+  for (auto &node : nodes) {
+    auto op_vars = node->Op()->OutputArgumentNames();
+    recv_vars.reserve(recv_vars.size() +
+                      std::distance(op_vars.begin(), op_vars.end()));
+    recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end());
+  }
+  return recv_vars;
+}
+
+bool IsDistTrainOp(ir::Node *node, const std::vector<std::string> &send_vars,
+                   const std::vector<std::string> &recv_vars) {
+  if (send_vars.size() == 0 || recv_vars.size() == 0) {
+    return false;
+  }
+
+  /**
+   * Check any of opvars contains `.block` and in sendvars
+   */
+  auto checker = [](const std::vector<std::string> &opvars,
+                    const std::vector<std::string> &rpc_vars) -> bool {
+    for (auto &var : opvars) {
+      // a variable name with the suffix `.block` means it's a splited
+      // variable by (DistributeTranspiler)
+      // [python/paddle/fluid/transpiler/distribute_transpiler.py]
+      if (var.find(".block") != std::string::npos &&
+          std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) {
+        return true;
+      }
+    }
+    return false;
+  };
+
+  std::vector<std::string> input_var_names;
+  std::vector<std::string> output_var_names;
+  for (ir::Node *input : node->inputs) {
+    input_var_names.push_back(input->Name());
+  }
+  for (ir::Node *output : node->outputs) {
+    output_var_names.push_back(output->Name());
+  }
+
+  return checker(output_var_names, send_vars) ||
+         checker(input_var_names, recv_vars);
+}
+
 Graph::Graph(const ProgramDesc &program) : program_(program) {
   VLOG(3) << "block in program:" << program_.Size();
   std::unordered_map<std::string, VarDesc *> all_vars;
@@ -61,6 +123,64 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
       var->inputs.push_back(node);
     }
   }
+
+  std::vector<ir::Node *> send_ops;
+  ir::Node *send_bar = nullptr;
+  std::vector<ir::Node *> recv_ops;
+  ir::Node *fetch_bar = nullptr;
+  for (ir::Node *node : Nodes()) {
+    if (node->Name() == "send") {
+      send_ops.push_back(node);
+    } else if (node->Name() == "send_barrier") {
+      PADDLE_ENFORCE(!send_bar, "only has one send barrier");
+      send_bar = node;
+    } else if (node->Name() == "recv") {
+      recv_ops.push_back(node);
+    } else if (node->Name() == "fetch_barrier") {
+      PADDLE_ENFORCE(!fetch_bar, "only has one fetch barrier");
+      fetch_bar = node;
+    }
+  }
+  if (send_bar) {
+    for (ir::Node *send : send_ops) {
+      ir::Node *dep_var = CreateControlDepVar();
+      send->outputs.push_back(dep_var);
+      dep_var->inputs.push_back(send);
+      send_bar->inputs.push_back(dep_var);
+      dep_var->outputs.push_back(send_bar);
+    }
+    for (ir::Node *recv : recv_ops) {
+      ir::Node *dep_var = CreateControlDepVar();
+      recv->inputs.push_back(dep_var);
+      dep_var->outputs.push_back(recv);
+      send_bar->outputs.push_back(dep_var);
+      dep_var->inputs.push_back(send_bar);
+    }
+  }
+  if (fetch_bar) {
+    for (ir::Node *recv : recv_ops) {
+      ir::Node *dep_var = CreateControlDepVar();
+      recv->outputs.push_back(dep_var);
+      dep_var->inputs.push_back(recv);
+      fetch_bar->inputs.push_back(dep_var);
+      dep_var->outputs.push_back(fetch_bar);
+    }
+  }
+
+  std::vector<std::string> send_vars = FindDistTrainSendVars(send_ops);
+  std::vector<std::string> recv_vars = FindDistTrainRecvVars(recv_ops);
+  for (ir::Node *node : Nodes()) {
+    if (IsDistTrainOp(node, send_vars, recv_vars)) {
+      if (fetch_bar && node->Name() == "concat") {
+        ir::Node *dep_var = CreateControlDepVar();
+        fetch_bar->outputs.push_back(dep_var);
+        dep_var->inputs.push_back(fetch_bar);
+        node->inputs.push_back(dep_var);
+        dep_var->outputs.push_back(node);
+      }
+    }
+  }
+
   /**
    * We only handle write after read(WAR), since it should not have a write
    * after write in program. If there are write after write operators, we need

paddle/fluid/inference/CMakeLists.txt

@@ -14,8 +14,15 @@ cc_library(paddle_fluid_api
 
 get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
 
+# paddle_fluid_origin exclude inference api interface
+cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api)
+
+if(NOT APPLE)
+  add_subdirectory(api)
+endif()
+
 # Create static library
-cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api)
+cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api paddle_inference_api)
 if(NOT APPLE)
   # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
   set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.sym")
@@ -24,20 +31,29 @@ endif()
 
 # Create shared library
 cc_library(paddle_fluid_shared SHARED
-    SRCS io.cc
+    SRCS io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc
     DEPS ${fluid_modules} paddle_fluid_api)
 
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
 if(NOT APPLE)
   # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.
   set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map")
   set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
+  # check symbol hidden
+  FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake
+    "execute_process(COMMAND bash -c \"${CMAKE_CURRENT_SOURCE_DIR}/check_symbol.sh"
+    " ${CMAKE_CURRENT_BINARY_DIR}/libpaddle_fluid.so\" RESULT_VARIABLE symbol_res)\n"
+    "if(NOT \"\${symbol_res}\" STREQUAL \"0\")\n"
+    "  message(FATAL_ERROR \"Check symbol failed.\")\n"
+    "endif()\n")
+  add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol"
+    COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake"
+    DEPENDS paddle_fluid_shared)
+  add_custom_target(check_symbol ALL DEPENDS "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol")
 endif()
 
 if(WITH_TESTING)
   # both tests/book and analysis depends the models that generated by python/paddle/fluid/tests/book
   add_subdirectory(tests/book)
 endif()
-if(NOT APPLE)
-  add_subdirectory(api)
-endif()

paddle/fluid/inference/api/CMakeLists.txt

@@ -42,35 +42,8 @@ function(inference_api_test TARGET_NAME)
     endif(WITH_TESTING)
 endfunction(inference_api_test)
 
-cc_library(paddle_inference_api
-    SRCS api.cc api_impl.cc
-    DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
-if(NOT APPLE)
-  set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/api.sym")
-  set_target_properties(paddle_inference_api PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
-endif()
-
-# Here the shared library doesn't depend on other fluid libraries, or double free will occur.
-cc_library(paddle_inference_api_shared SHARED
-    SRCS api.cc api_impl.cc)
-add_dependencies(paddle_inference_api_shared ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
-set_target_properties(paddle_inference_api_shared PROPERTIES OUTPUT_NAME paddle_inference_api)
+cc_library(paddle_inference_api SRCS api.cc api_impl.cc DEPS lod_tensor)
 
-if(NOT APPLE)
-  set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/api.map")
-  set_target_properties(paddle_inference_api_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
-  FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake
-    "execute_process(COMMAND bash -c \"${CMAKE_CURRENT_SOURCE_DIR}/check_symbol.sh"
-    " ${CMAKE_CURRENT_BINARY_DIR}/libpaddle_inference_api.so\" RESULT_VARIABLE symbol_res)\n"
-    "if(NOT \"\${symbol_res}\" STREQUAL \"0\")\n"
-    "  message(FATAL_ERROR \"Check symbol failed.\")\n"
-    "endif()\n")
-  add_custom_command(
-    OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol"
-    COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake"
-    DEPENDS paddle_inference_api_shared)
-  add_custom_target(check_symbol ALL DEPENDS "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol")
-endif()
 
 cc_test(test_paddle_inference_api
         SRCS api_tester.cc